]>
Commit | Line | Data |
---|---|---|
a497ee34 | 1 | // SPDX-License-Identifier: GPL-2.0-or-later |
aee69d78 PV |
2 | /* |
3 | * Budget Fair Queueing (BFQ) I/O scheduler. | |
4 | * | |
5 | * Based on ideas and code from CFQ: | |
6 | * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk> | |
7 | * | |
8 | * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it> | |
9 | * Paolo Valente <paolo.valente@unimore.it> | |
10 | * | |
11 | * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it> | |
12 | * Arianna Avanzini <avanzini@google.com> | |
13 | * | |
14 | * Copyright (C) 2017 Paolo Valente <paolo.valente@linaro.org> | |
15 | * | |
aee69d78 PV |
16 | * BFQ is a proportional-share I/O scheduler, with some extra |
17 | * low-latency capabilities. BFQ also supports full hierarchical | |
18 | * scheduling through cgroups. Next paragraphs provide an introduction | |
19 | * on BFQ inner workings. Details on BFQ benefits, usage and | |
20 | * limitations can be found in Documentation/block/bfq-iosched.txt. | |
21 | * | |
22 | * BFQ is a proportional-share storage-I/O scheduling algorithm based | |
23 | * on the slice-by-slice service scheme of CFQ. But BFQ assigns | |
24 | * budgets, measured in number of sectors, to processes instead of | |
25 | * time slices. The device is not granted to the in-service process | |
26 | * for a given time slice, but until it has exhausted its assigned | |
27 | * budget. This change from the time to the service domain enables BFQ | |
28 | * to distribute the device throughput among processes as desired, | |
29 | * without any distortion due to throughput fluctuations, or to device | |
30 | * internal queueing. BFQ uses an ad hoc internal scheduler, called | |
31 | * B-WF2Q+, to schedule processes according to their budgets. More | |
32 | * precisely, BFQ schedules queues associated with processes. Each | |
33 | * process/queue is assigned a user-configurable weight, and B-WF2Q+ | |
34 | * guarantees that each queue receives a fraction of the throughput | |
35 | * proportional to its weight. Thanks to the accurate policy of | |
36 | * B-WF2Q+, BFQ can afford to assign high budgets to I/O-bound | |
37 | * processes issuing sequential requests (to boost the throughput), | |
38 | * and yet guarantee a low latency to interactive and soft real-time | |
39 | * applications. | |
40 | * | |
41 | * In particular, to provide these low-latency guarantees, BFQ | |
42 | * explicitly privileges the I/O of two classes of time-sensitive | |
4029eef1 PV |
43 | * applications: interactive and soft real-time. In more detail, BFQ |
44 | * behaves this way if the low_latency parameter is set (default | |
45 | * configuration). This feature enables BFQ to provide applications in | |
46 | * these classes with a very low latency. | |
47 | * | |
48 | * To implement this feature, BFQ constantly tries to detect whether | |
49 | * the I/O requests in a bfq_queue come from an interactive or a soft | |
50 | * real-time application. For brevity, in these cases, the queue is | |
51 | * said to be interactive or soft real-time. In both cases, BFQ | |
52 | * privileges the service of the queue, over that of non-interactive | |
53 | * and non-soft-real-time queues. This privileging is performed, | |
54 | * mainly, by raising the weight of the queue. So, for brevity, we | |
55 | * call just weight-raising periods the time periods during which a | |
56 | * queue is privileged, because deemed interactive or soft real-time. | |
57 | * | |
58 | * The detection of soft real-time queues/applications is described in | |
59 | * detail in the comments on the function | |
60 | * bfq_bfqq_softrt_next_start. On the other hand, the detection of an | |
61 | * interactive queue works as follows: a queue is deemed interactive | |
62 | * if it is constantly non empty only for a limited time interval, | |
63 | * after which it does become empty. The queue may be deemed | |
64 | * interactive again (for a limited time), if it restarts being | |
65 | * constantly non empty, provided that this happens only after the | |
66 | * queue has remained empty for a given minimum idle time. | |
67 | * | |
68 | * By default, BFQ computes automatically the above maximum time | |
69 | * interval, i.e., the time interval after which a constantly | |
70 | * non-empty queue stops being deemed interactive. Since a queue is | |
71 | * weight-raised while it is deemed interactive, this maximum time | |
72 | * interval happens to coincide with the (maximum) duration of the | |
73 | * weight-raising for interactive queues. | |
74 | * | |
75 | * Finally, BFQ also features additional heuristics for | |
aee69d78 PV |
76 | * preserving both a low latency and a high throughput on NCQ-capable, |
77 | * rotational or flash-based devices, and to get the job done quickly | |
78 | * for applications consisting in many I/O-bound processes. | |
79 | * | |
43c1b3d6 PV |
80 | * NOTE: if the main or only goal, with a given device, is to achieve |
81 | * the maximum-possible throughput at all times, then do switch off | |
82 | * all low-latency heuristics for that device, by setting low_latency | |
83 | * to 0. | |
84 | * | |
4029eef1 PV |
85 | * BFQ is described in [1], where also a reference to the initial, |
86 | * more theoretical paper on BFQ can be found. The interested reader | |
87 | * can find in the latter paper full details on the main algorithm, as | |
88 | * well as formulas of the guarantees and formal proofs of all the | |
89 | * properties. With respect to the version of BFQ presented in these | |
90 | * papers, this implementation adds a few more heuristics, such as the | |
91 | * ones that guarantee a low latency to interactive and soft real-time | |
92 | * applications, and a hierarchical extension based on H-WF2Q+. | |
aee69d78 PV |
93 | * |
94 | * B-WF2Q+ is based on WF2Q+, which is described in [2], together with | |
95 | * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+ | |
96 | * with O(log N) complexity derives from the one introduced with EEVDF | |
97 | * in [3]. | |
98 | * | |
99 | * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O | |
100 | * Scheduler", Proceedings of the First Workshop on Mobile System | |
101 | * Technologies (MST-2015), May 2015. | |
102 | * http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf | |
103 | * | |
104 | * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing | |
105 | * Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689, | |
106 | * Oct 1997. | |
107 | * | |
108 | * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz | |
109 | * | |
110 | * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline | |
111 | * First: A Flexible and Accurate Mechanism for Proportional Share | |
112 | * Resource Allocation", technical report. | |
113 | * | |
114 | * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf | |
115 | */ | |
116 | #include <linux/module.h> | |
117 | #include <linux/slab.h> | |
118 | #include <linux/blkdev.h> | |
e21b7a0b | 119 | #include <linux/cgroup.h> |
aee69d78 PV |
120 | #include <linux/elevator.h> |
121 | #include <linux/ktime.h> | |
122 | #include <linux/rbtree.h> | |
123 | #include <linux/ioprio.h> | |
124 | #include <linux/sbitmap.h> | |
125 | #include <linux/delay.h> | |
126 | ||
127 | #include "blk.h" | |
128 | #include "blk-mq.h" | |
129 | #include "blk-mq-tag.h" | |
130 | #include "blk-mq-sched.h" | |
ea25da48 | 131 | #include "bfq-iosched.h" |
b5dc5d4d | 132 | #include "blk-wbt.h" |
aee69d78 | 133 | |
ea25da48 PV |
134 | #define BFQ_BFQQ_FNS(name) \ |
135 | void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \ | |
136 | { \ | |
137 | __set_bit(BFQQF_##name, &(bfqq)->flags); \ | |
138 | } \ | |
139 | void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \ | |
140 | { \ | |
141 | __clear_bit(BFQQF_##name, &(bfqq)->flags); \ | |
142 | } \ | |
143 | int bfq_bfqq_##name(const struct bfq_queue *bfqq) \ | |
144 | { \ | |
145 | return test_bit(BFQQF_##name, &(bfqq)->flags); \ | |
44e44a1b PV |
146 | } |
147 | ||
ea25da48 PV |
148 | BFQ_BFQQ_FNS(just_created); |
149 | BFQ_BFQQ_FNS(busy); | |
150 | BFQ_BFQQ_FNS(wait_request); | |
151 | BFQ_BFQQ_FNS(non_blocking_wait_rq); | |
152 | BFQ_BFQQ_FNS(fifo_expire); | |
d5be3fef | 153 | BFQ_BFQQ_FNS(has_short_ttime); |
ea25da48 PV |
154 | BFQ_BFQQ_FNS(sync); |
155 | BFQ_BFQQ_FNS(IO_bound); | |
156 | BFQ_BFQQ_FNS(in_large_burst); | |
157 | BFQ_BFQQ_FNS(coop); | |
158 | BFQ_BFQQ_FNS(split_coop); | |
159 | BFQ_BFQQ_FNS(softrt_update); | |
160 | #undef BFQ_BFQQ_FNS \ | |
aee69d78 | 161 | |
ea25da48 PV |
162 | /* Expiration time of sync (0) and async (1) requests, in ns. */ |
163 | static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 }; | |
aee69d78 | 164 | |
ea25da48 PV |
165 | /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */ |
166 | static const int bfq_back_max = 16 * 1024; | |
aee69d78 | 167 | |
ea25da48 PV |
168 | /* Penalty of a backwards seek, in number of sectors. */ |
169 | static const int bfq_back_penalty = 2; | |
e21b7a0b | 170 | |
ea25da48 PV |
171 | /* Idling period duration, in ns. */ |
172 | static u64 bfq_slice_idle = NSEC_PER_SEC / 125; | |
aee69d78 | 173 | |
ea25da48 PV |
174 | /* Minimum number of assigned budgets for which stats are safe to compute. */ |
175 | static const int bfq_stats_min_budgets = 194; | |
aee69d78 | 176 | |
ea25da48 PV |
177 | /* Default maximum budget values, in sectors and number of requests. */ |
178 | static const int bfq_default_max_budget = 16 * 1024; | |
e21b7a0b | 179 | |
ea25da48 | 180 | /* |
d5801088 PV |
181 | * When a sync request is dispatched, the queue that contains that |
182 | * request, and all the ancestor entities of that queue, are charged | |
636b8fe8 | 183 | * with the number of sectors of the request. In contrast, if the |
d5801088 PV |
184 | * request is async, then the queue and its ancestor entities are |
185 | * charged with the number of sectors of the request, multiplied by | |
186 | * the factor below. This throttles the bandwidth for async I/O, | |
187 | * w.r.t. to sync I/O, and it is done to counter the tendency of async | |
188 | * writes to steal I/O throughput to reads. | |
189 | * | |
190 | * The current value of this parameter is the result of a tuning with | |
191 | * several hardware and software configurations. We tried to find the | |
192 | * lowest value for which writes do not cause noticeable problems to | |
193 | * reads. In fact, the lower this parameter, the stabler I/O control, | |
194 | * in the following respect. The lower this parameter is, the less | |
195 | * the bandwidth enjoyed by a group decreases | |
196 | * - when the group does writes, w.r.t. to when it does reads; | |
197 | * - when other groups do reads, w.r.t. to when they do writes. | |
ea25da48 | 198 | */ |
d5801088 | 199 | static const int bfq_async_charge_factor = 3; |
aee69d78 | 200 | |
ea25da48 PV |
201 | /* Default timeout values, in jiffies, approximating CFQ defaults. */ |
202 | const int bfq_timeout = HZ / 8; | |
aee69d78 | 203 | |
7b8fa3b9 PV |
204 | /* |
205 | * Time limit for merging (see comments in bfq_setup_cooperator). Set | |
206 | * to the slowest value that, in our tests, proved to be effective in | |
207 | * removing false positives, while not causing true positives to miss | |
208 | * queue merging. | |
209 | * | |
210 | * As can be deduced from the low time limit below, queue merging, if | |
636b8fe8 | 211 | * successful, happens at the very beginning of the I/O of the involved |
7b8fa3b9 PV |
212 | * cooperating processes, as a consequence of the arrival of the very |
213 | * first requests from each cooperator. After that, there is very | |
214 | * little chance to find cooperators. | |
215 | */ | |
216 | static const unsigned long bfq_merge_time_limit = HZ/10; | |
217 | ||
ea25da48 | 218 | static struct kmem_cache *bfq_pool; |
e21b7a0b | 219 | |
ea25da48 PV |
220 | /* Below this threshold (in ns), we consider thinktime immediate. */ |
221 | #define BFQ_MIN_TT (2 * NSEC_PER_MSEC) | |
e21b7a0b | 222 | |
ea25da48 | 223 | /* hw_tag detection: parallel requests threshold and min samples needed. */ |
a3c92560 | 224 | #define BFQ_HW_QUEUE_THRESHOLD 3 |
ea25da48 | 225 | #define BFQ_HW_QUEUE_SAMPLES 32 |
aee69d78 | 226 | |
ea25da48 PV |
227 | #define BFQQ_SEEK_THR (sector_t)(8 * 100) |
228 | #define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32) | |
d87447d8 PV |
229 | #define BFQ_RQ_SEEKY(bfqd, last_pos, rq) \ |
230 | (get_sdist(last_pos, rq) > \ | |
231 | BFQQ_SEEK_THR && \ | |
232 | (!blk_queue_nonrot(bfqd->queue) || \ | |
233 | blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT)) | |
ea25da48 | 234 | #define BFQQ_CLOSE_THR (sector_t)(8 * 1024) |
f0ba5ea2 | 235 | #define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 19) |
7074f076 PV |
236 | /* |
237 | * Sync random I/O is likely to be confused with soft real-time I/O, | |
238 | * because it is characterized by limited throughput and apparently | |
239 | * isochronous arrival pattern. To avoid false positives, queues | |
240 | * containing only random (seeky) I/O are prevented from being tagged | |
241 | * as soft real-time. | |
242 | */ | |
243 | #define BFQQ_TOTALLY_SEEKY(bfqq) (bfqq->seek_history & -1) | |
aee69d78 | 244 | |
ea25da48 PV |
245 | /* Min number of samples required to perform peak-rate update */ |
246 | #define BFQ_RATE_MIN_SAMPLES 32 | |
247 | /* Min observation time interval required to perform a peak-rate update (ns) */ | |
248 | #define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC) | |
249 | /* Target observation time interval for a peak-rate update (ns) */ | |
250 | #define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC | |
aee69d78 | 251 | |
bc56e2ca PV |
252 | /* |
253 | * Shift used for peak-rate fixed precision calculations. | |
254 | * With | |
255 | * - the current shift: 16 positions | |
256 | * - the current type used to store rate: u32 | |
257 | * - the current unit of measure for rate: [sectors/usec], or, more precisely, | |
258 | * [(sectors/usec) / 2^BFQ_RATE_SHIFT] to take into account the shift, | |
259 | * the range of rates that can be stored is | |
260 | * [1 / 2^BFQ_RATE_SHIFT, 2^(32 - BFQ_RATE_SHIFT)] sectors/usec = | |
261 | * [1 / 2^16, 2^16] sectors/usec = [15e-6, 65536] sectors/usec = | |
262 | * [15, 65G] sectors/sec | |
263 | * Which, assuming a sector size of 512B, corresponds to a range of | |
264 | * [7.5K, 33T] B/sec | |
265 | */ | |
ea25da48 | 266 | #define BFQ_RATE_SHIFT 16 |
aee69d78 | 267 | |
ea25da48 | 268 | /* |
4029eef1 PV |
269 | * When configured for computing the duration of the weight-raising |
270 | * for interactive queues automatically (see the comments at the | |
271 | * beginning of this file), BFQ does it using the following formula: | |
e24f1c24 PV |
272 | * duration = (ref_rate / r) * ref_wr_duration, |
273 | * where r is the peak rate of the device, and ref_rate and | |
274 | * ref_wr_duration are two reference parameters. In particular, | |
275 | * ref_rate is the peak rate of the reference storage device (see | |
276 | * below), and ref_wr_duration is about the maximum time needed, with | |
277 | * BFQ and while reading two files in parallel, to load typical large | |
278 | * applications on the reference device (see the comments on | |
279 | * max_service_from_wr below, for more details on how ref_wr_duration | |
280 | * is obtained). In practice, the slower/faster the device at hand | |
281 | * is, the more/less it takes to load applications with respect to the | |
4029eef1 PV |
282 | * reference device. Accordingly, the longer/shorter BFQ grants |
283 | * weight raising to interactive applications. | |
ea25da48 | 284 | * |
e24f1c24 PV |
285 | * BFQ uses two different reference pairs (ref_rate, ref_wr_duration), |
286 | * depending on whether the device is rotational or non-rotational. | |
ea25da48 | 287 | * |
e24f1c24 PV |
288 | * In the following definitions, ref_rate[0] and ref_wr_duration[0] |
289 | * are the reference values for a rotational device, whereas | |
290 | * ref_rate[1] and ref_wr_duration[1] are the reference values for a | |
291 | * non-rotational device. The reference rates are not the actual peak | |
292 | * rates of the devices used as a reference, but slightly lower | |
293 | * values. The reason for using slightly lower values is that the | |
294 | * peak-rate estimator tends to yield slightly lower values than the | |
295 | * actual peak rate (it can yield the actual peak rate only if there | |
296 | * is only one process doing I/O, and the process does sequential | |
297 | * I/O). | |
ea25da48 | 298 | * |
e24f1c24 PV |
299 | * The reference peak rates are measured in sectors/usec, left-shifted |
300 | * by BFQ_RATE_SHIFT. | |
ea25da48 | 301 | */ |
e24f1c24 | 302 | static int ref_rate[2] = {14000, 33000}; |
ea25da48 | 303 | /* |
e24f1c24 PV |
304 | * To improve readability, a conversion function is used to initialize |
305 | * the following array, which entails that the array can be | |
306 | * initialized only in a function. | |
ea25da48 | 307 | */ |
e24f1c24 | 308 | static int ref_wr_duration[2]; |
aee69d78 | 309 | |
8a8747dc PV |
310 | /* |
311 | * BFQ uses the above-detailed, time-based weight-raising mechanism to | |
312 | * privilege interactive tasks. This mechanism is vulnerable to the | |
313 | * following false positives: I/O-bound applications that will go on | |
314 | * doing I/O for much longer than the duration of weight | |
315 | * raising. These applications have basically no benefit from being | |
316 | * weight-raised at the beginning of their I/O. On the opposite end, | |
317 | * while being weight-raised, these applications | |
318 | * a) unjustly steal throughput to applications that may actually need | |
319 | * low latency; | |
320 | * b) make BFQ uselessly perform device idling; device idling results | |
321 | * in loss of device throughput with most flash-based storage, and may | |
322 | * increase latencies when used purposelessly. | |
323 | * | |
324 | * BFQ tries to reduce these problems, by adopting the following | |
325 | * countermeasure. To introduce this countermeasure, we need first to | |
326 | * finish explaining how the duration of weight-raising for | |
327 | * interactive tasks is computed. | |
328 | * | |
329 | * For a bfq_queue deemed as interactive, the duration of weight | |
330 | * raising is dynamically adjusted, as a function of the estimated | |
331 | * peak rate of the device, so as to be equal to the time needed to | |
332 | * execute the 'largest' interactive task we benchmarked so far. By | |
333 | * largest task, we mean the task for which each involved process has | |
334 | * to do more I/O than for any of the other tasks we benchmarked. This | |
335 | * reference interactive task is the start-up of LibreOffice Writer, | |
336 | * and in this task each process/bfq_queue needs to have at most ~110K | |
337 | * sectors transferred. | |
338 | * | |
339 | * This last piece of information enables BFQ to reduce the actual | |
340 | * duration of weight-raising for at least one class of I/O-bound | |
341 | * applications: those doing sequential or quasi-sequential I/O. An | |
342 | * example is file copy. In fact, once started, the main I/O-bound | |
343 | * processes of these applications usually consume the above 110K | |
344 | * sectors in much less time than the processes of an application that | |
345 | * is starting, because these I/O-bound processes will greedily devote | |
346 | * almost all their CPU cycles only to their target, | |
347 | * throughput-friendly I/O operations. This is even more true if BFQ | |
348 | * happens to be underestimating the device peak rate, and thus | |
349 | * overestimating the duration of weight raising. But, according to | |
350 | * our measurements, once transferred 110K sectors, these processes | |
351 | * have no right to be weight-raised any longer. | |
352 | * | |
353 | * Basing on the last consideration, BFQ ends weight-raising for a | |
354 | * bfq_queue if the latter happens to have received an amount of | |
355 | * service at least equal to the following constant. The constant is | |
356 | * set to slightly more than 110K, to have a minimum safety margin. | |
357 | * | |
358 | * This early ending of weight-raising reduces the amount of time | |
359 | * during which interactive false positives cause the two problems | |
360 | * described at the beginning of these comments. | |
361 | */ | |
362 | static const unsigned long max_service_from_wr = 120000; | |
363 | ||
12cd3a2f | 364 | #define RQ_BIC(rq) icq_to_bic((rq)->elv.priv[0]) |
ea25da48 | 365 | #define RQ_BFQQ(rq) ((rq)->elv.priv[1]) |
aee69d78 | 366 | |
ea25da48 | 367 | struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync) |
e21b7a0b | 368 | { |
ea25da48 | 369 | return bic->bfqq[is_sync]; |
aee69d78 PV |
370 | } |
371 | ||
ea25da48 | 372 | void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync) |
aee69d78 | 373 | { |
ea25da48 | 374 | bic->bfqq[is_sync] = bfqq; |
aee69d78 PV |
375 | } |
376 | ||
ea25da48 | 377 | struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic) |
aee69d78 | 378 | { |
ea25da48 | 379 | return bic->icq.q->elevator->elevator_data; |
e21b7a0b | 380 | } |
aee69d78 | 381 | |
ea25da48 PV |
382 | /** |
383 | * icq_to_bic - convert iocontext queue structure to bfq_io_cq. | |
384 | * @icq: the iocontext queue. | |
385 | */ | |
386 | static struct bfq_io_cq *icq_to_bic(struct io_cq *icq) | |
e21b7a0b | 387 | { |
ea25da48 PV |
388 | /* bic->icq is the first member, %NULL will convert to %NULL */ |
389 | return container_of(icq, struct bfq_io_cq, icq); | |
e21b7a0b | 390 | } |
aee69d78 | 391 | |
ea25da48 PV |
392 | /** |
393 | * bfq_bic_lookup - search into @ioc a bic associated to @bfqd. | |
394 | * @bfqd: the lookup key. | |
395 | * @ioc: the io_context of the process doing I/O. | |
396 | * @q: the request queue. | |
397 | */ | |
398 | static struct bfq_io_cq *bfq_bic_lookup(struct bfq_data *bfqd, | |
399 | struct io_context *ioc, | |
400 | struct request_queue *q) | |
e21b7a0b | 401 | { |
ea25da48 PV |
402 | if (ioc) { |
403 | unsigned long flags; | |
404 | struct bfq_io_cq *icq; | |
aee69d78 | 405 | |
0d945c1f | 406 | spin_lock_irqsave(&q->queue_lock, flags); |
ea25da48 | 407 | icq = icq_to_bic(ioc_lookup_icq(ioc, q)); |
0d945c1f | 408 | spin_unlock_irqrestore(&q->queue_lock, flags); |
aee69d78 | 409 | |
ea25da48 | 410 | return icq; |
e21b7a0b | 411 | } |
e21b7a0b | 412 | |
ea25da48 | 413 | return NULL; |
aee69d78 PV |
414 | } |
415 | ||
ea25da48 PV |
416 | /* |
417 | * Scheduler run of queue, if there are requests pending and no one in the | |
418 | * driver that will restart queueing. | |
419 | */ | |
420 | void bfq_schedule_dispatch(struct bfq_data *bfqd) | |
aee69d78 | 421 | { |
ea25da48 PV |
422 | if (bfqd->queued != 0) { |
423 | bfq_log(bfqd, "schedule dispatch"); | |
424 | blk_mq_run_hw_queues(bfqd->queue, true); | |
e21b7a0b | 425 | } |
aee69d78 PV |
426 | } |
427 | ||
428 | #define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE) | |
429 | #define bfq_class_rt(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_RT) | |
430 | ||
431 | #define bfq_sample_valid(samples) ((samples) > 80) | |
432 | ||
aee69d78 PV |
433 | /* |
434 | * Lifted from AS - choose which of rq1 and rq2 that is best served now. | |
636b8fe8 | 435 | * We choose the request that is closer to the head right now. Distance |
aee69d78 PV |
436 | * behind the head is penalized and only allowed to a certain extent. |
437 | */ | |
438 | static struct request *bfq_choose_req(struct bfq_data *bfqd, | |
439 | struct request *rq1, | |
440 | struct request *rq2, | |
441 | sector_t last) | |
442 | { | |
443 | sector_t s1, s2, d1 = 0, d2 = 0; | |
444 | unsigned long back_max; | |
445 | #define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */ | |
446 | #define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */ | |
447 | unsigned int wrap = 0; /* bit mask: requests behind the disk head? */ | |
448 | ||
449 | if (!rq1 || rq1 == rq2) | |
450 | return rq2; | |
451 | if (!rq2) | |
452 | return rq1; | |
453 | ||
454 | if (rq_is_sync(rq1) && !rq_is_sync(rq2)) | |
455 | return rq1; | |
456 | else if (rq_is_sync(rq2) && !rq_is_sync(rq1)) | |
457 | return rq2; | |
458 | if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META)) | |
459 | return rq1; | |
460 | else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META)) | |
461 | return rq2; | |
462 | ||
463 | s1 = blk_rq_pos(rq1); | |
464 | s2 = blk_rq_pos(rq2); | |
465 | ||
466 | /* | |
467 | * By definition, 1KiB is 2 sectors. | |
468 | */ | |
469 | back_max = bfqd->bfq_back_max * 2; | |
470 | ||
471 | /* | |
472 | * Strict one way elevator _except_ in the case where we allow | |
473 | * short backward seeks which are biased as twice the cost of a | |
474 | * similar forward seek. | |
475 | */ | |
476 | if (s1 >= last) | |
477 | d1 = s1 - last; | |
478 | else if (s1 + back_max >= last) | |
479 | d1 = (last - s1) * bfqd->bfq_back_penalty; | |
480 | else | |
481 | wrap |= BFQ_RQ1_WRAP; | |
482 | ||
483 | if (s2 >= last) | |
484 | d2 = s2 - last; | |
485 | else if (s2 + back_max >= last) | |
486 | d2 = (last - s2) * bfqd->bfq_back_penalty; | |
487 | else | |
488 | wrap |= BFQ_RQ2_WRAP; | |
489 | ||
490 | /* Found required data */ | |
491 | ||
492 | /* | |
493 | * By doing switch() on the bit mask "wrap" we avoid having to | |
494 | * check two variables for all permutations: --> faster! | |
495 | */ | |
496 | switch (wrap) { | |
497 | case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ | |
498 | if (d1 < d2) | |
499 | return rq1; | |
500 | else if (d2 < d1) | |
501 | return rq2; | |
502 | ||
503 | if (s1 >= s2) | |
504 | return rq1; | |
505 | else | |
506 | return rq2; | |
507 | ||
508 | case BFQ_RQ2_WRAP: | |
509 | return rq1; | |
510 | case BFQ_RQ1_WRAP: | |
511 | return rq2; | |
512 | case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */ | |
513 | default: | |
514 | /* | |
515 | * Since both rqs are wrapped, | |
516 | * start with the one that's further behind head | |
517 | * (--> only *one* back seek required), | |
518 | * since back seek takes more time than forward. | |
519 | */ | |
520 | if (s1 <= s2) | |
521 | return rq1; | |
522 | else | |
523 | return rq2; | |
524 | } | |
525 | } | |
526 | ||
a52a69ea PV |
527 | /* |
528 | * Async I/O can easily starve sync I/O (both sync reads and sync | |
529 | * writes), by consuming all tags. Similarly, storms of sync writes, | |
530 | * such as those that sync(2) may trigger, can starve sync reads. | |
531 | * Limit depths of async I/O and sync writes so as to counter both | |
532 | * problems. | |
533 | */ | |
534 | static void bfq_limit_depth(unsigned int op, struct blk_mq_alloc_data *data) | |
535 | { | |
a52a69ea | 536 | struct bfq_data *bfqd = data->q->elevator->elevator_data; |
a52a69ea PV |
537 | |
538 | if (op_is_sync(op) && !op_is_write(op)) | |
539 | return; | |
540 | ||
a52a69ea PV |
541 | data->shallow_depth = |
542 | bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(op)]; | |
543 | ||
544 | bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u", | |
545 | __func__, bfqd->wr_busy_queues, op_is_sync(op), | |
546 | data->shallow_depth); | |
547 | } | |
548 | ||
36eca894 AA |
549 | static struct bfq_queue * |
550 | bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root, | |
551 | sector_t sector, struct rb_node **ret_parent, | |
552 | struct rb_node ***rb_link) | |
553 | { | |
554 | struct rb_node **p, *parent; | |
555 | struct bfq_queue *bfqq = NULL; | |
556 | ||
557 | parent = NULL; | |
558 | p = &root->rb_node; | |
559 | while (*p) { | |
560 | struct rb_node **n; | |
561 | ||
562 | parent = *p; | |
563 | bfqq = rb_entry(parent, struct bfq_queue, pos_node); | |
564 | ||
565 | /* | |
566 | * Sort strictly based on sector. Smallest to the left, | |
567 | * largest to the right. | |
568 | */ | |
569 | if (sector > blk_rq_pos(bfqq->next_rq)) | |
570 | n = &(*p)->rb_right; | |
571 | else if (sector < blk_rq_pos(bfqq->next_rq)) | |
572 | n = &(*p)->rb_left; | |
573 | else | |
574 | break; | |
575 | p = n; | |
576 | bfqq = NULL; | |
577 | } | |
578 | ||
579 | *ret_parent = parent; | |
580 | if (rb_link) | |
581 | *rb_link = p; | |
582 | ||
583 | bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d", | |
584 | (unsigned long long)sector, | |
585 | bfqq ? bfqq->pid : 0); | |
586 | ||
587 | return bfqq; | |
588 | } | |
589 | ||
7b8fa3b9 PV |
590 | static bool bfq_too_late_for_merging(struct bfq_queue *bfqq) |
591 | { | |
592 | return bfqq->service_from_backlogged > 0 && | |
593 | time_is_before_jiffies(bfqq->first_IO_time + | |
594 | bfq_merge_time_limit); | |
595 | } | |
596 | ||
8cacc5ab PV |
597 | /* |
598 | * The following function is not marked as __cold because it is | |
599 | * actually cold, but for the same performance goal described in the | |
600 | * comments on the likely() at the beginning of | |
601 | * bfq_setup_cooperator(). Unexpectedly, to reach an even lower | |
602 | * execution time for the case where this function is not invoked, we | |
603 | * had to add an unlikely() in each involved if(). | |
604 | */ | |
605 | void __cold | |
606 | bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq) | |
36eca894 AA |
607 | { |
608 | struct rb_node **p, *parent; | |
609 | struct bfq_queue *__bfqq; | |
610 | ||
611 | if (bfqq->pos_root) { | |
612 | rb_erase(&bfqq->pos_node, bfqq->pos_root); | |
613 | bfqq->pos_root = NULL; | |
614 | } | |
615 | ||
7b8fa3b9 PV |
616 | /* |
617 | * bfqq cannot be merged any longer (see comments in | |
618 | * bfq_setup_cooperator): no point in adding bfqq into the | |
619 | * position tree. | |
620 | */ | |
621 | if (bfq_too_late_for_merging(bfqq)) | |
622 | return; | |
623 | ||
36eca894 AA |
624 | if (bfq_class_idle(bfqq)) |
625 | return; | |
626 | if (!bfqq->next_rq) | |
627 | return; | |
628 | ||
629 | bfqq->pos_root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree; | |
630 | __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root, | |
631 | blk_rq_pos(bfqq->next_rq), &parent, &p); | |
632 | if (!__bfqq) { | |
633 | rb_link_node(&bfqq->pos_node, parent, p); | |
634 | rb_insert_color(&bfqq->pos_node, bfqq->pos_root); | |
635 | } else | |
636 | bfqq->pos_root = NULL; | |
637 | } | |
638 | ||
1de0c4cd | 639 | /* |
fb53ac6c PV |
640 | * The following function returns false either if every active queue |
641 | * must receive the same share of the throughput (symmetric scenario), | |
642 | * or, as a special case, if bfqq must receive a share of the | |
643 | * throughput lower than or equal to the share that every other active | |
644 | * queue must receive. If bfqq does sync I/O, then these are the only | |
645 | * two cases where bfqq happens to be guaranteed its share of the | |
646 | * throughput even if I/O dispatching is not plugged when bfqq remains | |
647 | * temporarily empty (for more details, see the comments in the | |
648 | * function bfq_better_to_idle()). For this reason, the return value | |
649 | * of this function is used to check whether I/O-dispatch plugging can | |
650 | * be avoided. | |
1de0c4cd | 651 | * |
fb53ac6c | 652 | * The above first case (symmetric scenario) occurs when: |
1de0c4cd | 653 | * 1) all active queues have the same weight, |
73d58118 | 654 | * 2) all active queues belong to the same I/O-priority class, |
1de0c4cd | 655 | * 3) all active groups at the same level in the groups tree have the same |
73d58118 PV |
656 | * weight, |
657 | * 4) all active groups at the same level in the groups tree have the same | |
1de0c4cd AA |
658 | * number of children. |
659 | * | |
2d29c9f8 FM |
660 | * Unfortunately, keeping the necessary state for evaluating exactly |
661 | * the last two symmetry sub-conditions above would be quite complex | |
73d58118 PV |
662 | * and time consuming. Therefore this function evaluates, instead, |
663 | * only the following stronger three sub-conditions, for which it is | |
2d29c9f8 | 664 | * much easier to maintain the needed state: |
1de0c4cd | 665 | * 1) all active queues have the same weight, |
73d58118 PV |
666 | * 2) all active queues belong to the same I/O-priority class, |
667 | * 3) there are no active groups. | |
2d29c9f8 FM |
668 | * In particular, the last condition is always true if hierarchical |
669 | * support or the cgroups interface are not enabled, thus no state | |
670 | * needs to be maintained in this case. | |
1de0c4cd | 671 | */ |
fb53ac6c PV |
672 | static bool bfq_asymmetric_scenario(struct bfq_data *bfqd, |
673 | struct bfq_queue *bfqq) | |
1de0c4cd | 674 | { |
fb53ac6c PV |
675 | bool smallest_weight = bfqq && |
676 | bfqq->weight_counter && | |
677 | bfqq->weight_counter == | |
678 | container_of( | |
679 | rb_first_cached(&bfqd->queue_weights_tree), | |
680 | struct bfq_weight_counter, | |
681 | weights_node); | |
682 | ||
73d58118 PV |
683 | /* |
684 | * For queue weights to differ, queue_weights_tree must contain | |
685 | * at least two nodes. | |
686 | */ | |
fb53ac6c PV |
687 | bool varied_queue_weights = !smallest_weight && |
688 | !RB_EMPTY_ROOT(&bfqd->queue_weights_tree.rb_root) && | |
689 | (bfqd->queue_weights_tree.rb_root.rb_node->rb_left || | |
690 | bfqd->queue_weights_tree.rb_root.rb_node->rb_right); | |
73d58118 PV |
691 | |
692 | bool multiple_classes_busy = | |
693 | (bfqd->busy_queues[0] && bfqd->busy_queues[1]) || | |
694 | (bfqd->busy_queues[0] && bfqd->busy_queues[2]) || | |
695 | (bfqd->busy_queues[1] && bfqd->busy_queues[2]); | |
696 | ||
fb53ac6c | 697 | return varied_queue_weights || multiple_classes_busy |
42b1bd33 | 698 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
73d58118 PV |
699 | || bfqd->num_groups_with_pending_reqs > 0 |
700 | #endif | |
fb53ac6c | 701 | ; |
1de0c4cd AA |
702 | } |
703 | ||
704 | /* | |
705 | * If the weight-counter tree passed as input contains no counter for | |
2d29c9f8 | 706 | * the weight of the input queue, then add that counter; otherwise just |
1de0c4cd AA |
707 | * increment the existing counter. |
708 | * | |
709 | * Note that weight-counter trees contain few nodes in mostly symmetric | |
710 | * scenarios. For example, if all queues have the same weight, then the | |
711 | * weight-counter tree for the queues may contain at most one node. | |
712 | * This holds even if low_latency is on, because weight-raised queues | |
713 | * are not inserted in the tree. | |
714 | * In most scenarios, the rate at which nodes are created/destroyed | |
715 | * should be low too. | |
716 | */ | |
2d29c9f8 | 717 | void bfq_weights_tree_add(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
fb53ac6c | 718 | struct rb_root_cached *root) |
1de0c4cd | 719 | { |
2d29c9f8 | 720 | struct bfq_entity *entity = &bfqq->entity; |
fb53ac6c PV |
721 | struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL; |
722 | bool leftmost = true; | |
1de0c4cd AA |
723 | |
724 | /* | |
2d29c9f8 | 725 | * Do not insert if the queue is already associated with a |
1de0c4cd | 726 | * counter, which happens if: |
2d29c9f8 | 727 | * 1) a request arrival has caused the queue to become both |
1de0c4cd AA |
728 | * non-weight-raised, and hence change its weight, and |
729 | * backlogged; in this respect, each of the two events | |
730 | * causes an invocation of this function, | |
2d29c9f8 | 731 | * 2) this is the invocation of this function caused by the |
1de0c4cd AA |
732 | * second event. This second invocation is actually useless, |
733 | * and we handle this fact by exiting immediately. More | |
734 | * efficient or clearer solutions might possibly be adopted. | |
735 | */ | |
2d29c9f8 | 736 | if (bfqq->weight_counter) |
1de0c4cd AA |
737 | return; |
738 | ||
739 | while (*new) { | |
740 | struct bfq_weight_counter *__counter = container_of(*new, | |
741 | struct bfq_weight_counter, | |
742 | weights_node); | |
743 | parent = *new; | |
744 | ||
745 | if (entity->weight == __counter->weight) { | |
2d29c9f8 | 746 | bfqq->weight_counter = __counter; |
1de0c4cd AA |
747 | goto inc_counter; |
748 | } | |
749 | if (entity->weight < __counter->weight) | |
750 | new = &((*new)->rb_left); | |
fb53ac6c | 751 | else { |
1de0c4cd | 752 | new = &((*new)->rb_right); |
fb53ac6c PV |
753 | leftmost = false; |
754 | } | |
1de0c4cd AA |
755 | } |
756 | ||
2d29c9f8 FM |
757 | bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter), |
758 | GFP_ATOMIC); | |
1de0c4cd AA |
759 | |
760 | /* | |
761 | * In the unlucky event of an allocation failure, we just | |
2d29c9f8 | 762 | * exit. This will cause the weight of queue to not be |
fb53ac6c | 763 | * considered in bfq_asymmetric_scenario, which, in its turn, |
73d58118 PV |
764 | * causes the scenario to be deemed wrongly symmetric in case |
765 | * bfqq's weight would have been the only weight making the | |
766 | * scenario asymmetric. On the bright side, no unbalance will | |
767 | * however occur when bfqq becomes inactive again (the | |
768 | * invocation of this function is triggered by an activation | |
769 | * of queue). In fact, bfq_weights_tree_remove does nothing | |
770 | * if !bfqq->weight_counter. | |
1de0c4cd | 771 | */ |
2d29c9f8 | 772 | if (unlikely(!bfqq->weight_counter)) |
1de0c4cd AA |
773 | return; |
774 | ||
2d29c9f8 FM |
775 | bfqq->weight_counter->weight = entity->weight; |
776 | rb_link_node(&bfqq->weight_counter->weights_node, parent, new); | |
fb53ac6c PV |
777 | rb_insert_color_cached(&bfqq->weight_counter->weights_node, root, |
778 | leftmost); | |
1de0c4cd AA |
779 | |
780 | inc_counter: | |
2d29c9f8 | 781 | bfqq->weight_counter->num_active++; |
9dee8b3b | 782 | bfqq->ref++; |
1de0c4cd AA |
783 | } |
784 | ||
785 | /* | |
2d29c9f8 | 786 | * Decrement the weight counter associated with the queue, and, if the |
1de0c4cd AA |
787 | * counter reaches 0, remove the counter from the tree. |
788 | * See the comments to the function bfq_weights_tree_add() for considerations | |
789 | * about overhead. | |
790 | */ | |
0471559c | 791 | void __bfq_weights_tree_remove(struct bfq_data *bfqd, |
2d29c9f8 | 792 | struct bfq_queue *bfqq, |
fb53ac6c | 793 | struct rb_root_cached *root) |
1de0c4cd | 794 | { |
2d29c9f8 | 795 | if (!bfqq->weight_counter) |
1de0c4cd AA |
796 | return; |
797 | ||
2d29c9f8 FM |
798 | bfqq->weight_counter->num_active--; |
799 | if (bfqq->weight_counter->num_active > 0) | |
1de0c4cd AA |
800 | goto reset_entity_pointer; |
801 | ||
fb53ac6c | 802 | rb_erase_cached(&bfqq->weight_counter->weights_node, root); |
2d29c9f8 | 803 | kfree(bfqq->weight_counter); |
1de0c4cd AA |
804 | |
805 | reset_entity_pointer: | |
2d29c9f8 | 806 | bfqq->weight_counter = NULL; |
9dee8b3b | 807 | bfq_put_queue(bfqq); |
1de0c4cd AA |
808 | } |
809 | ||
0471559c | 810 | /* |
2d29c9f8 FM |
811 | * Invoke __bfq_weights_tree_remove on bfqq and decrement the number |
812 | * of active groups for each queue's inactive parent entity. | |
0471559c PV |
813 | */ |
814 | void bfq_weights_tree_remove(struct bfq_data *bfqd, | |
815 | struct bfq_queue *bfqq) | |
816 | { | |
817 | struct bfq_entity *entity = bfqq->entity.parent; | |
818 | ||
0471559c PV |
819 | for_each_entity(entity) { |
820 | struct bfq_sched_data *sd = entity->my_sched_data; | |
821 | ||
822 | if (sd->next_in_service || sd->in_service_entity) { | |
823 | /* | |
824 | * entity is still active, because either | |
825 | * next_in_service or in_service_entity is not | |
826 | * NULL (see the comments on the definition of | |
827 | * next_in_service for details on why | |
828 | * in_service_entity must be checked too). | |
829 | * | |
2d29c9f8 FM |
830 | * As a consequence, its parent entities are |
831 | * active as well, and thus this loop must | |
832 | * stop here. | |
0471559c PV |
833 | */ |
834 | break; | |
835 | } | |
ba7aeae5 PV |
836 | |
837 | /* | |
838 | * The decrement of num_groups_with_pending_reqs is | |
839 | * not performed immediately upon the deactivation of | |
840 | * entity, but it is delayed to when it also happens | |
841 | * that the first leaf descendant bfqq of entity gets | |
842 | * all its pending requests completed. The following | |
843 | * instructions perform this delayed decrement, if | |
844 | * needed. See the comments on | |
845 | * num_groups_with_pending_reqs for details. | |
846 | */ | |
847 | if (entity->in_groups_with_pending_reqs) { | |
848 | entity->in_groups_with_pending_reqs = false; | |
849 | bfqd->num_groups_with_pending_reqs--; | |
850 | } | |
0471559c | 851 | } |
9dee8b3b PV |
852 | |
853 | /* | |
854 | * Next function is invoked last, because it causes bfqq to be | |
855 | * freed if the following holds: bfqq is not in service and | |
856 | * has no dispatched request. DO NOT use bfqq after the next | |
857 | * function invocation. | |
858 | */ | |
859 | __bfq_weights_tree_remove(bfqd, bfqq, | |
860 | &bfqd->queue_weights_tree); | |
0471559c PV |
861 | } |
862 | ||
aee69d78 PV |
863 | /* |
864 | * Return expired entry, or NULL to just start from scratch in rbtree. | |
865 | */ | |
866 | static struct request *bfq_check_fifo(struct bfq_queue *bfqq, | |
867 | struct request *last) | |
868 | { | |
869 | struct request *rq; | |
870 | ||
871 | if (bfq_bfqq_fifo_expire(bfqq)) | |
872 | return NULL; | |
873 | ||
874 | bfq_mark_bfqq_fifo_expire(bfqq); | |
875 | ||
876 | rq = rq_entry_fifo(bfqq->fifo.next); | |
877 | ||
878 | if (rq == last || ktime_get_ns() < rq->fifo_time) | |
879 | return NULL; | |
880 | ||
881 | bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq); | |
882 | return rq; | |
883 | } | |
884 | ||
885 | static struct request *bfq_find_next_rq(struct bfq_data *bfqd, | |
886 | struct bfq_queue *bfqq, | |
887 | struct request *last) | |
888 | { | |
889 | struct rb_node *rbnext = rb_next(&last->rb_node); | |
890 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
891 | struct request *next, *prev = NULL; | |
892 | ||
893 | /* Follow expired path, else get first next available. */ | |
894 | next = bfq_check_fifo(bfqq, last); | |
895 | if (next) | |
896 | return next; | |
897 | ||
898 | if (rbprev) | |
899 | prev = rb_entry_rq(rbprev); | |
900 | ||
901 | if (rbnext) | |
902 | next = rb_entry_rq(rbnext); | |
903 | else { | |
904 | rbnext = rb_first(&bfqq->sort_list); | |
905 | if (rbnext && rbnext != &last->rb_node) | |
906 | next = rb_entry_rq(rbnext); | |
907 | } | |
908 | ||
909 | return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last)); | |
910 | } | |
911 | ||
c074170e | 912 | /* see the definition of bfq_async_charge_factor for details */ |
aee69d78 PV |
913 | static unsigned long bfq_serv_to_charge(struct request *rq, |
914 | struct bfq_queue *bfqq) | |
915 | { | |
02a6d787 | 916 | if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1 || |
fb53ac6c | 917 | bfq_asymmetric_scenario(bfqq->bfqd, bfqq)) |
c074170e PV |
918 | return blk_rq_sectors(rq); |
919 | ||
d5801088 | 920 | return blk_rq_sectors(rq) * bfq_async_charge_factor; |
aee69d78 PV |
921 | } |
922 | ||
923 | /** | |
924 | * bfq_updated_next_req - update the queue after a new next_rq selection. | |
925 | * @bfqd: the device data the queue belongs to. | |
926 | * @bfqq: the queue to update. | |
927 | * | |
928 | * If the first request of a queue changes we make sure that the queue | |
929 | * has enough budget to serve at least its first request (if the | |
930 | * request has grown). We do this because if the queue has not enough | |
931 | * budget for its first request, it has to go through two dispatch | |
932 | * rounds to actually get it dispatched. | |
933 | */ | |
934 | static void bfq_updated_next_req(struct bfq_data *bfqd, | |
935 | struct bfq_queue *bfqq) | |
936 | { | |
937 | struct bfq_entity *entity = &bfqq->entity; | |
938 | struct request *next_rq = bfqq->next_rq; | |
939 | unsigned long new_budget; | |
940 | ||
941 | if (!next_rq) | |
942 | return; | |
943 | ||
944 | if (bfqq == bfqd->in_service_queue) | |
945 | /* | |
946 | * In order not to break guarantees, budgets cannot be | |
947 | * changed after an entity has been selected. | |
948 | */ | |
949 | return; | |
950 | ||
f3218ad8 PV |
951 | new_budget = max_t(unsigned long, |
952 | max_t(unsigned long, bfqq->max_budget, | |
953 | bfq_serv_to_charge(next_rq, bfqq)), | |
954 | entity->service); | |
aee69d78 PV |
955 | if (entity->budget != new_budget) { |
956 | entity->budget = new_budget; | |
957 | bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu", | |
958 | new_budget); | |
80294c3b | 959 | bfq_requeue_bfqq(bfqd, bfqq, false); |
aee69d78 PV |
960 | } |
961 | } | |
962 | ||
3e2bdd6d PV |
963 | static unsigned int bfq_wr_duration(struct bfq_data *bfqd) |
964 | { | |
965 | u64 dur; | |
966 | ||
967 | if (bfqd->bfq_wr_max_time > 0) | |
968 | return bfqd->bfq_wr_max_time; | |
969 | ||
e24f1c24 | 970 | dur = bfqd->rate_dur_prod; |
3e2bdd6d PV |
971 | do_div(dur, bfqd->peak_rate); |
972 | ||
973 | /* | |
d450542e DS |
974 | * Limit duration between 3 and 25 seconds. The upper limit |
975 | * has been conservatively set after the following worst case: | |
976 | * on a QEMU/KVM virtual machine | |
977 | * - running in a slow PC | |
978 | * - with a virtual disk stacked on a slow low-end 5400rpm HDD | |
979 | * - serving a heavy I/O workload, such as the sequential reading | |
980 | * of several files | |
981 | * mplayer took 23 seconds to start, if constantly weight-raised. | |
982 | * | |
636b8fe8 | 983 | * As for higher values than that accommodating the above bad |
d450542e DS |
984 | * scenario, tests show that higher values would often yield |
985 | * the opposite of the desired result, i.e., would worsen | |
986 | * responsiveness by allowing non-interactive applications to | |
987 | * preserve weight raising for too long. | |
3e2bdd6d PV |
988 | * |
989 | * On the other end, lower values than 3 seconds make it | |
990 | * difficult for most interactive tasks to complete their jobs | |
991 | * before weight-raising finishes. | |
992 | */ | |
d450542e | 993 | return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000)); |
3e2bdd6d PV |
994 | } |
995 | ||
996 | /* switch back from soft real-time to interactive weight raising */ | |
997 | static void switch_back_to_interactive_wr(struct bfq_queue *bfqq, | |
998 | struct bfq_data *bfqd) | |
999 | { | |
1000 | bfqq->wr_coeff = bfqd->bfq_wr_coeff; | |
1001 | bfqq->wr_cur_max_time = bfq_wr_duration(bfqd); | |
1002 | bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt; | |
1003 | } | |
1004 | ||
36eca894 | 1005 | static void |
13c931bd PV |
1006 | bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd, |
1007 | struct bfq_io_cq *bic, bool bfq_already_existing) | |
36eca894 | 1008 | { |
13c931bd PV |
1009 | unsigned int old_wr_coeff = bfqq->wr_coeff; |
1010 | bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq); | |
1011 | ||
d5be3fef PV |
1012 | if (bic->saved_has_short_ttime) |
1013 | bfq_mark_bfqq_has_short_ttime(bfqq); | |
36eca894 | 1014 | else |
d5be3fef | 1015 | bfq_clear_bfqq_has_short_ttime(bfqq); |
36eca894 AA |
1016 | |
1017 | if (bic->saved_IO_bound) | |
1018 | bfq_mark_bfqq_IO_bound(bfqq); | |
1019 | else | |
1020 | bfq_clear_bfqq_IO_bound(bfqq); | |
1021 | ||
fffca087 | 1022 | bfqq->entity.new_weight = bic->saved_weight; |
36eca894 AA |
1023 | bfqq->ttime = bic->saved_ttime; |
1024 | bfqq->wr_coeff = bic->saved_wr_coeff; | |
1025 | bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt; | |
1026 | bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish; | |
1027 | bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time; | |
1028 | ||
e1b2324d | 1029 | if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) || |
36eca894 | 1030 | time_is_before_jiffies(bfqq->last_wr_start_finish + |
e1b2324d | 1031 | bfqq->wr_cur_max_time))) { |
3e2bdd6d PV |
1032 | if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time && |
1033 | !bfq_bfqq_in_large_burst(bfqq) && | |
1034 | time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt + | |
1035 | bfq_wr_duration(bfqd))) { | |
1036 | switch_back_to_interactive_wr(bfqq, bfqd); | |
1037 | } else { | |
1038 | bfqq->wr_coeff = 1; | |
1039 | bfq_log_bfqq(bfqq->bfqd, bfqq, | |
1040 | "resume state: switching off wr"); | |
1041 | } | |
36eca894 AA |
1042 | } |
1043 | ||
1044 | /* make sure weight will be updated, however we got here */ | |
1045 | bfqq->entity.prio_changed = 1; | |
13c931bd PV |
1046 | |
1047 | if (likely(!busy)) | |
1048 | return; | |
1049 | ||
1050 | if (old_wr_coeff == 1 && bfqq->wr_coeff > 1) | |
1051 | bfqd->wr_busy_queues++; | |
1052 | else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1) | |
1053 | bfqd->wr_busy_queues--; | |
36eca894 AA |
1054 | } |
1055 | ||
1056 | static int bfqq_process_refs(struct bfq_queue *bfqq) | |
1057 | { | |
9dee8b3b PV |
1058 | return bfqq->ref - bfqq->allocated - bfqq->entity.on_st - |
1059 | (bfqq->weight_counter != NULL); | |
36eca894 AA |
1060 | } |
1061 | ||
e1b2324d AA |
1062 | /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */ |
1063 | static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq) | |
1064 | { | |
1065 | struct bfq_queue *item; | |
1066 | struct hlist_node *n; | |
1067 | ||
1068 | hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node) | |
1069 | hlist_del_init(&item->burst_list_node); | |
84a74689 PV |
1070 | |
1071 | /* | |
1072 | * Start the creation of a new burst list only if there is no | |
1073 | * active queue. See comments on the conditional invocation of | |
1074 | * bfq_handle_burst(). | |
1075 | */ | |
1076 | if (bfq_tot_busy_queues(bfqd) == 0) { | |
1077 | hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list); | |
1078 | bfqd->burst_size = 1; | |
1079 | } else | |
1080 | bfqd->burst_size = 0; | |
1081 | ||
e1b2324d AA |
1082 | bfqd->burst_parent_entity = bfqq->entity.parent; |
1083 | } | |
1084 | ||
1085 | /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */ | |
1086 | static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq) | |
1087 | { | |
1088 | /* Increment burst size to take into account also bfqq */ | |
1089 | bfqd->burst_size++; | |
1090 | ||
1091 | if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) { | |
1092 | struct bfq_queue *pos, *bfqq_item; | |
1093 | struct hlist_node *n; | |
1094 | ||
1095 | /* | |
1096 | * Enough queues have been activated shortly after each | |
1097 | * other to consider this burst as large. | |
1098 | */ | |
1099 | bfqd->large_burst = true; | |
1100 | ||
1101 | /* | |
1102 | * We can now mark all queues in the burst list as | |
1103 | * belonging to a large burst. | |
1104 | */ | |
1105 | hlist_for_each_entry(bfqq_item, &bfqd->burst_list, | |
1106 | burst_list_node) | |
1107 | bfq_mark_bfqq_in_large_burst(bfqq_item); | |
1108 | bfq_mark_bfqq_in_large_burst(bfqq); | |
1109 | ||
1110 | /* | |
1111 | * From now on, and until the current burst finishes, any | |
1112 | * new queue being activated shortly after the last queue | |
1113 | * was inserted in the burst can be immediately marked as | |
1114 | * belonging to a large burst. So the burst list is not | |
1115 | * needed any more. Remove it. | |
1116 | */ | |
1117 | hlist_for_each_entry_safe(pos, n, &bfqd->burst_list, | |
1118 | burst_list_node) | |
1119 | hlist_del_init(&pos->burst_list_node); | |
1120 | } else /* | |
1121 | * Burst not yet large: add bfqq to the burst list. Do | |
1122 | * not increment the ref counter for bfqq, because bfqq | |
1123 | * is removed from the burst list before freeing bfqq | |
1124 | * in put_queue. | |
1125 | */ | |
1126 | hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list); | |
1127 | } | |
1128 | ||
1129 | /* | |
1130 | * If many queues belonging to the same group happen to be created | |
1131 | * shortly after each other, then the processes associated with these | |
1132 | * queues have typically a common goal. In particular, bursts of queue | |
1133 | * creations are usually caused by services or applications that spawn | |
1134 | * many parallel threads/processes. Examples are systemd during boot, | |
1135 | * or git grep. To help these processes get their job done as soon as | |
1136 | * possible, it is usually better to not grant either weight-raising | |
84a74689 PV |
1137 | * or device idling to their queues, unless these queues must be |
1138 | * protected from the I/O flowing through other active queues. | |
e1b2324d AA |
1139 | * |
1140 | * In this comment we describe, firstly, the reasons why this fact | |
1141 | * holds, and, secondly, the next function, which implements the main | |
1142 | * steps needed to properly mark these queues so that they can then be | |
1143 | * treated in a different way. | |
1144 | * | |
1145 | * The above services or applications benefit mostly from a high | |
1146 | * throughput: the quicker the requests of the activated queues are | |
1147 | * cumulatively served, the sooner the target job of these queues gets | |
1148 | * completed. As a consequence, weight-raising any of these queues, | |
1149 | * which also implies idling the device for it, is almost always | |
84a74689 PV |
1150 | * counterproductive, unless there are other active queues to isolate |
1151 | * these new queues from. If there no other active queues, then | |
1152 | * weight-raising these new queues just lowers throughput in most | |
1153 | * cases. | |
e1b2324d AA |
1154 | * |
1155 | * On the other hand, a burst of queue creations may be caused also by | |
1156 | * the start of an application that does not consist of a lot of | |
1157 | * parallel I/O-bound threads. In fact, with a complex application, | |
1158 | * several short processes may need to be executed to start-up the | |
1159 | * application. In this respect, to start an application as quickly as | |
1160 | * possible, the best thing to do is in any case to privilege the I/O | |
1161 | * related to the application with respect to all other | |
1162 | * I/O. Therefore, the best strategy to start as quickly as possible | |
1163 | * an application that causes a burst of queue creations is to | |
1164 | * weight-raise all the queues created during the burst. This is the | |
1165 | * exact opposite of the best strategy for the other type of bursts. | |
1166 | * | |
1167 | * In the end, to take the best action for each of the two cases, the | |
1168 | * two types of bursts need to be distinguished. Fortunately, this | |
1169 | * seems relatively easy, by looking at the sizes of the bursts. In | |
1170 | * particular, we found a threshold such that only bursts with a | |
1171 | * larger size than that threshold are apparently caused by | |
1172 | * services or commands such as systemd or git grep. For brevity, | |
1173 | * hereafter we call just 'large' these bursts. BFQ *does not* | |
1174 | * weight-raise queues whose creation occurs in a large burst. In | |
1175 | * addition, for each of these queues BFQ performs or does not perform | |
1176 | * idling depending on which choice boosts the throughput more. The | |
1177 | * exact choice depends on the device and request pattern at | |
1178 | * hand. | |
1179 | * | |
1180 | * Unfortunately, false positives may occur while an interactive task | |
1181 | * is starting (e.g., an application is being started). The | |
1182 | * consequence is that the queues associated with the task do not | |
1183 | * enjoy weight raising as expected. Fortunately these false positives | |
1184 | * are very rare. They typically occur if some service happens to | |
1185 | * start doing I/O exactly when the interactive task starts. | |
1186 | * | |
84a74689 PV |
1187 | * Turning back to the next function, it is invoked only if there are |
1188 | * no active queues (apart from active queues that would belong to the | |
1189 | * same, possible burst bfqq would belong to), and it implements all | |
1190 | * the steps needed to detect the occurrence of a large burst and to | |
1191 | * properly mark all the queues belonging to it (so that they can then | |
1192 | * be treated in a different way). This goal is achieved by | |
1193 | * maintaining a "burst list" that holds, temporarily, the queues that | |
1194 | * belong to the burst in progress. The list is then used to mark | |
1195 | * these queues as belonging to a large burst if the burst does become | |
1196 | * large. The main steps are the following. | |
e1b2324d AA |
1197 | * |
1198 | * . when the very first queue is created, the queue is inserted into the | |
1199 | * list (as it could be the first queue in a possible burst) | |
1200 | * | |
1201 | * . if the current burst has not yet become large, and a queue Q that does | |
1202 | * not yet belong to the burst is activated shortly after the last time | |
1203 | * at which a new queue entered the burst list, then the function appends | |
1204 | * Q to the burst list | |
1205 | * | |
1206 | * . if, as a consequence of the previous step, the burst size reaches | |
1207 | * the large-burst threshold, then | |
1208 | * | |
1209 | * . all the queues in the burst list are marked as belonging to a | |
1210 | * large burst | |
1211 | * | |
1212 | * . the burst list is deleted; in fact, the burst list already served | |
1213 | * its purpose (keeping temporarily track of the queues in a burst, | |
1214 | * so as to be able to mark them as belonging to a large burst in the | |
1215 | * previous sub-step), and now is not needed any more | |
1216 | * | |
1217 | * . the device enters a large-burst mode | |
1218 | * | |
1219 | * . if a queue Q that does not belong to the burst is created while | |
1220 | * the device is in large-burst mode and shortly after the last time | |
1221 | * at which a queue either entered the burst list or was marked as | |
1222 | * belonging to the current large burst, then Q is immediately marked | |
1223 | * as belonging to a large burst. | |
1224 | * | |
1225 | * . if a queue Q that does not belong to the burst is created a while | |
1226 | * later, i.e., not shortly after, than the last time at which a queue | |
1227 | * either entered the burst list or was marked as belonging to the | |
1228 | * current large burst, then the current burst is deemed as finished and: | |
1229 | * | |
1230 | * . the large-burst mode is reset if set | |
1231 | * | |
1232 | * . the burst list is emptied | |
1233 | * | |
1234 | * . Q is inserted in the burst list, as Q may be the first queue | |
1235 | * in a possible new burst (then the burst list contains just Q | |
1236 | * after this step). | |
1237 | */ | |
1238 | static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq) | |
1239 | { | |
1240 | /* | |
1241 | * If bfqq is already in the burst list or is part of a large | |
1242 | * burst, or finally has just been split, then there is | |
1243 | * nothing else to do. | |
1244 | */ | |
1245 | if (!hlist_unhashed(&bfqq->burst_list_node) || | |
1246 | bfq_bfqq_in_large_burst(bfqq) || | |
1247 | time_is_after_eq_jiffies(bfqq->split_time + | |
1248 | msecs_to_jiffies(10))) | |
1249 | return; | |
1250 | ||
1251 | /* | |
1252 | * If bfqq's creation happens late enough, or bfqq belongs to | |
1253 | * a different group than the burst group, then the current | |
1254 | * burst is finished, and related data structures must be | |
1255 | * reset. | |
1256 | * | |
1257 | * In this respect, consider the special case where bfqq is | |
1258 | * the very first queue created after BFQ is selected for this | |
1259 | * device. In this case, last_ins_in_burst and | |
1260 | * burst_parent_entity are not yet significant when we get | |
1261 | * here. But it is easy to verify that, whether or not the | |
1262 | * following condition is true, bfqq will end up being | |
1263 | * inserted into the burst list. In particular the list will | |
1264 | * happen to contain only bfqq. And this is exactly what has | |
1265 | * to happen, as bfqq may be the first queue of the first | |
1266 | * burst. | |
1267 | */ | |
1268 | if (time_is_before_jiffies(bfqd->last_ins_in_burst + | |
1269 | bfqd->bfq_burst_interval) || | |
1270 | bfqq->entity.parent != bfqd->burst_parent_entity) { | |
1271 | bfqd->large_burst = false; | |
1272 | bfq_reset_burst_list(bfqd, bfqq); | |
1273 | goto end; | |
1274 | } | |
1275 | ||
1276 | /* | |
1277 | * If we get here, then bfqq is being activated shortly after the | |
1278 | * last queue. So, if the current burst is also large, we can mark | |
1279 | * bfqq as belonging to this large burst immediately. | |
1280 | */ | |
1281 | if (bfqd->large_burst) { | |
1282 | bfq_mark_bfqq_in_large_burst(bfqq); | |
1283 | goto end; | |
1284 | } | |
1285 | ||
1286 | /* | |
1287 | * If we get here, then a large-burst state has not yet been | |
1288 | * reached, but bfqq is being activated shortly after the last | |
1289 | * queue. Then we add bfqq to the burst. | |
1290 | */ | |
1291 | bfq_add_to_burst(bfqd, bfqq); | |
1292 | end: | |
1293 | /* | |
1294 | * At this point, bfqq either has been added to the current | |
1295 | * burst or has caused the current burst to terminate and a | |
1296 | * possible new burst to start. In particular, in the second | |
1297 | * case, bfqq has become the first queue in the possible new | |
1298 | * burst. In both cases last_ins_in_burst needs to be moved | |
1299 | * forward. | |
1300 | */ | |
1301 | bfqd->last_ins_in_burst = jiffies; | |
1302 | } | |
1303 | ||
aee69d78 PV |
1304 | static int bfq_bfqq_budget_left(struct bfq_queue *bfqq) |
1305 | { | |
1306 | struct bfq_entity *entity = &bfqq->entity; | |
1307 | ||
1308 | return entity->budget - entity->service; | |
1309 | } | |
1310 | ||
1311 | /* | |
1312 | * If enough samples have been computed, return the current max budget | |
1313 | * stored in bfqd, which is dynamically updated according to the | |
1314 | * estimated disk peak rate; otherwise return the default max budget | |
1315 | */ | |
1316 | static int bfq_max_budget(struct bfq_data *bfqd) | |
1317 | { | |
1318 | if (bfqd->budgets_assigned < bfq_stats_min_budgets) | |
1319 | return bfq_default_max_budget; | |
1320 | else | |
1321 | return bfqd->bfq_max_budget; | |
1322 | } | |
1323 | ||
1324 | /* | |
1325 | * Return min budget, which is a fraction of the current or default | |
1326 | * max budget (trying with 1/32) | |
1327 | */ | |
1328 | static int bfq_min_budget(struct bfq_data *bfqd) | |
1329 | { | |
1330 | if (bfqd->budgets_assigned < bfq_stats_min_budgets) | |
1331 | return bfq_default_max_budget / 32; | |
1332 | else | |
1333 | return bfqd->bfq_max_budget / 32; | |
1334 | } | |
1335 | ||
aee69d78 PV |
1336 | /* |
1337 | * The next function, invoked after the input queue bfqq switches from | |
1338 | * idle to busy, updates the budget of bfqq. The function also tells | |
1339 | * whether the in-service queue should be expired, by returning | |
1340 | * true. The purpose of expiring the in-service queue is to give bfqq | |
1341 | * the chance to possibly preempt the in-service queue, and the reason | |
44e44a1b PV |
1342 | * for preempting the in-service queue is to achieve one of the two |
1343 | * goals below. | |
aee69d78 | 1344 | * |
44e44a1b PV |
1345 | * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has |
1346 | * expired because it has remained idle. In particular, bfqq may have | |
1347 | * expired for one of the following two reasons: | |
aee69d78 PV |
1348 | * |
1349 | * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling | |
1350 | * and did not make it to issue a new request before its last | |
1351 | * request was served; | |
1352 | * | |
1353 | * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue | |
1354 | * a new request before the expiration of the idling-time. | |
1355 | * | |
1356 | * Even if bfqq has expired for one of the above reasons, the process | |
1357 | * associated with the queue may be however issuing requests greedily, | |
1358 | * and thus be sensitive to the bandwidth it receives (bfqq may have | |
1359 | * remained idle for other reasons: CPU high load, bfqq not enjoying | |
1360 | * idling, I/O throttling somewhere in the path from the process to | |
1361 | * the I/O scheduler, ...). But if, after every expiration for one of | |
1362 | * the above two reasons, bfqq has to wait for the service of at least | |
1363 | * one full budget of another queue before being served again, then | |
1364 | * bfqq is likely to get a much lower bandwidth or resource time than | |
1365 | * its reserved ones. To address this issue, two countermeasures need | |
1366 | * to be taken. | |
1367 | * | |
1368 | * First, the budget and the timestamps of bfqq need to be updated in | |
1369 | * a special way on bfqq reactivation: they need to be updated as if | |
1370 | * bfqq did not remain idle and did not expire. In fact, if they are | |
1371 | * computed as if bfqq expired and remained idle until reactivation, | |
1372 | * then the process associated with bfqq is treated as if, instead of | |
1373 | * being greedy, it stopped issuing requests when bfqq remained idle, | |
1374 | * and restarts issuing requests only on this reactivation. In other | |
1375 | * words, the scheduler does not help the process recover the "service | |
1376 | * hole" between bfqq expiration and reactivation. As a consequence, | |
1377 | * the process receives a lower bandwidth than its reserved one. In | |
1378 | * contrast, to recover this hole, the budget must be updated as if | |
1379 | * bfqq was not expired at all before this reactivation, i.e., it must | |
1380 | * be set to the value of the remaining budget when bfqq was | |
1381 | * expired. Along the same line, timestamps need to be assigned the | |
1382 | * value they had the last time bfqq was selected for service, i.e., | |
1383 | * before last expiration. Thus timestamps need to be back-shifted | |
1384 | * with respect to their normal computation (see [1] for more details | |
1385 | * on this tricky aspect). | |
1386 | * | |
1387 | * Secondly, to allow the process to recover the hole, the in-service | |
1388 | * queue must be expired too, to give bfqq the chance to preempt it | |
1389 | * immediately. In fact, if bfqq has to wait for a full budget of the | |
1390 | * in-service queue to be completed, then it may become impossible to | |
1391 | * let the process recover the hole, even if the back-shifted | |
1392 | * timestamps of bfqq are lower than those of the in-service queue. If | |
1393 | * this happens for most or all of the holes, then the process may not | |
1394 | * receive its reserved bandwidth. In this respect, it is worth noting | |
1395 | * that, being the service of outstanding requests unpreemptible, a | |
1396 | * little fraction of the holes may however be unrecoverable, thereby | |
1397 | * causing a little loss of bandwidth. | |
1398 | * | |
1399 | * The last important point is detecting whether bfqq does need this | |
1400 | * bandwidth recovery. In this respect, the next function deems the | |
1401 | * process associated with bfqq greedy, and thus allows it to recover | |
1402 | * the hole, if: 1) the process is waiting for the arrival of a new | |
1403 | * request (which implies that bfqq expired for one of the above two | |
1404 | * reasons), and 2) such a request has arrived soon. The first | |
1405 | * condition is controlled through the flag non_blocking_wait_rq, | |
1406 | * while the second through the flag arrived_in_time. If both | |
1407 | * conditions hold, then the function computes the budget in the | |
1408 | * above-described special way, and signals that the in-service queue | |
1409 | * should be expired. Timestamp back-shifting is done later in | |
1410 | * __bfq_activate_entity. | |
44e44a1b PV |
1411 | * |
1412 | * 2. Reduce latency. Even if timestamps are not backshifted to let | |
1413 | * the process associated with bfqq recover a service hole, bfqq may | |
1414 | * however happen to have, after being (re)activated, a lower finish | |
1415 | * timestamp than the in-service queue. That is, the next budget of | |
1416 | * bfqq may have to be completed before the one of the in-service | |
1417 | * queue. If this is the case, then preempting the in-service queue | |
1418 | * allows this goal to be achieved, apart from the unpreemptible, | |
1419 | * outstanding requests mentioned above. | |
1420 | * | |
1421 | * Unfortunately, regardless of which of the above two goals one wants | |
1422 | * to achieve, service trees need first to be updated to know whether | |
1423 | * the in-service queue must be preempted. To have service trees | |
1424 | * correctly updated, the in-service queue must be expired and | |
1425 | * rescheduled, and bfqq must be scheduled too. This is one of the | |
1426 | * most costly operations (in future versions, the scheduling | |
1427 | * mechanism may be re-designed in such a way to make it possible to | |
1428 | * know whether preemption is needed without needing to update service | |
1429 | * trees). In addition, queue preemptions almost always cause random | |
1430 | * I/O, and thus loss of throughput. Because of these facts, the next | |
1431 | * function adopts the following simple scheme to avoid both costly | |
1432 | * operations and too frequent preemptions: it requests the expiration | |
1433 | * of the in-service queue (unconditionally) only for queues that need | |
1434 | * to recover a hole, or that either are weight-raised or deserve to | |
1435 | * be weight-raised. | |
aee69d78 PV |
1436 | */ |
1437 | static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd, | |
1438 | struct bfq_queue *bfqq, | |
44e44a1b PV |
1439 | bool arrived_in_time, |
1440 | bool wr_or_deserves_wr) | |
aee69d78 PV |
1441 | { |
1442 | struct bfq_entity *entity = &bfqq->entity; | |
1443 | ||
218cb897 PV |
1444 | /* |
1445 | * In the next compound condition, we check also whether there | |
1446 | * is some budget left, because otherwise there is no point in | |
1447 | * trying to go on serving bfqq with this same budget: bfqq | |
1448 | * would be expired immediately after being selected for | |
1449 | * service. This would only cause useless overhead. | |
1450 | */ | |
1451 | if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time && | |
1452 | bfq_bfqq_budget_left(bfqq) > 0) { | |
aee69d78 PV |
1453 | /* |
1454 | * We do not clear the flag non_blocking_wait_rq here, as | |
1455 | * the latter is used in bfq_activate_bfqq to signal | |
1456 | * that timestamps need to be back-shifted (and is | |
1457 | * cleared right after). | |
1458 | */ | |
1459 | ||
1460 | /* | |
1461 | * In next assignment we rely on that either | |
1462 | * entity->service or entity->budget are not updated | |
1463 | * on expiration if bfqq is empty (see | |
1464 | * __bfq_bfqq_recalc_budget). Thus both quantities | |
1465 | * remain unchanged after such an expiration, and the | |
1466 | * following statement therefore assigns to | |
1467 | * entity->budget the remaining budget on such an | |
9fae8dd5 | 1468 | * expiration. |
aee69d78 PV |
1469 | */ |
1470 | entity->budget = min_t(unsigned long, | |
1471 | bfq_bfqq_budget_left(bfqq), | |
1472 | bfqq->max_budget); | |
1473 | ||
9fae8dd5 PV |
1474 | /* |
1475 | * At this point, we have used entity->service to get | |
1476 | * the budget left (needed for updating | |
1477 | * entity->budget). Thus we finally can, and have to, | |
1478 | * reset entity->service. The latter must be reset | |
1479 | * because bfqq would otherwise be charged again for | |
1480 | * the service it has received during its previous | |
1481 | * service slot(s). | |
1482 | */ | |
1483 | entity->service = 0; | |
1484 | ||
aee69d78 PV |
1485 | return true; |
1486 | } | |
1487 | ||
9fae8dd5 PV |
1488 | /* |
1489 | * We can finally complete expiration, by setting service to 0. | |
1490 | */ | |
1491 | entity->service = 0; | |
aee69d78 PV |
1492 | entity->budget = max_t(unsigned long, bfqq->max_budget, |
1493 | bfq_serv_to_charge(bfqq->next_rq, bfqq)); | |
1494 | bfq_clear_bfqq_non_blocking_wait_rq(bfqq); | |
44e44a1b PV |
1495 | return wr_or_deserves_wr; |
1496 | } | |
1497 | ||
4baa8bb1 PV |
1498 | /* |
1499 | * Return the farthest past time instant according to jiffies | |
1500 | * macros. | |
1501 | */ | |
1502 | static unsigned long bfq_smallest_from_now(void) | |
1503 | { | |
1504 | return jiffies - MAX_JIFFY_OFFSET; | |
1505 | } | |
1506 | ||
44e44a1b PV |
1507 | static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd, |
1508 | struct bfq_queue *bfqq, | |
1509 | unsigned int old_wr_coeff, | |
1510 | bool wr_or_deserves_wr, | |
77b7dcea | 1511 | bool interactive, |
e1b2324d | 1512 | bool in_burst, |
77b7dcea | 1513 | bool soft_rt) |
44e44a1b PV |
1514 | { |
1515 | if (old_wr_coeff == 1 && wr_or_deserves_wr) { | |
1516 | /* start a weight-raising period */ | |
77b7dcea | 1517 | if (interactive) { |
8a8747dc | 1518 | bfqq->service_from_wr = 0; |
77b7dcea PV |
1519 | bfqq->wr_coeff = bfqd->bfq_wr_coeff; |
1520 | bfqq->wr_cur_max_time = bfq_wr_duration(bfqd); | |
1521 | } else { | |
4baa8bb1 PV |
1522 | /* |
1523 | * No interactive weight raising in progress | |
1524 | * here: assign minus infinity to | |
1525 | * wr_start_at_switch_to_srt, to make sure | |
1526 | * that, at the end of the soft-real-time | |
1527 | * weight raising periods that is starting | |
1528 | * now, no interactive weight-raising period | |
1529 | * may be wrongly considered as still in | |
1530 | * progress (and thus actually started by | |
1531 | * mistake). | |
1532 | */ | |
1533 | bfqq->wr_start_at_switch_to_srt = | |
1534 | bfq_smallest_from_now(); | |
77b7dcea PV |
1535 | bfqq->wr_coeff = bfqd->bfq_wr_coeff * |
1536 | BFQ_SOFTRT_WEIGHT_FACTOR; | |
1537 | bfqq->wr_cur_max_time = | |
1538 | bfqd->bfq_wr_rt_max_time; | |
1539 | } | |
44e44a1b PV |
1540 | |
1541 | /* | |
1542 | * If needed, further reduce budget to make sure it is | |
1543 | * close to bfqq's backlog, so as to reduce the | |
1544 | * scheduling-error component due to a too large | |
1545 | * budget. Do not care about throughput consequences, | |
1546 | * but only about latency. Finally, do not assign a | |
1547 | * too small budget either, to avoid increasing | |
1548 | * latency by causing too frequent expirations. | |
1549 | */ | |
1550 | bfqq->entity.budget = min_t(unsigned long, | |
1551 | bfqq->entity.budget, | |
1552 | 2 * bfq_min_budget(bfqd)); | |
1553 | } else if (old_wr_coeff > 1) { | |
77b7dcea PV |
1554 | if (interactive) { /* update wr coeff and duration */ |
1555 | bfqq->wr_coeff = bfqd->bfq_wr_coeff; | |
1556 | bfqq->wr_cur_max_time = bfq_wr_duration(bfqd); | |
e1b2324d AA |
1557 | } else if (in_burst) |
1558 | bfqq->wr_coeff = 1; | |
1559 | else if (soft_rt) { | |
77b7dcea PV |
1560 | /* |
1561 | * The application is now or still meeting the | |
1562 | * requirements for being deemed soft rt. We | |
1563 | * can then correctly and safely (re)charge | |
1564 | * the weight-raising duration for the | |
1565 | * application with the weight-raising | |
1566 | * duration for soft rt applications. | |
1567 | * | |
1568 | * In particular, doing this recharge now, i.e., | |
1569 | * before the weight-raising period for the | |
1570 | * application finishes, reduces the probability | |
1571 | * of the following negative scenario: | |
1572 | * 1) the weight of a soft rt application is | |
1573 | * raised at startup (as for any newly | |
1574 | * created application), | |
1575 | * 2) since the application is not interactive, | |
1576 | * at a certain time weight-raising is | |
1577 | * stopped for the application, | |
1578 | * 3) at that time the application happens to | |
1579 | * still have pending requests, and hence | |
1580 | * is destined to not have a chance to be | |
1581 | * deemed soft rt before these requests are | |
1582 | * completed (see the comments to the | |
1583 | * function bfq_bfqq_softrt_next_start() | |
1584 | * for details on soft rt detection), | |
1585 | * 4) these pending requests experience a high | |
1586 | * latency because the application is not | |
1587 | * weight-raised while they are pending. | |
1588 | */ | |
1589 | if (bfqq->wr_cur_max_time != | |
1590 | bfqd->bfq_wr_rt_max_time) { | |
1591 | bfqq->wr_start_at_switch_to_srt = | |
1592 | bfqq->last_wr_start_finish; | |
1593 | ||
1594 | bfqq->wr_cur_max_time = | |
1595 | bfqd->bfq_wr_rt_max_time; | |
1596 | bfqq->wr_coeff = bfqd->bfq_wr_coeff * | |
1597 | BFQ_SOFTRT_WEIGHT_FACTOR; | |
1598 | } | |
1599 | bfqq->last_wr_start_finish = jiffies; | |
1600 | } | |
44e44a1b PV |
1601 | } |
1602 | } | |
1603 | ||
1604 | static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd, | |
1605 | struct bfq_queue *bfqq) | |
1606 | { | |
1607 | return bfqq->dispatched == 0 && | |
1608 | time_is_before_jiffies( | |
1609 | bfqq->budget_timeout + | |
1610 | bfqd->bfq_wr_min_idle_time); | |
aee69d78 PV |
1611 | } |
1612 | ||
1613 | static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd, | |
1614 | struct bfq_queue *bfqq, | |
44e44a1b PV |
1615 | int old_wr_coeff, |
1616 | struct request *rq, | |
1617 | bool *interactive) | |
aee69d78 | 1618 | { |
e1b2324d AA |
1619 | bool soft_rt, in_burst, wr_or_deserves_wr, |
1620 | bfqq_wants_to_preempt, | |
44e44a1b | 1621 | idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq), |
aee69d78 PV |
1622 | /* |
1623 | * See the comments on | |
1624 | * bfq_bfqq_update_budg_for_activation for | |
1625 | * details on the usage of the next variable. | |
1626 | */ | |
1627 | arrived_in_time = ktime_get_ns() <= | |
1628 | bfqq->ttime.last_end_request + | |
1629 | bfqd->bfq_slice_idle * 3; | |
1630 | ||
e21b7a0b | 1631 | |
aee69d78 | 1632 | /* |
44e44a1b PV |
1633 | * bfqq deserves to be weight-raised if: |
1634 | * - it is sync, | |
e1b2324d | 1635 | * - it does not belong to a large burst, |
36eca894 AA |
1636 | * - it has been idle for enough time or is soft real-time, |
1637 | * - is linked to a bfq_io_cq (it is not shared in any sense). | |
44e44a1b | 1638 | */ |
e1b2324d | 1639 | in_burst = bfq_bfqq_in_large_burst(bfqq); |
77b7dcea | 1640 | soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 && |
7074f076 | 1641 | !BFQQ_TOTALLY_SEEKY(bfqq) && |
e1b2324d | 1642 | !in_burst && |
f6c3ca0e DS |
1643 | time_is_before_jiffies(bfqq->soft_rt_next_start) && |
1644 | bfqq->dispatched == 0; | |
e1b2324d | 1645 | *interactive = !in_burst && idle_for_long_time; |
44e44a1b PV |
1646 | wr_or_deserves_wr = bfqd->low_latency && |
1647 | (bfqq->wr_coeff > 1 || | |
36eca894 AA |
1648 | (bfq_bfqq_sync(bfqq) && |
1649 | bfqq->bic && (*interactive || soft_rt))); | |
44e44a1b PV |
1650 | |
1651 | /* | |
1652 | * Using the last flag, update budget and check whether bfqq | |
1653 | * may want to preempt the in-service queue. | |
aee69d78 PV |
1654 | */ |
1655 | bfqq_wants_to_preempt = | |
1656 | bfq_bfqq_update_budg_for_activation(bfqd, bfqq, | |
44e44a1b PV |
1657 | arrived_in_time, |
1658 | wr_or_deserves_wr); | |
aee69d78 | 1659 | |
e1b2324d AA |
1660 | /* |
1661 | * If bfqq happened to be activated in a burst, but has been | |
1662 | * idle for much more than an interactive queue, then we | |
1663 | * assume that, in the overall I/O initiated in the burst, the | |
1664 | * I/O associated with bfqq is finished. So bfqq does not need | |
1665 | * to be treated as a queue belonging to a burst | |
1666 | * anymore. Accordingly, we reset bfqq's in_large_burst flag | |
1667 | * if set, and remove bfqq from the burst list if it's | |
1668 | * there. We do not decrement burst_size, because the fact | |
1669 | * that bfqq does not need to belong to the burst list any | |
1670 | * more does not invalidate the fact that bfqq was created in | |
1671 | * a burst. | |
1672 | */ | |
1673 | if (likely(!bfq_bfqq_just_created(bfqq)) && | |
1674 | idle_for_long_time && | |
1675 | time_is_before_jiffies( | |
1676 | bfqq->budget_timeout + | |
1677 | msecs_to_jiffies(10000))) { | |
1678 | hlist_del_init(&bfqq->burst_list_node); | |
1679 | bfq_clear_bfqq_in_large_burst(bfqq); | |
1680 | } | |
1681 | ||
1682 | bfq_clear_bfqq_just_created(bfqq); | |
1683 | ||
1684 | ||
aee69d78 PV |
1685 | if (!bfq_bfqq_IO_bound(bfqq)) { |
1686 | if (arrived_in_time) { | |
1687 | bfqq->requests_within_timer++; | |
1688 | if (bfqq->requests_within_timer >= | |
1689 | bfqd->bfq_requests_within_timer) | |
1690 | bfq_mark_bfqq_IO_bound(bfqq); | |
1691 | } else | |
1692 | bfqq->requests_within_timer = 0; | |
1693 | } | |
1694 | ||
44e44a1b | 1695 | if (bfqd->low_latency) { |
36eca894 AA |
1696 | if (unlikely(time_is_after_jiffies(bfqq->split_time))) |
1697 | /* wraparound */ | |
1698 | bfqq->split_time = | |
1699 | jiffies - bfqd->bfq_wr_min_idle_time - 1; | |
1700 | ||
1701 | if (time_is_before_jiffies(bfqq->split_time + | |
1702 | bfqd->bfq_wr_min_idle_time)) { | |
1703 | bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq, | |
1704 | old_wr_coeff, | |
1705 | wr_or_deserves_wr, | |
1706 | *interactive, | |
e1b2324d | 1707 | in_burst, |
36eca894 AA |
1708 | soft_rt); |
1709 | ||
1710 | if (old_wr_coeff != bfqq->wr_coeff) | |
1711 | bfqq->entity.prio_changed = 1; | |
1712 | } | |
44e44a1b PV |
1713 | } |
1714 | ||
77b7dcea PV |
1715 | bfqq->last_idle_bklogged = jiffies; |
1716 | bfqq->service_from_backlogged = 0; | |
1717 | bfq_clear_bfqq_softrt_update(bfqq); | |
1718 | ||
aee69d78 PV |
1719 | bfq_add_bfqq_busy(bfqd, bfqq); |
1720 | ||
1721 | /* | |
1722 | * Expire in-service queue only if preemption may be needed | |
1723 | * for guarantees. In this respect, the function | |
1724 | * next_queue_may_preempt just checks a simple, necessary | |
1725 | * condition, and not a sufficient condition based on | |
1726 | * timestamps. In fact, for the latter condition to be | |
1727 | * evaluated, timestamps would need first to be updated, and | |
1728 | * this operation is quite costly (see the comments on the | |
1729 | * function bfq_bfqq_update_budg_for_activation). | |
1730 | */ | |
1731 | if (bfqd->in_service_queue && bfqq_wants_to_preempt && | |
77b7dcea | 1732 | bfqd->in_service_queue->wr_coeff < bfqq->wr_coeff && |
aee69d78 PV |
1733 | next_queue_may_preempt(bfqd)) |
1734 | bfq_bfqq_expire(bfqd, bfqd->in_service_queue, | |
1735 | false, BFQQE_PREEMPTED); | |
1736 | } | |
1737 | ||
1738 | static void bfq_add_request(struct request *rq) | |
1739 | { | |
1740 | struct bfq_queue *bfqq = RQ_BFQQ(rq); | |
1741 | struct bfq_data *bfqd = bfqq->bfqd; | |
1742 | struct request *next_rq, *prev; | |
44e44a1b PV |
1743 | unsigned int old_wr_coeff = bfqq->wr_coeff; |
1744 | bool interactive = false; | |
aee69d78 PV |
1745 | |
1746 | bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq)); | |
1747 | bfqq->queued[rq_is_sync(rq)]++; | |
1748 | bfqd->queued++; | |
1749 | ||
2341d662 PV |
1750 | if (RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_bfqq_sync(bfqq)) { |
1751 | /* | |
1752 | * Periodically reset inject limit, to make sure that | |
1753 | * the latter eventually drops in case workload | |
1754 | * changes, see step (3) in the comments on | |
1755 | * bfq_update_inject_limit(). | |
1756 | */ | |
1757 | if (time_is_before_eq_jiffies(bfqq->decrease_time_jif + | |
1758 | msecs_to_jiffies(1000))) { | |
1759 | /* invalidate baseline total service time */ | |
1760 | bfqq->last_serv_time_ns = 0; | |
1761 | ||
1762 | /* | |
1763 | * Reset pointer in case we are waiting for | |
1764 | * some request completion. | |
1765 | */ | |
1766 | bfqd->waited_rq = NULL; | |
1767 | ||
1768 | /* | |
1769 | * If bfqq has a short think time, then start | |
1770 | * by setting the inject limit to 0 | |
1771 | * prudentially, because the service time of | |
1772 | * an injected I/O request may be higher than | |
1773 | * the think time of bfqq, and therefore, if | |
1774 | * one request was injected when bfqq remains | |
1775 | * empty, this injected request might delay | |
1776 | * the service of the next I/O request for | |
1777 | * bfqq significantly. In case bfqq can | |
1778 | * actually tolerate some injection, then the | |
1779 | * adaptive update will however raise the | |
1780 | * limit soon. This lucky circumstance holds | |
1781 | * exactly because bfqq has a short think | |
1782 | * time, and thus, after remaining empty, is | |
1783 | * likely to get new I/O enqueued---and then | |
1784 | * completed---before being expired. This is | |
1785 | * the very pattern that gives the | |
1786 | * limit-update algorithm the chance to | |
1787 | * measure the effect of injection on request | |
1788 | * service times, and then to update the limit | |
1789 | * accordingly. | |
1790 | * | |
1791 | * On the opposite end, if bfqq has a long | |
1792 | * think time, then start directly by 1, | |
1793 | * because: | |
1794 | * a) on the bright side, keeping at most one | |
1795 | * request in service in the drive is unlikely | |
1796 | * to cause any harm to the latency of bfqq's | |
1797 | * requests, as the service time of a single | |
1798 | * request is likely to be lower than the | |
1799 | * think time of bfqq; | |
1800 | * b) on the downside, after becoming empty, | |
1801 | * bfqq is likely to expire before getting its | |
1802 | * next request. With this request arrival | |
1803 | * pattern, it is very hard to sample total | |
1804 | * service times and update the inject limit | |
1805 | * accordingly (see comments on | |
1806 | * bfq_update_inject_limit()). So the limit is | |
1807 | * likely to be never, or at least seldom, | |
1808 | * updated. As a consequence, by setting the | |
1809 | * limit to 1, we avoid that no injection ever | |
1810 | * occurs with bfqq. On the downside, this | |
1811 | * proactive step further reduces chances to | |
1812 | * actually compute the baseline total service | |
1813 | * time. Thus it reduces chances to execute the | |
1814 | * limit-update algorithm and possibly raise the | |
1815 | * limit to more than 1. | |
1816 | */ | |
1817 | if (bfq_bfqq_has_short_ttime(bfqq)) | |
1818 | bfqq->inject_limit = 0; | |
1819 | else | |
1820 | bfqq->inject_limit = 1; | |
1821 | bfqq->decrease_time_jif = jiffies; | |
1822 | } | |
1823 | ||
1824 | /* | |
1825 | * The following conditions must hold to setup a new | |
1826 | * sampling of total service time, and then a new | |
1827 | * update of the inject limit: | |
1828 | * - bfqq is in service, because the total service | |
1829 | * time is evaluated only for the I/O requests of | |
1830 | * the queues in service; | |
1831 | * - this is the right occasion to compute or to | |
1832 | * lower the baseline total service time, because | |
1833 | * there are actually no requests in the drive, | |
1834 | * or | |
1835 | * the baseline total service time is available, and | |
1836 | * this is the right occasion to compute the other | |
1837 | * quantity needed to update the inject limit, i.e., | |
1838 | * the total service time caused by the amount of | |
1839 | * injection allowed by the current value of the | |
1840 | * limit. It is the right occasion because injection | |
1841 | * has actually been performed during the service | |
1842 | * hole, and there are still in-flight requests, | |
1843 | * which are very likely to be exactly the injected | |
1844 | * requests, or part of them; | |
1845 | * - the minimum interval for sampling the total | |
1846 | * service time and updating the inject limit has | |
1847 | * elapsed. | |
1848 | */ | |
1849 | if (bfqq == bfqd->in_service_queue && | |
1850 | (bfqd->rq_in_driver == 0 || | |
1851 | (bfqq->last_serv_time_ns > 0 && | |
1852 | bfqd->rqs_injected && bfqd->rq_in_driver > 0)) && | |
1853 | time_is_before_eq_jiffies(bfqq->decrease_time_jif + | |
1854 | msecs_to_jiffies(100))) { | |
1855 | bfqd->last_empty_occupied_ns = ktime_get_ns(); | |
1856 | /* | |
1857 | * Start the state machine for measuring the | |
1858 | * total service time of rq: setting | |
1859 | * wait_dispatch will cause bfqd->waited_rq to | |
1860 | * be set when rq will be dispatched. | |
1861 | */ | |
1862 | bfqd->wait_dispatch = true; | |
1863 | bfqd->rqs_injected = false; | |
1864 | } | |
1865 | } | |
1866 | ||
aee69d78 PV |
1867 | elv_rb_add(&bfqq->sort_list, rq); |
1868 | ||
1869 | /* | |
1870 | * Check if this request is a better next-serve candidate. | |
1871 | */ | |
1872 | prev = bfqq->next_rq; | |
1873 | next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position); | |
1874 | bfqq->next_rq = next_rq; | |
1875 | ||
36eca894 AA |
1876 | /* |
1877 | * Adjust priority tree position, if next_rq changes. | |
8cacc5ab | 1878 | * See comments on bfq_pos_tree_add_move() for the unlikely(). |
36eca894 | 1879 | */ |
8cacc5ab | 1880 | if (unlikely(!bfqd->nonrot_with_queueing && prev != bfqq->next_rq)) |
36eca894 AA |
1881 | bfq_pos_tree_add_move(bfqd, bfqq); |
1882 | ||
aee69d78 | 1883 | if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */ |
44e44a1b PV |
1884 | bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff, |
1885 | rq, &interactive); | |
1886 | else { | |
1887 | if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) && | |
1888 | time_is_before_jiffies( | |
1889 | bfqq->last_wr_start_finish + | |
1890 | bfqd->bfq_wr_min_inter_arr_async)) { | |
1891 | bfqq->wr_coeff = bfqd->bfq_wr_coeff; | |
1892 | bfqq->wr_cur_max_time = bfq_wr_duration(bfqd); | |
1893 | ||
cfd69712 | 1894 | bfqd->wr_busy_queues++; |
44e44a1b PV |
1895 | bfqq->entity.prio_changed = 1; |
1896 | } | |
1897 | if (prev != bfqq->next_rq) | |
1898 | bfq_updated_next_req(bfqd, bfqq); | |
1899 | } | |
1900 | ||
1901 | /* | |
1902 | * Assign jiffies to last_wr_start_finish in the following | |
1903 | * cases: | |
1904 | * | |
1905 | * . if bfqq is not going to be weight-raised, because, for | |
1906 | * non weight-raised queues, last_wr_start_finish stores the | |
1907 | * arrival time of the last request; as of now, this piece | |
1908 | * of information is used only for deciding whether to | |
1909 | * weight-raise async queues | |
1910 | * | |
1911 | * . if bfqq is not weight-raised, because, if bfqq is now | |
1912 | * switching to weight-raised, then last_wr_start_finish | |
1913 | * stores the time when weight-raising starts | |
1914 | * | |
1915 | * . if bfqq is interactive, because, regardless of whether | |
1916 | * bfqq is currently weight-raised, the weight-raising | |
1917 | * period must start or restart (this case is considered | |
1918 | * separately because it is not detected by the above | |
1919 | * conditions, if bfqq is already weight-raised) | |
77b7dcea PV |
1920 | * |
1921 | * last_wr_start_finish has to be updated also if bfqq is soft | |
1922 | * real-time, because the weight-raising period is constantly | |
1923 | * restarted on idle-to-busy transitions for these queues, but | |
1924 | * this is already done in bfq_bfqq_handle_idle_busy_switch if | |
1925 | * needed. | |
44e44a1b PV |
1926 | */ |
1927 | if (bfqd->low_latency && | |
1928 | (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive)) | |
1929 | bfqq->last_wr_start_finish = jiffies; | |
aee69d78 PV |
1930 | } |
1931 | ||
1932 | static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd, | |
1933 | struct bio *bio, | |
1934 | struct request_queue *q) | |
1935 | { | |
1936 | struct bfq_queue *bfqq = bfqd->bio_bfqq; | |
1937 | ||
1938 | ||
1939 | if (bfqq) | |
1940 | return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio)); | |
1941 | ||
1942 | return NULL; | |
1943 | } | |
1944 | ||
ab0e43e9 PV |
1945 | static sector_t get_sdist(sector_t last_pos, struct request *rq) |
1946 | { | |
1947 | if (last_pos) | |
1948 | return abs(blk_rq_pos(rq) - last_pos); | |
1949 | ||
1950 | return 0; | |
1951 | } | |
1952 | ||
aee69d78 PV |
1953 | #if 0 /* Still not clear if we can do without next two functions */ |
1954 | static void bfq_activate_request(struct request_queue *q, struct request *rq) | |
1955 | { | |
1956 | struct bfq_data *bfqd = q->elevator->elevator_data; | |
1957 | ||
1958 | bfqd->rq_in_driver++; | |
aee69d78 PV |
1959 | } |
1960 | ||
1961 | static void bfq_deactivate_request(struct request_queue *q, struct request *rq) | |
1962 | { | |
1963 | struct bfq_data *bfqd = q->elevator->elevator_data; | |
1964 | ||
1965 | bfqd->rq_in_driver--; | |
1966 | } | |
1967 | #endif | |
1968 | ||
1969 | static void bfq_remove_request(struct request_queue *q, | |
1970 | struct request *rq) | |
1971 | { | |
1972 | struct bfq_queue *bfqq = RQ_BFQQ(rq); | |
1973 | struct bfq_data *bfqd = bfqq->bfqd; | |
1974 | const int sync = rq_is_sync(rq); | |
1975 | ||
1976 | if (bfqq->next_rq == rq) { | |
1977 | bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq); | |
1978 | bfq_updated_next_req(bfqd, bfqq); | |
1979 | } | |
1980 | ||
1981 | if (rq->queuelist.prev != &rq->queuelist) | |
1982 | list_del_init(&rq->queuelist); | |
1983 | bfqq->queued[sync]--; | |
1984 | bfqd->queued--; | |
1985 | elv_rb_del(&bfqq->sort_list, rq); | |
1986 | ||
1987 | elv_rqhash_del(q, rq); | |
1988 | if (q->last_merge == rq) | |
1989 | q->last_merge = NULL; | |
1990 | ||
1991 | if (RB_EMPTY_ROOT(&bfqq->sort_list)) { | |
1992 | bfqq->next_rq = NULL; | |
1993 | ||
1994 | if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) { | |
e21b7a0b | 1995 | bfq_del_bfqq_busy(bfqd, bfqq, false); |
aee69d78 PV |
1996 | /* |
1997 | * bfqq emptied. In normal operation, when | |
1998 | * bfqq is empty, bfqq->entity.service and | |
1999 | * bfqq->entity.budget must contain, | |
2000 | * respectively, the service received and the | |
2001 | * budget used last time bfqq emptied. These | |
2002 | * facts do not hold in this case, as at least | |
2003 | * this last removal occurred while bfqq is | |
2004 | * not in service. To avoid inconsistencies, | |
2005 | * reset both bfqq->entity.service and | |
2006 | * bfqq->entity.budget, if bfqq has still a | |
2007 | * process that may issue I/O requests to it. | |
2008 | */ | |
2009 | bfqq->entity.budget = bfqq->entity.service = 0; | |
2010 | } | |
36eca894 AA |
2011 | |
2012 | /* | |
2013 | * Remove queue from request-position tree as it is empty. | |
2014 | */ | |
2015 | if (bfqq->pos_root) { | |
2016 | rb_erase(&bfqq->pos_node, bfqq->pos_root); | |
2017 | bfqq->pos_root = NULL; | |
2018 | } | |
05e90283 | 2019 | } else { |
8cacc5ab PV |
2020 | /* see comments on bfq_pos_tree_add_move() for the unlikely() */ |
2021 | if (unlikely(!bfqd->nonrot_with_queueing)) | |
2022 | bfq_pos_tree_add_move(bfqd, bfqq); | |
aee69d78 PV |
2023 | } |
2024 | ||
2025 | if (rq->cmd_flags & REQ_META) | |
2026 | bfqq->meta_pending--; | |
e21b7a0b | 2027 | |
aee69d78 PV |
2028 | } |
2029 | ||
2030 | static bool bfq_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio) | |
2031 | { | |
2032 | struct request_queue *q = hctx->queue; | |
2033 | struct bfq_data *bfqd = q->elevator->elevator_data; | |
2034 | struct request *free = NULL; | |
2035 | /* | |
2036 | * bfq_bic_lookup grabs the queue_lock: invoke it now and | |
2037 | * store its return value for later use, to avoid nesting | |
2038 | * queue_lock inside the bfqd->lock. We assume that the bic | |
2039 | * returned by bfq_bic_lookup does not go away before | |
2040 | * bfqd->lock is taken. | |
2041 | */ | |
2042 | struct bfq_io_cq *bic = bfq_bic_lookup(bfqd, current->io_context, q); | |
2043 | bool ret; | |
2044 | ||
2045 | spin_lock_irq(&bfqd->lock); | |
2046 | ||
2047 | if (bic) | |
2048 | bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf)); | |
2049 | else | |
2050 | bfqd->bio_bfqq = NULL; | |
2051 | bfqd->bio_bic = bic; | |
2052 | ||
2053 | ret = blk_mq_sched_try_merge(q, bio, &free); | |
2054 | ||
2055 | if (free) | |
2056 | blk_mq_free_request(free); | |
2057 | spin_unlock_irq(&bfqd->lock); | |
2058 | ||
2059 | return ret; | |
2060 | } | |
2061 | ||
2062 | static int bfq_request_merge(struct request_queue *q, struct request **req, | |
2063 | struct bio *bio) | |
2064 | { | |
2065 | struct bfq_data *bfqd = q->elevator->elevator_data; | |
2066 | struct request *__rq; | |
2067 | ||
2068 | __rq = bfq_find_rq_fmerge(bfqd, bio, q); | |
2069 | if (__rq && elv_bio_merge_ok(__rq, bio)) { | |
2070 | *req = __rq; | |
2071 | return ELEVATOR_FRONT_MERGE; | |
2072 | } | |
2073 | ||
2074 | return ELEVATOR_NO_MERGE; | |
2075 | } | |
2076 | ||
18e5a57d PV |
2077 | static struct bfq_queue *bfq_init_rq(struct request *rq); |
2078 | ||
aee69d78 PV |
2079 | static void bfq_request_merged(struct request_queue *q, struct request *req, |
2080 | enum elv_merge type) | |
2081 | { | |
2082 | if (type == ELEVATOR_FRONT_MERGE && | |
2083 | rb_prev(&req->rb_node) && | |
2084 | blk_rq_pos(req) < | |
2085 | blk_rq_pos(container_of(rb_prev(&req->rb_node), | |
2086 | struct request, rb_node))) { | |
18e5a57d | 2087 | struct bfq_queue *bfqq = bfq_init_rq(req); |
aee69d78 PV |
2088 | struct bfq_data *bfqd = bfqq->bfqd; |
2089 | struct request *prev, *next_rq; | |
2090 | ||
2091 | /* Reposition request in its sort_list */ | |
2092 | elv_rb_del(&bfqq->sort_list, req); | |
2093 | elv_rb_add(&bfqq->sort_list, req); | |
2094 | ||
2095 | /* Choose next request to be served for bfqq */ | |
2096 | prev = bfqq->next_rq; | |
2097 | next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req, | |
2098 | bfqd->last_position); | |
2099 | bfqq->next_rq = next_rq; | |
2100 | /* | |
36eca894 AA |
2101 | * If next_rq changes, update both the queue's budget to |
2102 | * fit the new request and the queue's position in its | |
2103 | * rq_pos_tree. | |
aee69d78 | 2104 | */ |
36eca894 | 2105 | if (prev != bfqq->next_rq) { |
aee69d78 | 2106 | bfq_updated_next_req(bfqd, bfqq); |
8cacc5ab PV |
2107 | /* |
2108 | * See comments on bfq_pos_tree_add_move() for | |
2109 | * the unlikely(). | |
2110 | */ | |
2111 | if (unlikely(!bfqd->nonrot_with_queueing)) | |
2112 | bfq_pos_tree_add_move(bfqd, bfqq); | |
36eca894 | 2113 | } |
aee69d78 PV |
2114 | } |
2115 | } | |
2116 | ||
8abfa4d6 PV |
2117 | /* |
2118 | * This function is called to notify the scheduler that the requests | |
2119 | * rq and 'next' have been merged, with 'next' going away. BFQ | |
2120 | * exploits this hook to address the following issue: if 'next' has a | |
2121 | * fifo_time lower that rq, then the fifo_time of rq must be set to | |
2122 | * the value of 'next', to not forget the greater age of 'next'. | |
8abfa4d6 PV |
2123 | * |
2124 | * NOTE: in this function we assume that rq is in a bfq_queue, basing | |
2125 | * on that rq is picked from the hash table q->elevator->hash, which, | |
2126 | * in its turn, is filled only with I/O requests present in | |
2127 | * bfq_queues, while BFQ is in use for the request queue q. In fact, | |
2128 | * the function that fills this hash table (elv_rqhash_add) is called | |
2129 | * only by bfq_insert_request. | |
2130 | */ | |
aee69d78 PV |
2131 | static void bfq_requests_merged(struct request_queue *q, struct request *rq, |
2132 | struct request *next) | |
2133 | { | |
18e5a57d PV |
2134 | struct bfq_queue *bfqq = bfq_init_rq(rq), |
2135 | *next_bfqq = bfq_init_rq(next); | |
aee69d78 | 2136 | |
aee69d78 PV |
2137 | /* |
2138 | * If next and rq belong to the same bfq_queue and next is older | |
2139 | * than rq, then reposition rq in the fifo (by substituting next | |
2140 | * with rq). Otherwise, if next and rq belong to different | |
2141 | * bfq_queues, never reposition rq: in fact, we would have to | |
2142 | * reposition it with respect to next's position in its own fifo, | |
2143 | * which would most certainly be too expensive with respect to | |
2144 | * the benefits. | |
2145 | */ | |
2146 | if (bfqq == next_bfqq && | |
2147 | !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && | |
2148 | next->fifo_time < rq->fifo_time) { | |
2149 | list_del_init(&rq->queuelist); | |
2150 | list_replace_init(&next->queuelist, &rq->queuelist); | |
2151 | rq->fifo_time = next->fifo_time; | |
2152 | } | |
2153 | ||
2154 | if (bfqq->next_rq == next) | |
2155 | bfqq->next_rq = rq; | |
2156 | ||
e21b7a0b | 2157 | bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags); |
aee69d78 PV |
2158 | } |
2159 | ||
44e44a1b PV |
2160 | /* Must be called with bfqq != NULL */ |
2161 | static void bfq_bfqq_end_wr(struct bfq_queue *bfqq) | |
2162 | { | |
cfd69712 PV |
2163 | if (bfq_bfqq_busy(bfqq)) |
2164 | bfqq->bfqd->wr_busy_queues--; | |
44e44a1b PV |
2165 | bfqq->wr_coeff = 1; |
2166 | bfqq->wr_cur_max_time = 0; | |
77b7dcea | 2167 | bfqq->last_wr_start_finish = jiffies; |
44e44a1b PV |
2168 | /* |
2169 | * Trigger a weight change on the next invocation of | |
2170 | * __bfq_entity_update_weight_prio. | |
2171 | */ | |
2172 | bfqq->entity.prio_changed = 1; | |
2173 | } | |
2174 | ||
ea25da48 PV |
2175 | void bfq_end_wr_async_queues(struct bfq_data *bfqd, |
2176 | struct bfq_group *bfqg) | |
44e44a1b PV |
2177 | { |
2178 | int i, j; | |
2179 | ||
2180 | for (i = 0; i < 2; i++) | |
2181 | for (j = 0; j < IOPRIO_BE_NR; j++) | |
2182 | if (bfqg->async_bfqq[i][j]) | |
2183 | bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]); | |
2184 | if (bfqg->async_idle_bfqq) | |
2185 | bfq_bfqq_end_wr(bfqg->async_idle_bfqq); | |
2186 | } | |
2187 | ||
2188 | static void bfq_end_wr(struct bfq_data *bfqd) | |
2189 | { | |
2190 | struct bfq_queue *bfqq; | |
2191 | ||
2192 | spin_lock_irq(&bfqd->lock); | |
2193 | ||
2194 | list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list) | |
2195 | bfq_bfqq_end_wr(bfqq); | |
2196 | list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list) | |
2197 | bfq_bfqq_end_wr(bfqq); | |
2198 | bfq_end_wr_async(bfqd); | |
2199 | ||
2200 | spin_unlock_irq(&bfqd->lock); | |
2201 | } | |
2202 | ||
36eca894 AA |
2203 | static sector_t bfq_io_struct_pos(void *io_struct, bool request) |
2204 | { | |
2205 | if (request) | |
2206 | return blk_rq_pos(io_struct); | |
2207 | else | |
2208 | return ((struct bio *)io_struct)->bi_iter.bi_sector; | |
2209 | } | |
2210 | ||
2211 | static int bfq_rq_close_to_sector(void *io_struct, bool request, | |
2212 | sector_t sector) | |
2213 | { | |
2214 | return abs(bfq_io_struct_pos(io_struct, request) - sector) <= | |
2215 | BFQQ_CLOSE_THR; | |
2216 | } | |
2217 | ||
2218 | static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd, | |
2219 | struct bfq_queue *bfqq, | |
2220 | sector_t sector) | |
2221 | { | |
2222 | struct rb_root *root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree; | |
2223 | struct rb_node *parent, *node; | |
2224 | struct bfq_queue *__bfqq; | |
2225 | ||
2226 | if (RB_EMPTY_ROOT(root)) | |
2227 | return NULL; | |
2228 | ||
2229 | /* | |
2230 | * First, if we find a request starting at the end of the last | |
2231 | * request, choose it. | |
2232 | */ | |
2233 | __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL); | |
2234 | if (__bfqq) | |
2235 | return __bfqq; | |
2236 | ||
2237 | /* | |
2238 | * If the exact sector wasn't found, the parent of the NULL leaf | |
2239 | * will contain the closest sector (rq_pos_tree sorted by | |
2240 | * next_request position). | |
2241 | */ | |
2242 | __bfqq = rb_entry(parent, struct bfq_queue, pos_node); | |
2243 | if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector)) | |
2244 | return __bfqq; | |
2245 | ||
2246 | if (blk_rq_pos(__bfqq->next_rq) < sector) | |
2247 | node = rb_next(&__bfqq->pos_node); | |
2248 | else | |
2249 | node = rb_prev(&__bfqq->pos_node); | |
2250 | if (!node) | |
2251 | return NULL; | |
2252 | ||
2253 | __bfqq = rb_entry(node, struct bfq_queue, pos_node); | |
2254 | if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector)) | |
2255 | return __bfqq; | |
2256 | ||
2257 | return NULL; | |
2258 | } | |
2259 | ||
2260 | static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd, | |
2261 | struct bfq_queue *cur_bfqq, | |
2262 | sector_t sector) | |
2263 | { | |
2264 | struct bfq_queue *bfqq; | |
2265 | ||
2266 | /* | |
2267 | * We shall notice if some of the queues are cooperating, | |
2268 | * e.g., working closely on the same area of the device. In | |
2269 | * that case, we can group them together and: 1) don't waste | |
2270 | * time idling, and 2) serve the union of their requests in | |
2271 | * the best possible order for throughput. | |
2272 | */ | |
2273 | bfqq = bfqq_find_close(bfqd, cur_bfqq, sector); | |
2274 | if (!bfqq || bfqq == cur_bfqq) | |
2275 | return NULL; | |
2276 | ||
2277 | return bfqq; | |
2278 | } | |
2279 | ||
2280 | static struct bfq_queue * | |
2281 | bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq) | |
2282 | { | |
2283 | int process_refs, new_process_refs; | |
2284 | struct bfq_queue *__bfqq; | |
2285 | ||
2286 | /* | |
2287 | * If there are no process references on the new_bfqq, then it is | |
2288 | * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain | |
2289 | * may have dropped their last reference (not just their last process | |
2290 | * reference). | |
2291 | */ | |
2292 | if (!bfqq_process_refs(new_bfqq)) | |
2293 | return NULL; | |
2294 | ||
2295 | /* Avoid a circular list and skip interim queue merges. */ | |
2296 | while ((__bfqq = new_bfqq->new_bfqq)) { | |
2297 | if (__bfqq == bfqq) | |
2298 | return NULL; | |
2299 | new_bfqq = __bfqq; | |
2300 | } | |
2301 | ||
2302 | process_refs = bfqq_process_refs(bfqq); | |
2303 | new_process_refs = bfqq_process_refs(new_bfqq); | |
2304 | /* | |
2305 | * If the process for the bfqq has gone away, there is no | |
2306 | * sense in merging the queues. | |
2307 | */ | |
2308 | if (process_refs == 0 || new_process_refs == 0) | |
2309 | return NULL; | |
2310 | ||
2311 | bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d", | |
2312 | new_bfqq->pid); | |
2313 | ||
2314 | /* | |
2315 | * Merging is just a redirection: the requests of the process | |
2316 | * owning one of the two queues are redirected to the other queue. | |
2317 | * The latter queue, in its turn, is set as shared if this is the | |
2318 | * first time that the requests of some process are redirected to | |
2319 | * it. | |
2320 | * | |
6fa3e8d3 PV |
2321 | * We redirect bfqq to new_bfqq and not the opposite, because |
2322 | * we are in the context of the process owning bfqq, thus we | |
2323 | * have the io_cq of this process. So we can immediately | |
2324 | * configure this io_cq to redirect the requests of the | |
2325 | * process to new_bfqq. In contrast, the io_cq of new_bfqq is | |
2326 | * not available any more (new_bfqq->bic == NULL). | |
36eca894 | 2327 | * |
6fa3e8d3 PV |
2328 | * Anyway, even in case new_bfqq coincides with the in-service |
2329 | * queue, redirecting requests the in-service queue is the | |
2330 | * best option, as we feed the in-service queue with new | |
2331 | * requests close to the last request served and, by doing so, | |
2332 | * are likely to increase the throughput. | |
36eca894 AA |
2333 | */ |
2334 | bfqq->new_bfqq = new_bfqq; | |
2335 | new_bfqq->ref += process_refs; | |
2336 | return new_bfqq; | |
2337 | } | |
2338 | ||
2339 | static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq, | |
2340 | struct bfq_queue *new_bfqq) | |
2341 | { | |
7b8fa3b9 PV |
2342 | if (bfq_too_late_for_merging(new_bfqq)) |
2343 | return false; | |
2344 | ||
36eca894 AA |
2345 | if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) || |
2346 | (bfqq->ioprio_class != new_bfqq->ioprio_class)) | |
2347 | return false; | |
2348 | ||
2349 | /* | |
2350 | * If either of the queues has already been detected as seeky, | |
2351 | * then merging it with the other queue is unlikely to lead to | |
2352 | * sequential I/O. | |
2353 | */ | |
2354 | if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq)) | |
2355 | return false; | |
2356 | ||
2357 | /* | |
2358 | * Interleaved I/O is known to be done by (some) applications | |
2359 | * only for reads, so it does not make sense to merge async | |
2360 | * queues. | |
2361 | */ | |
2362 | if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq)) | |
2363 | return false; | |
2364 | ||
2365 | return true; | |
2366 | } | |
2367 | ||
36eca894 AA |
2368 | /* |
2369 | * Attempt to schedule a merge of bfqq with the currently in-service | |
2370 | * queue or with a close queue among the scheduled queues. Return | |
2371 | * NULL if no merge was scheduled, a pointer to the shared bfq_queue | |
2372 | * structure otherwise. | |
2373 | * | |
2374 | * The OOM queue is not allowed to participate to cooperation: in fact, since | |
2375 | * the requests temporarily redirected to the OOM queue could be redirected | |
2376 | * again to dedicated queues at any time, the state needed to correctly | |
2377 | * handle merging with the OOM queue would be quite complex and expensive | |
2378 | * to maintain. Besides, in such a critical condition as an out of memory, | |
2379 | * the benefits of queue merging may be little relevant, or even negligible. | |
2380 | * | |
36eca894 AA |
2381 | * WARNING: queue merging may impair fairness among non-weight raised |
2382 | * queues, for at least two reasons: 1) the original weight of a | |
2383 | * merged queue may change during the merged state, 2) even being the | |
2384 | * weight the same, a merged queue may be bloated with many more | |
2385 | * requests than the ones produced by its originally-associated | |
2386 | * process. | |
2387 | */ | |
2388 | static struct bfq_queue * | |
2389 | bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
2390 | void *io_struct, bool request) | |
2391 | { | |
2392 | struct bfq_queue *in_service_bfqq, *new_bfqq; | |
2393 | ||
8cacc5ab PV |
2394 | /* |
2395 | * Do not perform queue merging if the device is non | |
2396 | * rotational and performs internal queueing. In fact, such a | |
2397 | * device reaches a high speed through internal parallelism | |
2398 | * and pipelining. This means that, to reach a high | |
2399 | * throughput, it must have many requests enqueued at the same | |
2400 | * time. But, in this configuration, the internal scheduling | |
2401 | * algorithm of the device does exactly the job of queue | |
2402 | * merging: it reorders requests so as to obtain as much as | |
2403 | * possible a sequential I/O pattern. As a consequence, with | |
2404 | * the workload generated by processes doing interleaved I/O, | |
2405 | * the throughput reached by the device is likely to be the | |
2406 | * same, with and without queue merging. | |
2407 | * | |
2408 | * Disabling merging also provides a remarkable benefit in | |
2409 | * terms of throughput. Merging tends to make many workloads | |
2410 | * artificially more uneven, because of shared queues | |
2411 | * remaining non empty for incomparably more time than | |
2412 | * non-merged queues. This may accentuate workload | |
2413 | * asymmetries. For example, if one of the queues in a set of | |
2414 | * merged queues has a higher weight than a normal queue, then | |
2415 | * the shared queue may inherit such a high weight and, by | |
2416 | * staying almost always active, may force BFQ to perform I/O | |
2417 | * plugging most of the time. This evidently makes it harder | |
2418 | * for BFQ to let the device reach a high throughput. | |
2419 | * | |
2420 | * Finally, the likely() macro below is not used because one | |
2421 | * of the two branches is more likely than the other, but to | |
2422 | * have the code path after the following if() executed as | |
2423 | * fast as possible for the case of a non rotational device | |
2424 | * with queueing. We want it because this is the fastest kind | |
2425 | * of device. On the opposite end, the likely() may lengthen | |
2426 | * the execution time of BFQ for the case of slower devices | |
2427 | * (rotational or at least without queueing). But in this case | |
2428 | * the execution time of BFQ matters very little, if not at | |
2429 | * all. | |
2430 | */ | |
2431 | if (likely(bfqd->nonrot_with_queueing)) | |
2432 | return NULL; | |
2433 | ||
7b8fa3b9 PV |
2434 | /* |
2435 | * Prevent bfqq from being merged if it has been created too | |
2436 | * long ago. The idea is that true cooperating processes, and | |
2437 | * thus their associated bfq_queues, are supposed to be | |
2438 | * created shortly after each other. This is the case, e.g., | |
2439 | * for KVM/QEMU and dump I/O threads. Basing on this | |
2440 | * assumption, the following filtering greatly reduces the | |
2441 | * probability that two non-cooperating processes, which just | |
2442 | * happen to do close I/O for some short time interval, have | |
2443 | * their queues merged by mistake. | |
2444 | */ | |
2445 | if (bfq_too_late_for_merging(bfqq)) | |
2446 | return NULL; | |
2447 | ||
36eca894 AA |
2448 | if (bfqq->new_bfqq) |
2449 | return bfqq->new_bfqq; | |
2450 | ||
4403e4e4 | 2451 | if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq)) |
36eca894 AA |
2452 | return NULL; |
2453 | ||
2454 | /* If there is only one backlogged queue, don't search. */ | |
73d58118 | 2455 | if (bfq_tot_busy_queues(bfqd) == 1) |
36eca894 AA |
2456 | return NULL; |
2457 | ||
2458 | in_service_bfqq = bfqd->in_service_queue; | |
2459 | ||
4403e4e4 AR |
2460 | if (in_service_bfqq && in_service_bfqq != bfqq && |
2461 | likely(in_service_bfqq != &bfqd->oom_bfqq) && | |
058fdecc PV |
2462 | bfq_rq_close_to_sector(io_struct, request, |
2463 | bfqd->in_serv_last_pos) && | |
36eca894 AA |
2464 | bfqq->entity.parent == in_service_bfqq->entity.parent && |
2465 | bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) { | |
2466 | new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq); | |
2467 | if (new_bfqq) | |
2468 | return new_bfqq; | |
2469 | } | |
2470 | /* | |
2471 | * Check whether there is a cooperator among currently scheduled | |
2472 | * queues. The only thing we need is that the bio/request is not | |
2473 | * NULL, as we need it to establish whether a cooperator exists. | |
2474 | */ | |
36eca894 AA |
2475 | new_bfqq = bfq_find_close_cooperator(bfqd, bfqq, |
2476 | bfq_io_struct_pos(io_struct, request)); | |
2477 | ||
4403e4e4 | 2478 | if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) && |
36eca894 AA |
2479 | bfq_may_be_close_cooperator(bfqq, new_bfqq)) |
2480 | return bfq_setup_merge(bfqq, new_bfqq); | |
2481 | ||
2482 | return NULL; | |
2483 | } | |
2484 | ||
2485 | static void bfq_bfqq_save_state(struct bfq_queue *bfqq) | |
2486 | { | |
2487 | struct bfq_io_cq *bic = bfqq->bic; | |
2488 | ||
2489 | /* | |
2490 | * If !bfqq->bic, the queue is already shared or its requests | |
2491 | * have already been redirected to a shared queue; both idle window | |
2492 | * and weight raising state have already been saved. Do nothing. | |
2493 | */ | |
2494 | if (!bic) | |
2495 | return; | |
2496 | ||
fffca087 | 2497 | bic->saved_weight = bfqq->entity.orig_weight; |
36eca894 | 2498 | bic->saved_ttime = bfqq->ttime; |
d5be3fef | 2499 | bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq); |
36eca894 | 2500 | bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq); |
e1b2324d AA |
2501 | bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq); |
2502 | bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node); | |
894df937 | 2503 | if (unlikely(bfq_bfqq_just_created(bfqq) && |
1be6e8a9 AR |
2504 | !bfq_bfqq_in_large_burst(bfqq) && |
2505 | bfqq->bfqd->low_latency)) { | |
894df937 PV |
2506 | /* |
2507 | * bfqq being merged right after being created: bfqq | |
2508 | * would have deserved interactive weight raising, but | |
2509 | * did not make it to be set in a weight-raised state, | |
2510 | * because of this early merge. Store directly the | |
2511 | * weight-raising state that would have been assigned | |
2512 | * to bfqq, so that to avoid that bfqq unjustly fails | |
2513 | * to enjoy weight raising if split soon. | |
2514 | */ | |
2515 | bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff; | |
2516 | bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd); | |
2517 | bic->saved_last_wr_start_finish = jiffies; | |
2518 | } else { | |
2519 | bic->saved_wr_coeff = bfqq->wr_coeff; | |
2520 | bic->saved_wr_start_at_switch_to_srt = | |
2521 | bfqq->wr_start_at_switch_to_srt; | |
2522 | bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish; | |
2523 | bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time; | |
2524 | } | |
36eca894 AA |
2525 | } |
2526 | ||
36eca894 AA |
2527 | static void |
2528 | bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic, | |
2529 | struct bfq_queue *bfqq, struct bfq_queue *new_bfqq) | |
2530 | { | |
2531 | bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu", | |
2532 | (unsigned long)new_bfqq->pid); | |
2533 | /* Save weight raising and idle window of the merged queues */ | |
2534 | bfq_bfqq_save_state(bfqq); | |
2535 | bfq_bfqq_save_state(new_bfqq); | |
2536 | if (bfq_bfqq_IO_bound(bfqq)) | |
2537 | bfq_mark_bfqq_IO_bound(new_bfqq); | |
2538 | bfq_clear_bfqq_IO_bound(bfqq); | |
2539 | ||
2540 | /* | |
2541 | * If bfqq is weight-raised, then let new_bfqq inherit | |
2542 | * weight-raising. To reduce false positives, neglect the case | |
2543 | * where bfqq has just been created, but has not yet made it | |
2544 | * to be weight-raised (which may happen because EQM may merge | |
2545 | * bfqq even before bfq_add_request is executed for the first | |
e1b2324d AA |
2546 | * time for bfqq). Handling this case would however be very |
2547 | * easy, thanks to the flag just_created. | |
36eca894 AA |
2548 | */ |
2549 | if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) { | |
2550 | new_bfqq->wr_coeff = bfqq->wr_coeff; | |
2551 | new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time; | |
2552 | new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish; | |
2553 | new_bfqq->wr_start_at_switch_to_srt = | |
2554 | bfqq->wr_start_at_switch_to_srt; | |
2555 | if (bfq_bfqq_busy(new_bfqq)) | |
2556 | bfqd->wr_busy_queues++; | |
2557 | new_bfqq->entity.prio_changed = 1; | |
2558 | } | |
2559 | ||
2560 | if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */ | |
2561 | bfqq->wr_coeff = 1; | |
2562 | bfqq->entity.prio_changed = 1; | |
2563 | if (bfq_bfqq_busy(bfqq)) | |
2564 | bfqd->wr_busy_queues--; | |
2565 | } | |
2566 | ||
2567 | bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d", | |
2568 | bfqd->wr_busy_queues); | |
2569 | ||
36eca894 AA |
2570 | /* |
2571 | * Merge queues (that is, let bic redirect its requests to new_bfqq) | |
2572 | */ | |
2573 | bic_set_bfqq(bic, new_bfqq, 1); | |
2574 | bfq_mark_bfqq_coop(new_bfqq); | |
2575 | /* | |
2576 | * new_bfqq now belongs to at least two bics (it is a shared queue): | |
2577 | * set new_bfqq->bic to NULL. bfqq either: | |
2578 | * - does not belong to any bic any more, and hence bfqq->bic must | |
2579 | * be set to NULL, or | |
2580 | * - is a queue whose owning bics have already been redirected to a | |
2581 | * different queue, hence the queue is destined to not belong to | |
2582 | * any bic soon and bfqq->bic is already NULL (therefore the next | |
2583 | * assignment causes no harm). | |
2584 | */ | |
2585 | new_bfqq->bic = NULL; | |
1e66413c FP |
2586 | /* |
2587 | * If the queue is shared, the pid is the pid of one of the associated | |
2588 | * processes. Which pid depends on the exact sequence of merge events | |
2589 | * the queue underwent. So printing such a pid is useless and confusing | |
2590 | * because it reports a random pid between those of the associated | |
2591 | * processes. | |
2592 | * We mark such a queue with a pid -1, and then print SHARED instead of | |
2593 | * a pid in logging messages. | |
2594 | */ | |
2595 | new_bfqq->pid = -1; | |
36eca894 AA |
2596 | bfqq->bic = NULL; |
2597 | /* release process reference to bfqq */ | |
2598 | bfq_put_queue(bfqq); | |
2599 | } | |
2600 | ||
aee69d78 PV |
2601 | static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq, |
2602 | struct bio *bio) | |
2603 | { | |
2604 | struct bfq_data *bfqd = q->elevator->elevator_data; | |
2605 | bool is_sync = op_is_sync(bio->bi_opf); | |
36eca894 | 2606 | struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq; |
aee69d78 PV |
2607 | |
2608 | /* | |
2609 | * Disallow merge of a sync bio into an async request. | |
2610 | */ | |
2611 | if (is_sync && !rq_is_sync(rq)) | |
2612 | return false; | |
2613 | ||
2614 | /* | |
2615 | * Lookup the bfqq that this bio will be queued with. Allow | |
2616 | * merge only if rq is queued there. | |
2617 | */ | |
2618 | if (!bfqq) | |
2619 | return false; | |
2620 | ||
36eca894 AA |
2621 | /* |
2622 | * We take advantage of this function to perform an early merge | |
2623 | * of the queues of possible cooperating processes. | |
2624 | */ | |
2625 | new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false); | |
2626 | if (new_bfqq) { | |
2627 | /* | |
2628 | * bic still points to bfqq, then it has not yet been | |
2629 | * redirected to some other bfq_queue, and a queue | |
636b8fe8 AR |
2630 | * merge between bfqq and new_bfqq can be safely |
2631 | * fulfilled, i.e., bic can be redirected to new_bfqq | |
36eca894 AA |
2632 | * and bfqq can be put. |
2633 | */ | |
2634 | bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq, | |
2635 | new_bfqq); | |
2636 | /* | |
2637 | * If we get here, bio will be queued into new_queue, | |
2638 | * so use new_bfqq to decide whether bio and rq can be | |
2639 | * merged. | |
2640 | */ | |
2641 | bfqq = new_bfqq; | |
2642 | ||
2643 | /* | |
2644 | * Change also bqfd->bio_bfqq, as | |
2645 | * bfqd->bio_bic now points to new_bfqq, and | |
2646 | * this function may be invoked again (and then may | |
2647 | * use again bqfd->bio_bfqq). | |
2648 | */ | |
2649 | bfqd->bio_bfqq = bfqq; | |
2650 | } | |
2651 | ||
aee69d78 PV |
2652 | return bfqq == RQ_BFQQ(rq); |
2653 | } | |
2654 | ||
44e44a1b PV |
2655 | /* |
2656 | * Set the maximum time for the in-service queue to consume its | |
2657 | * budget. This prevents seeky processes from lowering the throughput. | |
2658 | * In practice, a time-slice service scheme is used with seeky | |
2659 | * processes. | |
2660 | */ | |
2661 | static void bfq_set_budget_timeout(struct bfq_data *bfqd, | |
2662 | struct bfq_queue *bfqq) | |
2663 | { | |
77b7dcea PV |
2664 | unsigned int timeout_coeff; |
2665 | ||
2666 | if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time) | |
2667 | timeout_coeff = 1; | |
2668 | else | |
2669 | timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight; | |
2670 | ||
44e44a1b PV |
2671 | bfqd->last_budget_start = ktime_get(); |
2672 | ||
2673 | bfqq->budget_timeout = jiffies + | |
77b7dcea | 2674 | bfqd->bfq_timeout * timeout_coeff; |
44e44a1b PV |
2675 | } |
2676 | ||
aee69d78 PV |
2677 | static void __bfq_set_in_service_queue(struct bfq_data *bfqd, |
2678 | struct bfq_queue *bfqq) | |
2679 | { | |
2680 | if (bfqq) { | |
aee69d78 PV |
2681 | bfq_clear_bfqq_fifo_expire(bfqq); |
2682 | ||
2683 | bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8; | |
2684 | ||
77b7dcea PV |
2685 | if (time_is_before_jiffies(bfqq->last_wr_start_finish) && |
2686 | bfqq->wr_coeff > 1 && | |
2687 | bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time && | |
2688 | time_is_before_jiffies(bfqq->budget_timeout)) { | |
2689 | /* | |
2690 | * For soft real-time queues, move the start | |
2691 | * of the weight-raising period forward by the | |
2692 | * time the queue has not received any | |
2693 | * service. Otherwise, a relatively long | |
2694 | * service delay is likely to cause the | |
2695 | * weight-raising period of the queue to end, | |
2696 | * because of the short duration of the | |
2697 | * weight-raising period of a soft real-time | |
2698 | * queue. It is worth noting that this move | |
2699 | * is not so dangerous for the other queues, | |
2700 | * because soft real-time queues are not | |
2701 | * greedy. | |
2702 | * | |
2703 | * To not add a further variable, we use the | |
2704 | * overloaded field budget_timeout to | |
2705 | * determine for how long the queue has not | |
2706 | * received service, i.e., how much time has | |
2707 | * elapsed since the queue expired. However, | |
2708 | * this is a little imprecise, because | |
2709 | * budget_timeout is set to jiffies if bfqq | |
2710 | * not only expires, but also remains with no | |
2711 | * request. | |
2712 | */ | |
2713 | if (time_after(bfqq->budget_timeout, | |
2714 | bfqq->last_wr_start_finish)) | |
2715 | bfqq->last_wr_start_finish += | |
2716 | jiffies - bfqq->budget_timeout; | |
2717 | else | |
2718 | bfqq->last_wr_start_finish = jiffies; | |
2719 | } | |
2720 | ||
44e44a1b | 2721 | bfq_set_budget_timeout(bfqd, bfqq); |
aee69d78 PV |
2722 | bfq_log_bfqq(bfqd, bfqq, |
2723 | "set_in_service_queue, cur-budget = %d", | |
2724 | bfqq->entity.budget); | |
2725 | } | |
2726 | ||
2727 | bfqd->in_service_queue = bfqq; | |
2728 | } | |
2729 | ||
2730 | /* | |
2731 | * Get and set a new queue for service. | |
2732 | */ | |
2733 | static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd) | |
2734 | { | |
2735 | struct bfq_queue *bfqq = bfq_get_next_queue(bfqd); | |
2736 | ||
2737 | __bfq_set_in_service_queue(bfqd, bfqq); | |
2738 | return bfqq; | |
2739 | } | |
2740 | ||
aee69d78 PV |
2741 | static void bfq_arm_slice_timer(struct bfq_data *bfqd) |
2742 | { | |
2743 | struct bfq_queue *bfqq = bfqd->in_service_queue; | |
aee69d78 PV |
2744 | u32 sl; |
2745 | ||
aee69d78 PV |
2746 | bfq_mark_bfqq_wait_request(bfqq); |
2747 | ||
2748 | /* | |
2749 | * We don't want to idle for seeks, but we do want to allow | |
2750 | * fair distribution of slice time for a process doing back-to-back | |
2751 | * seeks. So allow a little bit of time for him to submit a new rq. | |
2752 | */ | |
2753 | sl = bfqd->bfq_slice_idle; | |
2754 | /* | |
1de0c4cd AA |
2755 | * Unless the queue is being weight-raised or the scenario is |
2756 | * asymmetric, grant only minimum idle time if the queue | |
2757 | * is seeky. A long idling is preserved for a weight-raised | |
2758 | * queue, or, more in general, in an asymmetric scenario, | |
2759 | * because a long idling is needed for guaranteeing to a queue | |
2760 | * its reserved share of the throughput (in particular, it is | |
2761 | * needed if the queue has a higher weight than some other | |
2762 | * queue). | |
aee69d78 | 2763 | */ |
1de0c4cd | 2764 | if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 && |
fb53ac6c | 2765 | !bfq_asymmetric_scenario(bfqd, bfqq)) |
aee69d78 | 2766 | sl = min_t(u64, sl, BFQ_MIN_TT); |
778c02a2 PV |
2767 | else if (bfqq->wr_coeff > 1) |
2768 | sl = max_t(u32, sl, 20ULL * NSEC_PER_MSEC); | |
aee69d78 PV |
2769 | |
2770 | bfqd->last_idling_start = ktime_get(); | |
2341d662 PV |
2771 | bfqd->last_idling_start_jiffies = jiffies; |
2772 | ||
aee69d78 PV |
2773 | hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl), |
2774 | HRTIMER_MODE_REL); | |
e21b7a0b | 2775 | bfqg_stats_set_start_idle_time(bfqq_group(bfqq)); |
aee69d78 PV |
2776 | } |
2777 | ||
ab0e43e9 PV |
2778 | /* |
2779 | * In autotuning mode, max_budget is dynamically recomputed as the | |
2780 | * amount of sectors transferred in timeout at the estimated peak | |
2781 | * rate. This enables BFQ to utilize a full timeslice with a full | |
2782 | * budget, even if the in-service queue is served at peak rate. And | |
2783 | * this maximises throughput with sequential workloads. | |
2784 | */ | |
2785 | static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd) | |
2786 | { | |
2787 | return (u64)bfqd->peak_rate * USEC_PER_MSEC * | |
2788 | jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT; | |
2789 | } | |
2790 | ||
44e44a1b PV |
2791 | /* |
2792 | * Update parameters related to throughput and responsiveness, as a | |
2793 | * function of the estimated peak rate. See comments on | |
e24f1c24 | 2794 | * bfq_calc_max_budget(), and on the ref_wr_duration array. |
44e44a1b PV |
2795 | */ |
2796 | static void update_thr_responsiveness_params(struct bfq_data *bfqd) | |
2797 | { | |
e24f1c24 | 2798 | if (bfqd->bfq_user_max_budget == 0) { |
44e44a1b PV |
2799 | bfqd->bfq_max_budget = |
2800 | bfq_calc_max_budget(bfqd); | |
e24f1c24 | 2801 | bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget); |
44e44a1b | 2802 | } |
44e44a1b PV |
2803 | } |
2804 | ||
ab0e43e9 PV |
2805 | static void bfq_reset_rate_computation(struct bfq_data *bfqd, |
2806 | struct request *rq) | |
2807 | { | |
2808 | if (rq != NULL) { /* new rq dispatch now, reset accordingly */ | |
2809 | bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns(); | |
2810 | bfqd->peak_rate_samples = 1; | |
2811 | bfqd->sequential_samples = 0; | |
2812 | bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size = | |
2813 | blk_rq_sectors(rq); | |
2814 | } else /* no new rq dispatched, just reset the number of samples */ | |
2815 | bfqd->peak_rate_samples = 0; /* full re-init on next disp. */ | |
2816 | ||
2817 | bfq_log(bfqd, | |
2818 | "reset_rate_computation at end, sample %u/%u tot_sects %llu", | |
2819 | bfqd->peak_rate_samples, bfqd->sequential_samples, | |
2820 | bfqd->tot_sectors_dispatched); | |
2821 | } | |
2822 | ||
2823 | static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq) | |
2824 | { | |
2825 | u32 rate, weight, divisor; | |
2826 | ||
2827 | /* | |
2828 | * For the convergence property to hold (see comments on | |
2829 | * bfq_update_peak_rate()) and for the assessment to be | |
2830 | * reliable, a minimum number of samples must be present, and | |
2831 | * a minimum amount of time must have elapsed. If not so, do | |
2832 | * not compute new rate. Just reset parameters, to get ready | |
2833 | * for a new evaluation attempt. | |
2834 | */ | |
2835 | if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES || | |
2836 | bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL) | |
2837 | goto reset_computation; | |
2838 | ||
2839 | /* | |
2840 | * If a new request completion has occurred after last | |
2841 | * dispatch, then, to approximate the rate at which requests | |
2842 | * have been served by the device, it is more precise to | |
2843 | * extend the observation interval to the last completion. | |
2844 | */ | |
2845 | bfqd->delta_from_first = | |
2846 | max_t(u64, bfqd->delta_from_first, | |
2847 | bfqd->last_completion - bfqd->first_dispatch); | |
2848 | ||
2849 | /* | |
2850 | * Rate computed in sects/usec, and not sects/nsec, for | |
2851 | * precision issues. | |
2852 | */ | |
2853 | rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT, | |
2854 | div_u64(bfqd->delta_from_first, NSEC_PER_USEC)); | |
2855 | ||
2856 | /* | |
2857 | * Peak rate not updated if: | |
2858 | * - the percentage of sequential dispatches is below 3/4 of the | |
2859 | * total, and rate is below the current estimated peak rate | |
2860 | * - rate is unreasonably high (> 20M sectors/sec) | |
2861 | */ | |
2862 | if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 && | |
2863 | rate <= bfqd->peak_rate) || | |
2864 | rate > 20<<BFQ_RATE_SHIFT) | |
2865 | goto reset_computation; | |
2866 | ||
2867 | /* | |
2868 | * We have to update the peak rate, at last! To this purpose, | |
2869 | * we use a low-pass filter. We compute the smoothing constant | |
2870 | * of the filter as a function of the 'weight' of the new | |
2871 | * measured rate. | |
2872 | * | |
2873 | * As can be seen in next formulas, we define this weight as a | |
2874 | * quantity proportional to how sequential the workload is, | |
2875 | * and to how long the observation time interval is. | |
2876 | * | |
2877 | * The weight runs from 0 to 8. The maximum value of the | |
2878 | * weight, 8, yields the minimum value for the smoothing | |
2879 | * constant. At this minimum value for the smoothing constant, | |
2880 | * the measured rate contributes for half of the next value of | |
2881 | * the estimated peak rate. | |
2882 | * | |
2883 | * So, the first step is to compute the weight as a function | |
2884 | * of how sequential the workload is. Note that the weight | |
2885 | * cannot reach 9, because bfqd->sequential_samples cannot | |
2886 | * become equal to bfqd->peak_rate_samples, which, in its | |
2887 | * turn, holds true because bfqd->sequential_samples is not | |
2888 | * incremented for the first sample. | |
2889 | */ | |
2890 | weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples; | |
2891 | ||
2892 | /* | |
2893 | * Second step: further refine the weight as a function of the | |
2894 | * duration of the observation interval. | |
2895 | */ | |
2896 | weight = min_t(u32, 8, | |
2897 | div_u64(weight * bfqd->delta_from_first, | |
2898 | BFQ_RATE_REF_INTERVAL)); | |
2899 | ||
2900 | /* | |
2901 | * Divisor ranging from 10, for minimum weight, to 2, for | |
2902 | * maximum weight. | |
2903 | */ | |
2904 | divisor = 10 - weight; | |
2905 | ||
2906 | /* | |
2907 | * Finally, update peak rate: | |
2908 | * | |
2909 | * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor | |
2910 | */ | |
2911 | bfqd->peak_rate *= divisor-1; | |
2912 | bfqd->peak_rate /= divisor; | |
2913 | rate /= divisor; /* smoothing constant alpha = 1/divisor */ | |
2914 | ||
2915 | bfqd->peak_rate += rate; | |
bc56e2ca PV |
2916 | |
2917 | /* | |
2918 | * For a very slow device, bfqd->peak_rate can reach 0 (see | |
2919 | * the minimum representable values reported in the comments | |
2920 | * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid | |
2921 | * divisions by zero where bfqd->peak_rate is used as a | |
2922 | * divisor. | |
2923 | */ | |
2924 | bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate); | |
2925 | ||
44e44a1b | 2926 | update_thr_responsiveness_params(bfqd); |
ab0e43e9 PV |
2927 | |
2928 | reset_computation: | |
2929 | bfq_reset_rate_computation(bfqd, rq); | |
2930 | } | |
2931 | ||
2932 | /* | |
2933 | * Update the read/write peak rate (the main quantity used for | |
2934 | * auto-tuning, see update_thr_responsiveness_params()). | |
2935 | * | |
2936 | * It is not trivial to estimate the peak rate (correctly): because of | |
2937 | * the presence of sw and hw queues between the scheduler and the | |
2938 | * device components that finally serve I/O requests, it is hard to | |
2939 | * say exactly when a given dispatched request is served inside the | |
2940 | * device, and for how long. As a consequence, it is hard to know | |
2941 | * precisely at what rate a given set of requests is actually served | |
2942 | * by the device. | |
2943 | * | |
2944 | * On the opposite end, the dispatch time of any request is trivially | |
2945 | * available, and, from this piece of information, the "dispatch rate" | |
2946 | * of requests can be immediately computed. So, the idea in the next | |
2947 | * function is to use what is known, namely request dispatch times | |
2948 | * (plus, when useful, request completion times), to estimate what is | |
2949 | * unknown, namely in-device request service rate. | |
2950 | * | |
2951 | * The main issue is that, because of the above facts, the rate at | |
2952 | * which a certain set of requests is dispatched over a certain time | |
2953 | * interval can vary greatly with respect to the rate at which the | |
2954 | * same requests are then served. But, since the size of any | |
2955 | * intermediate queue is limited, and the service scheme is lossless | |
2956 | * (no request is silently dropped), the following obvious convergence | |
2957 | * property holds: the number of requests dispatched MUST become | |
2958 | * closer and closer to the number of requests completed as the | |
2959 | * observation interval grows. This is the key property used in | |
2960 | * the next function to estimate the peak service rate as a function | |
2961 | * of the observed dispatch rate. The function assumes to be invoked | |
2962 | * on every request dispatch. | |
2963 | */ | |
2964 | static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq) | |
2965 | { | |
2966 | u64 now_ns = ktime_get_ns(); | |
2967 | ||
2968 | if (bfqd->peak_rate_samples == 0) { /* first dispatch */ | |
2969 | bfq_log(bfqd, "update_peak_rate: goto reset, samples %d", | |
2970 | bfqd->peak_rate_samples); | |
2971 | bfq_reset_rate_computation(bfqd, rq); | |
2972 | goto update_last_values; /* will add one sample */ | |
2973 | } | |
2974 | ||
2975 | /* | |
2976 | * Device idle for very long: the observation interval lasting | |
2977 | * up to this dispatch cannot be a valid observation interval | |
2978 | * for computing a new peak rate (similarly to the late- | |
2979 | * completion event in bfq_completed_request()). Go to | |
2980 | * update_rate_and_reset to have the following three steps | |
2981 | * taken: | |
2982 | * - close the observation interval at the last (previous) | |
2983 | * request dispatch or completion | |
2984 | * - compute rate, if possible, for that observation interval | |
2985 | * - start a new observation interval with this dispatch | |
2986 | */ | |
2987 | if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC && | |
2988 | bfqd->rq_in_driver == 0) | |
2989 | goto update_rate_and_reset; | |
2990 | ||
2991 | /* Update sampling information */ | |
2992 | bfqd->peak_rate_samples++; | |
2993 | ||
2994 | if ((bfqd->rq_in_driver > 0 || | |
2995 | now_ns - bfqd->last_completion < BFQ_MIN_TT) | |
d87447d8 | 2996 | && !BFQ_RQ_SEEKY(bfqd, bfqd->last_position, rq)) |
ab0e43e9 PV |
2997 | bfqd->sequential_samples++; |
2998 | ||
2999 | bfqd->tot_sectors_dispatched += blk_rq_sectors(rq); | |
3000 | ||
3001 | /* Reset max observed rq size every 32 dispatches */ | |
3002 | if (likely(bfqd->peak_rate_samples % 32)) | |
3003 | bfqd->last_rq_max_size = | |
3004 | max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size); | |
3005 | else | |
3006 | bfqd->last_rq_max_size = blk_rq_sectors(rq); | |
3007 | ||
3008 | bfqd->delta_from_first = now_ns - bfqd->first_dispatch; | |
3009 | ||
3010 | /* Target observation interval not yet reached, go on sampling */ | |
3011 | if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL) | |
3012 | goto update_last_values; | |
3013 | ||
3014 | update_rate_and_reset: | |
3015 | bfq_update_rate_reset(bfqd, rq); | |
3016 | update_last_values: | |
3017 | bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); | |
058fdecc PV |
3018 | if (RQ_BFQQ(rq) == bfqd->in_service_queue) |
3019 | bfqd->in_serv_last_pos = bfqd->last_position; | |
ab0e43e9 PV |
3020 | bfqd->last_dispatch = now_ns; |
3021 | } | |
3022 | ||
aee69d78 PV |
3023 | /* |
3024 | * Remove request from internal lists. | |
3025 | */ | |
3026 | static void bfq_dispatch_remove(struct request_queue *q, struct request *rq) | |
3027 | { | |
3028 | struct bfq_queue *bfqq = RQ_BFQQ(rq); | |
3029 | ||
3030 | /* | |
3031 | * For consistency, the next instruction should have been | |
3032 | * executed after removing the request from the queue and | |
3033 | * dispatching it. We execute instead this instruction before | |
3034 | * bfq_remove_request() (and hence introduce a temporary | |
3035 | * inconsistency), for efficiency. In fact, should this | |
3036 | * dispatch occur for a non in-service bfqq, this anticipated | |
3037 | * increment prevents two counters related to bfqq->dispatched | |
3038 | * from risking to be, first, uselessly decremented, and then | |
3039 | * incremented again when the (new) value of bfqq->dispatched | |
3040 | * happens to be taken into account. | |
3041 | */ | |
3042 | bfqq->dispatched++; | |
ab0e43e9 | 3043 | bfq_update_peak_rate(q->elevator->elevator_data, rq); |
aee69d78 PV |
3044 | |
3045 | bfq_remove_request(q, rq); | |
3046 | } | |
3047 | ||
eed47d19 | 3048 | static bool __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
aee69d78 | 3049 | { |
36eca894 AA |
3050 | /* |
3051 | * If this bfqq is shared between multiple processes, check | |
3052 | * to make sure that those processes are still issuing I/Os | |
3053 | * within the mean seek distance. If not, it may be time to | |
3054 | * break the queues apart again. | |
3055 | */ | |
3056 | if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq)) | |
3057 | bfq_mark_bfqq_split_coop(bfqq); | |
3058 | ||
44e44a1b PV |
3059 | if (RB_EMPTY_ROOT(&bfqq->sort_list)) { |
3060 | if (bfqq->dispatched == 0) | |
3061 | /* | |
3062 | * Overloading budget_timeout field to store | |
3063 | * the time at which the queue remains with no | |
3064 | * backlog and no outstanding request; used by | |
3065 | * the weight-raising mechanism. | |
3066 | */ | |
3067 | bfqq->budget_timeout = jiffies; | |
3068 | ||
e21b7a0b | 3069 | bfq_del_bfqq_busy(bfqd, bfqq, true); |
36eca894 | 3070 | } else { |
80294c3b | 3071 | bfq_requeue_bfqq(bfqd, bfqq, true); |
36eca894 AA |
3072 | /* |
3073 | * Resort priority tree of potential close cooperators. | |
8cacc5ab | 3074 | * See comments on bfq_pos_tree_add_move() for the unlikely(). |
36eca894 | 3075 | */ |
8cacc5ab PV |
3076 | if (unlikely(!bfqd->nonrot_with_queueing)) |
3077 | bfq_pos_tree_add_move(bfqd, bfqq); | |
36eca894 | 3078 | } |
e21b7a0b AA |
3079 | |
3080 | /* | |
3081 | * All in-service entities must have been properly deactivated | |
3082 | * or requeued before executing the next function, which | |
eed47d19 PV |
3083 | * resets all in-service entities as no more in service. This |
3084 | * may cause bfqq to be freed. If this happens, the next | |
3085 | * function returns true. | |
e21b7a0b | 3086 | */ |
eed47d19 | 3087 | return __bfq_bfqd_reset_in_service(bfqd); |
aee69d78 PV |
3088 | } |
3089 | ||
3090 | /** | |
3091 | * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior. | |
3092 | * @bfqd: device data. | |
3093 | * @bfqq: queue to update. | |
3094 | * @reason: reason for expiration. | |
3095 | * | |
3096 | * Handle the feedback on @bfqq budget at queue expiration. | |
3097 | * See the body for detailed comments. | |
3098 | */ | |
3099 | static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd, | |
3100 | struct bfq_queue *bfqq, | |
3101 | enum bfqq_expiration reason) | |
3102 | { | |
3103 | struct request *next_rq; | |
3104 | int budget, min_budget; | |
3105 | ||
aee69d78 PV |
3106 | min_budget = bfq_min_budget(bfqd); |
3107 | ||
44e44a1b PV |
3108 | if (bfqq->wr_coeff == 1) |
3109 | budget = bfqq->max_budget; | |
3110 | else /* | |
3111 | * Use a constant, low budget for weight-raised queues, | |
3112 | * to help achieve a low latency. Keep it slightly higher | |
3113 | * than the minimum possible budget, to cause a little | |
3114 | * bit fewer expirations. | |
3115 | */ | |
3116 | budget = 2 * min_budget; | |
3117 | ||
aee69d78 PV |
3118 | bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d", |
3119 | bfqq->entity.budget, bfq_bfqq_budget_left(bfqq)); | |
3120 | bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d", | |
3121 | budget, bfq_min_budget(bfqd)); | |
3122 | bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d", | |
3123 | bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue)); | |
3124 | ||
44e44a1b | 3125 | if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) { |
aee69d78 PV |
3126 | switch (reason) { |
3127 | /* | |
3128 | * Caveat: in all the following cases we trade latency | |
3129 | * for throughput. | |
3130 | */ | |
3131 | case BFQQE_TOO_IDLE: | |
54b60456 PV |
3132 | /* |
3133 | * This is the only case where we may reduce | |
3134 | * the budget: if there is no request of the | |
3135 | * process still waiting for completion, then | |
3136 | * we assume (tentatively) that the timer has | |
3137 | * expired because the batch of requests of | |
3138 | * the process could have been served with a | |
3139 | * smaller budget. Hence, betting that | |
3140 | * process will behave in the same way when it | |
3141 | * becomes backlogged again, we reduce its | |
3142 | * next budget. As long as we guess right, | |
3143 | * this budget cut reduces the latency | |
3144 | * experienced by the process. | |
3145 | * | |
3146 | * However, if there are still outstanding | |
3147 | * requests, then the process may have not yet | |
3148 | * issued its next request just because it is | |
3149 | * still waiting for the completion of some of | |
3150 | * the still outstanding ones. So in this | |
3151 | * subcase we do not reduce its budget, on the | |
3152 | * contrary we increase it to possibly boost | |
3153 | * the throughput, as discussed in the | |
3154 | * comments to the BUDGET_TIMEOUT case. | |
3155 | */ | |
3156 | if (bfqq->dispatched > 0) /* still outstanding reqs */ | |
3157 | budget = min(budget * 2, bfqd->bfq_max_budget); | |
3158 | else { | |
3159 | if (budget > 5 * min_budget) | |
3160 | budget -= 4 * min_budget; | |
3161 | else | |
3162 | budget = min_budget; | |
3163 | } | |
aee69d78 PV |
3164 | break; |
3165 | case BFQQE_BUDGET_TIMEOUT: | |
54b60456 PV |
3166 | /* |
3167 | * We double the budget here because it gives | |
3168 | * the chance to boost the throughput if this | |
3169 | * is not a seeky process (and has bumped into | |
3170 | * this timeout because of, e.g., ZBR). | |
3171 | */ | |
3172 | budget = min(budget * 2, bfqd->bfq_max_budget); | |
aee69d78 PV |
3173 | break; |
3174 | case BFQQE_BUDGET_EXHAUSTED: | |
3175 | /* | |
3176 | * The process still has backlog, and did not | |
3177 | * let either the budget timeout or the disk | |
3178 | * idling timeout expire. Hence it is not | |
3179 | * seeky, has a short thinktime and may be | |
3180 | * happy with a higher budget too. So | |
3181 | * definitely increase the budget of this good | |
3182 | * candidate to boost the disk throughput. | |
3183 | */ | |
54b60456 | 3184 | budget = min(budget * 4, bfqd->bfq_max_budget); |
aee69d78 PV |
3185 | break; |
3186 | case BFQQE_NO_MORE_REQUESTS: | |
3187 | /* | |
3188 | * For queues that expire for this reason, it | |
3189 | * is particularly important to keep the | |
3190 | * budget close to the actual service they | |
3191 | * need. Doing so reduces the timestamp | |
3192 | * misalignment problem described in the | |
3193 | * comments in the body of | |
3194 | * __bfq_activate_entity. In fact, suppose | |
3195 | * that a queue systematically expires for | |
3196 | * BFQQE_NO_MORE_REQUESTS and presents a | |
3197 | * new request in time to enjoy timestamp | |
3198 | * back-shifting. The larger the budget of the | |
3199 | * queue is with respect to the service the | |
3200 | * queue actually requests in each service | |
3201 | * slot, the more times the queue can be | |
3202 | * reactivated with the same virtual finish | |
3203 | * time. It follows that, even if this finish | |
3204 | * time is pushed to the system virtual time | |
3205 | * to reduce the consequent timestamp | |
3206 | * misalignment, the queue unjustly enjoys for | |
3207 | * many re-activations a lower finish time | |
3208 | * than all newly activated queues. | |
3209 | * | |
3210 | * The service needed by bfqq is measured | |
3211 | * quite precisely by bfqq->entity.service. | |
3212 | * Since bfqq does not enjoy device idling, | |
3213 | * bfqq->entity.service is equal to the number | |
3214 | * of sectors that the process associated with | |
3215 | * bfqq requested to read/write before waiting | |
3216 | * for request completions, or blocking for | |
3217 | * other reasons. | |
3218 | */ | |
3219 | budget = max_t(int, bfqq->entity.service, min_budget); | |
3220 | break; | |
3221 | default: | |
3222 | return; | |
3223 | } | |
44e44a1b | 3224 | } else if (!bfq_bfqq_sync(bfqq)) { |
aee69d78 PV |
3225 | /* |
3226 | * Async queues get always the maximum possible | |
3227 | * budget, as for them we do not care about latency | |
3228 | * (in addition, their ability to dispatch is limited | |
3229 | * by the charging factor). | |
3230 | */ | |
3231 | budget = bfqd->bfq_max_budget; | |
3232 | } | |
3233 | ||
3234 | bfqq->max_budget = budget; | |
3235 | ||
3236 | if (bfqd->budgets_assigned >= bfq_stats_min_budgets && | |
3237 | !bfqd->bfq_user_max_budget) | |
3238 | bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget); | |
3239 | ||
3240 | /* | |
3241 | * If there is still backlog, then assign a new budget, making | |
3242 | * sure that it is large enough for the next request. Since | |
3243 | * the finish time of bfqq must be kept in sync with the | |
3244 | * budget, be sure to call __bfq_bfqq_expire() *after* this | |
3245 | * update. | |
3246 | * | |
3247 | * If there is no backlog, then no need to update the budget; | |
3248 | * it will be updated on the arrival of a new request. | |
3249 | */ | |
3250 | next_rq = bfqq->next_rq; | |
3251 | if (next_rq) | |
3252 | bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget, | |
3253 | bfq_serv_to_charge(next_rq, bfqq)); | |
3254 | ||
3255 | bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d", | |
3256 | next_rq ? blk_rq_sectors(next_rq) : 0, | |
3257 | bfqq->entity.budget); | |
3258 | } | |
3259 | ||
aee69d78 | 3260 | /* |
ab0e43e9 PV |
3261 | * Return true if the process associated with bfqq is "slow". The slow |
3262 | * flag is used, in addition to the budget timeout, to reduce the | |
3263 | * amount of service provided to seeky processes, and thus reduce | |
3264 | * their chances to lower the throughput. More details in the comments | |
3265 | * on the function bfq_bfqq_expire(). | |
3266 | * | |
3267 | * An important observation is in order: as discussed in the comments | |
3268 | * on the function bfq_update_peak_rate(), with devices with internal | |
3269 | * queues, it is hard if ever possible to know when and for how long | |
3270 | * an I/O request is processed by the device (apart from the trivial | |
3271 | * I/O pattern where a new request is dispatched only after the | |
3272 | * previous one has been completed). This makes it hard to evaluate | |
3273 | * the real rate at which the I/O requests of each bfq_queue are | |
3274 | * served. In fact, for an I/O scheduler like BFQ, serving a | |
3275 | * bfq_queue means just dispatching its requests during its service | |
3276 | * slot (i.e., until the budget of the queue is exhausted, or the | |
3277 | * queue remains idle, or, finally, a timeout fires). But, during the | |
3278 | * service slot of a bfq_queue, around 100 ms at most, the device may | |
3279 | * be even still processing requests of bfq_queues served in previous | |
3280 | * service slots. On the opposite end, the requests of the in-service | |
3281 | * bfq_queue may be completed after the service slot of the queue | |
3282 | * finishes. | |
3283 | * | |
3284 | * Anyway, unless more sophisticated solutions are used | |
3285 | * (where possible), the sum of the sizes of the requests dispatched | |
3286 | * during the service slot of a bfq_queue is probably the only | |
3287 | * approximation available for the service received by the bfq_queue | |
3288 | * during its service slot. And this sum is the quantity used in this | |
3289 | * function to evaluate the I/O speed of a process. | |
aee69d78 | 3290 | */ |
ab0e43e9 PV |
3291 | static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
3292 | bool compensate, enum bfqq_expiration reason, | |
3293 | unsigned long *delta_ms) | |
aee69d78 | 3294 | { |
ab0e43e9 PV |
3295 | ktime_t delta_ktime; |
3296 | u32 delta_usecs; | |
3297 | bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */ | |
aee69d78 | 3298 | |
ab0e43e9 | 3299 | if (!bfq_bfqq_sync(bfqq)) |
aee69d78 PV |
3300 | return false; |
3301 | ||
3302 | if (compensate) | |
ab0e43e9 | 3303 | delta_ktime = bfqd->last_idling_start; |
aee69d78 | 3304 | else |
ab0e43e9 PV |
3305 | delta_ktime = ktime_get(); |
3306 | delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start); | |
3307 | delta_usecs = ktime_to_us(delta_ktime); | |
aee69d78 PV |
3308 | |
3309 | /* don't use too short time intervals */ | |
ab0e43e9 PV |
3310 | if (delta_usecs < 1000) { |
3311 | if (blk_queue_nonrot(bfqd->queue)) | |
3312 | /* | |
3313 | * give same worst-case guarantees as idling | |
3314 | * for seeky | |
3315 | */ | |
3316 | *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC; | |
3317 | else /* charge at least one seek */ | |
3318 | *delta_ms = bfq_slice_idle / NSEC_PER_MSEC; | |
3319 | ||
3320 | return slow; | |
3321 | } | |
aee69d78 | 3322 | |
ab0e43e9 | 3323 | *delta_ms = delta_usecs / USEC_PER_MSEC; |
aee69d78 PV |
3324 | |
3325 | /* | |
ab0e43e9 PV |
3326 | * Use only long (> 20ms) intervals to filter out excessive |
3327 | * spikes in service rate estimation. | |
aee69d78 | 3328 | */ |
ab0e43e9 PV |
3329 | if (delta_usecs > 20000) { |
3330 | /* | |
3331 | * Caveat for rotational devices: processes doing I/O | |
3332 | * in the slower disk zones tend to be slow(er) even | |
3333 | * if not seeky. In this respect, the estimated peak | |
3334 | * rate is likely to be an average over the disk | |
3335 | * surface. Accordingly, to not be too harsh with | |
3336 | * unlucky processes, a process is deemed slow only if | |
3337 | * its rate has been lower than half of the estimated | |
3338 | * peak rate. | |
3339 | */ | |
3340 | slow = bfqq->entity.service < bfqd->bfq_max_budget / 2; | |
aee69d78 PV |
3341 | } |
3342 | ||
ab0e43e9 | 3343 | bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow); |
aee69d78 | 3344 | |
ab0e43e9 | 3345 | return slow; |
aee69d78 PV |
3346 | } |
3347 | ||
77b7dcea PV |
3348 | /* |
3349 | * To be deemed as soft real-time, an application must meet two | |
3350 | * requirements. First, the application must not require an average | |
3351 | * bandwidth higher than the approximate bandwidth required to playback or | |
3352 | * record a compressed high-definition video. | |
3353 | * The next function is invoked on the completion of the last request of a | |
3354 | * batch, to compute the next-start time instant, soft_rt_next_start, such | |
3355 | * that, if the next request of the application does not arrive before | |
3356 | * soft_rt_next_start, then the above requirement on the bandwidth is met. | |
3357 | * | |
3358 | * The second requirement is that the request pattern of the application is | |
3359 | * isochronous, i.e., that, after issuing a request or a batch of requests, | |
3360 | * the application stops issuing new requests until all its pending requests | |
3361 | * have been completed. After that, the application may issue a new batch, | |
3362 | * and so on. | |
3363 | * For this reason the next function is invoked to compute | |
3364 | * soft_rt_next_start only for applications that meet this requirement, | |
3365 | * whereas soft_rt_next_start is set to infinity for applications that do | |
3366 | * not. | |
3367 | * | |
a34b0244 PV |
3368 | * Unfortunately, even a greedy (i.e., I/O-bound) application may |
3369 | * happen to meet, occasionally or systematically, both the above | |
3370 | * bandwidth and isochrony requirements. This may happen at least in | |
3371 | * the following circumstances. First, if the CPU load is high. The | |
3372 | * application may stop issuing requests while the CPUs are busy | |
3373 | * serving other processes, then restart, then stop again for a while, | |
3374 | * and so on. The other circumstances are related to the storage | |
3375 | * device: the storage device is highly loaded or reaches a low-enough | |
3376 | * throughput with the I/O of the application (e.g., because the I/O | |
3377 | * is random and/or the device is slow). In all these cases, the | |
3378 | * I/O of the application may be simply slowed down enough to meet | |
3379 | * the bandwidth and isochrony requirements. To reduce the probability | |
3380 | * that greedy applications are deemed as soft real-time in these | |
3381 | * corner cases, a further rule is used in the computation of | |
3382 | * soft_rt_next_start: the return value of this function is forced to | |
3383 | * be higher than the maximum between the following two quantities. | |
3384 | * | |
3385 | * (a) Current time plus: (1) the maximum time for which the arrival | |
3386 | * of a request is waited for when a sync queue becomes idle, | |
3387 | * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We | |
3388 | * postpone for a moment the reason for adding a few extra | |
3389 | * jiffies; we get back to it after next item (b). Lower-bounding | |
3390 | * the return value of this function with the current time plus | |
3391 | * bfqd->bfq_slice_idle tends to filter out greedy applications, | |
3392 | * because the latter issue their next request as soon as possible | |
3393 | * after the last one has been completed. In contrast, a soft | |
3394 | * real-time application spends some time processing data, after a | |
3395 | * batch of its requests has been completed. | |
3396 | * | |
3397 | * (b) Current value of bfqq->soft_rt_next_start. As pointed out | |
3398 | * above, greedy applications may happen to meet both the | |
3399 | * bandwidth and isochrony requirements under heavy CPU or | |
3400 | * storage-device load. In more detail, in these scenarios, these | |
3401 | * applications happen, only for limited time periods, to do I/O | |
3402 | * slowly enough to meet all the requirements described so far, | |
3403 | * including the filtering in above item (a). These slow-speed | |
3404 | * time intervals are usually interspersed between other time | |
3405 | * intervals during which these applications do I/O at a very high | |
3406 | * speed. Fortunately, exactly because of the high speed of the | |
3407 | * I/O in the high-speed intervals, the values returned by this | |
3408 | * function happen to be so high, near the end of any such | |
3409 | * high-speed interval, to be likely to fall *after* the end of | |
3410 | * the low-speed time interval that follows. These high values are | |
3411 | * stored in bfqq->soft_rt_next_start after each invocation of | |
3412 | * this function. As a consequence, if the last value of | |
3413 | * bfqq->soft_rt_next_start is constantly used to lower-bound the | |
3414 | * next value that this function may return, then, from the very | |
3415 | * beginning of a low-speed interval, bfqq->soft_rt_next_start is | |
3416 | * likely to be constantly kept so high that any I/O request | |
3417 | * issued during the low-speed interval is considered as arriving | |
3418 | * to soon for the application to be deemed as soft | |
3419 | * real-time. Then, in the high-speed interval that follows, the | |
3420 | * application will not be deemed as soft real-time, just because | |
3421 | * it will do I/O at a high speed. And so on. | |
3422 | * | |
3423 | * Getting back to the filtering in item (a), in the following two | |
3424 | * cases this filtering might be easily passed by a greedy | |
3425 | * application, if the reference quantity was just | |
3426 | * bfqd->bfq_slice_idle: | |
3427 | * 1) HZ is so low that the duration of a jiffy is comparable to or | |
3428 | * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow | |
3429 | * devices with HZ=100. The time granularity may be so coarse | |
3430 | * that the approximation, in jiffies, of bfqd->bfq_slice_idle | |
3431 | * is rather lower than the exact value. | |
77b7dcea PV |
3432 | * 2) jiffies, instead of increasing at a constant rate, may stop increasing |
3433 | * for a while, then suddenly 'jump' by several units to recover the lost | |
3434 | * increments. This seems to happen, e.g., inside virtual machines. | |
a34b0244 PV |
3435 | * To address this issue, in the filtering in (a) we do not use as a |
3436 | * reference time interval just bfqd->bfq_slice_idle, but | |
3437 | * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the | |
3438 | * minimum number of jiffies for which the filter seems to be quite | |
3439 | * precise also in embedded systems and KVM/QEMU virtual machines. | |
77b7dcea PV |
3440 | */ |
3441 | static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd, | |
3442 | struct bfq_queue *bfqq) | |
3443 | { | |
a34b0244 PV |
3444 | return max3(bfqq->soft_rt_next_start, |
3445 | bfqq->last_idle_bklogged + | |
3446 | HZ * bfqq->service_from_backlogged / | |
3447 | bfqd->bfq_wr_max_softrt_rate, | |
3448 | jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4); | |
77b7dcea PV |
3449 | } |
3450 | ||
aee69d78 PV |
3451 | /** |
3452 | * bfq_bfqq_expire - expire a queue. | |
3453 | * @bfqd: device owning the queue. | |
3454 | * @bfqq: the queue to expire. | |
3455 | * @compensate: if true, compensate for the time spent idling. | |
3456 | * @reason: the reason causing the expiration. | |
3457 | * | |
c074170e PV |
3458 | * If the process associated with bfqq does slow I/O (e.g., because it |
3459 | * issues random requests), we charge bfqq with the time it has been | |
3460 | * in service instead of the service it has received (see | |
3461 | * bfq_bfqq_charge_time for details on how this goal is achieved). As | |
3462 | * a consequence, bfqq will typically get higher timestamps upon | |
3463 | * reactivation, and hence it will be rescheduled as if it had | |
3464 | * received more service than what it has actually received. In the | |
3465 | * end, bfqq receives less service in proportion to how slowly its | |
3466 | * associated process consumes its budgets (and hence how seriously it | |
3467 | * tends to lower the throughput). In addition, this time-charging | |
3468 | * strategy guarantees time fairness among slow processes. In | |
3469 | * contrast, if the process associated with bfqq is not slow, we | |
3470 | * charge bfqq exactly with the service it has received. | |
aee69d78 | 3471 | * |
c074170e PV |
3472 | * Charging time to the first type of queues and the exact service to |
3473 | * the other has the effect of using the WF2Q+ policy to schedule the | |
3474 | * former on a timeslice basis, without violating service domain | |
3475 | * guarantees among the latter. | |
aee69d78 | 3476 | */ |
ea25da48 PV |
3477 | void bfq_bfqq_expire(struct bfq_data *bfqd, |
3478 | struct bfq_queue *bfqq, | |
3479 | bool compensate, | |
3480 | enum bfqq_expiration reason) | |
aee69d78 PV |
3481 | { |
3482 | bool slow; | |
ab0e43e9 PV |
3483 | unsigned long delta = 0; |
3484 | struct bfq_entity *entity = &bfqq->entity; | |
aee69d78 PV |
3485 | |
3486 | /* | |
ab0e43e9 | 3487 | * Check whether the process is slow (see bfq_bfqq_is_slow). |
aee69d78 | 3488 | */ |
ab0e43e9 | 3489 | slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta); |
aee69d78 PV |
3490 | |
3491 | /* | |
c074170e PV |
3492 | * As above explained, charge slow (typically seeky) and |
3493 | * timed-out queues with the time and not the service | |
3494 | * received, to favor sequential workloads. | |
3495 | * | |
3496 | * Processes doing I/O in the slower disk zones will tend to | |
3497 | * be slow(er) even if not seeky. Therefore, since the | |
3498 | * estimated peak rate is actually an average over the disk | |
3499 | * surface, these processes may timeout just for bad luck. To | |
3500 | * avoid punishing them, do not charge time to processes that | |
3501 | * succeeded in consuming at least 2/3 of their budget. This | |
3502 | * allows BFQ to preserve enough elasticity to still perform | |
3503 | * bandwidth, and not time, distribution with little unlucky | |
3504 | * or quasi-sequential processes. | |
aee69d78 | 3505 | */ |
44e44a1b PV |
3506 | if (bfqq->wr_coeff == 1 && |
3507 | (slow || | |
3508 | (reason == BFQQE_BUDGET_TIMEOUT && | |
3509 | bfq_bfqq_budget_left(bfqq) >= entity->budget / 3))) | |
c074170e | 3510 | bfq_bfqq_charge_time(bfqd, bfqq, delta); |
aee69d78 PV |
3511 | |
3512 | if (reason == BFQQE_TOO_IDLE && | |
ab0e43e9 | 3513 | entity->service <= 2 * entity->budget / 10) |
aee69d78 PV |
3514 | bfq_clear_bfqq_IO_bound(bfqq); |
3515 | ||
44e44a1b PV |
3516 | if (bfqd->low_latency && bfqq->wr_coeff == 1) |
3517 | bfqq->last_wr_start_finish = jiffies; | |
3518 | ||
77b7dcea PV |
3519 | if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 && |
3520 | RB_EMPTY_ROOT(&bfqq->sort_list)) { | |
3521 | /* | |
3522 | * If we get here, and there are no outstanding | |
3523 | * requests, then the request pattern is isochronous | |
3524 | * (see the comments on the function | |
3525 | * bfq_bfqq_softrt_next_start()). Thus we can compute | |
20cd3245 PV |
3526 | * soft_rt_next_start. And we do it, unless bfqq is in |
3527 | * interactive weight raising. We do not do it in the | |
3528 | * latter subcase, for the following reason. bfqq may | |
3529 | * be conveying the I/O needed to load a soft | |
3530 | * real-time application. Such an application will | |
3531 | * actually exhibit a soft real-time I/O pattern after | |
3532 | * it finally starts doing its job. But, if | |
3533 | * soft_rt_next_start is computed here for an | |
3534 | * interactive bfqq, and bfqq had received a lot of | |
3535 | * service before remaining with no outstanding | |
3536 | * request (likely to happen on a fast device), then | |
3537 | * soft_rt_next_start would be assigned such a high | |
3538 | * value that, for a very long time, bfqq would be | |
3539 | * prevented from being possibly considered as soft | |
3540 | * real time. | |
3541 | * | |
3542 | * If, instead, the queue still has outstanding | |
3543 | * requests, then we have to wait for the completion | |
3544 | * of all the outstanding requests to discover whether | |
3545 | * the request pattern is actually isochronous. | |
77b7dcea | 3546 | */ |
20cd3245 PV |
3547 | if (bfqq->dispatched == 0 && |
3548 | bfqq->wr_coeff != bfqd->bfq_wr_coeff) | |
77b7dcea PV |
3549 | bfqq->soft_rt_next_start = |
3550 | bfq_bfqq_softrt_next_start(bfqd, bfqq); | |
20cd3245 | 3551 | else if (bfqq->dispatched > 0) { |
77b7dcea PV |
3552 | /* |
3553 | * Schedule an update of soft_rt_next_start to when | |
3554 | * the task may be discovered to be isochronous. | |
3555 | */ | |
3556 | bfq_mark_bfqq_softrt_update(bfqq); | |
3557 | } | |
3558 | } | |
3559 | ||
aee69d78 | 3560 | bfq_log_bfqq(bfqd, bfqq, |
d5be3fef PV |
3561 | "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason, |
3562 | slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq)); | |
aee69d78 | 3563 | |
2341d662 PV |
3564 | /* |
3565 | * bfqq expired, so no total service time needs to be computed | |
3566 | * any longer: reset state machine for measuring total service | |
3567 | * times. | |
3568 | */ | |
3569 | bfqd->rqs_injected = bfqd->wait_dispatch = false; | |
3570 | bfqd->waited_rq = NULL; | |
3571 | ||
aee69d78 PV |
3572 | /* |
3573 | * Increase, decrease or leave budget unchanged according to | |
3574 | * reason. | |
3575 | */ | |
3576 | __bfq_bfqq_recalc_budget(bfqd, bfqq, reason); | |
eed47d19 PV |
3577 | if (__bfq_bfqq_expire(bfqd, bfqq)) |
3578 | /* bfqq is gone, no more actions on it */ | |
9fae8dd5 PV |
3579 | return; |
3580 | ||
aee69d78 | 3581 | /* mark bfqq as waiting a request only if a bic still points to it */ |
9fae8dd5 | 3582 | if (!bfq_bfqq_busy(bfqq) && |
aee69d78 | 3583 | reason != BFQQE_BUDGET_TIMEOUT && |
9fae8dd5 | 3584 | reason != BFQQE_BUDGET_EXHAUSTED) { |
aee69d78 | 3585 | bfq_mark_bfqq_non_blocking_wait_rq(bfqq); |
9fae8dd5 PV |
3586 | /* |
3587 | * Not setting service to 0, because, if the next rq | |
3588 | * arrives in time, the queue will go on receiving | |
3589 | * service with this same budget (as if it never expired) | |
3590 | */ | |
3591 | } else | |
3592 | entity->service = 0; | |
8a511ba5 PV |
3593 | |
3594 | /* | |
3595 | * Reset the received-service counter for every parent entity. | |
3596 | * Differently from what happens with bfqq->entity.service, | |
3597 | * the resetting of this counter never needs to be postponed | |
3598 | * for parent entities. In fact, in case bfqq may have a | |
3599 | * chance to go on being served using the last, partially | |
3600 | * consumed budget, bfqq->entity.service needs to be kept, | |
3601 | * because if bfqq then actually goes on being served using | |
3602 | * the same budget, the last value of bfqq->entity.service is | |
3603 | * needed to properly decrement bfqq->entity.budget by the | |
3604 | * portion already consumed. In contrast, it is not necessary | |
3605 | * to keep entity->service for parent entities too, because | |
3606 | * the bubble up of the new value of bfqq->entity.budget will | |
3607 | * make sure that the budgets of parent entities are correct, | |
3608 | * even in case bfqq and thus parent entities go on receiving | |
3609 | * service with the same budget. | |
3610 | */ | |
3611 | entity = entity->parent; | |
3612 | for_each_entity(entity) | |
3613 | entity->service = 0; | |
aee69d78 PV |
3614 | } |
3615 | ||
3616 | /* | |
3617 | * Budget timeout is not implemented through a dedicated timer, but | |
3618 | * just checked on request arrivals and completions, as well as on | |
3619 | * idle timer expirations. | |
3620 | */ | |
3621 | static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq) | |
3622 | { | |
44e44a1b | 3623 | return time_is_before_eq_jiffies(bfqq->budget_timeout); |
aee69d78 PV |
3624 | } |
3625 | ||
3626 | /* | |
3627 | * If we expire a queue that is actively waiting (i.e., with the | |
3628 | * device idled) for the arrival of a new request, then we may incur | |
3629 | * the timestamp misalignment problem described in the body of the | |
3630 | * function __bfq_activate_entity. Hence we return true only if this | |
3631 | * condition does not hold, or if the queue is slow enough to deserve | |
3632 | * only to be kicked off for preserving a high throughput. | |
3633 | */ | |
3634 | static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq) | |
3635 | { | |
3636 | bfq_log_bfqq(bfqq->bfqd, bfqq, | |
3637 | "may_budget_timeout: wait_request %d left %d timeout %d", | |
3638 | bfq_bfqq_wait_request(bfqq), | |
3639 | bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3, | |
3640 | bfq_bfqq_budget_timeout(bfqq)); | |
3641 | ||
3642 | return (!bfq_bfqq_wait_request(bfqq) || | |
3643 | bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3) | |
3644 | && | |
3645 | bfq_bfqq_budget_timeout(bfqq); | |
3646 | } | |
3647 | ||
05c2f5c3 PV |
3648 | static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd, |
3649 | struct bfq_queue *bfqq) | |
aee69d78 | 3650 | { |
edaf9428 PV |
3651 | bool rot_without_queueing = |
3652 | !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag, | |
3653 | bfqq_sequential_and_IO_bound, | |
05c2f5c3 | 3654 | idling_boosts_thr; |
d5be3fef | 3655 | |
edaf9428 PV |
3656 | bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) && |
3657 | bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq); | |
3658 | ||
aee69d78 | 3659 | /* |
44e44a1b PV |
3660 | * The next variable takes into account the cases where idling |
3661 | * boosts the throughput. | |
3662 | * | |
e01eff01 PV |
3663 | * The value of the variable is computed considering, first, that |
3664 | * idling is virtually always beneficial for the throughput if: | |
edaf9428 PV |
3665 | * (a) the device is not NCQ-capable and rotational, or |
3666 | * (b) regardless of the presence of NCQ, the device is rotational and | |
3667 | * the request pattern for bfqq is I/O-bound and sequential, or | |
3668 | * (c) regardless of whether it is rotational, the device is | |
3669 | * not NCQ-capable and the request pattern for bfqq is | |
3670 | * I/O-bound and sequential. | |
bf2b79e7 PV |
3671 | * |
3672 | * Secondly, and in contrast to the above item (b), idling an | |
3673 | * NCQ-capable flash-based device would not boost the | |
e01eff01 | 3674 | * throughput even with sequential I/O; rather it would lower |
bf2b79e7 PV |
3675 | * the throughput in proportion to how fast the device |
3676 | * is. Accordingly, the next variable is true if any of the | |
edaf9428 PV |
3677 | * above conditions (a), (b) or (c) is true, and, in |
3678 | * particular, happens to be false if bfqd is an NCQ-capable | |
3679 | * flash-based device. | |
aee69d78 | 3680 | */ |
edaf9428 PV |
3681 | idling_boosts_thr = rot_without_queueing || |
3682 | ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) && | |
3683 | bfqq_sequential_and_IO_bound); | |
aee69d78 | 3684 | |
cfd69712 | 3685 | /* |
05c2f5c3 | 3686 | * The return value of this function is equal to that of |
cfd69712 PV |
3687 | * idling_boosts_thr, unless a special case holds. In this |
3688 | * special case, described below, idling may cause problems to | |
3689 | * weight-raised queues. | |
3690 | * | |
3691 | * When the request pool is saturated (e.g., in the presence | |
3692 | * of write hogs), if the processes associated with | |
3693 | * non-weight-raised queues ask for requests at a lower rate, | |
3694 | * then processes associated with weight-raised queues have a | |
3695 | * higher probability to get a request from the pool | |
3696 | * immediately (or at least soon) when they need one. Thus | |
3697 | * they have a higher probability to actually get a fraction | |
3698 | * of the device throughput proportional to their high | |
3699 | * weight. This is especially true with NCQ-capable drives, | |
3700 | * which enqueue several requests in advance, and further | |
3701 | * reorder internally-queued requests. | |
3702 | * | |
05c2f5c3 PV |
3703 | * For this reason, we force to false the return value if |
3704 | * there are weight-raised busy queues. In this case, and if | |
3705 | * bfqq is not weight-raised, this guarantees that the device | |
3706 | * is not idled for bfqq (if, instead, bfqq is weight-raised, | |
3707 | * then idling will be guaranteed by another variable, see | |
3708 | * below). Combined with the timestamping rules of BFQ (see | |
3709 | * [1] for details), this behavior causes bfqq, and hence any | |
3710 | * sync non-weight-raised queue, to get a lower number of | |
3711 | * requests served, and thus to ask for a lower number of | |
3712 | * requests from the request pool, before the busy | |
3713 | * weight-raised queues get served again. This often mitigates | |
3714 | * starvation problems in the presence of heavy write | |
3715 | * workloads and NCQ, thereby guaranteeing a higher | |
3716 | * application and system responsiveness in these hostile | |
3717 | * scenarios. | |
3718 | */ | |
3719 | return idling_boosts_thr && | |
cfd69712 | 3720 | bfqd->wr_busy_queues == 0; |
05c2f5c3 | 3721 | } |
cfd69712 | 3722 | |
530c4cbb | 3723 | /* |
fb53ac6c PV |
3724 | * There is a case where idling does not have to be performed for |
3725 | * throughput concerns, but to preserve the throughput share of | |
3726 | * the process associated with bfqq. | |
530c4cbb PV |
3727 | * |
3728 | * To introduce this case, we can note that allowing the drive | |
3729 | * to enqueue more than one request at a time, and hence | |
3730 | * delegating de facto final scheduling decisions to the | |
3731 | * drive's internal scheduler, entails loss of control on the | |
3732 | * actual request service order. In particular, the critical | |
3733 | * situation is when requests from different processes happen | |
3734 | * to be present, at the same time, in the internal queue(s) | |
3735 | * of the drive. In such a situation, the drive, by deciding | |
3736 | * the service order of the internally-queued requests, does | |
3737 | * determine also the actual throughput distribution among | |
3738 | * these processes. But the drive typically has no notion or | |
3739 | * concern about per-process throughput distribution, and | |
3740 | * makes its decisions only on a per-request basis. Therefore, | |
3741 | * the service distribution enforced by the drive's internal | |
fb53ac6c PV |
3742 | * scheduler is likely to coincide with the desired throughput |
3743 | * distribution only in a completely symmetric, or favorably | |
3744 | * skewed scenario where: | |
3745 | * (i-a) each of these processes must get the same throughput as | |
3746 | * the others, | |
3747 | * (i-b) in case (i-a) does not hold, it holds that the process | |
3748 | * associated with bfqq must receive a lower or equal | |
3749 | * throughput than any of the other processes; | |
3750 | * (ii) the I/O of each process has the same properties, in | |
3751 | * terms of locality (sequential or random), direction | |
3752 | * (reads or writes), request sizes, greediness | |
3753 | * (from I/O-bound to sporadic), and so on; | |
3754 | ||
3755 | * In fact, in such a scenario, the drive tends to treat the requests | |
3756 | * of each process in about the same way as the requests of the | |
3757 | * others, and thus to provide each of these processes with about the | |
3758 | * same throughput. This is exactly the desired throughput | |
3759 | * distribution if (i-a) holds, or, if (i-b) holds instead, this is an | |
3760 | * even more convenient distribution for (the process associated with) | |
3761 | * bfqq. | |
3762 | * | |
3763 | * In contrast, in any asymmetric or unfavorable scenario, device | |
3764 | * idling (I/O-dispatch plugging) is certainly needed to guarantee | |
3765 | * that bfqq receives its assigned fraction of the device throughput | |
3766 | * (see [1] for details). | |
3767 | * | |
3768 | * The problem is that idling may significantly reduce throughput with | |
3769 | * certain combinations of types of I/O and devices. An important | |
3770 | * example is sync random I/O on flash storage with command | |
3771 | * queueing. So, unless bfqq falls in cases where idling also boosts | |
3772 | * throughput, it is important to check conditions (i-a), i(-b) and | |
3773 | * (ii) accurately, so as to avoid idling when not strictly needed for | |
3774 | * service guarantees. | |
530c4cbb | 3775 | * |
fb53ac6c PV |
3776 | * Unfortunately, it is extremely difficult to thoroughly check |
3777 | * condition (ii). And, in case there are active groups, it becomes | |
3778 | * very difficult to check conditions (i-a) and (i-b) too. In fact, | |
3779 | * if there are active groups, then, for conditions (i-a) or (i-b) to | |
3780 | * become false 'indirectly', it is enough that an active group | |
3781 | * contains more active processes or sub-groups than some other active | |
3782 | * group. More precisely, for conditions (i-a) or (i-b) to become | |
3783 | * false because of such a group, it is not even necessary that the | |
3784 | * group is (still) active: it is sufficient that, even if the group | |
3785 | * has become inactive, some of its descendant processes still have | |
3786 | * some request already dispatched but still waiting for | |
3787 | * completion. In fact, requests have still to be guaranteed their | |
3788 | * share of the throughput even after being dispatched. In this | |
3789 | * respect, it is easy to show that, if a group frequently becomes | |
3790 | * inactive while still having in-flight requests, and if, when this | |
3791 | * happens, the group is not considered in the calculation of whether | |
3792 | * the scenario is asymmetric, then the group may fail to be | |
3793 | * guaranteed its fair share of the throughput (basically because | |
3794 | * idling may not be performed for the descendant processes of the | |
3795 | * group, but it had to be). We address this issue with the following | |
3796 | * bi-modal behavior, implemented in the function | |
3797 | * bfq_asymmetric_scenario(). | |
530c4cbb PV |
3798 | * |
3799 | * If there are groups with requests waiting for completion | |
3800 | * (as commented above, some of these groups may even be | |
3801 | * already inactive), then the scenario is tagged as | |
3802 | * asymmetric, conservatively, without checking any of the | |
fb53ac6c | 3803 | * conditions (i-a), (i-b) or (ii). So the device is idled for bfqq. |
530c4cbb PV |
3804 | * This behavior matches also the fact that groups are created |
3805 | * exactly if controlling I/O is a primary concern (to | |
3806 | * preserve bandwidth and latency guarantees). | |
3807 | * | |
fb53ac6c PV |
3808 | * On the opposite end, if there are no groups with requests waiting |
3809 | * for completion, then only conditions (i-a) and (i-b) are actually | |
3810 | * controlled, i.e., provided that conditions (i-a) or (i-b) holds, | |
3811 | * idling is not performed, regardless of whether condition (ii) | |
3812 | * holds. In other words, only if conditions (i-a) and (i-b) do not | |
3813 | * hold, then idling is allowed, and the device tends to be prevented | |
3814 | * from queueing many requests, possibly of several processes. Since | |
3815 | * there are no groups with requests waiting for completion, then, to | |
3816 | * control conditions (i-a) and (i-b) it is enough to check just | |
3817 | * whether all the queues with requests waiting for completion also | |
3818 | * have the same weight. | |
530c4cbb PV |
3819 | * |
3820 | * Not checking condition (ii) evidently exposes bfqq to the | |
3821 | * risk of getting less throughput than its fair share. | |
3822 | * However, for queues with the same weight, a further | |
3823 | * mechanism, preemption, mitigates or even eliminates this | |
3824 | * problem. And it does so without consequences on overall | |
3825 | * throughput. This mechanism and its benefits are explained | |
3826 | * in the next three paragraphs. | |
3827 | * | |
3828 | * Even if a queue, say Q, is expired when it remains idle, Q | |
3829 | * can still preempt the new in-service queue if the next | |
3830 | * request of Q arrives soon (see the comments on | |
3831 | * bfq_bfqq_update_budg_for_activation). If all queues and | |
3832 | * groups have the same weight, this form of preemption, | |
3833 | * combined with the hole-recovery heuristic described in the | |
3834 | * comments on function bfq_bfqq_update_budg_for_activation, | |
3835 | * are enough to preserve a correct bandwidth distribution in | |
3836 | * the mid term, even without idling. In fact, even if not | |
3837 | * idling allows the internal queues of the device to contain | |
3838 | * many requests, and thus to reorder requests, we can rather | |
3839 | * safely assume that the internal scheduler still preserves a | |
3840 | * minimum of mid-term fairness. | |
3841 | * | |
3842 | * More precisely, this preemption-based, idleless approach | |
3843 | * provides fairness in terms of IOPS, and not sectors per | |
3844 | * second. This can be seen with a simple example. Suppose | |
3845 | * that there are two queues with the same weight, but that | |
3846 | * the first queue receives requests of 8 sectors, while the | |
3847 | * second queue receives requests of 1024 sectors. In | |
3848 | * addition, suppose that each of the two queues contains at | |
3849 | * most one request at a time, which implies that each queue | |
3850 | * always remains idle after it is served. Finally, after | |
3851 | * remaining idle, each queue receives very quickly a new | |
3852 | * request. It follows that the two queues are served | |
3853 | * alternatively, preempting each other if needed. This | |
3854 | * implies that, although both queues have the same weight, | |
3855 | * the queue with large requests receives a service that is | |
3856 | * 1024/8 times as high as the service received by the other | |
3857 | * queue. | |
3858 | * | |
3859 | * The motivation for using preemption instead of idling (for | |
3860 | * queues with the same weight) is that, by not idling, | |
3861 | * service guarantees are preserved (completely or at least in | |
3862 | * part) without minimally sacrificing throughput. And, if | |
3863 | * there is no active group, then the primary expectation for | |
3864 | * this device is probably a high throughput. | |
3865 | * | |
3866 | * We are now left only with explaining the additional | |
3867 | * compound condition that is checked below for deciding | |
3868 | * whether the scenario is asymmetric. To explain this | |
3869 | * compound condition, we need to add that the function | |
fb53ac6c | 3870 | * bfq_asymmetric_scenario checks the weights of only |
530c4cbb PV |
3871 | * non-weight-raised queues, for efficiency reasons (see |
3872 | * comments on bfq_weights_tree_add()). Then the fact that | |
3873 | * bfqq is weight-raised is checked explicitly here. More | |
3874 | * precisely, the compound condition below takes into account | |
3875 | * also the fact that, even if bfqq is being weight-raised, | |
3876 | * the scenario is still symmetric if all queues with requests | |
3877 | * waiting for completion happen to be | |
3878 | * weight-raised. Actually, we should be even more precise | |
3879 | * here, and differentiate between interactive weight raising | |
3880 | * and soft real-time weight raising. | |
3881 | * | |
3882 | * As a side note, it is worth considering that the above | |
3883 | * device-idling countermeasures may however fail in the | |
3884 | * following unlucky scenario: if idling is (correctly) | |
3885 | * disabled in a time period during which all symmetry | |
3886 | * sub-conditions hold, and hence the device is allowed to | |
3887 | * enqueue many requests, but at some later point in time some | |
3888 | * sub-condition stops to hold, then it may become impossible | |
3889 | * to let requests be served in the desired order until all | |
3890 | * the requests already queued in the device have been served. | |
3891 | */ | |
05c2f5c3 PV |
3892 | static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd, |
3893 | struct bfq_queue *bfqq) | |
3894 | { | |
530c4cbb PV |
3895 | return (bfqq->wr_coeff > 1 && |
3896 | bfqd->wr_busy_queues < | |
3897 | bfq_tot_busy_queues(bfqd)) || | |
fb53ac6c | 3898 | bfq_asymmetric_scenario(bfqd, bfqq); |
05c2f5c3 PV |
3899 | } |
3900 | ||
3901 | /* | |
3902 | * For a queue that becomes empty, device idling is allowed only if | |
3903 | * this function returns true for that queue. As a consequence, since | |
3904 | * device idling plays a critical role for both throughput boosting | |
3905 | * and service guarantees, the return value of this function plays a | |
3906 | * critical role as well. | |
3907 | * | |
3908 | * In a nutshell, this function returns true only if idling is | |
3909 | * beneficial for throughput or, even if detrimental for throughput, | |
3910 | * idling is however necessary to preserve service guarantees (low | |
3911 | * latency, desired throughput distribution, ...). In particular, on | |
3912 | * NCQ-capable devices, this function tries to return false, so as to | |
3913 | * help keep the drives' internal queues full, whenever this helps the | |
3914 | * device boost the throughput without causing any service-guarantee | |
3915 | * issue. | |
3916 | * | |
3917 | * Most of the issues taken into account to get the return value of | |
3918 | * this function are not trivial. We discuss these issues in the two | |
3919 | * functions providing the main pieces of information needed by this | |
3920 | * function. | |
3921 | */ | |
3922 | static bool bfq_better_to_idle(struct bfq_queue *bfqq) | |
3923 | { | |
3924 | struct bfq_data *bfqd = bfqq->bfqd; | |
3925 | bool idling_boosts_thr_with_no_issue, idling_needed_for_service_guar; | |
3926 | ||
3927 | if (unlikely(bfqd->strict_guarantees)) | |
3928 | return true; | |
3929 | ||
3930 | /* | |
3931 | * Idling is performed only if slice_idle > 0. In addition, we | |
3932 | * do not idle if | |
3933 | * (a) bfqq is async | |
3934 | * (b) bfqq is in the idle io prio class: in this case we do | |
3935 | * not idle because we want to minimize the bandwidth that | |
3936 | * queues in this class can steal to higher-priority queues | |
3937 | */ | |
3938 | if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) || | |
3939 | bfq_class_idle(bfqq)) | |
3940 | return false; | |
3941 | ||
3942 | idling_boosts_thr_with_no_issue = | |
3943 | idling_boosts_thr_without_issues(bfqd, bfqq); | |
3944 | ||
3945 | idling_needed_for_service_guar = | |
3946 | idling_needed_for_service_guarantees(bfqd, bfqq); | |
e1b2324d | 3947 | |
44e44a1b | 3948 | /* |
05c2f5c3 | 3949 | * We have now the two components we need to compute the |
d5be3fef PV |
3950 | * return value of the function, which is true only if idling |
3951 | * either boosts the throughput (without issues), or is | |
3952 | * necessary to preserve service guarantees. | |
aee69d78 | 3953 | */ |
05c2f5c3 PV |
3954 | return idling_boosts_thr_with_no_issue || |
3955 | idling_needed_for_service_guar; | |
aee69d78 PV |
3956 | } |
3957 | ||
3958 | /* | |
277a4a9b | 3959 | * If the in-service queue is empty but the function bfq_better_to_idle |
aee69d78 PV |
3960 | * returns true, then: |
3961 | * 1) the queue must remain in service and cannot be expired, and | |
3962 | * 2) the device must be idled to wait for the possible arrival of a new | |
3963 | * request for the queue. | |
277a4a9b | 3964 | * See the comments on the function bfq_better_to_idle for the reasons |
aee69d78 | 3965 | * why performing device idling is the best choice to boost the throughput |
277a4a9b | 3966 | * and preserve service guarantees when bfq_better_to_idle itself |
aee69d78 PV |
3967 | * returns true. |
3968 | */ | |
3969 | static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq) | |
3970 | { | |
277a4a9b | 3971 | return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq); |
aee69d78 PV |
3972 | } |
3973 | ||
2341d662 PV |
3974 | /* |
3975 | * This function chooses the queue from which to pick the next extra | |
3976 | * I/O request to inject, if it finds a compatible queue. See the | |
3977 | * comments on bfq_update_inject_limit() for details on the injection | |
3978 | * mechanism, and for the definitions of the quantities mentioned | |
3979 | * below. | |
3980 | */ | |
3981 | static struct bfq_queue * | |
3982 | bfq_choose_bfqq_for_injection(struct bfq_data *bfqd) | |
d0edc247 | 3983 | { |
2341d662 PV |
3984 | struct bfq_queue *bfqq, *in_serv_bfqq = bfqd->in_service_queue; |
3985 | unsigned int limit = in_serv_bfqq->inject_limit; | |
3986 | /* | |
3987 | * If | |
3988 | * - bfqq is not weight-raised and therefore does not carry | |
3989 | * time-critical I/O, | |
3990 | * or | |
3991 | * - regardless of whether bfqq is weight-raised, bfqq has | |
3992 | * however a long think time, during which it can absorb the | |
3993 | * effect of an appropriate number of extra I/O requests | |
3994 | * from other queues (see bfq_update_inject_limit for | |
3995 | * details on the computation of this number); | |
3996 | * then injection can be performed without restrictions. | |
3997 | */ | |
3998 | bool in_serv_always_inject = in_serv_bfqq->wr_coeff == 1 || | |
3999 | !bfq_bfqq_has_short_ttime(in_serv_bfqq); | |
d0edc247 PV |
4000 | |
4001 | /* | |
2341d662 PV |
4002 | * If |
4003 | * - the baseline total service time could not be sampled yet, | |
4004 | * so the inject limit happens to be still 0, and | |
4005 | * - a lot of time has elapsed since the plugging of I/O | |
4006 | * dispatching started, so drive speed is being wasted | |
4007 | * significantly; | |
4008 | * then temporarily raise inject limit to one request. | |
4009 | */ | |
4010 | if (limit == 0 && in_serv_bfqq->last_serv_time_ns == 0 && | |
4011 | bfq_bfqq_wait_request(in_serv_bfqq) && | |
4012 | time_is_before_eq_jiffies(bfqd->last_idling_start_jiffies + | |
4013 | bfqd->bfq_slice_idle) | |
4014 | ) | |
4015 | limit = 1; | |
4016 | ||
4017 | if (bfqd->rq_in_driver >= limit) | |
4018 | return NULL; | |
4019 | ||
4020 | /* | |
4021 | * Linear search of the source queue for injection; but, with | |
4022 | * a high probability, very few steps are needed to find a | |
4023 | * candidate queue, i.e., a queue with enough budget left for | |
4024 | * its next request. In fact: | |
d0edc247 PV |
4025 | * - BFQ dynamically updates the budget of every queue so as |
4026 | * to accommodate the expected backlog of the queue; | |
4027 | * - if a queue gets all its requests dispatched as injected | |
4028 | * service, then the queue is removed from the active list | |
2341d662 PV |
4029 | * (and re-added only if it gets new requests, but then it |
4030 | * is assigned again enough budget for its new backlog). | |
d0edc247 PV |
4031 | */ |
4032 | list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list) | |
4033 | if (!RB_EMPTY_ROOT(&bfqq->sort_list) && | |
2341d662 | 4034 | (in_serv_always_inject || bfqq->wr_coeff > 1) && |
d0edc247 | 4035 | bfq_serv_to_charge(bfqq->next_rq, bfqq) <= |
2341d662 PV |
4036 | bfq_bfqq_budget_left(bfqq)) { |
4037 | /* | |
4038 | * Allow for only one large in-flight request | |
4039 | * on non-rotational devices, for the | |
4040 | * following reason. On non-rotationl drives, | |
4041 | * large requests take much longer than | |
4042 | * smaller requests to be served. In addition, | |
4043 | * the drive prefers to serve large requests | |
4044 | * w.r.t. to small ones, if it can choose. So, | |
4045 | * having more than one large requests queued | |
4046 | * in the drive may easily make the next first | |
4047 | * request of the in-service queue wait for so | |
4048 | * long to break bfqq's service guarantees. On | |
4049 | * the bright side, large requests let the | |
4050 | * drive reach a very high throughput, even if | |
4051 | * there is only one in-flight large request | |
4052 | * at a time. | |
4053 | */ | |
4054 | if (blk_queue_nonrot(bfqd->queue) && | |
4055 | blk_rq_sectors(bfqq->next_rq) >= | |
4056 | BFQQ_SECT_THR_NONROT) | |
4057 | limit = min_t(unsigned int, 1, limit); | |
4058 | else | |
4059 | limit = in_serv_bfqq->inject_limit; | |
4060 | ||
4061 | if (bfqd->rq_in_driver < limit) { | |
4062 | bfqd->rqs_injected = true; | |
4063 | return bfqq; | |
4064 | } | |
4065 | } | |
d0edc247 PV |
4066 | |
4067 | return NULL; | |
4068 | } | |
4069 | ||
aee69d78 PV |
4070 | /* |
4071 | * Select a queue for service. If we have a current queue in service, | |
4072 | * check whether to continue servicing it, or retrieve and set a new one. | |
4073 | */ | |
4074 | static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd) | |
4075 | { | |
4076 | struct bfq_queue *bfqq; | |
4077 | struct request *next_rq; | |
4078 | enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT; | |
4079 | ||
4080 | bfqq = bfqd->in_service_queue; | |
4081 | if (!bfqq) | |
4082 | goto new_queue; | |
4083 | ||
4084 | bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue"); | |
4085 | ||
4420b095 PV |
4086 | /* |
4087 | * Do not expire bfqq for budget timeout if bfqq may be about | |
4088 | * to enjoy device idling. The reason why, in this case, we | |
4089 | * prevent bfqq from expiring is the same as in the comments | |
4090 | * on the case where bfq_bfqq_must_idle() returns true, in | |
4091 | * bfq_completed_request(). | |
4092 | */ | |
aee69d78 | 4093 | if (bfq_may_expire_for_budg_timeout(bfqq) && |
aee69d78 PV |
4094 | !bfq_bfqq_must_idle(bfqq)) |
4095 | goto expire; | |
4096 | ||
4097 | check_queue: | |
4098 | /* | |
4099 | * This loop is rarely executed more than once. Even when it | |
4100 | * happens, it is much more convenient to re-execute this loop | |
4101 | * than to return NULL and trigger a new dispatch to get a | |
4102 | * request served. | |
4103 | */ | |
4104 | next_rq = bfqq->next_rq; | |
4105 | /* | |
4106 | * If bfqq has requests queued and it has enough budget left to | |
4107 | * serve them, keep the queue, otherwise expire it. | |
4108 | */ | |
4109 | if (next_rq) { | |
4110 | if (bfq_serv_to_charge(next_rq, bfqq) > | |
4111 | bfq_bfqq_budget_left(bfqq)) { | |
4112 | /* | |
4113 | * Expire the queue for budget exhaustion, | |
4114 | * which makes sure that the next budget is | |
4115 | * enough to serve the next request, even if | |
4116 | * it comes from the fifo expired path. | |
4117 | */ | |
4118 | reason = BFQQE_BUDGET_EXHAUSTED; | |
4119 | goto expire; | |
4120 | } else { | |
4121 | /* | |
4122 | * The idle timer may be pending because we may | |
4123 | * not disable disk idling even when a new request | |
4124 | * arrives. | |
4125 | */ | |
4126 | if (bfq_bfqq_wait_request(bfqq)) { | |
4127 | /* | |
4128 | * If we get here: 1) at least a new request | |
4129 | * has arrived but we have not disabled the | |
4130 | * timer because the request was too small, | |
4131 | * 2) then the block layer has unplugged | |
4132 | * the device, causing the dispatch to be | |
4133 | * invoked. | |
4134 | * | |
4135 | * Since the device is unplugged, now the | |
4136 | * requests are probably large enough to | |
4137 | * provide a reasonable throughput. | |
4138 | * So we disable idling. | |
4139 | */ | |
4140 | bfq_clear_bfqq_wait_request(bfqq); | |
4141 | hrtimer_try_to_cancel(&bfqd->idle_slice_timer); | |
4142 | } | |
4143 | goto keep_queue; | |
4144 | } | |
4145 | } | |
4146 | ||
4147 | /* | |
4148 | * No requests pending. However, if the in-service queue is idling | |
4149 | * for a new request, or has requests waiting for a completion and | |
4150 | * may idle after their completion, then keep it anyway. | |
d0edc247 | 4151 | * |
2341d662 PV |
4152 | * Yet, inject service from other queues if it boosts |
4153 | * throughput and is possible. | |
aee69d78 PV |
4154 | */ |
4155 | if (bfq_bfqq_wait_request(bfqq) || | |
277a4a9b | 4156 | (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) { |
2341d662 PV |
4157 | struct bfq_queue *async_bfqq = |
4158 | bfqq->bic && bfqq->bic->bfqq[0] && | |
4159 | bfq_bfqq_busy(bfqq->bic->bfqq[0]) ? | |
4160 | bfqq->bic->bfqq[0] : NULL; | |
4161 | ||
4162 | /* | |
4163 | * If the process associated with bfqq has also async | |
4164 | * I/O pending, then inject it | |
4165 | * unconditionally. Injecting I/O from the same | |
4166 | * process can cause no harm to the process. On the | |
4167 | * contrary, it can only increase bandwidth and reduce | |
4168 | * latency for the process. | |
4169 | */ | |
4170 | if (async_bfqq && | |
4171 | icq_to_bic(async_bfqq->next_rq->elv.icq) == bfqq->bic && | |
4172 | bfq_serv_to_charge(async_bfqq->next_rq, async_bfqq) <= | |
4173 | bfq_bfqq_budget_left(async_bfqq)) | |
4174 | bfqq = bfqq->bic->bfqq[0]; | |
4175 | else if (!idling_boosts_thr_without_issues(bfqd, bfqq) && | |
4176 | (bfqq->wr_coeff == 1 || bfqd->wr_busy_queues > 1 || | |
4177 | !bfq_bfqq_has_short_ttime(bfqq))) | |
d0edc247 PV |
4178 | bfqq = bfq_choose_bfqq_for_injection(bfqd); |
4179 | else | |
4180 | bfqq = NULL; | |
4181 | ||
aee69d78 PV |
4182 | goto keep_queue; |
4183 | } | |
4184 | ||
4185 | reason = BFQQE_NO_MORE_REQUESTS; | |
4186 | expire: | |
4187 | bfq_bfqq_expire(bfqd, bfqq, false, reason); | |
4188 | new_queue: | |
4189 | bfqq = bfq_set_in_service_queue(bfqd); | |
4190 | if (bfqq) { | |
4191 | bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue"); | |
4192 | goto check_queue; | |
4193 | } | |
4194 | keep_queue: | |
4195 | if (bfqq) | |
4196 | bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue"); | |
4197 | else | |
4198 | bfq_log(bfqd, "select_queue: no queue returned"); | |
4199 | ||
4200 | return bfqq; | |
4201 | } | |
4202 | ||
44e44a1b PV |
4203 | static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
4204 | { | |
4205 | struct bfq_entity *entity = &bfqq->entity; | |
4206 | ||
4207 | if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */ | |
4208 | bfq_log_bfqq(bfqd, bfqq, | |
4209 | "raising period dur %u/%u msec, old coeff %u, w %d(%d)", | |
4210 | jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish), | |
4211 | jiffies_to_msecs(bfqq->wr_cur_max_time), | |
4212 | bfqq->wr_coeff, | |
4213 | bfqq->entity.weight, bfqq->entity.orig_weight); | |
4214 | ||
4215 | if (entity->prio_changed) | |
4216 | bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change"); | |
4217 | ||
4218 | /* | |
e1b2324d AA |
4219 | * If the queue was activated in a burst, or too much |
4220 | * time has elapsed from the beginning of this | |
4221 | * weight-raising period, then end weight raising. | |
44e44a1b | 4222 | */ |
e1b2324d AA |
4223 | if (bfq_bfqq_in_large_burst(bfqq)) |
4224 | bfq_bfqq_end_wr(bfqq); | |
4225 | else if (time_is_before_jiffies(bfqq->last_wr_start_finish + | |
4226 | bfqq->wr_cur_max_time)) { | |
77b7dcea PV |
4227 | if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time || |
4228 | time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt + | |
e1b2324d | 4229 | bfq_wr_duration(bfqd))) |
77b7dcea PV |
4230 | bfq_bfqq_end_wr(bfqq); |
4231 | else { | |
3e2bdd6d | 4232 | switch_back_to_interactive_wr(bfqq, bfqd); |
77b7dcea PV |
4233 | bfqq->entity.prio_changed = 1; |
4234 | } | |
44e44a1b | 4235 | } |
8a8747dc PV |
4236 | if (bfqq->wr_coeff > 1 && |
4237 | bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time && | |
4238 | bfqq->service_from_wr > max_service_from_wr) { | |
4239 | /* see comments on max_service_from_wr */ | |
4240 | bfq_bfqq_end_wr(bfqq); | |
4241 | } | |
44e44a1b | 4242 | } |
431b17f9 PV |
4243 | /* |
4244 | * To improve latency (for this or other queues), immediately | |
4245 | * update weight both if it must be raised and if it must be | |
4246 | * lowered. Since, entity may be on some active tree here, and | |
4247 | * might have a pending change of its ioprio class, invoke | |
4248 | * next function with the last parameter unset (see the | |
4249 | * comments on the function). | |
4250 | */ | |
44e44a1b | 4251 | if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1)) |
431b17f9 PV |
4252 | __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity), |
4253 | entity, false); | |
44e44a1b PV |
4254 | } |
4255 | ||
aee69d78 PV |
4256 | /* |
4257 | * Dispatch next request from bfqq. | |
4258 | */ | |
4259 | static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd, | |
4260 | struct bfq_queue *bfqq) | |
4261 | { | |
4262 | struct request *rq = bfqq->next_rq; | |
4263 | unsigned long service_to_charge; | |
4264 | ||
4265 | service_to_charge = bfq_serv_to_charge(rq, bfqq); | |
4266 | ||
4267 | bfq_bfqq_served(bfqq, service_to_charge); | |
4268 | ||
2341d662 PV |
4269 | if (bfqq == bfqd->in_service_queue && bfqd->wait_dispatch) { |
4270 | bfqd->wait_dispatch = false; | |
4271 | bfqd->waited_rq = rq; | |
4272 | } | |
aee69d78 | 4273 | |
2341d662 | 4274 | bfq_dispatch_remove(bfqd->queue, rq); |
d0edc247 | 4275 | |
2341d662 | 4276 | if (bfqq != bfqd->in_service_queue) |
d0edc247 | 4277 | goto return_rq; |
d0edc247 | 4278 | |
44e44a1b PV |
4279 | /* |
4280 | * If weight raising has to terminate for bfqq, then next | |
4281 | * function causes an immediate update of bfqq's weight, | |
4282 | * without waiting for next activation. As a consequence, on | |
4283 | * expiration, bfqq will be timestamped as if has never been | |
4284 | * weight-raised during this service slot, even if it has | |
4285 | * received part or even most of the service as a | |
4286 | * weight-raised queue. This inflates bfqq's timestamps, which | |
4287 | * is beneficial, as bfqq is then more willing to leave the | |
4288 | * device immediately to possible other weight-raised queues. | |
4289 | */ | |
4290 | bfq_update_wr_data(bfqd, bfqq); | |
4291 | ||
aee69d78 PV |
4292 | /* |
4293 | * Expire bfqq, pretending that its budget expired, if bfqq | |
4294 | * belongs to CLASS_IDLE and other queues are waiting for | |
4295 | * service. | |
4296 | */ | |
73d58118 | 4297 | if (!(bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq))) |
d0edc247 | 4298 | goto return_rq; |
aee69d78 | 4299 | |
aee69d78 | 4300 | bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED); |
d0edc247 PV |
4301 | |
4302 | return_rq: | |
aee69d78 PV |
4303 | return rq; |
4304 | } | |
4305 | ||
4306 | static bool bfq_has_work(struct blk_mq_hw_ctx *hctx) | |
4307 | { | |
4308 | struct bfq_data *bfqd = hctx->queue->elevator->elevator_data; | |
4309 | ||
4310 | /* | |
4311 | * Avoiding lock: a race on bfqd->busy_queues should cause at | |
4312 | * most a call to dispatch for nothing | |
4313 | */ | |
4314 | return !list_empty_careful(&bfqd->dispatch) || | |
73d58118 | 4315 | bfq_tot_busy_queues(bfqd) > 0; |
aee69d78 PV |
4316 | } |
4317 | ||
4318 | static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx) | |
4319 | { | |
4320 | struct bfq_data *bfqd = hctx->queue->elevator->elevator_data; | |
4321 | struct request *rq = NULL; | |
4322 | struct bfq_queue *bfqq = NULL; | |
4323 | ||
4324 | if (!list_empty(&bfqd->dispatch)) { | |
4325 | rq = list_first_entry(&bfqd->dispatch, struct request, | |
4326 | queuelist); | |
4327 | list_del_init(&rq->queuelist); | |
4328 | ||
4329 | bfqq = RQ_BFQQ(rq); | |
4330 | ||
4331 | if (bfqq) { | |
4332 | /* | |
4333 | * Increment counters here, because this | |
4334 | * dispatch does not follow the standard | |
4335 | * dispatch flow (where counters are | |
4336 | * incremented) | |
4337 | */ | |
4338 | bfqq->dispatched++; | |
4339 | ||
4340 | goto inc_in_driver_start_rq; | |
4341 | } | |
4342 | ||
4343 | /* | |
a7877390 PV |
4344 | * We exploit the bfq_finish_requeue_request hook to |
4345 | * decrement rq_in_driver, but | |
4346 | * bfq_finish_requeue_request will not be invoked on | |
4347 | * this request. So, to avoid unbalance, just start | |
4348 | * this request, without incrementing rq_in_driver. As | |
4349 | * a negative consequence, rq_in_driver is deceptively | |
4350 | * lower than it should be while this request is in | |
4351 | * service. This may cause bfq_schedule_dispatch to be | |
4352 | * invoked uselessly. | |
aee69d78 PV |
4353 | * |
4354 | * As for implementing an exact solution, the | |
a7877390 PV |
4355 | * bfq_finish_requeue_request hook, if defined, is |
4356 | * probably invoked also on this request. So, by | |
4357 | * exploiting this hook, we could 1) increment | |
4358 | * rq_in_driver here, and 2) decrement it in | |
4359 | * bfq_finish_requeue_request. Such a solution would | |
4360 | * let the value of the counter be always accurate, | |
4361 | * but it would entail using an extra interface | |
4362 | * function. This cost seems higher than the benefit, | |
4363 | * being the frequency of non-elevator-private | |
aee69d78 PV |
4364 | * requests very low. |
4365 | */ | |
4366 | goto start_rq; | |
4367 | } | |
4368 | ||
73d58118 PV |
4369 | bfq_log(bfqd, "dispatch requests: %d busy queues", |
4370 | bfq_tot_busy_queues(bfqd)); | |
aee69d78 | 4371 | |
73d58118 | 4372 | if (bfq_tot_busy_queues(bfqd) == 0) |
aee69d78 PV |
4373 | goto exit; |
4374 | ||
4375 | /* | |
4376 | * Force device to serve one request at a time if | |
4377 | * strict_guarantees is true. Forcing this service scheme is | |
4378 | * currently the ONLY way to guarantee that the request | |
4379 | * service order enforced by the scheduler is respected by a | |
4380 | * queueing device. Otherwise the device is free even to make | |
4381 | * some unlucky request wait for as long as the device | |
4382 | * wishes. | |
4383 | * | |
4384 | * Of course, serving one request at at time may cause loss of | |
4385 | * throughput. | |
4386 | */ | |
4387 | if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0) | |
4388 | goto exit; | |
4389 | ||
4390 | bfqq = bfq_select_queue(bfqd); | |
4391 | if (!bfqq) | |
4392 | goto exit; | |
4393 | ||
4394 | rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq); | |
4395 | ||
4396 | if (rq) { | |
4397 | inc_in_driver_start_rq: | |
4398 | bfqd->rq_in_driver++; | |
4399 | start_rq: | |
4400 | rq->rq_flags |= RQF_STARTED; | |
4401 | } | |
4402 | exit: | |
4403 | return rq; | |
4404 | } | |
4405 | ||
a33801e8 | 4406 | #if defined(CONFIG_BFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP) |
9b25bd03 PV |
4407 | static void bfq_update_dispatch_stats(struct request_queue *q, |
4408 | struct request *rq, | |
4409 | struct bfq_queue *in_serv_queue, | |
4410 | bool idle_timer_disabled) | |
4411 | { | |
4412 | struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL; | |
aee69d78 | 4413 | |
24bfd19b | 4414 | if (!idle_timer_disabled && !bfqq) |
9b25bd03 | 4415 | return; |
24bfd19b PV |
4416 | |
4417 | /* | |
4418 | * rq and bfqq are guaranteed to exist until this function | |
4419 | * ends, for the following reasons. First, rq can be | |
4420 | * dispatched to the device, and then can be completed and | |
4421 | * freed, only after this function ends. Second, rq cannot be | |
4422 | * merged (and thus freed because of a merge) any longer, | |
4423 | * because it has already started. Thus rq cannot be freed | |
4424 | * before this function ends, and, since rq has a reference to | |
4425 | * bfqq, the same guarantee holds for bfqq too. | |
4426 | * | |
4427 | * In addition, the following queue lock guarantees that | |
4428 | * bfqq_group(bfqq) exists as well. | |
4429 | */ | |
0d945c1f | 4430 | spin_lock_irq(&q->queue_lock); |
24bfd19b PV |
4431 | if (idle_timer_disabled) |
4432 | /* | |
4433 | * Since the idle timer has been disabled, | |
4434 | * in_serv_queue contained some request when | |
4435 | * __bfq_dispatch_request was invoked above, which | |
4436 | * implies that rq was picked exactly from | |
4437 | * in_serv_queue. Thus in_serv_queue == bfqq, and is | |
4438 | * therefore guaranteed to exist because of the above | |
4439 | * arguments. | |
4440 | */ | |
4441 | bfqg_stats_update_idle_time(bfqq_group(in_serv_queue)); | |
4442 | if (bfqq) { | |
4443 | struct bfq_group *bfqg = bfqq_group(bfqq); | |
4444 | ||
4445 | bfqg_stats_update_avg_queue_size(bfqg); | |
4446 | bfqg_stats_set_start_empty_time(bfqg); | |
4447 | bfqg_stats_update_io_remove(bfqg, rq->cmd_flags); | |
4448 | } | |
0d945c1f | 4449 | spin_unlock_irq(&q->queue_lock); |
9b25bd03 PV |
4450 | } |
4451 | #else | |
4452 | static inline void bfq_update_dispatch_stats(struct request_queue *q, | |
4453 | struct request *rq, | |
4454 | struct bfq_queue *in_serv_queue, | |
4455 | bool idle_timer_disabled) {} | |
24bfd19b PV |
4456 | #endif |
4457 | ||
9b25bd03 PV |
4458 | static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx) |
4459 | { | |
4460 | struct bfq_data *bfqd = hctx->queue->elevator->elevator_data; | |
4461 | struct request *rq; | |
4462 | struct bfq_queue *in_serv_queue; | |
4463 | bool waiting_rq, idle_timer_disabled; | |
4464 | ||
4465 | spin_lock_irq(&bfqd->lock); | |
4466 | ||
4467 | in_serv_queue = bfqd->in_service_queue; | |
4468 | waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue); | |
4469 | ||
4470 | rq = __bfq_dispatch_request(hctx); | |
4471 | ||
4472 | idle_timer_disabled = | |
4473 | waiting_rq && !bfq_bfqq_wait_request(in_serv_queue); | |
4474 | ||
4475 | spin_unlock_irq(&bfqd->lock); | |
4476 | ||
4477 | bfq_update_dispatch_stats(hctx->queue, rq, in_serv_queue, | |
4478 | idle_timer_disabled); | |
4479 | ||
aee69d78 PV |
4480 | return rq; |
4481 | } | |
4482 | ||
4483 | /* | |
4484 | * Task holds one reference to the queue, dropped when task exits. Each rq | |
4485 | * in-flight on this queue also holds a reference, dropped when rq is freed. | |
4486 | * | |
4487 | * Scheduler lock must be held here. Recall not to use bfqq after calling | |
4488 | * this function on it. | |
4489 | */ | |
ea25da48 | 4490 | void bfq_put_queue(struct bfq_queue *bfqq) |
aee69d78 | 4491 | { |
e21b7a0b AA |
4492 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
4493 | struct bfq_group *bfqg = bfqq_group(bfqq); | |
4494 | #endif | |
4495 | ||
aee69d78 PV |
4496 | if (bfqq->bfqd) |
4497 | bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d", | |
4498 | bfqq, bfqq->ref); | |
4499 | ||
4500 | bfqq->ref--; | |
4501 | if (bfqq->ref) | |
4502 | return; | |
4503 | ||
99fead8d | 4504 | if (!hlist_unhashed(&bfqq->burst_list_node)) { |
e1b2324d | 4505 | hlist_del_init(&bfqq->burst_list_node); |
99fead8d PV |
4506 | /* |
4507 | * Decrement also burst size after the removal, if the | |
4508 | * process associated with bfqq is exiting, and thus | |
4509 | * does not contribute to the burst any longer. This | |
4510 | * decrement helps filter out false positives of large | |
4511 | * bursts, when some short-lived process (often due to | |
4512 | * the execution of commands by some service) happens | |
4513 | * to start and exit while a complex application is | |
4514 | * starting, and thus spawning several processes that | |
4515 | * do I/O (and that *must not* be treated as a large | |
4516 | * burst, see comments on bfq_handle_burst). | |
4517 | * | |
4518 | * In particular, the decrement is performed only if: | |
4519 | * 1) bfqq is not a merged queue, because, if it is, | |
4520 | * then this free of bfqq is not triggered by the exit | |
4521 | * of the process bfqq is associated with, but exactly | |
4522 | * by the fact that bfqq has just been merged. | |
4523 | * 2) burst_size is greater than 0, to handle | |
4524 | * unbalanced decrements. Unbalanced decrements may | |
4525 | * happen in te following case: bfqq is inserted into | |
4526 | * the current burst list--without incrementing | |
4527 | * bust_size--because of a split, but the current | |
4528 | * burst list is not the burst list bfqq belonged to | |
4529 | * (see comments on the case of a split in | |
4530 | * bfq_set_request). | |
4531 | */ | |
4532 | if (bfqq->bic && bfqq->bfqd->burst_size > 0) | |
4533 | bfqq->bfqd->burst_size--; | |
7cb04004 | 4534 | } |
e21b7a0b | 4535 | |
aee69d78 | 4536 | kmem_cache_free(bfq_pool, bfqq); |
e21b7a0b | 4537 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
8f9bebc3 | 4538 | bfqg_and_blkg_put(bfqg); |
e21b7a0b | 4539 | #endif |
aee69d78 PV |
4540 | } |
4541 | ||
36eca894 AA |
4542 | static void bfq_put_cooperator(struct bfq_queue *bfqq) |
4543 | { | |
4544 | struct bfq_queue *__bfqq, *next; | |
4545 | ||
4546 | /* | |
4547 | * If this queue was scheduled to merge with another queue, be | |
4548 | * sure to drop the reference taken on that queue (and others in | |
4549 | * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs. | |
4550 | */ | |
4551 | __bfqq = bfqq->new_bfqq; | |
4552 | while (__bfqq) { | |
4553 | if (__bfqq == bfqq) | |
4554 | break; | |
4555 | next = __bfqq->new_bfqq; | |
4556 | bfq_put_queue(__bfqq); | |
4557 | __bfqq = next; | |
4558 | } | |
4559 | } | |
4560 | ||
aee69d78 PV |
4561 | static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
4562 | { | |
4563 | if (bfqq == bfqd->in_service_queue) { | |
4564 | __bfq_bfqq_expire(bfqd, bfqq); | |
4565 | bfq_schedule_dispatch(bfqd); | |
4566 | } | |
4567 | ||
4568 | bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref); | |
4569 | ||
36eca894 AA |
4570 | bfq_put_cooperator(bfqq); |
4571 | ||
aee69d78 PV |
4572 | bfq_put_queue(bfqq); /* release process reference */ |
4573 | } | |
4574 | ||
4575 | static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync) | |
4576 | { | |
4577 | struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync); | |
4578 | struct bfq_data *bfqd; | |
4579 | ||
4580 | if (bfqq) | |
4581 | bfqd = bfqq->bfqd; /* NULL if scheduler already exited */ | |
4582 | ||
4583 | if (bfqq && bfqd) { | |
4584 | unsigned long flags; | |
4585 | ||
4586 | spin_lock_irqsave(&bfqd->lock, flags); | |
4587 | bfq_exit_bfqq(bfqd, bfqq); | |
4588 | bic_set_bfqq(bic, NULL, is_sync); | |
6fa3e8d3 | 4589 | spin_unlock_irqrestore(&bfqd->lock, flags); |
aee69d78 PV |
4590 | } |
4591 | } | |
4592 | ||
4593 | static void bfq_exit_icq(struct io_cq *icq) | |
4594 | { | |
4595 | struct bfq_io_cq *bic = icq_to_bic(icq); | |
4596 | ||
4597 | bfq_exit_icq_bfqq(bic, true); | |
4598 | bfq_exit_icq_bfqq(bic, false); | |
4599 | } | |
4600 | ||
4601 | /* | |
4602 | * Update the entity prio values; note that the new values will not | |
4603 | * be used until the next (re)activation. | |
4604 | */ | |
4605 | static void | |
4606 | bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic) | |
4607 | { | |
4608 | struct task_struct *tsk = current; | |
4609 | int ioprio_class; | |
4610 | struct bfq_data *bfqd = bfqq->bfqd; | |
4611 | ||
4612 | if (!bfqd) | |
4613 | return; | |
4614 | ||
4615 | ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio); | |
4616 | switch (ioprio_class) { | |
4617 | default: | |
4618 | dev_err(bfqq->bfqd->queue->backing_dev_info->dev, | |
4619 | "bfq: bad prio class %d\n", ioprio_class); | |
fa393d1b | 4620 | /* fall through */ |
aee69d78 PV |
4621 | case IOPRIO_CLASS_NONE: |
4622 | /* | |
4623 | * No prio set, inherit CPU scheduling settings. | |
4624 | */ | |
4625 | bfqq->new_ioprio = task_nice_ioprio(tsk); | |
4626 | bfqq->new_ioprio_class = task_nice_ioclass(tsk); | |
4627 | break; | |
4628 | case IOPRIO_CLASS_RT: | |
4629 | bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio); | |
4630 | bfqq->new_ioprio_class = IOPRIO_CLASS_RT; | |
4631 | break; | |
4632 | case IOPRIO_CLASS_BE: | |
4633 | bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio); | |
4634 | bfqq->new_ioprio_class = IOPRIO_CLASS_BE; | |
4635 | break; | |
4636 | case IOPRIO_CLASS_IDLE: | |
4637 | bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE; | |
4638 | bfqq->new_ioprio = 7; | |
aee69d78 PV |
4639 | break; |
4640 | } | |
4641 | ||
4642 | if (bfqq->new_ioprio >= IOPRIO_BE_NR) { | |
4643 | pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n", | |
4644 | bfqq->new_ioprio); | |
4645 | bfqq->new_ioprio = IOPRIO_BE_NR; | |
4646 | } | |
4647 | ||
4648 | bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio); | |
4649 | bfqq->entity.prio_changed = 1; | |
4650 | } | |
4651 | ||
ea25da48 PV |
4652 | static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd, |
4653 | struct bio *bio, bool is_sync, | |
4654 | struct bfq_io_cq *bic); | |
4655 | ||
aee69d78 PV |
4656 | static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio) |
4657 | { | |
4658 | struct bfq_data *bfqd = bic_to_bfqd(bic); | |
4659 | struct bfq_queue *bfqq; | |
4660 | int ioprio = bic->icq.ioc->ioprio; | |
4661 | ||
4662 | /* | |
4663 | * This condition may trigger on a newly created bic, be sure to | |
4664 | * drop the lock before returning. | |
4665 | */ | |
4666 | if (unlikely(!bfqd) || likely(bic->ioprio == ioprio)) | |
4667 | return; | |
4668 | ||
4669 | bic->ioprio = ioprio; | |
4670 | ||
4671 | bfqq = bic_to_bfqq(bic, false); | |
4672 | if (bfqq) { | |
4673 | /* release process reference on this queue */ | |
4674 | bfq_put_queue(bfqq); | |
4675 | bfqq = bfq_get_queue(bfqd, bio, BLK_RW_ASYNC, bic); | |
4676 | bic_set_bfqq(bic, bfqq, false); | |
4677 | } | |
4678 | ||
4679 | bfqq = bic_to_bfqq(bic, true); | |
4680 | if (bfqq) | |
4681 | bfq_set_next_ioprio_data(bfqq, bic); | |
4682 | } | |
4683 | ||
4684 | static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
4685 | struct bfq_io_cq *bic, pid_t pid, int is_sync) | |
4686 | { | |
4687 | RB_CLEAR_NODE(&bfqq->entity.rb_node); | |
4688 | INIT_LIST_HEAD(&bfqq->fifo); | |
e1b2324d | 4689 | INIT_HLIST_NODE(&bfqq->burst_list_node); |
aee69d78 PV |
4690 | |
4691 | bfqq->ref = 0; | |
4692 | bfqq->bfqd = bfqd; | |
4693 | ||
4694 | if (bic) | |
4695 | bfq_set_next_ioprio_data(bfqq, bic); | |
4696 | ||
4697 | if (is_sync) { | |
d5be3fef PV |
4698 | /* |
4699 | * No need to mark as has_short_ttime if in | |
4700 | * idle_class, because no device idling is performed | |
4701 | * for queues in idle class | |
4702 | */ | |
aee69d78 | 4703 | if (!bfq_class_idle(bfqq)) |
d5be3fef PV |
4704 | /* tentatively mark as has_short_ttime */ |
4705 | bfq_mark_bfqq_has_short_ttime(bfqq); | |
aee69d78 | 4706 | bfq_mark_bfqq_sync(bfqq); |
e1b2324d | 4707 | bfq_mark_bfqq_just_created(bfqq); |
aee69d78 PV |
4708 | } else |
4709 | bfq_clear_bfqq_sync(bfqq); | |
4710 | ||
4711 | /* set end request to minus infinity from now */ | |
4712 | bfqq->ttime.last_end_request = ktime_get_ns() + 1; | |
4713 | ||
4714 | bfq_mark_bfqq_IO_bound(bfqq); | |
4715 | ||
4716 | bfqq->pid = pid; | |
4717 | ||
4718 | /* Tentative initial value to trade off between thr and lat */ | |
54b60456 | 4719 | bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3; |
aee69d78 | 4720 | bfqq->budget_timeout = bfq_smallest_from_now(); |
aee69d78 | 4721 | |
44e44a1b | 4722 | bfqq->wr_coeff = 1; |
36eca894 | 4723 | bfqq->last_wr_start_finish = jiffies; |
77b7dcea | 4724 | bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now(); |
36eca894 | 4725 | bfqq->split_time = bfq_smallest_from_now(); |
77b7dcea PV |
4726 | |
4727 | /* | |
a34b0244 PV |
4728 | * To not forget the possibly high bandwidth consumed by a |
4729 | * process/queue in the recent past, | |
4730 | * bfq_bfqq_softrt_next_start() returns a value at least equal | |
4731 | * to the current value of bfqq->soft_rt_next_start (see | |
4732 | * comments on bfq_bfqq_softrt_next_start). Set | |
4733 | * soft_rt_next_start to now, to mean that bfqq has consumed | |
4734 | * no bandwidth so far. | |
77b7dcea | 4735 | */ |
a34b0244 | 4736 | bfqq->soft_rt_next_start = jiffies; |
44e44a1b | 4737 | |
aee69d78 PV |
4738 | /* first request is almost certainly seeky */ |
4739 | bfqq->seek_history = 1; | |
4740 | } | |
4741 | ||
4742 | static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd, | |
e21b7a0b | 4743 | struct bfq_group *bfqg, |
aee69d78 PV |
4744 | int ioprio_class, int ioprio) |
4745 | { | |
4746 | switch (ioprio_class) { | |
4747 | case IOPRIO_CLASS_RT: | |
e21b7a0b | 4748 | return &bfqg->async_bfqq[0][ioprio]; |
aee69d78 PV |
4749 | case IOPRIO_CLASS_NONE: |
4750 | ioprio = IOPRIO_NORM; | |
4751 | /* fall through */ | |
4752 | case IOPRIO_CLASS_BE: | |
e21b7a0b | 4753 | return &bfqg->async_bfqq[1][ioprio]; |
aee69d78 | 4754 | case IOPRIO_CLASS_IDLE: |
e21b7a0b | 4755 | return &bfqg->async_idle_bfqq; |
aee69d78 PV |
4756 | default: |
4757 | return NULL; | |
4758 | } | |
4759 | } | |
4760 | ||
4761 | static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd, | |
4762 | struct bio *bio, bool is_sync, | |
4763 | struct bfq_io_cq *bic) | |
4764 | { | |
4765 | const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio); | |
4766 | const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio); | |
4767 | struct bfq_queue **async_bfqq = NULL; | |
4768 | struct bfq_queue *bfqq; | |
e21b7a0b | 4769 | struct bfq_group *bfqg; |
aee69d78 PV |
4770 | |
4771 | rcu_read_lock(); | |
4772 | ||
0fe061b9 | 4773 | bfqg = bfq_find_set_group(bfqd, __bio_blkcg(bio)); |
e21b7a0b AA |
4774 | if (!bfqg) { |
4775 | bfqq = &bfqd->oom_bfqq; | |
4776 | goto out; | |
4777 | } | |
4778 | ||
aee69d78 | 4779 | if (!is_sync) { |
e21b7a0b | 4780 | async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class, |
aee69d78 PV |
4781 | ioprio); |
4782 | bfqq = *async_bfqq; | |
4783 | if (bfqq) | |
4784 | goto out; | |
4785 | } | |
4786 | ||
4787 | bfqq = kmem_cache_alloc_node(bfq_pool, | |
4788 | GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN, | |
4789 | bfqd->queue->node); | |
4790 | ||
4791 | if (bfqq) { | |
4792 | bfq_init_bfqq(bfqd, bfqq, bic, current->pid, | |
4793 | is_sync); | |
e21b7a0b | 4794 | bfq_init_entity(&bfqq->entity, bfqg); |
aee69d78 PV |
4795 | bfq_log_bfqq(bfqd, bfqq, "allocated"); |
4796 | } else { | |
4797 | bfqq = &bfqd->oom_bfqq; | |
4798 | bfq_log_bfqq(bfqd, bfqq, "using oom bfqq"); | |
4799 | goto out; | |
4800 | } | |
4801 | ||
4802 | /* | |
4803 | * Pin the queue now that it's allocated, scheduler exit will | |
4804 | * prune it. | |
4805 | */ | |
4806 | if (async_bfqq) { | |
e21b7a0b AA |
4807 | bfqq->ref++; /* |
4808 | * Extra group reference, w.r.t. sync | |
4809 | * queue. This extra reference is removed | |
4810 | * only if bfqq->bfqg disappears, to | |
4811 | * guarantee that this queue is not freed | |
4812 | * until its group goes away. | |
4813 | */ | |
4814 | bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d", | |
aee69d78 PV |
4815 | bfqq, bfqq->ref); |
4816 | *async_bfqq = bfqq; | |
4817 | } | |
4818 | ||
4819 | out: | |
4820 | bfqq->ref++; /* get a process reference to this queue */ | |
4821 | bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, bfqq->ref); | |
4822 | rcu_read_unlock(); | |
4823 | return bfqq; | |
4824 | } | |
4825 | ||
4826 | static void bfq_update_io_thinktime(struct bfq_data *bfqd, | |
4827 | struct bfq_queue *bfqq) | |
4828 | { | |
4829 | struct bfq_ttime *ttime = &bfqq->ttime; | |
4830 | u64 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request; | |
4831 | ||
4832 | elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle); | |
4833 | ||
4834 | ttime->ttime_samples = (7*bfqq->ttime.ttime_samples + 256) / 8; | |
4835 | ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8); | |
4836 | ttime->ttime_mean = div64_ul(ttime->ttime_total + 128, | |
4837 | ttime->ttime_samples); | |
4838 | } | |
4839 | ||
4840 | static void | |
4841 | bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
4842 | struct request *rq) | |
4843 | { | |
aee69d78 | 4844 | bfqq->seek_history <<= 1; |
d87447d8 | 4845 | bfqq->seek_history |= BFQ_RQ_SEEKY(bfqd, bfqq->last_request_pos, rq); |
7074f076 PV |
4846 | |
4847 | if (bfqq->wr_coeff > 1 && | |
4848 | bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time && | |
4849 | BFQQ_TOTALLY_SEEKY(bfqq)) | |
4850 | bfq_bfqq_end_wr(bfqq); | |
aee69d78 PV |
4851 | } |
4852 | ||
d5be3fef PV |
4853 | static void bfq_update_has_short_ttime(struct bfq_data *bfqd, |
4854 | struct bfq_queue *bfqq, | |
4855 | struct bfq_io_cq *bic) | |
aee69d78 | 4856 | { |
d5be3fef | 4857 | bool has_short_ttime = true; |
aee69d78 | 4858 | |
d5be3fef PV |
4859 | /* |
4860 | * No need to update has_short_ttime if bfqq is async or in | |
4861 | * idle io prio class, or if bfq_slice_idle is zero, because | |
4862 | * no device idling is performed for bfqq in this case. | |
4863 | */ | |
4864 | if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) || | |
4865 | bfqd->bfq_slice_idle == 0) | |
aee69d78 PV |
4866 | return; |
4867 | ||
36eca894 AA |
4868 | /* Idle window just restored, statistics are meaningless. */ |
4869 | if (time_is_after_eq_jiffies(bfqq->split_time + | |
4870 | bfqd->bfq_wr_min_idle_time)) | |
4871 | return; | |
4872 | ||
d5be3fef PV |
4873 | /* Think time is infinite if no process is linked to |
4874 | * bfqq. Otherwise check average think time to | |
4875 | * decide whether to mark as has_short_ttime | |
4876 | */ | |
aee69d78 | 4877 | if (atomic_read(&bic->icq.ioc->active_ref) == 0 || |
d5be3fef PV |
4878 | (bfq_sample_valid(bfqq->ttime.ttime_samples) && |
4879 | bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle)) | |
4880 | has_short_ttime = false; | |
4881 | ||
4882 | bfq_log_bfqq(bfqd, bfqq, "update_has_short_ttime: has_short_ttime %d", | |
4883 | has_short_ttime); | |
aee69d78 | 4884 | |
d5be3fef PV |
4885 | if (has_short_ttime) |
4886 | bfq_mark_bfqq_has_short_ttime(bfqq); | |
aee69d78 | 4887 | else |
d5be3fef | 4888 | bfq_clear_bfqq_has_short_ttime(bfqq); |
aee69d78 PV |
4889 | } |
4890 | ||
4891 | /* | |
4892 | * Called when a new fs request (rq) is added to bfqq. Check if there's | |
4893 | * something we should do about it. | |
4894 | */ | |
4895 | static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
4896 | struct request *rq) | |
4897 | { | |
4898 | struct bfq_io_cq *bic = RQ_BIC(rq); | |
4899 | ||
4900 | if (rq->cmd_flags & REQ_META) | |
4901 | bfqq->meta_pending++; | |
4902 | ||
4903 | bfq_update_io_thinktime(bfqd, bfqq); | |
d5be3fef | 4904 | bfq_update_has_short_ttime(bfqd, bfqq, bic); |
aee69d78 | 4905 | bfq_update_io_seektime(bfqd, bfqq, rq); |
aee69d78 PV |
4906 | |
4907 | bfq_log_bfqq(bfqd, bfqq, | |
d5be3fef PV |
4908 | "rq_enqueued: has_short_ttime=%d (seeky %d)", |
4909 | bfq_bfqq_has_short_ttime(bfqq), BFQQ_SEEKY(bfqq)); | |
aee69d78 PV |
4910 | |
4911 | bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); | |
4912 | ||
4913 | if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) { | |
4914 | bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 && | |
4915 | blk_rq_sectors(rq) < 32; | |
4916 | bool budget_timeout = bfq_bfqq_budget_timeout(bfqq); | |
4917 | ||
4918 | /* | |
ac8b0cb4 PV |
4919 | * There is just this request queued: if |
4920 | * - the request is small, and | |
4921 | * - we are idling to boost throughput, and | |
4922 | * - the queue is not to be expired, | |
4923 | * then just exit. | |
aee69d78 PV |
4924 | * |
4925 | * In this way, if the device is being idled to wait | |
4926 | * for a new request from the in-service queue, we | |
4927 | * avoid unplugging the device and committing the | |
ac8b0cb4 PV |
4928 | * device to serve just a small request. In contrast |
4929 | * we wait for the block layer to decide when to | |
4930 | * unplug the device: hopefully, new requests will be | |
4931 | * merged to this one quickly, then the device will be | |
4932 | * unplugged and larger requests will be dispatched. | |
aee69d78 | 4933 | */ |
ac8b0cb4 PV |
4934 | if (small_req && idling_boosts_thr_without_issues(bfqd, bfqq) && |
4935 | !budget_timeout) | |
aee69d78 PV |
4936 | return; |
4937 | ||
4938 | /* | |
ac8b0cb4 PV |
4939 | * A large enough request arrived, or idling is being |
4940 | * performed to preserve service guarantees, or | |
4941 | * finally the queue is to be expired: in all these | |
4942 | * cases disk idling is to be stopped, so clear | |
4943 | * wait_request flag and reset timer. | |
aee69d78 PV |
4944 | */ |
4945 | bfq_clear_bfqq_wait_request(bfqq); | |
4946 | hrtimer_try_to_cancel(&bfqd->idle_slice_timer); | |
4947 | ||
4948 | /* | |
4949 | * The queue is not empty, because a new request just | |
4950 | * arrived. Hence we can safely expire the queue, in | |
4951 | * case of budget timeout, without risking that the | |
4952 | * timestamps of the queue are not updated correctly. | |
4953 | * See [1] for more details. | |
4954 | */ | |
4955 | if (budget_timeout) | |
4956 | bfq_bfqq_expire(bfqd, bfqq, false, | |
4957 | BFQQE_BUDGET_TIMEOUT); | |
4958 | } | |
4959 | } | |
4960 | ||
24bfd19b PV |
4961 | /* returns true if it causes the idle timer to be disabled */ |
4962 | static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq) | |
aee69d78 | 4963 | { |
36eca894 AA |
4964 | struct bfq_queue *bfqq = RQ_BFQQ(rq), |
4965 | *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true); | |
24bfd19b | 4966 | bool waiting, idle_timer_disabled = false; |
36eca894 AA |
4967 | |
4968 | if (new_bfqq) { | |
36eca894 AA |
4969 | /* |
4970 | * Release the request's reference to the old bfqq | |
4971 | * and make sure one is taken to the shared queue. | |
4972 | */ | |
4973 | new_bfqq->allocated++; | |
4974 | bfqq->allocated--; | |
4975 | new_bfqq->ref++; | |
4976 | /* | |
4977 | * If the bic associated with the process | |
4978 | * issuing this request still points to bfqq | |
4979 | * (and thus has not been already redirected | |
4980 | * to new_bfqq or even some other bfq_queue), | |
4981 | * then complete the merge and redirect it to | |
4982 | * new_bfqq. | |
4983 | */ | |
4984 | if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq) | |
4985 | bfq_merge_bfqqs(bfqd, RQ_BIC(rq), | |
4986 | bfqq, new_bfqq); | |
894df937 PV |
4987 | |
4988 | bfq_clear_bfqq_just_created(bfqq); | |
36eca894 AA |
4989 | /* |
4990 | * rq is about to be enqueued into new_bfqq, | |
4991 | * release rq reference on bfqq | |
4992 | */ | |
4993 | bfq_put_queue(bfqq); | |
4994 | rq->elv.priv[1] = new_bfqq; | |
4995 | bfqq = new_bfqq; | |
4996 | } | |
aee69d78 | 4997 | |
24bfd19b | 4998 | waiting = bfqq && bfq_bfqq_wait_request(bfqq); |
aee69d78 | 4999 | bfq_add_request(rq); |
24bfd19b | 5000 | idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq); |
aee69d78 PV |
5001 | |
5002 | rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)]; | |
5003 | list_add_tail(&rq->queuelist, &bfqq->fifo); | |
5004 | ||
5005 | bfq_rq_enqueued(bfqd, bfqq, rq); | |
24bfd19b PV |
5006 | |
5007 | return idle_timer_disabled; | |
aee69d78 PV |
5008 | } |
5009 | ||
9b25bd03 PV |
5010 | #if defined(CONFIG_BFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP) |
5011 | static void bfq_update_insert_stats(struct request_queue *q, | |
5012 | struct bfq_queue *bfqq, | |
5013 | bool idle_timer_disabled, | |
5014 | unsigned int cmd_flags) | |
5015 | { | |
5016 | if (!bfqq) | |
5017 | return; | |
5018 | ||
5019 | /* | |
5020 | * bfqq still exists, because it can disappear only after | |
5021 | * either it is merged with another queue, or the process it | |
5022 | * is associated with exits. But both actions must be taken by | |
5023 | * the same process currently executing this flow of | |
5024 | * instructions. | |
5025 | * | |
5026 | * In addition, the following queue lock guarantees that | |
5027 | * bfqq_group(bfqq) exists as well. | |
5028 | */ | |
0d945c1f | 5029 | spin_lock_irq(&q->queue_lock); |
9b25bd03 PV |
5030 | bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags); |
5031 | if (idle_timer_disabled) | |
5032 | bfqg_stats_update_idle_time(bfqq_group(bfqq)); | |
0d945c1f | 5033 | spin_unlock_irq(&q->queue_lock); |
9b25bd03 PV |
5034 | } |
5035 | #else | |
5036 | static inline void bfq_update_insert_stats(struct request_queue *q, | |
5037 | struct bfq_queue *bfqq, | |
5038 | bool idle_timer_disabled, | |
5039 | unsigned int cmd_flags) {} | |
5040 | #endif | |
5041 | ||
aee69d78 PV |
5042 | static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
5043 | bool at_head) | |
5044 | { | |
5045 | struct request_queue *q = hctx->queue; | |
5046 | struct bfq_data *bfqd = q->elevator->elevator_data; | |
18e5a57d | 5047 | struct bfq_queue *bfqq; |
24bfd19b PV |
5048 | bool idle_timer_disabled = false; |
5049 | unsigned int cmd_flags; | |
aee69d78 PV |
5050 | |
5051 | spin_lock_irq(&bfqd->lock); | |
5052 | if (blk_mq_sched_try_insert_merge(q, rq)) { | |
5053 | spin_unlock_irq(&bfqd->lock); | |
5054 | return; | |
5055 | } | |
5056 | ||
5057 | spin_unlock_irq(&bfqd->lock); | |
5058 | ||
5059 | blk_mq_sched_request_inserted(rq); | |
5060 | ||
5061 | spin_lock_irq(&bfqd->lock); | |
18e5a57d | 5062 | bfqq = bfq_init_rq(rq); |
aee69d78 PV |
5063 | if (at_head || blk_rq_is_passthrough(rq)) { |
5064 | if (at_head) | |
5065 | list_add(&rq->queuelist, &bfqd->dispatch); | |
5066 | else | |
5067 | list_add_tail(&rq->queuelist, &bfqd->dispatch); | |
18e5a57d | 5068 | } else { /* bfqq is assumed to be non null here */ |
24bfd19b | 5069 | idle_timer_disabled = __bfq_insert_request(bfqd, rq); |
614822f8 LM |
5070 | /* |
5071 | * Update bfqq, because, if a queue merge has occurred | |
5072 | * in __bfq_insert_request, then rq has been | |
5073 | * redirected into a new queue. | |
5074 | */ | |
5075 | bfqq = RQ_BFQQ(rq); | |
aee69d78 PV |
5076 | |
5077 | if (rq_mergeable(rq)) { | |
5078 | elv_rqhash_add(q, rq); | |
5079 | if (!q->last_merge) | |
5080 | q->last_merge = rq; | |
5081 | } | |
5082 | } | |
5083 | ||
24bfd19b PV |
5084 | /* |
5085 | * Cache cmd_flags before releasing scheduler lock, because rq | |
5086 | * may disappear afterwards (for example, because of a request | |
5087 | * merge). | |
5088 | */ | |
5089 | cmd_flags = rq->cmd_flags; | |
9b25bd03 | 5090 | |
6fa3e8d3 | 5091 | spin_unlock_irq(&bfqd->lock); |
24bfd19b | 5092 | |
9b25bd03 PV |
5093 | bfq_update_insert_stats(q, bfqq, idle_timer_disabled, |
5094 | cmd_flags); | |
aee69d78 PV |
5095 | } |
5096 | ||
5097 | static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx, | |
5098 | struct list_head *list, bool at_head) | |
5099 | { | |
5100 | while (!list_empty(list)) { | |
5101 | struct request *rq; | |
5102 | ||
5103 | rq = list_first_entry(list, struct request, queuelist); | |
5104 | list_del_init(&rq->queuelist); | |
5105 | bfq_insert_request(hctx, rq, at_head); | |
5106 | } | |
5107 | } | |
5108 | ||
5109 | static void bfq_update_hw_tag(struct bfq_data *bfqd) | |
5110 | { | |
b3c34981 PV |
5111 | struct bfq_queue *bfqq = bfqd->in_service_queue; |
5112 | ||
aee69d78 PV |
5113 | bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver, |
5114 | bfqd->rq_in_driver); | |
5115 | ||
5116 | if (bfqd->hw_tag == 1) | |
5117 | return; | |
5118 | ||
5119 | /* | |
5120 | * This sample is valid if the number of outstanding requests | |
5121 | * is large enough to allow a queueing behavior. Note that the | |
5122 | * sum is not exact, as it's not taking into account deactivated | |
5123 | * requests. | |
5124 | */ | |
a3c92560 | 5125 | if (bfqd->rq_in_driver + bfqd->queued <= BFQ_HW_QUEUE_THRESHOLD) |
aee69d78 PV |
5126 | return; |
5127 | ||
b3c34981 PV |
5128 | /* |
5129 | * If active queue hasn't enough requests and can idle, bfq might not | |
5130 | * dispatch sufficient requests to hardware. Don't zero hw_tag in this | |
5131 | * case | |
5132 | */ | |
5133 | if (bfqq && bfq_bfqq_has_short_ttime(bfqq) && | |
5134 | bfqq->dispatched + bfqq->queued[0] + bfqq->queued[1] < | |
5135 | BFQ_HW_QUEUE_THRESHOLD && | |
5136 | bfqd->rq_in_driver < BFQ_HW_QUEUE_THRESHOLD) | |
5137 | return; | |
5138 | ||
aee69d78 PV |
5139 | if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES) |
5140 | return; | |
5141 | ||
5142 | bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD; | |
5143 | bfqd->max_rq_in_driver = 0; | |
5144 | bfqd->hw_tag_samples = 0; | |
8cacc5ab PV |
5145 | |
5146 | bfqd->nonrot_with_queueing = | |
5147 | blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag; | |
aee69d78 PV |
5148 | } |
5149 | ||
5150 | static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd) | |
5151 | { | |
ab0e43e9 PV |
5152 | u64 now_ns; |
5153 | u32 delta_us; | |
5154 | ||
aee69d78 PV |
5155 | bfq_update_hw_tag(bfqd); |
5156 | ||
5157 | bfqd->rq_in_driver--; | |
5158 | bfqq->dispatched--; | |
5159 | ||
44e44a1b PV |
5160 | if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) { |
5161 | /* | |
5162 | * Set budget_timeout (which we overload to store the | |
5163 | * time at which the queue remains with no backlog and | |
5164 | * no outstanding request; used by the weight-raising | |
5165 | * mechanism). | |
5166 | */ | |
5167 | bfqq->budget_timeout = jiffies; | |
1de0c4cd | 5168 | |
0471559c | 5169 | bfq_weights_tree_remove(bfqd, bfqq); |
44e44a1b PV |
5170 | } |
5171 | ||
ab0e43e9 PV |
5172 | now_ns = ktime_get_ns(); |
5173 | ||
5174 | bfqq->ttime.last_end_request = now_ns; | |
5175 | ||
5176 | /* | |
5177 | * Using us instead of ns, to get a reasonable precision in | |
5178 | * computing rate in next check. | |
5179 | */ | |
5180 | delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC); | |
5181 | ||
5182 | /* | |
5183 | * If the request took rather long to complete, and, according | |
5184 | * to the maximum request size recorded, this completion latency | |
5185 | * implies that the request was certainly served at a very low | |
5186 | * rate (less than 1M sectors/sec), then the whole observation | |
5187 | * interval that lasts up to this time instant cannot be a | |
5188 | * valid time interval for computing a new peak rate. Invoke | |
5189 | * bfq_update_rate_reset to have the following three steps | |
5190 | * taken: | |
5191 | * - close the observation interval at the last (previous) | |
5192 | * request dispatch or completion | |
5193 | * - compute rate, if possible, for that observation interval | |
5194 | * - reset to zero samples, which will trigger a proper | |
5195 | * re-initialization of the observation interval on next | |
5196 | * dispatch | |
5197 | */ | |
5198 | if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC && | |
5199 | (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us < | |
5200 | 1UL<<(BFQ_RATE_SHIFT - 10)) | |
5201 | bfq_update_rate_reset(bfqd, NULL); | |
5202 | bfqd->last_completion = now_ns; | |
aee69d78 | 5203 | |
77b7dcea PV |
5204 | /* |
5205 | * If we are waiting to discover whether the request pattern | |
5206 | * of the task associated with the queue is actually | |
5207 | * isochronous, and both requisites for this condition to hold | |
5208 | * are now satisfied, then compute soft_rt_next_start (see the | |
5209 | * comments on the function bfq_bfqq_softrt_next_start()). We | |
20cd3245 PV |
5210 | * do not compute soft_rt_next_start if bfqq is in interactive |
5211 | * weight raising (see the comments in bfq_bfqq_expire() for | |
5212 | * an explanation). We schedule this delayed update when bfqq | |
5213 | * expires, if it still has in-flight requests. | |
77b7dcea PV |
5214 | */ |
5215 | if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 && | |
20cd3245 PV |
5216 | RB_EMPTY_ROOT(&bfqq->sort_list) && |
5217 | bfqq->wr_coeff != bfqd->bfq_wr_coeff) | |
77b7dcea PV |
5218 | bfqq->soft_rt_next_start = |
5219 | bfq_bfqq_softrt_next_start(bfqd, bfqq); | |
5220 | ||
aee69d78 PV |
5221 | /* |
5222 | * If this is the in-service queue, check if it needs to be expired, | |
5223 | * or if we want to idle in case it has no pending requests. | |
5224 | */ | |
5225 | if (bfqd->in_service_queue == bfqq) { | |
4420b095 PV |
5226 | if (bfq_bfqq_must_idle(bfqq)) { |
5227 | if (bfqq->dispatched == 0) | |
5228 | bfq_arm_slice_timer(bfqd); | |
5229 | /* | |
5230 | * If we get here, we do not expire bfqq, even | |
5231 | * if bfqq was in budget timeout or had no | |
5232 | * more requests (as controlled in the next | |
5233 | * conditional instructions). The reason for | |
5234 | * not expiring bfqq is as follows. | |
5235 | * | |
5236 | * Here bfqq->dispatched > 0 holds, but | |
5237 | * bfq_bfqq_must_idle() returned true. This | |
5238 | * implies that, even if no request arrives | |
5239 | * for bfqq before bfqq->dispatched reaches 0, | |
5240 | * bfqq will, however, not be expired on the | |
5241 | * completion event that causes bfqq->dispatch | |
5242 | * to reach zero. In contrast, on this event, | |
5243 | * bfqq will start enjoying device idling | |
5244 | * (I/O-dispatch plugging). | |
5245 | * | |
5246 | * But, if we expired bfqq here, bfqq would | |
5247 | * not have the chance to enjoy device idling | |
5248 | * when bfqq->dispatched finally reaches | |
5249 | * zero. This would expose bfqq to violation | |
5250 | * of its reserved service guarantees. | |
5251 | */ | |
aee69d78 PV |
5252 | return; |
5253 | } else if (bfq_may_expire_for_budg_timeout(bfqq)) | |
5254 | bfq_bfqq_expire(bfqd, bfqq, false, | |
5255 | BFQQE_BUDGET_TIMEOUT); | |
5256 | else if (RB_EMPTY_ROOT(&bfqq->sort_list) && | |
5257 | (bfqq->dispatched == 0 || | |
277a4a9b | 5258 | !bfq_better_to_idle(bfqq))) |
aee69d78 PV |
5259 | bfq_bfqq_expire(bfqd, bfqq, false, |
5260 | BFQQE_NO_MORE_REQUESTS); | |
5261 | } | |
3f7cb4f4 HT |
5262 | |
5263 | if (!bfqd->rq_in_driver) | |
5264 | bfq_schedule_dispatch(bfqd); | |
aee69d78 PV |
5265 | } |
5266 | ||
a7877390 | 5267 | static void bfq_finish_requeue_request_body(struct bfq_queue *bfqq) |
aee69d78 PV |
5268 | { |
5269 | bfqq->allocated--; | |
5270 | ||
5271 | bfq_put_queue(bfqq); | |
5272 | } | |
5273 | ||
2341d662 PV |
5274 | /* |
5275 | * The processes associated with bfqq may happen to generate their | |
5276 | * cumulative I/O at a lower rate than the rate at which the device | |
5277 | * could serve the same I/O. This is rather probable, e.g., if only | |
5278 | * one process is associated with bfqq and the device is an SSD. It | |
5279 | * results in bfqq becoming often empty while in service. In this | |
5280 | * respect, if BFQ is allowed to switch to another queue when bfqq | |
5281 | * remains empty, then the device goes on being fed with I/O requests, | |
5282 | * and the throughput is not affected. In contrast, if BFQ is not | |
5283 | * allowed to switch to another queue---because bfqq is sync and | |
5284 | * I/O-dispatch needs to be plugged while bfqq is temporarily | |
5285 | * empty---then, during the service of bfqq, there will be frequent | |
5286 | * "service holes", i.e., time intervals during which bfqq gets empty | |
5287 | * and the device can only consume the I/O already queued in its | |
5288 | * hardware queues. During service holes, the device may even get to | |
5289 | * remaining idle. In the end, during the service of bfqq, the device | |
5290 | * is driven at a lower speed than the one it can reach with the kind | |
5291 | * of I/O flowing through bfqq. | |
5292 | * | |
5293 | * To counter this loss of throughput, BFQ implements a "request | |
5294 | * injection mechanism", which tries to fill the above service holes | |
5295 | * with I/O requests taken from other queues. The hard part in this | |
5296 | * mechanism is finding the right amount of I/O to inject, so as to | |
5297 | * both boost throughput and not break bfqq's bandwidth and latency | |
5298 | * guarantees. In this respect, the mechanism maintains a per-queue | |
5299 | * inject limit, computed as below. While bfqq is empty, the injection | |
5300 | * mechanism dispatches extra I/O requests only until the total number | |
5301 | * of I/O requests in flight---i.e., already dispatched but not yet | |
5302 | * completed---remains lower than this limit. | |
5303 | * | |
5304 | * A first definition comes in handy to introduce the algorithm by | |
5305 | * which the inject limit is computed. We define as first request for | |
5306 | * bfqq, an I/O request for bfqq that arrives while bfqq is in | |
5307 | * service, and causes bfqq to switch from empty to non-empty. The | |
5308 | * algorithm updates the limit as a function of the effect of | |
5309 | * injection on the service times of only the first requests of | |
5310 | * bfqq. The reason for this restriction is that these are the | |
5311 | * requests whose service time is affected most, because they are the | |
5312 | * first to arrive after injection possibly occurred. | |
5313 | * | |
5314 | * To evaluate the effect of injection, the algorithm measures the | |
5315 | * "total service time" of first requests. We define as total service | |
5316 | * time of an I/O request, the time that elapses since when the | |
5317 | * request is enqueued into bfqq, to when it is completed. This | |
5318 | * quantity allows the whole effect of injection to be measured. It is | |
5319 | * easy to see why. Suppose that some requests of other queues are | |
5320 | * actually injected while bfqq is empty, and that a new request R | |
5321 | * then arrives for bfqq. If the device does start to serve all or | |
5322 | * part of the injected requests during the service hole, then, | |
5323 | * because of this extra service, it may delay the next invocation of | |
5324 | * the dispatch hook of BFQ. Then, even after R gets eventually | |
5325 | * dispatched, the device may delay the actual service of R if it is | |
5326 | * still busy serving the extra requests, or if it decides to serve, | |
5327 | * before R, some extra request still present in its queues. As a | |
5328 | * conclusion, the cumulative extra delay caused by injection can be | |
5329 | * easily evaluated by just comparing the total service time of first | |
5330 | * requests with and without injection. | |
5331 | * | |
5332 | * The limit-update algorithm works as follows. On the arrival of a | |
5333 | * first request of bfqq, the algorithm measures the total time of the | |
5334 | * request only if one of the three cases below holds, and, for each | |
5335 | * case, it updates the limit as described below: | |
5336 | * | |
5337 | * (1) If there is no in-flight request. This gives a baseline for the | |
5338 | * total service time of the requests of bfqq. If the baseline has | |
5339 | * not been computed yet, then, after computing it, the limit is | |
5340 | * set to 1, to start boosting throughput, and to prepare the | |
5341 | * ground for the next case. If the baseline has already been | |
5342 | * computed, then it is updated, in case it results to be lower | |
5343 | * than the previous value. | |
5344 | * | |
5345 | * (2) If the limit is higher than 0 and there are in-flight | |
5346 | * requests. By comparing the total service time in this case with | |
5347 | * the above baseline, it is possible to know at which extent the | |
5348 | * current value of the limit is inflating the total service | |
5349 | * time. If the inflation is below a certain threshold, then bfqq | |
5350 | * is assumed to be suffering from no perceivable loss of its | |
5351 | * service guarantees, and the limit is even tentatively | |
5352 | * increased. If the inflation is above the threshold, then the | |
5353 | * limit is decreased. Due to the lack of any hysteresis, this | |
5354 | * logic makes the limit oscillate even in steady workload | |
5355 | * conditions. Yet we opted for it, because it is fast in reaching | |
5356 | * the best value for the limit, as a function of the current I/O | |
5357 | * workload. To reduce oscillations, this step is disabled for a | |
5358 | * short time interval after the limit happens to be decreased. | |
5359 | * | |
5360 | * (3) Periodically, after resetting the limit, to make sure that the | |
5361 | * limit eventually drops in case the workload changes. This is | |
5362 | * needed because, after the limit has gone safely up for a | |
5363 | * certain workload, it is impossible to guess whether the | |
5364 | * baseline total service time may have changed, without measuring | |
5365 | * it again without injection. A more effective version of this | |
5366 | * step might be to just sample the baseline, by interrupting | |
5367 | * injection only once, and then to reset/lower the limit only if | |
5368 | * the total service time with the current limit does happen to be | |
5369 | * too large. | |
5370 | * | |
5371 | * More details on each step are provided in the comments on the | |
5372 | * pieces of code that implement these steps: the branch handling the | |
5373 | * transition from empty to non empty in bfq_add_request(), the branch | |
5374 | * handling injection in bfq_select_queue(), and the function | |
5375 | * bfq_choose_bfqq_for_injection(). These comments also explain some | |
5376 | * exceptions, made by the injection mechanism in some special cases. | |
5377 | */ | |
5378 | static void bfq_update_inject_limit(struct bfq_data *bfqd, | |
5379 | struct bfq_queue *bfqq) | |
5380 | { | |
5381 | u64 tot_time_ns = ktime_get_ns() - bfqd->last_empty_occupied_ns; | |
5382 | unsigned int old_limit = bfqq->inject_limit; | |
5383 | ||
5384 | if (bfqq->last_serv_time_ns > 0) { | |
5385 | u64 threshold = (bfqq->last_serv_time_ns * 3)>>1; | |
5386 | ||
5387 | if (tot_time_ns >= threshold && old_limit > 0) { | |
5388 | bfqq->inject_limit--; | |
5389 | bfqq->decrease_time_jif = jiffies; | |
5390 | } else if (tot_time_ns < threshold && | |
5391 | old_limit < bfqd->max_rq_in_driver<<1) | |
5392 | bfqq->inject_limit++; | |
5393 | } | |
5394 | ||
5395 | /* | |
5396 | * Either we still have to compute the base value for the | |
5397 | * total service time, and there seem to be the right | |
5398 | * conditions to do it, or we can lower the last base value | |
5399 | * computed. | |
5400 | */ | |
5401 | if ((bfqq->last_serv_time_ns == 0 && bfqd->rq_in_driver == 0) || | |
5402 | tot_time_ns < bfqq->last_serv_time_ns) { | |
5403 | bfqq->last_serv_time_ns = tot_time_ns; | |
5404 | /* | |
5405 | * Now we certainly have a base value: make sure we | |
5406 | * start trying injection. | |
5407 | */ | |
5408 | bfqq->inject_limit = max_t(unsigned int, 1, old_limit); | |
5409 | } | |
5410 | ||
5411 | /* update complete, not waiting for any request completion any longer */ | |
5412 | bfqd->waited_rq = NULL; | |
5413 | } | |
5414 | ||
a7877390 PV |
5415 | /* |
5416 | * Handle either a requeue or a finish for rq. The things to do are | |
5417 | * the same in both cases: all references to rq are to be dropped. In | |
5418 | * particular, rq is considered completed from the point of view of | |
5419 | * the scheduler. | |
5420 | */ | |
5421 | static void bfq_finish_requeue_request(struct request *rq) | |
aee69d78 | 5422 | { |
a7877390 | 5423 | struct bfq_queue *bfqq = RQ_BFQQ(rq); |
5bbf4e5a CH |
5424 | struct bfq_data *bfqd; |
5425 | ||
a7877390 PV |
5426 | /* |
5427 | * Requeue and finish hooks are invoked in blk-mq without | |
5428 | * checking whether the involved request is actually still | |
5429 | * referenced in the scheduler. To handle this fact, the | |
5430 | * following two checks make this function exit in case of | |
5431 | * spurious invocations, for which there is nothing to do. | |
5432 | * | |
5433 | * First, check whether rq has nothing to do with an elevator. | |
5434 | */ | |
5435 | if (unlikely(!(rq->rq_flags & RQF_ELVPRIV))) | |
5436 | return; | |
5437 | ||
5438 | /* | |
5439 | * rq either is not associated with any icq, or is an already | |
5440 | * requeued request that has not (yet) been re-inserted into | |
5441 | * a bfq_queue. | |
5442 | */ | |
5443 | if (!rq->elv.icq || !bfqq) | |
5bbf4e5a CH |
5444 | return; |
5445 | ||
5bbf4e5a | 5446 | bfqd = bfqq->bfqd; |
aee69d78 | 5447 | |
e21b7a0b AA |
5448 | if (rq->rq_flags & RQF_STARTED) |
5449 | bfqg_stats_update_completion(bfqq_group(bfqq), | |
522a7775 OS |
5450 | rq->start_time_ns, |
5451 | rq->io_start_time_ns, | |
e21b7a0b | 5452 | rq->cmd_flags); |
aee69d78 PV |
5453 | |
5454 | if (likely(rq->rq_flags & RQF_STARTED)) { | |
5455 | unsigned long flags; | |
5456 | ||
5457 | spin_lock_irqsave(&bfqd->lock, flags); | |
5458 | ||
2341d662 PV |
5459 | if (rq == bfqd->waited_rq) |
5460 | bfq_update_inject_limit(bfqd, bfqq); | |
5461 | ||
aee69d78 | 5462 | bfq_completed_request(bfqq, bfqd); |
a7877390 | 5463 | bfq_finish_requeue_request_body(bfqq); |
aee69d78 | 5464 | |
6fa3e8d3 | 5465 | spin_unlock_irqrestore(&bfqd->lock, flags); |
aee69d78 PV |
5466 | } else { |
5467 | /* | |
5468 | * Request rq may be still/already in the scheduler, | |
a7877390 PV |
5469 | * in which case we need to remove it (this should |
5470 | * never happen in case of requeue). And we cannot | |
aee69d78 PV |
5471 | * defer such a check and removal, to avoid |
5472 | * inconsistencies in the time interval from the end | |
5473 | * of this function to the start of the deferred work. | |
5474 | * This situation seems to occur only in process | |
5475 | * context, as a consequence of a merge. In the | |
5476 | * current version of the code, this implies that the | |
5477 | * lock is held. | |
5478 | */ | |
5479 | ||
614822f8 | 5480 | if (!RB_EMPTY_NODE(&rq->rb_node)) { |
7b9e9361 | 5481 | bfq_remove_request(rq->q, rq); |
614822f8 LM |
5482 | bfqg_stats_update_io_remove(bfqq_group(bfqq), |
5483 | rq->cmd_flags); | |
5484 | } | |
a7877390 | 5485 | bfq_finish_requeue_request_body(bfqq); |
aee69d78 PV |
5486 | } |
5487 | ||
a7877390 PV |
5488 | /* |
5489 | * Reset private fields. In case of a requeue, this allows | |
5490 | * this function to correctly do nothing if it is spuriously | |
5491 | * invoked again on this same request (see the check at the | |
5492 | * beginning of the function). Probably, a better general | |
5493 | * design would be to prevent blk-mq from invoking the requeue | |
5494 | * or finish hooks of an elevator, for a request that is not | |
5495 | * referred by that elevator. | |
5496 | * | |
5497 | * Resetting the following fields would break the | |
5498 | * request-insertion logic if rq is re-inserted into a bfq | |
5499 | * internal queue, without a re-preparation. Here we assume | |
5500 | * that re-insertions of requeued requests, without | |
5501 | * re-preparation, can happen only for pass_through or at_head | |
5502 | * requests (which are not re-inserted into bfq internal | |
5503 | * queues). | |
5504 | */ | |
aee69d78 PV |
5505 | rq->elv.priv[0] = NULL; |
5506 | rq->elv.priv[1] = NULL; | |
5507 | } | |
5508 | ||
36eca894 AA |
5509 | /* |
5510 | * Returns NULL if a new bfqq should be allocated, or the old bfqq if this | |
5511 | * was the last process referring to that bfqq. | |
5512 | */ | |
5513 | static struct bfq_queue * | |
5514 | bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq) | |
5515 | { | |
5516 | bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue"); | |
5517 | ||
5518 | if (bfqq_process_refs(bfqq) == 1) { | |
5519 | bfqq->pid = current->pid; | |
5520 | bfq_clear_bfqq_coop(bfqq); | |
5521 | bfq_clear_bfqq_split_coop(bfqq); | |
5522 | return bfqq; | |
5523 | } | |
5524 | ||
5525 | bic_set_bfqq(bic, NULL, 1); | |
5526 | ||
5527 | bfq_put_cooperator(bfqq); | |
5528 | ||
5529 | bfq_put_queue(bfqq); | |
5530 | return NULL; | |
5531 | } | |
5532 | ||
5533 | static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd, | |
5534 | struct bfq_io_cq *bic, | |
5535 | struct bio *bio, | |
5536 | bool split, bool is_sync, | |
5537 | bool *new_queue) | |
5538 | { | |
5539 | struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync); | |
5540 | ||
5541 | if (likely(bfqq && bfqq != &bfqd->oom_bfqq)) | |
5542 | return bfqq; | |
5543 | ||
5544 | if (new_queue) | |
5545 | *new_queue = true; | |
5546 | ||
5547 | if (bfqq) | |
5548 | bfq_put_queue(bfqq); | |
5549 | bfqq = bfq_get_queue(bfqd, bio, is_sync, bic); | |
5550 | ||
5551 | bic_set_bfqq(bic, bfqq, is_sync); | |
e1b2324d AA |
5552 | if (split && is_sync) { |
5553 | if ((bic->was_in_burst_list && bfqd->large_burst) || | |
5554 | bic->saved_in_large_burst) | |
5555 | bfq_mark_bfqq_in_large_burst(bfqq); | |
5556 | else { | |
5557 | bfq_clear_bfqq_in_large_burst(bfqq); | |
5558 | if (bic->was_in_burst_list) | |
99fead8d PV |
5559 | /* |
5560 | * If bfqq was in the current | |
5561 | * burst list before being | |
5562 | * merged, then we have to add | |
5563 | * it back. And we do not need | |
5564 | * to increase burst_size, as | |
5565 | * we did not decrement | |
5566 | * burst_size when we removed | |
5567 | * bfqq from the burst list as | |
5568 | * a consequence of a merge | |
5569 | * (see comments in | |
5570 | * bfq_put_queue). In this | |
5571 | * respect, it would be rather | |
5572 | * costly to know whether the | |
5573 | * current burst list is still | |
5574 | * the same burst list from | |
5575 | * which bfqq was removed on | |
5576 | * the merge. To avoid this | |
5577 | * cost, if bfqq was in a | |
5578 | * burst list, then we add | |
5579 | * bfqq to the current burst | |
5580 | * list without any further | |
5581 | * check. This can cause | |
5582 | * inappropriate insertions, | |
5583 | * but rarely enough to not | |
5584 | * harm the detection of large | |
5585 | * bursts significantly. | |
5586 | */ | |
e1b2324d AA |
5587 | hlist_add_head(&bfqq->burst_list_node, |
5588 | &bfqd->burst_list); | |
5589 | } | |
36eca894 | 5590 | bfqq->split_time = jiffies; |
e1b2324d | 5591 | } |
36eca894 AA |
5592 | |
5593 | return bfqq; | |
5594 | } | |
5595 | ||
aee69d78 | 5596 | /* |
18e5a57d PV |
5597 | * Only reset private fields. The actual request preparation will be |
5598 | * performed by bfq_init_rq, when rq is either inserted or merged. See | |
5599 | * comments on bfq_init_rq for the reason behind this delayed | |
5600 | * preparation. | |
aee69d78 | 5601 | */ |
5bbf4e5a | 5602 | static void bfq_prepare_request(struct request *rq, struct bio *bio) |
18e5a57d PV |
5603 | { |
5604 | /* | |
5605 | * Regardless of whether we have an icq attached, we have to | |
5606 | * clear the scheduler pointers, as they might point to | |
5607 | * previously allocated bic/bfqq structs. | |
5608 | */ | |
5609 | rq->elv.priv[0] = rq->elv.priv[1] = NULL; | |
5610 | } | |
5611 | ||
5612 | /* | |
5613 | * If needed, init rq, allocate bfq data structures associated with | |
5614 | * rq, and increment reference counters in the destination bfq_queue | |
5615 | * for rq. Return the destination bfq_queue for rq, or NULL is rq is | |
5616 | * not associated with any bfq_queue. | |
5617 | * | |
5618 | * This function is invoked by the functions that perform rq insertion | |
5619 | * or merging. One may have expected the above preparation operations | |
5620 | * to be performed in bfq_prepare_request, and not delayed to when rq | |
5621 | * is inserted or merged. The rationale behind this delayed | |
5622 | * preparation is that, after the prepare_request hook is invoked for | |
5623 | * rq, rq may still be transformed into a request with no icq, i.e., a | |
5624 | * request not associated with any queue. No bfq hook is invoked to | |
636b8fe8 | 5625 | * signal this transformation. As a consequence, should these |
18e5a57d PV |
5626 | * preparation operations be performed when the prepare_request hook |
5627 | * is invoked, and should rq be transformed one moment later, bfq | |
5628 | * would end up in an inconsistent state, because it would have | |
5629 | * incremented some queue counters for an rq destined to | |
5630 | * transformation, without any chance to correctly lower these | |
5631 | * counters back. In contrast, no transformation can still happen for | |
5632 | * rq after rq has been inserted or merged. So, it is safe to execute | |
5633 | * these preparation operations when rq is finally inserted or merged. | |
5634 | */ | |
5635 | static struct bfq_queue *bfq_init_rq(struct request *rq) | |
aee69d78 | 5636 | { |
5bbf4e5a | 5637 | struct request_queue *q = rq->q; |
18e5a57d | 5638 | struct bio *bio = rq->bio; |
aee69d78 | 5639 | struct bfq_data *bfqd = q->elevator->elevator_data; |
9f210738 | 5640 | struct bfq_io_cq *bic; |
aee69d78 PV |
5641 | const int is_sync = rq_is_sync(rq); |
5642 | struct bfq_queue *bfqq; | |
36eca894 | 5643 | bool new_queue = false; |
13c931bd | 5644 | bool bfqq_already_existing = false, split = false; |
aee69d78 | 5645 | |
18e5a57d PV |
5646 | if (unlikely(!rq->elv.icq)) |
5647 | return NULL; | |
5648 | ||
72961c4e | 5649 | /* |
18e5a57d PV |
5650 | * Assuming that elv.priv[1] is set only if everything is set |
5651 | * for this rq. This holds true, because this function is | |
5652 | * invoked only for insertion or merging, and, after such | |
5653 | * events, a request cannot be manipulated any longer before | |
5654 | * being removed from bfq. | |
72961c4e | 5655 | */ |
18e5a57d PV |
5656 | if (rq->elv.priv[1]) |
5657 | return rq->elv.priv[1]; | |
72961c4e | 5658 | |
9f210738 | 5659 | bic = icq_to_bic(rq->elv.icq); |
aee69d78 | 5660 | |
8c9ff1ad CIK |
5661 | bfq_check_ioprio_change(bic, bio); |
5662 | ||
e21b7a0b AA |
5663 | bfq_bic_update_cgroup(bic, bio); |
5664 | ||
36eca894 AA |
5665 | bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync, |
5666 | &new_queue); | |
5667 | ||
5668 | if (likely(!new_queue)) { | |
5669 | /* If the queue was seeky for too long, break it apart. */ | |
5670 | if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) { | |
5671 | bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq"); | |
e1b2324d AA |
5672 | |
5673 | /* Update bic before losing reference to bfqq */ | |
5674 | if (bfq_bfqq_in_large_burst(bfqq)) | |
5675 | bic->saved_in_large_burst = true; | |
5676 | ||
36eca894 | 5677 | bfqq = bfq_split_bfqq(bic, bfqq); |
6fa3e8d3 | 5678 | split = true; |
36eca894 AA |
5679 | |
5680 | if (!bfqq) | |
5681 | bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, | |
5682 | true, is_sync, | |
5683 | NULL); | |
13c931bd PV |
5684 | else |
5685 | bfqq_already_existing = true; | |
36eca894 | 5686 | } |
aee69d78 PV |
5687 | } |
5688 | ||
5689 | bfqq->allocated++; | |
5690 | bfqq->ref++; | |
5691 | bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d", | |
5692 | rq, bfqq, bfqq->ref); | |
5693 | ||
5694 | rq->elv.priv[0] = bic; | |
5695 | rq->elv.priv[1] = bfqq; | |
5696 | ||
36eca894 AA |
5697 | /* |
5698 | * If a bfq_queue has only one process reference, it is owned | |
5699 | * by only this bic: we can then set bfqq->bic = bic. in | |
5700 | * addition, if the queue has also just been split, we have to | |
5701 | * resume its state. | |
5702 | */ | |
5703 | if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) { | |
5704 | bfqq->bic = bic; | |
6fa3e8d3 | 5705 | if (split) { |
36eca894 AA |
5706 | /* |
5707 | * The queue has just been split from a shared | |
5708 | * queue: restore the idle window and the | |
5709 | * possible weight raising period. | |
5710 | */ | |
13c931bd PV |
5711 | bfq_bfqq_resume_state(bfqq, bfqd, bic, |
5712 | bfqq_already_existing); | |
36eca894 AA |
5713 | } |
5714 | } | |
5715 | ||
84a74689 PV |
5716 | /* |
5717 | * Consider bfqq as possibly belonging to a burst of newly | |
5718 | * created queues only if: | |
5719 | * 1) A burst is actually happening (bfqd->burst_size > 0) | |
5720 | * or | |
5721 | * 2) There is no other active queue. In fact, if, in | |
5722 | * contrast, there are active queues not belonging to the | |
5723 | * possible burst bfqq may belong to, then there is no gain | |
5724 | * in considering bfqq as belonging to a burst, and | |
5725 | * therefore in not weight-raising bfqq. See comments on | |
5726 | * bfq_handle_burst(). | |
5727 | * | |
5728 | * This filtering also helps eliminating false positives, | |
5729 | * occurring when bfqq does not belong to an actual large | |
5730 | * burst, but some background task (e.g., a service) happens | |
5731 | * to trigger the creation of new queues very close to when | |
5732 | * bfqq and its possible companion queues are created. See | |
5733 | * comments on bfq_handle_burst() for further details also on | |
5734 | * this issue. | |
5735 | */ | |
5736 | if (unlikely(bfq_bfqq_just_created(bfqq) && | |
5737 | (bfqd->burst_size > 0 || | |
5738 | bfq_tot_busy_queues(bfqd) == 0))) | |
e1b2324d AA |
5739 | bfq_handle_burst(bfqd, bfqq); |
5740 | ||
18e5a57d | 5741 | return bfqq; |
aee69d78 PV |
5742 | } |
5743 | ||
5744 | static void bfq_idle_slice_timer_body(struct bfq_queue *bfqq) | |
5745 | { | |
5746 | struct bfq_data *bfqd = bfqq->bfqd; | |
5747 | enum bfqq_expiration reason; | |
5748 | unsigned long flags; | |
5749 | ||
5750 | spin_lock_irqsave(&bfqd->lock, flags); | |
5751 | bfq_clear_bfqq_wait_request(bfqq); | |
5752 | ||
5753 | if (bfqq != bfqd->in_service_queue) { | |
5754 | spin_unlock_irqrestore(&bfqd->lock, flags); | |
5755 | return; | |
5756 | } | |
5757 | ||
5758 | if (bfq_bfqq_budget_timeout(bfqq)) | |
5759 | /* | |
5760 | * Also here the queue can be safely expired | |
5761 | * for budget timeout without wasting | |
5762 | * guarantees | |
5763 | */ | |
5764 | reason = BFQQE_BUDGET_TIMEOUT; | |
5765 | else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0) | |
5766 | /* | |
5767 | * The queue may not be empty upon timer expiration, | |
5768 | * because we may not disable the timer when the | |
5769 | * first request of the in-service queue arrives | |
5770 | * during disk idling. | |
5771 | */ | |
5772 | reason = BFQQE_TOO_IDLE; | |
5773 | else | |
5774 | goto schedule_dispatch; | |
5775 | ||
5776 | bfq_bfqq_expire(bfqd, bfqq, true, reason); | |
5777 | ||
5778 | schedule_dispatch: | |
6fa3e8d3 | 5779 | spin_unlock_irqrestore(&bfqd->lock, flags); |
aee69d78 PV |
5780 | bfq_schedule_dispatch(bfqd); |
5781 | } | |
5782 | ||
5783 | /* | |
5784 | * Handler of the expiration of the timer running if the in-service queue | |
5785 | * is idling inside its time slice. | |
5786 | */ | |
5787 | static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer) | |
5788 | { | |
5789 | struct bfq_data *bfqd = container_of(timer, struct bfq_data, | |
5790 | idle_slice_timer); | |
5791 | struct bfq_queue *bfqq = bfqd->in_service_queue; | |
5792 | ||
5793 | /* | |
5794 | * Theoretical race here: the in-service queue can be NULL or | |
5795 | * different from the queue that was idling if a new request | |
5796 | * arrives for the current queue and there is a full dispatch | |
5797 | * cycle that changes the in-service queue. This can hardly | |
5798 | * happen, but in the worst case we just expire a queue too | |
5799 | * early. | |
5800 | */ | |
5801 | if (bfqq) | |
5802 | bfq_idle_slice_timer_body(bfqq); | |
5803 | ||
5804 | return HRTIMER_NORESTART; | |
5805 | } | |
5806 | ||
5807 | static void __bfq_put_async_bfqq(struct bfq_data *bfqd, | |
5808 | struct bfq_queue **bfqq_ptr) | |
5809 | { | |
5810 | struct bfq_queue *bfqq = *bfqq_ptr; | |
5811 | ||
5812 | bfq_log(bfqd, "put_async_bfqq: %p", bfqq); | |
5813 | if (bfqq) { | |
e21b7a0b AA |
5814 | bfq_bfqq_move(bfqd, bfqq, bfqd->root_group); |
5815 | ||
aee69d78 PV |
5816 | bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d", |
5817 | bfqq, bfqq->ref); | |
5818 | bfq_put_queue(bfqq); | |
5819 | *bfqq_ptr = NULL; | |
5820 | } | |
5821 | } | |
5822 | ||
5823 | /* | |
e21b7a0b AA |
5824 | * Release all the bfqg references to its async queues. If we are |
5825 | * deallocating the group these queues may still contain requests, so | |
5826 | * we reparent them to the root cgroup (i.e., the only one that will | |
5827 | * exist for sure until all the requests on a device are gone). | |
aee69d78 | 5828 | */ |
ea25da48 | 5829 | void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg) |
aee69d78 PV |
5830 | { |
5831 | int i, j; | |
5832 | ||
5833 | for (i = 0; i < 2; i++) | |
5834 | for (j = 0; j < IOPRIO_BE_NR; j++) | |
e21b7a0b | 5835 | __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]); |
aee69d78 | 5836 | |
e21b7a0b | 5837 | __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq); |
aee69d78 PV |
5838 | } |
5839 | ||
f0635b8a JA |
5840 | /* |
5841 | * See the comments on bfq_limit_depth for the purpose of | |
483b7bf2 | 5842 | * the depths set in the function. Return minimum shallow depth we'll use. |
f0635b8a | 5843 | */ |
483b7bf2 JA |
5844 | static unsigned int bfq_update_depths(struct bfq_data *bfqd, |
5845 | struct sbitmap_queue *bt) | |
f0635b8a | 5846 | { |
483b7bf2 JA |
5847 | unsigned int i, j, min_shallow = UINT_MAX; |
5848 | ||
f0635b8a JA |
5849 | /* |
5850 | * In-word depths if no bfq_queue is being weight-raised: | |
5851 | * leaving 25% of tags only for sync reads. | |
5852 | * | |
5853 | * In next formulas, right-shift the value | |
bd7d4ef6 JA |
5854 | * (1U<<bt->sb.shift), instead of computing directly |
5855 | * (1U<<(bt->sb.shift - something)), to be robust against | |
5856 | * any possible value of bt->sb.shift, without having to | |
f0635b8a JA |
5857 | * limit 'something'. |
5858 | */ | |
5859 | /* no more than 50% of tags for async I/O */ | |
bd7d4ef6 | 5860 | bfqd->word_depths[0][0] = max((1U << bt->sb.shift) >> 1, 1U); |
f0635b8a JA |
5861 | /* |
5862 | * no more than 75% of tags for sync writes (25% extra tags | |
5863 | * w.r.t. async I/O, to prevent async I/O from starving sync | |
5864 | * writes) | |
5865 | */ | |
bd7d4ef6 | 5866 | bfqd->word_depths[0][1] = max(((1U << bt->sb.shift) * 3) >> 2, 1U); |
f0635b8a JA |
5867 | |
5868 | /* | |
5869 | * In-word depths in case some bfq_queue is being weight- | |
5870 | * raised: leaving ~63% of tags for sync reads. This is the | |
5871 | * highest percentage for which, in our tests, application | |
5872 | * start-up times didn't suffer from any regression due to tag | |
5873 | * shortage. | |
5874 | */ | |
5875 | /* no more than ~18% of tags for async I/O */ | |
bd7d4ef6 | 5876 | bfqd->word_depths[1][0] = max(((1U << bt->sb.shift) * 3) >> 4, 1U); |
f0635b8a | 5877 | /* no more than ~37% of tags for sync writes (~20% extra tags) */ |
bd7d4ef6 | 5878 | bfqd->word_depths[1][1] = max(((1U << bt->sb.shift) * 6) >> 4, 1U); |
483b7bf2 JA |
5879 | |
5880 | for (i = 0; i < 2; i++) | |
5881 | for (j = 0; j < 2; j++) | |
5882 | min_shallow = min(min_shallow, bfqd->word_depths[i][j]); | |
5883 | ||
5884 | return min_shallow; | |
f0635b8a JA |
5885 | } |
5886 | ||
77f1e0a5 | 5887 | static void bfq_depth_updated(struct blk_mq_hw_ctx *hctx) |
f0635b8a JA |
5888 | { |
5889 | struct bfq_data *bfqd = hctx->queue->elevator->elevator_data; | |
5890 | struct blk_mq_tags *tags = hctx->sched_tags; | |
483b7bf2 | 5891 | unsigned int min_shallow; |
f0635b8a | 5892 | |
483b7bf2 JA |
5893 | min_shallow = bfq_update_depths(bfqd, &tags->bitmap_tags); |
5894 | sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, min_shallow); | |
77f1e0a5 JA |
5895 | } |
5896 | ||
5897 | static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index) | |
5898 | { | |
5899 | bfq_depth_updated(hctx); | |
f0635b8a JA |
5900 | return 0; |
5901 | } | |
5902 | ||
aee69d78 PV |
5903 | static void bfq_exit_queue(struct elevator_queue *e) |
5904 | { | |
5905 | struct bfq_data *bfqd = e->elevator_data; | |
5906 | struct bfq_queue *bfqq, *n; | |
5907 | ||
5908 | hrtimer_cancel(&bfqd->idle_slice_timer); | |
5909 | ||
5910 | spin_lock_irq(&bfqd->lock); | |
5911 | list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list) | |
e21b7a0b | 5912 | bfq_deactivate_bfqq(bfqd, bfqq, false, false); |
aee69d78 PV |
5913 | spin_unlock_irq(&bfqd->lock); |
5914 | ||
5915 | hrtimer_cancel(&bfqd->idle_slice_timer); | |
5916 | ||
8abef10b | 5917 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
0d52af59 PV |
5918 | /* release oom-queue reference to root group */ |
5919 | bfqg_and_blkg_put(bfqd->root_group); | |
5920 | ||
e21b7a0b AA |
5921 | blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq); |
5922 | #else | |
5923 | spin_lock_irq(&bfqd->lock); | |
5924 | bfq_put_async_queues(bfqd, bfqd->root_group); | |
5925 | kfree(bfqd->root_group); | |
5926 | spin_unlock_irq(&bfqd->lock); | |
5927 | #endif | |
5928 | ||
aee69d78 PV |
5929 | kfree(bfqd); |
5930 | } | |
5931 | ||
e21b7a0b AA |
5932 | static void bfq_init_root_group(struct bfq_group *root_group, |
5933 | struct bfq_data *bfqd) | |
5934 | { | |
5935 | int i; | |
5936 | ||
5937 | #ifdef CONFIG_BFQ_GROUP_IOSCHED | |
5938 | root_group->entity.parent = NULL; | |
5939 | root_group->my_entity = NULL; | |
5940 | root_group->bfqd = bfqd; | |
5941 | #endif | |
36eca894 | 5942 | root_group->rq_pos_tree = RB_ROOT; |
e21b7a0b AA |
5943 | for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) |
5944 | root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT; | |
5945 | root_group->sched_data.bfq_class_idle_last_service = jiffies; | |
5946 | } | |
5947 | ||
aee69d78 PV |
5948 | static int bfq_init_queue(struct request_queue *q, struct elevator_type *e) |
5949 | { | |
5950 | struct bfq_data *bfqd; | |
5951 | struct elevator_queue *eq; | |
aee69d78 PV |
5952 | |
5953 | eq = elevator_alloc(q, e); | |
5954 | if (!eq) | |
5955 | return -ENOMEM; | |
5956 | ||
5957 | bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node); | |
5958 | if (!bfqd) { | |
5959 | kobject_put(&eq->kobj); | |
5960 | return -ENOMEM; | |
5961 | } | |
5962 | eq->elevator_data = bfqd; | |
5963 | ||
0d945c1f | 5964 | spin_lock_irq(&q->queue_lock); |
e21b7a0b | 5965 | q->elevator = eq; |
0d945c1f | 5966 | spin_unlock_irq(&q->queue_lock); |
e21b7a0b | 5967 | |
aee69d78 PV |
5968 | /* |
5969 | * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues. | |
5970 | * Grab a permanent reference to it, so that the normal code flow | |
5971 | * will not attempt to free it. | |
5972 | */ | |
5973 | bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0); | |
5974 | bfqd->oom_bfqq.ref++; | |
5975 | bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO; | |
5976 | bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE; | |
5977 | bfqd->oom_bfqq.entity.new_weight = | |
5978 | bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio); | |
e1b2324d AA |
5979 | |
5980 | /* oom_bfqq does not participate to bursts */ | |
5981 | bfq_clear_bfqq_just_created(&bfqd->oom_bfqq); | |
5982 | ||
aee69d78 PV |
5983 | /* |
5984 | * Trigger weight initialization, according to ioprio, at the | |
5985 | * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio | |
5986 | * class won't be changed any more. | |
5987 | */ | |
5988 | bfqd->oom_bfqq.entity.prio_changed = 1; | |
5989 | ||
5990 | bfqd->queue = q; | |
5991 | ||
e21b7a0b | 5992 | INIT_LIST_HEAD(&bfqd->dispatch); |
aee69d78 PV |
5993 | |
5994 | hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC, | |
5995 | HRTIMER_MODE_REL); | |
5996 | bfqd->idle_slice_timer.function = bfq_idle_slice_timer; | |
5997 | ||
fb53ac6c | 5998 | bfqd->queue_weights_tree = RB_ROOT_CACHED; |
ba7aeae5 | 5999 | bfqd->num_groups_with_pending_reqs = 0; |
1de0c4cd | 6000 | |
aee69d78 PV |
6001 | INIT_LIST_HEAD(&bfqd->active_list); |
6002 | INIT_LIST_HEAD(&bfqd->idle_list); | |
e1b2324d | 6003 | INIT_HLIST_HEAD(&bfqd->burst_list); |
aee69d78 PV |
6004 | |
6005 | bfqd->hw_tag = -1; | |
8cacc5ab | 6006 | bfqd->nonrot_with_queueing = blk_queue_nonrot(bfqd->queue); |
aee69d78 PV |
6007 | |
6008 | bfqd->bfq_max_budget = bfq_default_max_budget; | |
6009 | ||
6010 | bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0]; | |
6011 | bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1]; | |
6012 | bfqd->bfq_back_max = bfq_back_max; | |
6013 | bfqd->bfq_back_penalty = bfq_back_penalty; | |
6014 | bfqd->bfq_slice_idle = bfq_slice_idle; | |
aee69d78 PV |
6015 | bfqd->bfq_timeout = bfq_timeout; |
6016 | ||
6017 | bfqd->bfq_requests_within_timer = 120; | |
6018 | ||
e1b2324d AA |
6019 | bfqd->bfq_large_burst_thresh = 8; |
6020 | bfqd->bfq_burst_interval = msecs_to_jiffies(180); | |
6021 | ||
44e44a1b PV |
6022 | bfqd->low_latency = true; |
6023 | ||
6024 | /* | |
6025 | * Trade-off between responsiveness and fairness. | |
6026 | */ | |
6027 | bfqd->bfq_wr_coeff = 30; | |
77b7dcea | 6028 | bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300); |
44e44a1b PV |
6029 | bfqd->bfq_wr_max_time = 0; |
6030 | bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000); | |
6031 | bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500); | |
77b7dcea PV |
6032 | bfqd->bfq_wr_max_softrt_rate = 7000; /* |
6033 | * Approximate rate required | |
6034 | * to playback or record a | |
6035 | * high-definition compressed | |
6036 | * video. | |
6037 | */ | |
cfd69712 | 6038 | bfqd->wr_busy_queues = 0; |
44e44a1b PV |
6039 | |
6040 | /* | |
e24f1c24 PV |
6041 | * Begin by assuming, optimistically, that the device peak |
6042 | * rate is equal to 2/3 of the highest reference rate. | |
44e44a1b | 6043 | */ |
e24f1c24 PV |
6044 | bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] * |
6045 | ref_wr_duration[blk_queue_nonrot(bfqd->queue)]; | |
6046 | bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3; | |
44e44a1b | 6047 | |
aee69d78 | 6048 | spin_lock_init(&bfqd->lock); |
aee69d78 | 6049 | |
e21b7a0b AA |
6050 | /* |
6051 | * The invocation of the next bfq_create_group_hierarchy | |
6052 | * function is the head of a chain of function calls | |
6053 | * (bfq_create_group_hierarchy->blkcg_activate_policy-> | |
6054 | * blk_mq_freeze_queue) that may lead to the invocation of the | |
6055 | * has_work hook function. For this reason, | |
6056 | * bfq_create_group_hierarchy is invoked only after all | |
6057 | * scheduler data has been initialized, apart from the fields | |
6058 | * that can be initialized only after invoking | |
6059 | * bfq_create_group_hierarchy. This, in particular, enables | |
6060 | * has_work to correctly return false. Of course, to avoid | |
6061 | * other inconsistencies, the blk-mq stack must then refrain | |
6062 | * from invoking further scheduler hooks before this init | |
6063 | * function is finished. | |
6064 | */ | |
6065 | bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node); | |
6066 | if (!bfqd->root_group) | |
6067 | goto out_free; | |
6068 | bfq_init_root_group(bfqd->root_group, bfqd); | |
6069 | bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group); | |
6070 | ||
b5dc5d4d | 6071 | wbt_disable_default(q); |
aee69d78 | 6072 | return 0; |
e21b7a0b AA |
6073 | |
6074 | out_free: | |
6075 | kfree(bfqd); | |
6076 | kobject_put(&eq->kobj); | |
6077 | return -ENOMEM; | |
aee69d78 PV |
6078 | } |
6079 | ||
6080 | static void bfq_slab_kill(void) | |
6081 | { | |
6082 | kmem_cache_destroy(bfq_pool); | |
6083 | } | |
6084 | ||
6085 | static int __init bfq_slab_setup(void) | |
6086 | { | |
6087 | bfq_pool = KMEM_CACHE(bfq_queue, 0); | |
6088 | if (!bfq_pool) | |
6089 | return -ENOMEM; | |
6090 | return 0; | |
6091 | } | |
6092 | ||
6093 | static ssize_t bfq_var_show(unsigned int var, char *page) | |
6094 | { | |
6095 | return sprintf(page, "%u\n", var); | |
6096 | } | |
6097 | ||
2f79136b | 6098 | static int bfq_var_store(unsigned long *var, const char *page) |
aee69d78 PV |
6099 | { |
6100 | unsigned long new_val; | |
6101 | int ret = kstrtoul(page, 10, &new_val); | |
6102 | ||
2f79136b BVA |
6103 | if (ret) |
6104 | return ret; | |
6105 | *var = new_val; | |
6106 | return 0; | |
aee69d78 PV |
6107 | } |
6108 | ||
6109 | #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ | |
6110 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ | |
6111 | { \ | |
6112 | struct bfq_data *bfqd = e->elevator_data; \ | |
6113 | u64 __data = __VAR; \ | |
6114 | if (__CONV == 1) \ | |
6115 | __data = jiffies_to_msecs(__data); \ | |
6116 | else if (__CONV == 2) \ | |
6117 | __data = div_u64(__data, NSEC_PER_MSEC); \ | |
6118 | return bfq_var_show(__data, (page)); \ | |
6119 | } | |
6120 | SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2); | |
6121 | SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2); | |
6122 | SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0); | |
6123 | SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0); | |
6124 | SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2); | |
6125 | SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0); | |
6126 | SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1); | |
6127 | SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0); | |
44e44a1b | 6128 | SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0); |
aee69d78 PV |
6129 | #undef SHOW_FUNCTION |
6130 | ||
6131 | #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \ | |
6132 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ | |
6133 | { \ | |
6134 | struct bfq_data *bfqd = e->elevator_data; \ | |
6135 | u64 __data = __VAR; \ | |
6136 | __data = div_u64(__data, NSEC_PER_USEC); \ | |
6137 | return bfq_var_show(__data, (page)); \ | |
6138 | } | |
6139 | USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle); | |
6140 | #undef USEC_SHOW_FUNCTION | |
6141 | ||
6142 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ | |
6143 | static ssize_t \ | |
6144 | __FUNC(struct elevator_queue *e, const char *page, size_t count) \ | |
6145 | { \ | |
6146 | struct bfq_data *bfqd = e->elevator_data; \ | |
1530486c | 6147 | unsigned long __data, __min = (MIN), __max = (MAX); \ |
2f79136b BVA |
6148 | int ret; \ |
6149 | \ | |
6150 | ret = bfq_var_store(&__data, (page)); \ | |
6151 | if (ret) \ | |
6152 | return ret; \ | |
1530486c BVA |
6153 | if (__data < __min) \ |
6154 | __data = __min; \ | |
6155 | else if (__data > __max) \ | |
6156 | __data = __max; \ | |
aee69d78 PV |
6157 | if (__CONV == 1) \ |
6158 | *(__PTR) = msecs_to_jiffies(__data); \ | |
6159 | else if (__CONV == 2) \ | |
6160 | *(__PTR) = (u64)__data * NSEC_PER_MSEC; \ | |
6161 | else \ | |
6162 | *(__PTR) = __data; \ | |
235f8da1 | 6163 | return count; \ |
aee69d78 PV |
6164 | } |
6165 | STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1, | |
6166 | INT_MAX, 2); | |
6167 | STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1, | |
6168 | INT_MAX, 2); | |
6169 | STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0); | |
6170 | STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1, | |
6171 | INT_MAX, 0); | |
6172 | STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2); | |
6173 | #undef STORE_FUNCTION | |
6174 | ||
6175 | #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \ | |
6176 | static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\ | |
6177 | { \ | |
6178 | struct bfq_data *bfqd = e->elevator_data; \ | |
1530486c | 6179 | unsigned long __data, __min = (MIN), __max = (MAX); \ |
2f79136b BVA |
6180 | int ret; \ |
6181 | \ | |
6182 | ret = bfq_var_store(&__data, (page)); \ | |
6183 | if (ret) \ | |
6184 | return ret; \ | |
1530486c BVA |
6185 | if (__data < __min) \ |
6186 | __data = __min; \ | |
6187 | else if (__data > __max) \ | |
6188 | __data = __max; \ | |
aee69d78 | 6189 | *(__PTR) = (u64)__data * NSEC_PER_USEC; \ |
235f8da1 | 6190 | return count; \ |
aee69d78 PV |
6191 | } |
6192 | USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0, | |
6193 | UINT_MAX); | |
6194 | #undef USEC_STORE_FUNCTION | |
6195 | ||
aee69d78 PV |
6196 | static ssize_t bfq_max_budget_store(struct elevator_queue *e, |
6197 | const char *page, size_t count) | |
6198 | { | |
6199 | struct bfq_data *bfqd = e->elevator_data; | |
2f79136b BVA |
6200 | unsigned long __data; |
6201 | int ret; | |
235f8da1 | 6202 | |
2f79136b BVA |
6203 | ret = bfq_var_store(&__data, (page)); |
6204 | if (ret) | |
6205 | return ret; | |
aee69d78 PV |
6206 | |
6207 | if (__data == 0) | |
ab0e43e9 | 6208 | bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd); |
aee69d78 PV |
6209 | else { |
6210 | if (__data > INT_MAX) | |
6211 | __data = INT_MAX; | |
6212 | bfqd->bfq_max_budget = __data; | |
6213 | } | |
6214 | ||
6215 | bfqd->bfq_user_max_budget = __data; | |
6216 | ||
235f8da1 | 6217 | return count; |
aee69d78 PV |
6218 | } |
6219 | ||
6220 | /* | |
6221 | * Leaving this name to preserve name compatibility with cfq | |
6222 | * parameters, but this timeout is used for both sync and async. | |
6223 | */ | |
6224 | static ssize_t bfq_timeout_sync_store(struct elevator_queue *e, | |
6225 | const char *page, size_t count) | |
6226 | { | |
6227 | struct bfq_data *bfqd = e->elevator_data; | |
2f79136b BVA |
6228 | unsigned long __data; |
6229 | int ret; | |
235f8da1 | 6230 | |
2f79136b BVA |
6231 | ret = bfq_var_store(&__data, (page)); |
6232 | if (ret) | |
6233 | return ret; | |
aee69d78 PV |
6234 | |
6235 | if (__data < 1) | |
6236 | __data = 1; | |
6237 | else if (__data > INT_MAX) | |
6238 | __data = INT_MAX; | |
6239 | ||
6240 | bfqd->bfq_timeout = msecs_to_jiffies(__data); | |
6241 | if (bfqd->bfq_user_max_budget == 0) | |
ab0e43e9 | 6242 | bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd); |
aee69d78 | 6243 | |
235f8da1 | 6244 | return count; |
aee69d78 PV |
6245 | } |
6246 | ||
6247 | static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e, | |
6248 | const char *page, size_t count) | |
6249 | { | |
6250 | struct bfq_data *bfqd = e->elevator_data; | |
2f79136b BVA |
6251 | unsigned long __data; |
6252 | int ret; | |
235f8da1 | 6253 | |
2f79136b BVA |
6254 | ret = bfq_var_store(&__data, (page)); |
6255 | if (ret) | |
6256 | return ret; | |
aee69d78 PV |
6257 | |
6258 | if (__data > 1) | |
6259 | __data = 1; | |
6260 | if (!bfqd->strict_guarantees && __data == 1 | |
6261 | && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC) | |
6262 | bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC; | |
6263 | ||
6264 | bfqd->strict_guarantees = __data; | |
6265 | ||
235f8da1 | 6266 | return count; |
aee69d78 PV |
6267 | } |
6268 | ||
44e44a1b PV |
6269 | static ssize_t bfq_low_latency_store(struct elevator_queue *e, |
6270 | const char *page, size_t count) | |
6271 | { | |
6272 | struct bfq_data *bfqd = e->elevator_data; | |
2f79136b BVA |
6273 | unsigned long __data; |
6274 | int ret; | |
235f8da1 | 6275 | |
2f79136b BVA |
6276 | ret = bfq_var_store(&__data, (page)); |
6277 | if (ret) | |
6278 | return ret; | |
44e44a1b PV |
6279 | |
6280 | if (__data > 1) | |
6281 | __data = 1; | |
6282 | if (__data == 0 && bfqd->low_latency != 0) | |
6283 | bfq_end_wr(bfqd); | |
6284 | bfqd->low_latency = __data; | |
6285 | ||
235f8da1 | 6286 | return count; |
44e44a1b PV |
6287 | } |
6288 | ||
aee69d78 PV |
6289 | #define BFQ_ATTR(name) \ |
6290 | __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store) | |
6291 | ||
6292 | static struct elv_fs_entry bfq_attrs[] = { | |
6293 | BFQ_ATTR(fifo_expire_sync), | |
6294 | BFQ_ATTR(fifo_expire_async), | |
6295 | BFQ_ATTR(back_seek_max), | |
6296 | BFQ_ATTR(back_seek_penalty), | |
6297 | BFQ_ATTR(slice_idle), | |
6298 | BFQ_ATTR(slice_idle_us), | |
6299 | BFQ_ATTR(max_budget), | |
6300 | BFQ_ATTR(timeout_sync), | |
6301 | BFQ_ATTR(strict_guarantees), | |
44e44a1b | 6302 | BFQ_ATTR(low_latency), |
aee69d78 PV |
6303 | __ATTR_NULL |
6304 | }; | |
6305 | ||
6306 | static struct elevator_type iosched_bfq_mq = { | |
f9cd4bfe | 6307 | .ops = { |
a52a69ea | 6308 | .limit_depth = bfq_limit_depth, |
5bbf4e5a | 6309 | .prepare_request = bfq_prepare_request, |
a7877390 PV |
6310 | .requeue_request = bfq_finish_requeue_request, |
6311 | .finish_request = bfq_finish_requeue_request, | |
aee69d78 PV |
6312 | .exit_icq = bfq_exit_icq, |
6313 | .insert_requests = bfq_insert_requests, | |
6314 | .dispatch_request = bfq_dispatch_request, | |
6315 | .next_request = elv_rb_latter_request, | |
6316 | .former_request = elv_rb_former_request, | |
6317 | .allow_merge = bfq_allow_bio_merge, | |
6318 | .bio_merge = bfq_bio_merge, | |
6319 | .request_merge = bfq_request_merge, | |
6320 | .requests_merged = bfq_requests_merged, | |
6321 | .request_merged = bfq_request_merged, | |
6322 | .has_work = bfq_has_work, | |
77f1e0a5 | 6323 | .depth_updated = bfq_depth_updated, |
f0635b8a | 6324 | .init_hctx = bfq_init_hctx, |
aee69d78 PV |
6325 | .init_sched = bfq_init_queue, |
6326 | .exit_sched = bfq_exit_queue, | |
6327 | }, | |
6328 | ||
aee69d78 PV |
6329 | .icq_size = sizeof(struct bfq_io_cq), |
6330 | .icq_align = __alignof__(struct bfq_io_cq), | |
6331 | .elevator_attrs = bfq_attrs, | |
6332 | .elevator_name = "bfq", | |
6333 | .elevator_owner = THIS_MODULE, | |
6334 | }; | |
26b4cf24 | 6335 | MODULE_ALIAS("bfq-iosched"); |
aee69d78 PV |
6336 | |
6337 | static int __init bfq_init(void) | |
6338 | { | |
6339 | int ret; | |
6340 | ||
e21b7a0b AA |
6341 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
6342 | ret = blkcg_policy_register(&blkcg_policy_bfq); | |
6343 | if (ret) | |
6344 | return ret; | |
6345 | #endif | |
6346 | ||
aee69d78 PV |
6347 | ret = -ENOMEM; |
6348 | if (bfq_slab_setup()) | |
6349 | goto err_pol_unreg; | |
6350 | ||
44e44a1b PV |
6351 | /* |
6352 | * Times to load large popular applications for the typical | |
6353 | * systems installed on the reference devices (see the | |
e24f1c24 PV |
6354 | * comments before the definition of the next |
6355 | * array). Actually, we use slightly lower values, as the | |
44e44a1b PV |
6356 | * estimated peak rate tends to be smaller than the actual |
6357 | * peak rate. The reason for this last fact is that estimates | |
6358 | * are computed over much shorter time intervals than the long | |
6359 | * intervals typically used for benchmarking. Why? First, to | |
6360 | * adapt more quickly to variations. Second, because an I/O | |
6361 | * scheduler cannot rely on a peak-rate-evaluation workload to | |
6362 | * be run for a long time. | |
6363 | */ | |
e24f1c24 PV |
6364 | ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */ |
6365 | ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */ | |
44e44a1b | 6366 | |
aee69d78 PV |
6367 | ret = elv_register(&iosched_bfq_mq); |
6368 | if (ret) | |
37dcd657 | 6369 | goto slab_kill; |
aee69d78 PV |
6370 | |
6371 | return 0; | |
6372 | ||
37dcd657 | 6373 | slab_kill: |
6374 | bfq_slab_kill(); | |
aee69d78 | 6375 | err_pol_unreg: |
e21b7a0b AA |
6376 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
6377 | blkcg_policy_unregister(&blkcg_policy_bfq); | |
6378 | #endif | |
aee69d78 PV |
6379 | return ret; |
6380 | } | |
6381 | ||
6382 | static void __exit bfq_exit(void) | |
6383 | { | |
6384 | elv_unregister(&iosched_bfq_mq); | |
e21b7a0b AA |
6385 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
6386 | blkcg_policy_unregister(&blkcg_policy_bfq); | |
6387 | #endif | |
aee69d78 PV |
6388 | bfq_slab_kill(); |
6389 | } | |
6390 | ||
6391 | module_init(bfq_init); | |
6392 | module_exit(bfq_exit); | |
6393 | ||
6394 | MODULE_AUTHOR("Paolo Valente"); | |
6395 | MODULE_LICENSE("GPL"); | |
6396 | MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler"); |