forked from ~ljy/RK356X_SDK_RELEASE

blame | history | patch | commit | commitdiff
hc
2023-12-11 d2ccde1c8e90d38cee87a1b0309ad2827f3fd30d 
 kernel/block/blk-mq-sched.c
....@@ -1,3 +1,4 @@
1
+// SPDX-License-Identifier: GPL-2.0
12 /*
23  * blk-mq scheduling framework
34  *
....@@ -6,6 +7,7 @@
67 #include <linux/kernel.h>
78 #include <linux/module.h>
89 #include <linux/blk-mq.h>
10
+#include <linux/list_sort.h>
911 
1012 #include <trace/events/block.h>
1113 
....@@ -16,30 +18,22 @@
1618 #include "blk-mq-tag.h"
1719 #include "blk-wbt.h"
1820 
19
-void blk_mq_sched_free_hctx_data(struct request_queue *q,
20
-				 void (*exit)(struct blk_mq_hw_ctx *))
21
-{
22
-	struct blk_mq_hw_ctx *hctx;
23
-	int i;
24
-
25
-	queue_for_each_hw_ctx(q, hctx, i) {
26
-		if (exit && hctx->sched_data)
27
-			exit(hctx);
28
-		kfree(hctx->sched_data);
29
-		hctx->sched_data = NULL;
30
-	}
31
-}
32
-EXPORT_SYMBOL_GPL(blk_mq_sched_free_hctx_data);
33
-
34
-void blk_mq_sched_assign_ioc(struct request *rq, struct bio *bio)
21
+void blk_mq_sched_assign_ioc(struct request *rq)
3522 {
3623 	struct request_queue *q = rq->q;
37
-	struct io_context *ioc = rq_ioc(bio);
24
+	struct io_context *ioc;
3825 	struct io_cq *icq;
3926 
40
-	spin_lock_irq(q->queue_lock);
27
+	/*
28
+	 * May not have an IO context if it's a passthrough request
29
+	 */
30
+	ioc = current->io_context;
31
+	if (!ioc)
32
+		return;
33
+
34
+	spin_lock_irq(&q->queue_lock);
4135 	icq = ioc_lookup_icq(ioc, q);
42
-	spin_unlock_irq(q->queue_lock);
36
+	spin_unlock_irq(&q->queue_lock);
4337 
4438 	if (!icq) {
4539 		icq = ioc_create_icq(ioc, q, GFP_ATOMIC);
....@@ -81,30 +75,85 @@
8175 	blk_mq_run_hw_queue(hctx, true);
8276 }
8377 
78
+static int sched_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
79
+{
80
+	struct request *rqa = container_of(a, struct request, queuelist);
81
+	struct request *rqb = container_of(b, struct request, queuelist);
82
+
83
+	return rqa->mq_hctx > rqb->mq_hctx;
84
+}
85
+
86
+static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list)
87
+{
88
+	struct blk_mq_hw_ctx *hctx =
89
+		list_first_entry(rq_list, struct request, queuelist)->mq_hctx;
90
+	struct request *rq;
91
+	LIST_HEAD(hctx_list);
92
+	unsigned int count = 0;
93
+
94
+	list_for_each_entry(rq, rq_list, queuelist) {
95
+		if (rq->mq_hctx != hctx) {
96
+			list_cut_before(&hctx_list, rq_list, &rq->queuelist);
97
+			goto dispatch;
98
+		}
99
+		count++;
100
+	}
101
+	list_splice_tail_init(rq_list, &hctx_list);
102
+
103
+dispatch:
104
+	return blk_mq_dispatch_rq_list(hctx, &hctx_list, count);
105
+}
106
+
107
+#define BLK_MQ_BUDGET_DELAY	3		/* ms units */
108
+
84109 /*
85110  * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
86111  * its queue by itself in its completion handler, so we don't need to
87112  * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
113
+ *
114
+ * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
115
+ * be run again.  This is necessary to avoid starving flushes.
88116  */
89
-static void blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
117
+static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
90118 {
91119 	struct request_queue *q = hctx->queue;
92120 	struct elevator_queue *e = q->elevator;
121
+	bool multi_hctxs = false, run_queue = false;
122
+	bool dispatched = false, busy = false;
123
+	unsigned int max_dispatch;
93124 	LIST_HEAD(rq_list);
125
+	int count = 0;
126
+
127
+	if (hctx->dispatch_busy)
128
+		max_dispatch = 1;
129
+	else
130
+		max_dispatch = hctx->queue->nr_requests;
94131 
95132 	do {
96133 		struct request *rq;
97134 
98
-		if (e->type->ops.mq.has_work &&
99
-				!e->type->ops.mq.has_work(hctx))
135
+		if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
100136 			break;
101137 
102
-		if (!blk_mq_get_dispatch_budget(hctx))
138
+		if (!list_empty_careful(&hctx->dispatch)) {
139
+			busy = true;
140
+			break;
141
+		}
142
+
143
+		if (!blk_mq_get_dispatch_budget(q))
103144 			break;
104145 
105
-		rq = e->type->ops.mq.dispatch_request(hctx);
146
+		rq = e->type->ops.dispatch_request(hctx);
106147 		if (!rq) {
107
-			blk_mq_put_dispatch_budget(hctx);
148
+			blk_mq_put_dispatch_budget(q);
149
+			/*
150
+			 * We're releasing without dispatching. Holding the
151
+			 * budget could have blocked any "hctx"s with the
152
+			 * same queue and if we didn't dispatch then there's
153
+			 * no guarantee anyone will kick the queue.  Kick it
154
+			 * ourselves.
155
+			 */
156
+			run_queue = true;
108157 			break;
109158 		}
110159 
....@@ -113,14 +162,57 @@
113162 		 * if this rq won't be queued to driver via .queue_rq()
114163 		 * in blk_mq_dispatch_rq_list().
115164 		 */
116
-		list_add(&rq->queuelist, &rq_list);
117
-	} while (blk_mq_dispatch_rq_list(q, &rq_list, true));
165
+		list_add_tail(&rq->queuelist, &rq_list);
166
+		if (rq->mq_hctx != hctx)
167
+			multi_hctxs = true;
168
+	} while (++count < max_dispatch);
169
+
170
+	if (!count) {
171
+		if (run_queue)
172
+			blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
173
+	} else if (multi_hctxs) {
174
+		/*
175
+		 * Requests from different hctx may be dequeued from some
176
+		 * schedulers, such as bfq and deadline.
177
+		 *
178
+		 * Sort the requests in the list according to their hctx,
179
+		 * dispatch batching requests from same hctx at a time.
180
+		 */
181
+		list_sort(NULL, &rq_list, sched_rq_cmp);
182
+		do {
183
+			dispatched |= blk_mq_dispatch_hctx_list(&rq_list);
184
+		} while (!list_empty(&rq_list));
185
+	} else {
186
+		dispatched = blk_mq_dispatch_rq_list(hctx, &rq_list, count);
187
+	}
188
+
189
+	if (busy)
190
+		return -EAGAIN;
191
+	return !!dispatched;
192
+}
193
+
194
+static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
195
+{
196
+	unsigned long end = jiffies + HZ;
197
+	int ret;
198
+
199
+	do {
200
+		ret = __blk_mq_do_dispatch_sched(hctx);
201
+		if (ret != 1)
202
+			break;
203
+		if (need_resched() || time_is_before_jiffies(end)) {
204
+			blk_mq_delay_run_hw_queue(hctx, 0);
205
+			break;
206
+		}
207
+	} while (1);
208
+
209
+	return ret;
118210 }
119211 
120212 static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
121213 					  struct blk_mq_ctx *ctx)
122214 {
123
-	unsigned idx = ctx->index_hw;
215
+	unsigned short idx = ctx->index_hw[hctx->type];
124216 
125217 	if (++idx == hctx->nr_ctx)
126218 		idx = 0;
....@@ -132,25 +224,41 @@
132224  * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
133225  * its queue by itself in its completion handler, so we don't need to
134226  * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
227
+ *
228
+ * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
229
+ * be run again.  This is necessary to avoid starving flushes.
135230  */
136
-static void blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
231
+static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
137232 {
138233 	struct request_queue *q = hctx->queue;
139234 	LIST_HEAD(rq_list);
140235 	struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
236
+	int ret = 0;
237
+	struct request *rq;
141238 
142239 	do {
143
-		struct request *rq;
240
+		if (!list_empty_careful(&hctx->dispatch)) {
241
+			ret = -EAGAIN;
242
+			break;
243
+		}
144244 
145245 		if (!sbitmap_any_bit_set(&hctx->ctx_map))
146246 			break;
147247 
148
-		if (!blk_mq_get_dispatch_budget(hctx))
248
+		if (!blk_mq_get_dispatch_budget(q))
149249 			break;
150250 
151251 		rq = blk_mq_dequeue_from_ctx(hctx, ctx);
152252 		if (!rq) {
153
-			blk_mq_put_dispatch_budget(hctx);
253
+			blk_mq_put_dispatch_budget(q);
254
+			/*
255
+			 * We're releasing without dispatching. Holding the
256
+			 * budget could have blocked any "hctx"s with the
257
+			 * same queue and if we didn't dispatch then there's
258
+			 * no guarantee anyone will kick the queue.  Kick it
259
+			 * ourselves.
260
+			 */
261
+			blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
154262 			break;
155263 		}
156264 
....@@ -164,23 +272,19 @@
164272 		/* round robin for fair dispatch */
165273 		ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
166274 
167
-	} while (blk_mq_dispatch_rq_list(q, &rq_list, true));
275
+	} while (blk_mq_dispatch_rq_list(rq->mq_hctx, &rq_list, 1));
168276 
169277 	WRITE_ONCE(hctx->dispatch_from, ctx);
278
+	return ret;
170279 }
171280 
172
-void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
281
+static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
173282 {
174283 	struct request_queue *q = hctx->queue;
175284 	struct elevator_queue *e = q->elevator;
176
-	const bool has_sched_dispatch = e && e->type->ops.mq.dispatch_request;
285
+	const bool has_sched_dispatch = e && e->type->ops.dispatch_request;
286
+	int ret = 0;
177287 	LIST_HEAD(rq_list);
178
-
179
-	/* RCU or SRCU read lock is needed before checking quiesced flag */
180
-	if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
181
-		return;
182
-
183
-	hctx->run++;
184288 
185289 	/*
186290 	 * If we have previous entries on our dispatch list, grab them first for
....@@ -208,136 +312,78 @@
208312 	 */
209313 	if (!list_empty(&rq_list)) {
210314 		blk_mq_sched_mark_restart_hctx(hctx);
211
-		if (blk_mq_dispatch_rq_list(q, &rq_list, false)) {
315
+		if (blk_mq_dispatch_rq_list(hctx, &rq_list, 0)) {
212316 			if (has_sched_dispatch)
213
-				blk_mq_do_dispatch_sched(hctx);
317
+				ret = blk_mq_do_dispatch_sched(hctx);
214318 			else
215
-				blk_mq_do_dispatch_ctx(hctx);
319
+				ret = blk_mq_do_dispatch_ctx(hctx);
216320 		}
217321 	} else if (has_sched_dispatch) {
218
-		blk_mq_do_dispatch_sched(hctx);
322
+		ret = blk_mq_do_dispatch_sched(hctx);
219323 	} else if (hctx->dispatch_busy) {
220324 		/* dequeue request one by one from sw queue if queue is busy */
221
-		blk_mq_do_dispatch_ctx(hctx);
325
+		ret = blk_mq_do_dispatch_ctx(hctx);
222326 	} else {
223327 		blk_mq_flush_busy_ctxs(hctx, &rq_list);
224
-		blk_mq_dispatch_rq_list(q, &rq_list, false);
328
+		blk_mq_dispatch_rq_list(hctx, &rq_list, 0);
225329 	}
330
+
331
+	return ret;
226332 }
227333 
228
-bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
229
-			    struct request **merged_request)
334
+void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
230335 {
231
-	struct request *rq;
336
+	struct request_queue *q = hctx->queue;
232337 
233
-	switch (elv_merge(q, &rq, bio)) {
234
-	case ELEVATOR_BACK_MERGE:
235
-		if (!blk_mq_sched_allow_merge(q, rq, bio))
236
-			return false;
237
-		if (!bio_attempt_back_merge(q, rq, bio))
238
-			return false;
239
-		*merged_request = attempt_back_merge(q, rq);
240
-		if (!*merged_request)
241
-			elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
242
-		return true;
243
-	case ELEVATOR_FRONT_MERGE:
244
-		if (!blk_mq_sched_allow_merge(q, rq, bio))
245
-			return false;
246
-		if (!bio_attempt_front_merge(q, rq, bio))
247
-			return false;
248
-		*merged_request = attempt_front_merge(q, rq);
249
-		if (!*merged_request)
250
-			elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
251
-		return true;
252
-	case ELEVATOR_DISCARD_MERGE:
253
-		return bio_attempt_discard_merge(q, rq, bio);
254
-	default:
255
-		return false;
338
+	/* RCU or SRCU read lock is needed before checking quiesced flag */
339
+	if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
340
+		return;
341
+
342
+	hctx->run++;
343
+
344
+	/*
345
+	 * A return of -EAGAIN is an indication that hctx->dispatch is not
346
+	 * empty and we must run again in order to avoid starving flushes.
347
+	 */
348
+	if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) {
349
+		if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN)
350
+			blk_mq_run_hw_queue(hctx, true);
256351 	}
257352 }
258
-EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
259353 
260
-/*
261
- * Iterate list of requests and see if we can merge this bio with any
262
- * of them.
263
- */
264
-bool blk_mq_bio_list_merge(struct request_queue *q, struct list_head *list,
265
-			   struct bio *bio)
266
-{
267
-	struct request *rq;
268
-	int checked = 8;
269
-
270
-	list_for_each_entry_reverse(rq, list, queuelist) {
271
-		bool merged = false;
272
-
273
-		if (!checked--)
274
-			break;
275
-
276
-		if (!blk_rq_merge_ok(rq, bio))
277
-			continue;
278
-
279
-		switch (blk_try_merge(rq, bio)) {
280
-		case ELEVATOR_BACK_MERGE:
281
-			if (blk_mq_sched_allow_merge(q, rq, bio))
282
-				merged = bio_attempt_back_merge(q, rq, bio);
283
-			break;
284
-		case ELEVATOR_FRONT_MERGE:
285
-			if (blk_mq_sched_allow_merge(q, rq, bio))
286
-				merged = bio_attempt_front_merge(q, rq, bio);
287
-			break;
288
-		case ELEVATOR_DISCARD_MERGE:
289
-			merged = bio_attempt_discard_merge(q, rq, bio);
290
-			break;
291
-		default:
292
-			continue;
293
-		}
294
-
295
-		return merged;
296
-	}
297
-
298
-	return false;
299
-}
300
-EXPORT_SYMBOL_GPL(blk_mq_bio_list_merge);
301
-
302
-/*
303
- * Reverse check our software queue for entries that we could potentially
304
- * merge with. Currently includes a hand-wavy stop count of 8, to not spend
305
- * too much time checking for merges.
306
- */
307
-static bool blk_mq_attempt_merge(struct request_queue *q,
308
-				 struct blk_mq_ctx *ctx, struct bio *bio)
309
-{
310
-	lockdep_assert_held(&ctx->lock);
311
-
312
-	if (blk_mq_bio_list_merge(q, &ctx->rq_list, bio)) {
313
-		ctx->rq_merged++;
314
-		return true;
315
-	}
316
-
317
-	return false;
318
-}
319
-
320
-bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
354
+bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio,
355
+		unsigned int nr_segs)
321356 {
322357 	struct elevator_queue *e = q->elevator;
323
-	struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
324
-	struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
358
+	struct blk_mq_ctx *ctx;
359
+	struct blk_mq_hw_ctx *hctx;
325360 	bool ret = false;
361
+	enum hctx_type type;
326362 
327
-	if (e && e->type->ops.mq.bio_merge) {
328
-		blk_mq_put_ctx(ctx);
329
-		return e->type->ops.mq.bio_merge(hctx, bio);
363
+	if (e && e->type->ops.bio_merge)
364
+		return e->type->ops.bio_merge(q, bio, nr_segs);
365
+
366
+	ctx = blk_mq_get_ctx(q);
367
+	hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
368
+	type = hctx->type;
369
+	if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE) ||
370
+	    list_empty_careful(&ctx->rq_lists[type]))
371
+		return false;
372
+
373
+	/* default per sw-queue merge */
374
+	spin_lock(&ctx->lock);
375
+	/*
376
+	 * Reverse check our software queue for entries that we could
377
+	 * potentially merge with. Currently includes a hand-wavy stop
378
+	 * count of 8, to not spend too much time checking for merges.
379
+	 */
380
+	if (blk_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs)) {
381
+		ctx->rq_merged++;
382
+		ret = true;
330383 	}
331384 
332
-	if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
333
-			!list_empty_careful(&ctx->rq_list)) {
334
-		/* default per sw-queue merge */
335
-		spin_lock(&ctx->lock);
336
-		ret = blk_mq_attempt_merge(q, ctx, bio);
337
-		spin_unlock(&ctx->lock);
338
-	}
385
+	spin_unlock(&ctx->lock);
339386 
340
-	blk_mq_put_ctx(ctx);
341387 	return ret;
342388 }
343389 
....@@ -357,13 +403,19 @@
357403 				       bool has_sched,
358404 				       struct request *rq)
359405 {
360
-	/* dispatch flush rq directly */
361
-	if (rq->rq_flags & RQF_FLUSH_SEQ) {
362
-		spin_lock(&hctx->lock);
363
-		list_add(&rq->queuelist, &hctx->dispatch);
364
-		spin_unlock(&hctx->lock);
406
+	/*
407
+	 * dispatch flush and passthrough rq directly
408
+	 *
409
+	 * passthrough request has to be added to hctx->dispatch directly.
410
+	 * For some reason, device may be in one situation which can't
411
+	 * handle FS request, so STS_RESOURCE is always returned and the
412
+	 * FS request will be added to hctx->dispatch. However passthrough
413
+	 * request may be required at that time for fixing the problem. If
414
+	 * passthrough request is added to scheduler queue, there isn't any
415
+	 * chance to dispatch it given we prioritize requests in hctx->dispatch.
416
+	 */
417
+	if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq))
365418 		return true;
366
-	}
367419 
368420 	if (has_sched)
369421 		rq->rq_flags |= RQF_SORTED;
....@@ -377,24 +429,42 @@
377429 	struct request_queue *q = rq->q;
378430 	struct elevator_queue *e = q->elevator;
379431 	struct blk_mq_ctx *ctx = rq->mq_ctx;
380
-	struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
432
+	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
381433 
382
-	/* flush rq in flush machinery need to be dispatched directly */
383
-	if (!(rq->rq_flags & RQF_FLUSH_SEQ) && op_is_flush(rq->cmd_flags)) {
384
-		blk_insert_flush(rq);
434
+	WARN_ON(e && (rq->tag != BLK_MQ_NO_TAG));
435
+
436
+	if (blk_mq_sched_bypass_insert(hctx, !!e, rq)) {
437
+		/*
438
+		 * Firstly normal IO request is inserted to scheduler queue or
439
+		 * sw queue, meantime we add flush request to dispatch queue(
440
+		 * hctx->dispatch) directly and there is at most one in-flight
441
+		 * flush request for each hw queue, so it doesn't matter to add
442
+		 * flush request to tail or front of the dispatch queue.
443
+		 *
444
+		 * Secondly in case of NCQ, flush request belongs to non-NCQ
445
+		 * command, and queueing it will fail when there is any
446
+		 * in-flight normal IO request(NCQ command). When adding flush
447
+		 * rq to the front of hctx->dispatch, it is easier to introduce
448
+		 * extra time to flush rq's latency because of S_SCHED_RESTART
449
+		 * compared with adding to the tail of dispatch queue, then
450
+		 * chance of flush merge is increased, and less flush requests
451
+		 * will be issued to controller. It is observed that ~10% time
452
+		 * is saved in blktests block/004 on disk attached to AHCI/NCQ
453
+		 * drive when adding flush rq to the front of hctx->dispatch.
454
+		 *
455
+		 * Simply queue flush rq to the front of hctx->dispatch so that
456
+		 * intensive flush workloads can benefit in case of NCQ HW.
457
+		 */
458
+		at_head = (rq->rq_flags & RQF_FLUSH_SEQ) ? true : at_head;
459
+		blk_mq_request_bypass_insert(rq, at_head, false);
385460 		goto run;
386461 	}
387462 
388
-	WARN_ON(e && (rq->tag != -1));
389
-
390
-	if (blk_mq_sched_bypass_insert(hctx, !!e, rq))
391
-		goto run;
392
-
393
-	if (e && e->type->ops.mq.insert_requests) {
463
+	if (e && e->type->ops.insert_requests) {
394464 		LIST_HEAD(list);
395465 
396466 		list_add(&rq->queuelist, &list);
397
-		e->type->ops.mq.insert_requests(hctx, &list, at_head);
467
+		e->type->ops.insert_requests(hctx, &list, at_head);
398468 	} else {
399469 		spin_lock(&ctx->lock);
400470 		__blk_mq_insert_request(hctx, rq, at_head);
....@@ -406,15 +476,23 @@
406476 		blk_mq_run_hw_queue(hctx, async);
407477 }
408478 
409
-void blk_mq_sched_insert_requests(struct request_queue *q,
479
+void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx,
410480 				  struct blk_mq_ctx *ctx,
411481 				  struct list_head *list, bool run_queue_async)
412482 {
413
-	struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
414
-	struct elevator_queue *e = hctx->queue->elevator;
483
+	struct elevator_queue *e;
484
+	struct request_queue *q = hctx->queue;
415485 
416
-	if (e && e->type->ops.mq.insert_requests)
417
-		e->type->ops.mq.insert_requests(hctx, list, false);
486
+	/*
487
+	 * blk_mq_sched_insert_requests() is called from flush plug
488
+	 * context only, and hold one usage counter to prevent queue
489
+	 * from being released.
490
+	 */
491
+	percpu_ref_get(&q->q_usage_counter);
492
+
493
+	e = hctx->queue->elevator;
494
+	if (e && e->type->ops.insert_requests)
495
+		e->type->ops.insert_requests(hctx, list, false);
418496 	else {
419497 		/*
420498 		 * try to issue requests directly if the hw queue isn't
....@@ -424,21 +502,25 @@
424502 		if (!hctx->dispatch_busy && !e && !run_queue_async) {
425503 			blk_mq_try_issue_list_directly(hctx, list);
426504 			if (list_empty(list))
427
-				return;
505
+				goto out;
428506 		}
429507 		blk_mq_insert_requests(hctx, ctx, list);
430508 	}
431509 
432510 	blk_mq_run_hw_queue(hctx, run_queue_async);
511
+ out:
512
+	percpu_ref_put(&q->q_usage_counter);
433513 }
434514 
435515 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
436516 				   struct blk_mq_hw_ctx *hctx,
437517 				   unsigned int hctx_idx)
438518 {
519
+	unsigned int flags = set->flags & ~BLK_MQ_F_TAG_HCTX_SHARED;
520
+
439521 	if (hctx->sched_tags) {
440522 		blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
441
-		blk_mq_free_rq_map(hctx->sched_tags);
523
+		blk_mq_free_rq_map(hctx->sched_tags, flags);
442524 		hctx->sched_tags = NULL;
443525 	}
444526 }
....@@ -448,10 +530,12 @@
448530 				   unsigned int hctx_idx)
449531 {
450532 	struct blk_mq_tag_set *set = q->tag_set;
533
+	/* Clear HCTX_SHARED so tags are init'ed */
534
+	unsigned int flags = set->flags & ~BLK_MQ_F_TAG_HCTX_SHARED;
451535 	int ret;
452536 
453537 	hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
454
-					       set->reserved_tags);
538
+					       set->reserved_tags, flags);
455539 	if (!hctx->sched_tags)
456540 		return -ENOMEM;
457541 
....@@ -462,14 +546,21 @@
462546 	return ret;
463547 }
464548 
549
+/* called in queue's release handler, tagset has gone away */
465550 static void blk_mq_sched_tags_teardown(struct request_queue *q)
466551 {
467
-	struct blk_mq_tag_set *set = q->tag_set;
468552 	struct blk_mq_hw_ctx *hctx;
469553 	int i;
470554 
471
-	queue_for_each_hw_ctx(q, hctx, i)
472
-		blk_mq_sched_free_tags(set, hctx, i);
555
+	queue_for_each_hw_ctx(q, hctx, i) {
556
+		/* Clear HCTX_SHARED so tags are freed */
557
+		unsigned int flags = hctx->flags & ~BLK_MQ_F_TAG_HCTX_SHARED;
558
+
559
+		if (hctx->sched_tags) {
560
+			blk_mq_free_rq_map(hctx->sched_tags, flags);
561
+			hctx->sched_tags = NULL;
562
+		}
563
+	}
473564 }
474565 
475566 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
....@@ -499,17 +590,18 @@
499590 			goto err;
500591 	}
501592 
502
-	ret = e->ops.mq.init_sched(q, e);
593
+	ret = e->ops.init_sched(q, e);
503594 	if (ret)
504595 		goto err;
505596 
506597 	blk_mq_debugfs_register_sched(q);
507598 
508599 	queue_for_each_hw_ctx(q, hctx, i) {
509
-		if (e->ops.mq.init_hctx) {
510
-			ret = e->ops.mq.init_hctx(hctx, i);
600
+		if (e->ops.init_hctx) {
601
+			ret = e->ops.init_hctx(hctx, i);
511602 			if (ret) {
512603 				eq = q->elevator;
604
+				blk_mq_sched_free_requests(q);
513605 				blk_mq_exit_sched(q, eq);
514606 				kobject_put(&eq->kobj);
515607 				return ret;
....@@ -521,9 +613,25 @@
521613 	return 0;
522614 
523615 err:
616
+	blk_mq_sched_free_requests(q);
524617 	blk_mq_sched_tags_teardown(q);
525618 	q->elevator = NULL;
526619 	return ret;
620
+}
621
+
622
+/*
623
+ * called in either blk_queue_cleanup or elevator_switch, tagset
624
+ * is required for freeing requests
625
+ */
626
+void blk_mq_sched_free_requests(struct request_queue *q)
627
+{
628
+	struct blk_mq_hw_ctx *hctx;
629
+	int i;
630
+
631
+	queue_for_each_hw_ctx(q, hctx, i) {
632
+		if (hctx->sched_tags)
633
+			blk_mq_free_rqs(q->tag_set, hctx->sched_tags, i);
634
+	}
527635 }
528636 
529637 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
....@@ -533,14 +641,14 @@
533641 
534642 	queue_for_each_hw_ctx(q, hctx, i) {
535643 		blk_mq_debugfs_unregister_sched_hctx(hctx);
536
-		if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
537
-			e->type->ops.mq.exit_hctx(hctx, i);
644
+		if (e->type->ops.exit_hctx && hctx->sched_data) {
645
+			e->type->ops.exit_hctx(hctx, i);
538646 			hctx->sched_data = NULL;
539647 		}
540648 	}
541649 	blk_mq_debugfs_unregister_sched(q);
542
-	if (e->type->ops.mq.exit_sched)
543
-		e->type->ops.mq.exit_sched(e);
650
+	if (e->type->ops.exit_sched)
651
+		e->type->ops.exit_sched(e);
544652 	blk_mq_sched_tags_teardown(q);
545653 	q->elevator = NULL;
546654 }