path: root/block/blk-mq.c (plain)
blob: 24fc09cf7f175adc5e2bb3f616798a39a39292e8

1	/*
2	* Block multiqueue core code
3	*
4	* Copyright (C) 2013-2014 Jens Axboe
5	* Copyright (C) 2013-2014 Christoph Hellwig
6	*/
7	#include <linux/kernel.h>
8	#include <linux/module.h>
9	#include <linux/backing-dev.h>
10	#include <linux/bio.h>
11	#include <linux/blkdev.h>
12	#include <linux/kmemleak.h>
13	#include <linux/mm.h>
14	#include <linux/init.h>
15	#include <linux/slab.h>
16	#include <linux/workqueue.h>
17	#include <linux/smp.h>
18	#include <linux/llist.h>
19	#include <linux/list_sort.h>
20	#include <linux/cpu.h>
21	#include <linux/cache.h>
22	#include <linux/sched/sysctl.h>
23	#include <linux/delay.h>
24	#include <linux/crash_dump.h>
25	#include <linux/prefetch.h>
26
27	#include <trace/events/block.h>
28
29	#include <linux/blk-mq.h>
30	#include "blk.h"
31	#include "blk-mq.h"
32	#include "blk-mq-tag.h"
33
34	static DEFINE_MUTEX(all_q_mutex);
35	static LIST_HEAD(all_q_list);
36
37	/*
38	* Check if any of the ctx's have pending work in this hardware queue
39	*/
40	static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
41	{
42	return sbitmap_any_bit_set(&hctx->ctx_map);
43	}
44
45	/*
46	* Mark this ctx as having pending work in this hardware queue
47	*/
48	static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
49	struct blk_mq_ctx *ctx)
50	{
51	if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
52	sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
53	}
54
55	static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
56	struct blk_mq_ctx *ctx)
57	{
58	sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
59	}
60
61	void blk_mq_freeze_queue_start(struct request_queue *q)
62	{
63	int freeze_depth;
64
65	freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
66	if (freeze_depth == 1) {
67	percpu_ref_kill(&q->q_usage_counter);
68	blk_mq_run_hw_queues(q, false);
69	}
70	}
71	EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
72
73	static void blk_mq_freeze_queue_wait(struct request_queue *q)
74	{
75	wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
76	}
77
78	/*
79	* Guarantee no request is in use, so we can change any data structure of
80	* the queue afterward.
81	*/
82	void blk_freeze_queue(struct request_queue *q)
83	{
84	/*
85	* In the !blk_mq case we are only calling this to kill the
86	* q_usage_counter, otherwise this increases the freeze depth
87	* and waits for it to return to zero. For this reason there is
88	* no blk_unfreeze_queue(), and blk_freeze_queue() is not
89	* exported to drivers as the only user for unfreeze is blk_mq.
90	*/
91	blk_mq_freeze_queue_start(q);
92	blk_mq_freeze_queue_wait(q);
93	}
94
95	void blk_mq_freeze_queue(struct request_queue *q)
96	{
97	/*
98	* ...just an alias to keep freeze and unfreeze actions balanced
99	* in the blk_mq_* namespace
100	*/
101	blk_freeze_queue(q);
102	}
103	EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
104
105	void blk_mq_unfreeze_queue(struct request_queue *q)
106	{
107	int freeze_depth;
108
109	freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
110	WARN_ON_ONCE(freeze_depth < 0);
111	if (!freeze_depth) {
112	percpu_ref_reinit(&q->q_usage_counter);
113	wake_up_all(&q->mq_freeze_wq);
114	}
115	}
116	EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
117
118	void blk_mq_wake_waiters(struct request_queue *q)
119	{
120	struct blk_mq_hw_ctx *hctx;
121	unsigned int i;
122
123	queue_for_each_hw_ctx(q, hctx, i)
124	if (blk_mq_hw_queue_mapped(hctx))
125	blk_mq_tag_wakeup_all(hctx->tags, true);
126
127	/*
128	* If we are called because the queue has now been marked as
129	* dying, we need to ensure that processes currently waiting on
130	* the queue are notified as well.
131	*/
132	wake_up_all(&q->mq_freeze_wq);
133	}
134
135	bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
136	{
137	return blk_mq_has_free_tags(hctx->tags);
138	}
139	EXPORT_SYMBOL(blk_mq_can_queue);
140
141	static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
142	struct request *rq, int op,
143	unsigned int op_flags)
144	{
145	if (blk_queue_io_stat(q))
146	op_flags |= REQ_IO_STAT;
147
148	INIT_LIST_HEAD(&rq->queuelist);
149	/* csd/requeue_work/fifo_time is initialized before use */
150	rq->q = q;
151	rq->mq_ctx = ctx;
152	req_set_op_attrs(rq, op, op_flags);
153	/* do not touch atomic flags, it needs atomic ops against the timer */
154	rq->cpu = -1;
155	INIT_HLIST_NODE(&rq->hash);
156	RB_CLEAR_NODE(&rq->rb_node);
157	rq->rq_disk = NULL;
158	rq->part = NULL;
159	rq->start_time = jiffies;
160	#ifdef CONFIG_BLK_CGROUP
161	rq->rl = NULL;
162	set_start_time_ns(rq);
163	rq->io_start_time_ns = 0;
164	#endif
165	rq->nr_phys_segments = 0;
166	#if defined(CONFIG_BLK_DEV_INTEGRITY)
167	rq->nr_integrity_segments = 0;
168	#endif
169	rq->special = NULL;
170	/* tag was already set */
171	rq->errors = 0;
172
173	rq->cmd = rq->__cmd;
174
175	rq->extra_len = 0;
176	rq->sense_len = 0;
177	rq->resid_len = 0;
178	rq->sense = NULL;
179
180	INIT_LIST_HEAD(&rq->timeout_list);
181	rq->timeout = 0;
182
183	rq->end_io = NULL;
184	rq->end_io_data = NULL;
185	rq->next_rq = NULL;
186
187	ctx->rq_dispatched[rw_is_sync(op, op_flags)]++;
188	}
189
190	static struct request *
191	__blk_mq_alloc_request(struct blk_mq_alloc_data *data, int op, int op_flags)
192	{
193	struct request *rq;
194	unsigned int tag;
195
196	tag = blk_mq_get_tag(data);
197	if (tag != BLK_MQ_TAG_FAIL) {
198	rq = data->hctx->tags->rqs[tag];
199
200	if (blk_mq_tag_busy(data->hctx)) {
201	rq->cmd_flags = REQ_MQ_INFLIGHT;
202	atomic_inc(&data->hctx->nr_active);
203	}
204
205	rq->tag = tag;
206	blk_mq_rq_ctx_init(data->q, data->ctx, rq, op, op_flags);
207	return rq;
208	}
209
210	return NULL;
211	}
212
213	struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
214	unsigned int flags)
215	{
216	struct blk_mq_ctx *ctx;
217	struct blk_mq_hw_ctx *hctx;
218	struct request *rq;
219	struct blk_mq_alloc_data alloc_data;
220	int ret;
221
222	ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
223	if (ret)
224	return ERR_PTR(ret);
225
226	ctx = blk_mq_get_ctx(q);
227	hctx = blk_mq_map_queue(q, ctx->cpu);
228	blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
229	rq = __blk_mq_alloc_request(&alloc_data, rw, 0);
230	blk_mq_put_ctx(ctx);
231
232	if (!rq) {
233	blk_queue_exit(q);
234	return ERR_PTR(-EWOULDBLOCK);
235	}
236
237	rq->__data_len = 0;
238	rq->__sector = (sector_t) -1;
239	rq->bio = rq->biotail = NULL;
240	return rq;
241	}
242	EXPORT_SYMBOL(blk_mq_alloc_request);
243
244	struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
245	unsigned int flags, unsigned int hctx_idx)
246	{
247	struct blk_mq_hw_ctx *hctx;
248	struct blk_mq_ctx *ctx;
249	struct request *rq;
250	struct blk_mq_alloc_data alloc_data;
251	int ret;
252
253	/*
254	* If the tag allocator sleeps we could get an allocation for a
255	* different hardware context. No need to complicate the low level
256	* allocator for this for the rare use case of a command tied to
257	* a specific queue.
258	*/
259	if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
260	return ERR_PTR(-EINVAL);
261
262	if (hctx_idx >= q->nr_hw_queues)
263	return ERR_PTR(-EIO);
264
265	ret = blk_queue_enter(q, true);
266	if (ret)
267	return ERR_PTR(ret);
268
269	/*
270	* Check if the hardware context is actually mapped to anything.
271	* If not tell the caller that it should skip this queue.
272	*/
273	hctx = q->queue_hw_ctx[hctx_idx];
274	if (!blk_mq_hw_queue_mapped(hctx)) {
275	ret = -EXDEV;
276	goto out_queue_exit;
277	}
278	ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask));
279
280	blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
281	rq = __blk_mq_alloc_request(&alloc_data, rw, 0);
282	if (!rq) {
283	ret = -EWOULDBLOCK;
284	goto out_queue_exit;
285	}
286
287	return rq;
288
289	out_queue_exit:
290	blk_queue_exit(q);
291	return ERR_PTR(ret);
292	}
293	EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
294
295	static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
296	struct blk_mq_ctx *ctx, struct request *rq)
297	{
298	const int tag = rq->tag;
299	struct request_queue *q = rq->q;
300
301	if (rq->cmd_flags & REQ_MQ_INFLIGHT)
302	atomic_dec(&hctx->nr_active);
303	rq->cmd_flags = 0;
304
305	clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
306	blk_mq_put_tag(hctx, ctx, tag);
307	blk_queue_exit(q);
308	}
309
310	void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
311	{
312	struct blk_mq_ctx *ctx = rq->mq_ctx;
313
314	ctx->rq_completed[rq_is_sync(rq)]++;
315	__blk_mq_free_request(hctx, ctx, rq);
316
317	}
318	EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
319
320	void blk_mq_free_request(struct request *rq)
321	{
322	blk_mq_free_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq);
323	}
324	EXPORT_SYMBOL_GPL(blk_mq_free_request);
325
326	inline void __blk_mq_end_request(struct request *rq, int error)
327	{
328	blk_account_io_done(rq);
329
330	if (rq->end_io) {
331	rq->end_io(rq, error);
332	} else {
333	if (unlikely(blk_bidi_rq(rq)))
334	blk_mq_free_request(rq->next_rq);
335	blk_mq_free_request(rq);
336	}
337	}
338	EXPORT_SYMBOL(__blk_mq_end_request);
339
340	void blk_mq_end_request(struct request *rq, int error)
341	{
342	if (blk_update_request(rq, error, blk_rq_bytes(rq)))
343	BUG();
344	__blk_mq_end_request(rq, error);
345	}
346	EXPORT_SYMBOL(blk_mq_end_request);
347
348	static void __blk_mq_complete_request_remote(void *data)
349	{
350	struct request *rq = data;
351
352	rq->q->softirq_done_fn(rq);
353	}
354
355	static void blk_mq_ipi_complete_request(struct request *rq)
356	{
357	struct blk_mq_ctx *ctx = rq->mq_ctx;
358	bool shared = false;
359	int cpu;
360
361	if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
362	rq->q->softirq_done_fn(rq);
363	return;
364	}
365
366	cpu = get_cpu();
367	if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
368	shared = cpus_share_cache(cpu, ctx->cpu);
369
370	if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
371	rq->csd.func = __blk_mq_complete_request_remote;
372	rq->csd.info = rq;
373	rq->csd.flags = 0;
374	smp_call_function_single_async(ctx->cpu, &rq->csd);
375	} else {
376	rq->q->softirq_done_fn(rq);
377	}
378	put_cpu();
379	}
380
381	static void __blk_mq_complete_request(struct request *rq)
382	{
383	struct request_queue *q = rq->q;
384
385	if (!q->softirq_done_fn)
386	blk_mq_end_request(rq, rq->errors);
387	else
388	blk_mq_ipi_complete_request(rq);
389	}
390
391	/**
392	* blk_mq_complete_request - end I/O on a request
393	* @rq: the request being processed
394	*
395	* Description:
396	* Ends all I/O on a request. It does not handle partial completions.
397	* The actual completion happens out-of-order, through a IPI handler.
398	**/
399	void blk_mq_complete_request(struct request *rq, int error)
400	{
401	struct request_queue *q = rq->q;
402
403	if (unlikely(blk_should_fake_timeout(q)))
404	return;
405	if (!blk_mark_rq_complete(rq)) {
406	rq->errors = error;
407	__blk_mq_complete_request(rq);
408	}
409	}
410	EXPORT_SYMBOL(blk_mq_complete_request);
411
412	int blk_mq_request_started(struct request *rq)
413	{
414	return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
415	}
416	EXPORT_SYMBOL_GPL(blk_mq_request_started);
417
418	void blk_mq_start_request(struct request *rq)
419	{
420	struct request_queue *q = rq->q;
421
422	trace_block_rq_issue(q, rq);
423
424	rq->resid_len = blk_rq_bytes(rq);
425	if (unlikely(blk_bidi_rq(rq)))
426	rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
427
428	blk_add_timer(rq);
429
430	/*
431	* Ensure that ->deadline is visible before set the started
432	* flag and clear the completed flag.
433	*/
434	smp_mb__before_atomic();
435
436	/*
437	* Mark us as started and clear complete. Complete might have been
438	* set if requeue raced with timeout, which then marked it as
439	* complete. So be sure to clear complete again when we start
440	* the request, otherwise we'll ignore the completion event.
441	*/
442	if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
443	set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
444	if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
445	clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
446
447	if (q->dma_drain_size && blk_rq_bytes(rq)) {
448	/*
449	* Make sure space for the drain appears. We know we can do
450	* this because max_hw_segments has been adjusted to be one
451	* fewer than the device can handle.
452	*/
453	rq->nr_phys_segments++;
454	}
455	}
456	EXPORT_SYMBOL(blk_mq_start_request);
457
458	static void __blk_mq_requeue_request(struct request *rq)
459	{
460	struct request_queue *q = rq->q;
461
462	trace_block_rq_requeue(q, rq);
463
464	if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
465	if (q->dma_drain_size && blk_rq_bytes(rq))
466	rq->nr_phys_segments--;
467	}
468	}
469
470	void blk_mq_requeue_request(struct request *rq)
471	{
472	__blk_mq_requeue_request(rq);
473
474	BUG_ON(blk_queued_rq(rq));
475	blk_mq_add_to_requeue_list(rq, true);
476	}
477	EXPORT_SYMBOL(blk_mq_requeue_request);
478
479	static void blk_mq_requeue_work(struct work_struct *work)
480	{
481	struct request_queue *q =
482	container_of(work, struct request_queue, requeue_work.work);
483	LIST_HEAD(rq_list);
484	struct request *rq, *next;
485	unsigned long flags;
486
487	spin_lock_irqsave(&q->requeue_lock, flags);
488	list_splice_init(&q->requeue_list, &rq_list);
489	spin_unlock_irqrestore(&q->requeue_lock, flags);
490
491	list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
492	if (!(rq->cmd_flags & REQ_SOFTBARRIER))
493	continue;
494
495	rq->cmd_flags &= ~REQ_SOFTBARRIER;
496	list_del_init(&rq->queuelist);
497	blk_mq_insert_request(rq, true, false, false);
498	}
499
500	while (!list_empty(&rq_list)) {
501	rq = list_entry(rq_list.next, struct request, queuelist);
502	list_del_init(&rq->queuelist);
503	blk_mq_insert_request(rq, false, false, false);
504	}
505
506	/*
507	* Use the start variant of queue running here, so that running
508	* the requeue work will kick stopped queues.
509	*/
510	blk_mq_start_hw_queues(q);
511	}
512
513	void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
514	{
515	struct request_queue *q = rq->q;
516	unsigned long flags;
517
518	/*
519	* We abuse this flag that is otherwise used by the I/O scheduler to
520	* request head insertation from the workqueue.
521	*/
522	BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
523
524	spin_lock_irqsave(&q->requeue_lock, flags);
525	if (at_head) {
526	rq->cmd_flags |= REQ_SOFTBARRIER;
527	list_add(&rq->queuelist, &q->requeue_list);
528	} else {
529	list_add_tail(&rq->queuelist, &q->requeue_list);
530	}
531	spin_unlock_irqrestore(&q->requeue_lock, flags);
532	}
533	EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
534
535	void blk_mq_cancel_requeue_work(struct request_queue *q)
536	{
537	cancel_delayed_work_sync(&q->requeue_work);
538	}
539	EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
540
541	void blk_mq_kick_requeue_list(struct request_queue *q)
542	{
543	kblockd_schedule_delayed_work(&q->requeue_work, 0);
544	}
545	EXPORT_SYMBOL(blk_mq_kick_requeue_list);
546
547	void blk_mq_delay_kick_requeue_list(struct request_queue *q,
548	unsigned long msecs)
549	{
550	kblockd_schedule_delayed_work(&q->requeue_work,
551	msecs_to_jiffies(msecs));
552	}
553	EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
554
555	void blk_mq_abort_requeue_list(struct request_queue *q)
556	{
557	unsigned long flags;
558	LIST_HEAD(rq_list);
559
560	spin_lock_irqsave(&q->requeue_lock, flags);
561	list_splice_init(&q->requeue_list, &rq_list);
562	spin_unlock_irqrestore(&q->requeue_lock, flags);
563
564	while (!list_empty(&rq_list)) {
565	struct request *rq;
566
567	rq = list_first_entry(&rq_list, struct request, queuelist);
568	list_del_init(&rq->queuelist);
569	rq->errors = -EIO;
570	blk_mq_end_request(rq, rq->errors);
571	}
572	}
573	EXPORT_SYMBOL(blk_mq_abort_requeue_list);
574
575	struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
576	{
577	if (tag < tags->nr_tags) {
578	prefetch(tags->rqs[tag]);
579	return tags->rqs[tag];
580	}
581
582	return NULL;
583	}
584	EXPORT_SYMBOL(blk_mq_tag_to_rq);
585
586	struct blk_mq_timeout_data {
587	unsigned long next;
588	unsigned int next_set;
589	};
590
591	void blk_mq_rq_timed_out(struct request *req, bool reserved)
592	{
593	struct blk_mq_ops *ops = req->q->mq_ops;
594	enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
595
596	/*
597	* We know that complete is set at this point. If STARTED isn't set
598	* anymore, then the request isn't active and the "timeout" should
599	* just be ignored. This can happen due to the bitflag ordering.
600	* Timeout first checks if STARTED is set, and if it is, assumes
601	* the request is active. But if we race with completion, then
602	* we both flags will get cleared. So check here again, and ignore
603	* a timeout event with a request that isn't active.
604	*/
605	if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
606	return;
607
608	if (ops->timeout)
609	ret = ops->timeout(req, reserved);
610
611	switch (ret) {
612	case BLK_EH_HANDLED:
613	__blk_mq_complete_request(req);
614	break;
615	case BLK_EH_RESET_TIMER:
616	blk_add_timer(req);
617	blk_clear_rq_complete(req);
618	break;
619	case BLK_EH_NOT_HANDLED:
620	break;
621	default:
622	printk(KERN_ERR "block: bad eh return: %d\n", ret);
623	break;
624	}
625	}
626
627	static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
628	struct request *rq, void *priv, bool reserved)
629	{
630	struct blk_mq_timeout_data *data = priv;
631
632	if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
633	return;
634
635	if (time_after_eq(jiffies, rq->deadline)) {
636	if (!blk_mark_rq_complete(rq))
637	blk_mq_rq_timed_out(rq, reserved);
638	} else if (!data->next_set || time_after(data->next, rq->deadline)) {
639	data->next = rq->deadline;
640	data->next_set = 1;
641	}
642	}
643
644	static void blk_mq_timeout_work(struct work_struct *work)
645	{
646	struct request_queue *q =
647	container_of(work, struct request_queue, timeout_work);
648	struct blk_mq_timeout_data data = {
649	.next = 0,
650	.next_set = 0,
651	};
652	int i;
653
654	/* A deadlock might occur if a request is stuck requiring a
655	* timeout at the same time a queue freeze is waiting
656	* completion, since the timeout code would not be able to
657	* acquire the queue reference here.
658	*
659	* That's why we don't use blk_queue_enter here; instead, we use
660	* percpu_ref_tryget directly, because we need to be able to
661	* obtain a reference even in the short window between the queue
662	* starting to freeze, by dropping the first reference in
663	* blk_mq_freeze_queue_start, and the moment the last request is
664	* consumed, marked by the instant q_usage_counter reaches
665	* zero.
666	*/
667	if (!percpu_ref_tryget(&q->q_usage_counter))
668	return;
669
670	blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
671
672	if (data.next_set) {
673	data.next = blk_rq_timeout(round_jiffies_up(data.next));
674	mod_timer(&q->timeout, data.next);
675	} else {
676	struct blk_mq_hw_ctx *hctx;
677
678	queue_for_each_hw_ctx(q, hctx, i) {
679	/* the hctx may be unmapped, so check it here */
680	if (blk_mq_hw_queue_mapped(hctx))
681	blk_mq_tag_idle(hctx);
682	}
683	}
684	blk_queue_exit(q);
685	}
686
687	/*
688	* Reverse check our software queue for entries that we could potentially
689	* merge with. Currently includes a hand-wavy stop count of 8, to not spend
690	* too much time checking for merges.
691	*/
692	static bool blk_mq_attempt_merge(struct request_queue *q,
693	struct blk_mq_ctx *ctx, struct bio *bio)
694	{
695	struct request *rq;
696	int checked = 8;
697
698	list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
699	int el_ret;
700
701	if (!checked--)
702	break;
703
704	if (!blk_rq_merge_ok(rq, bio))
705	continue;
706
707	el_ret = blk_try_merge(rq, bio);
708	if (el_ret == ELEVATOR_BACK_MERGE) {
709	if (bio_attempt_back_merge(q, rq, bio)) {
710	ctx->rq_merged++;
711	return true;
712	}
713	break;
714	} else if (el_ret == ELEVATOR_FRONT_MERGE) {
715	if (bio_attempt_front_merge(q, rq, bio)) {
716	ctx->rq_merged++;
717	return true;
718	}
719	break;
720	}
721	}
722
723	return false;
724	}
725
726	struct flush_busy_ctx_data {
727	struct blk_mq_hw_ctx *hctx;
728	struct list_head *list;
729	};
730
731	static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
732	{
733	struct flush_busy_ctx_data *flush_data = data;
734	struct blk_mq_hw_ctx *hctx = flush_data->hctx;
735	struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
736
737	sbitmap_clear_bit(sb, bitnr);
738	spin_lock(&ctx->lock);
739	list_splice_tail_init(&ctx->rq_list, flush_data->list);
740	spin_unlock(&ctx->lock);
741	return true;
742	}
743
744	/*
745	* Process software queues that have been marked busy, splicing them
746	* to the for-dispatch
747	*/
748	static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
749	{
750	struct flush_busy_ctx_data data = {
751	.hctx = hctx,
752	.list = list,
753	};
754
755	sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
756	}
757
758	static inline unsigned int queued_to_index(unsigned int queued)
759	{
760	if (!queued)
761	return 0;
762
763	return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
764	}
765
766	/*
767	* Run this hardware queue, pulling any software queues mapped to it in.
768	* Note that this function currently has various problems around ordering
769	* of IO. In particular, we'd like FIFO behaviour on handling existing
770	* items on the hctx->dispatch list. Ignore that for now.
771	*/
772	static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
773	{
774	struct request_queue *q = hctx->queue;
775	struct request *rq;
776	LIST_HEAD(rq_list);
777	LIST_HEAD(driver_list);
778	struct list_head *dptr;
779	int queued;
780
781	if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
782	return;
783
784	WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
785	cpu_online(hctx->next_cpu));
786
787	hctx->run++;
788
789	/*
790	* Touch any software queue that has pending entries.
791	*/
792	flush_busy_ctxs(hctx, &rq_list);
793
794	/*
795	* If we have previous entries on our dispatch list, grab them
796	* and stuff them at the front for more fair dispatch.
797	*/
798	if (!list_empty_careful(&hctx->dispatch)) {
799	spin_lock(&hctx->lock);
800	if (!list_empty(&hctx->dispatch))
801	list_splice_init(&hctx->dispatch, &rq_list);
802	spin_unlock(&hctx->lock);
803	}
804
805	/*
806	* Start off with dptr being NULL, so we start the first request
807	* immediately, even if we have more pending.
808	*/
809	dptr = NULL;
810
811	/*
812	* Now process all the entries, sending them to the driver.
813	*/
814	queued = 0;
815	while (!list_empty(&rq_list)) {
816	struct blk_mq_queue_data bd;
817	int ret;
818
819	rq = list_first_entry(&rq_list, struct request, queuelist);
820	list_del_init(&rq->queuelist);
821
822	bd.rq = rq;
823	bd.list = dptr;
824	bd.last = list_empty(&rq_list);
825
826	ret = q->mq_ops->queue_rq(hctx, &bd);
827	switch (ret) {
828	case BLK_MQ_RQ_QUEUE_OK:
829	queued++;
830	break;
831	case BLK_MQ_RQ_QUEUE_BUSY:
832	list_add(&rq->queuelist, &rq_list);
833	__blk_mq_requeue_request(rq);
834	break;
835	default:
836	pr_err("blk-mq: bad return on queue: %d\n", ret);
837	case BLK_MQ_RQ_QUEUE_ERROR:
838	rq->errors = -EIO;
839	blk_mq_end_request(rq, rq->errors);
840	break;
841	}
842
843	if (ret == BLK_MQ_RQ_QUEUE_BUSY)
844	break;
845
846	/*
847	* We've done the first request. If we have more than 1
848	* left in the list, set dptr to defer issue.
849	*/
850	if (!dptr && rq_list.next != rq_list.prev)
851	dptr = &driver_list;
852	}
853
854	hctx->dispatched[queued_to_index(queued)]++;
855
856	/*
857	* Any items that need requeuing? Stuff them into hctx->dispatch,
858	* that is where we will continue on next queue run.
859	*/
860	if (!list_empty(&rq_list)) {
861	spin_lock(&hctx->lock);
862	list_splice(&rq_list, &hctx->dispatch);
863	spin_unlock(&hctx->lock);
864	/*
865	* the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
866	* it's possible the queue is stopped and restarted again
867	* before this. Queue restart will dispatch requests. And since
868	* requests in rq_list aren't added into hctx->dispatch yet,
869	* the requests in rq_list might get lost.
870	*
871	* blk_mq_run_hw_queue() already checks the STOPPED bit
872	**/
873	blk_mq_run_hw_queue(hctx, true);
874	}
875	}
876
877	/*
878	* It'd be great if the workqueue API had a way to pass
879	* in a mask and had some smarts for more clever placement.
880	* For now we just round-robin here, switching for every
881	* BLK_MQ_CPU_WORK_BATCH queued items.
882	*/
883	static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
884	{
885	if (hctx->queue->nr_hw_queues == 1)
886	return WORK_CPU_UNBOUND;
887
888	if (--hctx->next_cpu_batch <= 0) {
889	int next_cpu;
890
891	next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
892	if (next_cpu >= nr_cpu_ids)
893	next_cpu = cpumask_first(hctx->cpumask);
894
895	hctx->next_cpu = next_cpu;
896	hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
897	}
898
899	return hctx->next_cpu;
900	}
901
902	void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
903	{
904	if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
905	!blk_mq_hw_queue_mapped(hctx)))
906	return;
907
908	if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
909	int cpu = get_cpu();
910	if (cpumask_test_cpu(cpu, hctx->cpumask)) {
911	__blk_mq_run_hw_queue(hctx);
912	put_cpu();
913	return;
914	}
915
916	put_cpu();
917	}
918
919	kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work);
920	}
921
922	void blk_mq_run_hw_queues(struct request_queue *q, bool async)
923	{
924	struct blk_mq_hw_ctx *hctx;
925	int i;
926
927	queue_for_each_hw_ctx(q, hctx, i) {
928	if ((!blk_mq_hctx_has_pending(hctx) &&
929	list_empty_careful(&hctx->dispatch)) ||
930	test_bit(BLK_MQ_S_STOPPED, &hctx->state))
931	continue;
932
933	blk_mq_run_hw_queue(hctx, async);
934	}
935	}
936	EXPORT_SYMBOL(blk_mq_run_hw_queues);
937
938	void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
939	{
940	cancel_work(&hctx->run_work);
941	cancel_delayed_work(&hctx->delay_work);
942	set_bit(BLK_MQ_S_STOPPED, &hctx->state);
943	}
944	EXPORT_SYMBOL(blk_mq_stop_hw_queue);
945
946	void blk_mq_stop_hw_queues(struct request_queue *q)
947	{
948	struct blk_mq_hw_ctx *hctx;
949	int i;
950
951	queue_for_each_hw_ctx(q, hctx, i)
952	blk_mq_stop_hw_queue(hctx);
953	}
954	EXPORT_SYMBOL(blk_mq_stop_hw_queues);
955
956	void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
957	{
958	clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
959
960	blk_mq_run_hw_queue(hctx, false);
961	}
962	EXPORT_SYMBOL(blk_mq_start_hw_queue);
963
964	void blk_mq_start_hw_queues(struct request_queue *q)
965	{
966	struct blk_mq_hw_ctx *hctx;
967	int i;
968
969	queue_for_each_hw_ctx(q, hctx, i)
970	blk_mq_start_hw_queue(hctx);
971	}
972	EXPORT_SYMBOL(blk_mq_start_hw_queues);
973
974	void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
975	{
976	struct blk_mq_hw_ctx *hctx;
977	int i;
978
979	queue_for_each_hw_ctx(q, hctx, i) {
980	if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
981	continue;
982
983	clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
984	blk_mq_run_hw_queue(hctx, async);
985	}
986	}
987	EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
988
989	static void blk_mq_run_work_fn(struct work_struct *work)
990	{
991	struct blk_mq_hw_ctx *hctx;
992
993	hctx = container_of(work, struct blk_mq_hw_ctx, run_work);
994
995	__blk_mq_run_hw_queue(hctx);
996	}
997
998	static void blk_mq_delay_work_fn(struct work_struct *work)
999	{
1000	struct blk_mq_hw_ctx *hctx;
1001
1002	hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1003
1004	if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1005	__blk_mq_run_hw_queue(hctx);
1006	}
1007
1008	void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1009	{
1010	if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1011	return;
1012
1013	kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1014	&hctx->delay_work, msecs_to_jiffies(msecs));
1015	}
1016	EXPORT_SYMBOL(blk_mq_delay_queue);
1017
1018	static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1019	struct request *rq,
1020	bool at_head)
1021	{
1022	struct blk_mq_ctx *ctx = rq->mq_ctx;
1023
1024	trace_block_rq_insert(hctx->queue, rq);
1025
1026	if (at_head)
1027	list_add(&rq->queuelist, &ctx->rq_list);
1028	else
1029	list_add_tail(&rq->queuelist, &ctx->rq_list);
1030	}
1031
1032	static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
1033	struct request *rq, bool at_head)
1034	{
1035	struct blk_mq_ctx *ctx = rq->mq_ctx;
1036
1037	__blk_mq_insert_req_list(hctx, rq, at_head);
1038	blk_mq_hctx_mark_pending(hctx, ctx);
1039	}
1040
1041	void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1042	bool async)
1043	{
1044	struct blk_mq_ctx *ctx = rq->mq_ctx;
1045	struct request_queue *q = rq->q;
1046	struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
1047
1048	spin_lock(&ctx->lock);
1049	__blk_mq_insert_request(hctx, rq, at_head);
1050	spin_unlock(&ctx->lock);
1051
1052	if (run_queue)
1053	blk_mq_run_hw_queue(hctx, async);
1054	}
1055
1056	static void blk_mq_insert_requests(struct request_queue *q,
1057	struct blk_mq_ctx *ctx,
1058	struct list_head *list,
1059	int depth,
1060	bool from_schedule)
1061
1062	{
1063	struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
1064
1065	trace_block_unplug(q, depth, !from_schedule);
1066
1067	/*
1068	* preemption doesn't flush plug list, so it's possible ctx->cpu is
1069	* offline now
1070	*/
1071	spin_lock(&ctx->lock);
1072	while (!list_empty(list)) {
1073	struct request *rq;
1074
1075	rq = list_first_entry(list, struct request, queuelist);
1076	BUG_ON(rq->mq_ctx != ctx);
1077	list_del_init(&rq->queuelist);
1078	__blk_mq_insert_req_list(hctx, rq, false);
1079	}
1080	blk_mq_hctx_mark_pending(hctx, ctx);
1081	spin_unlock(&ctx->lock);
1082
1083	blk_mq_run_hw_queue(hctx, from_schedule);
1084	}
1085
1086	static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1087	{
1088	struct request *rqa = container_of(a, struct request, queuelist);
1089	struct request *rqb = container_of(b, struct request, queuelist);
1090
1091	return !(rqa->mq_ctx < rqb->mq_ctx ||
1092	(rqa->mq_ctx == rqb->mq_ctx &&
1093	blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1094	}
1095
1096	void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1097	{
1098	struct blk_mq_ctx *this_ctx;
1099	struct request_queue *this_q;
1100	struct request *rq;
1101	LIST_HEAD(list);
1102	LIST_HEAD(ctx_list);
1103	unsigned int depth;
1104
1105	list_splice_init(&plug->mq_list, &list);
1106
1107	list_sort(NULL, &list, plug_ctx_cmp);
1108
1109	this_q = NULL;
1110	this_ctx = NULL;
1111	depth = 0;
1112
1113	while (!list_empty(&list)) {
1114	rq = list_entry_rq(list.next);
1115	list_del_init(&rq->queuelist);
1116	BUG_ON(!rq->q);
1117	if (rq->mq_ctx != this_ctx) {
1118	if (this_ctx) {
1119	blk_mq_insert_requests(this_q, this_ctx,
1120	&ctx_list, depth,
1121	from_schedule);
1122	}
1123
1124	this_ctx = rq->mq_ctx;
1125	this_q = rq->q;
1126	depth = 0;
1127	}
1128
1129	depth++;
1130	list_add_tail(&rq->queuelist, &ctx_list);
1131	}
1132
1133	/*
1134	* If 'this_ctx' is set, we know we have entries to complete
1135	* on 'ctx_list'. Do those.
1136	*/
1137	if (this_ctx) {
1138	blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1139	from_schedule);
1140	}
1141	}
1142
1143	static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1144	{
1145	init_request_from_bio(rq, bio);
1146
1147	blk_account_io_start(rq, 1);
1148	}
1149
1150	static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1151	{
1152	return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1153	!blk_queue_nomerges(hctx->queue);
1154	}
1155
1156	static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1157	struct blk_mq_ctx *ctx,
1158	struct request *rq, struct bio *bio)
1159	{
1160	if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1161	blk_mq_bio_to_request(rq, bio);
1162	spin_lock(&ctx->lock);
1163	insert_rq:
1164	__blk_mq_insert_request(hctx, rq, false);
1165	spin_unlock(&ctx->lock);
1166	return false;
1167	} else {
1168	struct request_queue *q = hctx->queue;
1169
1170	spin_lock(&ctx->lock);
1171	if (!blk_mq_attempt_merge(q, ctx, bio)) {
1172	blk_mq_bio_to_request(rq, bio);
1173	goto insert_rq;
1174	}
1175
1176	spin_unlock(&ctx->lock);
1177	__blk_mq_free_request(hctx, ctx, rq);
1178	return true;
1179	}
1180	}
1181
1182	struct blk_map_ctx {
1183	struct blk_mq_hw_ctx *hctx;
1184	struct blk_mq_ctx *ctx;
1185	};
1186
1187	static struct request *blk_mq_map_request(struct request_queue *q,
1188	struct bio *bio,
1189	struct blk_map_ctx *data)
1190	{
1191	struct blk_mq_hw_ctx *hctx;
1192	struct blk_mq_ctx *ctx;
1193	struct request *rq;
1194	int op = bio_data_dir(bio);
1195	int op_flags = 0;
1196	struct blk_mq_alloc_data alloc_data;
1197
1198	blk_queue_enter_live(q);
1199	ctx = blk_mq_get_ctx(q);
1200	hctx = blk_mq_map_queue(q, ctx->cpu);
1201
1202	if (rw_is_sync(bio_op(bio), bio->bi_opf))
1203	op_flags |= REQ_SYNC;
1204
1205	trace_block_getrq(q, bio, op);
1206	blk_mq_set_alloc_data(&alloc_data, q, 0, ctx, hctx);
1207	rq = __blk_mq_alloc_request(&alloc_data, op, op_flags);
1208
1209	data->hctx = alloc_data.hctx;
1210	data->ctx = alloc_data.ctx;
1211	data->hctx->queued++;
1212	return rq;
1213	}
1214
1215	static int blk_mq_direct_issue_request(struct request *rq, blk_qc_t *cookie)
1216	{
1217	int ret;
1218	struct request_queue *q = rq->q;
1219	struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu);
1220	struct blk_mq_queue_data bd = {
1221	.rq = rq,
1222	.list = NULL,
1223	.last = 1
1224	};
1225	blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num);
1226
1227	/*
1228	* For OK queue, we are done. For error, kill it. Any other
1229	* error (busy), just add it to our list as we previously
1230	* would have done
1231	*/
1232	ret = q->mq_ops->queue_rq(hctx, &bd);
1233	if (ret == BLK_MQ_RQ_QUEUE_OK) {
1234	*cookie = new_cookie;
1235	return 0;
1236	}
1237
1238	__blk_mq_requeue_request(rq);
1239
1240	if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1241	*cookie = BLK_QC_T_NONE;
1242	rq->errors = -EIO;
1243	blk_mq_end_request(rq, rq->errors);
1244	return 0;
1245	}
1246
1247	return -1;
1248	}
1249
1250	/*
1251	* Multiple hardware queue variant. This will not use per-process plugs,
1252	* but will attempt to bypass the hctx queueing if we can go straight to
1253	* hardware for SYNC IO.
1254	*/
1255	static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1256	{
1257	const int is_sync = rw_is_sync(bio_op(bio), bio->bi_opf);
1258	const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA);
1259	struct blk_map_ctx data;
1260	struct request *rq;
1261	unsigned int request_count = 0;
1262	struct blk_plug *plug;
1263	struct request *same_queue_rq = NULL;
1264	blk_qc_t cookie;
1265
1266	blk_queue_bounce(q, &bio);
1267
1268	blk_queue_split(q, &bio, q->bio_split);
1269
1270	if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1271	bio_io_error(bio);
1272	return BLK_QC_T_NONE;
1273	}
1274
1275	if (!is_flush_fua && !blk_queue_nomerges(q) &&
1276	blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1277	return BLK_QC_T_NONE;
1278
1279	rq = blk_mq_map_request(q, bio, &data);
1280	if (unlikely(!rq))
1281	return BLK_QC_T_NONE;
1282
1283	cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1284
1285	if (unlikely(is_flush_fua)) {
1286	blk_mq_bio_to_request(rq, bio);
1287	blk_insert_flush(rq);
1288	goto run_queue;
1289	}
1290
1291	plug = current->plug;
1292	/*
1293	* If the driver supports defer issued based on 'last', then
1294	* queue it up like normal since we can potentially save some
1295	* CPU this way.
1296	*/
1297	if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1298	!(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1299	struct request *old_rq = NULL;
1300
1301	blk_mq_bio_to_request(rq, bio);
1302
1303	/*
1304	* We do limited pluging. If the bio can be merged, do that.
1305	* Otherwise the existing request in the plug list will be
1306	* issued. So the plug list will have one request at most
1307	*/
1308	if (plug) {
1309	/*
1310	* The plug list might get flushed before this. If that
1311	* happens, same_queue_rq is invalid and plug list is
1312	* empty
1313	*/
1314	if (same_queue_rq && !list_empty(&plug->mq_list)) {
1315	old_rq = same_queue_rq;
1316	list_del_init(&old_rq->queuelist);
1317	}
1318	list_add_tail(&rq->queuelist, &plug->mq_list);
1319	} else /* is_sync */
1320	old_rq = rq;
1321	blk_mq_put_ctx(data.ctx);
1322	if (!old_rq)
1323	goto done;
1324	if (test_bit(BLK_MQ_S_STOPPED, &data.hctx->state) ||
1325	blk_mq_direct_issue_request(old_rq, &cookie) != 0)
1326	blk_mq_insert_request(old_rq, false, true, true);
1327	goto done;
1328	}
1329
1330	if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1331	/*
1332	* For a SYNC request, send it to the hardware immediately. For
1333	* an ASYNC request, just ensure that we run it later on. The
1334	* latter allows for merging opportunities and more efficient
1335	* dispatching.
1336	*/
1337	run_queue:
1338	blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1339	}
1340	blk_mq_put_ctx(data.ctx);
1341	done:
1342	return cookie;
1343	}
1344
1345	/*
1346	* Single hardware queue variant. This will attempt to use any per-process
1347	* plug for merging and IO deferral.
1348	*/
1349	static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
1350	{
1351	const int is_sync = rw_is_sync(bio_op(bio), bio->bi_opf);
1352	const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA);
1353	struct blk_plug *plug;
1354	unsigned int request_count = 0;
1355	struct blk_map_ctx data;
1356	struct request *rq;
1357	blk_qc_t cookie;
1358
1359	blk_queue_bounce(q, &bio);
1360
1361	if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1362	bio_io_error(bio);
1363	return BLK_QC_T_NONE;
1364	}
1365
1366	blk_queue_split(q, &bio, q->bio_split);
1367
1368	if (!is_flush_fua && !blk_queue_nomerges(q)) {
1369	if (blk_attempt_plug_merge(q, bio, &request_count, NULL))
1370	return BLK_QC_T_NONE;
1371	} else
1372	request_count = blk_plug_queued_count(q);
1373
1374	rq = blk_mq_map_request(q, bio, &data);
1375	if (unlikely(!rq))
1376	return BLK_QC_T_NONE;
1377
1378	cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1379
1380	if (unlikely(is_flush_fua)) {
1381	blk_mq_bio_to_request(rq, bio);
1382	blk_insert_flush(rq);
1383	goto run_queue;
1384	}
1385
1386	/*
1387	* A task plug currently exists. Since this is completely lockless,
1388	* utilize that to temporarily store requests until the task is
1389	* either done or scheduled away.
1390	*/
1391	plug = current->plug;
1392	if (plug) {
1393	blk_mq_bio_to_request(rq, bio);
1394	if (!request_count)
1395	trace_block_plug(q);
1396
1397	blk_mq_put_ctx(data.ctx);
1398
1399	if (request_count >= BLK_MAX_REQUEST_COUNT) {
1400	blk_flush_plug_list(plug, false);
1401	trace_block_plug(q);
1402	}
1403
1404	list_add_tail(&rq->queuelist, &plug->mq_list);
1405	return cookie;
1406	}
1407
1408	if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1409	/*
1410	* For a SYNC request, send it to the hardware immediately. For
1411	* an ASYNC request, just ensure that we run it later on. The
1412	* latter allows for merging opportunities and more efficient
1413	* dispatching.
1414	*/
1415	run_queue:
1416	blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1417	}
1418
1419	blk_mq_put_ctx(data.ctx);
1420	return cookie;
1421	}
1422
1423	static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1424	struct blk_mq_tags *tags, unsigned int hctx_idx)
1425	{
1426	struct page *page;
1427
1428	if (tags->rqs && set->ops->exit_request) {
1429	int i;
1430
1431	for (i = 0; i < tags->nr_tags; i++) {
1432	if (!tags->rqs[i])
1433	continue;
1434	set->ops->exit_request(set->driver_data, tags->rqs[i],
1435	hctx_idx, i);
1436	tags->rqs[i] = NULL;
1437	}
1438	}
1439
1440	while (!list_empty(&tags->page_list)) {
1441	page = list_first_entry(&tags->page_list, struct page, lru);
1442	list_del_init(&page->lru);
1443	/*
1444	* Remove kmemleak object previously allocated in
1445	* blk_mq_init_rq_map().
1446	*/
1447	kmemleak_free(page_address(page));
1448	__free_pages(page, page->private);
1449	}
1450
1451	kfree(tags->rqs);
1452
1453	blk_mq_free_tags(tags);
1454	}
1455
1456	static size_t order_to_size(unsigned int order)
1457	{
1458	return (size_t)PAGE_SIZE << order;
1459	}
1460
1461	static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1462	unsigned int hctx_idx)
1463	{
1464	struct blk_mq_tags *tags;
1465	unsigned int i, j, entries_per_page, max_order = 4;
1466	size_t rq_size, left;
1467
1468	tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1469	set->numa_node,
1470	BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1471	if (!tags)
1472	return NULL;
1473
1474	INIT_LIST_HEAD(&tags->page_list);
1475
1476	tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1477	GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1478	set->numa_node);
1479	if (!tags->rqs) {
1480	blk_mq_free_tags(tags);
1481	return NULL;
1482	}
1483
1484	/*
1485	* rq_size is the size of the request plus driver payload, rounded
1486	* to the cacheline size
1487	*/
1488	rq_size = round_up(sizeof(struct request) + set->cmd_size,
1489	cache_line_size());
1490	left = rq_size * set->queue_depth;
1491
1492	for (i = 0; i < set->queue_depth; ) {
1493	int this_order = max_order;
1494	struct page *page;
1495	int to_do;
1496	void *p;
1497
1498	while (this_order && left < order_to_size(this_order - 1))
1499	this_order--;
1500
1501	do {
1502	page = alloc_pages_node(set->numa_node,
1503	GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1504	this_order);
1505	if (page)
1506	break;
1507	if (!this_order--)
1508	break;
1509	if (order_to_size(this_order) < rq_size)
1510	break;
1511	} while (1);
1512
1513	if (!page)
1514	goto fail;
1515
1516	page->private = this_order;
1517	list_add_tail(&page->lru, &tags->page_list);
1518
1519	p = page_address(page);
1520	/*
1521	* Allow kmemleak to scan these pages as they contain pointers
1522	* to additional allocations like via ops->init_request().
1523	*/
1524	kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1525	entries_per_page = order_to_size(this_order) / rq_size;
1526	to_do = min(entries_per_page, set->queue_depth - i);
1527	left -= to_do * rq_size;
1528	for (j = 0; j < to_do; j++) {
1529	tags->rqs[i] = p;
1530	if (set->ops->init_request) {
1531	if (set->ops->init_request(set->driver_data,
1532	tags->rqs[i], hctx_idx, i,
1533	set->numa_node)) {
1534	tags->rqs[i] = NULL;
1535	goto fail;
1536	}
1537	}
1538
1539	p += rq_size;
1540	i++;
1541	}
1542	}
1543	return tags;
1544
1545	fail:
1546	blk_mq_free_rq_map(set, tags, hctx_idx);
1547	return NULL;
1548	}
1549
1550	/*
1551	* 'cpu' is going away. splice any existing rq_list entries from this
1552	* software queue to the hw queue dispatch list, and ensure that it
1553	* gets run.
1554	*/
1555	static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1556	{
1557	struct blk_mq_hw_ctx *hctx;
1558	struct blk_mq_ctx *ctx;
1559	LIST_HEAD(tmp);
1560
1561	hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1562	ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1563
1564	spin_lock(&ctx->lock);
1565	if (!list_empty(&ctx->rq_list)) {
1566	list_splice_init(&ctx->rq_list, &tmp);
1567	blk_mq_hctx_clear_pending(hctx, ctx);
1568	}
1569	spin_unlock(&ctx->lock);
1570
1571	if (list_empty(&tmp))
1572	return 0;
1573
1574	spin_lock(&hctx->lock);
1575	list_splice_tail_init(&tmp, &hctx->dispatch);
1576	spin_unlock(&hctx->lock);
1577
1578	blk_mq_run_hw_queue(hctx, true);
1579	return 0;
1580	}
1581
1582	static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1583	{
1584	cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1585	&hctx->cpuhp_dead);
1586	}
1587
1588	/* hctx->ctxs will be freed in queue's release handler */
1589	static void blk_mq_exit_hctx(struct request_queue *q,
1590	struct blk_mq_tag_set *set,
1591	struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1592	{
1593	unsigned flush_start_tag = set->queue_depth;
1594
1595	if (blk_mq_hw_queue_mapped(hctx))
1596	blk_mq_tag_idle(hctx);
1597
1598	if (set->ops->exit_request)
1599	set->ops->exit_request(set->driver_data,
1600	hctx->fq->flush_rq, hctx_idx,
1601	flush_start_tag + hctx_idx);
1602
1603	if (set->ops->exit_hctx)
1604	set->ops->exit_hctx(hctx, hctx_idx);
1605
1606	blk_mq_remove_cpuhp(hctx);
1607	blk_free_flush_queue(hctx->fq);
1608	sbitmap_free(&hctx->ctx_map);
1609	}
1610
1611	static void blk_mq_exit_hw_queues(struct request_queue *q,
1612	struct blk_mq_tag_set *set, int nr_queue)
1613	{
1614	struct blk_mq_hw_ctx *hctx;
1615	unsigned int i;
1616
1617	queue_for_each_hw_ctx(q, hctx, i) {
1618	if (i == nr_queue)
1619	break;
1620	blk_mq_exit_hctx(q, set, hctx, i);
1621	}
1622	}
1623
1624	static void blk_mq_free_hw_queues(struct request_queue *q,
1625	struct blk_mq_tag_set *set)
1626	{
1627	struct blk_mq_hw_ctx *hctx;
1628	unsigned int i;
1629
1630	queue_for_each_hw_ctx(q, hctx, i)
1631	free_cpumask_var(hctx->cpumask);
1632	}
1633
1634	static int blk_mq_init_hctx(struct request_queue *q,
1635	struct blk_mq_tag_set *set,
1636	struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1637	{
1638	int node;
1639	unsigned flush_start_tag = set->queue_depth;
1640
1641	node = hctx->numa_node;
1642	if (node == NUMA_NO_NODE)
1643	node = hctx->numa_node = set->numa_node;
1644
1645	INIT_WORK(&hctx->run_work, blk_mq_run_work_fn);
1646	INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1647	spin_lock_init(&hctx->lock);
1648	INIT_LIST_HEAD(&hctx->dispatch);
1649	hctx->queue = q;
1650	hctx->queue_num = hctx_idx;
1651	hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1652
1653	cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1654
1655	hctx->tags = set->tags[hctx_idx];
1656
1657	/*
1658	* Allocate space for all possible cpus to avoid allocation at
1659	* runtime
1660	*/
1661	hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1662	GFP_KERNEL, node);
1663	if (!hctx->ctxs)
1664	goto unregister_cpu_notifier;
1665
1666	if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1667	node))
1668	goto free_ctxs;
1669
1670	hctx->nr_ctx = 0;
1671
1672	if (set->ops->init_hctx &&
1673	set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1674	goto free_bitmap;
1675
1676	hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1677	if (!hctx->fq)
1678	goto exit_hctx;
1679
1680	if (set->ops->init_request &&
1681	set->ops->init_request(set->driver_data,
1682	hctx->fq->flush_rq, hctx_idx,
1683	flush_start_tag + hctx_idx, node))
1684	goto free_fq;
1685
1686	return 0;
1687
1688	free_fq:
1689	kfree(hctx->fq);
1690	exit_hctx:
1691	if (set->ops->exit_hctx)
1692	set->ops->exit_hctx(hctx, hctx_idx);
1693	free_bitmap:
1694	sbitmap_free(&hctx->ctx_map);
1695	free_ctxs:
1696	kfree(hctx->ctxs);
1697	unregister_cpu_notifier:
1698	blk_mq_remove_cpuhp(hctx);
1699	return -1;
1700	}
1701
1702	static void blk_mq_init_cpu_queues(struct request_queue *q,
1703	unsigned int nr_hw_queues)
1704	{
1705	unsigned int i;
1706
1707	for_each_possible_cpu(i) {
1708	struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1709	struct blk_mq_hw_ctx *hctx;
1710
1711	__ctx->cpu = i;
1712	spin_lock_init(&__ctx->lock);
1713	INIT_LIST_HEAD(&__ctx->rq_list);
1714	__ctx->queue = q;
1715
1716	/* If the cpu isn't online, the cpu is mapped to first hctx */
1717	if (!cpu_online(i))
1718	continue;
1719
1720	hctx = blk_mq_map_queue(q, i);
1721
1722	/*
1723	* Set local node, IFF we have more than one hw queue. If
1724	* not, we remain on the home node of the device
1725	*/
1726	if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1727	hctx->numa_node = local_memory_node(cpu_to_node(i));
1728	}
1729	}
1730
1731	static void blk_mq_map_swqueue(struct request_queue *q,
1732	const struct cpumask *online_mask)
1733	{
1734	unsigned int i;
1735	struct blk_mq_hw_ctx *hctx;
1736	struct blk_mq_ctx *ctx;
1737	struct blk_mq_tag_set *set = q->tag_set;
1738
1739	/*
1740	* Avoid others reading imcomplete hctx->cpumask through sysfs
1741	*/
1742	mutex_lock(&q->sysfs_lock);
1743
1744	queue_for_each_hw_ctx(q, hctx, i) {
1745	cpumask_clear(hctx->cpumask);
1746	hctx->nr_ctx = 0;
1747	}
1748
1749	/*
1750	* Map software to hardware queues
1751	*/
1752	for_each_possible_cpu(i) {
1753	/* If the cpu isn't online, the cpu is mapped to first hctx */
1754	if (!cpumask_test_cpu(i, online_mask))
1755	continue;
1756
1757	ctx = per_cpu_ptr(q->queue_ctx, i);
1758	hctx = blk_mq_map_queue(q, i);
1759
1760	cpumask_set_cpu(i, hctx->cpumask);
1761	ctx->index_hw = hctx->nr_ctx;
1762	hctx->ctxs[hctx->nr_ctx++] = ctx;
1763	}
1764
1765	mutex_unlock(&q->sysfs_lock);
1766
1767	queue_for_each_hw_ctx(q, hctx, i) {
1768	/*
1769	* If no software queues are mapped to this hardware queue,
1770	* disable it and free the request entries.
1771	*/
1772	if (!hctx->nr_ctx) {
1773	if (set->tags[i]) {
1774	blk_mq_free_rq_map(set, set->tags[i], i);
1775	set->tags[i] = NULL;
1776	}
1777	hctx->tags = NULL;
1778	continue;
1779	}
1780
1781	/* unmapped hw queue can be remapped after CPU topo changed */
1782	if (!set->tags[i])
1783	set->tags[i] = blk_mq_init_rq_map(set, i);
1784	hctx->tags = set->tags[i];
1785	WARN_ON(!hctx->tags);
1786
1787	/*
1788	* Set the map size to the number of mapped software queues.
1789	* This is more accurate and more efficient than looping
1790	* over all possibly mapped software queues.
1791	*/
1792	sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
1793
1794	/*
1795	* Initialize batch roundrobin counts
1796	*/
1797	hctx->next_cpu = cpumask_first(hctx->cpumask);
1798	hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1799	}
1800	}
1801
1802	static void queue_set_hctx_shared(struct request_queue *q, bool shared)
1803	{
1804	struct blk_mq_hw_ctx *hctx;
1805	int i;
1806
1807	queue_for_each_hw_ctx(q, hctx, i) {
1808	if (shared)
1809	hctx->flags |= BLK_MQ_F_TAG_SHARED;
1810	else
1811	hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1812	}
1813	}
1814
1815	static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
1816	{
1817	struct request_queue *q;
1818
1819	list_for_each_entry(q, &set->tag_list, tag_set_list) {
1820	blk_mq_freeze_queue(q);
1821	queue_set_hctx_shared(q, shared);
1822	blk_mq_unfreeze_queue(q);
1823	}
1824	}
1825
1826	static void blk_mq_del_queue_tag_set(struct request_queue *q)
1827	{
1828	struct blk_mq_tag_set *set = q->tag_set;
1829
1830	mutex_lock(&set->tag_list_lock);
1831	list_del_init(&q->tag_set_list);
1832	if (list_is_singular(&set->tag_list)) {
1833	/* just transitioned to unshared */
1834	set->flags &= ~BLK_MQ_F_TAG_SHARED;
1835	/* update existing queue */
1836	blk_mq_update_tag_set_depth(set, false);
1837	}
1838	mutex_unlock(&set->tag_list_lock);
1839	}
1840
1841	static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1842	struct request_queue *q)
1843	{
1844	q->tag_set = set;
1845
1846	mutex_lock(&set->tag_list_lock);
1847
1848	/* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1849	if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
1850	set->flags |= BLK_MQ_F_TAG_SHARED;
1851	/* update existing queue */
1852	blk_mq_update_tag_set_depth(set, true);
1853	}
1854	if (set->flags & BLK_MQ_F_TAG_SHARED)
1855	queue_set_hctx_shared(q, true);
1856	list_add_tail(&q->tag_set_list, &set->tag_list);
1857
1858	mutex_unlock(&set->tag_list_lock);
1859	}
1860
1861	/*
1862	* It is the actual release handler for mq, but we do it from
1863	* request queue's release handler for avoiding use-after-free
1864	* and headache because q->mq_kobj shouldn't have been introduced,
1865	* but we can't group ctx/kctx kobj without it.
1866	*/
1867	void blk_mq_release(struct request_queue *q)
1868	{
1869	struct blk_mq_hw_ctx *hctx;
1870	unsigned int i;
1871
1872	/* hctx kobj stays in hctx */
1873	queue_for_each_hw_ctx(q, hctx, i) {
1874	if (!hctx)
1875	continue;
1876	kfree(hctx->ctxs);
1877	kfree(hctx);
1878	}
1879
1880	q->mq_map = NULL;
1881
1882	kfree(q->queue_hw_ctx);
1883
1884	/* ctx kobj stays in queue_ctx */
1885	free_percpu(q->queue_ctx);
1886	}
1887
1888	struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1889	{
1890	struct request_queue *uninit_q, *q;
1891
1892	uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1893	if (!uninit_q)
1894	return ERR_PTR(-ENOMEM);
1895
1896	q = blk_mq_init_allocated_queue(set, uninit_q);
1897	if (IS_ERR(q))
1898	blk_cleanup_queue(uninit_q);
1899
1900	return q;
1901	}
1902	EXPORT_SYMBOL(blk_mq_init_queue);
1903
1904	static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
1905	struct request_queue *q)
1906	{
1907	int i, j;
1908	struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
1909
1910	blk_mq_sysfs_unregister(q);
1911
1912	/* protect against switching io scheduler */
1913	mutex_lock(&q->sysfs_lock);
1914	for (i = 0; i < set->nr_hw_queues; i++) {
1915	int node;
1916
1917	if (hctxs[i])
1918	continue;
1919
1920	node = blk_mq_hw_queue_to_node(q->mq_map, i);
1921	hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1922	GFP_KERNEL, node);
1923	if (!hctxs[i])
1924	break;
1925
1926	if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1927	node)) {
1928	kfree(hctxs[i]);
1929	hctxs[i] = NULL;
1930	break;
1931	}
1932
1933	atomic_set(&hctxs[i]->nr_active, 0);
1934	hctxs[i]->numa_node = node;
1935	hctxs[i]->queue_num = i;
1936
1937	if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
1938	free_cpumask_var(hctxs[i]->cpumask);
1939	kfree(hctxs[i]);
1940	hctxs[i] = NULL;
1941	break;
1942	}
1943	blk_mq_hctx_kobj_init(hctxs[i]);
1944	}
1945	for (j = i; j < q->nr_hw_queues; j++) {
1946	struct blk_mq_hw_ctx *hctx = hctxs[j];
1947
1948	if (hctx) {
1949	if (hctx->tags) {
1950	blk_mq_free_rq_map(set, hctx->tags, j);
1951	set->tags[j] = NULL;
1952	}
1953	blk_mq_exit_hctx(q, set, hctx, j);
1954	free_cpumask_var(hctx->cpumask);
1955	kobject_put(&hctx->kobj);
1956	kfree(hctx->ctxs);
1957	kfree(hctx);
1958	hctxs[j] = NULL;
1959
1960	}
1961	}
1962	q->nr_hw_queues = i;
1963	mutex_unlock(&q->sysfs_lock);
1964	blk_mq_sysfs_register(q);
1965	}
1966
1967	struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1968	struct request_queue *q)
1969	{
1970	/* mark the queue as mq asap */
1971	q->mq_ops = set->ops;
1972
1973	q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
1974	if (!q->queue_ctx)
1975	goto err_exit;
1976
1977	/* init q->mq_kobj and sw queues' kobjects */
1978	blk_mq_sysfs_init(q);
1979
1980	q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
1981	GFP_KERNEL, set->numa_node);
1982	if (!q->queue_hw_ctx)
1983	goto err_percpu;
1984
1985	q->mq_map = set->mq_map;
1986
1987	blk_mq_realloc_hw_ctxs(set, q);
1988	if (!q->nr_hw_queues)
1989	goto err_hctxs;
1990
1991	INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
1992	blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
1993
1994	q->nr_queues = nr_cpu_ids;
1995
1996	q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1997
1998	if (!(set->flags & BLK_MQ_F_SG_MERGE))
1999	q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2000
2001	q->sg_reserved_size = INT_MAX;
2002
2003	INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2004	INIT_LIST_HEAD(&q->requeue_list);
2005	spin_lock_init(&q->requeue_lock);
2006
2007	if (q->nr_hw_queues > 1)
2008	blk_queue_make_request(q, blk_mq_make_request);
2009	else
2010	blk_queue_make_request(q, blk_sq_make_request);
2011
2012	/*
2013	* Do this after blk_queue_make_request() overrides it...
2014	*/
2015	q->nr_requests = set->queue_depth;
2016
2017	if (set->ops->complete)
2018	blk_queue_softirq_done(q, set->ops->complete);
2019
2020	blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2021
2022	get_online_cpus();
2023	mutex_lock(&all_q_mutex);
2024
2025	list_add_tail(&q->all_q_node, &all_q_list);
2026	blk_mq_add_queue_tag_set(set, q);
2027	blk_mq_map_swqueue(q, cpu_online_mask);
2028
2029	mutex_unlock(&all_q_mutex);
2030	put_online_cpus();
2031
2032	return q;
2033
2034	err_hctxs:
2035	kfree(q->queue_hw_ctx);
2036	err_percpu:
2037	free_percpu(q->queue_ctx);
2038	err_exit:
2039	q->mq_ops = NULL;
2040	return ERR_PTR(-ENOMEM);
2041	}
2042	EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2043
2044	void blk_mq_free_queue(struct request_queue *q)
2045	{
2046	struct blk_mq_tag_set *set = q->tag_set;
2047
2048	mutex_lock(&all_q_mutex);
2049	list_del_init(&q->all_q_node);
2050	mutex_unlock(&all_q_mutex);
2051
2052	blk_mq_del_queue_tag_set(q);
2053
2054	blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2055	blk_mq_free_hw_queues(q, set);
2056	}
2057
2058	/* Basically redo blk_mq_init_queue with queue frozen */
2059	static void blk_mq_queue_reinit(struct request_queue *q,
2060	const struct cpumask *online_mask)
2061	{
2062	WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2063
2064	blk_mq_sysfs_unregister(q);
2065
2066	/*
2067	* redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2068	* we should change hctx numa_node according to new topology (this
2069	* involves free and re-allocate memory, worthy doing?)
2070	*/
2071
2072	blk_mq_map_swqueue(q, online_mask);
2073
2074	blk_mq_sysfs_register(q);
2075	}
2076
2077	/*
2078	* New online cpumask which is going to be set in this hotplug event.
2079	* Declare this cpumasks as global as cpu-hotplug operation is invoked
2080	* one-by-one and dynamically allocating this could result in a failure.
2081	*/
2082	static struct cpumask cpuhp_online_new;
2083
2084	static void blk_mq_queue_reinit_work(void)
2085	{
2086	struct request_queue *q;
2087
2088	mutex_lock(&all_q_mutex);
2089	/*
2090	* We need to freeze and reinit all existing queues. Freezing
2091	* involves synchronous wait for an RCU grace period and doing it
2092	* one by one may take a long time. Start freezing all queues in
2093	* one swoop and then wait for the completions so that freezing can
2094	* take place in parallel.
2095	*/
2096	list_for_each_entry(q, &all_q_list, all_q_node)
2097	blk_mq_freeze_queue_start(q);
2098	list_for_each_entry(q, &all_q_list, all_q_node) {
2099	blk_mq_freeze_queue_wait(q);
2100
2101	/*
2102	* timeout handler can't touch hw queue during the
2103	* reinitialization
2104	*/
2105	del_timer_sync(&q->timeout);
2106	}
2107
2108	list_for_each_entry(q, &all_q_list, all_q_node)
2109	blk_mq_queue_reinit(q, &cpuhp_online_new);
2110
2111	list_for_each_entry(q, &all_q_list, all_q_node)
2112	blk_mq_unfreeze_queue(q);
2113
2114	mutex_unlock(&all_q_mutex);
2115	}
2116
2117	static int blk_mq_queue_reinit_dead(unsigned int cpu)
2118	{
2119	cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2120	blk_mq_queue_reinit_work();
2121	return 0;
2122	}
2123
2124	/*
2125	* Before hotadded cpu starts handling requests, new mappings must be
2126	* established. Otherwise, these requests in hw queue might never be
2127	* dispatched.
2128	*
2129	* For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2130	* for CPU0, and ctx1 for CPU1).
2131	*
2132	* Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2133	* and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2134	*
2135	* And then while running hw queue, flush_busy_ctxs() finds bit0 is set in
2136	* pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2137	* But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list
2138	* is ignored.
2139	*/
2140	static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2141	{
2142	cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2143	cpumask_set_cpu(cpu, &cpuhp_online_new);
2144	blk_mq_queue_reinit_work();
2145	return 0;
2146	}
2147
2148	static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2149	{
2150	int i;
2151
2152	for (i = 0; i < set->nr_hw_queues; i++) {
2153	set->tags[i] = blk_mq_init_rq_map(set, i);
2154	if (!set->tags[i])
2155	goto out_unwind;
2156	}
2157
2158	return 0;
2159
2160	out_unwind:
2161	while (--i >= 0)
2162	blk_mq_free_rq_map(set, set->tags[i], i);
2163
2164	return -ENOMEM;
2165	}
2166
2167	/*
2168	* Allocate the request maps associated with this tag_set. Note that this
2169	* may reduce the depth asked for, if memory is tight. set->queue_depth
2170	* will be updated to reflect the allocated depth.
2171	*/
2172	static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2173	{
2174	unsigned int depth;
2175	int err;
2176
2177	depth = set->queue_depth;
2178	do {
2179	err = __blk_mq_alloc_rq_maps(set);
2180	if (!err)
2181	break;
2182
2183	set->queue_depth >>= 1;
2184	if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2185	err = -ENOMEM;
2186	break;
2187	}
2188	} while (set->queue_depth);
2189
2190	if (!set->queue_depth || err) {
2191	pr_err("blk-mq: failed to allocate request map\n");
2192	return -ENOMEM;
2193	}
2194
2195	if (depth != set->queue_depth)
2196	pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2197	depth, set->queue_depth);
2198
2199	return 0;
2200	}
2201
2202	/*
2203	* Alloc a tag set to be associated with one or more request queues.
2204	* May fail with EINVAL for various error conditions. May adjust the
2205	* requested depth down, if if it too large. In that case, the set
2206	* value will be stored in set->queue_depth.
2207	*/
2208	int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2209	{
2210	int ret;
2211
2212	BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2213
2214	if (!set->nr_hw_queues)
2215	return -EINVAL;
2216	if (!set->queue_depth)
2217	return -EINVAL;
2218	if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2219	return -EINVAL;
2220
2221	if (!set->ops->queue_rq)
2222	return -EINVAL;
2223
2224	if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2225	pr_info("blk-mq: reduced tag depth to %u\n",
2226	BLK_MQ_MAX_DEPTH);
2227	set->queue_depth = BLK_MQ_MAX_DEPTH;
2228	}
2229
2230	/*
2231	* If a crashdump is active, then we are potentially in a very
2232	* memory constrained environment. Limit us to 1 queue and
2233	* 64 tags to prevent using too much memory.
2234	*/
2235	if (is_kdump_kernel()) {
2236	set->nr_hw_queues = 1;
2237	set->queue_depth = min(64U, set->queue_depth);
2238	}
2239	/*
2240	* There is no use for more h/w queues than cpus.
2241	*/
2242	if (set->nr_hw_queues > nr_cpu_ids)
2243	set->nr_hw_queues = nr_cpu_ids;
2244
2245	set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2246	GFP_KERNEL, set->numa_node);
2247	if (!set->tags)
2248	return -ENOMEM;
2249
2250	ret = -ENOMEM;
2251	set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2252	GFP_KERNEL, set->numa_node);
2253	if (!set->mq_map)
2254	goto out_free_tags;
2255
2256	if (set->ops->map_queues)
2257	ret = set->ops->map_queues(set);
2258	else
2259	ret = blk_mq_map_queues(set);
2260	if (ret)
2261	goto out_free_mq_map;
2262
2263	ret = blk_mq_alloc_rq_maps(set);
2264	if (ret)
2265	goto out_free_mq_map;
2266
2267	mutex_init(&set->tag_list_lock);
2268	INIT_LIST_HEAD(&set->tag_list);
2269
2270	return 0;
2271
2272	out_free_mq_map:
2273	kfree(set->mq_map);
2274	set->mq_map = NULL;
2275	out_free_tags:
2276	kfree(set->tags);
2277	set->tags = NULL;
2278	return ret;
2279	}
2280	EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2281
2282	void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2283	{
2284	int i;
2285
2286	for (i = 0; i < nr_cpu_ids; i++) {
2287	if (set->tags[i])
2288	blk_mq_free_rq_map(set, set->tags[i], i);
2289	}
2290
2291	kfree(set->mq_map);
2292	set->mq_map = NULL;
2293
2294	kfree(set->tags);
2295	set->tags = NULL;
2296	}
2297	EXPORT_SYMBOL(blk_mq_free_tag_set);
2298
2299	int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2300	{
2301	struct blk_mq_tag_set *set = q->tag_set;
2302	struct blk_mq_hw_ctx *hctx;
2303	int i, ret;
2304
2305	if (!set || nr > set->queue_depth)
2306	return -EINVAL;
2307
2308	ret = 0;
2309	queue_for_each_hw_ctx(q, hctx, i) {
2310	if (!hctx->tags)
2311	continue;
2312	ret = blk_mq_tag_update_depth(hctx->tags, nr);
2313	if (ret)
2314	break;
2315	}
2316
2317	if (!ret)
2318	q->nr_requests = nr;
2319
2320	return ret;
2321	}
2322
2323	void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2324	{
2325	struct request_queue *q;
2326
2327	if (nr_hw_queues > nr_cpu_ids)
2328	nr_hw_queues = nr_cpu_ids;
2329	if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2330	return;
2331
2332	list_for_each_entry(q, &set->tag_list, tag_set_list)
2333	blk_mq_freeze_queue(q);
2334
2335	set->nr_hw_queues = nr_hw_queues;
2336	list_for_each_entry(q, &set->tag_list, tag_set_list) {
2337	blk_mq_realloc_hw_ctxs(set, q);
2338
2339	if (q->nr_hw_queues > 1)
2340	blk_queue_make_request(q, blk_mq_make_request);
2341	else
2342	blk_queue_make_request(q, blk_sq_make_request);
2343
2344	blk_mq_queue_reinit(q, cpu_online_mask);
2345	}
2346
2347	list_for_each_entry(q, &set->tag_list, tag_set_list)
2348	blk_mq_unfreeze_queue(q);
2349	}
2350	EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2351
2352	void blk_mq_disable_hotplug(void)
2353	{
2354	mutex_lock(&all_q_mutex);
2355	}
2356
2357	void blk_mq_enable_hotplug(void)
2358	{
2359	mutex_unlock(&all_q_mutex);
2360	}
2361
2362	static int __init blk_mq_init(void)
2363	{
2364	cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2365	blk_mq_hctx_notify_dead);
2366
2367	cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2368	blk_mq_queue_reinit_prepare,
2369	blk_mq_queue_reinit_dead);
2370	return 0;
2371	}
2372	subsys_initcall(blk_mq_init);
2373