]>
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 | |
52 | * applications: interactive and soft real-time. This feature enables | |
53 | * BFQ to provide applications in these classes with a very low | |
54 | * latency. Finally, BFQ also features additional heuristics for | |
55 | * preserving both a low latency and a high throughput on NCQ-capable, | |
56 | * rotational or flash-based devices, and to get the job done quickly | |
57 | * for applications consisting in many I/O-bound processes. | |
58 | * | |
59 | * BFQ is described in [1], where also a reference to the initial, more | |
60 | * theoretical paper on BFQ can be found. The interested reader can find | |
61 | * in the latter paper full details on the main algorithm, as well as | |
62 | * formulas of the guarantees and formal proofs of all the properties. | |
63 | * With respect to the version of BFQ presented in these papers, this | |
64 | * implementation adds a few more heuristics, such as the one that | |
65 | * guarantees a low latency to soft real-time applications, and a | |
66 | * hierarchical extension based on H-WF2Q+. | |
67 | * | |
68 | * B-WF2Q+ is based on WF2Q+, which is described in [2], together with | |
69 | * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+ | |
70 | * with O(log N) complexity derives from the one introduced with EEVDF | |
71 | * in [3]. | |
72 | * | |
73 | * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O | |
74 | * Scheduler", Proceedings of the First Workshop on Mobile System | |
75 | * Technologies (MST-2015), May 2015. | |
76 | * http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf | |
77 | * | |
78 | * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing | |
79 | * Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689, | |
80 | * Oct 1997. | |
81 | * | |
82 | * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz | |
83 | * | |
84 | * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline | |
85 | * First: A Flexible and Accurate Mechanism for Proportional Share | |
86 | * Resource Allocation", technical report. | |
87 | * | |
88 | * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf | |
89 | */ | |
90 | #include <linux/module.h> | |
91 | #include <linux/slab.h> | |
92 | #include <linux/blkdev.h> | |
e21b7a0b | 93 | #include <linux/cgroup.h> |
aee69d78 PV |
94 | #include <linux/elevator.h> |
95 | #include <linux/ktime.h> | |
96 | #include <linux/rbtree.h> | |
97 | #include <linux/ioprio.h> | |
98 | #include <linux/sbitmap.h> | |
99 | #include <linux/delay.h> | |
100 | ||
101 | #include "blk.h" | |
102 | #include "blk-mq.h" | |
103 | #include "blk-mq-tag.h" | |
104 | #include "blk-mq-sched.h" | |
105 | #include <linux/blktrace_api.h> | |
106 | #include <linux/hrtimer.h> | |
107 | #include <linux/blk-cgroup.h> | |
108 | ||
109 | #define BFQ_IOPRIO_CLASSES 3 | |
110 | #define BFQ_CL_IDLE_TIMEOUT (HZ/5) | |
111 | ||
112 | #define BFQ_MIN_WEIGHT 1 | |
113 | #define BFQ_MAX_WEIGHT 1000 | |
114 | #define BFQ_WEIGHT_CONVERSION_COEFF 10 | |
115 | ||
116 | #define BFQ_DEFAULT_QUEUE_IOPRIO 4 | |
117 | ||
e21b7a0b | 118 | #define BFQ_WEIGHT_LEGACY_DFL 100 |
aee69d78 PV |
119 | #define BFQ_DEFAULT_GRP_IOPRIO 0 |
120 | #define BFQ_DEFAULT_GRP_CLASS IOPRIO_CLASS_BE | |
121 | ||
122 | struct bfq_entity; | |
123 | ||
124 | /** | |
125 | * struct bfq_service_tree - per ioprio_class service tree. | |
126 | * | |
127 | * Each service tree represents a B-WF2Q+ scheduler on its own. Each | |
128 | * ioprio_class has its own independent scheduler, and so its own | |
129 | * bfq_service_tree. All the fields are protected by the queue lock | |
130 | * of the containing bfqd. | |
131 | */ | |
132 | struct bfq_service_tree { | |
133 | /* tree for active entities (i.e., those backlogged) */ | |
134 | struct rb_root active; | |
135 | /* tree for idle entities (i.e., not backlogged, with V <= F_i)*/ | |
136 | struct rb_root idle; | |
137 | ||
138 | /* idle entity with minimum F_i */ | |
139 | struct bfq_entity *first_idle; | |
140 | /* idle entity with maximum F_i */ | |
141 | struct bfq_entity *last_idle; | |
142 | ||
143 | /* scheduler virtual time */ | |
144 | u64 vtime; | |
145 | /* scheduler weight sum; active and idle entities contribute to it */ | |
146 | unsigned long wsum; | |
147 | }; | |
148 | ||
149 | /** | |
150 | * struct bfq_sched_data - multi-class scheduler. | |
151 | * | |
152 | * bfq_sched_data is the basic scheduler queue. It supports three | |
e21b7a0b AA |
153 | * ioprio_classes, and can be used either as a toplevel queue or as an |
154 | * intermediate queue on a hierarchical setup. @next_in_service | |
155 | * points to the active entity of the sched_data service trees that | |
156 | * will be scheduled next. It is used to reduce the number of steps | |
157 | * needed for each hierarchical-schedule update. | |
aee69d78 PV |
158 | * |
159 | * The supported ioprio_classes are the same as in CFQ, in descending | |
160 | * priority order, IOPRIO_CLASS_RT, IOPRIO_CLASS_BE, IOPRIO_CLASS_IDLE. | |
161 | * Requests from higher priority queues are served before all the | |
162 | * requests from lower priority queues; among requests of the same | |
163 | * queue requests are served according to B-WF2Q+. | |
164 | * All the fields are protected by the queue lock of the containing bfqd. | |
165 | */ | |
166 | struct bfq_sched_data { | |
167 | /* entity in service */ | |
168 | struct bfq_entity *in_service_entity; | |
e21b7a0b | 169 | /* head-of-line entity (see comments above) */ |
aee69d78 PV |
170 | struct bfq_entity *next_in_service; |
171 | /* array of service trees, one per ioprio_class */ | |
172 | struct bfq_service_tree service_tree[BFQ_IOPRIO_CLASSES]; | |
e21b7a0b AA |
173 | /* last time CLASS_IDLE was served */ |
174 | unsigned long bfq_class_idle_last_service; | |
175 | ||
aee69d78 PV |
176 | }; |
177 | ||
178 | /** | |
179 | * struct bfq_entity - schedulable entity. | |
180 | * | |
e21b7a0b AA |
181 | * A bfq_entity is used to represent either a bfq_queue (leaf node in the |
182 | * cgroup hierarchy) or a bfq_group into the upper level scheduler. Each | |
183 | * entity belongs to the sched_data of the parent group in the cgroup | |
184 | * hierarchy. Non-leaf entities have also their own sched_data, stored | |
185 | * in @my_sched_data. | |
aee69d78 PV |
186 | * |
187 | * Each entity stores independently its priority values; this would | |
188 | * allow different weights on different devices, but this | |
189 | * functionality is not exported to userspace by now. Priorities and | |
190 | * weights are updated lazily, first storing the new values into the | |
191 | * new_* fields, then setting the @prio_changed flag. As soon as | |
192 | * there is a transition in the entity state that allows the priority | |
193 | * update to take place the effective and the requested priority | |
194 | * values are synchronized. | |
195 | * | |
e21b7a0b AA |
196 | * Unless cgroups are used, the weight value is calculated from the |
197 | * ioprio to export the same interface as CFQ. When dealing with | |
198 | * ``well-behaved'' queues (i.e., queues that do not spend too much | |
199 | * time to consume their budget and have true sequential behavior, and | |
200 | * when there are no external factors breaking anticipation) the | |
201 | * relative weights at each level of the cgroups hierarchy should be | |
202 | * guaranteed. All the fields are protected by the queue lock of the | |
203 | * containing bfqd. | |
aee69d78 PV |
204 | */ |
205 | struct bfq_entity { | |
206 | /* service_tree member */ | |
207 | struct rb_node rb_node; | |
208 | ||
209 | /* | |
e21b7a0b AA |
210 | * Flag, true if the entity is on a tree (either the active or |
211 | * the idle one of its service_tree) or is in service. | |
aee69d78 | 212 | */ |
e21b7a0b | 213 | bool on_st; |
aee69d78 PV |
214 | |
215 | /* B-WF2Q+ start and finish timestamps [sectors/weight] */ | |
216 | u64 start, finish; | |
217 | ||
218 | /* tree the entity is enqueued into; %NULL if not on a tree */ | |
219 | struct rb_root *tree; | |
220 | ||
221 | /* | |
222 | * minimum start time of the (active) subtree rooted at this | |
223 | * entity; used for O(log N) lookups into active trees | |
224 | */ | |
225 | u64 min_start; | |
226 | ||
227 | /* amount of service received during the last service slot */ | |
228 | int service; | |
229 | ||
230 | /* budget, used also to calculate F_i: F_i = S_i + @budget / @weight */ | |
231 | int budget; | |
232 | ||
233 | /* weight of the queue */ | |
234 | int weight; | |
235 | /* next weight if a change is in progress */ | |
236 | int new_weight; | |
237 | ||
238 | /* original weight, used to implement weight boosting */ | |
239 | int orig_weight; | |
240 | ||
241 | /* parent entity, for hierarchical scheduling */ | |
242 | struct bfq_entity *parent; | |
243 | ||
244 | /* | |
245 | * For non-leaf nodes in the hierarchy, the associated | |
246 | * scheduler queue, %NULL on leaf nodes. | |
247 | */ | |
248 | struct bfq_sched_data *my_sched_data; | |
249 | /* the scheduler queue this entity belongs to */ | |
250 | struct bfq_sched_data *sched_data; | |
251 | ||
252 | /* flag, set to request a weight, ioprio or ioprio_class change */ | |
253 | int prio_changed; | |
254 | }; | |
255 | ||
e21b7a0b AA |
256 | struct bfq_group; |
257 | ||
aee69d78 PV |
258 | /** |
259 | * struct bfq_ttime - per process thinktime stats. | |
260 | */ | |
261 | struct bfq_ttime { | |
262 | /* completion time of the last request */ | |
263 | u64 last_end_request; | |
264 | ||
265 | /* total process thinktime */ | |
266 | u64 ttime_total; | |
267 | /* number of thinktime samples */ | |
268 | unsigned long ttime_samples; | |
269 | /* average process thinktime */ | |
270 | u64 ttime_mean; | |
271 | }; | |
272 | ||
273 | /** | |
274 | * struct bfq_queue - leaf schedulable entity. | |
275 | * | |
276 | * A bfq_queue is a leaf request queue; it can be associated with an | |
e21b7a0b AA |
277 | * io_context or more, if it is async. @cgroup holds a reference to |
278 | * the cgroup, to be sure that it does not disappear while a bfqq | |
279 | * still references it (mostly to avoid races between request issuing | |
280 | * and task migration followed by cgroup destruction). All the fields | |
281 | * are protected by the queue lock of the containing bfqd. | |
aee69d78 PV |
282 | */ |
283 | struct bfq_queue { | |
284 | /* reference counter */ | |
285 | int ref; | |
286 | /* parent bfq_data */ | |
287 | struct bfq_data *bfqd; | |
288 | ||
289 | /* current ioprio and ioprio class */ | |
290 | unsigned short ioprio, ioprio_class; | |
291 | /* next ioprio and ioprio class if a change is in progress */ | |
292 | unsigned short new_ioprio, new_ioprio_class; | |
293 | ||
294 | /* sorted list of pending requests */ | |
295 | struct rb_root sort_list; | |
296 | /* if fifo isn't expired, next request to serve */ | |
297 | struct request *next_rq; | |
298 | /* number of sync and async requests queued */ | |
299 | int queued[2]; | |
300 | /* number of requests currently allocated */ | |
301 | int allocated; | |
302 | /* number of pending metadata requests */ | |
303 | int meta_pending; | |
304 | /* fifo list of requests in sort_list */ | |
305 | struct list_head fifo; | |
306 | ||
307 | /* entity representing this queue in the scheduler */ | |
308 | struct bfq_entity entity; | |
309 | ||
310 | /* maximum budget allowed from the feedback mechanism */ | |
311 | int max_budget; | |
312 | /* budget expiration (in jiffies) */ | |
313 | unsigned long budget_timeout; | |
314 | ||
315 | /* number of requests on the dispatch list or inside driver */ | |
316 | int dispatched; | |
317 | ||
318 | /* status flags */ | |
319 | unsigned long flags; | |
320 | ||
321 | /* node for active/idle bfqq list inside parent bfqd */ | |
322 | struct list_head bfqq_list; | |
323 | ||
324 | /* associated @bfq_ttime struct */ | |
325 | struct bfq_ttime ttime; | |
326 | ||
327 | /* bit vector: a 1 for each seeky requests in history */ | |
328 | u32 seek_history; | |
329 | /* position of the last request enqueued */ | |
330 | sector_t last_request_pos; | |
331 | ||
332 | /* Number of consecutive pairs of request completion and | |
333 | * arrival, such that the queue becomes idle after the | |
334 | * completion, but the next request arrives within an idle | |
335 | * time slice; used only if the queue's IO_bound flag has been | |
336 | * cleared. | |
337 | */ | |
338 | unsigned int requests_within_timer; | |
339 | ||
340 | /* pid of the process owning the queue, used for logging purposes */ | |
341 | pid_t pid; | |
44e44a1b PV |
342 | |
343 | /* current maximum weight-raising time for this queue */ | |
344 | unsigned long wr_cur_max_time; | |
345 | /* | |
346 | * Start time of the current weight-raising period if | |
347 | * the @bfq-queue is being weight-raised, otherwise | |
348 | * finish time of the last weight-raising period. | |
349 | */ | |
350 | unsigned long last_wr_start_finish; | |
351 | /* factor by which the weight of this queue is multiplied */ | |
352 | unsigned int wr_coeff; | |
aee69d78 PV |
353 | }; |
354 | ||
355 | /** | |
356 | * struct bfq_io_cq - per (request_queue, io_context) structure. | |
357 | */ | |
358 | struct bfq_io_cq { | |
359 | /* associated io_cq structure */ | |
360 | struct io_cq icq; /* must be the first member */ | |
361 | /* array of two process queues, the sync and the async */ | |
362 | struct bfq_queue *bfqq[2]; | |
363 | /* per (request_queue, blkcg) ioprio */ | |
364 | int ioprio; | |
e21b7a0b AA |
365 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
366 | uint64_t blkcg_serial_nr; /* the current blkcg serial */ | |
367 | #endif | |
aee69d78 PV |
368 | }; |
369 | ||
44e44a1b PV |
370 | enum bfq_device_speed { |
371 | BFQ_BFQD_FAST, | |
372 | BFQ_BFQD_SLOW, | |
373 | }; | |
374 | ||
aee69d78 PV |
375 | /** |
376 | * struct bfq_data - per-device data structure. | |
377 | * | |
378 | * All the fields are protected by @lock. | |
379 | */ | |
380 | struct bfq_data { | |
381 | /* device request queue */ | |
382 | struct request_queue *queue; | |
383 | /* dispatch queue */ | |
384 | struct list_head dispatch; | |
385 | ||
e21b7a0b AA |
386 | /* root bfq_group for the device */ |
387 | struct bfq_group *root_group; | |
aee69d78 PV |
388 | |
389 | /* | |
390 | * Number of bfq_queues containing requests (including the | |
391 | * queue in service, even if it is idling). | |
392 | */ | |
393 | int busy_queues; | |
394 | /* number of queued requests */ | |
395 | int queued; | |
396 | /* number of requests dispatched and waiting for completion */ | |
397 | int rq_in_driver; | |
398 | ||
399 | /* | |
400 | * Maximum number of requests in driver in the last | |
401 | * @hw_tag_samples completed requests. | |
402 | */ | |
403 | int max_rq_in_driver; | |
404 | /* number of samples used to calculate hw_tag */ | |
405 | int hw_tag_samples; | |
406 | /* flag set to one if the driver is showing a queueing behavior */ | |
407 | int hw_tag; | |
408 | ||
409 | /* number of budgets assigned */ | |
410 | int budgets_assigned; | |
411 | ||
412 | /* | |
413 | * Timer set when idling (waiting) for the next request from | |
414 | * the queue in service. | |
415 | */ | |
416 | struct hrtimer idle_slice_timer; | |
417 | ||
418 | /* bfq_queue in service */ | |
419 | struct bfq_queue *in_service_queue; | |
420 | /* bfq_io_cq (bic) associated with the @in_service_queue */ | |
421 | struct bfq_io_cq *in_service_bic; | |
422 | ||
423 | /* on-disk position of the last served request */ | |
424 | sector_t last_position; | |
425 | ||
ab0e43e9 PV |
426 | /* time of last request completion (ns) */ |
427 | u64 last_completion; | |
428 | ||
429 | /* time of first rq dispatch in current observation interval (ns) */ | |
430 | u64 first_dispatch; | |
431 | /* time of last rq dispatch in current observation interval (ns) */ | |
432 | u64 last_dispatch; | |
433 | ||
aee69d78 PV |
434 | /* beginning of the last budget */ |
435 | ktime_t last_budget_start; | |
436 | /* beginning of the last idle slice */ | |
437 | ktime_t last_idling_start; | |
ab0e43e9 PV |
438 | |
439 | /* number of samples in current observation interval */ | |
aee69d78 | 440 | int peak_rate_samples; |
ab0e43e9 PV |
441 | /* num of samples of seq dispatches in current observation interval */ |
442 | u32 sequential_samples; | |
443 | /* total num of sectors transferred in current observation interval */ | |
444 | u64 tot_sectors_dispatched; | |
445 | /* max rq size seen during current observation interval (sectors) */ | |
446 | u32 last_rq_max_size; | |
447 | /* time elapsed from first dispatch in current observ. interval (us) */ | |
448 | u64 delta_from_first; | |
aee69d78 | 449 | /* |
ab0e43e9 PV |
450 | * Current estimate of the device peak rate, measured in |
451 | * [BFQ_RATE_SHIFT * sectors/usec]. The left-shift by | |
452 | * BFQ_RATE_SHIFT is performed to increase precision in | |
aee69d78 PV |
453 | * fixed-point calculations. |
454 | */ | |
ab0e43e9 PV |
455 | u32 peak_rate; |
456 | ||
aee69d78 PV |
457 | /* maximum budget allotted to a bfq_queue before rescheduling */ |
458 | int bfq_max_budget; | |
459 | ||
460 | /* list of all the bfq_queues active on the device */ | |
461 | struct list_head active_list; | |
462 | /* list of all the bfq_queues idle on the device */ | |
463 | struct list_head idle_list; | |
464 | ||
465 | /* | |
466 | * Timeout for async/sync requests; when it fires, requests | |
467 | * are served in fifo order. | |
468 | */ | |
469 | u64 bfq_fifo_expire[2]; | |
470 | /* weight of backward seeks wrt forward ones */ | |
471 | unsigned int bfq_back_penalty; | |
472 | /* maximum allowed backward seek */ | |
473 | unsigned int bfq_back_max; | |
474 | /* maximum idling time */ | |
475 | u32 bfq_slice_idle; | |
aee69d78 PV |
476 | |
477 | /* user-configured max budget value (0 for auto-tuning) */ | |
478 | int bfq_user_max_budget; | |
479 | /* | |
480 | * Timeout for bfq_queues to consume their budget; used to | |
481 | * prevent seeky queues from imposing long latencies to | |
482 | * sequential or quasi-sequential ones (this also implies that | |
483 | * seeky queues cannot receive guarantees in the service | |
484 | * domain; after a timeout they are charged for the time they | |
485 | * have been in service, to preserve fairness among them, but | |
486 | * without service-domain guarantees). | |
487 | */ | |
488 | unsigned int bfq_timeout; | |
489 | ||
490 | /* | |
491 | * Number of consecutive requests that must be issued within | |
492 | * the idle time slice to set again idling to a queue which | |
493 | * was marked as non-I/O-bound (see the definition of the | |
494 | * IO_bound flag for further details). | |
495 | */ | |
496 | unsigned int bfq_requests_within_timer; | |
497 | ||
498 | /* | |
499 | * Force device idling whenever needed to provide accurate | |
500 | * service guarantees, without caring about throughput | |
501 | * issues. CAVEAT: this may even increase latencies, in case | |
502 | * of useless idling for processes that did stop doing I/O. | |
503 | */ | |
504 | bool strict_guarantees; | |
505 | ||
44e44a1b PV |
506 | /* if set to true, low-latency heuristics are enabled */ |
507 | bool low_latency; | |
508 | /* | |
509 | * Maximum factor by which the weight of a weight-raised queue | |
510 | * is multiplied. | |
511 | */ | |
512 | unsigned int bfq_wr_coeff; | |
513 | /* maximum duration of a weight-raising period (jiffies) */ | |
514 | unsigned int bfq_wr_max_time; | |
515 | /* | |
516 | * Minimum idle period after which weight-raising may be | |
517 | * reactivated for a queue (in jiffies). | |
518 | */ | |
519 | unsigned int bfq_wr_min_idle_time; | |
520 | /* | |
521 | * Minimum period between request arrivals after which | |
522 | * weight-raising may be reactivated for an already busy async | |
523 | * queue (in jiffies). | |
524 | */ | |
525 | unsigned long bfq_wr_min_inter_arr_async; | |
526 | /* | |
527 | * Cached value of the product R*T, used for computing the | |
528 | * maximum duration of weight raising automatically. | |
529 | */ | |
530 | u64 RT_prod; | |
531 | /* device-speed class for the low-latency heuristic */ | |
532 | enum bfq_device_speed device_speed; | |
533 | ||
aee69d78 PV |
534 | /* fallback dummy bfqq for extreme OOM conditions */ |
535 | struct bfq_queue oom_bfqq; | |
536 | ||
537 | spinlock_t lock; | |
538 | ||
539 | /* | |
540 | * bic associated with the task issuing current bio for | |
541 | * merging. This and the next field are used as a support to | |
542 | * be able to perform the bic lookup, needed by bio-merge | |
543 | * functions, before the scheduler lock is taken, and thus | |
544 | * avoid taking the request-queue lock while the scheduler | |
545 | * lock is being held. | |
546 | */ | |
547 | struct bfq_io_cq *bio_bic; | |
548 | /* bfqq associated with the task issuing current bio for merging */ | |
549 | struct bfq_queue *bio_bfqq; | |
550 | }; | |
551 | ||
552 | enum bfqq_state_flags { | |
553 | BFQQF_busy = 0, /* has requests or is in service */ | |
554 | BFQQF_wait_request, /* waiting for a request */ | |
555 | BFQQF_non_blocking_wait_rq, /* | |
556 | * waiting for a request | |
557 | * without idling the device | |
558 | */ | |
559 | BFQQF_fifo_expire, /* FIFO checked in this slice */ | |
560 | BFQQF_idle_window, /* slice idling enabled */ | |
561 | BFQQF_sync, /* synchronous queue */ | |
aee69d78 PV |
562 | BFQQF_IO_bound, /* |
563 | * bfqq has timed-out at least once | |
564 | * having consumed at most 2/10 of | |
565 | * its budget | |
566 | */ | |
567 | }; | |
568 | ||
569 | #define BFQ_BFQQ_FNS(name) \ | |
570 | static void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \ | |
571 | { \ | |
572 | __set_bit(BFQQF_##name, &(bfqq)->flags); \ | |
573 | } \ | |
574 | static void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \ | |
575 | { \ | |
576 | __clear_bit(BFQQF_##name, &(bfqq)->flags); \ | |
577 | } \ | |
578 | static int bfq_bfqq_##name(const struct bfq_queue *bfqq) \ | |
579 | { \ | |
580 | return test_bit(BFQQF_##name, &(bfqq)->flags); \ | |
581 | } | |
582 | ||
583 | BFQ_BFQQ_FNS(busy); | |
584 | BFQ_BFQQ_FNS(wait_request); | |
585 | BFQ_BFQQ_FNS(non_blocking_wait_rq); | |
586 | BFQ_BFQQ_FNS(fifo_expire); | |
587 | BFQ_BFQQ_FNS(idle_window); | |
588 | BFQ_BFQQ_FNS(sync); | |
aee69d78 PV |
589 | BFQ_BFQQ_FNS(IO_bound); |
590 | #undef BFQ_BFQQ_FNS | |
591 | ||
592 | /* Logging facilities. */ | |
e21b7a0b AA |
593 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
594 | static struct bfq_group *bfqq_group(struct bfq_queue *bfqq); | |
595 | static struct blkcg_gq *bfqg_to_blkg(struct bfq_group *bfqg); | |
596 | ||
597 | #define bfq_log_bfqq(bfqd, bfqq, fmt, args...) do { \ | |
598 | char __pbuf[128]; \ | |
599 | \ | |
600 | blkg_path(bfqg_to_blkg(bfqq_group(bfqq)), __pbuf, sizeof(__pbuf)); \ | |
601 | blk_add_trace_msg((bfqd)->queue, "bfq%d%c %s " fmt, (bfqq)->pid, \ | |
602 | bfq_bfqq_sync((bfqq)) ? 'S' : 'A', \ | |
603 | __pbuf, ##args); \ | |
604 | } while (0) | |
605 | ||
606 | #define bfq_log_bfqg(bfqd, bfqg, fmt, args...) do { \ | |
607 | char __pbuf[128]; \ | |
608 | \ | |
609 | blkg_path(bfqg_to_blkg(bfqg), __pbuf, sizeof(__pbuf)); \ | |
610 | blk_add_trace_msg((bfqd)->queue, "%s " fmt, __pbuf, ##args); \ | |
611 | } while (0) | |
612 | ||
613 | #else /* CONFIG_BFQ_GROUP_IOSCHED */ | |
614 | ||
615 | #define bfq_log_bfqq(bfqd, bfqq, fmt, args...) \ | |
616 | blk_add_trace_msg((bfqd)->queue, "bfq%d%c " fmt, (bfqq)->pid, \ | |
617 | bfq_bfqq_sync((bfqq)) ? 'S' : 'A', \ | |
618 | ##args) | |
619 | #define bfq_log_bfqg(bfqd, bfqg, fmt, args...) do {} while (0) | |
620 | ||
621 | #endif /* CONFIG_BFQ_GROUP_IOSCHED */ | |
aee69d78 PV |
622 | |
623 | #define bfq_log(bfqd, fmt, args...) \ | |
624 | blk_add_trace_msg((bfqd)->queue, "bfq " fmt, ##args) | |
625 | ||
626 | /* Expiration reasons. */ | |
627 | enum bfqq_expiration { | |
628 | BFQQE_TOO_IDLE = 0, /* | |
629 | * queue has been idling for | |
630 | * too long | |
631 | */ | |
632 | BFQQE_BUDGET_TIMEOUT, /* budget took too long to be used */ | |
633 | BFQQE_BUDGET_EXHAUSTED, /* budget consumed */ | |
634 | BFQQE_NO_MORE_REQUESTS, /* the queue has no more requests */ | |
635 | BFQQE_PREEMPTED /* preemption in progress */ | |
636 | }; | |
637 | ||
e21b7a0b AA |
638 | struct bfqg_stats { |
639 | #ifdef CONFIG_BFQ_GROUP_IOSCHED | |
640 | /* number of ios merged */ | |
641 | struct blkg_rwstat merged; | |
642 | /* total time spent on device in ns, may not be accurate w/ queueing */ | |
643 | struct blkg_rwstat service_time; | |
644 | /* total time spent waiting in scheduler queue in ns */ | |
645 | struct blkg_rwstat wait_time; | |
646 | /* number of IOs queued up */ | |
647 | struct blkg_rwstat queued; | |
648 | /* total disk time and nr sectors dispatched by this group */ | |
649 | struct blkg_stat time; | |
650 | /* sum of number of ios queued across all samples */ | |
651 | struct blkg_stat avg_queue_size_sum; | |
652 | /* count of samples taken for average */ | |
653 | struct blkg_stat avg_queue_size_samples; | |
654 | /* how many times this group has been removed from service tree */ | |
655 | struct blkg_stat dequeue; | |
656 | /* total time spent waiting for it to be assigned a timeslice. */ | |
657 | struct blkg_stat group_wait_time; | |
658 | /* time spent idling for this blkcg_gq */ | |
659 | struct blkg_stat idle_time; | |
660 | /* total time with empty current active q with other requests queued */ | |
661 | struct blkg_stat empty_time; | |
662 | /* fields after this shouldn't be cleared on stat reset */ | |
663 | uint64_t start_group_wait_time; | |
664 | uint64_t start_idle_time; | |
665 | uint64_t start_empty_time; | |
666 | uint16_t flags; | |
667 | #endif /* CONFIG_BFQ_GROUP_IOSCHED */ | |
668 | }; | |
669 | ||
670 | #ifdef CONFIG_BFQ_GROUP_IOSCHED | |
671 | ||
672 | /* | |
673 | * struct bfq_group_data - per-blkcg storage for the blkio subsystem. | |
674 | * | |
675 | * @ps: @blkcg_policy_storage that this structure inherits | |
676 | * @weight: weight of the bfq_group | |
677 | */ | |
678 | struct bfq_group_data { | |
679 | /* must be the first member */ | |
680 | struct blkcg_policy_data pd; | |
681 | ||
44e44a1b | 682 | unsigned int weight; |
e21b7a0b AA |
683 | }; |
684 | ||
685 | /** | |
686 | * struct bfq_group - per (device, cgroup) data structure. | |
687 | * @entity: schedulable entity to insert into the parent group sched_data. | |
688 | * @sched_data: own sched_data, to contain child entities (they may be | |
689 | * both bfq_queues and bfq_groups). | |
690 | * @bfqd: the bfq_data for the device this group acts upon. | |
691 | * @async_bfqq: array of async queues for all the tasks belonging to | |
692 | * the group, one queue per ioprio value per ioprio_class, | |
693 | * except for the idle class that has only one queue. | |
694 | * @async_idle_bfqq: async queue for the idle class (ioprio is ignored). | |
695 | * @my_entity: pointer to @entity, %NULL for the toplevel group; used | |
696 | * to avoid too many special cases during group creation/ | |
697 | * migration. | |
698 | * @stats: stats for this bfqg. | |
699 | * | |
700 | * Each (device, cgroup) pair has its own bfq_group, i.e., for each cgroup | |
701 | * there is a set of bfq_groups, each one collecting the lower-level | |
702 | * entities belonging to the group that are acting on the same device. | |
703 | * | |
704 | * Locking works as follows: | |
705 | * o @bfqd is protected by the queue lock, RCU is used to access it | |
706 | * from the readers. | |
707 | * o All the other fields are protected by the @bfqd queue lock. | |
708 | */ | |
709 | struct bfq_group { | |
710 | /* must be the first member */ | |
711 | struct blkg_policy_data pd; | |
712 | ||
713 | struct bfq_entity entity; | |
714 | struct bfq_sched_data sched_data; | |
715 | ||
716 | void *bfqd; | |
717 | ||
718 | struct bfq_queue *async_bfqq[2][IOPRIO_BE_NR]; | |
719 | struct bfq_queue *async_idle_bfqq; | |
720 | ||
721 | struct bfq_entity *my_entity; | |
722 | ||
723 | struct bfqg_stats stats; | |
724 | }; | |
725 | ||
726 | #else | |
727 | struct bfq_group { | |
728 | struct bfq_sched_data sched_data; | |
729 | ||
730 | struct bfq_queue *async_bfqq[2][IOPRIO_BE_NR]; | |
731 | struct bfq_queue *async_idle_bfqq; | |
732 | ||
733 | struct rb_root rq_pos_tree; | |
734 | }; | |
735 | #endif | |
736 | ||
aee69d78 PV |
737 | static struct bfq_queue *bfq_entity_to_bfqq(struct bfq_entity *entity); |
738 | ||
e21b7a0b AA |
739 | static unsigned int bfq_class_idx(struct bfq_entity *entity) |
740 | { | |
741 | struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); | |
742 | ||
743 | return bfqq ? bfqq->ioprio_class - 1 : | |
744 | BFQ_DEFAULT_GRP_CLASS - 1; | |
745 | } | |
746 | ||
aee69d78 PV |
747 | static struct bfq_service_tree * |
748 | bfq_entity_service_tree(struct bfq_entity *entity) | |
749 | { | |
750 | struct bfq_sched_data *sched_data = entity->sched_data; | |
e21b7a0b | 751 | unsigned int idx = bfq_class_idx(entity); |
aee69d78 PV |
752 | |
753 | return sched_data->service_tree + idx; | |
754 | } | |
755 | ||
756 | static struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync) | |
757 | { | |
758 | return bic->bfqq[is_sync]; | |
759 | } | |
760 | ||
761 | static void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, | |
762 | bool is_sync) | |
763 | { | |
764 | bic->bfqq[is_sync] = bfqq; | |
765 | } | |
766 | ||
767 | static struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic) | |
768 | { | |
769 | return bic->icq.q->elevator->elevator_data; | |
770 | } | |
771 | ||
772 | static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio); | |
773 | static void bfq_put_queue(struct bfq_queue *bfqq); | |
774 | static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd, | |
775 | struct bio *bio, bool is_sync, | |
776 | struct bfq_io_cq *bic); | |
44e44a1b PV |
777 | static void bfq_end_wr_async_queues(struct bfq_data *bfqd, |
778 | struct bfq_group *bfqg); | |
e21b7a0b | 779 | static void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg); |
aee69d78 PV |
780 | static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq); |
781 | ||
aee69d78 PV |
782 | /* Expiration time of sync (0) and async (1) requests, in ns. */ |
783 | static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 }; | |
784 | ||
785 | /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */ | |
786 | static const int bfq_back_max = 16 * 1024; | |
787 | ||
788 | /* Penalty of a backwards seek, in number of sectors. */ | |
789 | static const int bfq_back_penalty = 2; | |
790 | ||
791 | /* Idling period duration, in ns. */ | |
792 | static u64 bfq_slice_idle = NSEC_PER_SEC / 125; | |
793 | ||
794 | /* Minimum number of assigned budgets for which stats are safe to compute. */ | |
795 | static const int bfq_stats_min_budgets = 194; | |
796 | ||
797 | /* Default maximum budget values, in sectors and number of requests. */ | |
798 | static const int bfq_default_max_budget = 16 * 1024; | |
799 | ||
c074170e PV |
800 | /* |
801 | * Async to sync throughput distribution is controlled as follows: | |
802 | * when an async request is served, the entity is charged the number | |
803 | * of sectors of the request, multiplied by the factor below | |
804 | */ | |
805 | static const int bfq_async_charge_factor = 10; | |
806 | ||
aee69d78 PV |
807 | /* Default timeout values, in jiffies, approximating CFQ defaults. */ |
808 | static const int bfq_timeout = HZ / 8; | |
809 | ||
810 | static struct kmem_cache *bfq_pool; | |
811 | ||
ab0e43e9 | 812 | /* Below this threshold (in ns), we consider thinktime immediate. */ |
aee69d78 PV |
813 | #define BFQ_MIN_TT (2 * NSEC_PER_MSEC) |
814 | ||
815 | /* hw_tag detection: parallel requests threshold and min samples needed. */ | |
816 | #define BFQ_HW_QUEUE_THRESHOLD 4 | |
817 | #define BFQ_HW_QUEUE_SAMPLES 32 | |
818 | ||
819 | #define BFQQ_SEEK_THR (sector_t)(8 * 100) | |
820 | #define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32) | |
821 | #define BFQQ_CLOSE_THR (sector_t)(8 * 1024) | |
822 | #define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 32/8) | |
823 | ||
ab0e43e9 PV |
824 | /* Min number of samples required to perform peak-rate update */ |
825 | #define BFQ_RATE_MIN_SAMPLES 32 | |
826 | /* Min observation time interval required to perform a peak-rate update (ns) */ | |
827 | #define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC) | |
828 | /* Target observation time interval for a peak-rate update (ns) */ | |
829 | #define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC | |
aee69d78 PV |
830 | |
831 | /* Shift used for peak rate fixed precision calculations. */ | |
832 | #define BFQ_RATE_SHIFT 16 | |
833 | ||
44e44a1b PV |
834 | /* |
835 | * By default, BFQ computes the duration of the weight raising for | |
836 | * interactive applications automatically, using the following formula: | |
837 | * duration = (R / r) * T, where r is the peak rate of the device, and | |
838 | * R and T are two reference parameters. | |
839 | * In particular, R is the peak rate of the reference device (see below), | |
840 | * and T is a reference time: given the systems that are likely to be | |
841 | * installed on the reference device according to its speed class, T is | |
842 | * about the maximum time needed, under BFQ and while reading two files in | |
843 | * parallel, to load typical large applications on these systems. | |
844 | * In practice, the slower/faster the device at hand is, the more/less it | |
845 | * takes to load applications with respect to the reference device. | |
846 | * Accordingly, the longer/shorter BFQ grants weight raising to interactive | |
847 | * applications. | |
848 | * | |
849 | * BFQ uses four different reference pairs (R, T), depending on: | |
850 | * . whether the device is rotational or non-rotational; | |
851 | * . whether the device is slow, such as old or portable HDDs, as well as | |
852 | * SD cards, or fast, such as newer HDDs and SSDs. | |
853 | * | |
854 | * The device's speed class is dynamically (re)detected in | |
855 | * bfq_update_peak_rate() every time the estimated peak rate is updated. | |
856 | * | |
857 | * In the following definitions, R_slow[0]/R_fast[0] and | |
858 | * T_slow[0]/T_fast[0] are the reference values for a slow/fast | |
859 | * rotational device, whereas R_slow[1]/R_fast[1] and | |
860 | * T_slow[1]/T_fast[1] are the reference values for a slow/fast | |
861 | * non-rotational device. Finally, device_speed_thresh are the | |
862 | * thresholds used to switch between speed classes. The reference | |
863 | * rates are not the actual peak rates of the devices used as a | |
864 | * reference, but slightly lower values. The reason for using these | |
865 | * slightly lower values is that the peak-rate estimator tends to | |
866 | * yield slightly lower values than the actual peak rate (it can yield | |
867 | * the actual peak rate only if there is only one process doing I/O, | |
868 | * and the process does sequential I/O). | |
869 | * | |
870 | * Both the reference peak rates and the thresholds are measured in | |
871 | * sectors/usec, left-shifted by BFQ_RATE_SHIFT. | |
872 | */ | |
873 | static int R_slow[2] = {1000, 10700}; | |
874 | static int R_fast[2] = {14000, 33000}; | |
875 | /* | |
876 | * To improve readability, a conversion function is used to initialize the | |
877 | * following arrays, which entails that they can be initialized only in a | |
878 | * function. | |
879 | */ | |
880 | static int T_slow[2]; | |
881 | static int T_fast[2]; | |
882 | static int device_speed_thresh[2]; | |
883 | ||
aee69d78 PV |
884 | #define BFQ_SERVICE_TREE_INIT ((struct bfq_service_tree) \ |
885 | { RB_ROOT, RB_ROOT, NULL, NULL, 0, 0 }) | |
886 | ||
887 | #define RQ_BIC(rq) ((struct bfq_io_cq *) (rq)->elv.priv[0]) | |
888 | #define RQ_BFQQ(rq) ((rq)->elv.priv[1]) | |
889 | ||
890 | /** | |
891 | * icq_to_bic - convert iocontext queue structure to bfq_io_cq. | |
892 | * @icq: the iocontext queue. | |
893 | */ | |
894 | static struct bfq_io_cq *icq_to_bic(struct io_cq *icq) | |
895 | { | |
896 | /* bic->icq is the first member, %NULL will convert to %NULL */ | |
897 | return container_of(icq, struct bfq_io_cq, icq); | |
898 | } | |
899 | ||
900 | /** | |
901 | * bfq_bic_lookup - search into @ioc a bic associated to @bfqd. | |
902 | * @bfqd: the lookup key. | |
903 | * @ioc: the io_context of the process doing I/O. | |
904 | * @q: the request queue. | |
905 | */ | |
906 | static struct bfq_io_cq *bfq_bic_lookup(struct bfq_data *bfqd, | |
907 | struct io_context *ioc, | |
908 | struct request_queue *q) | |
909 | { | |
910 | if (ioc) { | |
911 | unsigned long flags; | |
912 | struct bfq_io_cq *icq; | |
913 | ||
914 | spin_lock_irqsave(q->queue_lock, flags); | |
915 | icq = icq_to_bic(ioc_lookup_icq(ioc, q)); | |
916 | spin_unlock_irqrestore(q->queue_lock, flags); | |
917 | ||
918 | return icq; | |
919 | } | |
920 | ||
921 | return NULL; | |
922 | } | |
923 | ||
924 | /* | |
e21b7a0b AA |
925 | * Scheduler run of queue, if there are requests pending and no one in the |
926 | * driver that will restart queueing. | |
927 | */ | |
928 | static void bfq_schedule_dispatch(struct bfq_data *bfqd) | |
929 | { | |
930 | if (bfqd->queued != 0) { | |
931 | bfq_log(bfqd, "schedule dispatch"); | |
932 | blk_mq_run_hw_queues(bfqd->queue, true); | |
933 | } | |
934 | } | |
935 | ||
936 | /** | |
937 | * bfq_gt - compare two timestamps. | |
938 | * @a: first ts. | |
939 | * @b: second ts. | |
940 | * | |
941 | * Return @a > @b, dealing with wrapping correctly. | |
942 | */ | |
943 | static int bfq_gt(u64 a, u64 b) | |
944 | { | |
945 | return (s64)(a - b) > 0; | |
946 | } | |
947 | ||
948 | static struct bfq_entity *bfq_root_active_entity(struct rb_root *tree) | |
949 | { | |
950 | struct rb_node *node = tree->rb_node; | |
951 | ||
952 | return rb_entry(node, struct bfq_entity, rb_node); | |
953 | } | |
954 | ||
955 | static struct bfq_entity *bfq_lookup_next_entity(struct bfq_sched_data *sd); | |
956 | ||
957 | static bool bfq_update_parent_budget(struct bfq_entity *next_in_service); | |
958 | ||
959 | /** | |
960 | * bfq_update_next_in_service - update sd->next_in_service | |
961 | * @sd: sched_data for which to perform the update. | |
962 | * @new_entity: if not NULL, pointer to the entity whose activation, | |
963 | * requeueing or repositionig triggered the invocation of | |
964 | * this function. | |
965 | * | |
966 | * This function is called to update sd->next_in_service, which, in | |
967 | * its turn, may change as a consequence of the insertion or | |
968 | * extraction of an entity into/from one of the active trees of | |
969 | * sd. These insertions/extractions occur as a consequence of | |
970 | * activations/deactivations of entities, with some activations being | |
971 | * 'true' activations, and other activations being requeueings (i.e., | |
972 | * implementing the second, requeueing phase of the mechanism used to | |
973 | * reposition an entity in its active tree; see comments on | |
974 | * __bfq_activate_entity and __bfq_requeue_entity for details). In | |
975 | * both the last two activation sub-cases, new_entity points to the | |
976 | * just activated or requeued entity. | |
977 | * | |
978 | * Returns true if sd->next_in_service changes in such a way that | |
979 | * entity->parent may become the next_in_service for its parent | |
980 | * entity. | |
aee69d78 | 981 | */ |
e21b7a0b AA |
982 | static bool bfq_update_next_in_service(struct bfq_sched_data *sd, |
983 | struct bfq_entity *new_entity) | |
984 | { | |
985 | struct bfq_entity *next_in_service = sd->next_in_service; | |
986 | bool parent_sched_may_change = false; | |
987 | ||
988 | /* | |
989 | * If this update is triggered by the activation, requeueing | |
990 | * or repositiong of an entity that does not coincide with | |
991 | * sd->next_in_service, then a full lookup in the active tree | |
992 | * can be avoided. In fact, it is enough to check whether the | |
993 | * just-modified entity has a higher priority than | |
994 | * sd->next_in_service, or, even if it has the same priority | |
995 | * as sd->next_in_service, is eligible and has a lower virtual | |
996 | * finish time than sd->next_in_service. If this compound | |
997 | * condition holds, then the new entity becomes the new | |
998 | * next_in_service. Otherwise no change is needed. | |
999 | */ | |
1000 | if (new_entity && new_entity != sd->next_in_service) { | |
1001 | /* | |
1002 | * Flag used to decide whether to replace | |
1003 | * sd->next_in_service with new_entity. Tentatively | |
1004 | * set to true, and left as true if | |
1005 | * sd->next_in_service is NULL. | |
1006 | */ | |
1007 | bool replace_next = true; | |
1008 | ||
1009 | /* | |
1010 | * If there is already a next_in_service candidate | |
1011 | * entity, then compare class priorities or timestamps | |
1012 | * to decide whether to replace sd->service_tree with | |
1013 | * new_entity. | |
1014 | */ | |
1015 | if (next_in_service) { | |
1016 | unsigned int new_entity_class_idx = | |
1017 | bfq_class_idx(new_entity); | |
1018 | struct bfq_service_tree *st = | |
1019 | sd->service_tree + new_entity_class_idx; | |
1020 | ||
1021 | /* | |
1022 | * For efficiency, evaluate the most likely | |
1023 | * sub-condition first. | |
1024 | */ | |
1025 | replace_next = | |
1026 | (new_entity_class_idx == | |
1027 | bfq_class_idx(next_in_service) | |
1028 | && | |
1029 | !bfq_gt(new_entity->start, st->vtime) | |
1030 | && | |
1031 | bfq_gt(next_in_service->finish, | |
1032 | new_entity->finish)) | |
1033 | || | |
1034 | new_entity_class_idx < | |
1035 | bfq_class_idx(next_in_service); | |
1036 | } | |
1037 | ||
1038 | if (replace_next) | |
1039 | next_in_service = new_entity; | |
1040 | } else /* invoked because of a deactivation: lookup needed */ | |
1041 | next_in_service = bfq_lookup_next_entity(sd); | |
1042 | ||
1043 | if (next_in_service) { | |
1044 | parent_sched_may_change = !sd->next_in_service || | |
1045 | bfq_update_parent_budget(next_in_service); | |
1046 | } | |
1047 | ||
1048 | sd->next_in_service = next_in_service; | |
1049 | ||
1050 | if (!next_in_service) | |
1051 | return parent_sched_may_change; | |
1052 | ||
1053 | return parent_sched_may_change; | |
1054 | } | |
1055 | ||
1056 | #ifdef CONFIG_BFQ_GROUP_IOSCHED | |
1057 | /* both next loops stop at one of the child entities of the root group */ | |
aee69d78 | 1058 | #define for_each_entity(entity) \ |
e21b7a0b | 1059 | for (; entity ; entity = entity->parent) |
aee69d78 | 1060 | |
e21b7a0b AA |
1061 | /* |
1062 | * For each iteration, compute parent in advance, so as to be safe if | |
1063 | * entity is deallocated during the iteration. Such a deallocation may | |
1064 | * happen as a consequence of a bfq_put_queue that frees the bfq_queue | |
1065 | * containing entity. | |
1066 | */ | |
aee69d78 | 1067 | #define for_each_entity_safe(entity, parent) \ |
e21b7a0b | 1068 | for (; entity && ({ parent = entity->parent; 1; }); entity = parent) |
aee69d78 | 1069 | |
e21b7a0b AA |
1070 | /* |
1071 | * Returns true if this budget changes may let next_in_service->parent | |
1072 | * become the next_in_service entity for its parent entity. | |
1073 | */ | |
1074 | static bool bfq_update_parent_budget(struct bfq_entity *next_in_service) | |
aee69d78 | 1075 | { |
e21b7a0b AA |
1076 | struct bfq_entity *bfqg_entity; |
1077 | struct bfq_group *bfqg; | |
1078 | struct bfq_sched_data *group_sd; | |
1079 | bool ret = false; | |
1080 | ||
1081 | group_sd = next_in_service->sched_data; | |
1082 | ||
1083 | bfqg = container_of(group_sd, struct bfq_group, sched_data); | |
1084 | /* | |
1085 | * bfq_group's my_entity field is not NULL only if the group | |
1086 | * is not the root group. We must not touch the root entity | |
1087 | * as it must never become an in-service entity. | |
1088 | */ | |
1089 | bfqg_entity = bfqg->my_entity; | |
1090 | if (bfqg_entity) { | |
1091 | if (bfqg_entity->budget > next_in_service->budget) | |
1092 | ret = true; | |
1093 | bfqg_entity->budget = next_in_service->budget; | |
1094 | } | |
1095 | ||
1096 | return ret; | |
1097 | } | |
1098 | ||
1099 | /* | |
1100 | * This function tells whether entity stops being a candidate for next | |
1101 | * service, according to the following logic. | |
1102 | * | |
1103 | * This function is invoked for an entity that is about to be set in | |
1104 | * service. If such an entity is a queue, then the entity is no longer | |
1105 | * a candidate for next service (i.e, a candidate entity to serve | |
1106 | * after the in-service entity is expired). The function then returns | |
1107 | * true. | |
1108 | */ | |
1109 | static bool bfq_no_longer_next_in_service(struct bfq_entity *entity) | |
1110 | { | |
1111 | if (bfq_entity_to_bfqq(entity)) | |
1112 | return true; | |
1113 | ||
1114 | return false; | |
aee69d78 PV |
1115 | } |
1116 | ||
e21b7a0b AA |
1117 | #else /* CONFIG_BFQ_GROUP_IOSCHED */ |
1118 | /* | |
1119 | * Next two macros are fake loops when cgroups support is not | |
1120 | * enabled. I fact, in such a case, there is only one level to go up | |
1121 | * (to reach the root group). | |
1122 | */ | |
1123 | #define for_each_entity(entity) \ | |
1124 | for (; entity ; entity = NULL) | |
1125 | ||
1126 | #define for_each_entity_safe(entity, parent) \ | |
1127 | for (parent = NULL; entity ; entity = parent) | |
1128 | ||
1129 | static bool bfq_update_parent_budget(struct bfq_entity *next_in_service) | |
aee69d78 | 1130 | { |
e21b7a0b | 1131 | return false; |
aee69d78 PV |
1132 | } |
1133 | ||
e21b7a0b | 1134 | static bool bfq_no_longer_next_in_service(struct bfq_entity *entity) |
aee69d78 | 1135 | { |
e21b7a0b | 1136 | return true; |
aee69d78 PV |
1137 | } |
1138 | ||
e21b7a0b AA |
1139 | #endif /* CONFIG_BFQ_GROUP_IOSCHED */ |
1140 | ||
aee69d78 PV |
1141 | /* |
1142 | * Shift for timestamp calculations. This actually limits the maximum | |
1143 | * service allowed in one timestamp delta (small shift values increase it), | |
1144 | * the maximum total weight that can be used for the queues in the system | |
1145 | * (big shift values increase it), and the period of virtual time | |
1146 | * wraparounds. | |
1147 | */ | |
1148 | #define WFQ_SERVICE_SHIFT 22 | |
1149 | ||
aee69d78 PV |
1150 | static struct bfq_queue *bfq_entity_to_bfqq(struct bfq_entity *entity) |
1151 | { | |
1152 | struct bfq_queue *bfqq = NULL; | |
1153 | ||
1154 | if (!entity->my_sched_data) | |
1155 | bfqq = container_of(entity, struct bfq_queue, entity); | |
1156 | ||
1157 | return bfqq; | |
1158 | } | |
1159 | ||
1160 | ||
1161 | /** | |
1162 | * bfq_delta - map service into the virtual time domain. | |
1163 | * @service: amount of service. | |
1164 | * @weight: scale factor (weight of an entity or weight sum). | |
1165 | */ | |
1166 | static u64 bfq_delta(unsigned long service, unsigned long weight) | |
1167 | { | |
1168 | u64 d = (u64)service << WFQ_SERVICE_SHIFT; | |
1169 | ||
1170 | do_div(d, weight); | |
1171 | return d; | |
1172 | } | |
1173 | ||
1174 | /** | |
1175 | * bfq_calc_finish - assign the finish time to an entity. | |
1176 | * @entity: the entity to act upon. | |
1177 | * @service: the service to be charged to the entity. | |
1178 | */ | |
1179 | static void bfq_calc_finish(struct bfq_entity *entity, unsigned long service) | |
1180 | { | |
1181 | struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); | |
1182 | ||
1183 | entity->finish = entity->start + | |
1184 | bfq_delta(service, entity->weight); | |
1185 | ||
1186 | if (bfqq) { | |
1187 | bfq_log_bfqq(bfqq->bfqd, bfqq, | |
1188 | "calc_finish: serv %lu, w %d", | |
1189 | service, entity->weight); | |
1190 | bfq_log_bfqq(bfqq->bfqd, bfqq, | |
1191 | "calc_finish: start %llu, finish %llu, delta %llu", | |
1192 | entity->start, entity->finish, | |
1193 | bfq_delta(service, entity->weight)); | |
1194 | } | |
1195 | } | |
1196 | ||
1197 | /** | |
1198 | * bfq_entity_of - get an entity from a node. | |
1199 | * @node: the node field of the entity. | |
1200 | * | |
1201 | * Convert a node pointer to the relative entity. This is used only | |
1202 | * to simplify the logic of some functions and not as the generic | |
1203 | * conversion mechanism because, e.g., in the tree walking functions, | |
1204 | * the check for a %NULL value would be redundant. | |
1205 | */ | |
1206 | static struct bfq_entity *bfq_entity_of(struct rb_node *node) | |
1207 | { | |
1208 | struct bfq_entity *entity = NULL; | |
1209 | ||
1210 | if (node) | |
1211 | entity = rb_entry(node, struct bfq_entity, rb_node); | |
1212 | ||
1213 | return entity; | |
1214 | } | |
1215 | ||
1216 | /** | |
1217 | * bfq_extract - remove an entity from a tree. | |
1218 | * @root: the tree root. | |
1219 | * @entity: the entity to remove. | |
1220 | */ | |
1221 | static void bfq_extract(struct rb_root *root, struct bfq_entity *entity) | |
1222 | { | |
1223 | entity->tree = NULL; | |
1224 | rb_erase(&entity->rb_node, root); | |
1225 | } | |
1226 | ||
1227 | /** | |
1228 | * bfq_idle_extract - extract an entity from the idle tree. | |
1229 | * @st: the service tree of the owning @entity. | |
1230 | * @entity: the entity being removed. | |
1231 | */ | |
1232 | static void bfq_idle_extract(struct bfq_service_tree *st, | |
1233 | struct bfq_entity *entity) | |
1234 | { | |
1235 | struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); | |
1236 | struct rb_node *next; | |
1237 | ||
1238 | if (entity == st->first_idle) { | |
1239 | next = rb_next(&entity->rb_node); | |
1240 | st->first_idle = bfq_entity_of(next); | |
1241 | } | |
1242 | ||
1243 | if (entity == st->last_idle) { | |
1244 | next = rb_prev(&entity->rb_node); | |
1245 | st->last_idle = bfq_entity_of(next); | |
1246 | } | |
1247 | ||
1248 | bfq_extract(&st->idle, entity); | |
1249 | ||
1250 | if (bfqq) | |
1251 | list_del(&bfqq->bfqq_list); | |
1252 | } | |
1253 | ||
1254 | /** | |
1255 | * bfq_insert - generic tree insertion. | |
1256 | * @root: tree root. | |
1257 | * @entity: entity to insert. | |
1258 | * | |
1259 | * This is used for the idle and the active tree, since they are both | |
1260 | * ordered by finish time. | |
1261 | */ | |
1262 | static void bfq_insert(struct rb_root *root, struct bfq_entity *entity) | |
1263 | { | |
1264 | struct bfq_entity *entry; | |
1265 | struct rb_node **node = &root->rb_node; | |
1266 | struct rb_node *parent = NULL; | |
1267 | ||
1268 | while (*node) { | |
1269 | parent = *node; | |
1270 | entry = rb_entry(parent, struct bfq_entity, rb_node); | |
1271 | ||
1272 | if (bfq_gt(entry->finish, entity->finish)) | |
1273 | node = &parent->rb_left; | |
1274 | else | |
1275 | node = &parent->rb_right; | |
1276 | } | |
1277 | ||
1278 | rb_link_node(&entity->rb_node, parent, node); | |
1279 | rb_insert_color(&entity->rb_node, root); | |
1280 | ||
1281 | entity->tree = root; | |
1282 | } | |
1283 | ||
1284 | /** | |
1285 | * bfq_update_min - update the min_start field of a entity. | |
1286 | * @entity: the entity to update. | |
1287 | * @node: one of its children. | |
1288 | * | |
1289 | * This function is called when @entity may store an invalid value for | |
1290 | * min_start due to updates to the active tree. The function assumes | |
1291 | * that the subtree rooted at @node (which may be its left or its right | |
1292 | * child) has a valid min_start value. | |
1293 | */ | |
1294 | static void bfq_update_min(struct bfq_entity *entity, struct rb_node *node) | |
1295 | { | |
1296 | struct bfq_entity *child; | |
1297 | ||
1298 | if (node) { | |
1299 | child = rb_entry(node, struct bfq_entity, rb_node); | |
1300 | if (bfq_gt(entity->min_start, child->min_start)) | |
1301 | entity->min_start = child->min_start; | |
1302 | } | |
1303 | } | |
1304 | ||
1305 | /** | |
1306 | * bfq_update_active_node - recalculate min_start. | |
1307 | * @node: the node to update. | |
1308 | * | |
1309 | * @node may have changed position or one of its children may have moved, | |
1310 | * this function updates its min_start value. The left and right subtrees | |
1311 | * are assumed to hold a correct min_start value. | |
1312 | */ | |
1313 | static void bfq_update_active_node(struct rb_node *node) | |
1314 | { | |
1315 | struct bfq_entity *entity = rb_entry(node, struct bfq_entity, rb_node); | |
1316 | ||
1317 | entity->min_start = entity->start; | |
1318 | bfq_update_min(entity, node->rb_right); | |
1319 | bfq_update_min(entity, node->rb_left); | |
1320 | } | |
1321 | ||
1322 | /** | |
1323 | * bfq_update_active_tree - update min_start for the whole active tree. | |
1324 | * @node: the starting node. | |
1325 | * | |
1326 | * @node must be the deepest modified node after an update. This function | |
1327 | * updates its min_start using the values held by its children, assuming | |
1328 | * that they did not change, and then updates all the nodes that may have | |
1329 | * changed in the path to the root. The only nodes that may have changed | |
1330 | * are the ones in the path or their siblings. | |
1331 | */ | |
1332 | static void bfq_update_active_tree(struct rb_node *node) | |
1333 | { | |
1334 | struct rb_node *parent; | |
1335 | ||
1336 | up: | |
1337 | bfq_update_active_node(node); | |
1338 | ||
1339 | parent = rb_parent(node); | |
1340 | if (!parent) | |
1341 | return; | |
1342 | ||
1343 | if (node == parent->rb_left && parent->rb_right) | |
1344 | bfq_update_active_node(parent->rb_right); | |
1345 | else if (parent->rb_left) | |
1346 | bfq_update_active_node(parent->rb_left); | |
1347 | ||
1348 | node = parent; | |
1349 | goto up; | |
1350 | } | |
1351 | ||
1352 | /** | |
1353 | * bfq_active_insert - insert an entity in the active tree of its | |
1354 | * group/device. | |
1355 | * @st: the service tree of the entity. | |
1356 | * @entity: the entity being inserted. | |
1357 | * | |
1358 | * The active tree is ordered by finish time, but an extra key is kept | |
1359 | * per each node, containing the minimum value for the start times of | |
1360 | * its children (and the node itself), so it's possible to search for | |
1361 | * the eligible node with the lowest finish time in logarithmic time. | |
1362 | */ | |
1363 | static void bfq_active_insert(struct bfq_service_tree *st, | |
1364 | struct bfq_entity *entity) | |
1365 | { | |
1366 | struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); | |
1367 | struct rb_node *node = &entity->rb_node; | |
e21b7a0b AA |
1368 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
1369 | struct bfq_sched_data *sd = NULL; | |
1370 | struct bfq_group *bfqg = NULL; | |
1371 | struct bfq_data *bfqd = NULL; | |
1372 | #endif | |
aee69d78 PV |
1373 | |
1374 | bfq_insert(&st->active, entity); | |
1375 | ||
1376 | if (node->rb_left) | |
1377 | node = node->rb_left; | |
1378 | else if (node->rb_right) | |
1379 | node = node->rb_right; | |
1380 | ||
1381 | bfq_update_active_tree(node); | |
1382 | ||
e21b7a0b AA |
1383 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
1384 | sd = entity->sched_data; | |
1385 | bfqg = container_of(sd, struct bfq_group, sched_data); | |
1386 | bfqd = (struct bfq_data *)bfqg->bfqd; | |
1387 | #endif | |
aee69d78 PV |
1388 | if (bfqq) |
1389 | list_add(&bfqq->bfqq_list, &bfqq->bfqd->active_list); | |
1390 | } | |
1391 | ||
1392 | /** | |
1393 | * bfq_ioprio_to_weight - calc a weight from an ioprio. | |
1394 | * @ioprio: the ioprio value to convert. | |
1395 | */ | |
1396 | static unsigned short bfq_ioprio_to_weight(int ioprio) | |
1397 | { | |
1398 | return (IOPRIO_BE_NR - ioprio) * BFQ_WEIGHT_CONVERSION_COEFF; | |
1399 | } | |
1400 | ||
1401 | /** | |
1402 | * bfq_weight_to_ioprio - calc an ioprio from a weight. | |
1403 | * @weight: the weight value to convert. | |
1404 | * | |
1405 | * To preserve as much as possible the old only-ioprio user interface, | |
1406 | * 0 is used as an escape ioprio value for weights (numerically) equal or | |
1407 | * larger than IOPRIO_BE_NR * BFQ_WEIGHT_CONVERSION_COEFF. | |
1408 | */ | |
1409 | static unsigned short bfq_weight_to_ioprio(int weight) | |
1410 | { | |
1411 | return max_t(int, 0, | |
1412 | IOPRIO_BE_NR * BFQ_WEIGHT_CONVERSION_COEFF - weight); | |
1413 | } | |
1414 | ||
1415 | static void bfq_get_entity(struct bfq_entity *entity) | |
1416 | { | |
1417 | struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); | |
1418 | ||
1419 | if (bfqq) { | |
1420 | bfqq->ref++; | |
1421 | bfq_log_bfqq(bfqq->bfqd, bfqq, "get_entity: %p %d", | |
1422 | bfqq, bfqq->ref); | |
1423 | } | |
1424 | } | |
1425 | ||
1426 | /** | |
1427 | * bfq_find_deepest - find the deepest node that an extraction can modify. | |
1428 | * @node: the node being removed. | |
1429 | * | |
1430 | * Do the first step of an extraction in an rb tree, looking for the | |
1431 | * node that will replace @node, and returning the deepest node that | |
1432 | * the following modifications to the tree can touch. If @node is the | |
1433 | * last node in the tree return %NULL. | |
1434 | */ | |
1435 | static struct rb_node *bfq_find_deepest(struct rb_node *node) | |
1436 | { | |
1437 | struct rb_node *deepest; | |
1438 | ||
1439 | if (!node->rb_right && !node->rb_left) | |
1440 | deepest = rb_parent(node); | |
1441 | else if (!node->rb_right) | |
1442 | deepest = node->rb_left; | |
1443 | else if (!node->rb_left) | |
1444 | deepest = node->rb_right; | |
1445 | else { | |
1446 | deepest = rb_next(node); | |
1447 | if (deepest->rb_right) | |
1448 | deepest = deepest->rb_right; | |
1449 | else if (rb_parent(deepest) != node) | |
1450 | deepest = rb_parent(deepest); | |
1451 | } | |
1452 | ||
1453 | return deepest; | |
1454 | } | |
1455 | ||
1456 | /** | |
1457 | * bfq_active_extract - remove an entity from the active tree. | |
1458 | * @st: the service_tree containing the tree. | |
1459 | * @entity: the entity being removed. | |
1460 | */ | |
1461 | static void bfq_active_extract(struct bfq_service_tree *st, | |
1462 | struct bfq_entity *entity) | |
1463 | { | |
1464 | struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); | |
1465 | struct rb_node *node; | |
e21b7a0b AA |
1466 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
1467 | struct bfq_sched_data *sd = NULL; | |
1468 | struct bfq_group *bfqg = NULL; | |
1469 | struct bfq_data *bfqd = NULL; | |
1470 | #endif | |
aee69d78 PV |
1471 | |
1472 | node = bfq_find_deepest(&entity->rb_node); | |
1473 | bfq_extract(&st->active, entity); | |
1474 | ||
1475 | if (node) | |
1476 | bfq_update_active_tree(node); | |
1477 | ||
e21b7a0b AA |
1478 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
1479 | sd = entity->sched_data; | |
1480 | bfqg = container_of(sd, struct bfq_group, sched_data); | |
1481 | bfqd = (struct bfq_data *)bfqg->bfqd; | |
1482 | #endif | |
aee69d78 PV |
1483 | if (bfqq) |
1484 | list_del(&bfqq->bfqq_list); | |
1485 | } | |
1486 | ||
1487 | /** | |
1488 | * bfq_idle_insert - insert an entity into the idle tree. | |
1489 | * @st: the service tree containing the tree. | |
1490 | * @entity: the entity to insert. | |
1491 | */ | |
1492 | static void bfq_idle_insert(struct bfq_service_tree *st, | |
1493 | struct bfq_entity *entity) | |
1494 | { | |
1495 | struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); | |
1496 | struct bfq_entity *first_idle = st->first_idle; | |
1497 | struct bfq_entity *last_idle = st->last_idle; | |
1498 | ||
1499 | if (!first_idle || bfq_gt(first_idle->finish, entity->finish)) | |
1500 | st->first_idle = entity; | |
1501 | if (!last_idle || bfq_gt(entity->finish, last_idle->finish)) | |
1502 | st->last_idle = entity; | |
1503 | ||
1504 | bfq_insert(&st->idle, entity); | |
1505 | ||
1506 | if (bfqq) | |
1507 | list_add(&bfqq->bfqq_list, &bfqq->bfqd->idle_list); | |
1508 | } | |
1509 | ||
1510 | /** | |
1511 | * bfq_forget_entity - do not consider entity any longer for scheduling | |
1512 | * @st: the service tree. | |
1513 | * @entity: the entity being removed. | |
1514 | * @is_in_service: true if entity is currently the in-service entity. | |
1515 | * | |
1516 | * Forget everything about @entity. In addition, if entity represents | |
1517 | * a queue, and the latter is not in service, then release the service | |
1518 | * reference to the queue (the one taken through bfq_get_entity). In | |
1519 | * fact, in this case, there is really no more service reference to | |
1520 | * the queue, as the latter is also outside any service tree. If, | |
1521 | * instead, the queue is in service, then __bfq_bfqd_reset_in_service | |
1522 | * will take care of putting the reference when the queue finally | |
1523 | * stops being served. | |
1524 | */ | |
1525 | static void bfq_forget_entity(struct bfq_service_tree *st, | |
1526 | struct bfq_entity *entity, | |
1527 | bool is_in_service) | |
1528 | { | |
1529 | struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); | |
1530 | ||
e21b7a0b | 1531 | entity->on_st = false; |
aee69d78 PV |
1532 | st->wsum -= entity->weight; |
1533 | if (bfqq && !is_in_service) | |
1534 | bfq_put_queue(bfqq); | |
1535 | } | |
1536 | ||
1537 | /** | |
1538 | * bfq_put_idle_entity - release the idle tree ref of an entity. | |
1539 | * @st: service tree for the entity. | |
1540 | * @entity: the entity being released. | |
1541 | */ | |
1542 | static void bfq_put_idle_entity(struct bfq_service_tree *st, | |
1543 | struct bfq_entity *entity) | |
1544 | { | |
1545 | bfq_idle_extract(st, entity); | |
1546 | bfq_forget_entity(st, entity, | |
1547 | entity == entity->sched_data->in_service_entity); | |
1548 | } | |
1549 | ||
1550 | /** | |
1551 | * bfq_forget_idle - update the idle tree if necessary. | |
1552 | * @st: the service tree to act upon. | |
1553 | * | |
1554 | * To preserve the global O(log N) complexity we only remove one entry here; | |
1555 | * as the idle tree will not grow indefinitely this can be done safely. | |
1556 | */ | |
1557 | static void bfq_forget_idle(struct bfq_service_tree *st) | |
1558 | { | |
1559 | struct bfq_entity *first_idle = st->first_idle; | |
1560 | struct bfq_entity *last_idle = st->last_idle; | |
1561 | ||
1562 | if (RB_EMPTY_ROOT(&st->active) && last_idle && | |
1563 | !bfq_gt(last_idle->finish, st->vtime)) { | |
1564 | /* | |
1565 | * Forget the whole idle tree, increasing the vtime past | |
1566 | * the last finish time of idle entities. | |
1567 | */ | |
1568 | st->vtime = last_idle->finish; | |
1569 | } | |
1570 | ||
1571 | if (first_idle && !bfq_gt(first_idle->finish, st->vtime)) | |
1572 | bfq_put_idle_entity(st, first_idle); | |
1573 | } | |
1574 | ||
1575 | static struct bfq_service_tree * | |
1576 | __bfq_entity_update_weight_prio(struct bfq_service_tree *old_st, | |
e21b7a0b | 1577 | struct bfq_entity *entity) |
aee69d78 PV |
1578 | { |
1579 | struct bfq_service_tree *new_st = old_st; | |
1580 | ||
1581 | if (entity->prio_changed) { | |
1582 | struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); | |
44e44a1b | 1583 | unsigned int prev_weight, new_weight; |
aee69d78 | 1584 | struct bfq_data *bfqd = NULL; |
e21b7a0b AA |
1585 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
1586 | struct bfq_sched_data *sd; | |
1587 | struct bfq_group *bfqg; | |
1588 | #endif | |
aee69d78 PV |
1589 | |
1590 | if (bfqq) | |
1591 | bfqd = bfqq->bfqd; | |
e21b7a0b AA |
1592 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
1593 | else { | |
1594 | sd = entity->my_sched_data; | |
1595 | bfqg = container_of(sd, struct bfq_group, sched_data); | |
1596 | bfqd = (struct bfq_data *)bfqg->bfqd; | |
1597 | } | |
1598 | #endif | |
aee69d78 PV |
1599 | |
1600 | old_st->wsum -= entity->weight; | |
1601 | ||
1602 | if (entity->new_weight != entity->orig_weight) { | |
1603 | if (entity->new_weight < BFQ_MIN_WEIGHT || | |
1604 | entity->new_weight > BFQ_MAX_WEIGHT) { | |
1605 | pr_crit("update_weight_prio: new_weight %d\n", | |
1606 | entity->new_weight); | |
1607 | if (entity->new_weight < BFQ_MIN_WEIGHT) | |
1608 | entity->new_weight = BFQ_MIN_WEIGHT; | |
1609 | else | |
1610 | entity->new_weight = BFQ_MAX_WEIGHT; | |
1611 | } | |
1612 | entity->orig_weight = entity->new_weight; | |
1613 | if (bfqq) | |
1614 | bfqq->ioprio = | |
1615 | bfq_weight_to_ioprio(entity->orig_weight); | |
1616 | } | |
1617 | ||
1618 | if (bfqq) | |
1619 | bfqq->ioprio_class = bfqq->new_ioprio_class; | |
1620 | entity->prio_changed = 0; | |
1621 | ||
1622 | /* | |
1623 | * NOTE: here we may be changing the weight too early, | |
1624 | * this will cause unfairness. The correct approach | |
1625 | * would have required additional complexity to defer | |
1626 | * weight changes to the proper time instants (i.e., | |
1627 | * when entity->finish <= old_st->vtime). | |
1628 | */ | |
1629 | new_st = bfq_entity_service_tree(entity); | |
1630 | ||
1631 | prev_weight = entity->weight; | |
44e44a1b PV |
1632 | new_weight = entity->orig_weight * |
1633 | (bfqq ? bfqq->wr_coeff : 1); | |
aee69d78 PV |
1634 | entity->weight = new_weight; |
1635 | ||
1636 | new_st->wsum += entity->weight; | |
1637 | ||
1638 | if (new_st != old_st) | |
1639 | entity->start = new_st->vtime; | |
1640 | } | |
1641 | ||
1642 | return new_st; | |
1643 | } | |
1644 | ||
e21b7a0b AA |
1645 | static void bfqg_stats_set_start_empty_time(struct bfq_group *bfqg); |
1646 | static struct bfq_group *bfqq_group(struct bfq_queue *bfqq); | |
1647 | ||
aee69d78 PV |
1648 | /** |
1649 | * bfq_bfqq_served - update the scheduler status after selection for | |
1650 | * service. | |
1651 | * @bfqq: the queue being served. | |
1652 | * @served: bytes to transfer. | |
1653 | * | |
1654 | * NOTE: this can be optimized, as the timestamps of upper level entities | |
1655 | * are synchronized every time a new bfqq is selected for service. By now, | |
1656 | * we keep it to better check consistency. | |
1657 | */ | |
1658 | static void bfq_bfqq_served(struct bfq_queue *bfqq, int served) | |
1659 | { | |
1660 | struct bfq_entity *entity = &bfqq->entity; | |
1661 | struct bfq_service_tree *st; | |
1662 | ||
1663 | for_each_entity(entity) { | |
1664 | st = bfq_entity_service_tree(entity); | |
1665 | ||
1666 | entity->service += served; | |
1667 | ||
1668 | st->vtime += bfq_delta(served, st->wsum); | |
1669 | bfq_forget_idle(st); | |
1670 | } | |
e21b7a0b | 1671 | bfqg_stats_set_start_empty_time(bfqq_group(bfqq)); |
aee69d78 PV |
1672 | bfq_log_bfqq(bfqq->bfqd, bfqq, "bfqq_served %d secs", served); |
1673 | } | |
1674 | ||
1675 | /** | |
c074170e PV |
1676 | * bfq_bfqq_charge_time - charge an amount of service equivalent to the length |
1677 | * of the time interval during which bfqq has been in | |
1678 | * service. | |
1679 | * @bfqd: the device | |
aee69d78 | 1680 | * @bfqq: the queue that needs a service update. |
c074170e | 1681 | * @time_ms: the amount of time during which the queue has received service |
aee69d78 | 1682 | * |
c074170e PV |
1683 | * If a queue does not consume its budget fast enough, then providing |
1684 | * the queue with service fairness may impair throughput, more or less | |
1685 | * severely. For this reason, queues that consume their budget slowly | |
1686 | * are provided with time fairness instead of service fairness. This | |
1687 | * goal is achieved through the BFQ scheduling engine, even if such an | |
1688 | * engine works in the service, and not in the time domain. The trick | |
1689 | * is charging these queues with an inflated amount of service, equal | |
1690 | * to the amount of service that they would have received during their | |
1691 | * service slot if they had been fast, i.e., if their requests had | |
1692 | * been dispatched at a rate equal to the estimated peak rate. | |
1693 | * | |
1694 | * It is worth noting that time fairness can cause important | |
1695 | * distortions in terms of bandwidth distribution, on devices with | |
1696 | * internal queueing. The reason is that I/O requests dispatched | |
1697 | * during the service slot of a queue may be served after that service | |
1698 | * slot is finished, and may have a total processing time loosely | |
1699 | * correlated with the duration of the service slot. This is | |
1700 | * especially true for short service slots. | |
aee69d78 | 1701 | */ |
c074170e PV |
1702 | static void bfq_bfqq_charge_time(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
1703 | unsigned long time_ms) | |
aee69d78 PV |
1704 | { |
1705 | struct bfq_entity *entity = &bfqq->entity; | |
c074170e PV |
1706 | int tot_serv_to_charge = entity->service; |
1707 | unsigned int timeout_ms = jiffies_to_msecs(bfq_timeout); | |
1708 | ||
1709 | if (time_ms > 0 && time_ms < timeout_ms) | |
1710 | tot_serv_to_charge = | |
1711 | (bfqd->bfq_max_budget * time_ms) / timeout_ms; | |
aee69d78 | 1712 | |
c074170e PV |
1713 | if (tot_serv_to_charge < entity->service) |
1714 | tot_serv_to_charge = entity->service; | |
aee69d78 | 1715 | |
c074170e PV |
1716 | /* Increase budget to avoid inconsistencies */ |
1717 | if (tot_serv_to_charge > entity->budget) | |
1718 | entity->budget = tot_serv_to_charge; | |
1719 | ||
1720 | bfq_bfqq_served(bfqq, | |
1721 | max_t(int, 0, tot_serv_to_charge - entity->service)); | |
aee69d78 PV |
1722 | } |
1723 | ||
e21b7a0b AA |
1724 | static void bfq_update_fin_time_enqueue(struct bfq_entity *entity, |
1725 | struct bfq_service_tree *st, | |
1726 | bool backshifted) | |
aee69d78 | 1727 | { |
44e44a1b PV |
1728 | struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
1729 | ||
aee69d78 PV |
1730 | st = __bfq_entity_update_weight_prio(st, entity); |
1731 | bfq_calc_finish(entity, entity->budget); | |
1732 | ||
1733 | /* | |
1734 | * If some queues enjoy backshifting for a while, then their | |
1735 | * (virtual) finish timestamps may happen to become lower and | |
1736 | * lower than the system virtual time. In particular, if | |
1737 | * these queues often happen to be idle for short time | |
1738 | * periods, and during such time periods other queues with | |
1739 | * higher timestamps happen to be busy, then the backshifted | |
1740 | * timestamps of the former queues can become much lower than | |
1741 | * the system virtual time. In fact, to serve the queues with | |
1742 | * higher timestamps while the ones with lower timestamps are | |
1743 | * idle, the system virtual time may be pushed-up to much | |
1744 | * higher values than the finish timestamps of the idle | |
1745 | * queues. As a consequence, the finish timestamps of all new | |
1746 | * or newly activated queues may end up being much larger than | |
1747 | * those of lucky queues with backshifted timestamps. The | |
1748 | * latter queues may then monopolize the device for a lot of | |
1749 | * time. This would simply break service guarantees. | |
1750 | * | |
1751 | * To reduce this problem, push up a little bit the | |
1752 | * backshifted timestamps of the queue associated with this | |
1753 | * entity (only a queue can happen to have the backshifted | |
1754 | * flag set): just enough to let the finish timestamp of the | |
1755 | * queue be equal to the current value of the system virtual | |
1756 | * time. This may introduce a little unfairness among queues | |
1757 | * with backshifted timestamps, but it does not break | |
1758 | * worst-case fairness guarantees. | |
44e44a1b PV |
1759 | * |
1760 | * As a special case, if bfqq is weight-raised, push up | |
1761 | * timestamps much less, to keep very low the probability that | |
1762 | * this push up causes the backshifted finish timestamps of | |
1763 | * weight-raised queues to become higher than the backshifted | |
1764 | * finish timestamps of non weight-raised queues. | |
aee69d78 PV |
1765 | */ |
1766 | if (backshifted && bfq_gt(st->vtime, entity->finish)) { | |
1767 | unsigned long delta = st->vtime - entity->finish; | |
1768 | ||
44e44a1b PV |
1769 | if (bfqq) |
1770 | delta /= bfqq->wr_coeff; | |
1771 | ||
aee69d78 PV |
1772 | entity->start += delta; |
1773 | entity->finish += delta; | |
1774 | } | |
1775 | ||
1776 | bfq_active_insert(st, entity); | |
1777 | } | |
1778 | ||
1779 | /** | |
e21b7a0b AA |
1780 | * __bfq_activate_entity - handle activation of entity. |
1781 | * @entity: the entity being activated. | |
1782 | * @non_blocking_wait_rq: true if entity was waiting for a request | |
1783 | * | |
1784 | * Called for a 'true' activation, i.e., if entity is not active and | |
1785 | * one of its children receives a new request. | |
1786 | * | |
1787 | * Basically, this function updates the timestamps of entity and | |
1788 | * inserts entity into its active tree, ater possible extracting it | |
1789 | * from its idle tree. | |
1790 | */ | |
1791 | static void __bfq_activate_entity(struct bfq_entity *entity, | |
1792 | bool non_blocking_wait_rq) | |
1793 | { | |
1794 | struct bfq_service_tree *st = bfq_entity_service_tree(entity); | |
1795 | bool backshifted = false; | |
1796 | unsigned long long min_vstart; | |
1797 | ||
1798 | /* See comments on bfq_fqq_update_budg_for_activation */ | |
1799 | if (non_blocking_wait_rq && bfq_gt(st->vtime, entity->finish)) { | |
1800 | backshifted = true; | |
1801 | min_vstart = entity->finish; | |
1802 | } else | |
1803 | min_vstart = st->vtime; | |
1804 | ||
1805 | if (entity->tree == &st->idle) { | |
1806 | /* | |
1807 | * Must be on the idle tree, bfq_idle_extract() will | |
1808 | * check for that. | |
1809 | */ | |
1810 | bfq_idle_extract(st, entity); | |
1811 | entity->start = bfq_gt(min_vstart, entity->finish) ? | |
1812 | min_vstart : entity->finish; | |
1813 | } else { | |
1814 | /* | |
1815 | * The finish time of the entity may be invalid, and | |
1816 | * it is in the past for sure, otherwise the queue | |
1817 | * would have been on the idle tree. | |
1818 | */ | |
1819 | entity->start = min_vstart; | |
1820 | st->wsum += entity->weight; | |
1821 | /* | |
1822 | * entity is about to be inserted into a service tree, | |
1823 | * and then set in service: get a reference to make | |
1824 | * sure entity does not disappear until it is no | |
1825 | * longer in service or scheduled for service. | |
1826 | */ | |
1827 | bfq_get_entity(entity); | |
1828 | ||
1829 | entity->on_st = true; | |
1830 | } | |
1831 | ||
1832 | bfq_update_fin_time_enqueue(entity, st, backshifted); | |
1833 | } | |
1834 | ||
1835 | /** | |
1836 | * __bfq_requeue_entity - handle requeueing or repositioning of an entity. | |
1837 | * @entity: the entity being requeued or repositioned. | |
1838 | * | |
1839 | * Requeueing is needed if this entity stops being served, which | |
1840 | * happens if a leaf descendant entity has expired. On the other hand, | |
1841 | * repositioning is needed if the next_inservice_entity for the child | |
1842 | * entity has changed. See the comments inside the function for | |
1843 | * details. | |
1844 | * | |
1845 | * Basically, this function: 1) removes entity from its active tree if | |
1846 | * present there, 2) updates the timestamps of entity and 3) inserts | |
1847 | * entity back into its active tree (in the new, right position for | |
1848 | * the new values of the timestamps). | |
1849 | */ | |
1850 | static void __bfq_requeue_entity(struct bfq_entity *entity) | |
1851 | { | |
1852 | struct bfq_sched_data *sd = entity->sched_data; | |
1853 | struct bfq_service_tree *st = bfq_entity_service_tree(entity); | |
1854 | ||
1855 | if (entity == sd->in_service_entity) { | |
1856 | /* | |
1857 | * We are requeueing the current in-service entity, | |
1858 | * which may have to be done for one of the following | |
1859 | * reasons: | |
1860 | * - entity represents the in-service queue, and the | |
1861 | * in-service queue is being requeued after an | |
1862 | * expiration; | |
1863 | * - entity represents a group, and its budget has | |
1864 | * changed because one of its child entities has | |
1865 | * just been either activated or requeued for some | |
1866 | * reason; the timestamps of the entity need then to | |
1867 | * be updated, and the entity needs to be enqueued | |
1868 | * or repositioned accordingly. | |
1869 | * | |
1870 | * In particular, before requeueing, the start time of | |
1871 | * the entity must be moved forward to account for the | |
1872 | * service that the entity has received while in | |
1873 | * service. This is done by the next instructions. The | |
1874 | * finish time will then be updated according to this | |
1875 | * new value of the start time, and to the budget of | |
1876 | * the entity. | |
1877 | */ | |
1878 | bfq_calc_finish(entity, entity->service); | |
1879 | entity->start = entity->finish; | |
1880 | /* | |
1881 | * In addition, if the entity had more than one child | |
1882 | * when set in service, then was not extracted from | |
1883 | * the active tree. This implies that the position of | |
1884 | * the entity in the active tree may need to be | |
1885 | * changed now, because we have just updated the start | |
1886 | * time of the entity, and we will update its finish | |
1887 | * time in a moment (the requeueing is then, more | |
1888 | * precisely, a repositioning in this case). To | |
1889 | * implement this repositioning, we: 1) dequeue the | |
1890 | * entity here, 2) update the finish time and | |
1891 | * requeue the entity according to the new | |
1892 | * timestamps below. | |
1893 | */ | |
1894 | if (entity->tree) | |
1895 | bfq_active_extract(st, entity); | |
1896 | } else { /* The entity is already active, and not in service */ | |
1897 | /* | |
1898 | * In this case, this function gets called only if the | |
1899 | * next_in_service entity below this entity has | |
1900 | * changed, and this change has caused the budget of | |
1901 | * this entity to change, which, finally implies that | |
1902 | * the finish time of this entity must be | |
1903 | * updated. Such an update may cause the scheduling, | |
1904 | * i.e., the position in the active tree, of this | |
1905 | * entity to change. We handle this change by: 1) | |
1906 | * dequeueing the entity here, 2) updating the finish | |
1907 | * time and requeueing the entity according to the new | |
1908 | * timestamps below. This is the same approach as the | |
1909 | * non-extracted-entity sub-case above. | |
1910 | */ | |
1911 | bfq_active_extract(st, entity); | |
1912 | } | |
1913 | ||
1914 | bfq_update_fin_time_enqueue(entity, st, false); | |
1915 | } | |
1916 | ||
1917 | static void __bfq_activate_requeue_entity(struct bfq_entity *entity, | |
1918 | struct bfq_sched_data *sd, | |
1919 | bool non_blocking_wait_rq) | |
1920 | { | |
1921 | struct bfq_service_tree *st = bfq_entity_service_tree(entity); | |
1922 | ||
1923 | if (sd->in_service_entity == entity || entity->tree == &st->active) | |
1924 | /* | |
1925 | * in service or already queued on the active tree, | |
1926 | * requeue or reposition | |
1927 | */ | |
1928 | __bfq_requeue_entity(entity); | |
1929 | else | |
1930 | /* | |
1931 | * Not in service and not queued on its active tree: | |
1932 | * the activity is idle and this is a true activation. | |
1933 | */ | |
1934 | __bfq_activate_entity(entity, non_blocking_wait_rq); | |
1935 | } | |
1936 | ||
1937 | ||
1938 | /** | |
1939 | * bfq_activate_entity - activate or requeue an entity representing a bfq_queue, | |
1940 | * and activate, requeue or reposition all ancestors | |
1941 | * for which such an update becomes necessary. | |
aee69d78 PV |
1942 | * @entity: the entity to activate. |
1943 | * @non_blocking_wait_rq: true if this entity was waiting for a request | |
e21b7a0b AA |
1944 | * @requeue: true if this is a requeue, which implies that bfqq is |
1945 | * being expired; thus ALL its ancestors stop being served and must | |
1946 | * therefore be requeued | |
aee69d78 | 1947 | */ |
e21b7a0b AA |
1948 | static void bfq_activate_requeue_entity(struct bfq_entity *entity, |
1949 | bool non_blocking_wait_rq, | |
1950 | bool requeue) | |
aee69d78 PV |
1951 | { |
1952 | struct bfq_sched_data *sd; | |
1953 | ||
1954 | for_each_entity(entity) { | |
aee69d78 | 1955 | sd = entity->sched_data; |
e21b7a0b AA |
1956 | __bfq_activate_requeue_entity(entity, sd, non_blocking_wait_rq); |
1957 | ||
1958 | if (!bfq_update_next_in_service(sd, entity) && !requeue) | |
aee69d78 PV |
1959 | break; |
1960 | } | |
1961 | } | |
1962 | ||
1963 | /** | |
1964 | * __bfq_deactivate_entity - deactivate an entity from its service tree. | |
1965 | * @entity: the entity to deactivate. | |
e21b7a0b AA |
1966 | * @ins_into_idle_tree: if false, the entity will not be put into the |
1967 | * idle tree. | |
aee69d78 | 1968 | * |
e21b7a0b AA |
1969 | * Deactivates an entity, independently from its previous state. Must |
1970 | * be invoked only if entity is on a service tree. Extracts the entity | |
1971 | * from that tree, and if necessary and allowed, puts it on the idle | |
1972 | * tree. | |
aee69d78 | 1973 | */ |
e21b7a0b AA |
1974 | static bool __bfq_deactivate_entity(struct bfq_entity *entity, |
1975 | bool ins_into_idle_tree) | |
aee69d78 PV |
1976 | { |
1977 | struct bfq_sched_data *sd = entity->sched_data; | |
1978 | struct bfq_service_tree *st = bfq_entity_service_tree(entity); | |
1979 | int is_in_service = entity == sd->in_service_entity; | |
aee69d78 | 1980 | |
e21b7a0b AA |
1981 | if (!entity->on_st) /* entity never activated, or already inactive */ |
1982 | return false; | |
aee69d78 | 1983 | |
e21b7a0b | 1984 | if (is_in_service) |
aee69d78 | 1985 | bfq_calc_finish(entity, entity->service); |
e21b7a0b AA |
1986 | |
1987 | if (entity->tree == &st->active) | |
aee69d78 | 1988 | bfq_active_extract(st, entity); |
e21b7a0b | 1989 | else if (!is_in_service && entity->tree == &st->idle) |
aee69d78 PV |
1990 | bfq_idle_extract(st, entity); |
1991 | ||
e21b7a0b | 1992 | if (!ins_into_idle_tree || !bfq_gt(entity->finish, st->vtime)) |
aee69d78 PV |
1993 | bfq_forget_entity(st, entity, is_in_service); |
1994 | else | |
1995 | bfq_idle_insert(st, entity); | |
1996 | ||
e21b7a0b | 1997 | return true; |
aee69d78 PV |
1998 | } |
1999 | ||
2000 | /** | |
e21b7a0b | 2001 | * bfq_deactivate_entity - deactivate an entity representing a bfq_queue. |
aee69d78 | 2002 | * @entity: the entity to deactivate. |
e21b7a0b | 2003 | * @ins_into_idle_tree: true if the entity can be put on the idle tree |
aee69d78 | 2004 | */ |
e21b7a0b AA |
2005 | static void bfq_deactivate_entity(struct bfq_entity *entity, |
2006 | bool ins_into_idle_tree, | |
2007 | bool expiration) | |
aee69d78 PV |
2008 | { |
2009 | struct bfq_sched_data *sd; | |
2010 | struct bfq_entity *parent = NULL; | |
2011 | ||
2012 | for_each_entity_safe(entity, parent) { | |
2013 | sd = entity->sched_data; | |
2014 | ||
e21b7a0b | 2015 | if (!__bfq_deactivate_entity(entity, ins_into_idle_tree)) { |
aee69d78 | 2016 | /* |
e21b7a0b AA |
2017 | * entity is not in any tree any more, so |
2018 | * this deactivation is a no-op, and there is | |
2019 | * nothing to change for upper-level entities | |
2020 | * (in case of expiration, this can never | |
2021 | * happen). | |
aee69d78 | 2022 | */ |
e21b7a0b AA |
2023 | return; |
2024 | } | |
2025 | ||
2026 | if (sd->next_in_service == entity) | |
2027 | /* | |
2028 | * entity was the next_in_service entity, | |
2029 | * then, since entity has just been | |
2030 | * deactivated, a new one must be found. | |
2031 | */ | |
2032 | bfq_update_next_in_service(sd, NULL); | |
aee69d78 PV |
2033 | |
2034 | if (sd->next_in_service) | |
2035 | /* | |
e21b7a0b AA |
2036 | * The parent entity is still backlogged, |
2037 | * because next_in_service is not NULL. So, no | |
2038 | * further upwards deactivation must be | |
2039 | * performed. Yet, next_in_service has | |
2040 | * changed. Then the schedule does need to be | |
2041 | * updated upwards. | |
aee69d78 | 2042 | */ |
e21b7a0b | 2043 | break; |
aee69d78 PV |
2044 | |
2045 | /* | |
e21b7a0b AA |
2046 | * If we get here, then the parent is no more |
2047 | * backlogged and we need to propagate the | |
2048 | * deactivation upwards. Thus let the loop go on. | |
aee69d78 | 2049 | */ |
aee69d78 | 2050 | |
e21b7a0b AA |
2051 | /* |
2052 | * Also let parent be queued into the idle tree on | |
2053 | * deactivation, to preserve service guarantees, and | |
2054 | * assuming that who invoked this function does not | |
2055 | * need parent entities too to be removed completely. | |
2056 | */ | |
2057 | ins_into_idle_tree = true; | |
2058 | } | |
aee69d78 | 2059 | |
e21b7a0b AA |
2060 | /* |
2061 | * If the deactivation loop is fully executed, then there are | |
2062 | * no more entities to touch and next loop is not executed at | |
2063 | * all. Otherwise, requeue remaining entities if they are | |
2064 | * about to stop receiving service, or reposition them if this | |
2065 | * is not the case. | |
2066 | */ | |
aee69d78 PV |
2067 | entity = parent; |
2068 | for_each_entity(entity) { | |
e21b7a0b AA |
2069 | /* |
2070 | * Invoke __bfq_requeue_entity on entity, even if | |
2071 | * already active, to requeue/reposition it in the | |
2072 | * active tree (because sd->next_in_service has | |
2073 | * changed) | |
2074 | */ | |
2075 | __bfq_requeue_entity(entity); | |
aee69d78 PV |
2076 | |
2077 | sd = entity->sched_data; | |
e21b7a0b AA |
2078 | if (!bfq_update_next_in_service(sd, entity) && |
2079 | !expiration) | |
2080 | /* | |
2081 | * next_in_service unchanged or not causing | |
2082 | * any change in entity->parent->sd, and no | |
2083 | * requeueing needed for expiration: stop | |
2084 | * here. | |
2085 | */ | |
aee69d78 PV |
2086 | break; |
2087 | } | |
2088 | } | |
2089 | ||
2090 | /** | |
e21b7a0b AA |
2091 | * bfq_calc_vtime_jump - compute the value to which the vtime should jump, |
2092 | * if needed, to have at least one entity eligible. | |
aee69d78 PV |
2093 | * @st: the service tree to act upon. |
2094 | * | |
e21b7a0b | 2095 | * Assumes that st is not empty. |
aee69d78 | 2096 | */ |
e21b7a0b | 2097 | static u64 bfq_calc_vtime_jump(struct bfq_service_tree *st) |
aee69d78 | 2098 | { |
e21b7a0b AA |
2099 | struct bfq_entity *root_entity = bfq_root_active_entity(&st->active); |
2100 | ||
2101 | if (bfq_gt(root_entity->min_start, st->vtime)) | |
2102 | return root_entity->min_start; | |
2103 | ||
2104 | return st->vtime; | |
2105 | } | |
aee69d78 | 2106 | |
e21b7a0b AA |
2107 | static void bfq_update_vtime(struct bfq_service_tree *st, u64 new_value) |
2108 | { | |
2109 | if (new_value > st->vtime) { | |
2110 | st->vtime = new_value; | |
aee69d78 PV |
2111 | bfq_forget_idle(st); |
2112 | } | |
2113 | } | |
2114 | ||
2115 | /** | |
2116 | * bfq_first_active_entity - find the eligible entity with | |
2117 | * the smallest finish time | |
2118 | * @st: the service tree to select from. | |
e21b7a0b | 2119 | * @vtime: the system virtual to use as a reference for eligibility |
aee69d78 PV |
2120 | * |
2121 | * This function searches the first schedulable entity, starting from the | |
2122 | * root of the tree and going on the left every time on this side there is | |
2123 | * a subtree with at least one eligible (start >= vtime) entity. The path on | |
2124 | * the right is followed only if a) the left subtree contains no eligible | |
2125 | * entities and b) no eligible entity has been found yet. | |
2126 | */ | |
e21b7a0b AA |
2127 | static struct bfq_entity *bfq_first_active_entity(struct bfq_service_tree *st, |
2128 | u64 vtime) | |
aee69d78 PV |
2129 | { |
2130 | struct bfq_entity *entry, *first = NULL; | |
2131 | struct rb_node *node = st->active.rb_node; | |
2132 | ||
2133 | while (node) { | |
2134 | entry = rb_entry(node, struct bfq_entity, rb_node); | |
2135 | left: | |
e21b7a0b | 2136 | if (!bfq_gt(entry->start, vtime)) |
aee69d78 PV |
2137 | first = entry; |
2138 | ||
2139 | if (node->rb_left) { | |
2140 | entry = rb_entry(node->rb_left, | |
2141 | struct bfq_entity, rb_node); | |
e21b7a0b | 2142 | if (!bfq_gt(entry->min_start, vtime)) { |
aee69d78 PV |
2143 | node = node->rb_left; |
2144 | goto left; | |
2145 | } | |
2146 | } | |
2147 | if (first) | |
2148 | break; | |
2149 | node = node->rb_right; | |
2150 | } | |
2151 | ||
e21b7a0b AA |
2152 | return first; |
2153 | } | |
2154 | ||
2155 | /** | |
2156 | * __bfq_lookup_next_entity - return the first eligible entity in @st. | |
2157 | * @st: the service tree. | |
2158 | * | |
2159 | * If there is no in-service entity for the sched_data st belongs to, | |
2160 | * then return the entity that will be set in service if: | |
2161 | * 1) the parent entity this st belongs to is set in service; | |
2162 | * 2) no entity belonging to such parent entity undergoes a state change | |
2163 | * that would influence the timestamps of the entity (e.g., becomes idle, | |
2164 | * becomes backlogged, changes its budget, ...). | |
2165 | * | |
2166 | * In this first case, update the virtual time in @st too (see the | |
2167 | * comments on this update inside the function). | |
2168 | * | |
2169 | * In constrast, if there is an in-service entity, then return the | |
2170 | * entity that would be set in service if not only the above | |
2171 | * conditions, but also the next one held true: the currently | |
2172 | * in-service entity, on expiration, | |
2173 | * 1) gets a finish time equal to the current one, or | |
2174 | * 2) is not eligible any more, or | |
2175 | * 3) is idle. | |
2176 | */ | |
2177 | static struct bfq_entity * | |
2178 | __bfq_lookup_next_entity(struct bfq_service_tree *st, bool in_service) | |
2179 | { | |
2180 | struct bfq_entity *entity; | |
2181 | u64 new_vtime; | |
2182 | ||
2183 | if (RB_EMPTY_ROOT(&st->active)) | |
2184 | return NULL; | |
2185 | ||
2186 | /* | |
2187 | * Get the value of the system virtual time for which at | |
2188 | * least one entity is eligible. | |
2189 | */ | |
2190 | new_vtime = bfq_calc_vtime_jump(st); | |
2191 | ||
2192 | /* | |
2193 | * If there is no in-service entity for the sched_data this | |
2194 | * active tree belongs to, then push the system virtual time | |
2195 | * up to the value that guarantees that at least one entity is | |
2196 | * eligible. If, instead, there is an in-service entity, then | |
2197 | * do not make any such update, because there is already an | |
2198 | * eligible entity, namely the in-service one (even if the | |
2199 | * entity is not on st, because it was extracted when set in | |
2200 | * service). | |
2201 | */ | |
2202 | if (!in_service) | |
2203 | bfq_update_vtime(st, new_vtime); | |
2204 | ||
2205 | entity = bfq_first_active_entity(st, new_vtime); | |
2206 | ||
2207 | return entity; | |
2208 | } | |
2209 | ||
2210 | /** | |
2211 | * bfq_lookup_next_entity - return the first eligible entity in @sd. | |
2212 | * @sd: the sched_data. | |
2213 | * | |
2214 | * This function is invoked when there has been a change in the trees | |
2215 | * for sd, and we need know what is the new next entity after this | |
2216 | * change. | |
2217 | */ | |
2218 | static struct bfq_entity *bfq_lookup_next_entity(struct bfq_sched_data *sd) | |
2219 | { | |
2220 | struct bfq_service_tree *st = sd->service_tree; | |
2221 | struct bfq_service_tree *idle_class_st = st + (BFQ_IOPRIO_CLASSES - 1); | |
2222 | struct bfq_entity *entity = NULL; | |
2223 | int class_idx = 0; | |
2224 | ||
2225 | /* | |
2226 | * Choose from idle class, if needed to guarantee a minimum | |
2227 | * bandwidth to this class (and if there is some active entity | |
2228 | * in idle class). This should also mitigate | |
2229 | * priority-inversion problems in case a low priority task is | |
2230 | * holding file system resources. | |
2231 | */ | |
2232 | if (time_is_before_jiffies(sd->bfq_class_idle_last_service + | |
2233 | BFQ_CL_IDLE_TIMEOUT)) { | |
2234 | if (!RB_EMPTY_ROOT(&idle_class_st->active)) | |
2235 | class_idx = BFQ_IOPRIO_CLASSES - 1; | |
2236 | /* About to be served if backlogged, or not yet backlogged */ | |
2237 | sd->bfq_class_idle_last_service = jiffies; | |
2238 | } | |
2239 | ||
2240 | /* | |
2241 | * Find the next entity to serve for the highest-priority | |
2242 | * class, unless the idle class needs to be served. | |
2243 | */ | |
2244 | for (; class_idx < BFQ_IOPRIO_CLASSES; class_idx++) { | |
2245 | entity = __bfq_lookup_next_entity(st + class_idx, | |
2246 | sd->in_service_entity); | |
2247 | ||
2248 | if (entity) | |
2249 | break; | |
2250 | } | |
2251 | ||
2252 | if (!entity) | |
2253 | return NULL; | |
2254 | ||
2255 | return entity; | |
2256 | } | |
2257 | ||
2258 | static bool next_queue_may_preempt(struct bfq_data *bfqd) | |
2259 | { | |
2260 | struct bfq_sched_data *sd = &bfqd->root_group->sched_data; | |
2261 | ||
2262 | return sd->next_in_service != sd->in_service_entity; | |
2263 | } | |
2264 | ||
2265 | /* | |
2266 | * Get next queue for service. | |
2267 | */ | |
2268 | static struct bfq_queue *bfq_get_next_queue(struct bfq_data *bfqd) | |
2269 | { | |
2270 | struct bfq_entity *entity = NULL; | |
2271 | struct bfq_sched_data *sd; | |
2272 | struct bfq_queue *bfqq; | |
2273 | ||
2274 | if (bfqd->busy_queues == 0) | |
2275 | return NULL; | |
2276 | ||
2277 | /* | |
2278 | * Traverse the path from the root to the leaf entity to | |
2279 | * serve. Set in service all the entities visited along the | |
2280 | * way. | |
2281 | */ | |
2282 | sd = &bfqd->root_group->sched_data; | |
2283 | for (; sd ; sd = entity->my_sched_data) { | |
2284 | /* | |
2285 | * WARNING. We are about to set the in-service entity | |
2286 | * to sd->next_in_service, i.e., to the (cached) value | |
2287 | * returned by bfq_lookup_next_entity(sd) the last | |
2288 | * time it was invoked, i.e., the last time when the | |
2289 | * service order in sd changed as a consequence of the | |
2290 | * activation or deactivation of an entity. In this | |
2291 | * respect, if we execute bfq_lookup_next_entity(sd) | |
2292 | * in this very moment, it may, although with low | |
2293 | * probability, yield a different entity than that | |
2294 | * pointed to by sd->next_in_service. This rare event | |
2295 | * happens in case there was no CLASS_IDLE entity to | |
2296 | * serve for sd when bfq_lookup_next_entity(sd) was | |
2297 | * invoked for the last time, while there is now one | |
2298 | * such entity. | |
2299 | * | |
2300 | * If the above event happens, then the scheduling of | |
2301 | * such entity in CLASS_IDLE is postponed until the | |
2302 | * service of the sd->next_in_service entity | |
2303 | * finishes. In fact, when the latter is expired, | |
2304 | * bfq_lookup_next_entity(sd) gets called again, | |
2305 | * exactly to update sd->next_in_service. | |
2306 | */ | |
2307 | ||
2308 | /* Make next_in_service entity become in_service_entity */ | |
2309 | entity = sd->next_in_service; | |
2310 | sd->in_service_entity = entity; | |
2311 | ||
2312 | /* | |
2313 | * Reset the accumulator of the amount of service that | |
2314 | * the entity is about to receive. | |
2315 | */ | |
2316 | entity->service = 0; | |
2317 | ||
2318 | /* | |
2319 | * If entity is no longer a candidate for next | |
2320 | * service, then we extract it from its active tree, | |
2321 | * for the following reason. To further boost the | |
2322 | * throughput in some special case, BFQ needs to know | |
2323 | * which is the next candidate entity to serve, while | |
2324 | * there is already an entity in service. In this | |
2325 | * respect, to make it easy to compute/update the next | |
2326 | * candidate entity to serve after the current | |
2327 | * candidate has been set in service, there is a case | |
2328 | * where it is necessary to extract the current | |
2329 | * candidate from its service tree. Such a case is | |
2330 | * when the entity just set in service cannot be also | |
2331 | * a candidate for next service. Details about when | |
2332 | * this conditions holds are reported in the comments | |
2333 | * on the function bfq_no_longer_next_in_service() | |
2334 | * invoked below. | |
2335 | */ | |
2336 | if (bfq_no_longer_next_in_service(entity)) | |
2337 | bfq_active_extract(bfq_entity_service_tree(entity), | |
2338 | entity); | |
2339 | ||
2340 | /* | |
2341 | * For the same reason why we may have just extracted | |
2342 | * entity from its active tree, we may need to update | |
2343 | * next_in_service for the sched_data of entity too, | |
2344 | * regardless of whether entity has been extracted. | |
2345 | * In fact, even if entity has not been extracted, a | |
2346 | * descendant entity may get extracted. Such an event | |
2347 | * would cause a change in next_in_service for the | |
2348 | * level of the descendant entity, and thus possibly | |
2349 | * back to upper levels. | |
2350 | * | |
2351 | * We cannot perform the resulting needed update | |
2352 | * before the end of this loop, because, to know which | |
2353 | * is the correct next-to-serve candidate entity for | |
2354 | * each level, we need first to find the leaf entity | |
2355 | * to set in service. In fact, only after we know | |
2356 | * which is the next-to-serve leaf entity, we can | |
2357 | * discover whether the parent entity of the leaf | |
2358 | * entity becomes the next-to-serve, and so on. | |
2359 | */ | |
2360 | ||
2361 | } | |
2362 | ||
2363 | bfqq = bfq_entity_to_bfqq(entity); | |
2364 | ||
2365 | /* | |
2366 | * We can finally update all next-to-serve entities along the | |
2367 | * path from the leaf entity just set in service to the root. | |
2368 | */ | |
2369 | for_each_entity(entity) { | |
2370 | struct bfq_sched_data *sd = entity->sched_data; | |
2371 | ||
2372 | if (!bfq_update_next_in_service(sd, NULL)) | |
2373 | break; | |
2374 | } | |
2375 | ||
2376 | return bfqq; | |
2377 | } | |
2378 | ||
2379 | static void __bfq_bfqd_reset_in_service(struct bfq_data *bfqd) | |
2380 | { | |
2381 | struct bfq_queue *in_serv_bfqq = bfqd->in_service_queue; | |
2382 | struct bfq_entity *in_serv_entity = &in_serv_bfqq->entity; | |
2383 | struct bfq_entity *entity = in_serv_entity; | |
2384 | ||
2385 | if (bfqd->in_service_bic) { | |
2386 | put_io_context(bfqd->in_service_bic->icq.ioc); | |
2387 | bfqd->in_service_bic = NULL; | |
2388 | } | |
2389 | ||
2390 | bfq_clear_bfqq_wait_request(in_serv_bfqq); | |
2391 | hrtimer_try_to_cancel(&bfqd->idle_slice_timer); | |
2392 | bfqd->in_service_queue = NULL; | |
2393 | ||
2394 | /* | |
2395 | * When this function is called, all in-service entities have | |
2396 | * been properly deactivated or requeued, so we can safely | |
2397 | * execute the final step: reset in_service_entity along the | |
2398 | * path from entity to the root. | |
2399 | */ | |
2400 | for_each_entity(entity) | |
2401 | entity->sched_data->in_service_entity = NULL; | |
2402 | ||
2403 | /* | |
2404 | * in_serv_entity is no longer in service, so, if it is in no | |
2405 | * service tree either, then release the service reference to | |
2406 | * the queue it represents (taken with bfq_get_entity). | |
2407 | */ | |
2408 | if (!in_serv_entity->on_st) | |
2409 | bfq_put_queue(in_serv_bfqq); | |
2410 | } | |
2411 | ||
2412 | static void bfq_deactivate_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
2413 | bool ins_into_idle_tree, bool expiration) | |
2414 | { | |
2415 | struct bfq_entity *entity = &bfqq->entity; | |
2416 | ||
2417 | bfq_deactivate_entity(entity, ins_into_idle_tree, expiration); | |
2418 | } | |
2419 | ||
2420 | static void bfq_activate_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq) | |
2421 | { | |
2422 | struct bfq_entity *entity = &bfqq->entity; | |
2423 | ||
2424 | bfq_activate_requeue_entity(entity, bfq_bfqq_non_blocking_wait_rq(bfqq), | |
2425 | false); | |
2426 | bfq_clear_bfqq_non_blocking_wait_rq(bfqq); | |
2427 | } | |
2428 | ||
2429 | static void bfq_requeue_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq) | |
2430 | { | |
2431 | struct bfq_entity *entity = &bfqq->entity; | |
2432 | ||
2433 | bfq_activate_requeue_entity(entity, false, | |
2434 | bfqq == bfqd->in_service_queue); | |
2435 | } | |
2436 | ||
2437 | static void bfqg_stats_update_dequeue(struct bfq_group *bfqg); | |
2438 | ||
2439 | /* | |
2440 | * Called when the bfqq no longer has requests pending, remove it from | |
2441 | * the service tree. As a special case, it can be invoked during an | |
2442 | * expiration. | |
2443 | */ | |
2444 | static void bfq_del_bfqq_busy(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
2445 | bool expiration) | |
2446 | { | |
2447 | bfq_log_bfqq(bfqd, bfqq, "del from busy"); | |
2448 | ||
2449 | bfq_clear_bfqq_busy(bfqq); | |
2450 | ||
2451 | bfqd->busy_queues--; | |
2452 | ||
2453 | bfqg_stats_update_dequeue(bfqq_group(bfqq)); | |
2454 | ||
2455 | bfq_deactivate_bfqq(bfqd, bfqq, true, expiration); | |
2456 | } | |
2457 | ||
2458 | /* | |
2459 | * Called when an inactive queue receives a new request. | |
2460 | */ | |
2461 | static void bfq_add_bfqq_busy(struct bfq_data *bfqd, struct bfq_queue *bfqq) | |
2462 | { | |
2463 | bfq_log_bfqq(bfqd, bfqq, "add to busy"); | |
2464 | ||
2465 | bfq_activate_bfqq(bfqd, bfqq); | |
2466 | ||
2467 | bfq_mark_bfqq_busy(bfqq); | |
2468 | bfqd->busy_queues++; | |
2469 | } | |
2470 | ||
2471 | #ifdef CONFIG_BFQ_GROUP_IOSCHED | |
2472 | ||
2473 | /* bfqg stats flags */ | |
2474 | enum bfqg_stats_flags { | |
2475 | BFQG_stats_waiting = 0, | |
2476 | BFQG_stats_idling, | |
2477 | BFQG_stats_empty, | |
2478 | }; | |
2479 | ||
2480 | #define BFQG_FLAG_FNS(name) \ | |
2481 | static void bfqg_stats_mark_##name(struct bfqg_stats *stats) \ | |
2482 | { \ | |
2483 | stats->flags |= (1 << BFQG_stats_##name); \ | |
2484 | } \ | |
2485 | static void bfqg_stats_clear_##name(struct bfqg_stats *stats) \ | |
2486 | { \ | |
2487 | stats->flags &= ~(1 << BFQG_stats_##name); \ | |
2488 | } \ | |
2489 | static int bfqg_stats_##name(struct bfqg_stats *stats) \ | |
2490 | { \ | |
2491 | return (stats->flags & (1 << BFQG_stats_##name)) != 0; \ | |
2492 | } \ | |
2493 | ||
2494 | BFQG_FLAG_FNS(waiting) | |
2495 | BFQG_FLAG_FNS(idling) | |
2496 | BFQG_FLAG_FNS(empty) | |
2497 | #undef BFQG_FLAG_FNS | |
2498 | ||
2499 | /* This should be called with the queue_lock held. */ | |
2500 | static void bfqg_stats_update_group_wait_time(struct bfqg_stats *stats) | |
2501 | { | |
2502 | unsigned long long now; | |
2503 | ||
2504 | if (!bfqg_stats_waiting(stats)) | |
2505 | return; | |
2506 | ||
2507 | now = sched_clock(); | |
2508 | if (time_after64(now, stats->start_group_wait_time)) | |
2509 | blkg_stat_add(&stats->group_wait_time, | |
2510 | now - stats->start_group_wait_time); | |
2511 | bfqg_stats_clear_waiting(stats); | |
2512 | } | |
2513 | ||
2514 | /* This should be called with the queue_lock held. */ | |
2515 | static void bfqg_stats_set_start_group_wait_time(struct bfq_group *bfqg, | |
2516 | struct bfq_group *curr_bfqg) | |
2517 | { | |
2518 | struct bfqg_stats *stats = &bfqg->stats; | |
2519 | ||
2520 | if (bfqg_stats_waiting(stats)) | |
2521 | return; | |
2522 | if (bfqg == curr_bfqg) | |
2523 | return; | |
2524 | stats->start_group_wait_time = sched_clock(); | |
2525 | bfqg_stats_mark_waiting(stats); | |
2526 | } | |
2527 | ||
2528 | /* This should be called with the queue_lock held. */ | |
2529 | static void bfqg_stats_end_empty_time(struct bfqg_stats *stats) | |
2530 | { | |
2531 | unsigned long long now; | |
2532 | ||
2533 | if (!bfqg_stats_empty(stats)) | |
2534 | return; | |
2535 | ||
2536 | now = sched_clock(); | |
2537 | if (time_after64(now, stats->start_empty_time)) | |
2538 | blkg_stat_add(&stats->empty_time, | |
2539 | now - stats->start_empty_time); | |
2540 | bfqg_stats_clear_empty(stats); | |
2541 | } | |
2542 | ||
2543 | static void bfqg_stats_update_dequeue(struct bfq_group *bfqg) | |
2544 | { | |
2545 | blkg_stat_add(&bfqg->stats.dequeue, 1); | |
2546 | } | |
2547 | ||
2548 | static void bfqg_stats_set_start_empty_time(struct bfq_group *bfqg) | |
2549 | { | |
2550 | struct bfqg_stats *stats = &bfqg->stats; | |
2551 | ||
2552 | if (blkg_rwstat_total(&stats->queued)) | |
2553 | return; | |
2554 | ||
2555 | /* | |
2556 | * group is already marked empty. This can happen if bfqq got new | |
2557 | * request in parent group and moved to this group while being added | |
2558 | * to service tree. Just ignore the event and move on. | |
2559 | */ | |
2560 | if (bfqg_stats_empty(stats)) | |
2561 | return; | |
2562 | ||
2563 | stats->start_empty_time = sched_clock(); | |
2564 | bfqg_stats_mark_empty(stats); | |
2565 | } | |
2566 | ||
2567 | static void bfqg_stats_update_idle_time(struct bfq_group *bfqg) | |
2568 | { | |
2569 | struct bfqg_stats *stats = &bfqg->stats; | |
2570 | ||
2571 | if (bfqg_stats_idling(stats)) { | |
2572 | unsigned long long now = sched_clock(); | |
2573 | ||
2574 | if (time_after64(now, stats->start_idle_time)) | |
2575 | blkg_stat_add(&stats->idle_time, | |
2576 | now - stats->start_idle_time); | |
2577 | bfqg_stats_clear_idling(stats); | |
2578 | } | |
2579 | } | |
2580 | ||
2581 | static void bfqg_stats_set_start_idle_time(struct bfq_group *bfqg) | |
2582 | { | |
2583 | struct bfqg_stats *stats = &bfqg->stats; | |
2584 | ||
2585 | stats->start_idle_time = sched_clock(); | |
2586 | bfqg_stats_mark_idling(stats); | |
2587 | } | |
2588 | ||
2589 | static void bfqg_stats_update_avg_queue_size(struct bfq_group *bfqg) | |
2590 | { | |
2591 | struct bfqg_stats *stats = &bfqg->stats; | |
2592 | ||
2593 | blkg_stat_add(&stats->avg_queue_size_sum, | |
2594 | blkg_rwstat_total(&stats->queued)); | |
2595 | blkg_stat_add(&stats->avg_queue_size_samples, 1); | |
2596 | bfqg_stats_update_group_wait_time(stats); | |
2597 | } | |
2598 | ||
2599 | /* | |
2600 | * blk-cgroup policy-related handlers | |
2601 | * The following functions help in converting between blk-cgroup | |
2602 | * internal structures and BFQ-specific structures. | |
2603 | */ | |
2604 | ||
2605 | static struct bfq_group *pd_to_bfqg(struct blkg_policy_data *pd) | |
2606 | { | |
2607 | return pd ? container_of(pd, struct bfq_group, pd) : NULL; | |
2608 | } | |
2609 | ||
2610 | static struct blkcg_gq *bfqg_to_blkg(struct bfq_group *bfqg) | |
2611 | { | |
2612 | return pd_to_blkg(&bfqg->pd); | |
2613 | } | |
2614 | ||
2615 | static struct blkcg_policy blkcg_policy_bfq; | |
2616 | ||
2617 | static struct bfq_group *blkg_to_bfqg(struct blkcg_gq *blkg) | |
2618 | { | |
2619 | return pd_to_bfqg(blkg_to_pd(blkg, &blkcg_policy_bfq)); | |
2620 | } | |
2621 | ||
2622 | /* | |
2623 | * bfq_group handlers | |
2624 | * The following functions help in navigating the bfq_group hierarchy | |
2625 | * by allowing to find the parent of a bfq_group or the bfq_group | |
2626 | * associated to a bfq_queue. | |
2627 | */ | |
2628 | ||
2629 | static struct bfq_group *bfqg_parent(struct bfq_group *bfqg) | |
2630 | { | |
2631 | struct blkcg_gq *pblkg = bfqg_to_blkg(bfqg)->parent; | |
2632 | ||
2633 | return pblkg ? blkg_to_bfqg(pblkg) : NULL; | |
2634 | } | |
2635 | ||
2636 | static struct bfq_group *bfqq_group(struct bfq_queue *bfqq) | |
2637 | { | |
2638 | struct bfq_entity *group_entity = bfqq->entity.parent; | |
2639 | ||
2640 | return group_entity ? container_of(group_entity, struct bfq_group, | |
2641 | entity) : | |
2642 | bfqq->bfqd->root_group; | |
2643 | } | |
2644 | ||
2645 | /* | |
2646 | * The following two functions handle get and put of a bfq_group by | |
2647 | * wrapping the related blk-cgroup hooks. | |
2648 | */ | |
2649 | ||
2650 | static void bfqg_get(struct bfq_group *bfqg) | |
2651 | { | |
2652 | return blkg_get(bfqg_to_blkg(bfqg)); | |
2653 | } | |
2654 | ||
2655 | static void bfqg_put(struct bfq_group *bfqg) | |
2656 | { | |
2657 | return blkg_put(bfqg_to_blkg(bfqg)); | |
2658 | } | |
2659 | ||
2660 | static void bfqg_stats_update_io_add(struct bfq_group *bfqg, | |
2661 | struct bfq_queue *bfqq, | |
2662 | unsigned int op) | |
2663 | { | |
2664 | blkg_rwstat_add(&bfqg->stats.queued, op, 1); | |
2665 | bfqg_stats_end_empty_time(&bfqg->stats); | |
2666 | if (!(bfqq == ((struct bfq_data *)bfqg->bfqd)->in_service_queue)) | |
2667 | bfqg_stats_set_start_group_wait_time(bfqg, bfqq_group(bfqq)); | |
2668 | } | |
2669 | ||
2670 | static void bfqg_stats_update_io_remove(struct bfq_group *bfqg, unsigned int op) | |
2671 | { | |
2672 | blkg_rwstat_add(&bfqg->stats.queued, op, -1); | |
2673 | } | |
2674 | ||
2675 | static void bfqg_stats_update_io_merged(struct bfq_group *bfqg, unsigned int op) | |
2676 | { | |
2677 | blkg_rwstat_add(&bfqg->stats.merged, op, 1); | |
2678 | } | |
2679 | ||
2680 | static void bfqg_stats_update_completion(struct bfq_group *bfqg, | |
2681 | uint64_t start_time, uint64_t io_start_time, | |
2682 | unsigned int op) | |
2683 | { | |
2684 | struct bfqg_stats *stats = &bfqg->stats; | |
2685 | unsigned long long now = sched_clock(); | |
2686 | ||
2687 | if (time_after64(now, io_start_time)) | |
2688 | blkg_rwstat_add(&stats->service_time, op, | |
2689 | now - io_start_time); | |
2690 | if (time_after64(io_start_time, start_time)) | |
2691 | blkg_rwstat_add(&stats->wait_time, op, | |
2692 | io_start_time - start_time); | |
2693 | } | |
2694 | ||
2695 | /* @stats = 0 */ | |
2696 | static void bfqg_stats_reset(struct bfqg_stats *stats) | |
2697 | { | |
2698 | /* queued stats shouldn't be cleared */ | |
2699 | blkg_rwstat_reset(&stats->merged); | |
2700 | blkg_rwstat_reset(&stats->service_time); | |
2701 | blkg_rwstat_reset(&stats->wait_time); | |
2702 | blkg_stat_reset(&stats->time); | |
2703 | blkg_stat_reset(&stats->avg_queue_size_sum); | |
2704 | blkg_stat_reset(&stats->avg_queue_size_samples); | |
2705 | blkg_stat_reset(&stats->dequeue); | |
2706 | blkg_stat_reset(&stats->group_wait_time); | |
2707 | blkg_stat_reset(&stats->idle_time); | |
2708 | blkg_stat_reset(&stats->empty_time); | |
2709 | } | |
2710 | ||
2711 | /* @to += @from */ | |
2712 | static void bfqg_stats_add_aux(struct bfqg_stats *to, struct bfqg_stats *from) | |
2713 | { | |
2714 | if (!to || !from) | |
2715 | return; | |
2716 | ||
2717 | /* queued stats shouldn't be cleared */ | |
2718 | blkg_rwstat_add_aux(&to->merged, &from->merged); | |
2719 | blkg_rwstat_add_aux(&to->service_time, &from->service_time); | |
2720 | blkg_rwstat_add_aux(&to->wait_time, &from->wait_time); | |
2721 | blkg_stat_add_aux(&from->time, &from->time); | |
2722 | blkg_stat_add_aux(&to->avg_queue_size_sum, &from->avg_queue_size_sum); | |
2723 | blkg_stat_add_aux(&to->avg_queue_size_samples, | |
2724 | &from->avg_queue_size_samples); | |
2725 | blkg_stat_add_aux(&to->dequeue, &from->dequeue); | |
2726 | blkg_stat_add_aux(&to->group_wait_time, &from->group_wait_time); | |
2727 | blkg_stat_add_aux(&to->idle_time, &from->idle_time); | |
2728 | blkg_stat_add_aux(&to->empty_time, &from->empty_time); | |
2729 | } | |
2730 | ||
2731 | /* | |
2732 | * Transfer @bfqg's stats to its parent's aux counts so that the ancestors' | |
2733 | * recursive stats can still account for the amount used by this bfqg after | |
2734 | * it's gone. | |
2735 | */ | |
2736 | static void bfqg_stats_xfer_dead(struct bfq_group *bfqg) | |
2737 | { | |
2738 | struct bfq_group *parent; | |
2739 | ||
2740 | if (!bfqg) /* root_group */ | |
2741 | return; | |
2742 | ||
2743 | parent = bfqg_parent(bfqg); | |
2744 | ||
2745 | lockdep_assert_held(bfqg_to_blkg(bfqg)->q->queue_lock); | |
2746 | ||
2747 | if (unlikely(!parent)) | |
2748 | return; | |
2749 | ||
2750 | bfqg_stats_add_aux(&parent->stats, &bfqg->stats); | |
2751 | bfqg_stats_reset(&bfqg->stats); | |
2752 | } | |
2753 | ||
2754 | static void bfq_init_entity(struct bfq_entity *entity, | |
2755 | struct bfq_group *bfqg) | |
2756 | { | |
2757 | struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); | |
2758 | ||
2759 | entity->weight = entity->new_weight; | |
2760 | entity->orig_weight = entity->new_weight; | |
2761 | if (bfqq) { | |
2762 | bfqq->ioprio = bfqq->new_ioprio; | |
2763 | bfqq->ioprio_class = bfqq->new_ioprio_class; | |
2764 | bfqg_get(bfqg); | |
2765 | } | |
2766 | entity->parent = bfqg->my_entity; /* NULL for root group */ | |
2767 | entity->sched_data = &bfqg->sched_data; | |
2768 | } | |
2769 | ||
2770 | static void bfqg_stats_exit(struct bfqg_stats *stats) | |
2771 | { | |
2772 | blkg_rwstat_exit(&stats->merged); | |
2773 | blkg_rwstat_exit(&stats->service_time); | |
2774 | blkg_rwstat_exit(&stats->wait_time); | |
2775 | blkg_rwstat_exit(&stats->queued); | |
2776 | blkg_stat_exit(&stats->time); | |
2777 | blkg_stat_exit(&stats->avg_queue_size_sum); | |
2778 | blkg_stat_exit(&stats->avg_queue_size_samples); | |
2779 | blkg_stat_exit(&stats->dequeue); | |
2780 | blkg_stat_exit(&stats->group_wait_time); | |
2781 | blkg_stat_exit(&stats->idle_time); | |
2782 | blkg_stat_exit(&stats->empty_time); | |
2783 | } | |
2784 | ||
2785 | static int bfqg_stats_init(struct bfqg_stats *stats, gfp_t gfp) | |
2786 | { | |
2787 | if (blkg_rwstat_init(&stats->merged, gfp) || | |
2788 | blkg_rwstat_init(&stats->service_time, gfp) || | |
2789 | blkg_rwstat_init(&stats->wait_time, gfp) || | |
2790 | blkg_rwstat_init(&stats->queued, gfp) || | |
2791 | blkg_stat_init(&stats->time, gfp) || | |
2792 | blkg_stat_init(&stats->avg_queue_size_sum, gfp) || | |
2793 | blkg_stat_init(&stats->avg_queue_size_samples, gfp) || | |
2794 | blkg_stat_init(&stats->dequeue, gfp) || | |
2795 | blkg_stat_init(&stats->group_wait_time, gfp) || | |
2796 | blkg_stat_init(&stats->idle_time, gfp) || | |
2797 | blkg_stat_init(&stats->empty_time, gfp)) { | |
2798 | bfqg_stats_exit(stats); | |
2799 | return -ENOMEM; | |
2800 | } | |
2801 | ||
2802 | return 0; | |
2803 | } | |
2804 | ||
2805 | static struct bfq_group_data *cpd_to_bfqgd(struct blkcg_policy_data *cpd) | |
2806 | { | |
2807 | return cpd ? container_of(cpd, struct bfq_group_data, pd) : NULL; | |
2808 | } | |
2809 | ||
2810 | static struct bfq_group_data *blkcg_to_bfqgd(struct blkcg *blkcg) | |
2811 | { | |
2812 | return cpd_to_bfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_bfq)); | |
2813 | } | |
2814 | ||
2815 | static struct blkcg_policy_data *bfq_cpd_alloc(gfp_t gfp) | |
2816 | { | |
2817 | struct bfq_group_data *bgd; | |
2818 | ||
2819 | bgd = kzalloc(sizeof(*bgd), gfp); | |
2820 | if (!bgd) | |
2821 | return NULL; | |
2822 | return &bgd->pd; | |
2823 | } | |
2824 | ||
2825 | static void bfq_cpd_init(struct blkcg_policy_data *cpd) | |
2826 | { | |
2827 | struct bfq_group_data *d = cpd_to_bfqgd(cpd); | |
2828 | ||
2829 | d->weight = cgroup_subsys_on_dfl(io_cgrp_subsys) ? | |
2830 | CGROUP_WEIGHT_DFL : BFQ_WEIGHT_LEGACY_DFL; | |
2831 | } | |
2832 | ||
2833 | static void bfq_cpd_free(struct blkcg_policy_data *cpd) | |
2834 | { | |
2835 | kfree(cpd_to_bfqgd(cpd)); | |
2836 | } | |
2837 | ||
2838 | static struct blkg_policy_data *bfq_pd_alloc(gfp_t gfp, int node) | |
2839 | { | |
2840 | struct bfq_group *bfqg; | |
2841 | ||
2842 | bfqg = kzalloc_node(sizeof(*bfqg), gfp, node); | |
2843 | if (!bfqg) | |
2844 | return NULL; | |
2845 | ||
2846 | if (bfqg_stats_init(&bfqg->stats, gfp)) { | |
2847 | kfree(bfqg); | |
2848 | return NULL; | |
2849 | } | |
2850 | ||
2851 | return &bfqg->pd; | |
2852 | } | |
2853 | ||
2854 | static void bfq_pd_init(struct blkg_policy_data *pd) | |
2855 | { | |
2856 | struct blkcg_gq *blkg = pd_to_blkg(pd); | |
2857 | struct bfq_group *bfqg = blkg_to_bfqg(blkg); | |
2858 | struct bfq_data *bfqd = blkg->q->elevator->elevator_data; | |
2859 | struct bfq_entity *entity = &bfqg->entity; | |
2860 | struct bfq_group_data *d = blkcg_to_bfqgd(blkg->blkcg); | |
2861 | ||
2862 | entity->orig_weight = entity->weight = entity->new_weight = d->weight; | |
2863 | entity->my_sched_data = &bfqg->sched_data; | |
2864 | bfqg->my_entity = entity; /* | |
2865 | * the root_group's will be set to NULL | |
2866 | * in bfq_init_queue() | |
2867 | */ | |
2868 | bfqg->bfqd = bfqd; | |
2869 | } | |
2870 | ||
2871 | static void bfq_pd_free(struct blkg_policy_data *pd) | |
2872 | { | |
2873 | struct bfq_group *bfqg = pd_to_bfqg(pd); | |
2874 | ||
2875 | bfqg_stats_exit(&bfqg->stats); | |
2876 | return kfree(bfqg); | |
2877 | } | |
2878 | ||
2879 | static void bfq_pd_reset_stats(struct blkg_policy_data *pd) | |
2880 | { | |
2881 | struct bfq_group *bfqg = pd_to_bfqg(pd); | |
2882 | ||
2883 | bfqg_stats_reset(&bfqg->stats); | |
2884 | } | |
2885 | ||
2886 | static void bfq_group_set_parent(struct bfq_group *bfqg, | |
2887 | struct bfq_group *parent) | |
2888 | { | |
2889 | struct bfq_entity *entity; | |
2890 | ||
2891 | entity = &bfqg->entity; | |
2892 | entity->parent = parent->my_entity; | |
2893 | entity->sched_data = &parent->sched_data; | |
2894 | } | |
2895 | ||
2896 | static struct bfq_group *bfq_lookup_bfqg(struct bfq_data *bfqd, | |
2897 | struct blkcg *blkcg) | |
2898 | { | |
2899 | struct blkcg_gq *blkg; | |
2900 | ||
2901 | blkg = blkg_lookup(blkcg, bfqd->queue); | |
2902 | if (likely(blkg)) | |
2903 | return blkg_to_bfqg(blkg); | |
2904 | return NULL; | |
2905 | } | |
2906 | ||
2907 | static struct bfq_group *bfq_find_set_group(struct bfq_data *bfqd, | |
2908 | struct blkcg *blkcg) | |
2909 | { | |
2910 | struct bfq_group *bfqg, *parent; | |
2911 | struct bfq_entity *entity; | |
2912 | ||
2913 | bfqg = bfq_lookup_bfqg(bfqd, blkcg); | |
2914 | ||
2915 | if (unlikely(!bfqg)) | |
2916 | return NULL; | |
2917 | ||
2918 | /* | |
2919 | * Update chain of bfq_groups as we might be handling a leaf group | |
2920 | * which, along with some of its relatives, has not been hooked yet | |
2921 | * to the private hierarchy of BFQ. | |
2922 | */ | |
2923 | entity = &bfqg->entity; | |
2924 | for_each_entity(entity) { | |
2925 | bfqg = container_of(entity, struct bfq_group, entity); | |
2926 | if (bfqg != bfqd->root_group) { | |
2927 | parent = bfqg_parent(bfqg); | |
2928 | if (!parent) | |
2929 | parent = bfqd->root_group; | |
2930 | bfq_group_set_parent(bfqg, parent); | |
2931 | } | |
2932 | } | |
2933 | ||
2934 | return bfqg; | |
2935 | } | |
2936 | ||
2937 | static void bfq_bfqq_expire(struct bfq_data *bfqd, | |
2938 | struct bfq_queue *bfqq, | |
2939 | bool compensate, | |
2940 | enum bfqq_expiration reason); | |
2941 | ||
2942 | /** | |
2943 | * bfq_bfqq_move - migrate @bfqq to @bfqg. | |
2944 | * @bfqd: queue descriptor. | |
2945 | * @bfqq: the queue to move. | |
2946 | * @bfqg: the group to move to. | |
2947 | * | |
2948 | * Move @bfqq to @bfqg, deactivating it from its old group and reactivating | |
2949 | * it on the new one. Avoid putting the entity on the old group idle tree. | |
2950 | * | |
2951 | * Must be called under the queue lock; the cgroup owning @bfqg must | |
2952 | * not disappear (by now this just means that we are called under | |
2953 | * rcu_read_lock()). | |
2954 | */ | |
2955 | static void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
2956 | struct bfq_group *bfqg) | |
2957 | { | |
2958 | struct bfq_entity *entity = &bfqq->entity; | |
2959 | ||
2960 | /* If bfqq is empty, then bfq_bfqq_expire also invokes | |
2961 | * bfq_del_bfqq_busy, thereby removing bfqq and its entity | |
2962 | * from data structures related to current group. Otherwise we | |
2963 | * need to remove bfqq explicitly with bfq_deactivate_bfqq, as | |
2964 | * we do below. | |
2965 | */ | |
2966 | if (bfqq == bfqd->in_service_queue) | |
2967 | bfq_bfqq_expire(bfqd, bfqd->in_service_queue, | |
2968 | false, BFQQE_PREEMPTED); | |
2969 | ||
2970 | if (bfq_bfqq_busy(bfqq)) | |
2971 | bfq_deactivate_bfqq(bfqd, bfqq, false, false); | |
2972 | else if (entity->on_st) | |
2973 | bfq_put_idle_entity(bfq_entity_service_tree(entity), entity); | |
2974 | bfqg_put(bfqq_group(bfqq)); | |
2975 | ||
2976 | /* | |
2977 | * Here we use a reference to bfqg. We don't need a refcounter | |
2978 | * as the cgroup reference will not be dropped, so that its | |
2979 | * destroy() callback will not be invoked. | |
2980 | */ | |
2981 | entity->parent = bfqg->my_entity; | |
2982 | entity->sched_data = &bfqg->sched_data; | |
2983 | bfqg_get(bfqg); | |
2984 | ||
2985 | if (bfq_bfqq_busy(bfqq)) | |
2986 | bfq_activate_bfqq(bfqd, bfqq); | |
2987 | ||
2988 | if (!bfqd->in_service_queue && !bfqd->rq_in_driver) | |
2989 | bfq_schedule_dispatch(bfqd); | |
2990 | } | |
2991 | ||
2992 | /** | |
2993 | * __bfq_bic_change_cgroup - move @bic to @cgroup. | |
2994 | * @bfqd: the queue descriptor. | |
2995 | * @bic: the bic to move. | |
2996 | * @blkcg: the blk-cgroup to move to. | |
2997 | * | |
2998 | * Move bic to blkcg, assuming that bfqd->queue is locked; the caller | |
2999 | * has to make sure that the reference to cgroup is valid across the call. | |
3000 | * | |
3001 | * NOTE: an alternative approach might have been to store the current | |
3002 | * cgroup in bfqq and getting a reference to it, reducing the lookup | |
3003 | * time here, at the price of slightly more complex code. | |
3004 | */ | |
3005 | static struct bfq_group *__bfq_bic_change_cgroup(struct bfq_data *bfqd, | |
3006 | struct bfq_io_cq *bic, | |
3007 | struct blkcg *blkcg) | |
3008 | { | |
3009 | struct bfq_queue *async_bfqq = bic_to_bfqq(bic, 0); | |
3010 | struct bfq_queue *sync_bfqq = bic_to_bfqq(bic, 1); | |
3011 | struct bfq_group *bfqg; | |
3012 | struct bfq_entity *entity; | |
3013 | ||
3014 | bfqg = bfq_find_set_group(bfqd, blkcg); | |
3015 | ||
3016 | if (unlikely(!bfqg)) | |
3017 | bfqg = bfqd->root_group; | |
3018 | ||
3019 | if (async_bfqq) { | |
3020 | entity = &async_bfqq->entity; | |
3021 | ||
3022 | if (entity->sched_data != &bfqg->sched_data) { | |
3023 | bic_set_bfqq(bic, NULL, 0); | |
3024 | bfq_log_bfqq(bfqd, async_bfqq, | |
3025 | "bic_change_group: %p %d", | |
3026 | async_bfqq, | |
3027 | async_bfqq->ref); | |
3028 | bfq_put_queue(async_bfqq); | |
3029 | } | |
3030 | } | |
3031 | ||
3032 | if (sync_bfqq) { | |
3033 | entity = &sync_bfqq->entity; | |
3034 | if (entity->sched_data != &bfqg->sched_data) | |
3035 | bfq_bfqq_move(bfqd, sync_bfqq, bfqg); | |
3036 | } | |
3037 | ||
3038 | return bfqg; | |
3039 | } | |
3040 | ||
3041 | static void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio) | |
3042 | { | |
3043 | struct bfq_data *bfqd = bic_to_bfqd(bic); | |
3044 | struct bfq_group *bfqg = NULL; | |
3045 | uint64_t serial_nr; | |
3046 | ||
3047 | rcu_read_lock(); | |
3048 | serial_nr = bio_blkcg(bio)->css.serial_nr; | |
3049 | ||
3050 | /* | |
3051 | * Check whether blkcg has changed. The condition may trigger | |
3052 | * spuriously on a newly created cic but there's no harm. | |
3053 | */ | |
3054 | if (unlikely(!bfqd) || likely(bic->blkcg_serial_nr == serial_nr)) | |
3055 | goto out; | |
3056 | ||
3057 | bfqg = __bfq_bic_change_cgroup(bfqd, bic, bio_blkcg(bio)); | |
3058 | bic->blkcg_serial_nr = serial_nr; | |
3059 | out: | |
3060 | rcu_read_unlock(); | |
3061 | } | |
3062 | ||
3063 | /** | |
3064 | * bfq_flush_idle_tree - deactivate any entity on the idle tree of @st. | |
3065 | * @st: the service tree being flushed. | |
3066 | */ | |
3067 | static void bfq_flush_idle_tree(struct bfq_service_tree *st) | |
3068 | { | |
3069 | struct bfq_entity *entity = st->first_idle; | |
3070 | ||
3071 | for (; entity ; entity = st->first_idle) | |
3072 | __bfq_deactivate_entity(entity, false); | |
3073 | } | |
3074 | ||
3075 | /** | |
3076 | * bfq_reparent_leaf_entity - move leaf entity to the root_group. | |
3077 | * @bfqd: the device data structure with the root group. | |
3078 | * @entity: the entity to move. | |
3079 | */ | |
3080 | static void bfq_reparent_leaf_entity(struct bfq_data *bfqd, | |
3081 | struct bfq_entity *entity) | |
3082 | { | |
3083 | struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); | |
3084 | ||
3085 | bfq_bfqq_move(bfqd, bfqq, bfqd->root_group); | |
aee69d78 PV |
3086 | } |
3087 | ||
3088 | /** | |
e21b7a0b AA |
3089 | * bfq_reparent_active_entities - move to the root group all active |
3090 | * entities. | |
3091 | * @bfqd: the device data structure with the root group. | |
3092 | * @bfqg: the group to move from. | |
3093 | * @st: the service tree with the entities. | |
aee69d78 | 3094 | * |
e21b7a0b | 3095 | * Needs queue_lock to be taken and reference to be valid over the call. |
aee69d78 | 3096 | */ |
e21b7a0b AA |
3097 | static void bfq_reparent_active_entities(struct bfq_data *bfqd, |
3098 | struct bfq_group *bfqg, | |
3099 | struct bfq_service_tree *st) | |
aee69d78 | 3100 | { |
e21b7a0b AA |
3101 | struct rb_root *active = &st->active; |
3102 | struct bfq_entity *entity = NULL; | |
aee69d78 | 3103 | |
e21b7a0b AA |
3104 | if (!RB_EMPTY_ROOT(&st->active)) |
3105 | entity = bfq_entity_of(rb_first(active)); | |
aee69d78 | 3106 | |
e21b7a0b AA |
3107 | for (; entity ; entity = bfq_entity_of(rb_first(active))) |
3108 | bfq_reparent_leaf_entity(bfqd, entity); | |
aee69d78 | 3109 | |
e21b7a0b AA |
3110 | if (bfqg->sched_data.in_service_entity) |
3111 | bfq_reparent_leaf_entity(bfqd, | |
3112 | bfqg->sched_data.in_service_entity); | |
aee69d78 PV |
3113 | } |
3114 | ||
3115 | /** | |
e21b7a0b AA |
3116 | * bfq_pd_offline - deactivate the entity associated with @pd, |
3117 | * and reparent its children entities. | |
3118 | * @pd: descriptor of the policy going offline. | |
aee69d78 | 3119 | * |
e21b7a0b AA |
3120 | * blkio already grabs the queue_lock for us, so no need to use |
3121 | * RCU-based magic | |
aee69d78 | 3122 | */ |
e21b7a0b | 3123 | static void bfq_pd_offline(struct blkg_policy_data *pd) |
aee69d78 | 3124 | { |
e21b7a0b AA |
3125 | struct bfq_service_tree *st; |
3126 | struct bfq_group *bfqg = pd_to_bfqg(pd); | |
3127 | struct bfq_data *bfqd = bfqg->bfqd; | |
3128 | struct bfq_entity *entity = bfqg->my_entity; | |
3129 | unsigned long flags; | |
3130 | int i; | |
aee69d78 | 3131 | |
e21b7a0b AA |
3132 | if (!entity) /* root group */ |
3133 | return; | |
3134 | ||
3135 | spin_lock_irqsave(&bfqd->lock, flags); | |
aee69d78 | 3136 | /* |
e21b7a0b AA |
3137 | * Empty all service_trees belonging to this group before |
3138 | * deactivating the group itself. | |
aee69d78 | 3139 | */ |
e21b7a0b AA |
3140 | for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) { |
3141 | st = bfqg->sched_data.service_tree + i; | |
3142 | ||
3143 | /* | |
3144 | * The idle tree may still contain bfq_queues belonging | |
3145 | * to exited task because they never migrated to a different | |
3146 | * cgroup from the one being destroyed now. No one else | |
3147 | * can access them so it's safe to act without any lock. | |
3148 | */ | |
3149 | bfq_flush_idle_tree(st); | |
3150 | ||
3151 | /* | |
3152 | * It may happen that some queues are still active | |
3153 | * (busy) upon group destruction (if the corresponding | |
3154 | * processes have been forced to terminate). We move | |
3155 | * all the leaf entities corresponding to these queues | |
3156 | * to the root_group. | |
3157 | * Also, it may happen that the group has an entity | |
3158 | * in service, which is disconnected from the active | |
3159 | * tree: it must be moved, too. | |
3160 | * There is no need to put the sync queues, as the | |
3161 | * scheduler has taken no reference. | |
3162 | */ | |
3163 | bfq_reparent_active_entities(bfqd, bfqg, st); | |
aee69d78 PV |
3164 | } |
3165 | ||
e21b7a0b AA |
3166 | __bfq_deactivate_entity(entity, false); |
3167 | bfq_put_async_queues(bfqd, bfqg); | |
3168 | ||
3169 | spin_unlock_irqrestore(&bfqd->lock, flags); | |
3170 | /* | |
3171 | * @blkg is going offline and will be ignored by | |
3172 | * blkg_[rw]stat_recursive_sum(). Transfer stats to the parent so | |
3173 | * that they don't get lost. If IOs complete after this point, the | |
3174 | * stats for them will be lost. Oh well... | |
3175 | */ | |
3176 | bfqg_stats_xfer_dead(bfqg); | |
aee69d78 PV |
3177 | } |
3178 | ||
44e44a1b PV |
3179 | static void bfq_end_wr_async(struct bfq_data *bfqd) |
3180 | { | |
3181 | struct blkcg_gq *blkg; | |
3182 | ||
3183 | list_for_each_entry(blkg, &bfqd->queue->blkg_list, q_node) { | |
3184 | struct bfq_group *bfqg = blkg_to_bfqg(blkg); | |
3185 | ||
3186 | bfq_end_wr_async_queues(bfqd, bfqg); | |
3187 | } | |
3188 | bfq_end_wr_async_queues(bfqd, bfqd->root_group); | |
3189 | } | |
3190 | ||
e21b7a0b | 3191 | static int bfq_io_show_weight(struct seq_file *sf, void *v) |
aee69d78 | 3192 | { |
e21b7a0b AA |
3193 | struct blkcg *blkcg = css_to_blkcg(seq_css(sf)); |
3194 | struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg); | |
3195 | unsigned int val = 0; | |
aee69d78 | 3196 | |
e21b7a0b AA |
3197 | if (bfqgd) |
3198 | val = bfqgd->weight; | |
aee69d78 | 3199 | |
e21b7a0b | 3200 | seq_printf(sf, "%u\n", val); |
aee69d78 | 3201 | |
e21b7a0b AA |
3202 | return 0; |
3203 | } | |
3204 | ||
3205 | static int bfq_io_set_weight_legacy(struct cgroup_subsys_state *css, | |
3206 | struct cftype *cftype, | |
3207 | u64 val) | |
aee69d78 | 3208 | { |
e21b7a0b AA |
3209 | struct blkcg *blkcg = css_to_blkcg(css); |
3210 | struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg); | |
3211 | struct blkcg_gq *blkg; | |
3212 | int ret = -ERANGE; | |
aee69d78 | 3213 | |
e21b7a0b AA |
3214 | if (val < BFQ_MIN_WEIGHT || val > BFQ_MAX_WEIGHT) |
3215 | return ret; | |
aee69d78 | 3216 | |
e21b7a0b AA |
3217 | ret = 0; |
3218 | spin_lock_irq(&blkcg->lock); | |
3219 | bfqgd->weight = (unsigned short)val; | |
3220 | hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) { | |
3221 | struct bfq_group *bfqg = blkg_to_bfqg(blkg); | |
3222 | ||
3223 | if (!bfqg) | |
3224 | continue; | |
3225 | /* | |
3226 | * Setting the prio_changed flag of the entity | |
3227 | * to 1 with new_weight == weight would re-set | |
3228 | * the value of the weight to its ioprio mapping. | |
3229 | * Set the flag only if necessary. | |
3230 | */ | |
3231 | if ((unsigned short)val != bfqg->entity.new_weight) { | |
3232 | bfqg->entity.new_weight = (unsigned short)val; | |
3233 | /* | |
3234 | * Make sure that the above new value has been | |
3235 | * stored in bfqg->entity.new_weight before | |
3236 | * setting the prio_changed flag. In fact, | |
3237 | * this flag may be read asynchronously (in | |
3238 | * critical sections protected by a different | |
3239 | * lock than that held here), and finding this | |
3240 | * flag set may cause the execution of the code | |
3241 | * for updating parameters whose value may | |
3242 | * depend also on bfqg->entity.new_weight (in | |
3243 | * __bfq_entity_update_weight_prio). | |
3244 | * This barrier makes sure that the new value | |
3245 | * of bfqg->entity.new_weight is correctly | |
3246 | * seen in that code. | |
3247 | */ | |
3248 | smp_wmb(); | |
3249 | bfqg->entity.prio_changed = 1; | |
3250 | } | |
aee69d78 | 3251 | } |
e21b7a0b | 3252 | spin_unlock_irq(&blkcg->lock); |
aee69d78 | 3253 | |
e21b7a0b AA |
3254 | return ret; |
3255 | } | |
aee69d78 | 3256 | |
e21b7a0b AA |
3257 | static ssize_t bfq_io_set_weight(struct kernfs_open_file *of, |
3258 | char *buf, size_t nbytes, | |
3259 | loff_t off) | |
3260 | { | |
3261 | u64 weight; | |
3262 | /* First unsigned long found in the file is used */ | |
3263 | int ret = kstrtoull(strim(buf), 0, &weight); | |
3264 | ||
3265 | if (ret) | |
3266 | return ret; | |
3267 | ||
3268 | return bfq_io_set_weight_legacy(of_css(of), NULL, weight); | |
aee69d78 PV |
3269 | } |
3270 | ||
e21b7a0b | 3271 | static int bfqg_print_stat(struct seq_file *sf, void *v) |
aee69d78 | 3272 | { |
e21b7a0b AA |
3273 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat, |
3274 | &blkcg_policy_bfq, seq_cft(sf)->private, false); | |
3275 | return 0; | |
3276 | } | |
aee69d78 | 3277 | |
e21b7a0b AA |
3278 | static int bfqg_print_rwstat(struct seq_file *sf, void *v) |
3279 | { | |
3280 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat, | |
3281 | &blkcg_policy_bfq, seq_cft(sf)->private, true); | |
3282 | return 0; | |
3283 | } | |
aee69d78 | 3284 | |
e21b7a0b AA |
3285 | static u64 bfqg_prfill_stat_recursive(struct seq_file *sf, |
3286 | struct blkg_policy_data *pd, int off) | |
3287 | { | |
3288 | u64 sum = blkg_stat_recursive_sum(pd_to_blkg(pd), | |
3289 | &blkcg_policy_bfq, off); | |
3290 | return __blkg_prfill_u64(sf, pd, sum); | |
3291 | } | |
aee69d78 | 3292 | |
e21b7a0b AA |
3293 | static u64 bfqg_prfill_rwstat_recursive(struct seq_file *sf, |
3294 | struct blkg_policy_data *pd, int off) | |
3295 | { | |
3296 | struct blkg_rwstat sum = blkg_rwstat_recursive_sum(pd_to_blkg(pd), | |
3297 | &blkcg_policy_bfq, | |
3298 | off); | |
3299 | return __blkg_prfill_rwstat(sf, pd, &sum); | |
aee69d78 PV |
3300 | } |
3301 | ||
e21b7a0b | 3302 | static int bfqg_print_stat_recursive(struct seq_file *sf, void *v) |
aee69d78 | 3303 | { |
e21b7a0b AA |
3304 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), |
3305 | bfqg_prfill_stat_recursive, &blkcg_policy_bfq, | |
3306 | seq_cft(sf)->private, false); | |
3307 | return 0; | |
3308 | } | |
aee69d78 | 3309 | |
e21b7a0b AA |
3310 | static int bfqg_print_rwstat_recursive(struct seq_file *sf, void *v) |
3311 | { | |
3312 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), | |
3313 | bfqg_prfill_rwstat_recursive, &blkcg_policy_bfq, | |
3314 | seq_cft(sf)->private, true); | |
3315 | return 0; | |
aee69d78 PV |
3316 | } |
3317 | ||
e21b7a0b AA |
3318 | static u64 bfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd, |
3319 | int off) | |
aee69d78 | 3320 | { |
e21b7a0b | 3321 | u64 sum = blkg_rwstat_total(&pd->blkg->stat_bytes); |
aee69d78 | 3322 | |
e21b7a0b | 3323 | return __blkg_prfill_u64(sf, pd, sum >> 9); |
aee69d78 PV |
3324 | } |
3325 | ||
e21b7a0b | 3326 | static int bfqg_print_stat_sectors(struct seq_file *sf, void *v) |
aee69d78 | 3327 | { |
e21b7a0b AA |
3328 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), |
3329 | bfqg_prfill_sectors, &blkcg_policy_bfq, 0, false); | |
3330 | return 0; | |
3331 | } | |
aee69d78 | 3332 | |
e21b7a0b AA |
3333 | static u64 bfqg_prfill_sectors_recursive(struct seq_file *sf, |
3334 | struct blkg_policy_data *pd, int off) | |
3335 | { | |
3336 | struct blkg_rwstat tmp = blkg_rwstat_recursive_sum(pd->blkg, NULL, | |
3337 | offsetof(struct blkcg_gq, stat_bytes)); | |
3338 | u64 sum = atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_READ]) + | |
3339 | atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_WRITE]); | |
aee69d78 | 3340 | |
e21b7a0b AA |
3341 | return __blkg_prfill_u64(sf, pd, sum >> 9); |
3342 | } | |
aee69d78 | 3343 | |
e21b7a0b AA |
3344 | static int bfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v) |
3345 | { | |
3346 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), | |
3347 | bfqg_prfill_sectors_recursive, &blkcg_policy_bfq, 0, | |
3348 | false); | |
3349 | return 0; | |
aee69d78 PV |
3350 | } |
3351 | ||
e21b7a0b AA |
3352 | static u64 bfqg_prfill_avg_queue_size(struct seq_file *sf, |
3353 | struct blkg_policy_data *pd, int off) | |
aee69d78 | 3354 | { |
e21b7a0b AA |
3355 | struct bfq_group *bfqg = pd_to_bfqg(pd); |
3356 | u64 samples = blkg_stat_read(&bfqg->stats.avg_queue_size_samples); | |
3357 | u64 v = 0; | |
aee69d78 | 3358 | |
e21b7a0b AA |
3359 | if (samples) { |
3360 | v = blkg_stat_read(&bfqg->stats.avg_queue_size_sum); | |
3361 | v = div64_u64(v, samples); | |
3362 | } | |
3363 | __blkg_prfill_u64(sf, pd, v); | |
3364 | return 0; | |
3365 | } | |
aee69d78 | 3366 | |
e21b7a0b AA |
3367 | /* print avg_queue_size */ |
3368 | static int bfqg_print_avg_queue_size(struct seq_file *sf, void *v) | |
3369 | { | |
3370 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), | |
3371 | bfqg_prfill_avg_queue_size, &blkcg_policy_bfq, | |
3372 | 0, false); | |
3373 | return 0; | |
3374 | } | |
3375 | ||
3376 | static struct bfq_group * | |
3377 | bfq_create_group_hierarchy(struct bfq_data *bfqd, int node) | |
3378 | { | |
3379 | int ret; | |
3380 | ||
3381 | ret = blkcg_activate_policy(bfqd->queue, &blkcg_policy_bfq); | |
3382 | if (ret) | |
3383 | return NULL; | |
3384 | ||
3385 | return blkg_to_bfqg(bfqd->queue->root_blkg); | |
aee69d78 PV |
3386 | } |
3387 | ||
e21b7a0b AA |
3388 | static struct cftype bfq_blkcg_legacy_files[] = { |
3389 | { | |
3390 | .name = "bfq.weight", | |
3391 | .flags = CFTYPE_NOT_ON_ROOT, | |
3392 | .seq_show = bfq_io_show_weight, | |
3393 | .write_u64 = bfq_io_set_weight_legacy, | |
3394 | }, | |
3395 | ||
3396 | /* statistics, covers only the tasks in the bfqg */ | |
3397 | { | |
3398 | .name = "bfq.time", | |
3399 | .private = offsetof(struct bfq_group, stats.time), | |
3400 | .seq_show = bfqg_print_stat, | |
3401 | }, | |
3402 | { | |
3403 | .name = "bfq.sectors", | |
3404 | .seq_show = bfqg_print_stat_sectors, | |
3405 | }, | |
3406 | { | |
3407 | .name = "bfq.io_service_bytes", | |
3408 | .private = (unsigned long)&blkcg_policy_bfq, | |
3409 | .seq_show = blkg_print_stat_bytes, | |
3410 | }, | |
3411 | { | |
3412 | .name = "bfq.io_serviced", | |
3413 | .private = (unsigned long)&blkcg_policy_bfq, | |
3414 | .seq_show = blkg_print_stat_ios, | |
3415 | }, | |
3416 | { | |
3417 | .name = "bfq.io_service_time", | |
3418 | .private = offsetof(struct bfq_group, stats.service_time), | |
3419 | .seq_show = bfqg_print_rwstat, | |
3420 | }, | |
3421 | { | |
3422 | .name = "bfq.io_wait_time", | |
3423 | .private = offsetof(struct bfq_group, stats.wait_time), | |
3424 | .seq_show = bfqg_print_rwstat, | |
3425 | }, | |
3426 | { | |
3427 | .name = "bfq.io_merged", | |
3428 | .private = offsetof(struct bfq_group, stats.merged), | |
3429 | .seq_show = bfqg_print_rwstat, | |
3430 | }, | |
3431 | { | |
3432 | .name = "bfq.io_queued", | |
3433 | .private = offsetof(struct bfq_group, stats.queued), | |
3434 | .seq_show = bfqg_print_rwstat, | |
3435 | }, | |
3436 | ||
3437 | /* the same statictics which cover the bfqg and its descendants */ | |
3438 | { | |
3439 | .name = "bfq.time_recursive", | |
3440 | .private = offsetof(struct bfq_group, stats.time), | |
3441 | .seq_show = bfqg_print_stat_recursive, | |
3442 | }, | |
3443 | { | |
3444 | .name = "bfq.sectors_recursive", | |
3445 | .seq_show = bfqg_print_stat_sectors_recursive, | |
3446 | }, | |
3447 | { | |
3448 | .name = "bfq.io_service_bytes_recursive", | |
3449 | .private = (unsigned long)&blkcg_policy_bfq, | |
3450 | .seq_show = blkg_print_stat_bytes_recursive, | |
3451 | }, | |
3452 | { | |
3453 | .name = "bfq.io_serviced_recursive", | |
3454 | .private = (unsigned long)&blkcg_policy_bfq, | |
3455 | .seq_show = blkg_print_stat_ios_recursive, | |
3456 | }, | |
3457 | { | |
3458 | .name = "bfq.io_service_time_recursive", | |
3459 | .private = offsetof(struct bfq_group, stats.service_time), | |
3460 | .seq_show = bfqg_print_rwstat_recursive, | |
3461 | }, | |
3462 | { | |
3463 | .name = "bfq.io_wait_time_recursive", | |
3464 | .private = offsetof(struct bfq_group, stats.wait_time), | |
3465 | .seq_show = bfqg_print_rwstat_recursive, | |
3466 | }, | |
3467 | { | |
3468 | .name = "bfq.io_merged_recursive", | |
3469 | .private = offsetof(struct bfq_group, stats.merged), | |
3470 | .seq_show = bfqg_print_rwstat_recursive, | |
3471 | }, | |
3472 | { | |
3473 | .name = "bfq.io_queued_recursive", | |
3474 | .private = offsetof(struct bfq_group, stats.queued), | |
3475 | .seq_show = bfqg_print_rwstat_recursive, | |
3476 | }, | |
3477 | { | |
3478 | .name = "bfq.avg_queue_size", | |
3479 | .seq_show = bfqg_print_avg_queue_size, | |
3480 | }, | |
3481 | { | |
3482 | .name = "bfq.group_wait_time", | |
3483 | .private = offsetof(struct bfq_group, stats.group_wait_time), | |
3484 | .seq_show = bfqg_print_stat, | |
3485 | }, | |
3486 | { | |
3487 | .name = "bfq.idle_time", | |
3488 | .private = offsetof(struct bfq_group, stats.idle_time), | |
3489 | .seq_show = bfqg_print_stat, | |
3490 | }, | |
3491 | { | |
3492 | .name = "bfq.empty_time", | |
3493 | .private = offsetof(struct bfq_group, stats.empty_time), | |
3494 | .seq_show = bfqg_print_stat, | |
3495 | }, | |
3496 | { | |
3497 | .name = "bfq.dequeue", | |
3498 | .private = offsetof(struct bfq_group, stats.dequeue), | |
3499 | .seq_show = bfqg_print_stat, | |
3500 | }, | |
3501 | { } /* terminate */ | |
3502 | }; | |
3503 | ||
3504 | static struct cftype bfq_blkg_files[] = { | |
3505 | { | |
3506 | .name = "bfq.weight", | |
3507 | .flags = CFTYPE_NOT_ON_ROOT, | |
3508 | .seq_show = bfq_io_show_weight, | |
3509 | .write = bfq_io_set_weight, | |
3510 | }, | |
3511 | {} /* terminate */ | |
3512 | }; | |
3513 | ||
3514 | #else /* CONFIG_BFQ_GROUP_IOSCHED */ | |
3515 | ||
3516 | static inline void bfqg_stats_update_io_add(struct bfq_group *bfqg, | |
3517 | struct bfq_queue *bfqq, unsigned int op) { } | |
3518 | static inline void | |
3519 | bfqg_stats_update_io_remove(struct bfq_group *bfqg, unsigned int op) { } | |
3520 | static inline void | |
3521 | bfqg_stats_update_io_merged(struct bfq_group *bfqg, unsigned int op) { } | |
3522 | static inline void bfqg_stats_update_completion(struct bfq_group *bfqg, | |
3523 | uint64_t start_time, uint64_t io_start_time, | |
3524 | unsigned int op) { } | |
3525 | static inline void | |
3526 | bfqg_stats_set_start_group_wait_time(struct bfq_group *bfqg, | |
3527 | struct bfq_group *curr_bfqg) { } | |
3528 | static inline void bfqg_stats_end_empty_time(struct bfqg_stats *stats) { } | |
3529 | static inline void bfqg_stats_update_dequeue(struct bfq_group *bfqg) { } | |
3530 | static inline void bfqg_stats_set_start_empty_time(struct bfq_group *bfqg) { } | |
3531 | static inline void bfqg_stats_update_idle_time(struct bfq_group *bfqg) { } | |
3532 | static inline void bfqg_stats_set_start_idle_time(struct bfq_group *bfqg) { } | |
3533 | static inline void bfqg_stats_update_avg_queue_size(struct bfq_group *bfqg) { } | |
3534 | ||
3535 | static void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
3536 | struct bfq_group *bfqg) {} | |
3537 | ||
3538 | static void bfq_init_entity(struct bfq_entity *entity, | |
3539 | struct bfq_group *bfqg) | |
aee69d78 PV |
3540 | { |
3541 | struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); | |
3542 | ||
3543 | entity->weight = entity->new_weight; | |
3544 | entity->orig_weight = entity->new_weight; | |
e21b7a0b AA |
3545 | if (bfqq) { |
3546 | bfqq->ioprio = bfqq->new_ioprio; | |
3547 | bfqq->ioprio_class = bfqq->new_ioprio_class; | |
3548 | } | |
3549 | entity->sched_data = &bfqg->sched_data; | |
3550 | } | |
3551 | ||
3552 | static void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio) {} | |
3553 | ||
44e44a1b PV |
3554 | static void bfq_end_wr_async(struct bfq_data *bfqd) |
3555 | { | |
3556 | bfq_end_wr_async_queues(bfqd, bfqd->root_group); | |
3557 | } | |
3558 | ||
e21b7a0b AA |
3559 | static struct bfq_group *bfq_find_set_group(struct bfq_data *bfqd, |
3560 | struct blkcg *blkcg) | |
3561 | { | |
3562 | return bfqd->root_group; | |
3563 | } | |
3564 | ||
3565 | static struct bfq_group *bfqq_group(struct bfq_queue *bfqq) | |
3566 | { | |
3567 | return bfqq->bfqd->root_group; | |
3568 | } | |
aee69d78 | 3569 | |
e21b7a0b AA |
3570 | static struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, |
3571 | int node) | |
3572 | { | |
3573 | struct bfq_group *bfqg; | |
3574 | int i; | |
3575 | ||
3576 | bfqg = kmalloc_node(sizeof(*bfqg), GFP_KERNEL | __GFP_ZERO, node); | |
3577 | if (!bfqg) | |
3578 | return NULL; | |
aee69d78 | 3579 | |
e21b7a0b AA |
3580 | for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) |
3581 | bfqg->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT; | |
3582 | ||
3583 | return bfqg; | |
aee69d78 | 3584 | } |
e21b7a0b | 3585 | #endif /* CONFIG_BFQ_GROUP_IOSCHED */ |
aee69d78 PV |
3586 | |
3587 | #define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE) | |
3588 | #define bfq_class_rt(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_RT) | |
3589 | ||
3590 | #define bfq_sample_valid(samples) ((samples) > 80) | |
3591 | ||
aee69d78 PV |
3592 | /* |
3593 | * Lifted from AS - choose which of rq1 and rq2 that is best served now. | |
3594 | * We choose the request that is closesr to the head right now. Distance | |
3595 | * behind the head is penalized and only allowed to a certain extent. | |
3596 | */ | |
3597 | static struct request *bfq_choose_req(struct bfq_data *bfqd, | |
3598 | struct request *rq1, | |
3599 | struct request *rq2, | |
3600 | sector_t last) | |
3601 | { | |
3602 | sector_t s1, s2, d1 = 0, d2 = 0; | |
3603 | unsigned long back_max; | |
3604 | #define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */ | |
3605 | #define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */ | |
3606 | unsigned int wrap = 0; /* bit mask: requests behind the disk head? */ | |
3607 | ||
3608 | if (!rq1 || rq1 == rq2) | |
3609 | return rq2; | |
3610 | if (!rq2) | |
3611 | return rq1; | |
3612 | ||
3613 | if (rq_is_sync(rq1) && !rq_is_sync(rq2)) | |
3614 | return rq1; | |
3615 | else if (rq_is_sync(rq2) && !rq_is_sync(rq1)) | |
3616 | return rq2; | |
3617 | if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META)) | |
3618 | return rq1; | |
3619 | else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META)) | |
3620 | return rq2; | |
3621 | ||
3622 | s1 = blk_rq_pos(rq1); | |
3623 | s2 = blk_rq_pos(rq2); | |
3624 | ||
3625 | /* | |
3626 | * By definition, 1KiB is 2 sectors. | |
3627 | */ | |
3628 | back_max = bfqd->bfq_back_max * 2; | |
3629 | ||
3630 | /* | |
3631 | * Strict one way elevator _except_ in the case where we allow | |
3632 | * short backward seeks which are biased as twice the cost of a | |
3633 | * similar forward seek. | |
3634 | */ | |
3635 | if (s1 >= last) | |
3636 | d1 = s1 - last; | |
3637 | else if (s1 + back_max >= last) | |
3638 | d1 = (last - s1) * bfqd->bfq_back_penalty; | |
3639 | else | |
3640 | wrap |= BFQ_RQ1_WRAP; | |
3641 | ||
3642 | if (s2 >= last) | |
3643 | d2 = s2 - last; | |
3644 | else if (s2 + back_max >= last) | |
3645 | d2 = (last - s2) * bfqd->bfq_back_penalty; | |
3646 | else | |
3647 | wrap |= BFQ_RQ2_WRAP; | |
3648 | ||
3649 | /* Found required data */ | |
3650 | ||
3651 | /* | |
3652 | * By doing switch() on the bit mask "wrap" we avoid having to | |
3653 | * check two variables for all permutations: --> faster! | |
3654 | */ | |
3655 | switch (wrap) { | |
3656 | case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ | |
3657 | if (d1 < d2) | |
3658 | return rq1; | |
3659 | else if (d2 < d1) | |
3660 | return rq2; | |
3661 | ||
3662 | if (s1 >= s2) | |
3663 | return rq1; | |
3664 | else | |
3665 | return rq2; | |
3666 | ||
3667 | case BFQ_RQ2_WRAP: | |
3668 | return rq1; | |
3669 | case BFQ_RQ1_WRAP: | |
3670 | return rq2; | |
3671 | case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */ | |
3672 | default: | |
3673 | /* | |
3674 | * Since both rqs are wrapped, | |
3675 | * start with the one that's further behind head | |
3676 | * (--> only *one* back seek required), | |
3677 | * since back seek takes more time than forward. | |
3678 | */ | |
3679 | if (s1 <= s2) | |
3680 | return rq1; | |
3681 | else | |
3682 | return rq2; | |
3683 | } | |
3684 | } | |
3685 | ||
3686 | /* | |
3687 | * Return expired entry, or NULL to just start from scratch in rbtree. | |
3688 | */ | |
3689 | static struct request *bfq_check_fifo(struct bfq_queue *bfqq, | |
3690 | struct request *last) | |
3691 | { | |
3692 | struct request *rq; | |
3693 | ||
3694 | if (bfq_bfqq_fifo_expire(bfqq)) | |
3695 | return NULL; | |
3696 | ||
3697 | bfq_mark_bfqq_fifo_expire(bfqq); | |
3698 | ||
3699 | rq = rq_entry_fifo(bfqq->fifo.next); | |
3700 | ||
3701 | if (rq == last || ktime_get_ns() < rq->fifo_time) | |
3702 | return NULL; | |
3703 | ||
3704 | bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq); | |
3705 | return rq; | |
3706 | } | |
3707 | ||
3708 | static struct request *bfq_find_next_rq(struct bfq_data *bfqd, | |
3709 | struct bfq_queue *bfqq, | |
3710 | struct request *last) | |
3711 | { | |
3712 | struct rb_node *rbnext = rb_next(&last->rb_node); | |
3713 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
3714 | struct request *next, *prev = NULL; | |
3715 | ||
3716 | /* Follow expired path, else get first next available. */ | |
3717 | next = bfq_check_fifo(bfqq, last); | |
3718 | if (next) | |
3719 | return next; | |
3720 | ||
3721 | if (rbprev) | |
3722 | prev = rb_entry_rq(rbprev); | |
3723 | ||
3724 | if (rbnext) | |
3725 | next = rb_entry_rq(rbnext); | |
3726 | else { | |
3727 | rbnext = rb_first(&bfqq->sort_list); | |
3728 | if (rbnext && rbnext != &last->rb_node) | |
3729 | next = rb_entry_rq(rbnext); | |
3730 | } | |
3731 | ||
3732 | return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last)); | |
3733 | } | |
3734 | ||
c074170e | 3735 | /* see the definition of bfq_async_charge_factor for details */ |
aee69d78 PV |
3736 | static unsigned long bfq_serv_to_charge(struct request *rq, |
3737 | struct bfq_queue *bfqq) | |
3738 | { | |
44e44a1b | 3739 | if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1) |
c074170e PV |
3740 | return blk_rq_sectors(rq); |
3741 | ||
3742 | return blk_rq_sectors(rq) * bfq_async_charge_factor; | |
aee69d78 PV |
3743 | } |
3744 | ||
3745 | /** | |
3746 | * bfq_updated_next_req - update the queue after a new next_rq selection. | |
3747 | * @bfqd: the device data the queue belongs to. | |
3748 | * @bfqq: the queue to update. | |
3749 | * | |
3750 | * If the first request of a queue changes we make sure that the queue | |
3751 | * has enough budget to serve at least its first request (if the | |
3752 | * request has grown). We do this because if the queue has not enough | |
3753 | * budget for its first request, it has to go through two dispatch | |
3754 | * rounds to actually get it dispatched. | |
3755 | */ | |
3756 | static void bfq_updated_next_req(struct bfq_data *bfqd, | |
3757 | struct bfq_queue *bfqq) | |
3758 | { | |
3759 | struct bfq_entity *entity = &bfqq->entity; | |
3760 | struct request *next_rq = bfqq->next_rq; | |
3761 | unsigned long new_budget; | |
3762 | ||
3763 | if (!next_rq) | |
3764 | return; | |
3765 | ||
3766 | if (bfqq == bfqd->in_service_queue) | |
3767 | /* | |
3768 | * In order not to break guarantees, budgets cannot be | |
3769 | * changed after an entity has been selected. | |
3770 | */ | |
3771 | return; | |
3772 | ||
3773 | new_budget = max_t(unsigned long, bfqq->max_budget, | |
3774 | bfq_serv_to_charge(next_rq, bfqq)); | |
3775 | if (entity->budget != new_budget) { | |
3776 | entity->budget = new_budget; | |
3777 | bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu", | |
3778 | new_budget); | |
e21b7a0b | 3779 | bfq_requeue_bfqq(bfqd, bfqq); |
aee69d78 PV |
3780 | } |
3781 | } | |
3782 | ||
3783 | static int bfq_bfqq_budget_left(struct bfq_queue *bfqq) | |
3784 | { | |
3785 | struct bfq_entity *entity = &bfqq->entity; | |
3786 | ||
3787 | return entity->budget - entity->service; | |
3788 | } | |
3789 | ||
3790 | /* | |
3791 | * If enough samples have been computed, return the current max budget | |
3792 | * stored in bfqd, which is dynamically updated according to the | |
3793 | * estimated disk peak rate; otherwise return the default max budget | |
3794 | */ | |
3795 | static int bfq_max_budget(struct bfq_data *bfqd) | |
3796 | { | |
3797 | if (bfqd->budgets_assigned < bfq_stats_min_budgets) | |
3798 | return bfq_default_max_budget; | |
3799 | else | |
3800 | return bfqd->bfq_max_budget; | |
3801 | } | |
3802 | ||
3803 | /* | |
3804 | * Return min budget, which is a fraction of the current or default | |
3805 | * max budget (trying with 1/32) | |
3806 | */ | |
3807 | static int bfq_min_budget(struct bfq_data *bfqd) | |
3808 | { | |
3809 | if (bfqd->budgets_assigned < bfq_stats_min_budgets) | |
3810 | return bfq_default_max_budget / 32; | |
3811 | else | |
3812 | return bfqd->bfq_max_budget / 32; | |
3813 | } | |
3814 | ||
3815 | static void bfq_bfqq_expire(struct bfq_data *bfqd, | |
3816 | struct bfq_queue *bfqq, | |
3817 | bool compensate, | |
3818 | enum bfqq_expiration reason); | |
3819 | ||
3820 | /* | |
3821 | * The next function, invoked after the input queue bfqq switches from | |
3822 | * idle to busy, updates the budget of bfqq. The function also tells | |
3823 | * whether the in-service queue should be expired, by returning | |
3824 | * true. The purpose of expiring the in-service queue is to give bfqq | |
3825 | * the chance to possibly preempt the in-service queue, and the reason | |
44e44a1b PV |
3826 | * for preempting the in-service queue is to achieve one of the two |
3827 | * goals below. | |
aee69d78 | 3828 | * |
44e44a1b PV |
3829 | * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has |
3830 | * expired because it has remained idle. In particular, bfqq may have | |
3831 | * expired for one of the following two reasons: | |
aee69d78 PV |
3832 | * |
3833 | * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling | |
3834 | * and did not make it to issue a new request before its last | |
3835 | * request was served; | |
3836 | * | |
3837 | * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue | |
3838 | * a new request before the expiration of the idling-time. | |
3839 | * | |
3840 | * Even if bfqq has expired for one of the above reasons, the process | |
3841 | * associated with the queue may be however issuing requests greedily, | |
3842 | * and thus be sensitive to the bandwidth it receives (bfqq may have | |
3843 | * remained idle for other reasons: CPU high load, bfqq not enjoying | |
3844 | * idling, I/O throttling somewhere in the path from the process to | |
3845 | * the I/O scheduler, ...). But if, after every expiration for one of | |
3846 | * the above two reasons, bfqq has to wait for the service of at least | |
3847 | * one full budget of another queue before being served again, then | |
3848 | * bfqq is likely to get a much lower bandwidth or resource time than | |
3849 | * its reserved ones. To address this issue, two countermeasures need | |
3850 | * to be taken. | |
3851 | * | |
3852 | * First, the budget and the timestamps of bfqq need to be updated in | |
3853 | * a special way on bfqq reactivation: they need to be updated as if | |
3854 | * bfqq did not remain idle and did not expire. In fact, if they are | |
3855 | * computed as if bfqq expired and remained idle until reactivation, | |
3856 | * then the process associated with bfqq is treated as if, instead of | |
3857 | * being greedy, it stopped issuing requests when bfqq remained idle, | |
3858 | * and restarts issuing requests only on this reactivation. In other | |
3859 | * words, the scheduler does not help the process recover the "service | |
3860 | * hole" between bfqq expiration and reactivation. As a consequence, | |
3861 | * the process receives a lower bandwidth than its reserved one. In | |
3862 | * contrast, to recover this hole, the budget must be updated as if | |
3863 | * bfqq was not expired at all before this reactivation, i.e., it must | |
3864 | * be set to the value of the remaining budget when bfqq was | |
3865 | * expired. Along the same line, timestamps need to be assigned the | |
3866 | * value they had the last time bfqq was selected for service, i.e., | |
3867 | * before last expiration. Thus timestamps need to be back-shifted | |
3868 | * with respect to their normal computation (see [1] for more details | |
3869 | * on this tricky aspect). | |
3870 | * | |
3871 | * Secondly, to allow the process to recover the hole, the in-service | |
3872 | * queue must be expired too, to give bfqq the chance to preempt it | |
3873 | * immediately. In fact, if bfqq has to wait for a full budget of the | |
3874 | * in-service queue to be completed, then it may become impossible to | |
3875 | * let the process recover the hole, even if the back-shifted | |
3876 | * timestamps of bfqq are lower than those of the in-service queue. If | |
3877 | * this happens for most or all of the holes, then the process may not | |
3878 | * receive its reserved bandwidth. In this respect, it is worth noting | |
3879 | * that, being the service of outstanding requests unpreemptible, a | |
3880 | * little fraction of the holes may however be unrecoverable, thereby | |
3881 | * causing a little loss of bandwidth. | |
3882 | * | |
3883 | * The last important point is detecting whether bfqq does need this | |
3884 | * bandwidth recovery. In this respect, the next function deems the | |
3885 | * process associated with bfqq greedy, and thus allows it to recover | |
3886 | * the hole, if: 1) the process is waiting for the arrival of a new | |
3887 | * request (which implies that bfqq expired for one of the above two | |
3888 | * reasons), and 2) such a request has arrived soon. The first | |
3889 | * condition is controlled through the flag non_blocking_wait_rq, | |
3890 | * while the second through the flag arrived_in_time. If both | |
3891 | * conditions hold, then the function computes the budget in the | |
3892 | * above-described special way, and signals that the in-service queue | |
3893 | * should be expired. Timestamp back-shifting is done later in | |
3894 | * __bfq_activate_entity. | |
44e44a1b PV |
3895 | * |
3896 | * 2. Reduce latency. Even if timestamps are not backshifted to let | |
3897 | * the process associated with bfqq recover a service hole, bfqq may | |
3898 | * however happen to have, after being (re)activated, a lower finish | |
3899 | * timestamp than the in-service queue. That is, the next budget of | |
3900 | * bfqq may have to be completed before the one of the in-service | |
3901 | * queue. If this is the case, then preempting the in-service queue | |
3902 | * allows this goal to be achieved, apart from the unpreemptible, | |
3903 | * outstanding requests mentioned above. | |
3904 | * | |
3905 | * Unfortunately, regardless of which of the above two goals one wants | |
3906 | * to achieve, service trees need first to be updated to know whether | |
3907 | * the in-service queue must be preempted. To have service trees | |
3908 | * correctly updated, the in-service queue must be expired and | |
3909 | * rescheduled, and bfqq must be scheduled too. This is one of the | |
3910 | * most costly operations (in future versions, the scheduling | |
3911 | * mechanism may be re-designed in such a way to make it possible to | |
3912 | * know whether preemption is needed without needing to update service | |
3913 | * trees). In addition, queue preemptions almost always cause random | |
3914 | * I/O, and thus loss of throughput. Because of these facts, the next | |
3915 | * function adopts the following simple scheme to avoid both costly | |
3916 | * operations and too frequent preemptions: it requests the expiration | |
3917 | * of the in-service queue (unconditionally) only for queues that need | |
3918 | * to recover a hole, or that either are weight-raised or deserve to | |
3919 | * be weight-raised. | |
aee69d78 PV |
3920 | */ |
3921 | static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd, | |
3922 | struct bfq_queue *bfqq, | |
44e44a1b PV |
3923 | bool arrived_in_time, |
3924 | bool wr_or_deserves_wr) | |
aee69d78 PV |
3925 | { |
3926 | struct bfq_entity *entity = &bfqq->entity; | |
3927 | ||
3928 | if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time) { | |
3929 | /* | |
3930 | * We do not clear the flag non_blocking_wait_rq here, as | |
3931 | * the latter is used in bfq_activate_bfqq to signal | |
3932 | * that timestamps need to be back-shifted (and is | |
3933 | * cleared right after). | |
3934 | */ | |
3935 | ||
3936 | /* | |
3937 | * In next assignment we rely on that either | |
3938 | * entity->service or entity->budget are not updated | |
3939 | * on expiration if bfqq is empty (see | |
3940 | * __bfq_bfqq_recalc_budget). Thus both quantities | |
3941 | * remain unchanged after such an expiration, and the | |
3942 | * following statement therefore assigns to | |
3943 | * entity->budget the remaining budget on such an | |
3944 | * expiration. For clarity, entity->service is not | |
3945 | * updated on expiration in any case, and, in normal | |
3946 | * operation, is reset only when bfqq is selected for | |
3947 | * service (see bfq_get_next_queue). | |
3948 | */ | |
3949 | entity->budget = min_t(unsigned long, | |
3950 | bfq_bfqq_budget_left(bfqq), | |
3951 | bfqq->max_budget); | |
3952 | ||
3953 | return true; | |
3954 | } | |
3955 | ||
3956 | entity->budget = max_t(unsigned long, bfqq->max_budget, | |
3957 | bfq_serv_to_charge(bfqq->next_rq, bfqq)); | |
3958 | bfq_clear_bfqq_non_blocking_wait_rq(bfqq); | |
44e44a1b PV |
3959 | return wr_or_deserves_wr; |
3960 | } | |
3961 | ||
3962 | static unsigned int bfq_wr_duration(struct bfq_data *bfqd) | |
3963 | { | |
3964 | u64 dur; | |
3965 | ||
3966 | if (bfqd->bfq_wr_max_time > 0) | |
3967 | return bfqd->bfq_wr_max_time; | |
3968 | ||
3969 | dur = bfqd->RT_prod; | |
3970 | do_div(dur, bfqd->peak_rate); | |
3971 | ||
3972 | /* | |
3973 | * Limit duration between 3 and 13 seconds. Tests show that | |
3974 | * higher values than 13 seconds often yield the opposite of | |
3975 | * the desired result, i.e., worsen responsiveness by letting | |
3976 | * non-interactive and non-soft-real-time applications | |
3977 | * preserve weight raising for a too long time interval. | |
3978 | * | |
3979 | * On the other end, lower values than 3 seconds make it | |
3980 | * difficult for most interactive tasks to complete their jobs | |
3981 | * before weight-raising finishes. | |
3982 | */ | |
3983 | if (dur > msecs_to_jiffies(13000)) | |
3984 | dur = msecs_to_jiffies(13000); | |
3985 | else if (dur < msecs_to_jiffies(3000)) | |
3986 | dur = msecs_to_jiffies(3000); | |
3987 | ||
3988 | return dur; | |
3989 | } | |
3990 | ||
3991 | static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd, | |
3992 | struct bfq_queue *bfqq, | |
3993 | unsigned int old_wr_coeff, | |
3994 | bool wr_or_deserves_wr, | |
3995 | bool interactive) | |
3996 | { | |
3997 | if (old_wr_coeff == 1 && wr_or_deserves_wr) { | |
3998 | /* start a weight-raising period */ | |
3999 | bfqq->wr_coeff = bfqd->bfq_wr_coeff; | |
4000 | /* update wr duration */ | |
4001 | bfqq->wr_cur_max_time = bfq_wr_duration(bfqd); | |
4002 | ||
4003 | /* | |
4004 | * If needed, further reduce budget to make sure it is | |
4005 | * close to bfqq's backlog, so as to reduce the | |
4006 | * scheduling-error component due to a too large | |
4007 | * budget. Do not care about throughput consequences, | |
4008 | * but only about latency. Finally, do not assign a | |
4009 | * too small budget either, to avoid increasing | |
4010 | * latency by causing too frequent expirations. | |
4011 | */ | |
4012 | bfqq->entity.budget = min_t(unsigned long, | |
4013 | bfqq->entity.budget, | |
4014 | 2 * bfq_min_budget(bfqd)); | |
4015 | } else if (old_wr_coeff > 1) { | |
4016 | /* update wr duration */ | |
4017 | bfqq->wr_cur_max_time = bfq_wr_duration(bfqd); | |
4018 | } | |
4019 | } | |
4020 | ||
4021 | static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd, | |
4022 | struct bfq_queue *bfqq) | |
4023 | { | |
4024 | return bfqq->dispatched == 0 && | |
4025 | time_is_before_jiffies( | |
4026 | bfqq->budget_timeout + | |
4027 | bfqd->bfq_wr_min_idle_time); | |
aee69d78 PV |
4028 | } |
4029 | ||
4030 | static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd, | |
4031 | struct bfq_queue *bfqq, | |
44e44a1b PV |
4032 | int old_wr_coeff, |
4033 | struct request *rq, | |
4034 | bool *interactive) | |
aee69d78 | 4035 | { |
44e44a1b PV |
4036 | bool wr_or_deserves_wr, bfqq_wants_to_preempt, |
4037 | idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq), | |
aee69d78 PV |
4038 | /* |
4039 | * See the comments on | |
4040 | * bfq_bfqq_update_budg_for_activation for | |
4041 | * details on the usage of the next variable. | |
4042 | */ | |
4043 | arrived_in_time = ktime_get_ns() <= | |
4044 | bfqq->ttime.last_end_request + | |
4045 | bfqd->bfq_slice_idle * 3; | |
4046 | ||
e21b7a0b AA |
4047 | bfqg_stats_update_io_add(bfqq_group(RQ_BFQQ(rq)), bfqq, rq->cmd_flags); |
4048 | ||
aee69d78 | 4049 | /* |
44e44a1b PV |
4050 | * bfqq deserves to be weight-raised if: |
4051 | * - it is sync, | |
4052 | * - it has been idle for enough time. | |
4053 | */ | |
4054 | *interactive = idle_for_long_time; | |
4055 | wr_or_deserves_wr = bfqd->low_latency && | |
4056 | (bfqq->wr_coeff > 1 || | |
4057 | (bfq_bfqq_sync(bfqq) && *interactive)); | |
4058 | ||
4059 | /* | |
4060 | * Using the last flag, update budget and check whether bfqq | |
4061 | * may want to preempt the in-service queue. | |
aee69d78 PV |
4062 | */ |
4063 | bfqq_wants_to_preempt = | |
4064 | bfq_bfqq_update_budg_for_activation(bfqd, bfqq, | |
44e44a1b PV |
4065 | arrived_in_time, |
4066 | wr_or_deserves_wr); | |
aee69d78 PV |
4067 | |
4068 | if (!bfq_bfqq_IO_bound(bfqq)) { | |
4069 | if (arrived_in_time) { | |
4070 | bfqq->requests_within_timer++; | |
4071 | if (bfqq->requests_within_timer >= | |
4072 | bfqd->bfq_requests_within_timer) | |
4073 | bfq_mark_bfqq_IO_bound(bfqq); | |
4074 | } else | |
4075 | bfqq->requests_within_timer = 0; | |
4076 | } | |
4077 | ||
44e44a1b PV |
4078 | if (bfqd->low_latency) { |
4079 | bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq, | |
4080 | old_wr_coeff, | |
4081 | wr_or_deserves_wr, | |
4082 | *interactive); | |
4083 | ||
4084 | if (old_wr_coeff != bfqq->wr_coeff) | |
4085 | bfqq->entity.prio_changed = 1; | |
4086 | } | |
4087 | ||
aee69d78 PV |
4088 | bfq_add_bfqq_busy(bfqd, bfqq); |
4089 | ||
4090 | /* | |
4091 | * Expire in-service queue only if preemption may be needed | |
4092 | * for guarantees. In this respect, the function | |
4093 | * next_queue_may_preempt just checks a simple, necessary | |
4094 | * condition, and not a sufficient condition based on | |
4095 | * timestamps. In fact, for the latter condition to be | |
4096 | * evaluated, timestamps would need first to be updated, and | |
4097 | * this operation is quite costly (see the comments on the | |
4098 | * function bfq_bfqq_update_budg_for_activation). | |
4099 | */ | |
4100 | if (bfqd->in_service_queue && bfqq_wants_to_preempt && | |
44e44a1b | 4101 | bfqd->in_service_queue->wr_coeff == 1 && |
aee69d78 PV |
4102 | next_queue_may_preempt(bfqd)) |
4103 | bfq_bfqq_expire(bfqd, bfqd->in_service_queue, | |
4104 | false, BFQQE_PREEMPTED); | |
4105 | } | |
4106 | ||
4107 | static void bfq_add_request(struct request *rq) | |
4108 | { | |
4109 | struct bfq_queue *bfqq = RQ_BFQQ(rq); | |
4110 | struct bfq_data *bfqd = bfqq->bfqd; | |
4111 | struct request *next_rq, *prev; | |
44e44a1b PV |
4112 | unsigned int old_wr_coeff = bfqq->wr_coeff; |
4113 | bool interactive = false; | |
aee69d78 PV |
4114 | |
4115 | bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq)); | |
4116 | bfqq->queued[rq_is_sync(rq)]++; | |
4117 | bfqd->queued++; | |
4118 | ||
4119 | elv_rb_add(&bfqq->sort_list, rq); | |
4120 | ||
4121 | /* | |
4122 | * Check if this request is a better next-serve candidate. | |
4123 | */ | |
4124 | prev = bfqq->next_rq; | |
4125 | next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position); | |
4126 | bfqq->next_rq = next_rq; | |
4127 | ||
4128 | if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */ | |
44e44a1b PV |
4129 | bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff, |
4130 | rq, &interactive); | |
4131 | else { | |
4132 | if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) && | |
4133 | time_is_before_jiffies( | |
4134 | bfqq->last_wr_start_finish + | |
4135 | bfqd->bfq_wr_min_inter_arr_async)) { | |
4136 | bfqq->wr_coeff = bfqd->bfq_wr_coeff; | |
4137 | bfqq->wr_cur_max_time = bfq_wr_duration(bfqd); | |
4138 | ||
4139 | bfqq->entity.prio_changed = 1; | |
4140 | } | |
4141 | if (prev != bfqq->next_rq) | |
4142 | bfq_updated_next_req(bfqd, bfqq); | |
4143 | } | |
4144 | ||
4145 | /* | |
4146 | * Assign jiffies to last_wr_start_finish in the following | |
4147 | * cases: | |
4148 | * | |
4149 | * . if bfqq is not going to be weight-raised, because, for | |
4150 | * non weight-raised queues, last_wr_start_finish stores the | |
4151 | * arrival time of the last request; as of now, this piece | |
4152 | * of information is used only for deciding whether to | |
4153 | * weight-raise async queues | |
4154 | * | |
4155 | * . if bfqq is not weight-raised, because, if bfqq is now | |
4156 | * switching to weight-raised, then last_wr_start_finish | |
4157 | * stores the time when weight-raising starts | |
4158 | * | |
4159 | * . if bfqq is interactive, because, regardless of whether | |
4160 | * bfqq is currently weight-raised, the weight-raising | |
4161 | * period must start or restart (this case is considered | |
4162 | * separately because it is not detected by the above | |
4163 | * conditions, if bfqq is already weight-raised) | |
4164 | */ | |
4165 | if (bfqd->low_latency && | |
4166 | (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive)) | |
4167 | bfqq->last_wr_start_finish = jiffies; | |
aee69d78 PV |
4168 | } |
4169 | ||
4170 | static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd, | |
4171 | struct bio *bio, | |
4172 | struct request_queue *q) | |
4173 | { | |
4174 | struct bfq_queue *bfqq = bfqd->bio_bfqq; | |
4175 | ||
4176 | ||
4177 | if (bfqq) | |
4178 | return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio)); | |
4179 | ||
4180 | return NULL; | |
4181 | } | |
4182 | ||
ab0e43e9 PV |
4183 | static sector_t get_sdist(sector_t last_pos, struct request *rq) |
4184 | { | |
4185 | if (last_pos) | |
4186 | return abs(blk_rq_pos(rq) - last_pos); | |
4187 | ||
4188 | return 0; | |
4189 | } | |
4190 | ||
aee69d78 PV |
4191 | #if 0 /* Still not clear if we can do without next two functions */ |
4192 | static void bfq_activate_request(struct request_queue *q, struct request *rq) | |
4193 | { | |
4194 | struct bfq_data *bfqd = q->elevator->elevator_data; | |
4195 | ||
4196 | bfqd->rq_in_driver++; | |
aee69d78 PV |
4197 | } |
4198 | ||
4199 | static void bfq_deactivate_request(struct request_queue *q, struct request *rq) | |
4200 | { | |
4201 | struct bfq_data *bfqd = q->elevator->elevator_data; | |
4202 | ||
4203 | bfqd->rq_in_driver--; | |
4204 | } | |
4205 | #endif | |
4206 | ||
4207 | static void bfq_remove_request(struct request_queue *q, | |
4208 | struct request *rq) | |
4209 | { | |
4210 | struct bfq_queue *bfqq = RQ_BFQQ(rq); | |
4211 | struct bfq_data *bfqd = bfqq->bfqd; | |
4212 | const int sync = rq_is_sync(rq); | |
4213 | ||
4214 | if (bfqq->next_rq == rq) { | |
4215 | bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq); | |
4216 | bfq_updated_next_req(bfqd, bfqq); | |
4217 | } | |
4218 | ||
4219 | if (rq->queuelist.prev != &rq->queuelist) | |
4220 | list_del_init(&rq->queuelist); | |
4221 | bfqq->queued[sync]--; | |
4222 | bfqd->queued--; | |
4223 | elv_rb_del(&bfqq->sort_list, rq); | |
4224 | ||
4225 | elv_rqhash_del(q, rq); | |
4226 | if (q->last_merge == rq) | |
4227 | q->last_merge = NULL; | |
4228 | ||
4229 | if (RB_EMPTY_ROOT(&bfqq->sort_list)) { | |
4230 | bfqq->next_rq = NULL; | |
4231 | ||
4232 | if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) { | |
e21b7a0b | 4233 | bfq_del_bfqq_busy(bfqd, bfqq, false); |
aee69d78 PV |
4234 | /* |
4235 | * bfqq emptied. In normal operation, when | |
4236 | * bfqq is empty, bfqq->entity.service and | |
4237 | * bfqq->entity.budget must contain, | |
4238 | * respectively, the service received and the | |
4239 | * budget used last time bfqq emptied. These | |
4240 | * facts do not hold in this case, as at least | |
4241 | * this last removal occurred while bfqq is | |
4242 | * not in service. To avoid inconsistencies, | |
4243 | * reset both bfqq->entity.service and | |
4244 | * bfqq->entity.budget, if bfqq has still a | |
4245 | * process that may issue I/O requests to it. | |
4246 | */ | |
4247 | bfqq->entity.budget = bfqq->entity.service = 0; | |
4248 | } | |
4249 | } | |
4250 | ||
4251 | if (rq->cmd_flags & REQ_META) | |
4252 | bfqq->meta_pending--; | |
e21b7a0b AA |
4253 | |
4254 | bfqg_stats_update_io_remove(bfqq_group(bfqq), rq->cmd_flags); | |
aee69d78 PV |
4255 | } |
4256 | ||
4257 | static bool bfq_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio) | |
4258 | { | |
4259 | struct request_queue *q = hctx->queue; | |
4260 | struct bfq_data *bfqd = q->elevator->elevator_data; | |
4261 | struct request *free = NULL; | |
4262 | /* | |
4263 | * bfq_bic_lookup grabs the queue_lock: invoke it now and | |
4264 | * store its return value for later use, to avoid nesting | |
4265 | * queue_lock inside the bfqd->lock. We assume that the bic | |
4266 | * returned by bfq_bic_lookup does not go away before | |
4267 | * bfqd->lock is taken. | |
4268 | */ | |
4269 | struct bfq_io_cq *bic = bfq_bic_lookup(bfqd, current->io_context, q); | |
4270 | bool ret; | |
4271 | ||
4272 | spin_lock_irq(&bfqd->lock); | |
4273 | ||
4274 | if (bic) | |
4275 | bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf)); | |
4276 | else | |
4277 | bfqd->bio_bfqq = NULL; | |
4278 | bfqd->bio_bic = bic; | |
4279 | ||
4280 | ret = blk_mq_sched_try_merge(q, bio, &free); | |
4281 | ||
4282 | if (free) | |
4283 | blk_mq_free_request(free); | |
4284 | spin_unlock_irq(&bfqd->lock); | |
4285 | ||
4286 | return ret; | |
4287 | } | |
4288 | ||
4289 | static int bfq_request_merge(struct request_queue *q, struct request **req, | |
4290 | struct bio *bio) | |
4291 | { | |
4292 | struct bfq_data *bfqd = q->elevator->elevator_data; | |
4293 | struct request *__rq; | |
4294 | ||
4295 | __rq = bfq_find_rq_fmerge(bfqd, bio, q); | |
4296 | if (__rq && elv_bio_merge_ok(__rq, bio)) { | |
4297 | *req = __rq; | |
4298 | return ELEVATOR_FRONT_MERGE; | |
4299 | } | |
4300 | ||
4301 | return ELEVATOR_NO_MERGE; | |
4302 | } | |
4303 | ||
4304 | static void bfq_request_merged(struct request_queue *q, struct request *req, | |
4305 | enum elv_merge type) | |
4306 | { | |
4307 | if (type == ELEVATOR_FRONT_MERGE && | |
4308 | rb_prev(&req->rb_node) && | |
4309 | blk_rq_pos(req) < | |
4310 | blk_rq_pos(container_of(rb_prev(&req->rb_node), | |
4311 | struct request, rb_node))) { | |
4312 | struct bfq_queue *bfqq = RQ_BFQQ(req); | |
4313 | struct bfq_data *bfqd = bfqq->bfqd; | |
4314 | struct request *prev, *next_rq; | |
4315 | ||
4316 | /* Reposition request in its sort_list */ | |
4317 | elv_rb_del(&bfqq->sort_list, req); | |
4318 | elv_rb_add(&bfqq->sort_list, req); | |
4319 | ||
4320 | /* Choose next request to be served for bfqq */ | |
4321 | prev = bfqq->next_rq; | |
4322 | next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req, | |
4323 | bfqd->last_position); | |
4324 | bfqq->next_rq = next_rq; | |
4325 | /* | |
4326 | * If next_rq changes, update the queue's budget to fit | |
4327 | * the new request. | |
4328 | */ | |
4329 | if (prev != bfqq->next_rq) | |
4330 | bfq_updated_next_req(bfqd, bfqq); | |
4331 | } | |
4332 | } | |
4333 | ||
4334 | static void bfq_requests_merged(struct request_queue *q, struct request *rq, | |
4335 | struct request *next) | |
4336 | { | |
4337 | struct bfq_queue *bfqq = RQ_BFQQ(rq), *next_bfqq = RQ_BFQQ(next); | |
4338 | ||
4339 | if (!RB_EMPTY_NODE(&rq->rb_node)) | |
e21b7a0b | 4340 | goto end; |
aee69d78 PV |
4341 | spin_lock_irq(&bfqq->bfqd->lock); |
4342 | ||
4343 | /* | |
4344 | * If next and rq belong to the same bfq_queue and next is older | |
4345 | * than rq, then reposition rq in the fifo (by substituting next | |
4346 | * with rq). Otherwise, if next and rq belong to different | |
4347 | * bfq_queues, never reposition rq: in fact, we would have to | |
4348 | * reposition it with respect to next's position in its own fifo, | |
4349 | * which would most certainly be too expensive with respect to | |
4350 | * the benefits. | |
4351 | */ | |
4352 | if (bfqq == next_bfqq && | |
4353 | !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && | |
4354 | next->fifo_time < rq->fifo_time) { | |
4355 | list_del_init(&rq->queuelist); | |
4356 | list_replace_init(&next->queuelist, &rq->queuelist); | |
4357 | rq->fifo_time = next->fifo_time; | |
4358 | } | |
4359 | ||
4360 | if (bfqq->next_rq == next) | |
4361 | bfqq->next_rq = rq; | |
4362 | ||
4363 | bfq_remove_request(q, next); | |
4364 | ||
4365 | spin_unlock_irq(&bfqq->bfqd->lock); | |
e21b7a0b AA |
4366 | end: |
4367 | bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags); | |
aee69d78 PV |
4368 | } |
4369 | ||
44e44a1b PV |
4370 | /* Must be called with bfqq != NULL */ |
4371 | static void bfq_bfqq_end_wr(struct bfq_queue *bfqq) | |
4372 | { | |
4373 | bfqq->wr_coeff = 1; | |
4374 | bfqq->wr_cur_max_time = 0; | |
4375 | /* | |
4376 | * Trigger a weight change on the next invocation of | |
4377 | * __bfq_entity_update_weight_prio. | |
4378 | */ | |
4379 | bfqq->entity.prio_changed = 1; | |
4380 | } | |
4381 | ||
4382 | static void bfq_end_wr_async_queues(struct bfq_data *bfqd, | |
4383 | struct bfq_group *bfqg) | |
4384 | { | |
4385 | int i, j; | |
4386 | ||
4387 | for (i = 0; i < 2; i++) | |
4388 | for (j = 0; j < IOPRIO_BE_NR; j++) | |
4389 | if (bfqg->async_bfqq[i][j]) | |
4390 | bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]); | |
4391 | if (bfqg->async_idle_bfqq) | |
4392 | bfq_bfqq_end_wr(bfqg->async_idle_bfqq); | |
4393 | } | |
4394 | ||
4395 | static void bfq_end_wr(struct bfq_data *bfqd) | |
4396 | { | |
4397 | struct bfq_queue *bfqq; | |
4398 | ||
4399 | spin_lock_irq(&bfqd->lock); | |
4400 | ||
4401 | list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list) | |
4402 | bfq_bfqq_end_wr(bfqq); | |
4403 | list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list) | |
4404 | bfq_bfqq_end_wr(bfqq); | |
4405 | bfq_end_wr_async(bfqd); | |
4406 | ||
4407 | spin_unlock_irq(&bfqd->lock); | |
4408 | } | |
4409 | ||
aee69d78 PV |
4410 | static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq, |
4411 | struct bio *bio) | |
4412 | { | |
4413 | struct bfq_data *bfqd = q->elevator->elevator_data; | |
4414 | bool is_sync = op_is_sync(bio->bi_opf); | |
4415 | struct bfq_queue *bfqq = bfqd->bio_bfqq; | |
4416 | ||
4417 | /* | |
4418 | * Disallow merge of a sync bio into an async request. | |
4419 | */ | |
4420 | if (is_sync && !rq_is_sync(rq)) | |
4421 | return false; | |
4422 | ||
4423 | /* | |
4424 | * Lookup the bfqq that this bio will be queued with. Allow | |
4425 | * merge only if rq is queued there. | |
4426 | */ | |
4427 | if (!bfqq) | |
4428 | return false; | |
4429 | ||
4430 | return bfqq == RQ_BFQQ(rq); | |
4431 | } | |
4432 | ||
44e44a1b PV |
4433 | /* |
4434 | * Set the maximum time for the in-service queue to consume its | |
4435 | * budget. This prevents seeky processes from lowering the throughput. | |
4436 | * In practice, a time-slice service scheme is used with seeky | |
4437 | * processes. | |
4438 | */ | |
4439 | static void bfq_set_budget_timeout(struct bfq_data *bfqd, | |
4440 | struct bfq_queue *bfqq) | |
4441 | { | |
4442 | bfqd->last_budget_start = ktime_get(); | |
4443 | ||
4444 | bfqq->budget_timeout = jiffies + | |
4445 | bfqd->bfq_timeout * | |
4446 | (bfqq->entity.weight / bfqq->entity.orig_weight); | |
4447 | } | |
4448 | ||
aee69d78 PV |
4449 | static void __bfq_set_in_service_queue(struct bfq_data *bfqd, |
4450 | struct bfq_queue *bfqq) | |
4451 | { | |
4452 | if (bfqq) { | |
e21b7a0b | 4453 | bfqg_stats_update_avg_queue_size(bfqq_group(bfqq)); |
aee69d78 PV |
4454 | bfq_clear_bfqq_fifo_expire(bfqq); |
4455 | ||
4456 | bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8; | |
4457 | ||
44e44a1b | 4458 | bfq_set_budget_timeout(bfqd, bfqq); |
aee69d78 PV |
4459 | bfq_log_bfqq(bfqd, bfqq, |
4460 | "set_in_service_queue, cur-budget = %d", | |
4461 | bfqq->entity.budget); | |
4462 | } | |
4463 | ||
4464 | bfqd->in_service_queue = bfqq; | |
4465 | } | |
4466 | ||
4467 | /* | |
4468 | * Get and set a new queue for service. | |
4469 | */ | |
4470 | static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd) | |
4471 | { | |
4472 | struct bfq_queue *bfqq = bfq_get_next_queue(bfqd); | |
4473 | ||
4474 | __bfq_set_in_service_queue(bfqd, bfqq); | |
4475 | return bfqq; | |
4476 | } | |
4477 | ||
aee69d78 PV |
4478 | static void bfq_arm_slice_timer(struct bfq_data *bfqd) |
4479 | { | |
4480 | struct bfq_queue *bfqq = bfqd->in_service_queue; | |
4481 | struct bfq_io_cq *bic; | |
4482 | u32 sl; | |
4483 | ||
4484 | /* Processes have exited, don't wait. */ | |
4485 | bic = bfqd->in_service_bic; | |
4486 | if (!bic || atomic_read(&bic->icq.ioc->active_ref) == 0) | |
4487 | return; | |
4488 | ||
4489 | bfq_mark_bfqq_wait_request(bfqq); | |
4490 | ||
4491 | /* | |
4492 | * We don't want to idle for seeks, but we do want to allow | |
4493 | * fair distribution of slice time for a process doing back-to-back | |
4494 | * seeks. So allow a little bit of time for him to submit a new rq. | |
4495 | */ | |
4496 | sl = bfqd->bfq_slice_idle; | |
4497 | /* | |
44e44a1b PV |
4498 | * Unless the queue is being weight-raised, grant only minimum |
4499 | * idle time if the queue is seeky. A long idling is preserved | |
4500 | * for a weight-raised queue, because it is needed for | |
4501 | * guaranteeing to the queue its reserved share of the | |
4502 | * throughput. | |
aee69d78 | 4503 | */ |
44e44a1b | 4504 | if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1) |
aee69d78 PV |
4505 | sl = min_t(u64, sl, BFQ_MIN_TT); |
4506 | ||
4507 | bfqd->last_idling_start = ktime_get(); | |
4508 | hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl), | |
4509 | HRTIMER_MODE_REL); | |
e21b7a0b | 4510 | bfqg_stats_set_start_idle_time(bfqq_group(bfqq)); |
aee69d78 PV |
4511 | } |
4512 | ||
ab0e43e9 PV |
4513 | /* |
4514 | * In autotuning mode, max_budget is dynamically recomputed as the | |
4515 | * amount of sectors transferred in timeout at the estimated peak | |
4516 | * rate. This enables BFQ to utilize a full timeslice with a full | |
4517 | * budget, even if the in-service queue is served at peak rate. And | |
4518 | * this maximises throughput with sequential workloads. | |
4519 | */ | |
4520 | static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd) | |
4521 | { | |
4522 | return (u64)bfqd->peak_rate * USEC_PER_MSEC * | |
4523 | jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT; | |
4524 | } | |
4525 | ||
44e44a1b PV |
4526 | /* |
4527 | * Update parameters related to throughput and responsiveness, as a | |
4528 | * function of the estimated peak rate. See comments on | |
4529 | * bfq_calc_max_budget(), and on T_slow and T_fast arrays. | |
4530 | */ | |
4531 | static void update_thr_responsiveness_params(struct bfq_data *bfqd) | |
4532 | { | |
4533 | int dev_type = blk_queue_nonrot(bfqd->queue); | |
4534 | ||
4535 | if (bfqd->bfq_user_max_budget == 0) | |
4536 | bfqd->bfq_max_budget = | |
4537 | bfq_calc_max_budget(bfqd); | |
4538 | ||
4539 | if (bfqd->device_speed == BFQ_BFQD_FAST && | |
4540 | bfqd->peak_rate < device_speed_thresh[dev_type]) { | |
4541 | bfqd->device_speed = BFQ_BFQD_SLOW; | |
4542 | bfqd->RT_prod = R_slow[dev_type] * | |
4543 | T_slow[dev_type]; | |
4544 | } else if (bfqd->device_speed == BFQ_BFQD_SLOW && | |
4545 | bfqd->peak_rate > device_speed_thresh[dev_type]) { | |
4546 | bfqd->device_speed = BFQ_BFQD_FAST; | |
4547 | bfqd->RT_prod = R_fast[dev_type] * | |
4548 | T_fast[dev_type]; | |
4549 | } | |
4550 | ||
4551 | bfq_log(bfqd, | |
4552 | "dev_type %s dev_speed_class = %s (%llu sects/sec), thresh %llu setcs/sec", | |
4553 | dev_type == 0 ? "ROT" : "NONROT", | |
4554 | bfqd->device_speed == BFQ_BFQD_FAST ? "FAST" : "SLOW", | |
4555 | bfqd->device_speed == BFQ_BFQD_FAST ? | |
4556 | (USEC_PER_SEC*(u64)R_fast[dev_type])>>BFQ_RATE_SHIFT : | |
4557 | (USEC_PER_SEC*(u64)R_slow[dev_type])>>BFQ_RATE_SHIFT, | |
4558 | (USEC_PER_SEC*(u64)device_speed_thresh[dev_type])>> | |
4559 | BFQ_RATE_SHIFT); | |
4560 | } | |
4561 | ||
ab0e43e9 PV |
4562 | static void bfq_reset_rate_computation(struct bfq_data *bfqd, |
4563 | struct request *rq) | |
4564 | { | |
4565 | if (rq != NULL) { /* new rq dispatch now, reset accordingly */ | |
4566 | bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns(); | |
4567 | bfqd->peak_rate_samples = 1; | |
4568 | bfqd->sequential_samples = 0; | |
4569 | bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size = | |
4570 | blk_rq_sectors(rq); | |
4571 | } else /* no new rq dispatched, just reset the number of samples */ | |
4572 | bfqd->peak_rate_samples = 0; /* full re-init on next disp. */ | |
4573 | ||
4574 | bfq_log(bfqd, | |
4575 | "reset_rate_computation at end, sample %u/%u tot_sects %llu", | |
4576 | bfqd->peak_rate_samples, bfqd->sequential_samples, | |
4577 | bfqd->tot_sectors_dispatched); | |
4578 | } | |
4579 | ||
4580 | static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq) | |
4581 | { | |
4582 | u32 rate, weight, divisor; | |
4583 | ||
4584 | /* | |
4585 | * For the convergence property to hold (see comments on | |
4586 | * bfq_update_peak_rate()) and for the assessment to be | |
4587 | * reliable, a minimum number of samples must be present, and | |
4588 | * a minimum amount of time must have elapsed. If not so, do | |
4589 | * not compute new rate. Just reset parameters, to get ready | |
4590 | * for a new evaluation attempt. | |
4591 | */ | |
4592 | if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES || | |
4593 | bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL) | |
4594 | goto reset_computation; | |
4595 | ||
4596 | /* | |
4597 | * If a new request completion has occurred after last | |
4598 | * dispatch, then, to approximate the rate at which requests | |
4599 | * have been served by the device, it is more precise to | |
4600 | * extend the observation interval to the last completion. | |
4601 | */ | |
4602 | bfqd->delta_from_first = | |
4603 | max_t(u64, bfqd->delta_from_first, | |
4604 | bfqd->last_completion - bfqd->first_dispatch); | |
4605 | ||
4606 | /* | |
4607 | * Rate computed in sects/usec, and not sects/nsec, for | |
4608 | * precision issues. | |
4609 | */ | |
4610 | rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT, | |
4611 | div_u64(bfqd->delta_from_first, NSEC_PER_USEC)); | |
4612 | ||
4613 | /* | |
4614 | * Peak rate not updated if: | |
4615 | * - the percentage of sequential dispatches is below 3/4 of the | |
4616 | * total, and rate is below the current estimated peak rate | |
4617 | * - rate is unreasonably high (> 20M sectors/sec) | |
4618 | */ | |
4619 | if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 && | |
4620 | rate <= bfqd->peak_rate) || | |
4621 | rate > 20<<BFQ_RATE_SHIFT) | |
4622 | goto reset_computation; | |
4623 | ||
4624 | /* | |
4625 | * We have to update the peak rate, at last! To this purpose, | |
4626 | * we use a low-pass filter. We compute the smoothing constant | |
4627 | * of the filter as a function of the 'weight' of the new | |
4628 | * measured rate. | |
4629 | * | |
4630 | * As can be seen in next formulas, we define this weight as a | |
4631 | * quantity proportional to how sequential the workload is, | |
4632 | * and to how long the observation time interval is. | |
4633 | * | |
4634 | * The weight runs from 0 to 8. The maximum value of the | |
4635 | * weight, 8, yields the minimum value for the smoothing | |
4636 | * constant. At this minimum value for the smoothing constant, | |
4637 | * the measured rate contributes for half of the next value of | |
4638 | * the estimated peak rate. | |
4639 | * | |
4640 | * So, the first step is to compute the weight as a function | |
4641 | * of how sequential the workload is. Note that the weight | |
4642 | * cannot reach 9, because bfqd->sequential_samples cannot | |
4643 | * become equal to bfqd->peak_rate_samples, which, in its | |
4644 | * turn, holds true because bfqd->sequential_samples is not | |
4645 | * incremented for the first sample. | |
4646 | */ | |
4647 | weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples; | |
4648 | ||
4649 | /* | |
4650 | * Second step: further refine the weight as a function of the | |
4651 | * duration of the observation interval. | |
4652 | */ | |
4653 | weight = min_t(u32, 8, | |
4654 | div_u64(weight * bfqd->delta_from_first, | |
4655 | BFQ_RATE_REF_INTERVAL)); | |
4656 | ||
4657 | /* | |
4658 | * Divisor ranging from 10, for minimum weight, to 2, for | |
4659 | * maximum weight. | |
4660 | */ | |
4661 | divisor = 10 - weight; | |
4662 | ||
4663 | /* | |
4664 | * Finally, update peak rate: | |
4665 | * | |
4666 | * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor | |
4667 | */ | |
4668 | bfqd->peak_rate *= divisor-1; | |
4669 | bfqd->peak_rate /= divisor; | |
4670 | rate /= divisor; /* smoothing constant alpha = 1/divisor */ | |
4671 | ||
4672 | bfqd->peak_rate += rate; | |
44e44a1b | 4673 | update_thr_responsiveness_params(bfqd); |
ab0e43e9 PV |
4674 | |
4675 | reset_computation: | |
4676 | bfq_reset_rate_computation(bfqd, rq); | |
4677 | } | |
4678 | ||
4679 | /* | |
4680 | * Update the read/write peak rate (the main quantity used for | |
4681 | * auto-tuning, see update_thr_responsiveness_params()). | |
4682 | * | |
4683 | * It is not trivial to estimate the peak rate (correctly): because of | |
4684 | * the presence of sw and hw queues between the scheduler and the | |
4685 | * device components that finally serve I/O requests, it is hard to | |
4686 | * say exactly when a given dispatched request is served inside the | |
4687 | * device, and for how long. As a consequence, it is hard to know | |
4688 | * precisely at what rate a given set of requests is actually served | |
4689 | * by the device. | |
4690 | * | |
4691 | * On the opposite end, the dispatch time of any request is trivially | |
4692 | * available, and, from this piece of information, the "dispatch rate" | |
4693 | * of requests can be immediately computed. So, the idea in the next | |
4694 | * function is to use what is known, namely request dispatch times | |
4695 | * (plus, when useful, request completion times), to estimate what is | |
4696 | * unknown, namely in-device request service rate. | |
4697 | * | |
4698 | * The main issue is that, because of the above facts, the rate at | |
4699 | * which a certain set of requests is dispatched over a certain time | |
4700 | * interval can vary greatly with respect to the rate at which the | |
4701 | * same requests are then served. But, since the size of any | |
4702 | * intermediate queue is limited, and the service scheme is lossless | |
4703 | * (no request is silently dropped), the following obvious convergence | |
4704 | * property holds: the number of requests dispatched MUST become | |
4705 | * closer and closer to the number of requests completed as the | |
4706 | * observation interval grows. This is the key property used in | |
4707 | * the next function to estimate the peak service rate as a function | |
4708 | * of the observed dispatch rate. The function assumes to be invoked | |
4709 | * on every request dispatch. | |
4710 | */ | |
4711 | static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq) | |
4712 | { | |
4713 | u64 now_ns = ktime_get_ns(); | |
4714 | ||
4715 | if (bfqd->peak_rate_samples == 0) { /* first dispatch */ | |
4716 | bfq_log(bfqd, "update_peak_rate: goto reset, samples %d", | |
4717 | bfqd->peak_rate_samples); | |
4718 | bfq_reset_rate_computation(bfqd, rq); | |
4719 | goto update_last_values; /* will add one sample */ | |
4720 | } | |
4721 | ||
4722 | /* | |
4723 | * Device idle for very long: the observation interval lasting | |
4724 | * up to this dispatch cannot be a valid observation interval | |
4725 | * for computing a new peak rate (similarly to the late- | |
4726 | * completion event in bfq_completed_request()). Go to | |
4727 | * update_rate_and_reset to have the following three steps | |
4728 | * taken: | |
4729 | * - close the observation interval at the last (previous) | |
4730 | * request dispatch or completion | |
4731 | * - compute rate, if possible, for that observation interval | |
4732 | * - start a new observation interval with this dispatch | |
4733 | */ | |
4734 | if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC && | |
4735 | bfqd->rq_in_driver == 0) | |
4736 | goto update_rate_and_reset; | |
4737 | ||
4738 | /* Update sampling information */ | |
4739 | bfqd->peak_rate_samples++; | |
4740 | ||
4741 | if ((bfqd->rq_in_driver > 0 || | |
4742 | now_ns - bfqd->last_completion < BFQ_MIN_TT) | |
4743 | && get_sdist(bfqd->last_position, rq) < BFQQ_SEEK_THR) | |
4744 | bfqd->sequential_samples++; | |
4745 | ||
4746 | bfqd->tot_sectors_dispatched += blk_rq_sectors(rq); | |
4747 | ||
4748 | /* Reset max observed rq size every 32 dispatches */ | |
4749 | if (likely(bfqd->peak_rate_samples % 32)) | |
4750 | bfqd->last_rq_max_size = | |
4751 | max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size); | |
4752 | else | |
4753 | bfqd->last_rq_max_size = blk_rq_sectors(rq); | |
4754 | ||
4755 | bfqd->delta_from_first = now_ns - bfqd->first_dispatch; | |
4756 | ||
4757 | /* Target observation interval not yet reached, go on sampling */ | |
4758 | if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL) | |
4759 | goto update_last_values; | |
4760 | ||
4761 | update_rate_and_reset: | |
4762 | bfq_update_rate_reset(bfqd, rq); | |
4763 | update_last_values: | |
4764 | bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); | |
4765 | bfqd->last_dispatch = now_ns; | |
4766 | } | |
4767 | ||
aee69d78 PV |
4768 | /* |
4769 | * Remove request from internal lists. | |
4770 | */ | |
4771 | static void bfq_dispatch_remove(struct request_queue *q, struct request *rq) | |
4772 | { | |
4773 | struct bfq_queue *bfqq = RQ_BFQQ(rq); | |
4774 | ||
4775 | /* | |
4776 | * For consistency, the next instruction should have been | |
4777 | * executed after removing the request from the queue and | |
4778 | * dispatching it. We execute instead this instruction before | |
4779 | * bfq_remove_request() (and hence introduce a temporary | |
4780 | * inconsistency), for efficiency. In fact, should this | |
4781 | * dispatch occur for a non in-service bfqq, this anticipated | |
4782 | * increment prevents two counters related to bfqq->dispatched | |
4783 | * from risking to be, first, uselessly decremented, and then | |
4784 | * incremented again when the (new) value of bfqq->dispatched | |
4785 | * happens to be taken into account. | |
4786 | */ | |
4787 | bfqq->dispatched++; | |
ab0e43e9 | 4788 | bfq_update_peak_rate(q->elevator->elevator_data, rq); |
aee69d78 PV |
4789 | |
4790 | bfq_remove_request(q, rq); | |
4791 | } | |
4792 | ||
4793 | static void __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq) | |
4794 | { | |
44e44a1b PV |
4795 | if (RB_EMPTY_ROOT(&bfqq->sort_list)) { |
4796 | if (bfqq->dispatched == 0) | |
4797 | /* | |
4798 | * Overloading budget_timeout field to store | |
4799 | * the time at which the queue remains with no | |
4800 | * backlog and no outstanding request; used by | |
4801 | * the weight-raising mechanism. | |
4802 | */ | |
4803 | bfqq->budget_timeout = jiffies; | |
4804 | ||
e21b7a0b | 4805 | bfq_del_bfqq_busy(bfqd, bfqq, true); |
44e44a1b | 4806 | } else |
e21b7a0b AA |
4807 | bfq_requeue_bfqq(bfqd, bfqq); |
4808 | ||
4809 | /* | |
4810 | * All in-service entities must have been properly deactivated | |
4811 | * or requeued before executing the next function, which | |
4812 | * resets all in-service entites as no more in service. | |
4813 | */ | |
4814 | __bfq_bfqd_reset_in_service(bfqd); | |
aee69d78 PV |
4815 | } |
4816 | ||
4817 | /** | |
4818 | * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior. | |
4819 | * @bfqd: device data. | |
4820 | * @bfqq: queue to update. | |
4821 | * @reason: reason for expiration. | |
4822 | * | |
4823 | * Handle the feedback on @bfqq budget at queue expiration. | |
4824 | * See the body for detailed comments. | |
4825 | */ | |
4826 | static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd, | |
4827 | struct bfq_queue *bfqq, | |
4828 | enum bfqq_expiration reason) | |
4829 | { | |
4830 | struct request *next_rq; | |
4831 | int budget, min_budget; | |
4832 | ||
aee69d78 PV |
4833 | min_budget = bfq_min_budget(bfqd); |
4834 | ||
44e44a1b PV |
4835 | if (bfqq->wr_coeff == 1) |
4836 | budget = bfqq->max_budget; | |
4837 | else /* | |
4838 | * Use a constant, low budget for weight-raised queues, | |
4839 | * to help achieve a low latency. Keep it slightly higher | |
4840 | * than the minimum possible budget, to cause a little | |
4841 | * bit fewer expirations. | |
4842 | */ | |
4843 | budget = 2 * min_budget; | |
4844 | ||
aee69d78 PV |
4845 | bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d", |
4846 | bfqq->entity.budget, bfq_bfqq_budget_left(bfqq)); | |
4847 | bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d", | |
4848 | budget, bfq_min_budget(bfqd)); | |
4849 | bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d", | |
4850 | bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue)); | |
4851 | ||
44e44a1b | 4852 | if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) { |
aee69d78 PV |
4853 | switch (reason) { |
4854 | /* | |
4855 | * Caveat: in all the following cases we trade latency | |
4856 | * for throughput. | |
4857 | */ | |
4858 | case BFQQE_TOO_IDLE: | |
54b60456 PV |
4859 | /* |
4860 | * This is the only case where we may reduce | |
4861 | * the budget: if there is no request of the | |
4862 | * process still waiting for completion, then | |
4863 | * we assume (tentatively) that the timer has | |
4864 | * expired because the batch of requests of | |
4865 | * the process could have been served with a | |
4866 | * smaller budget. Hence, betting that | |
4867 | * process will behave in the same way when it | |
4868 | * becomes backlogged again, we reduce its | |
4869 | * next budget. As long as we guess right, | |
4870 | * this budget cut reduces the latency | |
4871 | * experienced by the process. | |
4872 | * | |
4873 | * However, if there are still outstanding | |
4874 | * requests, then the process may have not yet | |
4875 | * issued its next request just because it is | |
4876 | * still waiting for the completion of some of | |
4877 | * the still outstanding ones. So in this | |
4878 | * subcase we do not reduce its budget, on the | |
4879 | * contrary we increase it to possibly boost | |
4880 | * the throughput, as discussed in the | |
4881 | * comments to the BUDGET_TIMEOUT case. | |
4882 | */ | |
4883 | if (bfqq->dispatched > 0) /* still outstanding reqs */ | |
4884 | budget = min(budget * 2, bfqd->bfq_max_budget); | |
4885 | else { | |
4886 | if (budget > 5 * min_budget) | |
4887 | budget -= 4 * min_budget; | |
4888 | else | |
4889 | budget = min_budget; | |
4890 | } | |
aee69d78 PV |
4891 | break; |
4892 | case BFQQE_BUDGET_TIMEOUT: | |
54b60456 PV |
4893 | /* |
4894 | * We double the budget here because it gives | |
4895 | * the chance to boost the throughput if this | |
4896 | * is not a seeky process (and has bumped into | |
4897 | * this timeout because of, e.g., ZBR). | |
4898 | */ | |
4899 | budget = min(budget * 2, bfqd->bfq_max_budget); | |
aee69d78 PV |
4900 | break; |
4901 | case BFQQE_BUDGET_EXHAUSTED: | |
4902 | /* | |
4903 | * The process still has backlog, and did not | |
4904 | * let either the budget timeout or the disk | |
4905 | * idling timeout expire. Hence it is not | |
4906 | * seeky, has a short thinktime and may be | |
4907 | * happy with a higher budget too. So | |
4908 | * definitely increase the budget of this good | |
4909 | * candidate to boost the disk throughput. | |
4910 | */ | |
54b60456 | 4911 | budget = min(budget * 4, bfqd->bfq_max_budget); |
aee69d78 PV |
4912 | break; |
4913 | case BFQQE_NO_MORE_REQUESTS: | |
4914 | /* | |
4915 | * For queues that expire for this reason, it | |
4916 | * is particularly important to keep the | |
4917 | * budget close to the actual service they | |
4918 | * need. Doing so reduces the timestamp | |
4919 | * misalignment problem described in the | |
4920 | * comments in the body of | |
4921 | * __bfq_activate_entity. In fact, suppose | |
4922 | * that a queue systematically expires for | |
4923 | * BFQQE_NO_MORE_REQUESTS and presents a | |
4924 | * new request in time to enjoy timestamp | |
4925 | * back-shifting. The larger the budget of the | |
4926 | * queue is with respect to the service the | |
4927 | * queue actually requests in each service | |
4928 | * slot, the more times the queue can be | |
4929 | * reactivated with the same virtual finish | |
4930 | * time. It follows that, even if this finish | |
4931 | * time is pushed to the system virtual time | |
4932 | * to reduce the consequent timestamp | |
4933 | * misalignment, the queue unjustly enjoys for | |
4934 | * many re-activations a lower finish time | |
4935 | * than all newly activated queues. | |
4936 | * | |
4937 | * The service needed by bfqq is measured | |
4938 | * quite precisely by bfqq->entity.service. | |
4939 | * Since bfqq does not enjoy device idling, | |
4940 | * bfqq->entity.service is equal to the number | |
4941 | * of sectors that the process associated with | |
4942 | * bfqq requested to read/write before waiting | |
4943 | * for request completions, or blocking for | |
4944 | * other reasons. | |
4945 | */ | |
4946 | budget = max_t(int, bfqq->entity.service, min_budget); | |
4947 | break; | |
4948 | default: | |
4949 | return; | |
4950 | } | |
44e44a1b | 4951 | } else if (!bfq_bfqq_sync(bfqq)) { |
aee69d78 PV |
4952 | /* |
4953 | * Async queues get always the maximum possible | |
4954 | * budget, as for them we do not care about latency | |
4955 | * (in addition, their ability to dispatch is limited | |
4956 | * by the charging factor). | |
4957 | */ | |
4958 | budget = bfqd->bfq_max_budget; | |
4959 | } | |
4960 | ||
4961 | bfqq->max_budget = budget; | |
4962 | ||
4963 | if (bfqd->budgets_assigned >= bfq_stats_min_budgets && | |
4964 | !bfqd->bfq_user_max_budget) | |
4965 | bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget); | |
4966 | ||
4967 | /* | |
4968 | * If there is still backlog, then assign a new budget, making | |
4969 | * sure that it is large enough for the next request. Since | |
4970 | * the finish time of bfqq must be kept in sync with the | |
4971 | * budget, be sure to call __bfq_bfqq_expire() *after* this | |
4972 | * update. | |
4973 | * | |
4974 | * If there is no backlog, then no need to update the budget; | |
4975 | * it will be updated on the arrival of a new request. | |
4976 | */ | |
4977 | next_rq = bfqq->next_rq; | |
4978 | if (next_rq) | |
4979 | bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget, | |
4980 | bfq_serv_to_charge(next_rq, bfqq)); | |
4981 | ||
4982 | bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d", | |
4983 | next_rq ? blk_rq_sectors(next_rq) : 0, | |
4984 | bfqq->entity.budget); | |
4985 | } | |
4986 | ||
aee69d78 | 4987 | /* |
ab0e43e9 PV |
4988 | * Return true if the process associated with bfqq is "slow". The slow |
4989 | * flag is used, in addition to the budget timeout, to reduce the | |
4990 | * amount of service provided to seeky processes, and thus reduce | |
4991 | * their chances to lower the throughput. More details in the comments | |
4992 | * on the function bfq_bfqq_expire(). | |
4993 | * | |
4994 | * An important observation is in order: as discussed in the comments | |
4995 | * on the function bfq_update_peak_rate(), with devices with internal | |
4996 | * queues, it is hard if ever possible to know when and for how long | |
4997 | * an I/O request is processed by the device (apart from the trivial | |
4998 | * I/O pattern where a new request is dispatched only after the | |
4999 | * previous one has been completed). This makes it hard to evaluate | |
5000 | * the real rate at which the I/O requests of each bfq_queue are | |
5001 | * served. In fact, for an I/O scheduler like BFQ, serving a | |
5002 | * bfq_queue means just dispatching its requests during its service | |
5003 | * slot (i.e., until the budget of the queue is exhausted, or the | |
5004 | * queue remains idle, or, finally, a timeout fires). But, during the | |
5005 | * service slot of a bfq_queue, around 100 ms at most, the device may | |
5006 | * be even still processing requests of bfq_queues served in previous | |
5007 | * service slots. On the opposite end, the requests of the in-service | |
5008 | * bfq_queue may be completed after the service slot of the queue | |
5009 | * finishes. | |
5010 | * | |
5011 | * Anyway, unless more sophisticated solutions are used | |
5012 | * (where possible), the sum of the sizes of the requests dispatched | |
5013 | * during the service slot of a bfq_queue is probably the only | |
5014 | * approximation available for the service received by the bfq_queue | |
5015 | * during its service slot. And this sum is the quantity used in this | |
5016 | * function to evaluate the I/O speed of a process. | |
aee69d78 | 5017 | */ |
ab0e43e9 PV |
5018 | static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
5019 | bool compensate, enum bfqq_expiration reason, | |
5020 | unsigned long *delta_ms) | |
aee69d78 | 5021 | { |
ab0e43e9 PV |
5022 | ktime_t delta_ktime; |
5023 | u32 delta_usecs; | |
5024 | bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */ | |
aee69d78 | 5025 | |
ab0e43e9 | 5026 | if (!bfq_bfqq_sync(bfqq)) |
aee69d78 PV |
5027 | return false; |
5028 | ||
5029 | if (compensate) | |
ab0e43e9 | 5030 | delta_ktime = bfqd->last_idling_start; |
aee69d78 | 5031 | else |
ab0e43e9 PV |
5032 | delta_ktime = ktime_get(); |
5033 | delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start); | |
5034 | delta_usecs = ktime_to_us(delta_ktime); | |
aee69d78 PV |
5035 | |
5036 | /* don't use too short time intervals */ | |
ab0e43e9 PV |
5037 | if (delta_usecs < 1000) { |
5038 | if (blk_queue_nonrot(bfqd->queue)) | |
5039 | /* | |
5040 | * give same worst-case guarantees as idling | |
5041 | * for seeky | |
5042 | */ | |
5043 | *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC; | |
5044 | else /* charge at least one seek */ | |
5045 | *delta_ms = bfq_slice_idle / NSEC_PER_MSEC; | |
5046 | ||
5047 | return slow; | |
5048 | } | |
aee69d78 | 5049 | |
ab0e43e9 | 5050 | *delta_ms = delta_usecs / USEC_PER_MSEC; |
aee69d78 PV |
5051 | |
5052 | /* | |
ab0e43e9 PV |
5053 | * Use only long (> 20ms) intervals to filter out excessive |
5054 | * spikes in service rate estimation. | |
aee69d78 | 5055 | */ |
ab0e43e9 PV |
5056 | if (delta_usecs > 20000) { |
5057 | /* | |
5058 | * Caveat for rotational devices: processes doing I/O | |
5059 | * in the slower disk zones tend to be slow(er) even | |
5060 | * if not seeky. In this respect, the estimated peak | |
5061 | * rate is likely to be an average over the disk | |
5062 | * surface. Accordingly, to not be too harsh with | |
5063 | * unlucky processes, a process is deemed slow only if | |
5064 | * its rate has been lower than half of the estimated | |
5065 | * peak rate. | |
5066 | */ | |
5067 | slow = bfqq->entity.service < bfqd->bfq_max_budget / 2; | |
aee69d78 PV |
5068 | } |
5069 | ||
ab0e43e9 | 5070 | bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow); |
aee69d78 | 5071 | |
ab0e43e9 | 5072 | return slow; |
aee69d78 PV |
5073 | } |
5074 | ||
5075 | /* | |
5076 | * Return the farthest past time instant according to jiffies | |
5077 | * macros. | |
5078 | */ | |
5079 | static unsigned long bfq_smallest_from_now(void) | |
5080 | { | |
5081 | return jiffies - MAX_JIFFY_OFFSET; | |
5082 | } | |
5083 | ||
5084 | /** | |
5085 | * bfq_bfqq_expire - expire a queue. | |
5086 | * @bfqd: device owning the queue. | |
5087 | * @bfqq: the queue to expire. | |
5088 | * @compensate: if true, compensate for the time spent idling. | |
5089 | * @reason: the reason causing the expiration. | |
5090 | * | |
c074170e PV |
5091 | * If the process associated with bfqq does slow I/O (e.g., because it |
5092 | * issues random requests), we charge bfqq with the time it has been | |
5093 | * in service instead of the service it has received (see | |
5094 | * bfq_bfqq_charge_time for details on how this goal is achieved). As | |
5095 | * a consequence, bfqq will typically get higher timestamps upon | |
5096 | * reactivation, and hence it will be rescheduled as if it had | |
5097 | * received more service than what it has actually received. In the | |
5098 | * end, bfqq receives less service in proportion to how slowly its | |
5099 | * associated process consumes its budgets (and hence how seriously it | |
5100 | * tends to lower the throughput). In addition, this time-charging | |
5101 | * strategy guarantees time fairness among slow processes. In | |
5102 | * contrast, if the process associated with bfqq is not slow, we | |
5103 | * charge bfqq exactly with the service it has received. | |
aee69d78 | 5104 | * |
c074170e PV |
5105 | * Charging time to the first type of queues and the exact service to |
5106 | * the other has the effect of using the WF2Q+ policy to schedule the | |
5107 | * former on a timeslice basis, without violating service domain | |
5108 | * guarantees among the latter. | |
aee69d78 PV |
5109 | */ |
5110 | static void bfq_bfqq_expire(struct bfq_data *bfqd, | |
5111 | struct bfq_queue *bfqq, | |
5112 | bool compensate, | |
5113 | enum bfqq_expiration reason) | |
5114 | { | |
5115 | bool slow; | |
ab0e43e9 PV |
5116 | unsigned long delta = 0; |
5117 | struct bfq_entity *entity = &bfqq->entity; | |
aee69d78 PV |
5118 | int ref; |
5119 | ||
5120 | /* | |
ab0e43e9 | 5121 | * Check whether the process is slow (see bfq_bfqq_is_slow). |
aee69d78 | 5122 | */ |
ab0e43e9 | 5123 | slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta); |
aee69d78 PV |
5124 | |
5125 | /* | |
c074170e PV |
5126 | * As above explained, charge slow (typically seeky) and |
5127 | * timed-out queues with the time and not the service | |
5128 | * received, to favor sequential workloads. | |
5129 | * | |
5130 | * Processes doing I/O in the slower disk zones will tend to | |
5131 | * be slow(er) even if not seeky. Therefore, since the | |
5132 | * estimated peak rate is actually an average over the disk | |
5133 | * surface, these processes may timeout just for bad luck. To | |
5134 | * avoid punishing them, do not charge time to processes that | |
5135 | * succeeded in consuming at least 2/3 of their budget. This | |
5136 | * allows BFQ to preserve enough elasticity to still perform | |
5137 | * bandwidth, and not time, distribution with little unlucky | |
5138 | * or quasi-sequential processes. | |
aee69d78 | 5139 | */ |
44e44a1b PV |
5140 | if (bfqq->wr_coeff == 1 && |
5141 | (slow || | |
5142 | (reason == BFQQE_BUDGET_TIMEOUT && | |
5143 | bfq_bfqq_budget_left(bfqq) >= entity->budget / 3))) | |
c074170e | 5144 | bfq_bfqq_charge_time(bfqd, bfqq, delta); |
aee69d78 PV |
5145 | |
5146 | if (reason == BFQQE_TOO_IDLE && | |
ab0e43e9 | 5147 | entity->service <= 2 * entity->budget / 10) |
aee69d78 PV |
5148 | bfq_clear_bfqq_IO_bound(bfqq); |
5149 | ||
44e44a1b PV |
5150 | if (bfqd->low_latency && bfqq->wr_coeff == 1) |
5151 | bfqq->last_wr_start_finish = jiffies; | |
5152 | ||
aee69d78 PV |
5153 | bfq_log_bfqq(bfqd, bfqq, |
5154 | "expire (%d, slow %d, num_disp %d, idle_win %d)", reason, | |
5155 | slow, bfqq->dispatched, bfq_bfqq_idle_window(bfqq)); | |
5156 | ||
5157 | /* | |
5158 | * Increase, decrease or leave budget unchanged according to | |
5159 | * reason. | |
5160 | */ | |
5161 | __bfq_bfqq_recalc_budget(bfqd, bfqq, reason); | |
5162 | ref = bfqq->ref; | |
5163 | __bfq_bfqq_expire(bfqd, bfqq); | |
5164 | ||
5165 | /* mark bfqq as waiting a request only if a bic still points to it */ | |
5166 | if (ref > 1 && !bfq_bfqq_busy(bfqq) && | |
5167 | reason != BFQQE_BUDGET_TIMEOUT && | |
5168 | reason != BFQQE_BUDGET_EXHAUSTED) | |
5169 | bfq_mark_bfqq_non_blocking_wait_rq(bfqq); | |
5170 | } | |
5171 | ||
5172 | /* | |
5173 | * Budget timeout is not implemented through a dedicated timer, but | |
5174 | * just checked on request arrivals and completions, as well as on | |
5175 | * idle timer expirations. | |
5176 | */ | |
5177 | static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq) | |
5178 | { | |
44e44a1b | 5179 | return time_is_before_eq_jiffies(bfqq->budget_timeout); |
aee69d78 PV |
5180 | } |
5181 | ||
5182 | /* | |
5183 | * If we expire a queue that is actively waiting (i.e., with the | |
5184 | * device idled) for the arrival of a new request, then we may incur | |
5185 | * the timestamp misalignment problem described in the body of the | |
5186 | * function __bfq_activate_entity. Hence we return true only if this | |
5187 | * condition does not hold, or if the queue is slow enough to deserve | |
5188 | * only to be kicked off for preserving a high throughput. | |
5189 | */ | |
5190 | static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq) | |
5191 | { | |
5192 | bfq_log_bfqq(bfqq->bfqd, bfqq, | |
5193 | "may_budget_timeout: wait_request %d left %d timeout %d", | |
5194 | bfq_bfqq_wait_request(bfqq), | |
5195 | bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3, | |
5196 | bfq_bfqq_budget_timeout(bfqq)); | |
5197 | ||
5198 | return (!bfq_bfqq_wait_request(bfqq) || | |
5199 | bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3) | |
5200 | && | |
5201 | bfq_bfqq_budget_timeout(bfqq); | |
5202 | } | |
5203 | ||
5204 | /* | |
5205 | * For a queue that becomes empty, device idling is allowed only if | |
44e44a1b PV |
5206 | * this function returns true for the queue. As a consequence, since |
5207 | * device idling plays a critical role in both throughput boosting and | |
5208 | * service guarantees, the return value of this function plays a | |
5209 | * critical role in both these aspects as well. | |
5210 | * | |
5211 | * In a nutshell, this function returns true only if idling is | |
5212 | * beneficial for throughput or, even if detrimental for throughput, | |
5213 | * idling is however necessary to preserve service guarantees (low | |
5214 | * latency, desired throughput distribution, ...). In particular, on | |
5215 | * NCQ-capable devices, this function tries to return false, so as to | |
5216 | * help keep the drives' internal queues full, whenever this helps the | |
5217 | * device boost the throughput without causing any service-guarantee | |
5218 | * issue. | |
5219 | * | |
5220 | * In more detail, the return value of this function is obtained by, | |
5221 | * first, computing a number of boolean variables that take into | |
5222 | * account throughput and service-guarantee issues, and, then, | |
5223 | * combining these variables in a logical expression. Most of the | |
5224 | * issues taken into account are not trivial. We discuss these issues | |
5225 | * individually while introducing the variables. | |
aee69d78 PV |
5226 | */ |
5227 | static bool bfq_bfqq_may_idle(struct bfq_queue *bfqq) | |
5228 | { | |
5229 | struct bfq_data *bfqd = bfqq->bfqd; | |
44e44a1b | 5230 | bool idling_boosts_thr, asymmetric_scenario; |
aee69d78 PV |
5231 | |
5232 | if (bfqd->strict_guarantees) | |
5233 | return true; | |
5234 | ||
5235 | /* | |
44e44a1b PV |
5236 | * The next variable takes into account the cases where idling |
5237 | * boosts the throughput. | |
5238 | * | |
5239 | * The value of the variable is computed considering that | |
aee69d78 PV |
5240 | * idling is usually beneficial for the throughput if: |
5241 | * (a) the device is not NCQ-capable, or | |
5242 | * (b) regardless of the presence of NCQ, the request pattern | |
5243 | * for bfqq is I/O-bound (possible throughput losses | |
5244 | * caused by granting idling to seeky queues are mitigated | |
5245 | * by the fact that, in all scenarios where boosting | |
5246 | * throughput is the best thing to do, i.e., in all | |
5247 | * symmetric scenarios, only a minimal idle time is | |
5248 | * allowed to seeky queues). | |
5249 | */ | |
5250 | idling_boosts_thr = !bfqd->hw_tag || bfq_bfqq_IO_bound(bfqq); | |
5251 | ||
5252 | /* | |
44e44a1b PV |
5253 | * There is then a case where idling must be performed not for |
5254 | * throughput concerns, but to preserve service guarantees. To | |
5255 | * introduce it, we can note that allowing the drive to | |
5256 | * enqueue more than one request at a time, and hence | |
5257 | * delegating de facto final scheduling decisions to the | |
5258 | * drive's internal scheduler, causes loss of control on the | |
5259 | * actual request service order. In particular, the critical | |
5260 | * situation is when requests from different processes happens | |
5261 | * to be present, at the same time, in the internal queue(s) | |
5262 | * of the drive. In such a situation, the drive, by deciding | |
5263 | * the service order of the internally-queued requests, does | |
5264 | * determine also the actual throughput distribution among | |
5265 | * these processes. But the drive typically has no notion or | |
5266 | * concern about per-process throughput distribution, and | |
5267 | * makes its decisions only on a per-request basis. Therefore, | |
5268 | * the service distribution enforced by the drive's internal | |
5269 | * scheduler is likely to coincide with the desired | |
5270 | * device-throughput distribution only in a completely | |
5271 | * symmetric scenario where: (i) each of these processes must | |
5272 | * get the same throughput as the others; (ii) all these | |
5273 | * processes have the same I/O pattern (either sequential or | |
5274 | * random). In fact, in such a scenario, the drive will tend | |
5275 | * to treat the requests of each of these processes in about | |
5276 | * the same way as the requests of the others, and thus to | |
5277 | * provide each of these processes with about the same | |
5278 | * throughput (which is exactly the desired throughput | |
5279 | * distribution). In contrast, in any asymmetric scenario, | |
5280 | * device idling is certainly needed to guarantee that bfqq | |
5281 | * receives its assigned fraction of the device throughput | |
5282 | * (see [1] for details). | |
5283 | * | |
5284 | * As for sub-condition (i), actually we check only whether | |
5285 | * bfqq is being weight-raised. In fact, if bfqq is not being | |
5286 | * weight-raised, we have that: | |
5287 | * - if the process associated with bfqq is not I/O-bound, then | |
5288 | * it is not either latency- or throughput-critical; therefore | |
5289 | * idling is not needed for bfqq; | |
5290 | * - if the process asociated with bfqq is I/O-bound, then | |
5291 | * idling is already granted with bfqq (see the comments on | |
5292 | * idling_boosts_thr). | |
5293 | * | |
5294 | * We do not check sub-condition (ii) at all, i.e., the next | |
5295 | * variable is true if and only if bfqq is being | |
5296 | * weight-raised. We do not need to control sub-condition (ii) | |
5297 | * for the following reason: | |
5298 | * - if bfqq is being weight-raised, then idling is already | |
5299 | * guaranteed to bfqq by sub-condition (i); | |
5300 | * - if bfqq is not being weight-raised, then idling is | |
5301 | * already guaranteed to bfqq (only) if it matters, i.e., if | |
5302 | * bfqq is associated to a currently I/O-bound process (see | |
5303 | * the above comment on sub-condition (i)). | |
5304 | * | |
5305 | * As a side note, it is worth considering that the above | |
5306 | * device-idling countermeasures may however fail in the | |
5307 | * following unlucky scenario: if idling is (correctly) | |
5308 | * disabled in a time period during which the symmetry | |
5309 | * sub-condition holds, and hence the device is allowed to | |
5310 | * enqueue many requests, but at some later point in time some | |
5311 | * sub-condition stops to hold, then it may become impossible | |
5312 | * to let requests be served in the desired order until all | |
5313 | * the requests already queued in the device have been served. | |
5314 | */ | |
5315 | asymmetric_scenario = bfqq->wr_coeff > 1; | |
5316 | ||
5317 | /* | |
5318 | * We have now all the components we need to compute the return | |
5319 | * value of the function, which is true only if both the following | |
5320 | * conditions hold: | |
aee69d78 | 5321 | * 1) bfqq is sync, because idling make sense only for sync queues; |
44e44a1b PV |
5322 | * 2) idling either boosts the throughput (without issues), or |
5323 | * is necessary to preserve service guarantees. | |
aee69d78 | 5324 | */ |
44e44a1b PV |
5325 | return bfq_bfqq_sync(bfqq) && |
5326 | (idling_boosts_thr || asymmetric_scenario); | |
aee69d78 PV |
5327 | } |
5328 | ||
5329 | /* | |
5330 | * If the in-service queue is empty but the function bfq_bfqq_may_idle | |
5331 | * returns true, then: | |
5332 | * 1) the queue must remain in service and cannot be expired, and | |
5333 | * 2) the device must be idled to wait for the possible arrival of a new | |
5334 | * request for the queue. | |
5335 | * See the comments on the function bfq_bfqq_may_idle for the reasons | |
5336 | * why performing device idling is the best choice to boost the throughput | |
5337 | * and preserve service guarantees when bfq_bfqq_may_idle itself | |
5338 | * returns true. | |
5339 | */ | |
5340 | static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq) | |
5341 | { | |
5342 | struct bfq_data *bfqd = bfqq->bfqd; | |
5343 | ||
5344 | return RB_EMPTY_ROOT(&bfqq->sort_list) && bfqd->bfq_slice_idle != 0 && | |
5345 | bfq_bfqq_may_idle(bfqq); | |
5346 | } | |
5347 | ||
5348 | /* | |
5349 | * Select a queue for service. If we have a current queue in service, | |
5350 | * check whether to continue servicing it, or retrieve and set a new one. | |
5351 | */ | |
5352 | static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd) | |
5353 | { | |
5354 | struct bfq_queue *bfqq; | |
5355 | struct request *next_rq; | |
5356 | enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT; | |
5357 | ||
5358 | bfqq = bfqd->in_service_queue; | |
5359 | if (!bfqq) | |
5360 | goto new_queue; | |
5361 | ||
5362 | bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue"); | |
5363 | ||
5364 | if (bfq_may_expire_for_budg_timeout(bfqq) && | |
5365 | !bfq_bfqq_wait_request(bfqq) && | |
5366 | !bfq_bfqq_must_idle(bfqq)) | |
5367 | goto expire; | |
5368 | ||
5369 | check_queue: | |
5370 | /* | |
5371 | * This loop is rarely executed more than once. Even when it | |
5372 | * happens, it is much more convenient to re-execute this loop | |
5373 | * than to return NULL and trigger a new dispatch to get a | |
5374 | * request served. | |
5375 | */ | |
5376 | next_rq = bfqq->next_rq; | |
5377 | /* | |
5378 | * If bfqq has requests queued and it has enough budget left to | |
5379 | * serve them, keep the queue, otherwise expire it. | |
5380 | */ | |
5381 | if (next_rq) { | |
5382 | if (bfq_serv_to_charge(next_rq, bfqq) > | |
5383 | bfq_bfqq_budget_left(bfqq)) { | |
5384 | /* | |
5385 | * Expire the queue for budget exhaustion, | |
5386 | * which makes sure that the next budget is | |
5387 | * enough to serve the next request, even if | |
5388 | * it comes from the fifo expired path. | |
5389 | */ | |
5390 | reason = BFQQE_BUDGET_EXHAUSTED; | |
5391 | goto expire; | |
5392 | } else { | |
5393 | /* | |
5394 | * The idle timer may be pending because we may | |
5395 | * not disable disk idling even when a new request | |
5396 | * arrives. | |
5397 | */ | |
5398 | if (bfq_bfqq_wait_request(bfqq)) { | |
5399 | /* | |
5400 | * If we get here: 1) at least a new request | |
5401 | * has arrived but we have not disabled the | |
5402 | * timer because the request was too small, | |
5403 | * 2) then the block layer has unplugged | |
5404 | * the device, causing the dispatch to be | |
5405 | * invoked. | |
5406 | * | |
5407 | * Since the device is unplugged, now the | |
5408 | * requests are probably large enough to | |
5409 | * provide a reasonable throughput. | |
5410 | * So we disable idling. | |
5411 | */ | |
5412 | bfq_clear_bfqq_wait_request(bfqq); | |
5413 | hrtimer_try_to_cancel(&bfqd->idle_slice_timer); | |
e21b7a0b | 5414 | bfqg_stats_update_idle_time(bfqq_group(bfqq)); |
aee69d78 PV |
5415 | } |
5416 | goto keep_queue; | |
5417 | } | |
5418 | } | |
5419 | ||
5420 | /* | |
5421 | * No requests pending. However, if the in-service queue is idling | |
5422 | * for a new request, or has requests waiting for a completion and | |
5423 | * may idle after their completion, then keep it anyway. | |
5424 | */ | |
5425 | if (bfq_bfqq_wait_request(bfqq) || | |
5426 | (bfqq->dispatched != 0 && bfq_bfqq_may_idle(bfqq))) { | |
5427 | bfqq = NULL; | |
5428 | goto keep_queue; | |
5429 | } | |
5430 | ||
5431 | reason = BFQQE_NO_MORE_REQUESTS; | |
5432 | expire: | |
5433 | bfq_bfqq_expire(bfqd, bfqq, false, reason); | |
5434 | new_queue: | |
5435 | bfqq = bfq_set_in_service_queue(bfqd); | |
5436 | if (bfqq) { | |
5437 | bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue"); | |
5438 | goto check_queue; | |
5439 | } | |
5440 | keep_queue: | |
5441 | if (bfqq) | |
5442 | bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue"); | |
5443 | else | |
5444 | bfq_log(bfqd, "select_queue: no queue returned"); | |
5445 | ||
5446 | return bfqq; | |
5447 | } | |
5448 | ||
44e44a1b PV |
5449 | static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
5450 | { | |
5451 | struct bfq_entity *entity = &bfqq->entity; | |
5452 | ||
5453 | if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */ | |
5454 | bfq_log_bfqq(bfqd, bfqq, | |
5455 | "raising period dur %u/%u msec, old coeff %u, w %d(%d)", | |
5456 | jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish), | |
5457 | jiffies_to_msecs(bfqq->wr_cur_max_time), | |
5458 | bfqq->wr_coeff, | |
5459 | bfqq->entity.weight, bfqq->entity.orig_weight); | |
5460 | ||
5461 | if (entity->prio_changed) | |
5462 | bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change"); | |
5463 | ||
5464 | /* | |
5465 | * If too much time has elapsed from the beginning of | |
5466 | * this weight-raising period, then end weight | |
5467 | * raising. | |
5468 | */ | |
5469 | if (time_is_before_jiffies(bfqq->last_wr_start_finish + | |
5470 | bfqq->wr_cur_max_time)) { | |
5471 | bfqq->last_wr_start_finish = jiffies; | |
5472 | bfq_log_bfqq(bfqd, bfqq, | |
5473 | "wrais ending at %lu, rais_max_time %u", | |
5474 | bfqq->last_wr_start_finish, | |
5475 | jiffies_to_msecs(bfqq->wr_cur_max_time)); | |
5476 | bfq_bfqq_end_wr(bfqq); | |
5477 | } | |
5478 | } | |
5479 | /* Update weight both if it must be raised and if it must be lowered */ | |
5480 | if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1)) | |
5481 | __bfq_entity_update_weight_prio( | |
5482 | bfq_entity_service_tree(entity), | |
5483 | entity); | |
5484 | } | |
5485 | ||
aee69d78 PV |
5486 | /* |
5487 | * Dispatch next request from bfqq. | |
5488 | */ | |
5489 | static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd, | |
5490 | struct bfq_queue *bfqq) | |
5491 | { | |
5492 | struct request *rq = bfqq->next_rq; | |
5493 | unsigned long service_to_charge; | |
5494 | ||
5495 | service_to_charge = bfq_serv_to_charge(rq, bfqq); | |
5496 | ||
5497 | bfq_bfqq_served(bfqq, service_to_charge); | |
5498 | ||
5499 | bfq_dispatch_remove(bfqd->queue, rq); | |
5500 | ||
44e44a1b PV |
5501 | /* |
5502 | * If weight raising has to terminate for bfqq, then next | |
5503 | * function causes an immediate update of bfqq's weight, | |
5504 | * without waiting for next activation. As a consequence, on | |
5505 | * expiration, bfqq will be timestamped as if has never been | |
5506 | * weight-raised during this service slot, even if it has | |
5507 | * received part or even most of the service as a | |
5508 | * weight-raised queue. This inflates bfqq's timestamps, which | |
5509 | * is beneficial, as bfqq is then more willing to leave the | |
5510 | * device immediately to possible other weight-raised queues. | |
5511 | */ | |
5512 | bfq_update_wr_data(bfqd, bfqq); | |
5513 | ||
aee69d78 PV |
5514 | if (!bfqd->in_service_bic) { |
5515 | atomic_long_inc(&RQ_BIC(rq)->icq.ioc->refcount); | |
5516 | bfqd->in_service_bic = RQ_BIC(rq); | |
5517 | } | |
5518 | ||
5519 | /* | |
5520 | * Expire bfqq, pretending that its budget expired, if bfqq | |
5521 | * belongs to CLASS_IDLE and other queues are waiting for | |
5522 | * service. | |
5523 | */ | |
5524 | if (bfqd->busy_queues > 1 && bfq_class_idle(bfqq)) | |
5525 | goto expire; | |
5526 | ||
5527 | return rq; | |
5528 | ||
5529 | expire: | |
5530 | bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED); | |
5531 | return rq; | |
5532 | } | |
5533 | ||
5534 | static bool bfq_has_work(struct blk_mq_hw_ctx *hctx) | |
5535 | { | |
5536 | struct bfq_data *bfqd = hctx->queue->elevator->elevator_data; | |
5537 | ||
5538 | /* | |
5539 | * Avoiding lock: a race on bfqd->busy_queues should cause at | |
5540 | * most a call to dispatch for nothing | |
5541 | */ | |
5542 | return !list_empty_careful(&bfqd->dispatch) || | |
5543 | bfqd->busy_queues > 0; | |
5544 | } | |
5545 | ||
5546 | static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx) | |
5547 | { | |
5548 | struct bfq_data *bfqd = hctx->queue->elevator->elevator_data; | |
5549 | struct request *rq = NULL; | |
5550 | struct bfq_queue *bfqq = NULL; | |
5551 | ||
5552 | if (!list_empty(&bfqd->dispatch)) { | |
5553 | rq = list_first_entry(&bfqd->dispatch, struct request, | |
5554 | queuelist); | |
5555 | list_del_init(&rq->queuelist); | |
5556 | ||
5557 | bfqq = RQ_BFQQ(rq); | |
5558 | ||
5559 | if (bfqq) { | |
5560 | /* | |
5561 | * Increment counters here, because this | |
5562 | * dispatch does not follow the standard | |
5563 | * dispatch flow (where counters are | |
5564 | * incremented) | |
5565 | */ | |
5566 | bfqq->dispatched++; | |
5567 | ||
5568 | goto inc_in_driver_start_rq; | |
5569 | } | |
5570 | ||
5571 | /* | |
5572 | * We exploit the put_rq_private hook to decrement | |
5573 | * rq_in_driver, but put_rq_private will not be | |
5574 | * invoked on this request. So, to avoid unbalance, | |
5575 | * just start this request, without incrementing | |
5576 | * rq_in_driver. As a negative consequence, | |
5577 | * rq_in_driver is deceptively lower than it should be | |
5578 | * while this request is in service. This may cause | |
5579 | * bfq_schedule_dispatch to be invoked uselessly. | |
5580 | * | |
5581 | * As for implementing an exact solution, the | |
5582 | * put_request hook, if defined, is probably invoked | |
5583 | * also on this request. So, by exploiting this hook, | |
5584 | * we could 1) increment rq_in_driver here, and 2) | |
5585 | * decrement it in put_request. Such a solution would | |
5586 | * let the value of the counter be always accurate, | |
5587 | * but it would entail using an extra interface | |
5588 | * function. This cost seems higher than the benefit, | |
5589 | * being the frequency of non-elevator-private | |
5590 | * requests very low. | |
5591 | */ | |
5592 | goto start_rq; | |
5593 | } | |
5594 | ||
5595 | bfq_log(bfqd, "dispatch requests: %d busy queues", bfqd->busy_queues); | |
5596 | ||
5597 | if (bfqd->busy_queues == 0) | |
5598 | goto exit; | |
5599 | ||
5600 | /* | |
5601 | * Force device to serve one request at a time if | |
5602 | * strict_guarantees is true. Forcing this service scheme is | |
5603 | * currently the ONLY way to guarantee that the request | |
5604 | * service order enforced by the scheduler is respected by a | |
5605 | * queueing device. Otherwise the device is free even to make | |
5606 | * some unlucky request wait for as long as the device | |
5607 | * wishes. | |
5608 | * | |
5609 | * Of course, serving one request at at time may cause loss of | |
5610 | * throughput. | |
5611 | */ | |
5612 | if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0) | |
5613 | goto exit; | |
5614 | ||
5615 | bfqq = bfq_select_queue(bfqd); | |
5616 | if (!bfqq) | |
5617 | goto exit; | |
5618 | ||
5619 | rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq); | |
5620 | ||
5621 | if (rq) { | |
5622 | inc_in_driver_start_rq: | |
5623 | bfqd->rq_in_driver++; | |
5624 | start_rq: | |
5625 | rq->rq_flags |= RQF_STARTED; | |
5626 | } | |
5627 | exit: | |
5628 | return rq; | |
5629 | } | |
5630 | ||
5631 | static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx) | |
5632 | { | |
5633 | struct bfq_data *bfqd = hctx->queue->elevator->elevator_data; | |
5634 | struct request *rq; | |
5635 | ||
5636 | spin_lock_irq(&bfqd->lock); | |
5637 | rq = __bfq_dispatch_request(hctx); | |
5638 | spin_unlock_irq(&bfqd->lock); | |
5639 | ||
5640 | return rq; | |
5641 | } | |
5642 | ||
5643 | /* | |
5644 | * Task holds one reference to the queue, dropped when task exits. Each rq | |
5645 | * in-flight on this queue also holds a reference, dropped when rq is freed. | |
5646 | * | |
5647 | * Scheduler lock must be held here. Recall not to use bfqq after calling | |
5648 | * this function on it. | |
5649 | */ | |
5650 | static void bfq_put_queue(struct bfq_queue *bfqq) | |
5651 | { | |
e21b7a0b AA |
5652 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
5653 | struct bfq_group *bfqg = bfqq_group(bfqq); | |
5654 | #endif | |
5655 | ||
aee69d78 PV |
5656 | if (bfqq->bfqd) |
5657 | bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d", | |
5658 | bfqq, bfqq->ref); | |
5659 | ||
5660 | bfqq->ref--; | |
5661 | if (bfqq->ref) | |
5662 | return; | |
5663 | ||
e21b7a0b AA |
5664 | bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p freed", bfqq); |
5665 | ||
aee69d78 | 5666 | kmem_cache_free(bfq_pool, bfqq); |
e21b7a0b AA |
5667 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
5668 | bfqg_put(bfqg); | |
5669 | #endif | |
aee69d78 PV |
5670 | } |
5671 | ||
5672 | static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq) | |
5673 | { | |
5674 | if (bfqq == bfqd->in_service_queue) { | |
5675 | __bfq_bfqq_expire(bfqd, bfqq); | |
5676 | bfq_schedule_dispatch(bfqd); | |
5677 | } | |
5678 | ||
5679 | bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref); | |
5680 | ||
5681 | bfq_put_queue(bfqq); /* release process reference */ | |
5682 | } | |
5683 | ||
5684 | static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync) | |
5685 | { | |
5686 | struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync); | |
5687 | struct bfq_data *bfqd; | |
5688 | ||
5689 | if (bfqq) | |
5690 | bfqd = bfqq->bfqd; /* NULL if scheduler already exited */ | |
5691 | ||
5692 | if (bfqq && bfqd) { | |
5693 | unsigned long flags; | |
5694 | ||
5695 | spin_lock_irqsave(&bfqd->lock, flags); | |
5696 | bfq_exit_bfqq(bfqd, bfqq); | |
5697 | bic_set_bfqq(bic, NULL, is_sync); | |
5698 | spin_unlock_irq(&bfqd->lock); | |
5699 | } | |
5700 | } | |
5701 | ||
5702 | static void bfq_exit_icq(struct io_cq *icq) | |
5703 | { | |
5704 | struct bfq_io_cq *bic = icq_to_bic(icq); | |
5705 | ||
5706 | bfq_exit_icq_bfqq(bic, true); | |
5707 | bfq_exit_icq_bfqq(bic, false); | |
5708 | } | |
5709 | ||
5710 | /* | |
5711 | * Update the entity prio values; note that the new values will not | |
5712 | * be used until the next (re)activation. | |
5713 | */ | |
5714 | static void | |
5715 | bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic) | |
5716 | { | |
5717 | struct task_struct *tsk = current; | |
5718 | int ioprio_class; | |
5719 | struct bfq_data *bfqd = bfqq->bfqd; | |
5720 | ||
5721 | if (!bfqd) | |
5722 | return; | |
5723 | ||
5724 | ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio); | |
5725 | switch (ioprio_class) { | |
5726 | default: | |
5727 | dev_err(bfqq->bfqd->queue->backing_dev_info->dev, | |
5728 | "bfq: bad prio class %d\n", ioprio_class); | |
5729 | case IOPRIO_CLASS_NONE: | |
5730 | /* | |
5731 | * No prio set, inherit CPU scheduling settings. | |
5732 | */ | |
5733 | bfqq->new_ioprio = task_nice_ioprio(tsk); | |
5734 | bfqq->new_ioprio_class = task_nice_ioclass(tsk); | |
5735 | break; | |
5736 | case IOPRIO_CLASS_RT: | |
5737 | bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio); | |
5738 | bfqq->new_ioprio_class = IOPRIO_CLASS_RT; | |
5739 | break; | |
5740 | case IOPRIO_CLASS_BE: | |
5741 | bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio); | |
5742 | bfqq->new_ioprio_class = IOPRIO_CLASS_BE; | |
5743 | break; | |
5744 | case IOPRIO_CLASS_IDLE: | |
5745 | bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE; | |
5746 | bfqq->new_ioprio = 7; | |
5747 | bfq_clear_bfqq_idle_window(bfqq); | |
5748 | break; | |
5749 | } | |
5750 | ||
5751 | if (bfqq->new_ioprio >= IOPRIO_BE_NR) { | |
5752 | pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n", | |
5753 | bfqq->new_ioprio); | |
5754 | bfqq->new_ioprio = IOPRIO_BE_NR; | |
5755 | } | |
5756 | ||
5757 | bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio); | |
5758 | bfqq->entity.prio_changed = 1; | |
5759 | } | |
5760 | ||
5761 | static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio) | |
5762 | { | |
5763 | struct bfq_data *bfqd = bic_to_bfqd(bic); | |
5764 | struct bfq_queue *bfqq; | |
5765 | int ioprio = bic->icq.ioc->ioprio; | |
5766 | ||
5767 | /* | |
5768 | * This condition may trigger on a newly created bic, be sure to | |
5769 | * drop the lock before returning. | |
5770 | */ | |
5771 | if (unlikely(!bfqd) || likely(bic->ioprio == ioprio)) | |
5772 | return; | |
5773 | ||
5774 | bic->ioprio = ioprio; | |
5775 | ||
5776 | bfqq = bic_to_bfqq(bic, false); | |
5777 | if (bfqq) { | |
5778 | /* release process reference on this queue */ | |
5779 | bfq_put_queue(bfqq); | |
5780 | bfqq = bfq_get_queue(bfqd, bio, BLK_RW_ASYNC, bic); | |
5781 | bic_set_bfqq(bic, bfqq, false); | |
5782 | } | |
5783 | ||
5784 | bfqq = bic_to_bfqq(bic, true); | |
5785 | if (bfqq) | |
5786 | bfq_set_next_ioprio_data(bfqq, bic); | |
5787 | } | |
5788 | ||
5789 | static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
5790 | struct bfq_io_cq *bic, pid_t pid, int is_sync) | |
5791 | { | |
5792 | RB_CLEAR_NODE(&bfqq->entity.rb_node); | |
5793 | INIT_LIST_HEAD(&bfqq->fifo); | |
5794 | ||
5795 | bfqq->ref = 0; | |
5796 | bfqq->bfqd = bfqd; | |
5797 | ||
5798 | if (bic) | |
5799 | bfq_set_next_ioprio_data(bfqq, bic); | |
5800 | ||
5801 | if (is_sync) { | |
5802 | if (!bfq_class_idle(bfqq)) | |
5803 | bfq_mark_bfqq_idle_window(bfqq); | |
5804 | bfq_mark_bfqq_sync(bfqq); | |
5805 | } else | |
5806 | bfq_clear_bfqq_sync(bfqq); | |
5807 | ||
5808 | /* set end request to minus infinity from now */ | |
5809 | bfqq->ttime.last_end_request = ktime_get_ns() + 1; | |
5810 | ||
5811 | bfq_mark_bfqq_IO_bound(bfqq); | |
5812 | ||
5813 | bfqq->pid = pid; | |
5814 | ||
5815 | /* Tentative initial value to trade off between thr and lat */ | |
54b60456 | 5816 | bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3; |
aee69d78 | 5817 | bfqq->budget_timeout = bfq_smallest_from_now(); |
aee69d78 | 5818 | |
44e44a1b PV |
5819 | bfqq->wr_coeff = 1; |
5820 | bfqq->last_wr_start_finish = bfq_smallest_from_now(); | |
5821 | ||
aee69d78 PV |
5822 | /* first request is almost certainly seeky */ |
5823 | bfqq->seek_history = 1; | |
5824 | } | |
5825 | ||
5826 | static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd, | |
e21b7a0b | 5827 | struct bfq_group *bfqg, |
aee69d78 PV |
5828 | int ioprio_class, int ioprio) |
5829 | { | |
5830 | switch (ioprio_class) { | |
5831 | case IOPRIO_CLASS_RT: | |
e21b7a0b | 5832 | return &bfqg->async_bfqq[0][ioprio]; |
aee69d78 PV |
5833 | case IOPRIO_CLASS_NONE: |
5834 | ioprio = IOPRIO_NORM; | |
5835 | /* fall through */ | |
5836 | case IOPRIO_CLASS_BE: | |
e21b7a0b | 5837 | return &bfqg->async_bfqq[1][ioprio]; |
aee69d78 | 5838 | case IOPRIO_CLASS_IDLE: |
e21b7a0b | 5839 | return &bfqg->async_idle_bfqq; |
aee69d78 PV |
5840 | default: |
5841 | return NULL; | |
5842 | } | |
5843 | } | |
5844 | ||
5845 | static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd, | |
5846 | struct bio *bio, bool is_sync, | |
5847 | struct bfq_io_cq *bic) | |
5848 | { | |
5849 | const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio); | |
5850 | const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio); | |
5851 | struct bfq_queue **async_bfqq = NULL; | |
5852 | struct bfq_queue *bfqq; | |
e21b7a0b | 5853 | struct bfq_group *bfqg; |
aee69d78 PV |
5854 | |
5855 | rcu_read_lock(); | |
5856 | ||
e21b7a0b AA |
5857 | bfqg = bfq_find_set_group(bfqd, bio_blkcg(bio)); |
5858 | if (!bfqg) { | |
5859 | bfqq = &bfqd->oom_bfqq; | |
5860 | goto out; | |
5861 | } | |
5862 | ||
aee69d78 | 5863 | if (!is_sync) { |
e21b7a0b | 5864 | async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class, |
aee69d78 PV |
5865 | ioprio); |
5866 | bfqq = *async_bfqq; | |
5867 | if (bfqq) | |
5868 | goto out; | |
5869 | } | |
5870 | ||
5871 | bfqq = kmem_cache_alloc_node(bfq_pool, | |
5872 | GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN, | |
5873 | bfqd->queue->node); | |
5874 | ||
5875 | if (bfqq) { | |
5876 | bfq_init_bfqq(bfqd, bfqq, bic, current->pid, | |
5877 | is_sync); | |
e21b7a0b | 5878 | bfq_init_entity(&bfqq->entity, bfqg); |
aee69d78 PV |
5879 | bfq_log_bfqq(bfqd, bfqq, "allocated"); |
5880 | } else { | |
5881 | bfqq = &bfqd->oom_bfqq; | |
5882 | bfq_log_bfqq(bfqd, bfqq, "using oom bfqq"); | |
5883 | goto out; | |
5884 | } | |
5885 | ||
5886 | /* | |
5887 | * Pin the queue now that it's allocated, scheduler exit will | |
5888 | * prune it. | |
5889 | */ | |
5890 | if (async_bfqq) { | |
e21b7a0b AA |
5891 | bfqq->ref++; /* |
5892 | * Extra group reference, w.r.t. sync | |
5893 | * queue. This extra reference is removed | |
5894 | * only if bfqq->bfqg disappears, to | |
5895 | * guarantee that this queue is not freed | |
5896 | * until its group goes away. | |
5897 | */ | |
5898 | bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d", | |
aee69d78 PV |
5899 | bfqq, bfqq->ref); |
5900 | *async_bfqq = bfqq; | |
5901 | } | |
5902 | ||
5903 | out: | |
5904 | bfqq->ref++; /* get a process reference to this queue */ | |
5905 | bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, bfqq->ref); | |
5906 | rcu_read_unlock(); | |
5907 | return bfqq; | |
5908 | } | |
5909 | ||
5910 | static void bfq_update_io_thinktime(struct bfq_data *bfqd, | |
5911 | struct bfq_queue *bfqq) | |
5912 | { | |
5913 | struct bfq_ttime *ttime = &bfqq->ttime; | |
5914 | u64 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request; | |
5915 | ||
5916 | elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle); | |
5917 | ||
5918 | ttime->ttime_samples = (7*bfqq->ttime.ttime_samples + 256) / 8; | |
5919 | ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8); | |
5920 | ttime->ttime_mean = div64_ul(ttime->ttime_total + 128, | |
5921 | ttime->ttime_samples); | |
5922 | } | |
5923 | ||
5924 | static void | |
5925 | bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
5926 | struct request *rq) | |
5927 | { | |
aee69d78 | 5928 | bfqq->seek_history <<= 1; |
ab0e43e9 PV |
5929 | bfqq->seek_history |= |
5930 | get_sdist(bfqq->last_request_pos, rq) > BFQQ_SEEK_THR && | |
aee69d78 PV |
5931 | (!blk_queue_nonrot(bfqd->queue) || |
5932 | blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT); | |
5933 | } | |
5934 | ||
5935 | /* | |
5936 | * Disable idle window if the process thinks too long or seeks so much that | |
5937 | * it doesn't matter. | |
5938 | */ | |
5939 | static void bfq_update_idle_window(struct bfq_data *bfqd, | |
5940 | struct bfq_queue *bfqq, | |
5941 | struct bfq_io_cq *bic) | |
5942 | { | |
5943 | int enable_idle; | |
5944 | ||
5945 | /* Don't idle for async or idle io prio class. */ | |
5946 | if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq)) | |
5947 | return; | |
5948 | ||
5949 | enable_idle = bfq_bfqq_idle_window(bfqq); | |
5950 | ||
5951 | if (atomic_read(&bic->icq.ioc->active_ref) == 0 || | |
5952 | bfqd->bfq_slice_idle == 0 || | |
5953 | (bfqd->hw_tag && BFQQ_SEEKY(bfqq))) | |
5954 | enable_idle = 0; | |
5955 | else if (bfq_sample_valid(bfqq->ttime.ttime_samples)) { | |
44e44a1b PV |
5956 | if (bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle && |
5957 | bfqq->wr_coeff == 1) | |
aee69d78 PV |
5958 | enable_idle = 0; |
5959 | else | |
5960 | enable_idle = 1; | |
5961 | } | |
5962 | bfq_log_bfqq(bfqd, bfqq, "update_idle_window: enable_idle %d", | |
5963 | enable_idle); | |
5964 | ||
5965 | if (enable_idle) | |
5966 | bfq_mark_bfqq_idle_window(bfqq); | |
5967 | else | |
5968 | bfq_clear_bfqq_idle_window(bfqq); | |
5969 | } | |
5970 | ||
5971 | /* | |
5972 | * Called when a new fs request (rq) is added to bfqq. Check if there's | |
5973 | * something we should do about it. | |
5974 | */ | |
5975 | static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
5976 | struct request *rq) | |
5977 | { | |
5978 | struct bfq_io_cq *bic = RQ_BIC(rq); | |
5979 | ||
5980 | if (rq->cmd_flags & REQ_META) | |
5981 | bfqq->meta_pending++; | |
5982 | ||
5983 | bfq_update_io_thinktime(bfqd, bfqq); | |
5984 | bfq_update_io_seektime(bfqd, bfqq, rq); | |
5985 | if (bfqq->entity.service > bfq_max_budget(bfqd) / 8 || | |
5986 | !BFQQ_SEEKY(bfqq)) | |
5987 | bfq_update_idle_window(bfqd, bfqq, bic); | |
5988 | ||
5989 | bfq_log_bfqq(bfqd, bfqq, | |
5990 | "rq_enqueued: idle_window=%d (seeky %d)", | |
5991 | bfq_bfqq_idle_window(bfqq), BFQQ_SEEKY(bfqq)); | |
5992 | ||
5993 | bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); | |
5994 | ||
5995 | if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) { | |
5996 | bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 && | |
5997 | blk_rq_sectors(rq) < 32; | |
5998 | bool budget_timeout = bfq_bfqq_budget_timeout(bfqq); | |
5999 | ||
6000 | /* | |
6001 | * There is just this request queued: if the request | |
6002 | * is small and the queue is not to be expired, then | |
6003 | * just exit. | |
6004 | * | |
6005 | * In this way, if the device is being idled to wait | |
6006 | * for a new request from the in-service queue, we | |
6007 | * avoid unplugging the device and committing the | |
6008 | * device to serve just a small request. On the | |
6009 | * contrary, we wait for the block layer to decide | |
6010 | * when to unplug the device: hopefully, new requests | |
6011 | * will be merged to this one quickly, then the device | |
6012 | * will be unplugged and larger requests will be | |
6013 | * dispatched. | |
6014 | */ | |
6015 | if (small_req && !budget_timeout) | |
6016 | return; | |
6017 | ||
6018 | /* | |
6019 | * A large enough request arrived, or the queue is to | |
6020 | * be expired: in both cases disk idling is to be | |
6021 | * stopped, so clear wait_request flag and reset | |
6022 | * timer. | |
6023 | */ | |
6024 | bfq_clear_bfqq_wait_request(bfqq); | |
6025 | hrtimer_try_to_cancel(&bfqd->idle_slice_timer); | |
e21b7a0b | 6026 | bfqg_stats_update_idle_time(bfqq_group(bfqq)); |
aee69d78 PV |
6027 | |
6028 | /* | |
6029 | * The queue is not empty, because a new request just | |
6030 | * arrived. Hence we can safely expire the queue, in | |
6031 | * case of budget timeout, without risking that the | |
6032 | * timestamps of the queue are not updated correctly. | |
6033 | * See [1] for more details. | |
6034 | */ | |
6035 | if (budget_timeout) | |
6036 | bfq_bfqq_expire(bfqd, bfqq, false, | |
6037 | BFQQE_BUDGET_TIMEOUT); | |
6038 | } | |
6039 | } | |
6040 | ||
6041 | static void __bfq_insert_request(struct bfq_data *bfqd, struct request *rq) | |
6042 | { | |
6043 | struct bfq_queue *bfqq = RQ_BFQQ(rq); | |
6044 | ||
6045 | bfq_add_request(rq); | |
6046 | ||
6047 | rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)]; | |
6048 | list_add_tail(&rq->queuelist, &bfqq->fifo); | |
6049 | ||
6050 | bfq_rq_enqueued(bfqd, bfqq, rq); | |
6051 | } | |
6052 | ||
6053 | static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, | |
6054 | bool at_head) | |
6055 | { | |
6056 | struct request_queue *q = hctx->queue; | |
6057 | struct bfq_data *bfqd = q->elevator->elevator_data; | |
6058 | ||
6059 | spin_lock_irq(&bfqd->lock); | |
6060 | if (blk_mq_sched_try_insert_merge(q, rq)) { | |
6061 | spin_unlock_irq(&bfqd->lock); | |
6062 | return; | |
6063 | } | |
6064 | ||
6065 | spin_unlock_irq(&bfqd->lock); | |
6066 | ||
6067 | blk_mq_sched_request_inserted(rq); | |
6068 | ||
6069 | spin_lock_irq(&bfqd->lock); | |
6070 | if (at_head || blk_rq_is_passthrough(rq)) { | |
6071 | if (at_head) | |
6072 | list_add(&rq->queuelist, &bfqd->dispatch); | |
6073 | else | |
6074 | list_add_tail(&rq->queuelist, &bfqd->dispatch); | |
6075 | } else { | |
6076 | __bfq_insert_request(bfqd, rq); | |
6077 | ||
6078 | if (rq_mergeable(rq)) { | |
6079 | elv_rqhash_add(q, rq); | |
6080 | if (!q->last_merge) | |
6081 | q->last_merge = rq; | |
6082 | } | |
6083 | } | |
6084 | ||
6085 | spin_unlock_irq(&bfqd->lock); | |
6086 | } | |
6087 | ||
6088 | static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx, | |
6089 | struct list_head *list, bool at_head) | |
6090 | { | |
6091 | while (!list_empty(list)) { | |
6092 | struct request *rq; | |
6093 | ||
6094 | rq = list_first_entry(list, struct request, queuelist); | |
6095 | list_del_init(&rq->queuelist); | |
6096 | bfq_insert_request(hctx, rq, at_head); | |
6097 | } | |
6098 | } | |
6099 | ||
6100 | static void bfq_update_hw_tag(struct bfq_data *bfqd) | |
6101 | { | |
6102 | bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver, | |
6103 | bfqd->rq_in_driver); | |
6104 | ||
6105 | if (bfqd->hw_tag == 1) | |
6106 | return; | |
6107 | ||
6108 | /* | |
6109 | * This sample is valid if the number of outstanding requests | |
6110 | * is large enough to allow a queueing behavior. Note that the | |
6111 | * sum is not exact, as it's not taking into account deactivated | |
6112 | * requests. | |
6113 | */ | |
6114 | if (bfqd->rq_in_driver + bfqd->queued < BFQ_HW_QUEUE_THRESHOLD) | |
6115 | return; | |
6116 | ||
6117 | if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES) | |
6118 | return; | |
6119 | ||
6120 | bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD; | |
6121 | bfqd->max_rq_in_driver = 0; | |
6122 | bfqd->hw_tag_samples = 0; | |
6123 | } | |
6124 | ||
6125 | static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd) | |
6126 | { | |
ab0e43e9 PV |
6127 | u64 now_ns; |
6128 | u32 delta_us; | |
6129 | ||
aee69d78 PV |
6130 | bfq_update_hw_tag(bfqd); |
6131 | ||
6132 | bfqd->rq_in_driver--; | |
6133 | bfqq->dispatched--; | |
6134 | ||
44e44a1b PV |
6135 | if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) { |
6136 | /* | |
6137 | * Set budget_timeout (which we overload to store the | |
6138 | * time at which the queue remains with no backlog and | |
6139 | * no outstanding request; used by the weight-raising | |
6140 | * mechanism). | |
6141 | */ | |
6142 | bfqq->budget_timeout = jiffies; | |
6143 | } | |
6144 | ||
ab0e43e9 PV |
6145 | now_ns = ktime_get_ns(); |
6146 | ||
6147 | bfqq->ttime.last_end_request = now_ns; | |
6148 | ||
6149 | /* | |
6150 | * Using us instead of ns, to get a reasonable precision in | |
6151 | * computing rate in next check. | |
6152 | */ | |
6153 | delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC); | |
6154 | ||
6155 | /* | |
6156 | * If the request took rather long to complete, and, according | |
6157 | * to the maximum request size recorded, this completion latency | |
6158 | * implies that the request was certainly served at a very low | |
6159 | * rate (less than 1M sectors/sec), then the whole observation | |
6160 | * interval that lasts up to this time instant cannot be a | |
6161 | * valid time interval for computing a new peak rate. Invoke | |
6162 | * bfq_update_rate_reset to have the following three steps | |
6163 | * taken: | |
6164 | * - close the observation interval at the last (previous) | |
6165 | * request dispatch or completion | |
6166 | * - compute rate, if possible, for that observation interval | |
6167 | * - reset to zero samples, which will trigger a proper | |
6168 | * re-initialization of the observation interval on next | |
6169 | * dispatch | |
6170 | */ | |
6171 | if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC && | |
6172 | (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us < | |
6173 | 1UL<<(BFQ_RATE_SHIFT - 10)) | |
6174 | bfq_update_rate_reset(bfqd, NULL); | |
6175 | bfqd->last_completion = now_ns; | |
aee69d78 PV |
6176 | |
6177 | /* | |
6178 | * If this is the in-service queue, check if it needs to be expired, | |
6179 | * or if we want to idle in case it has no pending requests. | |
6180 | */ | |
6181 | if (bfqd->in_service_queue == bfqq) { | |
44e44a1b | 6182 | if (bfqq->dispatched == 0 && bfq_bfqq_must_idle(bfqq)) { |
aee69d78 PV |
6183 | bfq_arm_slice_timer(bfqd); |
6184 | return; | |
6185 | } else if (bfq_may_expire_for_budg_timeout(bfqq)) | |
6186 | bfq_bfqq_expire(bfqd, bfqq, false, | |
6187 | BFQQE_BUDGET_TIMEOUT); | |
6188 | else if (RB_EMPTY_ROOT(&bfqq->sort_list) && | |
6189 | (bfqq->dispatched == 0 || | |
6190 | !bfq_bfqq_may_idle(bfqq))) | |
6191 | bfq_bfqq_expire(bfqd, bfqq, false, | |
6192 | BFQQE_NO_MORE_REQUESTS); | |
6193 | } | |
6194 | } | |
6195 | ||
6196 | static void bfq_put_rq_priv_body(struct bfq_queue *bfqq) | |
6197 | { | |
6198 | bfqq->allocated--; | |
6199 | ||
6200 | bfq_put_queue(bfqq); | |
6201 | } | |
6202 | ||
6203 | static void bfq_put_rq_private(struct request_queue *q, struct request *rq) | |
6204 | { | |
6205 | struct bfq_queue *bfqq = RQ_BFQQ(rq); | |
6206 | struct bfq_data *bfqd = bfqq->bfqd; | |
6207 | ||
e21b7a0b AA |
6208 | if (rq->rq_flags & RQF_STARTED) |
6209 | bfqg_stats_update_completion(bfqq_group(bfqq), | |
6210 | rq_start_time_ns(rq), | |
6211 | rq_io_start_time_ns(rq), | |
6212 | rq->cmd_flags); | |
aee69d78 PV |
6213 | |
6214 | if (likely(rq->rq_flags & RQF_STARTED)) { | |
6215 | unsigned long flags; | |
6216 | ||
6217 | spin_lock_irqsave(&bfqd->lock, flags); | |
6218 | ||
6219 | bfq_completed_request(bfqq, bfqd); | |
6220 | bfq_put_rq_priv_body(bfqq); | |
6221 | ||
6222 | spin_unlock_irqrestore(&bfqd->lock, flags); | |
6223 | } else { | |
6224 | /* | |
6225 | * Request rq may be still/already in the scheduler, | |
6226 | * in which case we need to remove it. And we cannot | |
6227 | * defer such a check and removal, to avoid | |
6228 | * inconsistencies in the time interval from the end | |
6229 | * of this function to the start of the deferred work. | |
6230 | * This situation seems to occur only in process | |
6231 | * context, as a consequence of a merge. In the | |
6232 | * current version of the code, this implies that the | |
6233 | * lock is held. | |
6234 | */ | |
6235 | ||
6236 | if (!RB_EMPTY_NODE(&rq->rb_node)) | |
6237 | bfq_remove_request(q, rq); | |
6238 | bfq_put_rq_priv_body(bfqq); | |
6239 | } | |
6240 | ||
6241 | rq->elv.priv[0] = NULL; | |
6242 | rq->elv.priv[1] = NULL; | |
6243 | } | |
6244 | ||
6245 | /* | |
6246 | * Allocate bfq data structures associated with this request. | |
6247 | */ | |
6248 | static int bfq_get_rq_private(struct request_queue *q, struct request *rq, | |
6249 | struct bio *bio) | |
6250 | { | |
6251 | struct bfq_data *bfqd = q->elevator->elevator_data; | |
6252 | struct bfq_io_cq *bic = icq_to_bic(rq->elv.icq); | |
6253 | const int is_sync = rq_is_sync(rq); | |
6254 | struct bfq_queue *bfqq; | |
6255 | ||
6256 | spin_lock_irq(&bfqd->lock); | |
6257 | ||
6258 | bfq_check_ioprio_change(bic, bio); | |
6259 | ||
6260 | if (!bic) | |
6261 | goto queue_fail; | |
6262 | ||
e21b7a0b AA |
6263 | bfq_bic_update_cgroup(bic, bio); |
6264 | ||
aee69d78 PV |
6265 | bfqq = bic_to_bfqq(bic, is_sync); |
6266 | if (!bfqq || bfqq == &bfqd->oom_bfqq) { | |
6267 | if (bfqq) | |
6268 | bfq_put_queue(bfqq); | |
6269 | bfqq = bfq_get_queue(bfqd, bio, is_sync, bic); | |
6270 | bic_set_bfqq(bic, bfqq, is_sync); | |
6271 | } | |
6272 | ||
6273 | bfqq->allocated++; | |
6274 | bfqq->ref++; | |
6275 | bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d", | |
6276 | rq, bfqq, bfqq->ref); | |
6277 | ||
6278 | rq->elv.priv[0] = bic; | |
6279 | rq->elv.priv[1] = bfqq; | |
6280 | ||
6281 | spin_unlock_irq(&bfqd->lock); | |
6282 | ||
6283 | return 0; | |
6284 | ||
6285 | queue_fail: | |
6286 | spin_unlock_irq(&bfqd->lock); | |
6287 | ||
6288 | return 1; | |
6289 | } | |
6290 | ||
6291 | static void bfq_idle_slice_timer_body(struct bfq_queue *bfqq) | |
6292 | { | |
6293 | struct bfq_data *bfqd = bfqq->bfqd; | |
6294 | enum bfqq_expiration reason; | |
6295 | unsigned long flags; | |
6296 | ||
6297 | spin_lock_irqsave(&bfqd->lock, flags); | |
6298 | bfq_clear_bfqq_wait_request(bfqq); | |
6299 | ||
6300 | if (bfqq != bfqd->in_service_queue) { | |
6301 | spin_unlock_irqrestore(&bfqd->lock, flags); | |
6302 | return; | |
6303 | } | |
6304 | ||
6305 | if (bfq_bfqq_budget_timeout(bfqq)) | |
6306 | /* | |
6307 | * Also here the queue can be safely expired | |
6308 | * for budget timeout without wasting | |
6309 | * guarantees | |
6310 | */ | |
6311 | reason = BFQQE_BUDGET_TIMEOUT; | |
6312 | else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0) | |
6313 | /* | |
6314 | * The queue may not be empty upon timer expiration, | |
6315 | * because we may not disable the timer when the | |
6316 | * first request of the in-service queue arrives | |
6317 | * during disk idling. | |
6318 | */ | |
6319 | reason = BFQQE_TOO_IDLE; | |
6320 | else | |
6321 | goto schedule_dispatch; | |
6322 | ||
6323 | bfq_bfqq_expire(bfqd, bfqq, true, reason); | |
6324 | ||
6325 | schedule_dispatch: | |
6326 | spin_unlock_irqrestore(&bfqd->lock, flags); | |
6327 | bfq_schedule_dispatch(bfqd); | |
6328 | } | |
6329 | ||
6330 | /* | |
6331 | * Handler of the expiration of the timer running if the in-service queue | |
6332 | * is idling inside its time slice. | |
6333 | */ | |
6334 | static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer) | |
6335 | { | |
6336 | struct bfq_data *bfqd = container_of(timer, struct bfq_data, | |
6337 | idle_slice_timer); | |
6338 | struct bfq_queue *bfqq = bfqd->in_service_queue; | |
6339 | ||
6340 | /* | |
6341 | * Theoretical race here: the in-service queue can be NULL or | |
6342 | * different from the queue that was idling if a new request | |
6343 | * arrives for the current queue and there is a full dispatch | |
6344 | * cycle that changes the in-service queue. This can hardly | |
6345 | * happen, but in the worst case we just expire a queue too | |
6346 | * early. | |
6347 | */ | |
6348 | if (bfqq) | |
6349 | bfq_idle_slice_timer_body(bfqq); | |
6350 | ||
6351 | return HRTIMER_NORESTART; | |
6352 | } | |
6353 | ||
6354 | static void __bfq_put_async_bfqq(struct bfq_data *bfqd, | |
6355 | struct bfq_queue **bfqq_ptr) | |
6356 | { | |
6357 | struct bfq_queue *bfqq = *bfqq_ptr; | |
6358 | ||
6359 | bfq_log(bfqd, "put_async_bfqq: %p", bfqq); | |
6360 | if (bfqq) { | |
e21b7a0b AA |
6361 | bfq_bfqq_move(bfqd, bfqq, bfqd->root_group); |
6362 | ||
aee69d78 PV |
6363 | bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d", |
6364 | bfqq, bfqq->ref); | |
6365 | bfq_put_queue(bfqq); | |
6366 | *bfqq_ptr = NULL; | |
6367 | } | |
6368 | } | |
6369 | ||
6370 | /* | |
e21b7a0b AA |
6371 | * Release all the bfqg references to its async queues. If we are |
6372 | * deallocating the group these queues may still contain requests, so | |
6373 | * we reparent them to the root cgroup (i.e., the only one that will | |
6374 | * exist for sure until all the requests on a device are gone). | |
aee69d78 | 6375 | */ |
e21b7a0b | 6376 | static void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg) |
aee69d78 PV |
6377 | { |
6378 | int i, j; | |
6379 | ||
6380 | for (i = 0; i < 2; i++) | |
6381 | for (j = 0; j < IOPRIO_BE_NR; j++) | |
e21b7a0b | 6382 | __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]); |
aee69d78 | 6383 | |
e21b7a0b | 6384 | __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq); |
aee69d78 PV |
6385 | } |
6386 | ||
6387 | static void bfq_exit_queue(struct elevator_queue *e) | |
6388 | { | |
6389 | struct bfq_data *bfqd = e->elevator_data; | |
6390 | struct bfq_queue *bfqq, *n; | |
6391 | ||
6392 | hrtimer_cancel(&bfqd->idle_slice_timer); | |
6393 | ||
6394 | spin_lock_irq(&bfqd->lock); | |
6395 | list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list) | |
e21b7a0b | 6396 | bfq_deactivate_bfqq(bfqd, bfqq, false, false); |
aee69d78 PV |
6397 | spin_unlock_irq(&bfqd->lock); |
6398 | ||
6399 | hrtimer_cancel(&bfqd->idle_slice_timer); | |
6400 | ||
e21b7a0b AA |
6401 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
6402 | blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq); | |
6403 | #else | |
6404 | spin_lock_irq(&bfqd->lock); | |
6405 | bfq_put_async_queues(bfqd, bfqd->root_group); | |
6406 | kfree(bfqd->root_group); | |
6407 | spin_unlock_irq(&bfqd->lock); | |
6408 | #endif | |
6409 | ||
aee69d78 PV |
6410 | kfree(bfqd); |
6411 | } | |
6412 | ||
e21b7a0b AA |
6413 | static void bfq_init_root_group(struct bfq_group *root_group, |
6414 | struct bfq_data *bfqd) | |
6415 | { | |
6416 | int i; | |
6417 | ||
6418 | #ifdef CONFIG_BFQ_GROUP_IOSCHED | |
6419 | root_group->entity.parent = NULL; | |
6420 | root_group->my_entity = NULL; | |
6421 | root_group->bfqd = bfqd; | |
6422 | #endif | |
6423 | for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) | |
6424 | root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT; | |
6425 | root_group->sched_data.bfq_class_idle_last_service = jiffies; | |
6426 | } | |
6427 | ||
aee69d78 PV |
6428 | static int bfq_init_queue(struct request_queue *q, struct elevator_type *e) |
6429 | { | |
6430 | struct bfq_data *bfqd; | |
6431 | struct elevator_queue *eq; | |
aee69d78 PV |
6432 | |
6433 | eq = elevator_alloc(q, e); | |
6434 | if (!eq) | |
6435 | return -ENOMEM; | |
6436 | ||
6437 | bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node); | |
6438 | if (!bfqd) { | |
6439 | kobject_put(&eq->kobj); | |
6440 | return -ENOMEM; | |
6441 | } | |
6442 | eq->elevator_data = bfqd; | |
6443 | ||
e21b7a0b AA |
6444 | spin_lock_irq(q->queue_lock); |
6445 | q->elevator = eq; | |
6446 | spin_unlock_irq(q->queue_lock); | |
6447 | ||
aee69d78 PV |
6448 | /* |
6449 | * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues. | |
6450 | * Grab a permanent reference to it, so that the normal code flow | |
6451 | * will not attempt to free it. | |
6452 | */ | |
6453 | bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0); | |
6454 | bfqd->oom_bfqq.ref++; | |
6455 | bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO; | |
6456 | bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE; | |
6457 | bfqd->oom_bfqq.entity.new_weight = | |
6458 | bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio); | |
6459 | /* | |
6460 | * Trigger weight initialization, according to ioprio, at the | |
6461 | * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio | |
6462 | * class won't be changed any more. | |
6463 | */ | |
6464 | bfqd->oom_bfqq.entity.prio_changed = 1; | |
6465 | ||
6466 | bfqd->queue = q; | |
6467 | ||
e21b7a0b | 6468 | INIT_LIST_HEAD(&bfqd->dispatch); |
aee69d78 PV |
6469 | |
6470 | hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC, | |
6471 | HRTIMER_MODE_REL); | |
6472 | bfqd->idle_slice_timer.function = bfq_idle_slice_timer; | |
6473 | ||
6474 | INIT_LIST_HEAD(&bfqd->active_list); | |
6475 | INIT_LIST_HEAD(&bfqd->idle_list); | |
6476 | ||
6477 | bfqd->hw_tag = -1; | |
6478 | ||
6479 | bfqd->bfq_max_budget = bfq_default_max_budget; | |
6480 | ||
6481 | bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0]; | |
6482 | bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1]; | |
6483 | bfqd->bfq_back_max = bfq_back_max; | |
6484 | bfqd->bfq_back_penalty = bfq_back_penalty; | |
6485 | bfqd->bfq_slice_idle = bfq_slice_idle; | |
aee69d78 PV |
6486 | bfqd->bfq_timeout = bfq_timeout; |
6487 | ||
6488 | bfqd->bfq_requests_within_timer = 120; | |
6489 | ||
44e44a1b PV |
6490 | bfqd->low_latency = true; |
6491 | ||
6492 | /* | |
6493 | * Trade-off between responsiveness and fairness. | |
6494 | */ | |
6495 | bfqd->bfq_wr_coeff = 30; | |
6496 | bfqd->bfq_wr_max_time = 0; | |
6497 | bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000); | |
6498 | bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500); | |
6499 | ||
6500 | /* | |
6501 | * Begin by assuming, optimistically, that the device is a | |
6502 | * high-speed one, and that its peak rate is equal to 2/3 of | |
6503 | * the highest reference rate. | |
6504 | */ | |
6505 | bfqd->RT_prod = R_fast[blk_queue_nonrot(bfqd->queue)] * | |
6506 | T_fast[blk_queue_nonrot(bfqd->queue)]; | |
6507 | bfqd->peak_rate = R_fast[blk_queue_nonrot(bfqd->queue)] * 2 / 3; | |
6508 | bfqd->device_speed = BFQ_BFQD_FAST; | |
6509 | ||
aee69d78 | 6510 | spin_lock_init(&bfqd->lock); |
aee69d78 | 6511 | |
e21b7a0b AA |
6512 | /* |
6513 | * The invocation of the next bfq_create_group_hierarchy | |
6514 | * function is the head of a chain of function calls | |
6515 | * (bfq_create_group_hierarchy->blkcg_activate_policy-> | |
6516 | * blk_mq_freeze_queue) that may lead to the invocation of the | |
6517 | * has_work hook function. For this reason, | |
6518 | * bfq_create_group_hierarchy is invoked only after all | |
6519 | * scheduler data has been initialized, apart from the fields | |
6520 | * that can be initialized only after invoking | |
6521 | * bfq_create_group_hierarchy. This, in particular, enables | |
6522 | * has_work to correctly return false. Of course, to avoid | |
6523 | * other inconsistencies, the blk-mq stack must then refrain | |
6524 | * from invoking further scheduler hooks before this init | |
6525 | * function is finished. | |
6526 | */ | |
6527 | bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node); | |
6528 | if (!bfqd->root_group) | |
6529 | goto out_free; | |
6530 | bfq_init_root_group(bfqd->root_group, bfqd); | |
6531 | bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group); | |
6532 | ||
aee69d78 PV |
6533 | |
6534 | return 0; | |
e21b7a0b AA |
6535 | |
6536 | out_free: | |
6537 | kfree(bfqd); | |
6538 | kobject_put(&eq->kobj); | |
6539 | return -ENOMEM; | |
aee69d78 PV |
6540 | } |
6541 | ||
6542 | static void bfq_slab_kill(void) | |
6543 | { | |
6544 | kmem_cache_destroy(bfq_pool); | |
6545 | } | |
6546 | ||
6547 | static int __init bfq_slab_setup(void) | |
6548 | { | |
6549 | bfq_pool = KMEM_CACHE(bfq_queue, 0); | |
6550 | if (!bfq_pool) | |
6551 | return -ENOMEM; | |
6552 | return 0; | |
6553 | } | |
6554 | ||
6555 | static ssize_t bfq_var_show(unsigned int var, char *page) | |
6556 | { | |
6557 | return sprintf(page, "%u\n", var); | |
6558 | } | |
6559 | ||
6560 | static ssize_t bfq_var_store(unsigned long *var, const char *page, | |
6561 | size_t count) | |
6562 | { | |
6563 | unsigned long new_val; | |
6564 | int ret = kstrtoul(page, 10, &new_val); | |
6565 | ||
6566 | if (ret == 0) | |
6567 | *var = new_val; | |
6568 | ||
6569 | return count; | |
6570 | } | |
6571 | ||
6572 | #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ | |
6573 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ | |
6574 | { \ | |
6575 | struct bfq_data *bfqd = e->elevator_data; \ | |
6576 | u64 __data = __VAR; \ | |
6577 | if (__CONV == 1) \ | |
6578 | __data = jiffies_to_msecs(__data); \ | |
6579 | else if (__CONV == 2) \ | |
6580 | __data = div_u64(__data, NSEC_PER_MSEC); \ | |
6581 | return bfq_var_show(__data, (page)); \ | |
6582 | } | |
6583 | SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2); | |
6584 | SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2); | |
6585 | SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0); | |
6586 | SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0); | |
6587 | SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2); | |
6588 | SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0); | |
6589 | SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1); | |
6590 | SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0); | |
44e44a1b | 6591 | SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0); |
aee69d78 PV |
6592 | #undef SHOW_FUNCTION |
6593 | ||
6594 | #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \ | |
6595 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ | |
6596 | { \ | |
6597 | struct bfq_data *bfqd = e->elevator_data; \ | |
6598 | u64 __data = __VAR; \ | |
6599 | __data = div_u64(__data, NSEC_PER_USEC); \ | |
6600 | return bfq_var_show(__data, (page)); \ | |
6601 | } | |
6602 | USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle); | |
6603 | #undef USEC_SHOW_FUNCTION | |
6604 | ||
6605 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ | |
6606 | static ssize_t \ | |
6607 | __FUNC(struct elevator_queue *e, const char *page, size_t count) \ | |
6608 | { \ | |
6609 | struct bfq_data *bfqd = e->elevator_data; \ | |
6610 | unsigned long uninitialized_var(__data); \ | |
6611 | int ret = bfq_var_store(&__data, (page), count); \ | |
6612 | if (__data < (MIN)) \ | |
6613 | __data = (MIN); \ | |
6614 | else if (__data > (MAX)) \ | |
6615 | __data = (MAX); \ | |
6616 | if (__CONV == 1) \ | |
6617 | *(__PTR) = msecs_to_jiffies(__data); \ | |
6618 | else if (__CONV == 2) \ | |
6619 | *(__PTR) = (u64)__data * NSEC_PER_MSEC; \ | |
6620 | else \ | |
6621 | *(__PTR) = __data; \ | |
6622 | return ret; \ | |
6623 | } | |
6624 | STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1, | |
6625 | INT_MAX, 2); | |
6626 | STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1, | |
6627 | INT_MAX, 2); | |
6628 | STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0); | |
6629 | STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1, | |
6630 | INT_MAX, 0); | |
6631 | STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2); | |
6632 | #undef STORE_FUNCTION | |
6633 | ||
6634 | #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \ | |
6635 | static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\ | |
6636 | { \ | |
6637 | struct bfq_data *bfqd = e->elevator_data; \ | |
6638 | unsigned long uninitialized_var(__data); \ | |
6639 | int ret = bfq_var_store(&__data, (page), count); \ | |
6640 | if (__data < (MIN)) \ | |
6641 | __data = (MIN); \ | |
6642 | else if (__data > (MAX)) \ | |
6643 | __data = (MAX); \ | |
6644 | *(__PTR) = (u64)__data * NSEC_PER_USEC; \ | |
6645 | return ret; \ | |
6646 | } | |
6647 | USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0, | |
6648 | UINT_MAX); | |
6649 | #undef USEC_STORE_FUNCTION | |
6650 | ||
aee69d78 PV |
6651 | static ssize_t bfq_max_budget_store(struct elevator_queue *e, |
6652 | const char *page, size_t count) | |
6653 | { | |
6654 | struct bfq_data *bfqd = e->elevator_data; | |
6655 | unsigned long uninitialized_var(__data); | |
6656 | int ret = bfq_var_store(&__data, (page), count); | |
6657 | ||
6658 | if (__data == 0) | |
ab0e43e9 | 6659 | bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd); |
aee69d78 PV |
6660 | else { |
6661 | if (__data > INT_MAX) | |
6662 | __data = INT_MAX; | |
6663 | bfqd->bfq_max_budget = __data; | |
6664 | } | |
6665 | ||
6666 | bfqd->bfq_user_max_budget = __data; | |
6667 | ||
6668 | return ret; | |
6669 | } | |
6670 | ||
6671 | /* | |
6672 | * Leaving this name to preserve name compatibility with cfq | |
6673 | * parameters, but this timeout is used for both sync and async. | |
6674 | */ | |
6675 | static ssize_t bfq_timeout_sync_store(struct elevator_queue *e, | |
6676 | const char *page, size_t count) | |
6677 | { | |
6678 | struct bfq_data *bfqd = e->elevator_data; | |
6679 | unsigned long uninitialized_var(__data); | |
6680 | int ret = bfq_var_store(&__data, (page), count); | |
6681 | ||
6682 | if (__data < 1) | |
6683 | __data = 1; | |
6684 | else if (__data > INT_MAX) | |
6685 | __data = INT_MAX; | |
6686 | ||
6687 | bfqd->bfq_timeout = msecs_to_jiffies(__data); | |
6688 | if (bfqd->bfq_user_max_budget == 0) | |
ab0e43e9 | 6689 | bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd); |
aee69d78 PV |
6690 | |
6691 | return ret; | |
6692 | } | |
6693 | ||
6694 | static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e, | |
6695 | const char *page, size_t count) | |
6696 | { | |
6697 | struct bfq_data *bfqd = e->elevator_data; | |
6698 | unsigned long uninitialized_var(__data); | |
6699 | int ret = bfq_var_store(&__data, (page), count); | |
6700 | ||
6701 | if (__data > 1) | |
6702 | __data = 1; | |
6703 | if (!bfqd->strict_guarantees && __data == 1 | |
6704 | && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC) | |
6705 | bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC; | |
6706 | ||
6707 | bfqd->strict_guarantees = __data; | |
6708 | ||
6709 | return ret; | |
6710 | } | |
6711 | ||
44e44a1b PV |
6712 | static ssize_t bfq_low_latency_store(struct elevator_queue *e, |
6713 | const char *page, size_t count) | |
6714 | { | |
6715 | struct bfq_data *bfqd = e->elevator_data; | |
6716 | unsigned long uninitialized_var(__data); | |
6717 | int ret = bfq_var_store(&__data, (page), count); | |
6718 | ||
6719 | if (__data > 1) | |
6720 | __data = 1; | |
6721 | if (__data == 0 && bfqd->low_latency != 0) | |
6722 | bfq_end_wr(bfqd); | |
6723 | bfqd->low_latency = __data; | |
6724 | ||
6725 | return ret; | |
6726 | } | |
6727 | ||
aee69d78 PV |
6728 | #define BFQ_ATTR(name) \ |
6729 | __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store) | |
6730 | ||
6731 | static struct elv_fs_entry bfq_attrs[] = { | |
6732 | BFQ_ATTR(fifo_expire_sync), | |
6733 | BFQ_ATTR(fifo_expire_async), | |
6734 | BFQ_ATTR(back_seek_max), | |
6735 | BFQ_ATTR(back_seek_penalty), | |
6736 | BFQ_ATTR(slice_idle), | |
6737 | BFQ_ATTR(slice_idle_us), | |
6738 | BFQ_ATTR(max_budget), | |
6739 | BFQ_ATTR(timeout_sync), | |
6740 | BFQ_ATTR(strict_guarantees), | |
44e44a1b | 6741 | BFQ_ATTR(low_latency), |
aee69d78 PV |
6742 | __ATTR_NULL |
6743 | }; | |
6744 | ||
6745 | static struct elevator_type iosched_bfq_mq = { | |
6746 | .ops.mq = { | |
6747 | .get_rq_priv = bfq_get_rq_private, | |
6748 | .put_rq_priv = bfq_put_rq_private, | |
6749 | .exit_icq = bfq_exit_icq, | |
6750 | .insert_requests = bfq_insert_requests, | |
6751 | .dispatch_request = bfq_dispatch_request, | |
6752 | .next_request = elv_rb_latter_request, | |
6753 | .former_request = elv_rb_former_request, | |
6754 | .allow_merge = bfq_allow_bio_merge, | |
6755 | .bio_merge = bfq_bio_merge, | |
6756 | .request_merge = bfq_request_merge, | |
6757 | .requests_merged = bfq_requests_merged, | |
6758 | .request_merged = bfq_request_merged, | |
6759 | .has_work = bfq_has_work, | |
6760 | .init_sched = bfq_init_queue, | |
6761 | .exit_sched = bfq_exit_queue, | |
6762 | }, | |
6763 | ||
6764 | .uses_mq = true, | |
6765 | .icq_size = sizeof(struct bfq_io_cq), | |
6766 | .icq_align = __alignof__(struct bfq_io_cq), | |
6767 | .elevator_attrs = bfq_attrs, | |
6768 | .elevator_name = "bfq", | |
6769 | .elevator_owner = THIS_MODULE, | |
6770 | }; | |
6771 | ||
e21b7a0b AA |
6772 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
6773 | static struct blkcg_policy blkcg_policy_bfq = { | |
6774 | .dfl_cftypes = bfq_blkg_files, | |
6775 | .legacy_cftypes = bfq_blkcg_legacy_files, | |
6776 | ||
6777 | .cpd_alloc_fn = bfq_cpd_alloc, | |
6778 | .cpd_init_fn = bfq_cpd_init, | |
6779 | .cpd_bind_fn = bfq_cpd_init, | |
6780 | .cpd_free_fn = bfq_cpd_free, | |
6781 | ||
6782 | .pd_alloc_fn = bfq_pd_alloc, | |
6783 | .pd_init_fn = bfq_pd_init, | |
6784 | .pd_offline_fn = bfq_pd_offline, | |
6785 | .pd_free_fn = bfq_pd_free, | |
6786 | .pd_reset_stats_fn = bfq_pd_reset_stats, | |
6787 | }; | |
6788 | #endif | |
6789 | ||
aee69d78 PV |
6790 | static int __init bfq_init(void) |
6791 | { | |
6792 | int ret; | |
6793 | ||
e21b7a0b AA |
6794 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
6795 | ret = blkcg_policy_register(&blkcg_policy_bfq); | |
6796 | if (ret) | |
6797 | return ret; | |
6798 | #endif | |
6799 | ||
aee69d78 PV |
6800 | ret = -ENOMEM; |
6801 | if (bfq_slab_setup()) | |
6802 | goto err_pol_unreg; | |
6803 | ||
44e44a1b PV |
6804 | /* |
6805 | * Times to load large popular applications for the typical | |
6806 | * systems installed on the reference devices (see the | |
6807 | * comments before the definitions of the next two | |
6808 | * arrays). Actually, we use slightly slower values, as the | |
6809 | * estimated peak rate tends to be smaller than the actual | |
6810 | * peak rate. The reason for this last fact is that estimates | |
6811 | * are computed over much shorter time intervals than the long | |
6812 | * intervals typically used for benchmarking. Why? First, to | |
6813 | * adapt more quickly to variations. Second, because an I/O | |
6814 | * scheduler cannot rely on a peak-rate-evaluation workload to | |
6815 | * be run for a long time. | |
6816 | */ | |
6817 | T_slow[0] = msecs_to_jiffies(3500); /* actually 4 sec */ | |
6818 | T_slow[1] = msecs_to_jiffies(6000); /* actually 6.5 sec */ | |
6819 | T_fast[0] = msecs_to_jiffies(7000); /* actually 8 sec */ | |
6820 | T_fast[1] = msecs_to_jiffies(2500); /* actually 3 sec */ | |
6821 | ||
6822 | /* | |
6823 | * Thresholds that determine the switch between speed classes | |
6824 | * (see the comments before the definition of the array | |
6825 | * device_speed_thresh). These thresholds are biased towards | |
6826 | * transitions to the fast class. This is safer than the | |
6827 | * opposite bias. In fact, a wrong transition to the slow | |
6828 | * class results in short weight-raising periods, because the | |
6829 | * speed of the device then tends to be higher that the | |
6830 | * reference peak rate. On the opposite end, a wrong | |
6831 | * transition to the fast class tends to increase | |
6832 | * weight-raising periods, because of the opposite reason. | |
6833 | */ | |
6834 | device_speed_thresh[0] = (4 * R_slow[0]) / 3; | |
6835 | device_speed_thresh[1] = (4 * R_slow[1]) / 3; | |
6836 | ||
aee69d78 PV |
6837 | ret = elv_register(&iosched_bfq_mq); |
6838 | if (ret) | |
6839 | goto err_pol_unreg; | |
6840 | ||
6841 | return 0; | |
6842 | ||
6843 | err_pol_unreg: | |
e21b7a0b AA |
6844 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
6845 | blkcg_policy_unregister(&blkcg_policy_bfq); | |
6846 | #endif | |
aee69d78 PV |
6847 | return ret; |
6848 | } | |
6849 | ||
6850 | static void __exit bfq_exit(void) | |
6851 | { | |
6852 | elv_unregister(&iosched_bfq_mq); | |
e21b7a0b AA |
6853 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
6854 | blkcg_policy_unregister(&blkcg_policy_bfq); | |
6855 | #endif | |
aee69d78 PV |
6856 | bfq_slab_kill(); |
6857 | } | |
6858 | ||
6859 | module_init(bfq_init); | |
6860 | module_exit(bfq_exit); | |
6861 | ||
6862 | MODULE_AUTHOR("Paolo Valente"); | |
6863 | MODULE_LICENSE("GPL"); | |
6864 | MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler"); |