Root/block/blk-mq.c

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/mm.h>
13#include <linux/init.h>
14#include <linux/slab.h>
15#include <linux/workqueue.h>
16#include <linux/smp.h>
17#include <linux/llist.h>
18#include <linux/list_sort.h>
19#include <linux/cpu.h>
20#include <linux/cache.h>
21#include <linux/sched/sysctl.h>
22#include <linux/delay.h>
23
24#include <trace/events/block.h>
25
26#include <linux/blk-mq.h>
27#include "blk.h"
28#include "blk-mq.h"
29#include "blk-mq-tag.h"
30
31static DEFINE_MUTEX(all_q_mutex);
32static LIST_HEAD(all_q_list);
33
34static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
35
36/*
37 * Check if any of the ctx's have pending work in this hardware queue
38 */
39static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
40{
41    unsigned int i;
42
43    for (i = 0; i < hctx->ctx_map.map_size; i++)
44        if (hctx->ctx_map.map[i].word)
45            return true;
46
47    return false;
48}
49
50static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
51                          struct blk_mq_ctx *ctx)
52{
53    return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
54}
55
56#define CTX_TO_BIT(hctx, ctx) \
57    ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
58
59/*
60 * Mark this ctx as having pending work in this hardware queue
61 */
62static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
63                     struct blk_mq_ctx *ctx)
64{
65    struct blk_align_bitmap *bm = get_bm(hctx, ctx);
66
67    if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
68        set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
69}
70
71static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
72                      struct blk_mq_ctx *ctx)
73{
74    struct blk_align_bitmap *bm = get_bm(hctx, ctx);
75
76    clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
77}
78
79static int blk_mq_queue_enter(struct request_queue *q)
80{
81    int ret;
82
83    __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
84    smp_wmb();
85
86    /* we have problems freezing the queue if it's initializing */
87    if (!blk_queue_dying(q) &&
88        (!blk_queue_bypass(q) || !blk_queue_init_done(q)))
89        return 0;
90
91    __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
92
93    spin_lock_irq(q->queue_lock);
94    ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq,
95        !blk_queue_bypass(q) || blk_queue_dying(q),
96        *q->queue_lock);
97    /* inc usage with lock hold to avoid freeze_queue runs here */
98    if (!ret && !blk_queue_dying(q))
99        __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
100    else if (blk_queue_dying(q))
101        ret = -ENODEV;
102    spin_unlock_irq(q->queue_lock);
103
104    return ret;
105}
106
107static void blk_mq_queue_exit(struct request_queue *q)
108{
109    __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
110}
111
112void blk_mq_drain_queue(struct request_queue *q)
113{
114    while (true) {
115        s64 count;
116
117        spin_lock_irq(q->queue_lock);
118        count = percpu_counter_sum(&q->mq_usage_counter);
119        spin_unlock_irq(q->queue_lock);
120
121        if (count == 0)
122            break;
123        blk_mq_start_hw_queues(q);
124        msleep(10);
125    }
126}
127
128/*
129 * Guarantee no request is in use, so we can change any data structure of
130 * the queue afterward.
131 */
132static void blk_mq_freeze_queue(struct request_queue *q)
133{
134    bool drain;
135
136    spin_lock_irq(q->queue_lock);
137    drain = !q->bypass_depth++;
138    queue_flag_set(QUEUE_FLAG_BYPASS, q);
139    spin_unlock_irq(q->queue_lock);
140
141    if (drain)
142        blk_mq_drain_queue(q);
143}
144
145static void blk_mq_unfreeze_queue(struct request_queue *q)
146{
147    bool wake = false;
148
149    spin_lock_irq(q->queue_lock);
150    if (!--q->bypass_depth) {
151        queue_flag_clear(QUEUE_FLAG_BYPASS, q);
152        wake = true;
153    }
154    WARN_ON_ONCE(q->bypass_depth < 0);
155    spin_unlock_irq(q->queue_lock);
156    if (wake)
157        wake_up_all(&q->mq_freeze_wq);
158}
159
160bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
161{
162    return blk_mq_has_free_tags(hctx->tags);
163}
164EXPORT_SYMBOL(blk_mq_can_queue);
165
166static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
167                   struct request *rq, unsigned int rw_flags)
168{
169    if (blk_queue_io_stat(q))
170        rw_flags |= REQ_IO_STAT;
171
172    INIT_LIST_HEAD(&rq->queuelist);
173    /* csd/requeue_work/fifo_time is initialized before use */
174    rq->q = q;
175    rq->mq_ctx = ctx;
176    rq->cmd_flags |= rw_flags;
177    /* do not touch atomic flags, it needs atomic ops against the timer */
178    rq->cpu = -1;
179    INIT_HLIST_NODE(&rq->hash);
180    RB_CLEAR_NODE(&rq->rb_node);
181    rq->rq_disk = NULL;
182    rq->part = NULL;
183    rq->start_time = jiffies;
184#ifdef CONFIG_BLK_CGROUP
185    rq->rl = NULL;
186    set_start_time_ns(rq);
187    rq->io_start_time_ns = 0;
188#endif
189    rq->nr_phys_segments = 0;
190#if defined(CONFIG_BLK_DEV_INTEGRITY)
191    rq->nr_integrity_segments = 0;
192#endif
193    rq->special = NULL;
194    /* tag was already set */
195    rq->errors = 0;
196
197    rq->extra_len = 0;
198    rq->sense_len = 0;
199    rq->resid_len = 0;
200    rq->sense = NULL;
201
202    INIT_LIST_HEAD(&rq->timeout_list);
203    rq->timeout = 0;
204
205    rq->end_io = NULL;
206    rq->end_io_data = NULL;
207    rq->next_rq = NULL;
208
209    ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
210}
211
212static struct request *
213__blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
214{
215    struct request *rq;
216    unsigned int tag;
217
218    tag = blk_mq_get_tag(data);
219    if (tag != BLK_MQ_TAG_FAIL) {
220        rq = data->hctx->tags->rqs[tag];
221
222        rq->cmd_flags = 0;
223        if (blk_mq_tag_busy(data->hctx)) {
224            rq->cmd_flags = REQ_MQ_INFLIGHT;
225            atomic_inc(&data->hctx->nr_active);
226        }
227
228        rq->tag = tag;
229        blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
230        return rq;
231    }
232
233    return NULL;
234}
235
236struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
237        bool reserved)
238{
239    struct blk_mq_ctx *ctx;
240    struct blk_mq_hw_ctx *hctx;
241    struct request *rq;
242    struct blk_mq_alloc_data alloc_data;
243
244    if (blk_mq_queue_enter(q))
245        return NULL;
246
247    ctx = blk_mq_get_ctx(q);
248    hctx = q->mq_ops->map_queue(q, ctx->cpu);
249    blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
250            reserved, ctx, hctx);
251
252    rq = __blk_mq_alloc_request(&alloc_data, rw);
253    if (!rq && (gfp & __GFP_WAIT)) {
254        __blk_mq_run_hw_queue(hctx);
255        blk_mq_put_ctx(ctx);
256
257        ctx = blk_mq_get_ctx(q);
258        hctx = q->mq_ops->map_queue(q, ctx->cpu);
259        blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
260                hctx);
261        rq = __blk_mq_alloc_request(&alloc_data, rw);
262        ctx = alloc_data.ctx;
263    }
264    blk_mq_put_ctx(ctx);
265    return rq;
266}
267EXPORT_SYMBOL(blk_mq_alloc_request);
268
269static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
270                  struct blk_mq_ctx *ctx, struct request *rq)
271{
272    const int tag = rq->tag;
273    struct request_queue *q = rq->q;
274
275    if (rq->cmd_flags & REQ_MQ_INFLIGHT)
276        atomic_dec(&hctx->nr_active);
277
278    clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
279    blk_mq_put_tag(hctx, tag, &ctx->last_tag);
280    blk_mq_queue_exit(q);
281}
282
283void blk_mq_free_request(struct request *rq)
284{
285    struct blk_mq_ctx *ctx = rq->mq_ctx;
286    struct blk_mq_hw_ctx *hctx;
287    struct request_queue *q = rq->q;
288
289    ctx->rq_completed[rq_is_sync(rq)]++;
290
291    hctx = q->mq_ops->map_queue(q, ctx->cpu);
292    __blk_mq_free_request(hctx, ctx, rq);
293}
294
295/*
296 * Clone all relevant state from a request that has been put on hold in
297 * the flush state machine into the preallocated flush request that hangs
298 * off the request queue.
299 *
300 * For a driver the flush request should be invisible, that's why we are
301 * impersonating the original request here.
302 */
303void blk_mq_clone_flush_request(struct request *flush_rq,
304        struct request *orig_rq)
305{
306    struct blk_mq_hw_ctx *hctx =
307        orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu);
308
309    flush_rq->mq_ctx = orig_rq->mq_ctx;
310    flush_rq->tag = orig_rq->tag;
311    memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq),
312        hctx->cmd_size);
313}
314
315inline void __blk_mq_end_io(struct request *rq, int error)
316{
317    blk_account_io_done(rq);
318
319    if (rq->end_io) {
320        rq->end_io(rq, error);
321    } else {
322        if (unlikely(blk_bidi_rq(rq)))
323            blk_mq_free_request(rq->next_rq);
324        blk_mq_free_request(rq);
325    }
326}
327EXPORT_SYMBOL(__blk_mq_end_io);
328
329void blk_mq_end_io(struct request *rq, int error)
330{
331    if (blk_update_request(rq, error, blk_rq_bytes(rq)))
332        BUG();
333    __blk_mq_end_io(rq, error);
334}
335EXPORT_SYMBOL(blk_mq_end_io);
336
337static void __blk_mq_complete_request_remote(void *data)
338{
339    struct request *rq = data;
340
341    rq->q->softirq_done_fn(rq);
342}
343
344static void blk_mq_ipi_complete_request(struct request *rq)
345{
346    struct blk_mq_ctx *ctx = rq->mq_ctx;
347    bool shared = false;
348    int cpu;
349
350    if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
351        rq->q->softirq_done_fn(rq);
352        return;
353    }
354
355    cpu = get_cpu();
356    if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
357        shared = cpus_share_cache(cpu, ctx->cpu);
358
359    if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
360        rq->csd.func = __blk_mq_complete_request_remote;
361        rq->csd.info = rq;
362        rq->csd.flags = 0;
363        smp_call_function_single_async(ctx->cpu, &rq->csd);
364    } else {
365        rq->q->softirq_done_fn(rq);
366    }
367    put_cpu();
368}
369
370void __blk_mq_complete_request(struct request *rq)
371{
372    struct request_queue *q = rq->q;
373
374    if (!q->softirq_done_fn)
375        blk_mq_end_io(rq, rq->errors);
376    else
377        blk_mq_ipi_complete_request(rq);
378}
379
380/**
381 * blk_mq_complete_request - end I/O on a request
382 * @rq: the request being processed
383 *
384 * Description:
385 * Ends all I/O on a request. It does not handle partial completions.
386 * The actual completion happens out-of-order, through a IPI handler.
387 **/
388void blk_mq_complete_request(struct request *rq)
389{
390    struct request_queue *q = rq->q;
391
392    if (unlikely(blk_should_fake_timeout(q)))
393        return;
394    if (!blk_mark_rq_complete(rq))
395        __blk_mq_complete_request(rq);
396}
397EXPORT_SYMBOL(blk_mq_complete_request);
398
399static void blk_mq_start_request(struct request *rq, bool last)
400{
401    struct request_queue *q = rq->q;
402
403    trace_block_rq_issue(q, rq);
404
405    rq->resid_len = blk_rq_bytes(rq);
406    if (unlikely(blk_bidi_rq(rq)))
407        rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
408
409    blk_add_timer(rq);
410
411    /*
412     * Mark us as started and clear complete. Complete might have been
413     * set if requeue raced with timeout, which then marked it as
414     * complete. So be sure to clear complete again when we start
415     * the request, otherwise we'll ignore the completion event.
416     */
417    if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
418        set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
419    if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
420        clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
421
422    if (q->dma_drain_size && blk_rq_bytes(rq)) {
423        /*
424         * Make sure space for the drain appears. We know we can do
425         * this because max_hw_segments has been adjusted to be one
426         * fewer than the device can handle.
427         */
428        rq->nr_phys_segments++;
429    }
430
431    /*
432     * Flag the last request in the series so that drivers know when IO
433     * should be kicked off, if they don't do it on a per-request basis.
434     *
435     * Note: the flag isn't the only condition drivers should do kick off.
436     * If drive is busy, the last request might not have the bit set.
437     */
438    if (last)
439        rq->cmd_flags |= REQ_END;
440}
441
442static void __blk_mq_requeue_request(struct request *rq)
443{
444    struct request_queue *q = rq->q;
445
446    trace_block_rq_requeue(q, rq);
447    clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
448
449    rq->cmd_flags &= ~REQ_END;
450
451    if (q->dma_drain_size && blk_rq_bytes(rq))
452        rq->nr_phys_segments--;
453}
454
455void blk_mq_requeue_request(struct request *rq)
456{
457    __blk_mq_requeue_request(rq);
458    blk_clear_rq_complete(rq);
459
460    BUG_ON(blk_queued_rq(rq));
461    blk_mq_add_to_requeue_list(rq, true);
462}
463EXPORT_SYMBOL(blk_mq_requeue_request);
464
465static void blk_mq_requeue_work(struct work_struct *work)
466{
467    struct request_queue *q =
468        container_of(work, struct request_queue, requeue_work);
469    LIST_HEAD(rq_list);
470    struct request *rq, *next;
471    unsigned long flags;
472
473    spin_lock_irqsave(&q->requeue_lock, flags);
474    list_splice_init(&q->requeue_list, &rq_list);
475    spin_unlock_irqrestore(&q->requeue_lock, flags);
476
477    list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
478        if (!(rq->cmd_flags & REQ_SOFTBARRIER))
479            continue;
480
481        rq->cmd_flags &= ~REQ_SOFTBARRIER;
482        list_del_init(&rq->queuelist);
483        blk_mq_insert_request(rq, true, false, false);
484    }
485
486    while (!list_empty(&rq_list)) {
487        rq = list_entry(rq_list.next, struct request, queuelist);
488        list_del_init(&rq->queuelist);
489        blk_mq_insert_request(rq, false, false, false);
490    }
491
492    blk_mq_run_queues(q, false);
493}
494
495void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
496{
497    struct request_queue *q = rq->q;
498    unsigned long flags;
499
500    /*
501     * We abuse this flag that is otherwise used by the I/O scheduler to
502     * request head insertation from the workqueue.
503     */
504    BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
505
506    spin_lock_irqsave(&q->requeue_lock, flags);
507    if (at_head) {
508        rq->cmd_flags |= REQ_SOFTBARRIER;
509        list_add(&rq->queuelist, &q->requeue_list);
510    } else {
511        list_add_tail(&rq->queuelist, &q->requeue_list);
512    }
513    spin_unlock_irqrestore(&q->requeue_lock, flags);
514}
515EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
516
517void blk_mq_kick_requeue_list(struct request_queue *q)
518{
519    kblockd_schedule_work(&q->requeue_work);
520}
521EXPORT_SYMBOL(blk_mq_kick_requeue_list);
522
523static inline bool is_flush_request(struct request *rq, unsigned int tag)
524{
525    return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
526            rq->q->flush_rq->tag == tag);
527}
528
529struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
530{
531    struct request *rq = tags->rqs[tag];
532
533    if (!is_flush_request(rq, tag))
534        return rq;
535
536    return rq->q->flush_rq;
537}
538EXPORT_SYMBOL(blk_mq_tag_to_rq);
539
540struct blk_mq_timeout_data {
541    struct blk_mq_hw_ctx *hctx;
542    unsigned long *next;
543    unsigned int *next_set;
544};
545
546static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
547{
548    struct blk_mq_timeout_data *data = __data;
549    struct blk_mq_hw_ctx *hctx = data->hctx;
550    unsigned int tag;
551
552     /* It may not be in flight yet (this is where
553     * the REQ_ATOMIC_STARTED flag comes in). The requests are
554     * statically allocated, so we know it's always safe to access the
555     * memory associated with a bit offset into ->rqs[].
556     */
557    tag = 0;
558    do {
559        struct request *rq;
560
561        tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
562        if (tag >= hctx->tags->nr_tags)
563            break;
564
565        rq = blk_mq_tag_to_rq(hctx->tags, tag++);
566        if (rq->q != hctx->queue)
567            continue;
568        if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
569            continue;
570
571        blk_rq_check_expired(rq, data->next, data->next_set);
572    } while (1);
573}
574
575static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
576                    unsigned long *next,
577                    unsigned int *next_set)
578{
579    struct blk_mq_timeout_data data = {
580        .hctx = hctx,
581        .next = next,
582        .next_set = next_set,
583    };
584
585    /*
586     * Ask the tagging code to iterate busy requests, so we can
587     * check them for timeout.
588     */
589    blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
590}
591
592static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq)
593{
594    struct request_queue *q = rq->q;
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, &rq->atomic_flags))
606        return BLK_EH_NOT_HANDLED;
607
608    if (!q->mq_ops->timeout)
609        return BLK_EH_RESET_TIMER;
610
611    return q->mq_ops->timeout(rq);
612}
613
614static void blk_mq_rq_timer(unsigned long data)
615{
616    struct request_queue *q = (struct request_queue *) data;
617    struct blk_mq_hw_ctx *hctx;
618    unsigned long next = 0;
619    int i, next_set = 0;
620
621    queue_for_each_hw_ctx(q, hctx, i) {
622        /*
623         * If not software queues are currently mapped to this
624         * hardware queue, there's nothing to check
625         */
626        if (!hctx->nr_ctx || !hctx->tags)
627            continue;
628
629        blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
630    }
631
632    if (next_set) {
633        next = blk_rq_timeout(round_jiffies_up(next));
634        mod_timer(&q->timeout, next);
635    } else {
636        queue_for_each_hw_ctx(q, hctx, i)
637            blk_mq_tag_idle(hctx);
638    }
639}
640
641/*
642 * Reverse check our software queue for entries that we could potentially
643 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
644 * too much time checking for merges.
645 */
646static bool blk_mq_attempt_merge(struct request_queue *q,
647                 struct blk_mq_ctx *ctx, struct bio *bio)
648{
649    struct request *rq;
650    int checked = 8;
651
652    list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
653        int el_ret;
654
655        if (!checked--)
656            break;
657
658        if (!blk_rq_merge_ok(rq, bio))
659            continue;
660
661        el_ret = blk_try_merge(rq, bio);
662        if (el_ret == ELEVATOR_BACK_MERGE) {
663            if (bio_attempt_back_merge(q, rq, bio)) {
664                ctx->rq_merged++;
665                return true;
666            }
667            break;
668        } else if (el_ret == ELEVATOR_FRONT_MERGE) {
669            if (bio_attempt_front_merge(q, rq, bio)) {
670                ctx->rq_merged++;
671                return true;
672            }
673            break;
674        }
675    }
676
677    return false;
678}
679
680/*
681 * Process software queues that have been marked busy, splicing them
682 * to the for-dispatch
683 */
684static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
685{
686    struct blk_mq_ctx *ctx;
687    int i;
688
689    for (i = 0; i < hctx->ctx_map.map_size; i++) {
690        struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
691        unsigned int off, bit;
692
693        if (!bm->word)
694            continue;
695
696        bit = 0;
697        off = i * hctx->ctx_map.bits_per_word;
698        do {
699            bit = find_next_bit(&bm->word, bm->depth, bit);
700            if (bit >= bm->depth)
701                break;
702
703            ctx = hctx->ctxs[bit + off];
704            clear_bit(bit, &bm->word);
705            spin_lock(&ctx->lock);
706            list_splice_tail_init(&ctx->rq_list, list);
707            spin_unlock(&ctx->lock);
708
709            bit++;
710        } while (1);
711    }
712}
713
714/*
715 * Run this hardware queue, pulling any software queues mapped to it in.
716 * Note that this function currently has various problems around ordering
717 * of IO. In particular, we'd like FIFO behaviour on handling existing
718 * items on the hctx->dispatch list. Ignore that for now.
719 */
720static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
721{
722    struct request_queue *q = hctx->queue;
723    struct request *rq;
724    LIST_HEAD(rq_list);
725    int queued;
726
727    WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
728
729    if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
730        return;
731
732    hctx->run++;
733
734    /*
735     * Touch any software queue that has pending entries.
736     */
737    flush_busy_ctxs(hctx, &rq_list);
738
739    /*
740     * If we have previous entries on our dispatch list, grab them
741     * and stuff them at the front for more fair dispatch.
742     */
743    if (!list_empty_careful(&hctx->dispatch)) {
744        spin_lock(&hctx->lock);
745        if (!list_empty(&hctx->dispatch))
746            list_splice_init(&hctx->dispatch, &rq_list);
747        spin_unlock(&hctx->lock);
748    }
749
750    /*
751     * Now process all the entries, sending them to the driver.
752     */
753    queued = 0;
754    while (!list_empty(&rq_list)) {
755        int ret;
756
757        rq = list_first_entry(&rq_list, struct request, queuelist);
758        list_del_init(&rq->queuelist);
759
760        blk_mq_start_request(rq, list_empty(&rq_list));
761
762        ret = q->mq_ops->queue_rq(hctx, rq);
763        switch (ret) {
764        case BLK_MQ_RQ_QUEUE_OK:
765            queued++;
766            continue;
767        case BLK_MQ_RQ_QUEUE_BUSY:
768            list_add(&rq->queuelist, &rq_list);
769            __blk_mq_requeue_request(rq);
770            break;
771        default:
772            pr_err("blk-mq: bad return on queue: %d\n", ret);
773        case BLK_MQ_RQ_QUEUE_ERROR:
774            rq->errors = -EIO;
775            blk_mq_end_io(rq, rq->errors);
776            break;
777        }
778
779        if (ret == BLK_MQ_RQ_QUEUE_BUSY)
780            break;
781    }
782
783    if (!queued)
784        hctx->dispatched[0]++;
785    else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
786        hctx->dispatched[ilog2(queued) + 1]++;
787
788    /*
789     * Any items that need requeuing? Stuff them into hctx->dispatch,
790     * that is where we will continue on next queue run.
791     */
792    if (!list_empty(&rq_list)) {
793        spin_lock(&hctx->lock);
794        list_splice(&rq_list, &hctx->dispatch);
795        spin_unlock(&hctx->lock);
796    }
797}
798
799/*
800 * It'd be great if the workqueue API had a way to pass
801 * in a mask and had some smarts for more clever placement.
802 * For now we just round-robin here, switching for every
803 * BLK_MQ_CPU_WORK_BATCH queued items.
804 */
805static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
806{
807    int cpu = hctx->next_cpu;
808
809    if (--hctx->next_cpu_batch <= 0) {
810        int next_cpu;
811
812        next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
813        if (next_cpu >= nr_cpu_ids)
814            next_cpu = cpumask_first(hctx->cpumask);
815
816        hctx->next_cpu = next_cpu;
817        hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
818    }
819
820    return cpu;
821}
822
823void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
824{
825    if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
826        return;
827
828    if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
829        __blk_mq_run_hw_queue(hctx);
830    else if (hctx->queue->nr_hw_queues == 1)
831        kblockd_schedule_delayed_work(&hctx->run_work, 0);
832    else {
833        unsigned int cpu;
834
835        cpu = blk_mq_hctx_next_cpu(hctx);
836        kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
837    }
838}
839
840void blk_mq_run_queues(struct request_queue *q, bool async)
841{
842    struct blk_mq_hw_ctx *hctx;
843    int i;
844
845    queue_for_each_hw_ctx(q, hctx, i) {
846        if ((!blk_mq_hctx_has_pending(hctx) &&
847            list_empty_careful(&hctx->dispatch)) ||
848            test_bit(BLK_MQ_S_STOPPED, &hctx->state))
849            continue;
850
851        preempt_disable();
852        blk_mq_run_hw_queue(hctx, async);
853        preempt_enable();
854    }
855}
856EXPORT_SYMBOL(blk_mq_run_queues);
857
858void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
859{
860    cancel_delayed_work(&hctx->run_work);
861    cancel_delayed_work(&hctx->delay_work);
862    set_bit(BLK_MQ_S_STOPPED, &hctx->state);
863}
864EXPORT_SYMBOL(blk_mq_stop_hw_queue);
865
866void blk_mq_stop_hw_queues(struct request_queue *q)
867{
868    struct blk_mq_hw_ctx *hctx;
869    int i;
870
871    queue_for_each_hw_ctx(q, hctx, i)
872        blk_mq_stop_hw_queue(hctx);
873}
874EXPORT_SYMBOL(blk_mq_stop_hw_queues);
875
876void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
877{
878    clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
879
880    preempt_disable();
881    blk_mq_run_hw_queue(hctx, false);
882    preempt_enable();
883}
884EXPORT_SYMBOL(blk_mq_start_hw_queue);
885
886void blk_mq_start_hw_queues(struct request_queue *q)
887{
888    struct blk_mq_hw_ctx *hctx;
889    int i;
890
891    queue_for_each_hw_ctx(q, hctx, i)
892        blk_mq_start_hw_queue(hctx);
893}
894EXPORT_SYMBOL(blk_mq_start_hw_queues);
895
896
897void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
898{
899    struct blk_mq_hw_ctx *hctx;
900    int i;
901
902    queue_for_each_hw_ctx(q, hctx, i) {
903        if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
904            continue;
905
906        clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
907        preempt_disable();
908        blk_mq_run_hw_queue(hctx, async);
909        preempt_enable();
910    }
911}
912EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
913
914static void blk_mq_run_work_fn(struct work_struct *work)
915{
916    struct blk_mq_hw_ctx *hctx;
917
918    hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
919
920    __blk_mq_run_hw_queue(hctx);
921}
922
923static void blk_mq_delay_work_fn(struct work_struct *work)
924{
925    struct blk_mq_hw_ctx *hctx;
926
927    hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
928
929    if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
930        __blk_mq_run_hw_queue(hctx);
931}
932
933void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
934{
935    unsigned long tmo = msecs_to_jiffies(msecs);
936
937    if (hctx->queue->nr_hw_queues == 1)
938        kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
939    else {
940        unsigned int cpu;
941
942        cpu = blk_mq_hctx_next_cpu(hctx);
943        kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
944    }
945}
946EXPORT_SYMBOL(blk_mq_delay_queue);
947
948static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
949                    struct request *rq, bool at_head)
950{
951    struct blk_mq_ctx *ctx = rq->mq_ctx;
952
953    trace_block_rq_insert(hctx->queue, rq);
954
955    if (at_head)
956        list_add(&rq->queuelist, &ctx->rq_list);
957    else
958        list_add_tail(&rq->queuelist, &ctx->rq_list);
959
960    blk_mq_hctx_mark_pending(hctx, ctx);
961}
962
963void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
964        bool async)
965{
966    struct request_queue *q = rq->q;
967    struct blk_mq_hw_ctx *hctx;
968    struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
969
970    current_ctx = blk_mq_get_ctx(q);
971    if (!cpu_online(ctx->cpu))
972        rq->mq_ctx = ctx = current_ctx;
973
974    hctx = q->mq_ops->map_queue(q, ctx->cpu);
975
976    if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) &&
977        !(rq->cmd_flags & (REQ_FLUSH_SEQ))) {
978        blk_insert_flush(rq);
979    } else {
980        spin_lock(&ctx->lock);
981        __blk_mq_insert_request(hctx, rq, at_head);
982        spin_unlock(&ctx->lock);
983    }
984
985    if (run_queue)
986        blk_mq_run_hw_queue(hctx, async);
987
988    blk_mq_put_ctx(current_ctx);
989}
990
991static void blk_mq_insert_requests(struct request_queue *q,
992                     struct blk_mq_ctx *ctx,
993                     struct list_head *list,
994                     int depth,
995                     bool from_schedule)
996
997{
998    struct blk_mq_hw_ctx *hctx;
999    struct blk_mq_ctx *current_ctx;
1000
1001    trace_block_unplug(q, depth, !from_schedule);
1002
1003    current_ctx = blk_mq_get_ctx(q);
1004
1005    if (!cpu_online(ctx->cpu))
1006        ctx = current_ctx;
1007    hctx = q->mq_ops->map_queue(q, ctx->cpu);
1008
1009    /*
1010     * preemption doesn't flush plug list, so it's possible ctx->cpu is
1011     * offline now
1012     */
1013    spin_lock(&ctx->lock);
1014    while (!list_empty(list)) {
1015        struct request *rq;
1016
1017        rq = list_first_entry(list, struct request, queuelist);
1018        list_del_init(&rq->queuelist);
1019        rq->mq_ctx = ctx;
1020        __blk_mq_insert_request(hctx, rq, false);
1021    }
1022    spin_unlock(&ctx->lock);
1023
1024    blk_mq_run_hw_queue(hctx, from_schedule);
1025    blk_mq_put_ctx(current_ctx);
1026}
1027
1028static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1029{
1030    struct request *rqa = container_of(a, struct request, queuelist);
1031    struct request *rqb = container_of(b, struct request, queuelist);
1032
1033    return !(rqa->mq_ctx < rqb->mq_ctx ||
1034         (rqa->mq_ctx == rqb->mq_ctx &&
1035          blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1036}
1037
1038void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1039{
1040    struct blk_mq_ctx *this_ctx;
1041    struct request_queue *this_q;
1042    struct request *rq;
1043    LIST_HEAD(list);
1044    LIST_HEAD(ctx_list);
1045    unsigned int depth;
1046
1047    list_splice_init(&plug->mq_list, &list);
1048
1049    list_sort(NULL, &list, plug_ctx_cmp);
1050
1051    this_q = NULL;
1052    this_ctx = NULL;
1053    depth = 0;
1054
1055    while (!list_empty(&list)) {
1056        rq = list_entry_rq(list.next);
1057        list_del_init(&rq->queuelist);
1058        BUG_ON(!rq->q);
1059        if (rq->mq_ctx != this_ctx) {
1060            if (this_ctx) {
1061                blk_mq_insert_requests(this_q, this_ctx,
1062                            &ctx_list, depth,
1063                            from_schedule);
1064            }
1065
1066            this_ctx = rq->mq_ctx;
1067            this_q = rq->q;
1068            depth = 0;
1069        }
1070
1071        depth++;
1072        list_add_tail(&rq->queuelist, &ctx_list);
1073    }
1074
1075    /*
1076     * If 'this_ctx' is set, we know we have entries to complete
1077     * on 'ctx_list'. Do those.
1078     */
1079    if (this_ctx) {
1080        blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1081                       from_schedule);
1082    }
1083}
1084
1085static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1086{
1087    init_request_from_bio(rq, bio);
1088
1089    if (blk_do_io_stat(rq))
1090        blk_account_io_start(rq, 1);
1091}
1092
1093static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1094                     struct blk_mq_ctx *ctx,
1095                     struct request *rq, struct bio *bio)
1096{
1097    struct request_queue *q = hctx->queue;
1098
1099    if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE)) {
1100        blk_mq_bio_to_request(rq, bio);
1101        spin_lock(&ctx->lock);
1102insert_rq:
1103        __blk_mq_insert_request(hctx, rq, false);
1104        spin_unlock(&ctx->lock);
1105        return false;
1106    } else {
1107        spin_lock(&ctx->lock);
1108        if (!blk_mq_attempt_merge(q, ctx, bio)) {
1109            blk_mq_bio_to_request(rq, bio);
1110            goto insert_rq;
1111        }
1112
1113        spin_unlock(&ctx->lock);
1114        __blk_mq_free_request(hctx, ctx, rq);
1115        return true;
1116    }
1117}
1118
1119struct blk_map_ctx {
1120    struct blk_mq_hw_ctx *hctx;
1121    struct blk_mq_ctx *ctx;
1122};
1123
1124static struct request *blk_mq_map_request(struct request_queue *q,
1125                      struct bio *bio,
1126                      struct blk_map_ctx *data)
1127{
1128    struct blk_mq_hw_ctx *hctx;
1129    struct blk_mq_ctx *ctx;
1130    struct request *rq;
1131    int rw = bio_data_dir(bio);
1132    struct blk_mq_alloc_data alloc_data;
1133
1134    if (unlikely(blk_mq_queue_enter(q))) {
1135        bio_endio(bio, -EIO);
1136        return NULL;
1137    }
1138
1139    ctx = blk_mq_get_ctx(q);
1140    hctx = q->mq_ops->map_queue(q, ctx->cpu);
1141
1142    if (rw_is_sync(bio->bi_rw))
1143        rw |= REQ_SYNC;
1144
1145    trace_block_getrq(q, bio, rw);
1146    blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1147            hctx);
1148    rq = __blk_mq_alloc_request(&alloc_data, rw);
1149    if (unlikely(!rq)) {
1150        __blk_mq_run_hw_queue(hctx);
1151        blk_mq_put_ctx(ctx);
1152        trace_block_sleeprq(q, bio, rw);
1153
1154        ctx = blk_mq_get_ctx(q);
1155        hctx = q->mq_ops->map_queue(q, ctx->cpu);
1156        blk_mq_set_alloc_data(&alloc_data, q,
1157                __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1158        rq = __blk_mq_alloc_request(&alloc_data, rw);
1159        ctx = alloc_data.ctx;
1160        hctx = alloc_data.hctx;
1161    }
1162
1163    hctx->queued++;
1164    data->hctx = hctx;
1165    data->ctx = ctx;
1166    return rq;
1167}
1168
1169/*
1170 * Multiple hardware queue variant. This will not use per-process plugs,
1171 * but will attempt to bypass the hctx queueing if we can go straight to
1172 * hardware for SYNC IO.
1173 */
1174static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1175{
1176    const int is_sync = rw_is_sync(bio->bi_rw);
1177    const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1178    struct blk_map_ctx data;
1179    struct request *rq;
1180
1181    blk_queue_bounce(q, &bio);
1182
1183    if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1184        bio_endio(bio, -EIO);
1185        return;
1186    }
1187
1188    rq = blk_mq_map_request(q, bio, &data);
1189    if (unlikely(!rq))
1190        return;
1191
1192    if (unlikely(is_flush_fua)) {
1193        blk_mq_bio_to_request(rq, bio);
1194        blk_insert_flush(rq);
1195        goto run_queue;
1196    }
1197
1198    if (is_sync) {
1199        int ret;
1200
1201        blk_mq_bio_to_request(rq, bio);
1202        blk_mq_start_request(rq, true);
1203
1204        /*
1205         * For OK queue, we are done. For error, kill it. Any other
1206         * error (busy), just add it to our list as we previously
1207         * would have done
1208         */
1209        ret = q->mq_ops->queue_rq(data.hctx, rq);
1210        if (ret == BLK_MQ_RQ_QUEUE_OK)
1211            goto done;
1212        else {
1213            __blk_mq_requeue_request(rq);
1214
1215            if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1216                rq->errors = -EIO;
1217                blk_mq_end_io(rq, rq->errors);
1218                goto done;
1219            }
1220        }
1221    }
1222
1223    if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1224        /*
1225         * For a SYNC request, send it to the hardware immediately. For
1226         * an ASYNC request, just ensure that we run it later on. The
1227         * latter allows for merging opportunities and more efficient
1228         * dispatching.
1229         */
1230run_queue:
1231        blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1232    }
1233done:
1234    blk_mq_put_ctx(data.ctx);
1235}
1236
1237/*
1238 * Single hardware queue variant. This will attempt to use any per-process
1239 * plug for merging and IO deferral.
1240 */
1241static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1242{
1243    const int is_sync = rw_is_sync(bio->bi_rw);
1244    const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1245    unsigned int use_plug, request_count = 0;
1246    struct blk_map_ctx data;
1247    struct request *rq;
1248
1249    /*
1250     * If we have multiple hardware queues, just go directly to
1251     * one of those for sync IO.
1252     */
1253    use_plug = !is_flush_fua && !is_sync;
1254
1255    blk_queue_bounce(q, &bio);
1256
1257    if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1258        bio_endio(bio, -EIO);
1259        return;
1260    }
1261
1262    if (use_plug && !blk_queue_nomerges(q) &&
1263        blk_attempt_plug_merge(q, bio, &request_count))
1264        return;
1265
1266    rq = blk_mq_map_request(q, bio, &data);
1267    if (unlikely(!rq))
1268        return;
1269
1270    if (unlikely(is_flush_fua)) {
1271        blk_mq_bio_to_request(rq, bio);
1272        blk_insert_flush(rq);
1273        goto run_queue;
1274    }
1275
1276    /*
1277     * A task plug currently exists. Since this is completely lockless,
1278     * utilize that to temporarily store requests until the task is
1279     * either done or scheduled away.
1280     */
1281    if (use_plug) {
1282        struct blk_plug *plug = current->plug;
1283
1284        if (plug) {
1285            blk_mq_bio_to_request(rq, bio);
1286            if (list_empty(&plug->mq_list))
1287                trace_block_plug(q);
1288            else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1289                blk_flush_plug_list(plug, false);
1290                trace_block_plug(q);
1291            }
1292            list_add_tail(&rq->queuelist, &plug->mq_list);
1293            blk_mq_put_ctx(data.ctx);
1294            return;
1295        }
1296    }
1297
1298    if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1299        /*
1300         * For a SYNC request, send it to the hardware immediately. For
1301         * an ASYNC request, just ensure that we run it later on. The
1302         * latter allows for merging opportunities and more efficient
1303         * dispatching.
1304         */
1305run_queue:
1306        blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1307    }
1308
1309    blk_mq_put_ctx(data.ctx);
1310}
1311
1312/*
1313 * Default mapping to a software queue, since we use one per CPU.
1314 */
1315struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1316{
1317    return q->queue_hw_ctx[q->mq_map[cpu]];
1318}
1319EXPORT_SYMBOL(blk_mq_map_queue);
1320
1321static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1322        struct blk_mq_tags *tags, unsigned int hctx_idx)
1323{
1324    struct page *page;
1325
1326    if (tags->rqs && set->ops->exit_request) {
1327        int i;
1328
1329        for (i = 0; i < tags->nr_tags; i++) {
1330            if (!tags->rqs[i])
1331                continue;
1332            set->ops->exit_request(set->driver_data, tags->rqs[i],
1333                        hctx_idx, i);
1334        }
1335    }
1336
1337    while (!list_empty(&tags->page_list)) {
1338        page = list_first_entry(&tags->page_list, struct page, lru);
1339        list_del_init(&page->lru);
1340        __free_pages(page, page->private);
1341    }
1342
1343    kfree(tags->rqs);
1344
1345    blk_mq_free_tags(tags);
1346}
1347
1348static size_t order_to_size(unsigned int order)
1349{
1350    return (size_t)PAGE_SIZE << order;
1351}
1352
1353static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1354        unsigned int hctx_idx)
1355{
1356    struct blk_mq_tags *tags;
1357    unsigned int i, j, entries_per_page, max_order = 4;
1358    size_t rq_size, left;
1359
1360    tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1361                set->numa_node);
1362    if (!tags)
1363        return NULL;
1364
1365    INIT_LIST_HEAD(&tags->page_list);
1366
1367    tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *),
1368                    GFP_KERNEL, set->numa_node);
1369    if (!tags->rqs) {
1370        blk_mq_free_tags(tags);
1371        return NULL;
1372    }
1373
1374    /*
1375     * rq_size is the size of the request plus driver payload, rounded
1376     * to the cacheline size
1377     */
1378    rq_size = round_up(sizeof(struct request) + set->cmd_size,
1379                cache_line_size());
1380    left = rq_size * set->queue_depth;
1381
1382    for (i = 0; i < set->queue_depth; ) {
1383        int this_order = max_order;
1384        struct page *page;
1385        int to_do;
1386        void *p;
1387
1388        while (left < order_to_size(this_order - 1) && this_order)
1389            this_order--;
1390
1391        do {
1392            page = alloc_pages_node(set->numa_node, GFP_KERNEL,
1393                        this_order);
1394            if (page)
1395                break;
1396            if (!this_order--)
1397                break;
1398            if (order_to_size(this_order) < rq_size)
1399                break;
1400        } while (1);
1401
1402        if (!page)
1403            goto fail;
1404
1405        page->private = this_order;
1406        list_add_tail(&page->lru, &tags->page_list);
1407
1408        p = page_address(page);
1409        entries_per_page = order_to_size(this_order) / rq_size;
1410        to_do = min(entries_per_page, set->queue_depth - i);
1411        left -= to_do * rq_size;
1412        for (j = 0; j < to_do; j++) {
1413            tags->rqs[i] = p;
1414            if (set->ops->init_request) {
1415                if (set->ops->init_request(set->driver_data,
1416                        tags->rqs[i], hctx_idx, i,
1417                        set->numa_node))
1418                    goto fail;
1419            }
1420
1421            p += rq_size;
1422            i++;
1423        }
1424    }
1425
1426    return tags;
1427
1428fail:
1429    pr_warn("%s: failed to allocate requests\n", __func__);
1430    blk_mq_free_rq_map(set, tags, hctx_idx);
1431    return NULL;
1432}
1433
1434static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1435{
1436    kfree(bitmap->map);
1437}
1438
1439static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1440{
1441    unsigned int bpw = 8, total, num_maps, i;
1442
1443    bitmap->bits_per_word = bpw;
1444
1445    num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1446    bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1447                    GFP_KERNEL, node);
1448    if (!bitmap->map)
1449        return -ENOMEM;
1450
1451    bitmap->map_size = num_maps;
1452
1453    total = nr_cpu_ids;
1454    for (i = 0; i < num_maps; i++) {
1455        bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1456        total -= bitmap->map[i].depth;
1457    }
1458
1459    return 0;
1460}
1461
1462static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1463{
1464    struct request_queue *q = hctx->queue;
1465    struct blk_mq_ctx *ctx;
1466    LIST_HEAD(tmp);
1467
1468    /*
1469     * Move ctx entries to new CPU, if this one is going away.
1470     */
1471    ctx = __blk_mq_get_ctx(q, cpu);
1472
1473    spin_lock(&ctx->lock);
1474    if (!list_empty(&ctx->rq_list)) {
1475        list_splice_init(&ctx->rq_list, &tmp);
1476        blk_mq_hctx_clear_pending(hctx, ctx);
1477    }
1478    spin_unlock(&ctx->lock);
1479
1480    if (list_empty(&tmp))
1481        return NOTIFY_OK;
1482
1483    ctx = blk_mq_get_ctx(q);
1484    spin_lock(&ctx->lock);
1485
1486    while (!list_empty(&tmp)) {
1487        struct request *rq;
1488
1489        rq = list_first_entry(&tmp, struct request, queuelist);
1490        rq->mq_ctx = ctx;
1491        list_move_tail(&rq->queuelist, &ctx->rq_list);
1492    }
1493
1494    hctx = q->mq_ops->map_queue(q, ctx->cpu);
1495    blk_mq_hctx_mark_pending(hctx, ctx);
1496
1497    spin_unlock(&ctx->lock);
1498
1499    blk_mq_run_hw_queue(hctx, true);
1500    blk_mq_put_ctx(ctx);
1501    return NOTIFY_OK;
1502}
1503
1504static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1505{
1506    struct request_queue *q = hctx->queue;
1507    struct blk_mq_tag_set *set = q->tag_set;
1508
1509    if (set->tags[hctx->queue_num])
1510        return NOTIFY_OK;
1511
1512    set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1513    if (!set->tags[hctx->queue_num])
1514        return NOTIFY_STOP;
1515
1516    hctx->tags = set->tags[hctx->queue_num];
1517    return NOTIFY_OK;
1518}
1519
1520static int blk_mq_hctx_notify(void *data, unsigned long action,
1521                  unsigned int cpu)
1522{
1523    struct blk_mq_hw_ctx *hctx = data;
1524
1525    if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1526        return blk_mq_hctx_cpu_offline(hctx, cpu);
1527    else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1528        return blk_mq_hctx_cpu_online(hctx, cpu);
1529
1530    return NOTIFY_OK;
1531}
1532
1533static void blk_mq_exit_hw_queues(struct request_queue *q,
1534        struct blk_mq_tag_set *set, int nr_queue)
1535{
1536    struct blk_mq_hw_ctx *hctx;
1537    unsigned int i;
1538
1539    queue_for_each_hw_ctx(q, hctx, i) {
1540        if (i == nr_queue)
1541            break;
1542
1543        blk_mq_tag_idle(hctx);
1544
1545        if (set->ops->exit_hctx)
1546            set->ops->exit_hctx(hctx, i);
1547
1548        blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1549        kfree(hctx->ctxs);
1550        blk_mq_free_bitmap(&hctx->ctx_map);
1551    }
1552
1553}
1554
1555static void blk_mq_free_hw_queues(struct request_queue *q,
1556        struct blk_mq_tag_set *set)
1557{
1558    struct blk_mq_hw_ctx *hctx;
1559    unsigned int i;
1560
1561    queue_for_each_hw_ctx(q, hctx, i) {
1562        free_cpumask_var(hctx->cpumask);
1563        kfree(hctx);
1564    }
1565}
1566
1567static int blk_mq_init_hw_queues(struct request_queue *q,
1568        struct blk_mq_tag_set *set)
1569{
1570    struct blk_mq_hw_ctx *hctx;
1571    unsigned int i;
1572
1573    /*
1574     * Initialize hardware queues
1575     */
1576    queue_for_each_hw_ctx(q, hctx, i) {
1577        int node;
1578
1579        node = hctx->numa_node;
1580        if (node == NUMA_NO_NODE)
1581            node = hctx->numa_node = set->numa_node;
1582
1583        INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1584        INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1585        spin_lock_init(&hctx->lock);
1586        INIT_LIST_HEAD(&hctx->dispatch);
1587        hctx->queue = q;
1588        hctx->queue_num = i;
1589        hctx->flags = set->flags;
1590        hctx->cmd_size = set->cmd_size;
1591
1592        blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1593                        blk_mq_hctx_notify, hctx);
1594        blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1595
1596        hctx->tags = set->tags[i];
1597
1598        /*
1599         * Allocate space for all possible cpus to avoid allocation in
1600         * runtime
1601         */
1602        hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1603                        GFP_KERNEL, node);
1604        if (!hctx->ctxs)
1605            break;
1606
1607        if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1608            break;
1609
1610        hctx->nr_ctx = 0;
1611
1612        if (set->ops->init_hctx &&
1613            set->ops->init_hctx(hctx, set->driver_data, i))
1614            break;
1615    }
1616
1617    if (i == q->nr_hw_queues)
1618        return 0;
1619
1620    /*
1621     * Init failed
1622     */
1623    blk_mq_exit_hw_queues(q, set, i);
1624
1625    return 1;
1626}
1627
1628static void blk_mq_init_cpu_queues(struct request_queue *q,
1629                   unsigned int nr_hw_queues)
1630{
1631    unsigned int i;
1632
1633    for_each_possible_cpu(i) {
1634        struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1635        struct blk_mq_hw_ctx *hctx;
1636
1637        memset(__ctx, 0, sizeof(*__ctx));
1638        __ctx->cpu = i;
1639        spin_lock_init(&__ctx->lock);
1640        INIT_LIST_HEAD(&__ctx->rq_list);
1641        __ctx->queue = q;
1642
1643        /* If the cpu isn't online, the cpu is mapped to first hctx */
1644        if (!cpu_online(i))
1645            continue;
1646
1647        hctx = q->mq_ops->map_queue(q, i);
1648        cpumask_set_cpu(i, hctx->cpumask);
1649        hctx->nr_ctx++;
1650
1651        /*
1652         * Set local node, IFF we have more than one hw queue. If
1653         * not, we remain on the home node of the device
1654         */
1655        if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1656            hctx->numa_node = cpu_to_node(i);
1657    }
1658}
1659
1660static void blk_mq_map_swqueue(struct request_queue *q)
1661{
1662    unsigned int i;
1663    struct blk_mq_hw_ctx *hctx;
1664    struct blk_mq_ctx *ctx;
1665
1666    queue_for_each_hw_ctx(q, hctx, i) {
1667        cpumask_clear(hctx->cpumask);
1668        hctx->nr_ctx = 0;
1669    }
1670
1671    /*
1672     * Map software to hardware queues
1673     */
1674    queue_for_each_ctx(q, ctx, i) {
1675        /* If the cpu isn't online, the cpu is mapped to first hctx */
1676        if (!cpu_online(i))
1677            continue;
1678
1679        hctx = q->mq_ops->map_queue(q, i);
1680        cpumask_set_cpu(i, hctx->cpumask);
1681        ctx->index_hw = hctx->nr_ctx;
1682        hctx->ctxs[hctx->nr_ctx++] = ctx;
1683    }
1684
1685    queue_for_each_hw_ctx(q, hctx, i) {
1686        /*
1687         * If not software queues are mapped to this hardware queue,
1688         * disable it and free the request entries
1689         */
1690        if (!hctx->nr_ctx) {
1691            struct blk_mq_tag_set *set = q->tag_set;
1692
1693            if (set->tags[i]) {
1694                blk_mq_free_rq_map(set, set->tags[i], i);
1695                set->tags[i] = NULL;
1696                hctx->tags = NULL;
1697            }
1698            continue;
1699        }
1700
1701        /*
1702         * Initialize batch roundrobin counts
1703         */
1704        hctx->next_cpu = cpumask_first(hctx->cpumask);
1705        hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1706    }
1707}
1708
1709static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1710{
1711    struct blk_mq_hw_ctx *hctx;
1712    struct request_queue *q;
1713    bool shared;
1714    int i;
1715
1716    if (set->tag_list.next == set->tag_list.prev)
1717        shared = false;
1718    else
1719        shared = true;
1720
1721    list_for_each_entry(q, &set->tag_list, tag_set_list) {
1722        blk_mq_freeze_queue(q);
1723
1724        queue_for_each_hw_ctx(q, hctx, i) {
1725            if (shared)
1726                hctx->flags |= BLK_MQ_F_TAG_SHARED;
1727            else
1728                hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1729        }
1730        blk_mq_unfreeze_queue(q);
1731    }
1732}
1733
1734static void blk_mq_del_queue_tag_set(struct request_queue *q)
1735{
1736    struct blk_mq_tag_set *set = q->tag_set;
1737
1738    blk_mq_freeze_queue(q);
1739
1740    mutex_lock(&set->tag_list_lock);
1741    list_del_init(&q->tag_set_list);
1742    blk_mq_update_tag_set_depth(set);
1743    mutex_unlock(&set->tag_list_lock);
1744
1745    blk_mq_unfreeze_queue(q);
1746}
1747
1748static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1749                     struct request_queue *q)
1750{
1751    q->tag_set = set;
1752
1753    mutex_lock(&set->tag_list_lock);
1754    list_add_tail(&q->tag_set_list, &set->tag_list);
1755    blk_mq_update_tag_set_depth(set);
1756    mutex_unlock(&set->tag_list_lock);
1757}
1758
1759struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1760{
1761    struct blk_mq_hw_ctx **hctxs;
1762    struct blk_mq_ctx __percpu *ctx;
1763    struct request_queue *q;
1764    unsigned int *map;
1765    int i;
1766
1767    ctx = alloc_percpu(struct blk_mq_ctx);
1768    if (!ctx)
1769        return ERR_PTR(-ENOMEM);
1770
1771    hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1772            set->numa_node);
1773
1774    if (!hctxs)
1775        goto err_percpu;
1776
1777    map = blk_mq_make_queue_map(set);
1778    if (!map)
1779        goto err_map;
1780
1781    for (i = 0; i < set->nr_hw_queues; i++) {
1782        int node = blk_mq_hw_queue_to_node(map, i);
1783
1784        hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1785                    GFP_KERNEL, node);
1786        if (!hctxs[i])
1787            goto err_hctxs;
1788
1789        if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1790            goto err_hctxs;
1791
1792        atomic_set(&hctxs[i]->nr_active, 0);
1793        hctxs[i]->numa_node = node;
1794        hctxs[i]->queue_num = i;
1795    }
1796
1797    q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1798    if (!q)
1799        goto err_hctxs;
1800
1801    if (percpu_counter_init(&q->mq_usage_counter, 0))
1802        goto err_map;
1803
1804    setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1805    blk_queue_rq_timeout(q, 30000);
1806
1807    q->nr_queues = nr_cpu_ids;
1808    q->nr_hw_queues = set->nr_hw_queues;
1809    q->mq_map = map;
1810
1811    q->queue_ctx = ctx;
1812    q->queue_hw_ctx = hctxs;
1813
1814    q->mq_ops = set->ops;
1815    q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1816
1817    if (!(set->flags & BLK_MQ_F_SG_MERGE))
1818        q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1819
1820    q->sg_reserved_size = INT_MAX;
1821
1822    INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1823    INIT_LIST_HEAD(&q->requeue_list);
1824    spin_lock_init(&q->requeue_lock);
1825
1826    if (q->nr_hw_queues > 1)
1827        blk_queue_make_request(q, blk_mq_make_request);
1828    else
1829        blk_queue_make_request(q, blk_sq_make_request);
1830
1831    blk_queue_rq_timed_out(q, blk_mq_rq_timed_out);
1832    if (set->timeout)
1833        blk_queue_rq_timeout(q, set->timeout);
1834
1835    /*
1836     * Do this after blk_queue_make_request() overrides it...
1837     */
1838    q->nr_requests = set->queue_depth;
1839
1840    if (set->ops->complete)
1841        blk_queue_softirq_done(q, set->ops->complete);
1842
1843    blk_mq_init_flush(q);
1844    blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1845
1846    q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1847                set->cmd_size, cache_line_size()),
1848                GFP_KERNEL);
1849    if (!q->flush_rq)
1850        goto err_hw;
1851
1852    if (blk_mq_init_hw_queues(q, set))
1853        goto err_flush_rq;
1854
1855    mutex_lock(&all_q_mutex);
1856    list_add_tail(&q->all_q_node, &all_q_list);
1857    mutex_unlock(&all_q_mutex);
1858
1859    blk_mq_add_queue_tag_set(set, q);
1860
1861    blk_mq_map_swqueue(q);
1862
1863    return q;
1864
1865err_flush_rq:
1866    kfree(q->flush_rq);
1867err_hw:
1868    blk_cleanup_queue(q);
1869err_hctxs:
1870    kfree(map);
1871    for (i = 0; i < set->nr_hw_queues; i++) {
1872        if (!hctxs[i])
1873            break;
1874        free_cpumask_var(hctxs[i]->cpumask);
1875        kfree(hctxs[i]);
1876    }
1877err_map:
1878    kfree(hctxs);
1879err_percpu:
1880    free_percpu(ctx);
1881    return ERR_PTR(-ENOMEM);
1882}
1883EXPORT_SYMBOL(blk_mq_init_queue);
1884
1885void blk_mq_free_queue(struct request_queue *q)
1886{
1887    struct blk_mq_tag_set *set = q->tag_set;
1888
1889    blk_mq_del_queue_tag_set(q);
1890
1891    blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1892    blk_mq_free_hw_queues(q, set);
1893
1894    percpu_counter_destroy(&q->mq_usage_counter);
1895
1896    free_percpu(q->queue_ctx);
1897    kfree(q->queue_hw_ctx);
1898    kfree(q->mq_map);
1899
1900    q->queue_ctx = NULL;
1901    q->queue_hw_ctx = NULL;
1902    q->mq_map = NULL;
1903
1904    mutex_lock(&all_q_mutex);
1905    list_del_init(&q->all_q_node);
1906    mutex_unlock(&all_q_mutex);
1907}
1908
1909/* Basically redo blk_mq_init_queue with queue frozen */
1910static void blk_mq_queue_reinit(struct request_queue *q)
1911{
1912    blk_mq_freeze_queue(q);
1913
1914    blk_mq_sysfs_unregister(q);
1915
1916    blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1917
1918    /*
1919     * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1920     * we should change hctx numa_node according to new topology (this
1921     * involves free and re-allocate memory, worthy doing?)
1922     */
1923
1924    blk_mq_map_swqueue(q);
1925
1926    blk_mq_sysfs_register(q);
1927
1928    blk_mq_unfreeze_queue(q);
1929}
1930
1931static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1932                      unsigned long action, void *hcpu)
1933{
1934    struct request_queue *q;
1935
1936    /*
1937     * Before new mappings are established, hotadded cpu might already
1938     * start handling requests. This doesn't break anything as we map
1939     * offline CPUs to first hardware queue. We will re-init the queue
1940     * below to get optimal settings.
1941     */
1942    if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1943        action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1944        return NOTIFY_OK;
1945
1946    mutex_lock(&all_q_mutex);
1947    list_for_each_entry(q, &all_q_list, all_q_node)
1948        blk_mq_queue_reinit(q);
1949    mutex_unlock(&all_q_mutex);
1950    return NOTIFY_OK;
1951}
1952
1953/*
1954 * Alloc a tag set to be associated with one or more request queues.
1955 * May fail with EINVAL for various error conditions. May adjust the
1956 * requested depth down, if if it too large. In that case, the set
1957 * value will be stored in set->queue_depth.
1958 */
1959int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
1960{
1961    int i;
1962
1963    if (!set->nr_hw_queues)
1964        return -EINVAL;
1965    if (!set->queue_depth)
1966        return -EINVAL;
1967    if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
1968        return -EINVAL;
1969
1970    if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
1971        return -EINVAL;
1972
1973    if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
1974        pr_info("blk-mq: reduced tag depth to %u\n",
1975            BLK_MQ_MAX_DEPTH);
1976        set->queue_depth = BLK_MQ_MAX_DEPTH;
1977    }
1978
1979    set->tags = kmalloc_node(set->nr_hw_queues *
1980                 sizeof(struct blk_mq_tags *),
1981                 GFP_KERNEL, set->numa_node);
1982    if (!set->tags)
1983        goto out;
1984
1985    for (i = 0; i < set->nr_hw_queues; i++) {
1986        set->tags[i] = blk_mq_init_rq_map(set, i);
1987        if (!set->tags[i])
1988            goto out_unwind;
1989    }
1990
1991    mutex_init(&set->tag_list_lock);
1992    INIT_LIST_HEAD(&set->tag_list);
1993
1994    return 0;
1995
1996out_unwind:
1997    while (--i >= 0)
1998        blk_mq_free_rq_map(set, set->tags[i], i);
1999out:
2000    return -ENOMEM;
2001}
2002EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2003
2004void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2005{
2006    int i;
2007
2008    for (i = 0; i < set->nr_hw_queues; i++) {
2009        if (set->tags[i])
2010            blk_mq_free_rq_map(set, set->tags[i], i);
2011    }
2012
2013    kfree(set->tags);
2014}
2015EXPORT_SYMBOL(blk_mq_free_tag_set);
2016
2017int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2018{
2019    struct blk_mq_tag_set *set = q->tag_set;
2020    struct blk_mq_hw_ctx *hctx;
2021    int i, ret;
2022
2023    if (!set || nr > set->queue_depth)
2024        return -EINVAL;
2025
2026    ret = 0;
2027    queue_for_each_hw_ctx(q, hctx, i) {
2028        ret = blk_mq_tag_update_depth(hctx->tags, nr);
2029        if (ret)
2030            break;
2031    }
2032
2033    if (!ret)
2034        q->nr_requests = nr;
2035
2036    return ret;
2037}
2038
2039void blk_mq_disable_hotplug(void)
2040{
2041    mutex_lock(&all_q_mutex);
2042}
2043
2044void blk_mq_enable_hotplug(void)
2045{
2046    mutex_unlock(&all_q_mutex);
2047}
2048
2049static int __init blk_mq_init(void)
2050{
2051    blk_mq_cpu_init();
2052
2053    /* Must be called after percpu_counter_hotcpu_callback() */
2054    hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
2055
2056    return 0;
2057}
2058subsys_initcall(blk_mq_init);
2059

Archive Download this file



interactive