On Tue, Aug 15, 2023 at 05:45:10PM +0800, Chengming Zhou wrote:
> On 2023/8/15 16:31, Ming Lei wrote:
> > On Tue, Aug 15, 2023 at 02:04:42PM +0800, chengming.zhou@xxxxxxxxx wrote:
> >> From: Chengming Zhou <zhouchengming@xxxxxxxxxxxxx>
> >>
> >> When doing io_uring benchmark on /dev/nullb0, it's easy to crash the
> >> kernel if poll requests timeout triggered, as reported by David. [1]
> >
> > Just be curious, how is the timeout triggered when running
> > "./fio/t/io_uring -r20 /dev/nullb0"?
>
> I tried "./fio/t/io_uring -r20 /dev/nullb0" multiple times, sometimes
> program exit ok, sometimes it dump many timeout messages and kernel BUG.
>
> I just used David's "./fio/t/io_uring -n4 /dev/nullb0", run a while
> then ctrl-C, the program will always dump many timeout messages.
>
> It seems that program exit is not clean, leave some requests in driver
> haven't been polled & completed? I don't know.
Just run a trace on this issue, looks it is caused by delayed __fput():
1) it is pretty easy to cause __fput(io_uring) too late in case of passing -n4
2) then io_ring_exit_work() is scheduled(for polling in-completed IOs) too late
since it is started in io_uring_release()
3) null_blk sets timeout as 5sec, so the issue can be triggered easier
But not see delayed __fput() triggered on nvme yet, still a bit
confused.
> >
> > David mentioned that the issue is triggered in 6.5-rc1, maybe one
> > regression?
> >
>
> I just tested using v6.4.9, found the same timeout and kernel BUG
> using "./fio/t/io_uring -n4 /dev/nullb0".
>
> >>
> >> BUG: kernel NULL pointer dereference, address: 0000000000000008
> >> Workqueue: kblockd blk_mq_timeout_work
> >> RIP: 0010:null_timeout_rq+0x4e/0x91
> >> Call Trace:
> >> ? __die_body+0x1a/0x5c
> >> ? page_fault_oops+0x6f/0x9c
> >> ? kernelmode_fixup_or_oops+0xc6/0xd6
> >> ? __bad_area_nosemaphore+0x44/0x1eb
> >> ? exc_page_fault+0xe2/0xf4
> >> ? asm_exc_page_fault+0x22/0x30
> >> ? null_timeout_rq+0x4e/0x91
> >> blk_mq_handle_expired+0x31/0x4b
> >> bt_iter+0x68/0x84
> >> ? bt_tags_iter+0x81/0x81
> >> __sbitmap_for_each_set.constprop.0+0xb0/0xf2
> >> ? __blk_mq_complete_request_remote+0xf/0xf
> >> bt_for_each+0x46/0x64
> >> ? __blk_mq_complete_request_remote+0xf/0xf
> >> ? percpu_ref_get_many+0xc/0x2a
> >> blk_mq_queue_tag_busy_iter+0x14d/0x18e
> >> blk_mq_timeout_work+0x95/0x127
> >> process_one_work+0x185/0x263
> >> worker_thread+0x1b5/0x227
> >> ? rescuer_thread+0x287/0x287
> >> kthread+0xfa/0x102
> >> ? kthread_complete_and_exit+0x1b/0x1b
> >> ret_from_fork+0x22/0x30
> >>
> >> This is indeed a race problem between null_timeout_rq() and null_poll().
> >>
> >> null_poll() null_timeout_rq()
> >> spin_lock(&nq->poll_lock)
> >> list_splice_init(&nq->poll_list, &list)
> >> spin_unlock(&nq->poll_lock)
> >>
> >> while (!list_empty(&list))
> >> req = list_first_entry()
> >> list_del_init()
> >> ...
> >> blk_mq_add_to_batch()
> >> // req->rq_next = NULL
> >> spin_lock(&nq->poll_lock)
> >>
> >> // rq->queuelist->next == NULL
> >> list_del_init(&rq->queuelist)
> >>
> >> spin_unlock(&nq->poll_lock)
> >>
> >> What's worse is that we don't call blk_mq_complete_request_remote()
> >> before blk_mq_add_to_batch(), so these completed requests have wrong
> >> rq->state == MQ_RQ_IN_FLIGHT. We can easily check this using bpftrace:
> >>
> >> ```
> >> bpftrace -e 'kretfunc:null_blk:null_poll {
> >> $iob=(struct io_comp_batch *)args->iob;
> >> @[$iob->req_list->state]=count();
> >> }'
> >>
> >> @[1]: 51708
> >> ```
> >>
> >> Fix these problems by setting requests state to MQ_RQ_COMPLETE under
> >> nq->poll_lock protection, in which null_timeout_rq() can safely detect
> >> this race and early return.
> >>
> >> [1] https://lore.kernel.org/all/3893581.1691785261@xxxxxxxxxxxxxxxxxxxxxx/
> >>
> >> Fixes: 0a593fbbc245 ("null_blk: poll queue support")
> >> Reported-by: David Howells <dhowells@xxxxxxxxxx>
> >> Signed-off-by: Chengming Zhou <zhouchengming@xxxxxxxxxxxxx>
> >> ---
> >> drivers/block/null_blk/main.c | 12 ++++++++++--
> >> 1 file changed, 10 insertions(+), 2 deletions(-)
> >>
> >> diff --git a/drivers/block/null_blk/main.c b/drivers/block/null_blk/main.c
> >> index 864013019d6b..968090935eb2 100644
> >> --- a/drivers/block/null_blk/main.c
> >> +++ b/drivers/block/null_blk/main.c
> >> @@ -1643,9 +1643,12 @@ static int null_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
> >> struct nullb_queue *nq = hctx->driver_data;
> >> LIST_HEAD(list);
> >> int nr = 0;
> >> + struct request *rq;
> >>
> >> spin_lock(&nq->poll_lock);
> >> list_splice_init(&nq->poll_list, &list);
> >> + list_for_each_entry(rq, &list, queuelist)
> >> + blk_mq_set_request_complete(rq);
> >> spin_unlock(&nq->poll_lock);
> >>
> >> while (!list_empty(&list)) {
> >> @@ -1671,16 +1674,21 @@ static enum blk_eh_timer_return null_timeout_rq(struct request *rq)
> >> struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
> >> struct nullb_cmd *cmd = blk_mq_rq_to_pdu(rq);
> >>
> >> - pr_info("rq %p timed out\n", rq);
> >> -
> >> if (hctx->type == HCTX_TYPE_POLL) {
> >> struct nullb_queue *nq = hctx->driver_data;
> >>
> >> spin_lock(&nq->poll_lock);
> >> + /* The request may have completed meanwhile. */
> >> + if (blk_mq_request_completed(rq)) {
> >> + spin_unlock(&nq->poll_lock);
> >> + return BLK_EH_DONE;
> >> + }
> >> list_del_init(&rq->queuelist);
> >> spin_unlock(&nq->poll_lock);
> >> }
> >
> > I think null_process_cmd() is needed for un-completed request.
> >
>
> The end of function will set BLK_STS_TIMEOUT error and complete request
> using blk_mq_complete_request(), not sure if null_process_cmd() is
> needed in this error case?
The null_blk fix looks fine:
Reviewed-by: Ming Lei <ming.lei@xxxxxxxxxx>
thanks,
Ming
