Commit | Line | Data |
---|---|---|
00e04393 OS |
1 | /* |
2 | * The Kyber I/O scheduler. Controls latency by throttling queue depths using | |
3 | * scalable techniques. | |
4 | * | |
5 | * Copyright (C) 2017 Facebook | |
6 | * | |
7 | * This program is free software; you can redistribute it and/or | |
8 | * modify it under the terms of the GNU General Public | |
9 | * License v2 as published by the Free Software Foundation. | |
10 | * | |
11 | * This program is distributed in the hope that it will be useful, | |
12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
14 | * General Public License for more details. | |
15 | * | |
16 | * You should have received a copy of the GNU General Public License | |
17 | * along with this program. If not, see <https://www.gnu.org/licenses/>. | |
18 | */ | |
19 | ||
20 | #include <linux/kernel.h> | |
21 | #include <linux/blkdev.h> | |
22 | #include <linux/blk-mq.h> | |
23 | #include <linux/elevator.h> | |
24 | #include <linux/module.h> | |
25 | #include <linux/sbitmap.h> | |
26 | ||
27 | #include "blk.h" | |
28 | #include "blk-mq.h" | |
16b738f6 | 29 | #include "blk-mq-debugfs.h" |
00e04393 OS |
30 | #include "blk-mq-sched.h" |
31 | #include "blk-mq-tag.h" | |
00e04393 | 32 | |
6c3b7af1 OS |
33 | #define CREATE_TRACE_POINTS |
34 | #include <trace/events/kyber.h> | |
35 | ||
6e25cb01 OS |
36 | /* |
37 | * Scheduling domains: the device is divided into multiple domains based on the | |
38 | * request type. | |
39 | */ | |
00e04393 OS |
40 | enum { |
41 | KYBER_READ, | |
6e25cb01 OS |
42 | KYBER_WRITE, |
43 | KYBER_DISCARD, | |
44 | KYBER_OTHER, | |
00e04393 OS |
45 | KYBER_NUM_DOMAINS, |
46 | }; | |
47 | ||
6c3b7af1 OS |
48 | static const char *kyber_domain_names[] = { |
49 | [KYBER_READ] = "READ", | |
50 | [KYBER_WRITE] = "WRITE", | |
51 | [KYBER_DISCARD] = "DISCARD", | |
52 | [KYBER_OTHER] = "OTHER", | |
53 | }; | |
54 | ||
00e04393 | 55 | enum { |
00e04393 OS |
56 | /* |
57 | * In order to prevent starvation of synchronous requests by a flood of | |
58 | * asynchronous requests, we reserve 25% of requests for synchronous | |
59 | * operations. | |
60 | */ | |
61 | KYBER_ASYNC_PERCENT = 75, | |
62 | }; | |
63 | ||
64 | /* | |
6e25cb01 | 65 | * Maximum device-wide depth for each scheduling domain. |
00e04393 | 66 | * |
6e25cb01 OS |
67 | * Even for fast devices with lots of tags like NVMe, you can saturate the |
68 | * device with only a fraction of the maximum possible queue depth. So, we cap | |
69 | * these to a reasonable value. | |
00e04393 OS |
70 | */ |
71 | static const unsigned int kyber_depth[] = { | |
72 | [KYBER_READ] = 256, | |
6e25cb01 OS |
73 | [KYBER_WRITE] = 128, |
74 | [KYBER_DISCARD] = 64, | |
75 | [KYBER_OTHER] = 16, | |
00e04393 OS |
76 | }; |
77 | ||
78 | /* | |
6e25cb01 OS |
79 | * Default latency targets for each scheduling domain. |
80 | */ | |
81 | static const u64 kyber_latency_targets[] = { | |
f0a0cddd OS |
82 | [KYBER_READ] = 2ULL * NSEC_PER_MSEC, |
83 | [KYBER_WRITE] = 10ULL * NSEC_PER_MSEC, | |
84 | [KYBER_DISCARD] = 5ULL * NSEC_PER_SEC, | |
6e25cb01 OS |
85 | }; |
86 | ||
87 | /* | |
88 | * Batch size (number of requests we'll dispatch in a row) for each scheduling | |
89 | * domain. | |
00e04393 OS |
90 | */ |
91 | static const unsigned int kyber_batch_size[] = { | |
92 | [KYBER_READ] = 16, | |
6e25cb01 OS |
93 | [KYBER_WRITE] = 8, |
94 | [KYBER_DISCARD] = 1, | |
95 | [KYBER_OTHER] = 1, | |
96 | }; | |
97 | ||
98 | /* | |
99 | * Requests latencies are recorded in a histogram with buckets defined relative | |
100 | * to the target latency: | |
101 | * | |
102 | * <= 1/4 * target latency | |
103 | * <= 1/2 * target latency | |
104 | * <= 3/4 * target latency | |
105 | * <= target latency | |
106 | * <= 1 1/4 * target latency | |
107 | * <= 1 1/2 * target latency | |
108 | * <= 1 3/4 * target latency | |
109 | * > 1 3/4 * target latency | |
110 | */ | |
111 | enum { | |
112 | /* | |
113 | * The width of the latency histogram buckets is | |
114 | * 1 / (1 << KYBER_LATENCY_SHIFT) * target latency. | |
115 | */ | |
116 | KYBER_LATENCY_SHIFT = 2, | |
117 | /* | |
118 | * The first (1 << KYBER_LATENCY_SHIFT) buckets are <= target latency, | |
119 | * thus, "good". | |
120 | */ | |
121 | KYBER_GOOD_BUCKETS = 1 << KYBER_LATENCY_SHIFT, | |
122 | /* There are also (1 << KYBER_LATENCY_SHIFT) "bad" buckets. */ | |
123 | KYBER_LATENCY_BUCKETS = 2 << KYBER_LATENCY_SHIFT, | |
124 | }; | |
125 | ||
126 | /* | |
127 | * We measure both the total latency and the I/O latency (i.e., latency after | |
128 | * submitting to the device). | |
129 | */ | |
130 | enum { | |
131 | KYBER_TOTAL_LATENCY, | |
132 | KYBER_IO_LATENCY, | |
133 | }; | |
134 | ||
6c3b7af1 OS |
135 | static const char *kyber_latency_type_names[] = { |
136 | [KYBER_TOTAL_LATENCY] = "total", | |
137 | [KYBER_IO_LATENCY] = "I/O", | |
138 | }; | |
139 | ||
6e25cb01 OS |
140 | /* |
141 | * Per-cpu latency histograms: total latency and I/O latency for each scheduling | |
142 | * domain except for KYBER_OTHER. | |
143 | */ | |
144 | struct kyber_cpu_latency { | |
145 | atomic_t buckets[KYBER_OTHER][2][KYBER_LATENCY_BUCKETS]; | |
00e04393 OS |
146 | }; |
147 | ||
a6088845 JW |
148 | /* |
149 | * There is a same mapping between ctx & hctx and kcq & khd, | |
150 | * we use request->mq_ctx->index_hw to index the kcq in khd. | |
151 | */ | |
152 | struct kyber_ctx_queue { | |
153 | /* | |
154 | * Used to ensure operations on rq_list and kcq_map to be an atmoic one. | |
155 | * Also protect the rqs on rq_list when merge. | |
156 | */ | |
157 | spinlock_t lock; | |
158 | struct list_head rq_list[KYBER_NUM_DOMAINS]; | |
159 | } ____cacheline_aligned_in_smp; | |
160 | ||
00e04393 | 161 | struct kyber_queue_data { |
6c3b7af1 OS |
162 | struct request_queue *q; |
163 | ||
00e04393 | 164 | /* |
6e25cb01 OS |
165 | * Each scheduling domain has a limited number of in-flight requests |
166 | * device-wide, limited by these tokens. | |
00e04393 OS |
167 | */ |
168 | struct sbitmap_queue domain_tokens[KYBER_NUM_DOMAINS]; | |
169 | ||
170 | /* | |
171 | * Async request percentage, converted to per-word depth for | |
172 | * sbitmap_get_shallow(). | |
173 | */ | |
174 | unsigned int async_depth; | |
175 | ||
6e25cb01 OS |
176 | struct kyber_cpu_latency __percpu *cpu_latency; |
177 | ||
178 | /* Timer for stats aggregation and adjusting domain tokens. */ | |
179 | struct timer_list timer; | |
180 | ||
181 | unsigned int latency_buckets[KYBER_OTHER][2][KYBER_LATENCY_BUCKETS]; | |
182 | ||
183 | unsigned long latency_timeout[KYBER_OTHER]; | |
184 | ||
185 | int domain_p99[KYBER_OTHER]; | |
186 | ||
00e04393 | 187 | /* Target latencies in nanoseconds. */ |
6e25cb01 | 188 | u64 latency_targets[KYBER_OTHER]; |
00e04393 OS |
189 | }; |
190 | ||
191 | struct kyber_hctx_data { | |
192 | spinlock_t lock; | |
193 | struct list_head rqs[KYBER_NUM_DOMAINS]; | |
194 | unsigned int cur_domain; | |
195 | unsigned int batching; | |
a6088845 JW |
196 | struct kyber_ctx_queue *kcqs; |
197 | struct sbitmap kcq_map[KYBER_NUM_DOMAINS]; | |
ac6424b9 | 198 | wait_queue_entry_t domain_wait[KYBER_NUM_DOMAINS]; |
fcf38cdf | 199 | struct sbq_wait_state *domain_ws[KYBER_NUM_DOMAINS]; |
00e04393 OS |
200 | atomic_t wait_index[KYBER_NUM_DOMAINS]; |
201 | }; | |
202 | ||
fcf38cdf OS |
203 | static int kyber_domain_wake(wait_queue_entry_t *wait, unsigned mode, int flags, |
204 | void *key); | |
205 | ||
a6088845 | 206 | static unsigned int kyber_sched_domain(unsigned int op) |
00e04393 | 207 | { |
6e25cb01 OS |
208 | switch (op & REQ_OP_MASK) { |
209 | case REQ_OP_READ: | |
00e04393 | 210 | return KYBER_READ; |
6e25cb01 OS |
211 | case REQ_OP_WRITE: |
212 | return KYBER_WRITE; | |
213 | case REQ_OP_DISCARD: | |
214 | return KYBER_DISCARD; | |
215 | default: | |
00e04393 | 216 | return KYBER_OTHER; |
6e25cb01 | 217 | } |
00e04393 OS |
218 | } |
219 | ||
6e25cb01 OS |
220 | static void flush_latency_buckets(struct kyber_queue_data *kqd, |
221 | struct kyber_cpu_latency *cpu_latency, | |
222 | unsigned int sched_domain, unsigned int type) | |
00e04393 | 223 | { |
6e25cb01 OS |
224 | unsigned int *buckets = kqd->latency_buckets[sched_domain][type]; |
225 | atomic_t *cpu_buckets = cpu_latency->buckets[sched_domain][type]; | |
226 | unsigned int bucket; | |
00e04393 | 227 | |
6e25cb01 OS |
228 | for (bucket = 0; bucket < KYBER_LATENCY_BUCKETS; bucket++) |
229 | buckets[bucket] += atomic_xchg(&cpu_buckets[bucket], 0); | |
00e04393 OS |
230 | } |
231 | ||
232 | /* | |
6e25cb01 OS |
233 | * Calculate the histogram bucket with the given percentile rank, or -1 if there |
234 | * aren't enough samples yet. | |
00e04393 | 235 | */ |
6e25cb01 OS |
236 | static int calculate_percentile(struct kyber_queue_data *kqd, |
237 | unsigned int sched_domain, unsigned int type, | |
238 | unsigned int percentile) | |
00e04393 | 239 | { |
6e25cb01 OS |
240 | unsigned int *buckets = kqd->latency_buckets[sched_domain][type]; |
241 | unsigned int bucket, samples = 0, percentile_samples; | |
242 | ||
243 | for (bucket = 0; bucket < KYBER_LATENCY_BUCKETS; bucket++) | |
244 | samples += buckets[bucket]; | |
245 | ||
246 | if (!samples) | |
247 | return -1; | |
00e04393 OS |
248 | |
249 | /* | |
6e25cb01 OS |
250 | * We do the calculation once we have 500 samples or one second passes |
251 | * since the first sample was recorded, whichever comes first. | |
00e04393 | 252 | */ |
6e25cb01 OS |
253 | if (!kqd->latency_timeout[sched_domain]) |
254 | kqd->latency_timeout[sched_domain] = max(jiffies + HZ, 1UL); | |
255 | if (samples < 500 && | |
256 | time_is_after_jiffies(kqd->latency_timeout[sched_domain])) { | |
257 | return -1; | |
258 | } | |
259 | kqd->latency_timeout[sched_domain] = 0; | |
00e04393 | 260 | |
6e25cb01 OS |
261 | percentile_samples = DIV_ROUND_UP(samples * percentile, 100); |
262 | for (bucket = 0; bucket < KYBER_LATENCY_BUCKETS - 1; bucket++) { | |
263 | if (buckets[bucket] >= percentile_samples) | |
00e04393 | 264 | break; |
6e25cb01 | 265 | percentile_samples -= buckets[bucket]; |
00e04393 | 266 | } |
6e25cb01 | 267 | memset(buckets, 0, sizeof(kqd->latency_buckets[sched_domain][type])); |
00e04393 | 268 | |
6c3b7af1 OS |
269 | trace_kyber_latency(kqd->q, kyber_domain_names[sched_domain], |
270 | kyber_latency_type_names[type], percentile, | |
271 | bucket + 1, 1 << KYBER_LATENCY_SHIFT, samples); | |
272 | ||
6e25cb01 OS |
273 | return bucket; |
274 | } | |
275 | ||
276 | static void kyber_resize_domain(struct kyber_queue_data *kqd, | |
277 | unsigned int sched_domain, unsigned int depth) | |
278 | { | |
00e04393 | 279 | depth = clamp(depth, 1U, kyber_depth[sched_domain]); |
6c3b7af1 | 280 | if (depth != kqd->domain_tokens[sched_domain].sb.depth) { |
00e04393 | 281 | sbitmap_queue_resize(&kqd->domain_tokens[sched_domain], depth); |
6c3b7af1 OS |
282 | trace_kyber_adjust(kqd->q, kyber_domain_names[sched_domain], |
283 | depth); | |
284 | } | |
00e04393 OS |
285 | } |
286 | ||
6e25cb01 OS |
287 | static void kyber_timer_fn(struct timer_list *t) |
288 | { | |
289 | struct kyber_queue_data *kqd = from_timer(kqd, t, timer); | |
290 | unsigned int sched_domain; | |
291 | int cpu; | |
292 | bool bad = false; | |
293 | ||
294 | /* Sum all of the per-cpu latency histograms. */ | |
295 | for_each_online_cpu(cpu) { | |
296 | struct kyber_cpu_latency *cpu_latency; | |
297 | ||
298 | cpu_latency = per_cpu_ptr(kqd->cpu_latency, cpu); | |
299 | for (sched_domain = 0; sched_domain < KYBER_OTHER; sched_domain++) { | |
300 | flush_latency_buckets(kqd, cpu_latency, sched_domain, | |
301 | KYBER_TOTAL_LATENCY); | |
302 | flush_latency_buckets(kqd, cpu_latency, sched_domain, | |
303 | KYBER_IO_LATENCY); | |
00e04393 OS |
304 | } |
305 | } | |
306 | ||
6e25cb01 OS |
307 | /* |
308 | * Check if any domains have a high I/O latency, which might indicate | |
309 | * congestion in the device. Note that we use the p90; we don't want to | |
310 | * be too sensitive to outliers here. | |
311 | */ | |
312 | for (sched_domain = 0; sched_domain < KYBER_OTHER; sched_domain++) { | |
313 | int p90; | |
00e04393 | 314 | |
6e25cb01 OS |
315 | p90 = calculate_percentile(kqd, sched_domain, KYBER_IO_LATENCY, |
316 | 90); | |
317 | if (p90 >= KYBER_GOOD_BUCKETS) | |
318 | bad = true; | |
319 | } | |
00e04393 OS |
320 | |
321 | /* | |
6e25cb01 OS |
322 | * Adjust the scheduling domain depths. If we determined that there was |
323 | * congestion, we throttle all domains with good latencies. Either way, | |
324 | * we ease up on throttling domains with bad latencies. | |
00e04393 | 325 | */ |
6e25cb01 OS |
326 | for (sched_domain = 0; sched_domain < KYBER_OTHER; sched_domain++) { |
327 | unsigned int orig_depth, depth; | |
328 | int p99; | |
329 | ||
330 | p99 = calculate_percentile(kqd, sched_domain, | |
331 | KYBER_TOTAL_LATENCY, 99); | |
332 | /* | |
333 | * This is kind of subtle: different domains will not | |
334 | * necessarily have enough samples to calculate the latency | |
335 | * percentiles during the same window, so we have to remember | |
336 | * the p99 for the next time we observe congestion; once we do, | |
337 | * we don't want to throttle again until we get more data, so we | |
338 | * reset it to -1. | |
339 | */ | |
340 | if (bad) { | |
341 | if (p99 < 0) | |
342 | p99 = kqd->domain_p99[sched_domain]; | |
343 | kqd->domain_p99[sched_domain] = -1; | |
344 | } else if (p99 >= 0) { | |
345 | kqd->domain_p99[sched_domain] = p99; | |
346 | } | |
347 | if (p99 < 0) | |
348 | continue; | |
349 | ||
350 | /* | |
351 | * If this domain has bad latency, throttle less. Otherwise, | |
352 | * throttle more iff we determined that there is congestion. | |
353 | * | |
354 | * The new depth is scaled linearly with the p99 latency vs the | |
355 | * latency target. E.g., if the p99 is 3/4 of the target, then | |
356 | * we throttle down to 3/4 of the current depth, and if the p99 | |
357 | * is 2x the target, then we double the depth. | |
358 | */ | |
359 | if (bad || p99 >= KYBER_GOOD_BUCKETS) { | |
360 | orig_depth = kqd->domain_tokens[sched_domain].sb.depth; | |
361 | depth = (orig_depth * (p99 + 1)) >> KYBER_LATENCY_SHIFT; | |
362 | kyber_resize_domain(kqd, sched_domain, depth); | |
363 | } | |
364 | } | |
00e04393 OS |
365 | } |
366 | ||
6e25cb01 | 367 | static unsigned int kyber_sched_tags_shift(struct request_queue *q) |
00e04393 OS |
368 | { |
369 | /* | |
370 | * All of the hardware queues have the same depth, so we can just grab | |
371 | * the shift of the first one. | |
372 | */ | |
6e25cb01 | 373 | return q->queue_hw_ctx[0]->sched_tags->bitmap_tags.sb.shift; |
a6088845 JW |
374 | } |
375 | ||
00e04393 OS |
376 | static struct kyber_queue_data *kyber_queue_data_alloc(struct request_queue *q) |
377 | { | |
378 | struct kyber_queue_data *kqd; | |
00e04393 OS |
379 | unsigned int shift; |
380 | int ret = -ENOMEM; | |
381 | int i; | |
382 | ||
6e25cb01 | 383 | kqd = kzalloc_node(sizeof(*kqd), GFP_KERNEL, q->node); |
00e04393 OS |
384 | if (!kqd) |
385 | goto err; | |
00e04393 | 386 | |
6c3b7af1 OS |
387 | kqd->q = q; |
388 | ||
6e25cb01 OS |
389 | kqd->cpu_latency = alloc_percpu_gfp(struct kyber_cpu_latency, |
390 | GFP_KERNEL | __GFP_ZERO); | |
391 | if (!kqd->cpu_latency) | |
00e04393 OS |
392 | goto err_kqd; |
393 | ||
6e25cb01 OS |
394 | timer_setup(&kqd->timer, kyber_timer_fn, 0); |
395 | ||
00e04393 OS |
396 | for (i = 0; i < KYBER_NUM_DOMAINS; i++) { |
397 | WARN_ON(!kyber_depth[i]); | |
398 | WARN_ON(!kyber_batch_size[i]); | |
399 | ret = sbitmap_queue_init_node(&kqd->domain_tokens[i], | |
fa2a1f60 OS |
400 | kyber_depth[i], -1, false, |
401 | GFP_KERNEL, q->node); | |
00e04393 OS |
402 | if (ret) { |
403 | while (--i >= 0) | |
404 | sbitmap_queue_free(&kqd->domain_tokens[i]); | |
6e25cb01 | 405 | goto err_buckets; |
00e04393 | 406 | } |
00e04393 OS |
407 | } |
408 | ||
6e25cb01 OS |
409 | for (i = 0; i < KYBER_OTHER; i++) { |
410 | kqd->domain_p99[i] = -1; | |
411 | kqd->latency_targets[i] = kyber_latency_targets[i]; | |
412 | } | |
00e04393 | 413 | |
6e25cb01 OS |
414 | shift = kyber_sched_tags_shift(q); |
415 | kqd->async_depth = (1U << shift) * KYBER_ASYNC_PERCENT / 100U; | |
00e04393 OS |
416 | |
417 | return kqd; | |
418 | ||
6e25cb01 OS |
419 | err_buckets: |
420 | free_percpu(kqd->cpu_latency); | |
00e04393 OS |
421 | err_kqd: |
422 | kfree(kqd); | |
423 | err: | |
424 | return ERR_PTR(ret); | |
425 | } | |
426 | ||
427 | static int kyber_init_sched(struct request_queue *q, struct elevator_type *e) | |
428 | { | |
429 | struct kyber_queue_data *kqd; | |
430 | struct elevator_queue *eq; | |
431 | ||
432 | eq = elevator_alloc(q, e); | |
433 | if (!eq) | |
434 | return -ENOMEM; | |
435 | ||
436 | kqd = kyber_queue_data_alloc(q); | |
437 | if (IS_ERR(kqd)) { | |
438 | kobject_put(&eq->kobj); | |
439 | return PTR_ERR(kqd); | |
440 | } | |
441 | ||
6e25cb01 OS |
442 | blk_stat_enable_accounting(q); |
443 | ||
00e04393 OS |
444 | eq->elevator_data = kqd; |
445 | q->elevator = eq; | |
446 | ||
00e04393 OS |
447 | return 0; |
448 | } | |
449 | ||
450 | static void kyber_exit_sched(struct elevator_queue *e) | |
451 | { | |
452 | struct kyber_queue_data *kqd = e->elevator_data; | |
00e04393 OS |
453 | int i; |
454 | ||
6e25cb01 | 455 | del_timer_sync(&kqd->timer); |
00e04393 OS |
456 | |
457 | for (i = 0; i < KYBER_NUM_DOMAINS; i++) | |
458 | sbitmap_queue_free(&kqd->domain_tokens[i]); | |
6e25cb01 | 459 | free_percpu(kqd->cpu_latency); |
00e04393 OS |
460 | kfree(kqd); |
461 | } | |
462 | ||
a6088845 JW |
463 | static void kyber_ctx_queue_init(struct kyber_ctx_queue *kcq) |
464 | { | |
465 | unsigned int i; | |
466 | ||
467 | spin_lock_init(&kcq->lock); | |
468 | for (i = 0; i < KYBER_NUM_DOMAINS; i++) | |
469 | INIT_LIST_HEAD(&kcq->rq_list[i]); | |
470 | } | |
471 | ||
00e04393 OS |
472 | static int kyber_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) |
473 | { | |
28820640 | 474 | struct kyber_queue_data *kqd = hctx->queue->elevator->elevator_data; |
00e04393 OS |
475 | struct kyber_hctx_data *khd; |
476 | int i; | |
477 | ||
478 | khd = kmalloc_node(sizeof(*khd), GFP_KERNEL, hctx->numa_node); | |
479 | if (!khd) | |
480 | return -ENOMEM; | |
481 | ||
a6088845 JW |
482 | khd->kcqs = kmalloc_array_node(hctx->nr_ctx, |
483 | sizeof(struct kyber_ctx_queue), | |
484 | GFP_KERNEL, hctx->numa_node); | |
485 | if (!khd->kcqs) | |
486 | goto err_khd; | |
487 | ||
488 | for (i = 0; i < hctx->nr_ctx; i++) | |
489 | kyber_ctx_queue_init(&khd->kcqs[i]); | |
490 | ||
491 | for (i = 0; i < KYBER_NUM_DOMAINS; i++) { | |
492 | if (sbitmap_init_node(&khd->kcq_map[i], hctx->nr_ctx, | |
493 | ilog2(8), GFP_KERNEL, hctx->numa_node)) { | |
494 | while (--i >= 0) | |
495 | sbitmap_free(&khd->kcq_map[i]); | |
496 | goto err_kcqs; | |
497 | } | |
498 | } | |
499 | ||
00e04393 OS |
500 | spin_lock_init(&khd->lock); |
501 | ||
502 | for (i = 0; i < KYBER_NUM_DOMAINS; i++) { | |
503 | INIT_LIST_HEAD(&khd->rqs[i]); | |
fcf38cdf OS |
504 | init_waitqueue_func_entry(&khd->domain_wait[i], |
505 | kyber_domain_wake); | |
506 | khd->domain_wait[i].private = hctx; | |
2055da97 | 507 | INIT_LIST_HEAD(&khd->domain_wait[i].entry); |
00e04393 OS |
508 | atomic_set(&khd->wait_index[i], 0); |
509 | } | |
510 | ||
511 | khd->cur_domain = 0; | |
512 | khd->batching = 0; | |
513 | ||
514 | hctx->sched_data = khd; | |
28820640 JA |
515 | sbitmap_queue_min_shallow_depth(&hctx->sched_tags->bitmap_tags, |
516 | kqd->async_depth); | |
00e04393 OS |
517 | |
518 | return 0; | |
a6088845 JW |
519 | |
520 | err_kcqs: | |
521 | kfree(khd->kcqs); | |
522 | err_khd: | |
523 | kfree(khd); | |
524 | return -ENOMEM; | |
00e04393 OS |
525 | } |
526 | ||
527 | static void kyber_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
528 | { | |
a6088845 JW |
529 | struct kyber_hctx_data *khd = hctx->sched_data; |
530 | int i; | |
531 | ||
532 | for (i = 0; i < KYBER_NUM_DOMAINS; i++) | |
533 | sbitmap_free(&khd->kcq_map[i]); | |
534 | kfree(khd->kcqs); | |
00e04393 OS |
535 | kfree(hctx->sched_data); |
536 | } | |
537 | ||
538 | static int rq_get_domain_token(struct request *rq) | |
539 | { | |
540 | return (long)rq->elv.priv[0]; | |
541 | } | |
542 | ||
543 | static void rq_set_domain_token(struct request *rq, int token) | |
544 | { | |
545 | rq->elv.priv[0] = (void *)(long)token; | |
546 | } | |
547 | ||
548 | static void rq_clear_domain_token(struct kyber_queue_data *kqd, | |
549 | struct request *rq) | |
550 | { | |
551 | unsigned int sched_domain; | |
552 | int nr; | |
553 | ||
554 | nr = rq_get_domain_token(rq); | |
555 | if (nr != -1) { | |
a6088845 | 556 | sched_domain = kyber_sched_domain(rq->cmd_flags); |
00e04393 OS |
557 | sbitmap_queue_clear(&kqd->domain_tokens[sched_domain], nr, |
558 | rq->mq_ctx->cpu); | |
559 | } | |
560 | } | |
561 | ||
5bbf4e5a | 562 | static void kyber_limit_depth(unsigned int op, struct blk_mq_alloc_data *data) |
00e04393 | 563 | { |
00e04393 OS |
564 | /* |
565 | * We use the scheduler tags as per-hardware queue queueing tokens. | |
566 | * Async requests can be limited at this stage. | |
567 | */ | |
5bbf4e5a CH |
568 | if (!op_is_sync(op)) { |
569 | struct kyber_queue_data *kqd = data->q->elevator->elevator_data; | |
570 | ||
00e04393 | 571 | data->shallow_depth = kqd->async_depth; |
5bbf4e5a CH |
572 | } |
573 | } | |
00e04393 | 574 | |
a6088845 JW |
575 | static bool kyber_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio) |
576 | { | |
577 | struct kyber_hctx_data *khd = hctx->sched_data; | |
578 | struct blk_mq_ctx *ctx = blk_mq_get_ctx(hctx->queue); | |
579 | struct kyber_ctx_queue *kcq = &khd->kcqs[ctx->index_hw]; | |
580 | unsigned int sched_domain = kyber_sched_domain(bio->bi_opf); | |
581 | struct list_head *rq_list = &kcq->rq_list[sched_domain]; | |
582 | bool merged; | |
583 | ||
584 | spin_lock(&kcq->lock); | |
585 | merged = blk_mq_bio_list_merge(hctx->queue, rq_list, bio); | |
586 | spin_unlock(&kcq->lock); | |
587 | blk_mq_put_ctx(ctx); | |
588 | ||
589 | return merged; | |
590 | } | |
591 | ||
5bbf4e5a CH |
592 | static void kyber_prepare_request(struct request *rq, struct bio *bio) |
593 | { | |
594 | rq_set_domain_token(rq, -1); | |
00e04393 OS |
595 | } |
596 | ||
a6088845 JW |
597 | static void kyber_insert_requests(struct blk_mq_hw_ctx *hctx, |
598 | struct list_head *rq_list, bool at_head) | |
599 | { | |
600 | struct kyber_hctx_data *khd = hctx->sched_data; | |
601 | struct request *rq, *next; | |
602 | ||
603 | list_for_each_entry_safe(rq, next, rq_list, queuelist) { | |
604 | unsigned int sched_domain = kyber_sched_domain(rq->cmd_flags); | |
605 | struct kyber_ctx_queue *kcq = &khd->kcqs[rq->mq_ctx->index_hw]; | |
606 | struct list_head *head = &kcq->rq_list[sched_domain]; | |
607 | ||
608 | spin_lock(&kcq->lock); | |
609 | if (at_head) | |
610 | list_move(&rq->queuelist, head); | |
611 | else | |
612 | list_move_tail(&rq->queuelist, head); | |
613 | sbitmap_set_bit(&khd->kcq_map[sched_domain], | |
614 | rq->mq_ctx->index_hw); | |
615 | blk_mq_sched_request_inserted(rq); | |
616 | spin_unlock(&kcq->lock); | |
617 | } | |
618 | } | |
619 | ||
7b9e9361 | 620 | static void kyber_finish_request(struct request *rq) |
00e04393 | 621 | { |
7b9e9361 | 622 | struct kyber_queue_data *kqd = rq->q->elevator->elevator_data; |
00e04393 OS |
623 | |
624 | rq_clear_domain_token(kqd, rq); | |
00e04393 OS |
625 | } |
626 | ||
6e25cb01 OS |
627 | static void add_latency_sample(struct kyber_cpu_latency *cpu_latency, |
628 | unsigned int sched_domain, unsigned int type, | |
629 | u64 target, u64 latency) | |
00e04393 | 630 | { |
6e25cb01 OS |
631 | unsigned int bucket; |
632 | u64 divisor; | |
00e04393 | 633 | |
6e25cb01 OS |
634 | if (latency > 0) { |
635 | divisor = max_t(u64, target >> KYBER_LATENCY_SHIFT, 1); | |
636 | bucket = min_t(unsigned int, div64_u64(latency - 1, divisor), | |
637 | KYBER_LATENCY_BUCKETS - 1); | |
638 | } else { | |
639 | bucket = 0; | |
00e04393 OS |
640 | } |
641 | ||
6e25cb01 OS |
642 | atomic_inc(&cpu_latency->buckets[sched_domain][type][bucket]); |
643 | } | |
00e04393 | 644 | |
6e25cb01 OS |
645 | static void kyber_completed_request(struct request *rq, u64 now) |
646 | { | |
647 | struct kyber_queue_data *kqd = rq->q->elevator->elevator_data; | |
648 | struct kyber_cpu_latency *cpu_latency; | |
649 | unsigned int sched_domain; | |
650 | u64 target; | |
651 | ||
652 | sched_domain = kyber_sched_domain(rq->cmd_flags); | |
653 | if (sched_domain == KYBER_OTHER) | |
00e04393 OS |
654 | return; |
655 | ||
6e25cb01 OS |
656 | cpu_latency = get_cpu_ptr(kqd->cpu_latency); |
657 | target = kqd->latency_targets[sched_domain]; | |
658 | add_latency_sample(cpu_latency, sched_domain, KYBER_TOTAL_LATENCY, | |
659 | target, now - rq->start_time_ns); | |
660 | add_latency_sample(cpu_latency, sched_domain, KYBER_IO_LATENCY, target, | |
661 | now - rq->io_start_time_ns); | |
662 | put_cpu_ptr(kqd->cpu_latency); | |
00e04393 | 663 | |
6e25cb01 | 664 | timer_reduce(&kqd->timer, jiffies + HZ / 10); |
00e04393 OS |
665 | } |
666 | ||
a6088845 JW |
667 | struct flush_kcq_data { |
668 | struct kyber_hctx_data *khd; | |
669 | unsigned int sched_domain; | |
670 | struct list_head *list; | |
671 | }; | |
672 | ||
673 | static bool flush_busy_kcq(struct sbitmap *sb, unsigned int bitnr, void *data) | |
00e04393 | 674 | { |
a6088845 JW |
675 | struct flush_kcq_data *flush_data = data; |
676 | struct kyber_ctx_queue *kcq = &flush_data->khd->kcqs[bitnr]; | |
00e04393 | 677 | |
a6088845 JW |
678 | spin_lock(&kcq->lock); |
679 | list_splice_tail_init(&kcq->rq_list[flush_data->sched_domain], | |
680 | flush_data->list); | |
681 | sbitmap_clear_bit(sb, bitnr); | |
682 | spin_unlock(&kcq->lock); | |
00e04393 | 683 | |
a6088845 JW |
684 | return true; |
685 | } | |
686 | ||
687 | static void kyber_flush_busy_kcqs(struct kyber_hctx_data *khd, | |
688 | unsigned int sched_domain, | |
689 | struct list_head *list) | |
690 | { | |
691 | struct flush_kcq_data data = { | |
692 | .khd = khd, | |
693 | .sched_domain = sched_domain, | |
694 | .list = list, | |
695 | }; | |
696 | ||
697 | sbitmap_for_each_set(&khd->kcq_map[sched_domain], | |
698 | flush_busy_kcq, &data); | |
00e04393 OS |
699 | } |
700 | ||
ac6424b9 | 701 | static int kyber_domain_wake(wait_queue_entry_t *wait, unsigned mode, int flags, |
00e04393 OS |
702 | void *key) |
703 | { | |
704 | struct blk_mq_hw_ctx *hctx = READ_ONCE(wait->private); | |
705 | ||
2055da97 | 706 | list_del_init(&wait->entry); |
00e04393 OS |
707 | blk_mq_run_hw_queue(hctx, true); |
708 | return 1; | |
709 | } | |
710 | ||
711 | static int kyber_get_domain_token(struct kyber_queue_data *kqd, | |
712 | struct kyber_hctx_data *khd, | |
713 | struct blk_mq_hw_ctx *hctx) | |
714 | { | |
715 | unsigned int sched_domain = khd->cur_domain; | |
716 | struct sbitmap_queue *domain_tokens = &kqd->domain_tokens[sched_domain]; | |
ac6424b9 | 717 | wait_queue_entry_t *wait = &khd->domain_wait[sched_domain]; |
00e04393 OS |
718 | struct sbq_wait_state *ws; |
719 | int nr; | |
720 | ||
721 | nr = __sbitmap_queue_get(domain_tokens); | |
00e04393 OS |
722 | |
723 | /* | |
724 | * If we failed to get a domain token, make sure the hardware queue is | |
725 | * run when one becomes available. Note that this is serialized on | |
726 | * khd->lock, but we still need to be careful about the waker. | |
727 | */ | |
fcf38cdf | 728 | if (nr < 0 && list_empty_careful(&wait->entry)) { |
00e04393 OS |
729 | ws = sbq_wait_ptr(domain_tokens, |
730 | &khd->wait_index[sched_domain]); | |
fcf38cdf | 731 | khd->domain_ws[sched_domain] = ws; |
00e04393 OS |
732 | add_wait_queue(&ws->wait, wait); |
733 | ||
734 | /* | |
735 | * Try again in case a token was freed before we got on the wait | |
fcf38cdf | 736 | * queue. |
00e04393 OS |
737 | */ |
738 | nr = __sbitmap_queue_get(domain_tokens); | |
fcf38cdf | 739 | } |
8cf46660 | 740 | |
fcf38cdf OS |
741 | /* |
742 | * If we got a token while we were on the wait queue, remove ourselves | |
743 | * from the wait queue to ensure that all wake ups make forward | |
744 | * progress. It's possible that the waker already deleted the entry | |
745 | * between the !list_empty_careful() check and us grabbing the lock, but | |
746 | * list_del_init() is okay with that. | |
747 | */ | |
748 | if (nr >= 0 && !list_empty_careful(&wait->entry)) { | |
749 | ws = khd->domain_ws[sched_domain]; | |
750 | spin_lock_irq(&ws->wait.lock); | |
751 | list_del_init(&wait->entry); | |
752 | spin_unlock_irq(&ws->wait.lock); | |
00e04393 | 753 | } |
fcf38cdf | 754 | |
00e04393 OS |
755 | return nr; |
756 | } | |
757 | ||
758 | static struct request * | |
759 | kyber_dispatch_cur_domain(struct kyber_queue_data *kqd, | |
760 | struct kyber_hctx_data *khd, | |
a6088845 | 761 | struct blk_mq_hw_ctx *hctx) |
00e04393 OS |
762 | { |
763 | struct list_head *rqs; | |
764 | struct request *rq; | |
765 | int nr; | |
766 | ||
767 | rqs = &khd->rqs[khd->cur_domain]; | |
00e04393 OS |
768 | |
769 | /* | |
a6088845 JW |
770 | * If we already have a flushed request, then we just need to get a |
771 | * token for it. Otherwise, if there are pending requests in the kcqs, | |
772 | * flush the kcqs, but only if we can get a token. If not, we should | |
773 | * leave the requests in the kcqs so that they can be merged. Note that | |
774 | * khd->lock serializes the flushes, so if we observed any bit set in | |
775 | * the kcq_map, we will always get a request. | |
00e04393 | 776 | */ |
a6088845 | 777 | rq = list_first_entry_or_null(rqs, struct request, queuelist); |
00e04393 OS |
778 | if (rq) { |
779 | nr = kyber_get_domain_token(kqd, khd, hctx); | |
780 | if (nr >= 0) { | |
781 | khd->batching++; | |
782 | rq_set_domain_token(rq, nr); | |
783 | list_del_init(&rq->queuelist); | |
784 | return rq; | |
6c3b7af1 OS |
785 | } else { |
786 | trace_kyber_throttled(kqd->q, | |
787 | kyber_domain_names[khd->cur_domain]); | |
00e04393 | 788 | } |
a6088845 JW |
789 | } else if (sbitmap_any_bit_set(&khd->kcq_map[khd->cur_domain])) { |
790 | nr = kyber_get_domain_token(kqd, khd, hctx); | |
791 | if (nr >= 0) { | |
792 | kyber_flush_busy_kcqs(khd, khd->cur_domain, rqs); | |
793 | rq = list_first_entry(rqs, struct request, queuelist); | |
794 | khd->batching++; | |
795 | rq_set_domain_token(rq, nr); | |
796 | list_del_init(&rq->queuelist); | |
797 | return rq; | |
6c3b7af1 OS |
798 | } else { |
799 | trace_kyber_throttled(kqd->q, | |
800 | kyber_domain_names[khd->cur_domain]); | |
a6088845 | 801 | } |
00e04393 OS |
802 | } |
803 | ||
804 | /* There were either no pending requests or no tokens. */ | |
805 | return NULL; | |
806 | } | |
807 | ||
808 | static struct request *kyber_dispatch_request(struct blk_mq_hw_ctx *hctx) | |
809 | { | |
810 | struct kyber_queue_data *kqd = hctx->queue->elevator->elevator_data; | |
811 | struct kyber_hctx_data *khd = hctx->sched_data; | |
00e04393 OS |
812 | struct request *rq; |
813 | int i; | |
814 | ||
815 | spin_lock(&khd->lock); | |
816 | ||
817 | /* | |
818 | * First, if we are still entitled to batch, try to dispatch a request | |
819 | * from the batch. | |
820 | */ | |
821 | if (khd->batching < kyber_batch_size[khd->cur_domain]) { | |
a6088845 | 822 | rq = kyber_dispatch_cur_domain(kqd, khd, hctx); |
00e04393 OS |
823 | if (rq) |
824 | goto out; | |
825 | } | |
826 | ||
827 | /* | |
828 | * Either, | |
829 | * 1. We were no longer entitled to a batch. | |
830 | * 2. The domain we were batching didn't have any requests. | |
831 | * 3. The domain we were batching was out of tokens. | |
832 | * | |
833 | * Start another batch. Note that this wraps back around to the original | |
834 | * domain if no other domains have requests or tokens. | |
835 | */ | |
836 | khd->batching = 0; | |
837 | for (i = 0; i < KYBER_NUM_DOMAINS; i++) { | |
838 | if (khd->cur_domain == KYBER_NUM_DOMAINS - 1) | |
839 | khd->cur_domain = 0; | |
840 | else | |
841 | khd->cur_domain++; | |
842 | ||
a6088845 | 843 | rq = kyber_dispatch_cur_domain(kqd, khd, hctx); |
00e04393 OS |
844 | if (rq) |
845 | goto out; | |
846 | } | |
847 | ||
848 | rq = NULL; | |
849 | out: | |
850 | spin_unlock(&khd->lock); | |
851 | return rq; | |
852 | } | |
853 | ||
854 | static bool kyber_has_work(struct blk_mq_hw_ctx *hctx) | |
855 | { | |
856 | struct kyber_hctx_data *khd = hctx->sched_data; | |
857 | int i; | |
858 | ||
859 | for (i = 0; i < KYBER_NUM_DOMAINS; i++) { | |
a6088845 JW |
860 | if (!list_empty_careful(&khd->rqs[i]) || |
861 | sbitmap_any_bit_set(&khd->kcq_map[i])) | |
00e04393 OS |
862 | return true; |
863 | } | |
a6088845 JW |
864 | |
865 | return false; | |
00e04393 OS |
866 | } |
867 | ||
6e25cb01 OS |
868 | #define KYBER_LAT_SHOW_STORE(domain, name) \ |
869 | static ssize_t kyber_##name##_lat_show(struct elevator_queue *e, \ | |
870 | char *page) \ | |
00e04393 OS |
871 | { \ |
872 | struct kyber_queue_data *kqd = e->elevator_data; \ | |
873 | \ | |
6e25cb01 | 874 | return sprintf(page, "%llu\n", kqd->latency_targets[domain]); \ |
00e04393 OS |
875 | } \ |
876 | \ | |
6e25cb01 OS |
877 | static ssize_t kyber_##name##_lat_store(struct elevator_queue *e, \ |
878 | const char *page, size_t count) \ | |
00e04393 OS |
879 | { \ |
880 | struct kyber_queue_data *kqd = e->elevator_data; \ | |
881 | unsigned long long nsec; \ | |
882 | int ret; \ | |
883 | \ | |
884 | ret = kstrtoull(page, 10, &nsec); \ | |
885 | if (ret) \ | |
886 | return ret; \ | |
887 | \ | |
6e25cb01 | 888 | kqd->latency_targets[domain] = nsec; \ |
00e04393 OS |
889 | \ |
890 | return count; \ | |
891 | } | |
6e25cb01 OS |
892 | KYBER_LAT_SHOW_STORE(KYBER_READ, read); |
893 | KYBER_LAT_SHOW_STORE(KYBER_WRITE, write); | |
00e04393 OS |
894 | #undef KYBER_LAT_SHOW_STORE |
895 | ||
896 | #define KYBER_LAT_ATTR(op) __ATTR(op##_lat_nsec, 0644, kyber_##op##_lat_show, kyber_##op##_lat_store) | |
897 | static struct elv_fs_entry kyber_sched_attrs[] = { | |
898 | KYBER_LAT_ATTR(read), | |
899 | KYBER_LAT_ATTR(write), | |
900 | __ATTR_NULL | |
901 | }; | |
902 | #undef KYBER_LAT_ATTR | |
903 | ||
16b738f6 OS |
904 | #ifdef CONFIG_BLK_DEBUG_FS |
905 | #define KYBER_DEBUGFS_DOMAIN_ATTRS(domain, name) \ | |
906 | static int kyber_##name##_tokens_show(void *data, struct seq_file *m) \ | |
907 | { \ | |
908 | struct request_queue *q = data; \ | |
909 | struct kyber_queue_data *kqd = q->elevator->elevator_data; \ | |
910 | \ | |
911 | sbitmap_queue_show(&kqd->domain_tokens[domain], m); \ | |
912 | return 0; \ | |
913 | } \ | |
914 | \ | |
915 | static void *kyber_##name##_rqs_start(struct seq_file *m, loff_t *pos) \ | |
916 | __acquires(&khd->lock) \ | |
917 | { \ | |
918 | struct blk_mq_hw_ctx *hctx = m->private; \ | |
919 | struct kyber_hctx_data *khd = hctx->sched_data; \ | |
920 | \ | |
921 | spin_lock(&khd->lock); \ | |
922 | return seq_list_start(&khd->rqs[domain], *pos); \ | |
923 | } \ | |
924 | \ | |
925 | static void *kyber_##name##_rqs_next(struct seq_file *m, void *v, \ | |
926 | loff_t *pos) \ | |
927 | { \ | |
928 | struct blk_mq_hw_ctx *hctx = m->private; \ | |
929 | struct kyber_hctx_data *khd = hctx->sched_data; \ | |
930 | \ | |
931 | return seq_list_next(v, &khd->rqs[domain], pos); \ | |
932 | } \ | |
933 | \ | |
934 | static void kyber_##name##_rqs_stop(struct seq_file *m, void *v) \ | |
935 | __releases(&khd->lock) \ | |
936 | { \ | |
937 | struct blk_mq_hw_ctx *hctx = m->private; \ | |
938 | struct kyber_hctx_data *khd = hctx->sched_data; \ | |
939 | \ | |
940 | spin_unlock(&khd->lock); \ | |
941 | } \ | |
942 | \ | |
943 | static const struct seq_operations kyber_##name##_rqs_seq_ops = { \ | |
944 | .start = kyber_##name##_rqs_start, \ | |
945 | .next = kyber_##name##_rqs_next, \ | |
946 | .stop = kyber_##name##_rqs_stop, \ | |
947 | .show = blk_mq_debugfs_rq_show, \ | |
948 | }; \ | |
949 | \ | |
950 | static int kyber_##name##_waiting_show(void *data, struct seq_file *m) \ | |
951 | { \ | |
952 | struct blk_mq_hw_ctx *hctx = data; \ | |
953 | struct kyber_hctx_data *khd = hctx->sched_data; \ | |
ac6424b9 | 954 | wait_queue_entry_t *wait = &khd->domain_wait[domain]; \ |
16b738f6 | 955 | \ |
2055da97 | 956 | seq_printf(m, "%d\n", !list_empty_careful(&wait->entry)); \ |
16b738f6 OS |
957 | return 0; \ |
958 | } | |
959 | KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_READ, read) | |
6e25cb01 OS |
960 | KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_WRITE, write) |
961 | KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_DISCARD, discard) | |
16b738f6 OS |
962 | KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_OTHER, other) |
963 | #undef KYBER_DEBUGFS_DOMAIN_ATTRS | |
964 | ||
965 | static int kyber_async_depth_show(void *data, struct seq_file *m) | |
966 | { | |
967 | struct request_queue *q = data; | |
968 | struct kyber_queue_data *kqd = q->elevator->elevator_data; | |
969 | ||
970 | seq_printf(m, "%u\n", kqd->async_depth); | |
971 | return 0; | |
972 | } | |
973 | ||
974 | static int kyber_cur_domain_show(void *data, struct seq_file *m) | |
975 | { | |
976 | struct blk_mq_hw_ctx *hctx = data; | |
977 | struct kyber_hctx_data *khd = hctx->sched_data; | |
978 | ||
6c3b7af1 | 979 | seq_printf(m, "%s\n", kyber_domain_names[khd->cur_domain]); |
16b738f6 OS |
980 | return 0; |
981 | } | |
982 | ||
983 | static int kyber_batching_show(void *data, struct seq_file *m) | |
984 | { | |
985 | struct blk_mq_hw_ctx *hctx = data; | |
986 | struct kyber_hctx_data *khd = hctx->sched_data; | |
987 | ||
988 | seq_printf(m, "%u\n", khd->batching); | |
989 | return 0; | |
990 | } | |
991 | ||
992 | #define KYBER_QUEUE_DOMAIN_ATTRS(name) \ | |
993 | {#name "_tokens", 0400, kyber_##name##_tokens_show} | |
994 | static const struct blk_mq_debugfs_attr kyber_queue_debugfs_attrs[] = { | |
995 | KYBER_QUEUE_DOMAIN_ATTRS(read), | |
6e25cb01 OS |
996 | KYBER_QUEUE_DOMAIN_ATTRS(write), |
997 | KYBER_QUEUE_DOMAIN_ATTRS(discard), | |
16b738f6 OS |
998 | KYBER_QUEUE_DOMAIN_ATTRS(other), |
999 | {"async_depth", 0400, kyber_async_depth_show}, | |
1000 | {}, | |
1001 | }; | |
1002 | #undef KYBER_QUEUE_DOMAIN_ATTRS | |
1003 | ||
1004 | #define KYBER_HCTX_DOMAIN_ATTRS(name) \ | |
1005 | {#name "_rqs", 0400, .seq_ops = &kyber_##name##_rqs_seq_ops}, \ | |
1006 | {#name "_waiting", 0400, kyber_##name##_waiting_show} | |
1007 | static const struct blk_mq_debugfs_attr kyber_hctx_debugfs_attrs[] = { | |
1008 | KYBER_HCTX_DOMAIN_ATTRS(read), | |
6e25cb01 OS |
1009 | KYBER_HCTX_DOMAIN_ATTRS(write), |
1010 | KYBER_HCTX_DOMAIN_ATTRS(discard), | |
16b738f6 OS |
1011 | KYBER_HCTX_DOMAIN_ATTRS(other), |
1012 | {"cur_domain", 0400, kyber_cur_domain_show}, | |
1013 | {"batching", 0400, kyber_batching_show}, | |
1014 | {}, | |
1015 | }; | |
1016 | #undef KYBER_HCTX_DOMAIN_ATTRS | |
1017 | #endif | |
1018 | ||
00e04393 OS |
1019 | static struct elevator_type kyber_sched = { |
1020 | .ops.mq = { | |
1021 | .init_sched = kyber_init_sched, | |
1022 | .exit_sched = kyber_exit_sched, | |
1023 | .init_hctx = kyber_init_hctx, | |
1024 | .exit_hctx = kyber_exit_hctx, | |
5bbf4e5a | 1025 | .limit_depth = kyber_limit_depth, |
a6088845 | 1026 | .bio_merge = kyber_bio_merge, |
5bbf4e5a | 1027 | .prepare_request = kyber_prepare_request, |
a6088845 | 1028 | .insert_requests = kyber_insert_requests, |
7b9e9361 | 1029 | .finish_request = kyber_finish_request, |
ba989a01 | 1030 | .requeue_request = kyber_finish_request, |
00e04393 OS |
1031 | .completed_request = kyber_completed_request, |
1032 | .dispatch_request = kyber_dispatch_request, | |
1033 | .has_work = kyber_has_work, | |
1034 | }, | |
1035 | .uses_mq = true, | |
16b738f6 OS |
1036 | #ifdef CONFIG_BLK_DEBUG_FS |
1037 | .queue_debugfs_attrs = kyber_queue_debugfs_attrs, | |
1038 | .hctx_debugfs_attrs = kyber_hctx_debugfs_attrs, | |
1039 | #endif | |
00e04393 OS |
1040 | .elevator_attrs = kyber_sched_attrs, |
1041 | .elevator_name = "kyber", | |
1042 | .elevator_owner = THIS_MODULE, | |
1043 | }; | |
1044 | ||
1045 | static int __init kyber_init(void) | |
1046 | { | |
1047 | return elv_register(&kyber_sched); | |
1048 | } | |
1049 | ||
1050 | static void __exit kyber_exit(void) | |
1051 | { | |
1052 | elv_unregister(&kyber_sched); | |
1053 | } | |
1054 | ||
1055 | module_init(kyber_init); | |
1056 | module_exit(kyber_exit); | |
1057 | ||
1058 | MODULE_AUTHOR("Omar Sandoval"); | |
1059 | MODULE_LICENSE("GPL"); | |
1060 | MODULE_DESCRIPTION("Kyber I/O scheduler"); |