linux/block/blk-mq-sched.c

680 lines
18 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0
/*
* blk-mq scheduling framework
*
* Copyright (C) 2016 Jens Axboe
*/
#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/blk-mq.h>
blk-mq: support batching dispatch in case of io More and more drivers want to get batching requests queued from block layer, such as mmc, and tcp based storage drivers. Also current in-tree users have virtio-scsi, virtio-blk and nvme. For none, we already support batching dispatch. But for io scheduler, every time we just take one request from scheduler and pass the single request to blk_mq_dispatch_rq_list(). This way makes batching dispatch not possible when io scheduler is applied. One reason is that we don't want to hurt sequential IO performance, becasue IO merge chance is reduced if more requests are dequeued from scheduler queue. Try to support batching dispatch for io scheduler by starting with the following simple approach: 1) still make sure we can get budget before dequeueing request 2) use hctx->dispatch_busy to evaluate if queue is busy, if it is busy we fackback to non-batching dispatch, otherwise dequeue as many as possible requests from scheduler, and pass them to blk_mq_dispatch_rq_list(). Wrt. 2), we use similar policy for none, and turns out that SCSI SSD performance got improved much. In future, maybe we can develop more intelligent algorithem for batching dispatch. Baolin has tested this patch and found that MMC performance is improved[3]. [1] https://lore.kernel.org/linux-block/20200512075501.GF1531898@T590/#r [2] https://lore.kernel.org/linux-block/fe6bd8b9-6ed9-b225-f80c-314746133722@grimberg.me/ [3] https://lore.kernel.org/linux-block/CADBw62o9eTQDJ9RvNgEqSpXmg6Xcq=2TxH0Hfxhp29uF2W=TXA@mail.gmail.com/ Signed-off-by: Ming Lei <ming.lei@redhat.com> Tested-by: Baolin Wang <baolin.wang7@gmail.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Cc: Sagi Grimberg <sagi@grimberg.me> Cc: Baolin Wang <baolin.wang7@gmail.com> Cc: Christoph Hellwig <hch@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-30 10:25:01 +00:00
#include <linux/list_sort.h>
#include <trace/events/block.h>
#include "blk.h"
#include "blk-mq.h"
#include "blk-mq-debugfs.h"
#include "blk-mq-sched.h"
#include "blk-mq-tag.h"
#include "blk-wbt.h"
void blk_mq_sched_assign_ioc(struct request *rq)
{
struct request_queue *q = rq->q;
struct io_context *ioc;
struct io_cq *icq;
/*
* May not have an IO context if it's a passthrough request
*/
ioc = current->io_context;
if (!ioc)
return;
spin_lock_irq(&q->queue_lock);
icq = ioc_lookup_icq(ioc, q);
spin_unlock_irq(&q->queue_lock);
if (!icq) {
icq = ioc_create_icq(ioc, q, GFP_ATOMIC);
if (!icq)
return;
}
get_io_context(icq->ioc);
rq->elv.icq = icq;
}
/*
* Mark a hardware queue as needing a restart. For shared queues, maintain
* a count of how many hardware queues are marked for restart.
*/
void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
{
if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
return;
2018-06-25 11:31:48 +00:00
set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
}
EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx);
void __blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
{
2018-06-25 11:31:48 +00:00
clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
blk-mq: order adding requests to hctx->dispatch and checking SCHED_RESTART SCHED_RESTART code path is relied to re-run queue for dispatch requests in hctx->dispatch. Meantime the SCHED_RSTART flag is checked when adding requests to hctx->dispatch. memory barriers have to be used for ordering the following two pair of OPs: 1) adding requests to hctx->dispatch and checking SCHED_RESTART in blk_mq_dispatch_rq_list() 2) clearing SCHED_RESTART and checking if there is request in hctx->dispatch in blk_mq_sched_restart(). Without the added memory barrier, either: 1) blk_mq_sched_restart() may miss requests added to hctx->dispatch meantime blk_mq_dispatch_rq_list() observes SCHED_RESTART, and not run queue in dispatch side or 2) blk_mq_dispatch_rq_list still sees SCHED_RESTART, and not run queue in dispatch side, meantime checking if there is request in hctx->dispatch from blk_mq_sched_restart() is missed. IO hang in ltp/fs_fill test is reported by kernel test robot: https://lkml.org/lkml/2020/7/26/77 Turns out it is caused by the above out-of-order OPs. And the IO hang can't be observed any more after applying this patch. Fixes: bd166ef183c2 ("blk-mq-sched: add framework for MQ capable IO schedulers") Reported-by: kernel test robot <rong.a.chen@intel.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Cc: Bart Van Assche <bvanassche@acm.org> Cc: Christoph Hellwig <hch@lst.de> Cc: David Jeffery <djeffery@redhat.com> Cc: <stable@vger.kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-08-17 10:01:15 +00:00
/*
* Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch)
* in blk_mq_run_hw_queue(). Its pair is the barrier in
* blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART,
* meantime new request added to hctx->dispatch is missed to check in
* blk_mq_run_hw_queue().
*/
smp_mb();
2018-06-25 11:31:48 +00:00
blk_mq_run_hw_queue(hctx, true);
}
static int sched_rq_cmp(void *priv, const struct list_head *a,
const struct list_head *b)
blk-mq: support batching dispatch in case of io More and more drivers want to get batching requests queued from block layer, such as mmc, and tcp based storage drivers. Also current in-tree users have virtio-scsi, virtio-blk and nvme. For none, we already support batching dispatch. But for io scheduler, every time we just take one request from scheduler and pass the single request to blk_mq_dispatch_rq_list(). This way makes batching dispatch not possible when io scheduler is applied. One reason is that we don't want to hurt sequential IO performance, becasue IO merge chance is reduced if more requests are dequeued from scheduler queue. Try to support batching dispatch for io scheduler by starting with the following simple approach: 1) still make sure we can get budget before dequeueing request 2) use hctx->dispatch_busy to evaluate if queue is busy, if it is busy we fackback to non-batching dispatch, otherwise dequeue as many as possible requests from scheduler, and pass them to blk_mq_dispatch_rq_list(). Wrt. 2), we use similar policy for none, and turns out that SCSI SSD performance got improved much. In future, maybe we can develop more intelligent algorithem for batching dispatch. Baolin has tested this patch and found that MMC performance is improved[3]. [1] https://lore.kernel.org/linux-block/20200512075501.GF1531898@T590/#r [2] https://lore.kernel.org/linux-block/fe6bd8b9-6ed9-b225-f80c-314746133722@grimberg.me/ [3] https://lore.kernel.org/linux-block/CADBw62o9eTQDJ9RvNgEqSpXmg6Xcq=2TxH0Hfxhp29uF2W=TXA@mail.gmail.com/ Signed-off-by: Ming Lei <ming.lei@redhat.com> Tested-by: Baolin Wang <baolin.wang7@gmail.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Cc: Sagi Grimberg <sagi@grimberg.me> Cc: Baolin Wang <baolin.wang7@gmail.com> Cc: Christoph Hellwig <hch@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-30 10:25:01 +00:00
{
struct request *rqa = container_of(a, struct request, queuelist);
struct request *rqb = container_of(b, struct request, queuelist);
return rqa->mq_hctx > rqb->mq_hctx;
}
static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list)
{
struct blk_mq_hw_ctx *hctx =
list_first_entry(rq_list, struct request, queuelist)->mq_hctx;
struct request *rq;
LIST_HEAD(hctx_list);
unsigned int count = 0;
list_for_each_entry(rq, rq_list, queuelist) {
if (rq->mq_hctx != hctx) {
list_cut_before(&hctx_list, rq_list, &rq->queuelist);
goto dispatch;
}
count++;
}
list_splice_tail_init(rq_list, &hctx_list);
dispatch:
return blk_mq_dispatch_rq_list(hctx, &hctx_list, count);
blk-mq: support batching dispatch in case of io More and more drivers want to get batching requests queued from block layer, such as mmc, and tcp based storage drivers. Also current in-tree users have virtio-scsi, virtio-blk and nvme. For none, we already support batching dispatch. But for io scheduler, every time we just take one request from scheduler and pass the single request to blk_mq_dispatch_rq_list(). This way makes batching dispatch not possible when io scheduler is applied. One reason is that we don't want to hurt sequential IO performance, becasue IO merge chance is reduced if more requests are dequeued from scheduler queue. Try to support batching dispatch for io scheduler by starting with the following simple approach: 1) still make sure we can get budget before dequeueing request 2) use hctx->dispatch_busy to evaluate if queue is busy, if it is busy we fackback to non-batching dispatch, otherwise dequeue as many as possible requests from scheduler, and pass them to blk_mq_dispatch_rq_list(). Wrt. 2), we use similar policy for none, and turns out that SCSI SSD performance got improved much. In future, maybe we can develop more intelligent algorithem for batching dispatch. Baolin has tested this patch and found that MMC performance is improved[3]. [1] https://lore.kernel.org/linux-block/20200512075501.GF1531898@T590/#r [2] https://lore.kernel.org/linux-block/fe6bd8b9-6ed9-b225-f80c-314746133722@grimberg.me/ [3] https://lore.kernel.org/linux-block/CADBw62o9eTQDJ9RvNgEqSpXmg6Xcq=2TxH0Hfxhp29uF2W=TXA@mail.gmail.com/ Signed-off-by: Ming Lei <ming.lei@redhat.com> Tested-by: Baolin Wang <baolin.wang7@gmail.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Cc: Sagi Grimberg <sagi@grimberg.me> Cc: Baolin Wang <baolin.wang7@gmail.com> Cc: Christoph Hellwig <hch@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-30 10:25:01 +00:00
}
blk-mq: Rerun dispatching in the case of budget contention If ever a thread running blk-mq code tries to get budget and fails it immediately stops doing work and assumes that whenever budget is freed up that queues will be kicked and whatever work the thread was trying to do will be tried again. One path where budget is freed and queues are kicked in the normal case can be seen in scsi_finish_command(). Specifically: - scsi_finish_command() - scsi_device_unbusy() - # Decrement "device_busy", AKA release budget - scsi_io_completion() - scsi_end_request() - blk_mq_run_hw_queues() The above is all well and good. The problem comes up when a thread claims the budget but then releases it without actually dispatching any work. Since we didn't schedule any work we'll never run the path of finishing work / kicking the queues. This isn't often actually a problem which is why this issue has existed for a while and nobody noticed. Specifically we only get into this situation when we unexpectedly found that we weren't going to do any work. Code that later receives new work kicks the queues. All good, right? The problem shows up, however, if timing is just wrong and we hit a race. To see this race let's think about the case where we only have a budget of 1 (only one thread can hold budget). Now imagine that a thread got budget and then decided not to dispatch work. It's about to call put_budget() but then the thread gets context switched out for a long, long time. While in this state, any and all kicks of the queue (like the when we received new work) will be no-ops because nobody can get budget. Finally the thread holding budget gets to run again and returns. All the normal kicks will have been no-ops and we have an I/O stall. As you can see from the above, you need just the right timing to see the race. To start with, the only case it happens if we thought we had work, actually managed to get the budget, but then actually didn't have work. That's pretty rare to start with. Even then, there's usually a very small amount of time between realizing that there's no work and putting the budget. During this small amount of time new work has to come in and the queue kick has to make it all the way to trying to get the budget and fail. It's pretty unlikely. One case where this could have failed is illustrated by an example of threads running blk_mq_do_dispatch_sched(): * Threads A and B both run has_work() at the same time with the same "hctx". Imagine has_work() is exact. There's no lock, so it's OK if Thread A and B both get back true. * Thread B gets interrupted for a long time right after it decides that there is work. Maybe its CPU gets an interrupt and the interrupt handler is slow. * Thread A runs, get budget, dispatches work. * Thread A's work finishes and budget is released. * Thread B finally runs again and gets budget. * Since Thread A already took care of the work and no new work has come in, Thread B will get NULL from dispatch_request(). I believe this is specifically why dispatch_request() is allowed to return NULL in the first place if has_work() must be exact. * Thread B will now be holding the budget and is about to call put_budget(), but hasn't called it yet. * Thread B gets interrupted for a long time (again). Dang interrupts. * Now Thread C (maybe with a different "hctx" but the same queue) comes along and runs blk_mq_do_dispatch_sched(). * Thread C won't do anything because it can't get budget. * Finally Thread B will run again and put the budget without kicking any queues. Even though the example above is with blk_mq_do_dispatch_sched() I believe the race is possible any time someone is holding budget but doesn't do work. Unfortunately, the unlikely has become more likely if you happen to be using the BFQ I/O scheduler. BFQ, by design, sometimes returns "true" for has_work() but then NULL for dispatch_request() and stays in this state for a while (currently up to 9 ms). Suddenly you only need one race to hit, not two races in a row. With my current setup this is easy to reproduce in reboot tests and traces have actually shown that we hit a race similar to the one described above. Note that we only need to fix blk_mq_do_dispatch_sched() and blk_mq_do_dispatch_ctx() and not the other places that put budget. In other cases we know that we have work to do on at least one "hctx" and code already exists to kick that "hctx"'s queue. When that work finally finishes all the queues will be kicked using the normal flow. One last note is that (at least in the SCSI case) budget is shared by all "hctx"s that have the same queue. Thus we need to make sure to kick the whole queue, not just re-run dispatching on a single "hctx". Signed-off-by: Douglas Anderson <dianders@chromium.org> Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-04-20 16:24:53 +00:00
#define BLK_MQ_BUDGET_DELAY 3 /* ms units */
/*
* Only SCSI implements .get_budget and .put_budget, and SCSI restarts
* its queue by itself in its completion handler, so we don't need to
* restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
*
* Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
* be run again. This is necessary to avoid starving flushes.
*/
blk-mq: support batching dispatch in case of io More and more drivers want to get batching requests queued from block layer, such as mmc, and tcp based storage drivers. Also current in-tree users have virtio-scsi, virtio-blk and nvme. For none, we already support batching dispatch. But for io scheduler, every time we just take one request from scheduler and pass the single request to blk_mq_dispatch_rq_list(). This way makes batching dispatch not possible when io scheduler is applied. One reason is that we don't want to hurt sequential IO performance, becasue IO merge chance is reduced if more requests are dequeued from scheduler queue. Try to support batching dispatch for io scheduler by starting with the following simple approach: 1) still make sure we can get budget before dequeueing request 2) use hctx->dispatch_busy to evaluate if queue is busy, if it is busy we fackback to non-batching dispatch, otherwise dequeue as many as possible requests from scheduler, and pass them to blk_mq_dispatch_rq_list(). Wrt. 2), we use similar policy for none, and turns out that SCSI SSD performance got improved much. In future, maybe we can develop more intelligent algorithem for batching dispatch. Baolin has tested this patch and found that MMC performance is improved[3]. [1] https://lore.kernel.org/linux-block/20200512075501.GF1531898@T590/#r [2] https://lore.kernel.org/linux-block/fe6bd8b9-6ed9-b225-f80c-314746133722@grimberg.me/ [3] https://lore.kernel.org/linux-block/CADBw62o9eTQDJ9RvNgEqSpXmg6Xcq=2TxH0Hfxhp29uF2W=TXA@mail.gmail.com/ Signed-off-by: Ming Lei <ming.lei@redhat.com> Tested-by: Baolin Wang <baolin.wang7@gmail.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Cc: Sagi Grimberg <sagi@grimberg.me> Cc: Baolin Wang <baolin.wang7@gmail.com> Cc: Christoph Hellwig <hch@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-30 10:25:01 +00:00
static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
{
struct request_queue *q = hctx->queue;
struct elevator_queue *e = q->elevator;
blk-mq: support batching dispatch in case of io More and more drivers want to get batching requests queued from block layer, such as mmc, and tcp based storage drivers. Also current in-tree users have virtio-scsi, virtio-blk and nvme. For none, we already support batching dispatch. But for io scheduler, every time we just take one request from scheduler and pass the single request to blk_mq_dispatch_rq_list(). This way makes batching dispatch not possible when io scheduler is applied. One reason is that we don't want to hurt sequential IO performance, becasue IO merge chance is reduced if more requests are dequeued from scheduler queue. Try to support batching dispatch for io scheduler by starting with the following simple approach: 1) still make sure we can get budget before dequeueing request 2) use hctx->dispatch_busy to evaluate if queue is busy, if it is busy we fackback to non-batching dispatch, otherwise dequeue as many as possible requests from scheduler, and pass them to blk_mq_dispatch_rq_list(). Wrt. 2), we use similar policy for none, and turns out that SCSI SSD performance got improved much. In future, maybe we can develop more intelligent algorithem for batching dispatch. Baolin has tested this patch and found that MMC performance is improved[3]. [1] https://lore.kernel.org/linux-block/20200512075501.GF1531898@T590/#r [2] https://lore.kernel.org/linux-block/fe6bd8b9-6ed9-b225-f80c-314746133722@grimberg.me/ [3] https://lore.kernel.org/linux-block/CADBw62o9eTQDJ9RvNgEqSpXmg6Xcq=2TxH0Hfxhp29uF2W=TXA@mail.gmail.com/ Signed-off-by: Ming Lei <ming.lei@redhat.com> Tested-by: Baolin Wang <baolin.wang7@gmail.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Cc: Sagi Grimberg <sagi@grimberg.me> Cc: Baolin Wang <baolin.wang7@gmail.com> Cc: Christoph Hellwig <hch@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-30 10:25:01 +00:00
bool multi_hctxs = false, run_queue = false;
bool dispatched = false, busy = false;
unsigned int max_dispatch;
LIST_HEAD(rq_list);
blk-mq: support batching dispatch in case of io More and more drivers want to get batching requests queued from block layer, such as mmc, and tcp based storage drivers. Also current in-tree users have virtio-scsi, virtio-blk and nvme. For none, we already support batching dispatch. But for io scheduler, every time we just take one request from scheduler and pass the single request to blk_mq_dispatch_rq_list(). This way makes batching dispatch not possible when io scheduler is applied. One reason is that we don't want to hurt sequential IO performance, becasue IO merge chance is reduced if more requests are dequeued from scheduler queue. Try to support batching dispatch for io scheduler by starting with the following simple approach: 1) still make sure we can get budget before dequeueing request 2) use hctx->dispatch_busy to evaluate if queue is busy, if it is busy we fackback to non-batching dispatch, otherwise dequeue as many as possible requests from scheduler, and pass them to blk_mq_dispatch_rq_list(). Wrt. 2), we use similar policy for none, and turns out that SCSI SSD performance got improved much. In future, maybe we can develop more intelligent algorithem for batching dispatch. Baolin has tested this patch and found that MMC performance is improved[3]. [1] https://lore.kernel.org/linux-block/20200512075501.GF1531898@T590/#r [2] https://lore.kernel.org/linux-block/fe6bd8b9-6ed9-b225-f80c-314746133722@grimberg.me/ [3] https://lore.kernel.org/linux-block/CADBw62o9eTQDJ9RvNgEqSpXmg6Xcq=2TxH0Hfxhp29uF2W=TXA@mail.gmail.com/ Signed-off-by: Ming Lei <ming.lei@redhat.com> Tested-by: Baolin Wang <baolin.wang7@gmail.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Cc: Sagi Grimberg <sagi@grimberg.me> Cc: Baolin Wang <baolin.wang7@gmail.com> Cc: Christoph Hellwig <hch@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-30 10:25:01 +00:00
int count = 0;
if (hctx->dispatch_busy)
max_dispatch = 1;
else
max_dispatch = hctx->queue->nr_requests;
do {
blk-mq: support batching dispatch in case of io More and more drivers want to get batching requests queued from block layer, such as mmc, and tcp based storage drivers. Also current in-tree users have virtio-scsi, virtio-blk and nvme. For none, we already support batching dispatch. But for io scheduler, every time we just take one request from scheduler and pass the single request to blk_mq_dispatch_rq_list(). This way makes batching dispatch not possible when io scheduler is applied. One reason is that we don't want to hurt sequential IO performance, becasue IO merge chance is reduced if more requests are dequeued from scheduler queue. Try to support batching dispatch for io scheduler by starting with the following simple approach: 1) still make sure we can get budget before dequeueing request 2) use hctx->dispatch_busy to evaluate if queue is busy, if it is busy we fackback to non-batching dispatch, otherwise dequeue as many as possible requests from scheduler, and pass them to blk_mq_dispatch_rq_list(). Wrt. 2), we use similar policy for none, and turns out that SCSI SSD performance got improved much. In future, maybe we can develop more intelligent algorithem for batching dispatch. Baolin has tested this patch and found that MMC performance is improved[3]. [1] https://lore.kernel.org/linux-block/20200512075501.GF1531898@T590/#r [2] https://lore.kernel.org/linux-block/fe6bd8b9-6ed9-b225-f80c-314746133722@grimberg.me/ [3] https://lore.kernel.org/linux-block/CADBw62o9eTQDJ9RvNgEqSpXmg6Xcq=2TxH0Hfxhp29uF2W=TXA@mail.gmail.com/ Signed-off-by: Ming Lei <ming.lei@redhat.com> Tested-by: Baolin Wang <baolin.wang7@gmail.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Cc: Sagi Grimberg <sagi@grimberg.me> Cc: Baolin Wang <baolin.wang7@gmail.com> Cc: Christoph Hellwig <hch@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-30 10:25:01 +00:00
struct request *rq;
int budget_token;
blk-mq: support batching dispatch in case of io More and more drivers want to get batching requests queued from block layer, such as mmc, and tcp based storage drivers. Also current in-tree users have virtio-scsi, virtio-blk and nvme. For none, we already support batching dispatch. But for io scheduler, every time we just take one request from scheduler and pass the single request to blk_mq_dispatch_rq_list(). This way makes batching dispatch not possible when io scheduler is applied. One reason is that we don't want to hurt sequential IO performance, becasue IO merge chance is reduced if more requests are dequeued from scheduler queue. Try to support batching dispatch for io scheduler by starting with the following simple approach: 1) still make sure we can get budget before dequeueing request 2) use hctx->dispatch_busy to evaluate if queue is busy, if it is busy we fackback to non-batching dispatch, otherwise dequeue as many as possible requests from scheduler, and pass them to blk_mq_dispatch_rq_list(). Wrt. 2), we use similar policy for none, and turns out that SCSI SSD performance got improved much. In future, maybe we can develop more intelligent algorithem for batching dispatch. Baolin has tested this patch and found that MMC performance is improved[3]. [1] https://lore.kernel.org/linux-block/20200512075501.GF1531898@T590/#r [2] https://lore.kernel.org/linux-block/fe6bd8b9-6ed9-b225-f80c-314746133722@grimberg.me/ [3] https://lore.kernel.org/linux-block/CADBw62o9eTQDJ9RvNgEqSpXmg6Xcq=2TxH0Hfxhp29uF2W=TXA@mail.gmail.com/ Signed-off-by: Ming Lei <ming.lei@redhat.com> Tested-by: Baolin Wang <baolin.wang7@gmail.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Cc: Sagi Grimberg <sagi@grimberg.me> Cc: Baolin Wang <baolin.wang7@gmail.com> Cc: Christoph Hellwig <hch@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-30 10:25:01 +00:00
if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
break;
if (!list_empty_careful(&hctx->dispatch)) {
blk-mq: support batching dispatch in case of io More and more drivers want to get batching requests queued from block layer, such as mmc, and tcp based storage drivers. Also current in-tree users have virtio-scsi, virtio-blk and nvme. For none, we already support batching dispatch. But for io scheduler, every time we just take one request from scheduler and pass the single request to blk_mq_dispatch_rq_list(). This way makes batching dispatch not possible when io scheduler is applied. One reason is that we don't want to hurt sequential IO performance, becasue IO merge chance is reduced if more requests are dequeued from scheduler queue. Try to support batching dispatch for io scheduler by starting with the following simple approach: 1) still make sure we can get budget before dequeueing request 2) use hctx->dispatch_busy to evaluate if queue is busy, if it is busy we fackback to non-batching dispatch, otherwise dequeue as many as possible requests from scheduler, and pass them to blk_mq_dispatch_rq_list(). Wrt. 2), we use similar policy for none, and turns out that SCSI SSD performance got improved much. In future, maybe we can develop more intelligent algorithem for batching dispatch. Baolin has tested this patch and found that MMC performance is improved[3]. [1] https://lore.kernel.org/linux-block/20200512075501.GF1531898@T590/#r [2] https://lore.kernel.org/linux-block/fe6bd8b9-6ed9-b225-f80c-314746133722@grimberg.me/ [3] https://lore.kernel.org/linux-block/CADBw62o9eTQDJ9RvNgEqSpXmg6Xcq=2TxH0Hfxhp29uF2W=TXA@mail.gmail.com/ Signed-off-by: Ming Lei <ming.lei@redhat.com> Tested-by: Baolin Wang <baolin.wang7@gmail.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Cc: Sagi Grimberg <sagi@grimberg.me> Cc: Baolin Wang <baolin.wang7@gmail.com> Cc: Christoph Hellwig <hch@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-30 10:25:01 +00:00
busy = true;
break;
}
budget_token = blk_mq_get_dispatch_budget(q);
if (budget_token < 0)
break;
rq = e->type->ops.dispatch_request(hctx);
if (!rq) {
blk_mq_put_dispatch_budget(q, budget_token);
blk-mq: Rerun dispatching in the case of budget contention If ever a thread running blk-mq code tries to get budget and fails it immediately stops doing work and assumes that whenever budget is freed up that queues will be kicked and whatever work the thread was trying to do will be tried again. One path where budget is freed and queues are kicked in the normal case can be seen in scsi_finish_command(). Specifically: - scsi_finish_command() - scsi_device_unbusy() - # Decrement "device_busy", AKA release budget - scsi_io_completion() - scsi_end_request() - blk_mq_run_hw_queues() The above is all well and good. The problem comes up when a thread claims the budget but then releases it without actually dispatching any work. Since we didn't schedule any work we'll never run the path of finishing work / kicking the queues. This isn't often actually a problem which is why this issue has existed for a while and nobody noticed. Specifically we only get into this situation when we unexpectedly found that we weren't going to do any work. Code that later receives new work kicks the queues. All good, right? The problem shows up, however, if timing is just wrong and we hit a race. To see this race let's think about the case where we only have a budget of 1 (only one thread can hold budget). Now imagine that a thread got budget and then decided not to dispatch work. It's about to call put_budget() but then the thread gets context switched out for a long, long time. While in this state, any and all kicks of the queue (like the when we received new work) will be no-ops because nobody can get budget. Finally the thread holding budget gets to run again and returns. All the normal kicks will have been no-ops and we have an I/O stall. As you can see from the above, you need just the right timing to see the race. To start with, the only case it happens if we thought we had work, actually managed to get the budget, but then actually didn't have work. That's pretty rare to start with. Even then, there's usually a very small amount of time between realizing that there's no work and putting the budget. During this small amount of time new work has to come in and the queue kick has to make it all the way to trying to get the budget and fail. It's pretty unlikely. One case where this could have failed is illustrated by an example of threads running blk_mq_do_dispatch_sched(): * Threads A and B both run has_work() at the same time with the same "hctx". Imagine has_work() is exact. There's no lock, so it's OK if Thread A and B both get back true. * Thread B gets interrupted for a long time right after it decides that there is work. Maybe its CPU gets an interrupt and the interrupt handler is slow. * Thread A runs, get budget, dispatches work. * Thread A's work finishes and budget is released. * Thread B finally runs again and gets budget. * Since Thread A already took care of the work and no new work has come in, Thread B will get NULL from dispatch_request(). I believe this is specifically why dispatch_request() is allowed to return NULL in the first place if has_work() must be exact. * Thread B will now be holding the budget and is about to call put_budget(), but hasn't called it yet. * Thread B gets interrupted for a long time (again). Dang interrupts. * Now Thread C (maybe with a different "hctx" but the same queue) comes along and runs blk_mq_do_dispatch_sched(). * Thread C won't do anything because it can't get budget. * Finally Thread B will run again and put the budget without kicking any queues. Even though the example above is with blk_mq_do_dispatch_sched() I believe the race is possible any time someone is holding budget but doesn't do work. Unfortunately, the unlikely has become more likely if you happen to be using the BFQ I/O scheduler. BFQ, by design, sometimes returns "true" for has_work() but then NULL for dispatch_request() and stays in this state for a while (currently up to 9 ms). Suddenly you only need one race to hit, not two races in a row. With my current setup this is easy to reproduce in reboot tests and traces have actually shown that we hit a race similar to the one described above. Note that we only need to fix blk_mq_do_dispatch_sched() and blk_mq_do_dispatch_ctx() and not the other places that put budget. In other cases we know that we have work to do on at least one "hctx" and code already exists to kick that "hctx"'s queue. When that work finally finishes all the queues will be kicked using the normal flow. One last note is that (at least in the SCSI case) budget is shared by all "hctx"s that have the same queue. Thus we need to make sure to kick the whole queue, not just re-run dispatching on a single "hctx". Signed-off-by: Douglas Anderson <dianders@chromium.org> Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-04-20 16:24:53 +00:00
/*
* We're releasing without dispatching. Holding the
* budget could have blocked any "hctx"s with the
* same queue and if we didn't dispatch then there's
* no guarantee anyone will kick the queue. Kick it
* ourselves.
*/
blk-mq: support batching dispatch in case of io More and more drivers want to get batching requests queued from block layer, such as mmc, and tcp based storage drivers. Also current in-tree users have virtio-scsi, virtio-blk and nvme. For none, we already support batching dispatch. But for io scheduler, every time we just take one request from scheduler and pass the single request to blk_mq_dispatch_rq_list(). This way makes batching dispatch not possible when io scheduler is applied. One reason is that we don't want to hurt sequential IO performance, becasue IO merge chance is reduced if more requests are dequeued from scheduler queue. Try to support batching dispatch for io scheduler by starting with the following simple approach: 1) still make sure we can get budget before dequeueing request 2) use hctx->dispatch_busy to evaluate if queue is busy, if it is busy we fackback to non-batching dispatch, otherwise dequeue as many as possible requests from scheduler, and pass them to blk_mq_dispatch_rq_list(). Wrt. 2), we use similar policy for none, and turns out that SCSI SSD performance got improved much. In future, maybe we can develop more intelligent algorithem for batching dispatch. Baolin has tested this patch and found that MMC performance is improved[3]. [1] https://lore.kernel.org/linux-block/20200512075501.GF1531898@T590/#r [2] https://lore.kernel.org/linux-block/fe6bd8b9-6ed9-b225-f80c-314746133722@grimberg.me/ [3] https://lore.kernel.org/linux-block/CADBw62o9eTQDJ9RvNgEqSpXmg6Xcq=2TxH0Hfxhp29uF2W=TXA@mail.gmail.com/ Signed-off-by: Ming Lei <ming.lei@redhat.com> Tested-by: Baolin Wang <baolin.wang7@gmail.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Cc: Sagi Grimberg <sagi@grimberg.me> Cc: Baolin Wang <baolin.wang7@gmail.com> Cc: Christoph Hellwig <hch@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-30 10:25:01 +00:00
run_queue = true;
break;
}
blk_mq_set_rq_budget_token(rq, budget_token);
/*
* Now this rq owns the budget which has to be released
* if this rq won't be queued to driver via .queue_rq()
* in blk_mq_dispatch_rq_list().
*/
blk-mq: support batching dispatch in case of io More and more drivers want to get batching requests queued from block layer, such as mmc, and tcp based storage drivers. Also current in-tree users have virtio-scsi, virtio-blk and nvme. For none, we already support batching dispatch. But for io scheduler, every time we just take one request from scheduler and pass the single request to blk_mq_dispatch_rq_list(). This way makes batching dispatch not possible when io scheduler is applied. One reason is that we don't want to hurt sequential IO performance, becasue IO merge chance is reduced if more requests are dequeued from scheduler queue. Try to support batching dispatch for io scheduler by starting with the following simple approach: 1) still make sure we can get budget before dequeueing request 2) use hctx->dispatch_busy to evaluate if queue is busy, if it is busy we fackback to non-batching dispatch, otherwise dequeue as many as possible requests from scheduler, and pass them to blk_mq_dispatch_rq_list(). Wrt. 2), we use similar policy for none, and turns out that SCSI SSD performance got improved much. In future, maybe we can develop more intelligent algorithem for batching dispatch. Baolin has tested this patch and found that MMC performance is improved[3]. [1] https://lore.kernel.org/linux-block/20200512075501.GF1531898@T590/#r [2] https://lore.kernel.org/linux-block/fe6bd8b9-6ed9-b225-f80c-314746133722@grimberg.me/ [3] https://lore.kernel.org/linux-block/CADBw62o9eTQDJ9RvNgEqSpXmg6Xcq=2TxH0Hfxhp29uF2W=TXA@mail.gmail.com/ Signed-off-by: Ming Lei <ming.lei@redhat.com> Tested-by: Baolin Wang <baolin.wang7@gmail.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Cc: Sagi Grimberg <sagi@grimberg.me> Cc: Baolin Wang <baolin.wang7@gmail.com> Cc: Christoph Hellwig <hch@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-30 10:25:01 +00:00
list_add_tail(&rq->queuelist, &rq_list);
block: Do not pull requests from the scheduler when we cannot dispatch them Provided the device driver does not implement dispatch budget accounting (which only SCSI does) the loop in __blk_mq_do_dispatch_sched() pulls requests from the IO scheduler as long as it is willing to give out any. That defeats scheduling heuristics inside the scheduler by creating false impression that the device can take more IO when it in fact cannot. For example with BFQ IO scheduler on top of virtio-blk device setting blkio cgroup weight has barely any impact on observed throughput of async IO because __blk_mq_do_dispatch_sched() always sucks out all the IO queued in BFQ. BFQ first submits IO from higher weight cgroups but when that is all dispatched, it will give out IO of lower weight cgroups as well. And then we have to wait for all this IO to be dispatched to the disk (which means lot of it actually has to complete) before the IO scheduler is queried again for dispatching more requests. This completely destroys any service differentiation. So grab request tag for a request pulled out of the IO scheduler already in __blk_mq_do_dispatch_sched() and do not pull any more requests if we cannot get it because we are unlikely to be able to dispatch it. That way only single request is going to wait in the dispatch list for some tag to free. Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jan Kara <jack@suse.cz> Link: https://lore.kernel.org/r/20210603104721.6309-1-jack@suse.cz Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-03 10:47:21 +00:00
count++;
blk-mq: support batching dispatch in case of io More and more drivers want to get batching requests queued from block layer, such as mmc, and tcp based storage drivers. Also current in-tree users have virtio-scsi, virtio-blk and nvme. For none, we already support batching dispatch. But for io scheduler, every time we just take one request from scheduler and pass the single request to blk_mq_dispatch_rq_list(). This way makes batching dispatch not possible when io scheduler is applied. One reason is that we don't want to hurt sequential IO performance, becasue IO merge chance is reduced if more requests are dequeued from scheduler queue. Try to support batching dispatch for io scheduler by starting with the following simple approach: 1) still make sure we can get budget before dequeueing request 2) use hctx->dispatch_busy to evaluate if queue is busy, if it is busy we fackback to non-batching dispatch, otherwise dequeue as many as possible requests from scheduler, and pass them to blk_mq_dispatch_rq_list(). Wrt. 2), we use similar policy for none, and turns out that SCSI SSD performance got improved much. In future, maybe we can develop more intelligent algorithem for batching dispatch. Baolin has tested this patch and found that MMC performance is improved[3]. [1] https://lore.kernel.org/linux-block/20200512075501.GF1531898@T590/#r [2] https://lore.kernel.org/linux-block/fe6bd8b9-6ed9-b225-f80c-314746133722@grimberg.me/ [3] https://lore.kernel.org/linux-block/CADBw62o9eTQDJ9RvNgEqSpXmg6Xcq=2TxH0Hfxhp29uF2W=TXA@mail.gmail.com/ Signed-off-by: Ming Lei <ming.lei@redhat.com> Tested-by: Baolin Wang <baolin.wang7@gmail.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Cc: Sagi Grimberg <sagi@grimberg.me> Cc: Baolin Wang <baolin.wang7@gmail.com> Cc: Christoph Hellwig <hch@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-30 10:25:01 +00:00
if (rq->mq_hctx != hctx)
multi_hctxs = true;
block: Do not pull requests from the scheduler when we cannot dispatch them Provided the device driver does not implement dispatch budget accounting (which only SCSI does) the loop in __blk_mq_do_dispatch_sched() pulls requests from the IO scheduler as long as it is willing to give out any. That defeats scheduling heuristics inside the scheduler by creating false impression that the device can take more IO when it in fact cannot. For example with BFQ IO scheduler on top of virtio-blk device setting blkio cgroup weight has barely any impact on observed throughput of async IO because __blk_mq_do_dispatch_sched() always sucks out all the IO queued in BFQ. BFQ first submits IO from higher weight cgroups but when that is all dispatched, it will give out IO of lower weight cgroups as well. And then we have to wait for all this IO to be dispatched to the disk (which means lot of it actually has to complete) before the IO scheduler is queried again for dispatching more requests. This completely destroys any service differentiation. So grab request tag for a request pulled out of the IO scheduler already in __blk_mq_do_dispatch_sched() and do not pull any more requests if we cannot get it because we are unlikely to be able to dispatch it. That way only single request is going to wait in the dispatch list for some tag to free. Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jan Kara <jack@suse.cz> Link: https://lore.kernel.org/r/20210603104721.6309-1-jack@suse.cz Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-03 10:47:21 +00:00
/*
* If we cannot get tag for the request, stop dequeueing
* requests from the IO scheduler. We are unlikely to be able
* to submit them anyway and it creates false impression for
* scheduling heuristics that the device can take more IO.
*/
if (!blk_mq_get_driver_tag(rq))
break;
} while (count < max_dispatch);
blk-mq: support batching dispatch in case of io More and more drivers want to get batching requests queued from block layer, such as mmc, and tcp based storage drivers. Also current in-tree users have virtio-scsi, virtio-blk and nvme. For none, we already support batching dispatch. But for io scheduler, every time we just take one request from scheduler and pass the single request to blk_mq_dispatch_rq_list(). This way makes batching dispatch not possible when io scheduler is applied. One reason is that we don't want to hurt sequential IO performance, becasue IO merge chance is reduced if more requests are dequeued from scheduler queue. Try to support batching dispatch for io scheduler by starting with the following simple approach: 1) still make sure we can get budget before dequeueing request 2) use hctx->dispatch_busy to evaluate if queue is busy, if it is busy we fackback to non-batching dispatch, otherwise dequeue as many as possible requests from scheduler, and pass them to blk_mq_dispatch_rq_list(). Wrt. 2), we use similar policy for none, and turns out that SCSI SSD performance got improved much. In future, maybe we can develop more intelligent algorithem for batching dispatch. Baolin has tested this patch and found that MMC performance is improved[3]. [1] https://lore.kernel.org/linux-block/20200512075501.GF1531898@T590/#r [2] https://lore.kernel.org/linux-block/fe6bd8b9-6ed9-b225-f80c-314746133722@grimberg.me/ [3] https://lore.kernel.org/linux-block/CADBw62o9eTQDJ9RvNgEqSpXmg6Xcq=2TxH0Hfxhp29uF2W=TXA@mail.gmail.com/ Signed-off-by: Ming Lei <ming.lei@redhat.com> Tested-by: Baolin Wang <baolin.wang7@gmail.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Cc: Sagi Grimberg <sagi@grimberg.me> Cc: Baolin Wang <baolin.wang7@gmail.com> Cc: Christoph Hellwig <hch@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-30 10:25:01 +00:00
if (!count) {
if (run_queue)
blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
} else if (multi_hctxs) {
/*
* Requests from different hctx may be dequeued from some
* schedulers, such as bfq and deadline.
*
* Sort the requests in the list according to their hctx,
* dispatch batching requests from same hctx at a time.
*/
list_sort(NULL, &rq_list, sched_rq_cmp);
do {
dispatched |= blk_mq_dispatch_hctx_list(&rq_list);
} while (!list_empty(&rq_list));
} else {
dispatched = blk_mq_dispatch_rq_list(hctx, &rq_list, count);
}
if (busy)
return -EAGAIN;
return !!dispatched;
}
static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
{
int ret;
do {
ret = __blk_mq_do_dispatch_sched(hctx);
} while (ret == 1);
return ret;
}
static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
struct blk_mq_ctx *ctx)
{
unsigned short idx = ctx->index_hw[hctx->type];
if (++idx == hctx->nr_ctx)
idx = 0;
return hctx->ctxs[idx];
}
/*
* Only SCSI implements .get_budget and .put_budget, and SCSI restarts
* its queue by itself in its completion handler, so we don't need to
* restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
*
* Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
* be run again. This is necessary to avoid starving flushes.
*/
static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
{
struct request_queue *q = hctx->queue;
LIST_HEAD(rq_list);
struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
int ret = 0;
struct request *rq;
do {
int budget_token;
if (!list_empty_careful(&hctx->dispatch)) {
ret = -EAGAIN;
break;
}
if (!sbitmap_any_bit_set(&hctx->ctx_map))
break;
budget_token = blk_mq_get_dispatch_budget(q);
if (budget_token < 0)
break;
rq = blk_mq_dequeue_from_ctx(hctx, ctx);
if (!rq) {
blk_mq_put_dispatch_budget(q, budget_token);
blk-mq: Rerun dispatching in the case of budget contention If ever a thread running blk-mq code tries to get budget and fails it immediately stops doing work and assumes that whenever budget is freed up that queues will be kicked and whatever work the thread was trying to do will be tried again. One path where budget is freed and queues are kicked in the normal case can be seen in scsi_finish_command(). Specifically: - scsi_finish_command() - scsi_device_unbusy() - # Decrement "device_busy", AKA release budget - scsi_io_completion() - scsi_end_request() - blk_mq_run_hw_queues() The above is all well and good. The problem comes up when a thread claims the budget but then releases it without actually dispatching any work. Since we didn't schedule any work we'll never run the path of finishing work / kicking the queues. This isn't often actually a problem which is why this issue has existed for a while and nobody noticed. Specifically we only get into this situation when we unexpectedly found that we weren't going to do any work. Code that later receives new work kicks the queues. All good, right? The problem shows up, however, if timing is just wrong and we hit a race. To see this race let's think about the case where we only have a budget of 1 (only one thread can hold budget). Now imagine that a thread got budget and then decided not to dispatch work. It's about to call put_budget() but then the thread gets context switched out for a long, long time. While in this state, any and all kicks of the queue (like the when we received new work) will be no-ops because nobody can get budget. Finally the thread holding budget gets to run again and returns. All the normal kicks will have been no-ops and we have an I/O stall. As you can see from the above, you need just the right timing to see the race. To start with, the only case it happens if we thought we had work, actually managed to get the budget, but then actually didn't have work. That's pretty rare to start with. Even then, there's usually a very small amount of time between realizing that there's no work and putting the budget. During this small amount of time new work has to come in and the queue kick has to make it all the way to trying to get the budget and fail. It's pretty unlikely. One case where this could have failed is illustrated by an example of threads running blk_mq_do_dispatch_sched(): * Threads A and B both run has_work() at the same time with the same "hctx". Imagine has_work() is exact. There's no lock, so it's OK if Thread A and B both get back true. * Thread B gets interrupted for a long time right after it decides that there is work. Maybe its CPU gets an interrupt and the interrupt handler is slow. * Thread A runs, get budget, dispatches work. * Thread A's work finishes and budget is released. * Thread B finally runs again and gets budget. * Since Thread A already took care of the work and no new work has come in, Thread B will get NULL from dispatch_request(). I believe this is specifically why dispatch_request() is allowed to return NULL in the first place if has_work() must be exact. * Thread B will now be holding the budget and is about to call put_budget(), but hasn't called it yet. * Thread B gets interrupted for a long time (again). Dang interrupts. * Now Thread C (maybe with a different "hctx" but the same queue) comes along and runs blk_mq_do_dispatch_sched(). * Thread C won't do anything because it can't get budget. * Finally Thread B will run again and put the budget without kicking any queues. Even though the example above is with blk_mq_do_dispatch_sched() I believe the race is possible any time someone is holding budget but doesn't do work. Unfortunately, the unlikely has become more likely if you happen to be using the BFQ I/O scheduler. BFQ, by design, sometimes returns "true" for has_work() but then NULL for dispatch_request() and stays in this state for a while (currently up to 9 ms). Suddenly you only need one race to hit, not two races in a row. With my current setup this is easy to reproduce in reboot tests and traces have actually shown that we hit a race similar to the one described above. Note that we only need to fix blk_mq_do_dispatch_sched() and blk_mq_do_dispatch_ctx() and not the other places that put budget. In other cases we know that we have work to do on at least one "hctx" and code already exists to kick that "hctx"'s queue. When that work finally finishes all the queues will be kicked using the normal flow. One last note is that (at least in the SCSI case) budget is shared by all "hctx"s that have the same queue. Thus we need to make sure to kick the whole queue, not just re-run dispatching on a single "hctx". Signed-off-by: Douglas Anderson <dianders@chromium.org> Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-04-20 16:24:53 +00:00
/*
* We're releasing without dispatching. Holding the
* budget could have blocked any "hctx"s with the
* same queue and if we didn't dispatch then there's
* no guarantee anyone will kick the queue. Kick it
* ourselves.
*/
blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
break;
}
blk_mq_set_rq_budget_token(rq, budget_token);
/*
* Now this rq owns the budget which has to be released
* if this rq won't be queued to driver via .queue_rq()
* in blk_mq_dispatch_rq_list().
*/
list_add(&rq->queuelist, &rq_list);
/* round robin for fair dispatch */
ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
} while (blk_mq_dispatch_rq_list(rq->mq_hctx, &rq_list, 1));
WRITE_ONCE(hctx->dispatch_from, ctx);
return ret;
}
static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
{
struct request_queue *q = hctx->queue;
const bool has_sched = q->elevator;
int ret = 0;
LIST_HEAD(rq_list);
/*
* If we have previous entries on our dispatch list, grab them first for
* more fair dispatch.
*/
if (!list_empty_careful(&hctx->dispatch)) {
spin_lock(&hctx->lock);
if (!list_empty(&hctx->dispatch))
list_splice_init(&hctx->dispatch, &rq_list);
spin_unlock(&hctx->lock);
}
/*
* Only ask the scheduler for requests, if we didn't have residual
* requests from the dispatch list. This is to avoid the case where
* we only ever dispatch a fraction of the requests available because
* of low device queue depth. Once we pull requests out of the IO
* scheduler, we can no longer merge or sort them. So it's best to
* leave them there for as long as we can. Mark the hw queue as
* needing a restart in that case.
*
* We want to dispatch from the scheduler if there was nothing
* on the dispatch list or we were able to dispatch from the
* dispatch list.
*/
if (!list_empty(&rq_list)) {
blk_mq_sched_mark_restart_hctx(hctx);
if (blk_mq_dispatch_rq_list(hctx, &rq_list, 0)) {
if (has_sched)
ret = blk_mq_do_dispatch_sched(hctx);
else
ret = blk_mq_do_dispatch_ctx(hctx);
}
} else if (has_sched) {
ret = blk_mq_do_dispatch_sched(hctx);
} else if (hctx->dispatch_busy) {
/* dequeue request one by one from sw queue if queue is busy */
ret = blk_mq_do_dispatch_ctx(hctx);
} else {
blk_mq_flush_busy_ctxs(hctx, &rq_list);
blk_mq_dispatch_rq_list(hctx, &rq_list, 0);
}
return ret;
}
void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
{
struct request_queue *q = hctx->queue;
/* RCU or SRCU read lock is needed before checking quiesced flag */
if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
return;
hctx->run++;
/*
* A return of -EAGAIN is an indication that hctx->dispatch is not
* empty and we must run again in order to avoid starving flushes.
*/
if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) {
if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN)
blk_mq_run_hw_queue(hctx, true);
}
}
bool blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio,
unsigned int nr_segs)
{
struct elevator_queue *e = q->elevator;
kyber: fix out of bounds access when preempted __blk_mq_sched_bio_merge() gets the ctx and hctx for the current CPU and passes the hctx to ->bio_merge(). kyber_bio_merge() then gets the ctx for the current CPU again and uses that to get the corresponding Kyber context in the passed hctx. However, the thread may be preempted between the two calls to blk_mq_get_ctx(), and the ctx returned the second time may no longer correspond to the passed hctx. This "works" accidentally most of the time, but it can cause us to read garbage if the second ctx came from an hctx with more ctx's than the first one (i.e., if ctx->index_hw[hctx->type] > hctx->nr_ctx). This manifested as this UBSAN array index out of bounds error reported by Jakub: UBSAN: array-index-out-of-bounds in ../kernel/locking/qspinlock.c:130:9 index 13106 is out of range for type 'long unsigned int [128]' Call Trace: dump_stack+0xa4/0xe5 ubsan_epilogue+0x5/0x40 __ubsan_handle_out_of_bounds.cold.13+0x2a/0x34 queued_spin_lock_slowpath+0x476/0x480 do_raw_spin_lock+0x1c2/0x1d0 kyber_bio_merge+0x112/0x180 blk_mq_submit_bio+0x1f5/0x1100 submit_bio_noacct+0x7b0/0x870 submit_bio+0xc2/0x3a0 btrfs_map_bio+0x4f0/0x9d0 btrfs_submit_data_bio+0x24e/0x310 submit_one_bio+0x7f/0xb0 submit_extent_page+0xc4/0x440 __extent_writepage_io+0x2b8/0x5e0 __extent_writepage+0x28d/0x6e0 extent_write_cache_pages+0x4d7/0x7a0 extent_writepages+0xa2/0x110 do_writepages+0x8f/0x180 __writeback_single_inode+0x99/0x7f0 writeback_sb_inodes+0x34e/0x790 __writeback_inodes_wb+0x9e/0x120 wb_writeback+0x4d2/0x660 wb_workfn+0x64d/0xa10 process_one_work+0x53a/0xa80 worker_thread+0x69/0x5b0 kthread+0x20b/0x240 ret_from_fork+0x1f/0x30 Only Kyber uses the hctx, so fix it by passing the request_queue to ->bio_merge() instead. BFQ and mq-deadline just use that, and Kyber can map the queues itself to avoid the mismatch. Fixes: a6088845c2bf ("block: kyber: make kyber more friendly with merging") Reported-by: Jakub Kicinski <kuba@kernel.org> Signed-off-by: Omar Sandoval <osandov@fb.com> Link: https://lore.kernel.org/r/c7598605401a48d5cfeadebb678abd10af22b83f.1620691329.git.osandov@fb.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-05-11 00:05:35 +00:00
struct blk_mq_ctx *ctx;
struct blk_mq_hw_ctx *hctx;
bool ret = false;
enum hctx_type type;
if (e && e->type->ops.bio_merge) {
ret = e->type->ops.bio_merge(q, bio, nr_segs);
goto out_put;
}
kyber: fix out of bounds access when preempted __blk_mq_sched_bio_merge() gets the ctx and hctx for the current CPU and passes the hctx to ->bio_merge(). kyber_bio_merge() then gets the ctx for the current CPU again and uses that to get the corresponding Kyber context in the passed hctx. However, the thread may be preempted between the two calls to blk_mq_get_ctx(), and the ctx returned the second time may no longer correspond to the passed hctx. This "works" accidentally most of the time, but it can cause us to read garbage if the second ctx came from an hctx with more ctx's than the first one (i.e., if ctx->index_hw[hctx->type] > hctx->nr_ctx). This manifested as this UBSAN array index out of bounds error reported by Jakub: UBSAN: array-index-out-of-bounds in ../kernel/locking/qspinlock.c:130:9 index 13106 is out of range for type 'long unsigned int [128]' Call Trace: dump_stack+0xa4/0xe5 ubsan_epilogue+0x5/0x40 __ubsan_handle_out_of_bounds.cold.13+0x2a/0x34 queued_spin_lock_slowpath+0x476/0x480 do_raw_spin_lock+0x1c2/0x1d0 kyber_bio_merge+0x112/0x180 blk_mq_submit_bio+0x1f5/0x1100 submit_bio_noacct+0x7b0/0x870 submit_bio+0xc2/0x3a0 btrfs_map_bio+0x4f0/0x9d0 btrfs_submit_data_bio+0x24e/0x310 submit_one_bio+0x7f/0xb0 submit_extent_page+0xc4/0x440 __extent_writepage_io+0x2b8/0x5e0 __extent_writepage+0x28d/0x6e0 extent_write_cache_pages+0x4d7/0x7a0 extent_writepages+0xa2/0x110 do_writepages+0x8f/0x180 __writeback_single_inode+0x99/0x7f0 writeback_sb_inodes+0x34e/0x790 __writeback_inodes_wb+0x9e/0x120 wb_writeback+0x4d2/0x660 wb_workfn+0x64d/0xa10 process_one_work+0x53a/0xa80 worker_thread+0x69/0x5b0 kthread+0x20b/0x240 ret_from_fork+0x1f/0x30 Only Kyber uses the hctx, so fix it by passing the request_queue to ->bio_merge() instead. BFQ and mq-deadline just use that, and Kyber can map the queues itself to avoid the mismatch. Fixes: a6088845c2bf ("block: kyber: make kyber more friendly with merging") Reported-by: Jakub Kicinski <kuba@kernel.org> Signed-off-by: Omar Sandoval <osandov@fb.com> Link: https://lore.kernel.org/r/c7598605401a48d5cfeadebb678abd10af22b83f.1620691329.git.osandov@fb.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-05-11 00:05:35 +00:00
ctx = blk_mq_get_ctx(q);
hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
type = hctx->type;
if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE) ||
list_empty_careful(&ctx->rq_lists[type]))
goto out_put;
/* default per sw-queue merge */
spin_lock(&ctx->lock);
/*
* Reverse check our software queue for entries that we could
* potentially merge with. Currently includes a hand-wavy stop
* count of 8, to not spend too much time checking for merges.
*/
if (blk_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs))
ret = true;
spin_unlock(&ctx->lock);
out_put:
return ret;
}
bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq,
struct list_head *free)
{
return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq, free);
}
EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
struct request *rq)
{
/*
* dispatch flush and passthrough rq directly
*
* passthrough request has to be added to hctx->dispatch directly.
* For some reason, device may be in one situation which can't
* handle FS request, so STS_RESOURCE is always returned and the
* FS request will be added to hctx->dispatch. However passthrough
* request may be required at that time for fixing the problem. If
* passthrough request is added to scheduler queue, there isn't any
* chance to dispatch it given we prioritize requests in hctx->dispatch.
*/
if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq))
return true;
return false;
}
void blk_mq_sched_insert_request(struct request *rq, bool at_head,
bool run_queue, bool async)
{
struct request_queue *q = rq->q;
struct elevator_queue *e = q->elevator;
struct blk_mq_ctx *ctx = rq->mq_ctx;
struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
WARN_ON(e && (rq->tag != BLK_MQ_NO_TAG));
if (blk_mq_sched_bypass_insert(hctx, rq)) {
/*
* Firstly normal IO request is inserted to scheduler queue or
* sw queue, meantime we add flush request to dispatch queue(
* hctx->dispatch) directly and there is at most one in-flight
* flush request for each hw queue, so it doesn't matter to add
* flush request to tail or front of the dispatch queue.
*
* Secondly in case of NCQ, flush request belongs to non-NCQ
* command, and queueing it will fail when there is any
* in-flight normal IO request(NCQ command). When adding flush
* rq to the front of hctx->dispatch, it is easier to introduce
* extra time to flush rq's latency because of S_SCHED_RESTART
* compared with adding to the tail of dispatch queue, then
* chance of flush merge is increased, and less flush requests
* will be issued to controller. It is observed that ~10% time
* is saved in blktests block/004 on disk attached to AHCI/NCQ
* drive when adding flush rq to the front of hctx->dispatch.
*
* Simply queue flush rq to the front of hctx->dispatch so that
* intensive flush workloads can benefit in case of NCQ HW.
*/
at_head = (rq->rq_flags & RQF_FLUSH_SEQ) ? true : at_head;
blk_mq_request_bypass_insert(rq, at_head, false);
goto run;
}
if (e) {
LIST_HEAD(list);
list_add(&rq->queuelist, &list);
e->type->ops.insert_requests(hctx, &list, at_head);
} else {
spin_lock(&ctx->lock);
__blk_mq_insert_request(hctx, rq, at_head);
spin_unlock(&ctx->lock);
}
run:
if (run_queue)
blk_mq_run_hw_queue(hctx, async);
}
void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx,
struct blk_mq_ctx *ctx,
struct list_head *list, bool run_queue_async)
{
struct elevator_queue *e;
blk-mq: grab .q_usage_counter when queuing request from plug code path Just like aio/io_uring, we need to grab 2 refcount for queuing one request, one is for submission, another is for completion. If the request isn't queued from plug code path, the refcount grabbed in generic_make_request() serves for submission. In theroy, this refcount should have been released after the sumission(async run queue) is done. blk_freeze_queue() works with blk_sync_queue() together for avoiding race between cleanup queue and IO submission, given async run queue activities are canceled because hctx->run_work is scheduled with the refcount held, so it is fine to not hold the refcount when running the run queue work function for dispatch IO. However, if request is staggered into plug list, and finally queued from plug code path, the refcount in submission side is actually missed. And we may start to run queue after queue is removed because the queue's kobject refcount isn't guaranteed to be grabbed in flushing plug list context, then kernel oops is triggered, see the following race: blk_mq_flush_plug_list(): blk_mq_sched_insert_requests() insert requests to sw queue or scheduler queue blk_mq_run_hw_queue Because of concurrent run queue, all requests inserted above may be completed before calling the above blk_mq_run_hw_queue. Then queue can be freed during the above blk_mq_run_hw_queue(). Fixes the issue by grab .q_usage_counter before calling blk_mq_sched_insert_requests() in blk_mq_flush_plug_list(). This way is safe because the queue is absolutely alive before inserting request. Cc: Dongli Zhang <dongli.zhang@oracle.com> Cc: James Smart <james.smart@broadcom.com> Cc: linux-scsi@vger.kernel.org, Cc: Martin K . Petersen <martin.petersen@oracle.com>, Cc: Christoph Hellwig <hch@lst.de>, Cc: James E . J . Bottomley <jejb@linux.vnet.ibm.com>, Reviewed-by: Bart Van Assche <bvanassche@acm.org> Tested-by: James Smart <james.smart@broadcom.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-30 01:52:23 +00:00
struct request_queue *q = hctx->queue;
/*
* blk_mq_sched_insert_requests() is called from flush plug
* context only, and hold one usage counter to prevent queue
* from being released.
*/
percpu_ref_get(&q->q_usage_counter);
e = hctx->queue->elevator;
if (e) {
e->type->ops.insert_requests(hctx, list, false);
} else {
/*
* try to issue requests directly if the hw queue isn't
* busy in case of 'none' scheduler, and this way may save
* us one extra enqueue & dequeue to sw queue.
*/
if (!hctx->dispatch_busy && !run_queue_async) {
blk_mq_try_issue_list_directly(hctx, list);
if (list_empty(list))
blk-mq: grab .q_usage_counter when queuing request from plug code path Just like aio/io_uring, we need to grab 2 refcount for queuing one request, one is for submission, another is for completion. If the request isn't queued from plug code path, the refcount grabbed in generic_make_request() serves for submission. In theroy, this refcount should have been released after the sumission(async run queue) is done. blk_freeze_queue() works with blk_sync_queue() together for avoiding race between cleanup queue and IO submission, given async run queue activities are canceled because hctx->run_work is scheduled with the refcount held, so it is fine to not hold the refcount when running the run queue work function for dispatch IO. However, if request is staggered into plug list, and finally queued from plug code path, the refcount in submission side is actually missed. And we may start to run queue after queue is removed because the queue's kobject refcount isn't guaranteed to be grabbed in flushing plug list context, then kernel oops is triggered, see the following race: blk_mq_flush_plug_list(): blk_mq_sched_insert_requests() insert requests to sw queue or scheduler queue blk_mq_run_hw_queue Because of concurrent run queue, all requests inserted above may be completed before calling the above blk_mq_run_hw_queue. Then queue can be freed during the above blk_mq_run_hw_queue(). Fixes the issue by grab .q_usage_counter before calling blk_mq_sched_insert_requests() in blk_mq_flush_plug_list(). This way is safe because the queue is absolutely alive before inserting request. Cc: Dongli Zhang <dongli.zhang@oracle.com> Cc: James Smart <james.smart@broadcom.com> Cc: linux-scsi@vger.kernel.org, Cc: Martin K . Petersen <martin.petersen@oracle.com>, Cc: Christoph Hellwig <hch@lst.de>, Cc: James E . J . Bottomley <jejb@linux.vnet.ibm.com>, Reviewed-by: Bart Van Assche <bvanassche@acm.org> Tested-by: James Smart <james.smart@broadcom.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-30 01:52:23 +00:00
goto out;
}
blk_mq_insert_requests(hctx, ctx, list);
}
blk_mq_run_hw_queue(hctx, run_queue_async);
blk-mq: grab .q_usage_counter when queuing request from plug code path Just like aio/io_uring, we need to grab 2 refcount for queuing one request, one is for submission, another is for completion. If the request isn't queued from plug code path, the refcount grabbed in generic_make_request() serves for submission. In theroy, this refcount should have been released after the sumission(async run queue) is done. blk_freeze_queue() works with blk_sync_queue() together for avoiding race between cleanup queue and IO submission, given async run queue activities are canceled because hctx->run_work is scheduled with the refcount held, so it is fine to not hold the refcount when running the run queue work function for dispatch IO. However, if request is staggered into plug list, and finally queued from plug code path, the refcount in submission side is actually missed. And we may start to run queue after queue is removed because the queue's kobject refcount isn't guaranteed to be grabbed in flushing plug list context, then kernel oops is triggered, see the following race: blk_mq_flush_plug_list(): blk_mq_sched_insert_requests() insert requests to sw queue or scheduler queue blk_mq_run_hw_queue Because of concurrent run queue, all requests inserted above may be completed before calling the above blk_mq_run_hw_queue. Then queue can be freed during the above blk_mq_run_hw_queue(). Fixes the issue by grab .q_usage_counter before calling blk_mq_sched_insert_requests() in blk_mq_flush_plug_list(). This way is safe because the queue is absolutely alive before inserting request. Cc: Dongli Zhang <dongli.zhang@oracle.com> Cc: James Smart <james.smart@broadcom.com> Cc: linux-scsi@vger.kernel.org, Cc: Martin K . Petersen <martin.petersen@oracle.com>, Cc: Christoph Hellwig <hch@lst.de>, Cc: James E . J . Bottomley <jejb@linux.vnet.ibm.com>, Reviewed-by: Bart Van Assche <bvanassche@acm.org> Tested-by: James Smart <james.smart@broadcom.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-30 01:52:23 +00:00
out:
percpu_ref_put(&q->q_usage_counter);
}
static int blk_mq_sched_alloc_map_and_rqs(struct request_queue *q,
struct blk_mq_hw_ctx *hctx,
unsigned int hctx_idx)
{
if (blk_mq_is_shared_tags(q->tag_set->flags)) {
hctx->sched_tags = q->sched_shared_tags;
return 0;
}
hctx->sched_tags = blk_mq_alloc_map_and_rqs(q->tag_set, hctx_idx,
q->nr_requests);
if (!hctx->sched_tags)
return -ENOMEM;
return 0;
}
static void blk_mq_exit_sched_shared_tags(struct request_queue *queue)
{
blk_mq_free_rq_map(queue->sched_shared_tags);
queue->sched_shared_tags = NULL;
}
block: free sched's request pool in blk_cleanup_queue In theory, IO scheduler belongs to request queue, and the request pool of sched tags belongs to the request queue too. However, the current tags allocation interfaces are re-used for both driver tags and sched tags, and driver tags is definitely host wide, and doesn't belong to any request queue, same with its request pool. So we need tagset instance for freeing request of sched tags. Meantime, blk_mq_free_tag_set() often follows blk_cleanup_queue() in case of non-BLK_MQ_F_TAG_SHARED, this way requires that request pool of sched tags to be freed before calling blk_mq_free_tag_set(). Commit 47cdee29ef9d94e ("block: move blk_exit_queue into __blk_release_queue") moves blk_exit_queue into __blk_release_queue for simplying the fast path in generic_make_request(), then causes oops during freeing requests of sched tags in __blk_release_queue(). Fix the above issue by move freeing request pool of sched tags into blk_cleanup_queue(), this way is safe becasue queue has been frozen and no any in-queue requests at that time. Freeing sched tags has to be kept in queue's release handler becasue there might be un-completed dispatch activity which might refer to sched tags. Cc: Bart Van Assche <bvanassche@acm.org> Cc: Christoph Hellwig <hch@lst.de> Fixes: 47cdee29ef9d94e485eb08f962c74943023a5271 ("block: move blk_exit_queue into __blk_release_queue") Tested-by: Yi Zhang <yi.zhang@redhat.com> Reported-by: kernel test robot <rong.a.chen@intel.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-06-04 13:08:02 +00:00
/* called in queue's release handler, tagset has gone away */
static void blk_mq_sched_tags_teardown(struct request_queue *q, unsigned int flags)
{
struct blk_mq_hw_ctx *hctx;
int i;
block: free sched's request pool in blk_cleanup_queue In theory, IO scheduler belongs to request queue, and the request pool of sched tags belongs to the request queue too. However, the current tags allocation interfaces are re-used for both driver tags and sched tags, and driver tags is definitely host wide, and doesn't belong to any request queue, same with its request pool. So we need tagset instance for freeing request of sched tags. Meantime, blk_mq_free_tag_set() often follows blk_cleanup_queue() in case of non-BLK_MQ_F_TAG_SHARED, this way requires that request pool of sched tags to be freed before calling blk_mq_free_tag_set(). Commit 47cdee29ef9d94e ("block: move blk_exit_queue into __blk_release_queue") moves blk_exit_queue into __blk_release_queue for simplying the fast path in generic_make_request(), then causes oops during freeing requests of sched tags in __blk_release_queue(). Fix the above issue by move freeing request pool of sched tags into blk_cleanup_queue(), this way is safe becasue queue has been frozen and no any in-queue requests at that time. Freeing sched tags has to be kept in queue's release handler becasue there might be un-completed dispatch activity which might refer to sched tags. Cc: Bart Van Assche <bvanassche@acm.org> Cc: Christoph Hellwig <hch@lst.de> Fixes: 47cdee29ef9d94e485eb08f962c74943023a5271 ("block: move blk_exit_queue into __blk_release_queue") Tested-by: Yi Zhang <yi.zhang@redhat.com> Reported-by: kernel test robot <rong.a.chen@intel.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-06-04 13:08:02 +00:00
queue_for_each_hw_ctx(q, hctx, i) {
if (hctx->sched_tags) {
if (!blk_mq_is_shared_tags(flags))
blk_mq_free_rq_map(hctx->sched_tags);
block: free sched's request pool in blk_cleanup_queue In theory, IO scheduler belongs to request queue, and the request pool of sched tags belongs to the request queue too. However, the current tags allocation interfaces are re-used for both driver tags and sched tags, and driver tags is definitely host wide, and doesn't belong to any request queue, same with its request pool. So we need tagset instance for freeing request of sched tags. Meantime, blk_mq_free_tag_set() often follows blk_cleanup_queue() in case of non-BLK_MQ_F_TAG_SHARED, this way requires that request pool of sched tags to be freed before calling blk_mq_free_tag_set(). Commit 47cdee29ef9d94e ("block: move blk_exit_queue into __blk_release_queue") moves blk_exit_queue into __blk_release_queue for simplying the fast path in generic_make_request(), then causes oops during freeing requests of sched tags in __blk_release_queue(). Fix the above issue by move freeing request pool of sched tags into blk_cleanup_queue(), this way is safe becasue queue has been frozen and no any in-queue requests at that time. Freeing sched tags has to be kept in queue's release handler becasue there might be un-completed dispatch activity which might refer to sched tags. Cc: Bart Van Assche <bvanassche@acm.org> Cc: Christoph Hellwig <hch@lst.de> Fixes: 47cdee29ef9d94e485eb08f962c74943023a5271 ("block: move blk_exit_queue into __blk_release_queue") Tested-by: Yi Zhang <yi.zhang@redhat.com> Reported-by: kernel test robot <rong.a.chen@intel.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-06-04 13:08:02 +00:00
hctx->sched_tags = NULL;
}
}
if (blk_mq_is_shared_tags(flags))
blk_mq_exit_sched_shared_tags(q);
}
static int blk_mq_init_sched_shared_tags(struct request_queue *queue)
2021-05-13 12:00:58 +00:00
{
struct blk_mq_tag_set *set = queue->tag_set;
/*
* Set initial depth at max so that we don't need to reallocate for
* updating nr_requests.
*/
queue->sched_shared_tags = blk_mq_alloc_map_and_rqs(set,
BLK_MQ_NO_HCTX_IDX,
MAX_SCHED_RQ);
if (!queue->sched_shared_tags)
return -ENOMEM;
2021-05-13 12:00:58 +00:00
blk_mq_tag_update_sched_shared_tags(queue);
2021-05-13 12:00:58 +00:00
return 0;
}
int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
{
unsigned int i, flags = q->tag_set->flags;
struct blk_mq_hw_ctx *hctx;
struct elevator_queue *eq;
int ret;
if (!e) {
q->elevator = NULL;
q->nr_requests = q->tag_set->queue_depth;
return 0;
}
/*
* Default to double of smaller one between hw queue_depth and 128,
* since we don't split into sync/async like the old code did.
* Additionally, this is a per-hw queue depth.
*/
q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
BLKDEV_DEFAULT_RQ);
if (blk_mq_is_shared_tags(flags)) {
ret = blk_mq_init_sched_shared_tags(q);
if (ret)
return ret;
2021-05-13 12:00:58 +00:00
}
queue_for_each_hw_ctx(q, hctx, i) {
ret = blk_mq_sched_alloc_map_and_rqs(q, hctx, i);
2021-05-13 12:00:58 +00:00
if (ret)
goto err_free_map_and_rqs;
}
ret = e->ops.init_sched(q, e);
if (ret)
goto err_free_map_and_rqs;
blk_mq_debugfs_register_sched(q);
queue_for_each_hw_ctx(q, hctx, i) {
if (e->ops.init_hctx) {
ret = e->ops.init_hctx(hctx, i);
if (ret) {
eq = q->elevator;
blk_mq_sched_free_rqs(q);
blk_mq_exit_sched(q, eq);
kobject_put(&eq->kobj);
return ret;
}
}
blk_mq_debugfs_register_sched_hctx(q, hctx);
}
return 0;
err_free_map_and_rqs:
blk_mq_sched_free_rqs(q);
blk_mq_sched_tags_teardown(q, flags);
q->elevator = NULL;
return ret;
}
block: free sched's request pool in blk_cleanup_queue In theory, IO scheduler belongs to request queue, and the request pool of sched tags belongs to the request queue too. However, the current tags allocation interfaces are re-used for both driver tags and sched tags, and driver tags is definitely host wide, and doesn't belong to any request queue, same with its request pool. So we need tagset instance for freeing request of sched tags. Meantime, blk_mq_free_tag_set() often follows blk_cleanup_queue() in case of non-BLK_MQ_F_TAG_SHARED, this way requires that request pool of sched tags to be freed before calling blk_mq_free_tag_set(). Commit 47cdee29ef9d94e ("block: move blk_exit_queue into __blk_release_queue") moves blk_exit_queue into __blk_release_queue for simplying the fast path in generic_make_request(), then causes oops during freeing requests of sched tags in __blk_release_queue(). Fix the above issue by move freeing request pool of sched tags into blk_cleanup_queue(), this way is safe becasue queue has been frozen and no any in-queue requests at that time. Freeing sched tags has to be kept in queue's release handler becasue there might be un-completed dispatch activity which might refer to sched tags. Cc: Bart Van Assche <bvanassche@acm.org> Cc: Christoph Hellwig <hch@lst.de> Fixes: 47cdee29ef9d94e485eb08f962c74943023a5271 ("block: move blk_exit_queue into __blk_release_queue") Tested-by: Yi Zhang <yi.zhang@redhat.com> Reported-by: kernel test robot <rong.a.chen@intel.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-06-04 13:08:02 +00:00
/*
* called in either blk_queue_cleanup or elevator_switch, tagset
* is required for freeing requests
*/
void blk_mq_sched_free_rqs(struct request_queue *q)
block: free sched's request pool in blk_cleanup_queue In theory, IO scheduler belongs to request queue, and the request pool of sched tags belongs to the request queue too. However, the current tags allocation interfaces are re-used for both driver tags and sched tags, and driver tags is definitely host wide, and doesn't belong to any request queue, same with its request pool. So we need tagset instance for freeing request of sched tags. Meantime, blk_mq_free_tag_set() often follows blk_cleanup_queue() in case of non-BLK_MQ_F_TAG_SHARED, this way requires that request pool of sched tags to be freed before calling blk_mq_free_tag_set(). Commit 47cdee29ef9d94e ("block: move blk_exit_queue into __blk_release_queue") moves blk_exit_queue into __blk_release_queue for simplying the fast path in generic_make_request(), then causes oops during freeing requests of sched tags in __blk_release_queue(). Fix the above issue by move freeing request pool of sched tags into blk_cleanup_queue(), this way is safe becasue queue has been frozen and no any in-queue requests at that time. Freeing sched tags has to be kept in queue's release handler becasue there might be un-completed dispatch activity which might refer to sched tags. Cc: Bart Van Assche <bvanassche@acm.org> Cc: Christoph Hellwig <hch@lst.de> Fixes: 47cdee29ef9d94e485eb08f962c74943023a5271 ("block: move blk_exit_queue into __blk_release_queue") Tested-by: Yi Zhang <yi.zhang@redhat.com> Reported-by: kernel test robot <rong.a.chen@intel.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-06-04 13:08:02 +00:00
{
struct blk_mq_hw_ctx *hctx;
int i;
if (blk_mq_is_shared_tags(q->tag_set->flags)) {
blk_mq_free_rqs(q->tag_set, q->sched_shared_tags,
BLK_MQ_NO_HCTX_IDX);
} else {
queue_for_each_hw_ctx(q, hctx, i) {
if (hctx->sched_tags)
blk_mq_free_rqs(q->tag_set,
hctx->sched_tags, i);
}
block: free sched's request pool in blk_cleanup_queue In theory, IO scheduler belongs to request queue, and the request pool of sched tags belongs to the request queue too. However, the current tags allocation interfaces are re-used for both driver tags and sched tags, and driver tags is definitely host wide, and doesn't belong to any request queue, same with its request pool. So we need tagset instance for freeing request of sched tags. Meantime, blk_mq_free_tag_set() often follows blk_cleanup_queue() in case of non-BLK_MQ_F_TAG_SHARED, this way requires that request pool of sched tags to be freed before calling blk_mq_free_tag_set(). Commit 47cdee29ef9d94e ("block: move blk_exit_queue into __blk_release_queue") moves blk_exit_queue into __blk_release_queue for simplying the fast path in generic_make_request(), then causes oops during freeing requests of sched tags in __blk_release_queue(). Fix the above issue by move freeing request pool of sched tags into blk_cleanup_queue(), this way is safe becasue queue has been frozen and no any in-queue requests at that time. Freeing sched tags has to be kept in queue's release handler becasue there might be un-completed dispatch activity which might refer to sched tags. Cc: Bart Van Assche <bvanassche@acm.org> Cc: Christoph Hellwig <hch@lst.de> Fixes: 47cdee29ef9d94e485eb08f962c74943023a5271 ("block: move blk_exit_queue into __blk_release_queue") Tested-by: Yi Zhang <yi.zhang@redhat.com> Reported-by: kernel test robot <rong.a.chen@intel.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-06-04 13:08:02 +00:00
}
}
void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
{
struct blk_mq_hw_ctx *hctx;
unsigned int i;
unsigned int flags = 0;
queue_for_each_hw_ctx(q, hctx, i) {
blk_mq_debugfs_unregister_sched_hctx(hctx);
if (e->type->ops.exit_hctx && hctx->sched_data) {
e->type->ops.exit_hctx(hctx, i);
hctx->sched_data = NULL;
}
flags = hctx->flags;
}
blk_mq_debugfs_unregister_sched(q);
if (e->type->ops.exit_sched)
e->type->ops.exit_sched(e);
blk_mq_sched_tags_teardown(q, flags);
q->elevator = NULL;
}