On 7/30/19 1:55 AM, Paolo Valente wrote:
Hi Guenter,
sorry for the delay (Dolomiti's fault).
I didn't consider that rq->elv-icq might have been NULL also
because of OOM. Thanks for spotting this issue.
As for the other places where the return value of bfq_init_rq is used,
unfortunately I think they matter too. Those other places are related
to request merging, which is the alternative destiny of requests
(instead of being just inserted). But, regardless of whether a
request is to be merged or inserted, that request may be destined to a
bfq_queue (possibly merged with a request already in a bfq_queue), and
a NULL return value by bfq_init_rq leads to a crash. I guess you can
reproduce your failure also for the merge case, by generating
sequential, direct I/O with queue depth > 1, and of course by enabling
failslab.
My assumption was that requests would only be merged if they are associated
with the same io context. In that case, that IO context isn't reallocated
with ioc_create_icq() but reused, and icq would thus never be NULL.
I guess that assumption was wrong.
If my considerations above are correct, do you want to propose a
complete fix yourself?
Sure, I'll send an updated patch.
Thanks,
Guenter
Thanks,
Paolo
Il giorno 28 lug 2019, alle ore 17:19, Guenter Roeck <linux@xxxxxxxxxxxx> ha scritto:
ping ... just in case this patch got lost in Paolo's queue.
Guenter
On Mon, Jul 22, 2019 at 10:30:48AM -0700, Guenter Roeck wrote:
In bfq_insert_request(), bfqq is initialized with:
bfqq = bfq_init_rq(rq);
In bfq_init_rq(), we find:
if (unlikely(!rq->elv.icq))
return NULL;
Indeed, rq->elv.icq can be NULL if the memory allocation in
create_task_io_context() failed.
A comment in bfq_insert_request() suggests that bfqq is supposed to be
non-NULL if 'at_head || blk_rq_is_passthrough(rq)' is false. Yet, as
debugging and practical experience shows, this is not the case in the
above situation.
This results in the following crash.
Unable to handle kernel NULL pointer dereference
at virtual address 00000000000001b0
...
Call trace:
bfq_setup_cooperator+0x44/0x134
bfq_insert_requests+0x10c/0x630
blk_mq_sched_insert_requests+0x60/0xb4
blk_mq_flush_plug_list+0x290/0x2d4
blk_flush_plug_list+0xe0/0x230
blk_finish_plug+0x30/0x40
generic_writepages+0x60/0x94
blkdev_writepages+0x24/0x30
do_writepages+0x74/0xac
__filemap_fdatawrite_range+0x94/0xc8
file_write_and_wait_range+0x44/0xa0
blkdev_fsync+0x38/0x68
vfs_fsync_range+0x68/0x80
do_fsync+0x44/0x80
The problem is relatively easy to reproduce by running an image with
failslab enabled, such as with:
cd /sys/kernel/debug/failslab
echo 10 > probability
echo 300 > times
Avoid the problem by checking if bfqq is NULL before using it. With the
NULL check in place, requests with missing io context are queued
immediately, and the crash is no longer seen.
Fixes: 18e5a57d79878 ("block, bfq: postpone rq preparation to insert or merge")
Reported-by: Hsin-Yi Wang <hsinyi@xxxxxxxxxx>
Cc: Hsin-Yi Wang <hsinyi@xxxxxxxxxx>
Cc: Nicolas Boichat <drinkcat@xxxxxxxxxxxx>
Cc: Doug Anderson <dianders@xxxxxxxxxxxx>
Signed-off-by: Guenter Roeck <linux@xxxxxxxxxxxx>
---
block/bfq-iosched.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/block/bfq-iosched.c b/block/bfq-iosched.c
index 72860325245a..56f3f4227010 100644
--- a/block/bfq-iosched.c
+++ b/block/bfq-iosched.c
@@ -5417,7 +5417,7 @@ static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
spin_lock_irq(&bfqd->lock);
bfqq = bfq_init_rq(rq);
- if (at_head || blk_rq_is_passthrough(rq)) {
+ if (!bfqq || at_head || blk_rq_is_passthrough(rq)) {
if (at_head)
list_add(&rq->queuelist, &bfqd->dispatch);
else
--
2.7.4
