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