[PATCH] blk-mq: fix kernel panic during iterating over flush request

[PATCH] blk-mq: fix kernel panic during iterating over flush request

[Ming Lei]

For fixing use-after-free during iterating over requests, we grabbed
request's refcount before calling ->fn in commit 2e315dc07df0 ("blk-mq:
grab rq->refcount before calling ->fn in blk_mq_tagset_busy_iter").
Turns out this way may cause kernel panic when iterating over one flush
request:

1) old flush request's tag is just released, and this tag is reused by
one new request, but ->rqs[] isn't updated yet

2) the flush request can be re-used for submitting one new flush command,
so blk_rq_init() is called at the same time

3) meantime blk_mq_queue_tag_busy_iter() is called, and old flush request
is retrieved from ->rqs[tag]; when blk_mq_put_rq_ref() is called,
flush_rq->end_io may not be updated yet, so NULL pointer dereference
is triggered in blk_mq_put_rq_ref().

Fix the issue by calling refcount_set(&flush_rq->ref, 1) after
flush_rq->end_io is set. So far the only other caller of blk_rq_init() is
scsi_ioctl_reset() in which the request doesn't enter block IO stack and
the request reference count isn't used, so the change is safe.

Fixes: 2e315dc07df0 ("blk-mq: grab rq->refcount before calling ->fn in
blk_mq_tagset_busy_iter")
Reported-by: "Blank-Burian, Markus, Dr." <blankburian@uni-muenster.de>
Tested-by: "Blank-Burian, Markus, Dr." <blankburian@uni-muenster.de>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
---
 block/blk-core.c  | 1 -
 block/blk-flush.c | 8 ++++++++
 2 files changed, 8 insertions(+), 1 deletion(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index 0874bc2fcdb4..b5098739f72a 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -121,7 +121,6 @@ void blk_rq_init(struct request_queue *q, struct request *rq)
 	rq->internal_tag = BLK_MQ_NO_TAG;
 	rq->start_time_ns = ktime_get_ns();
 	rq->part = NULL;
-	refcount_set(&rq->ref, 1);
 	blk_crypto_rq_set_defaults(rq);
 }
 EXPORT_SYMBOL(blk_rq_init);
diff --git a/block/blk-flush.c b/block/blk-flush.c
index 1002f6c58181..4912c8dbb1d8 100644
--- a/block/blk-flush.c
+++ b/block/blk-flush.c
@@ -329,6 +329,14 @@ static void blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq,
 	flush_rq->rq_flags |= RQF_FLUSH_SEQ;
 	flush_rq->rq_disk = first_rq->rq_disk;
 	flush_rq->end_io = flush_end_io;
+	/*
+	 * Order WRITE ->end_io and WRITE rq->ref, and its pair is the one
+	 * implied in refcount_inc_not_zero() called from
+	 * blk_mq_find_and_get_req(), which orders WRITE/READ flush_rq->ref
+	 * and READ flush_rq->end_io
+	 */
+	smp_wmb();
+	refcount_set(&flush_rq->ref, 1);
 
 	blk_flush_queue_rq(flush_rq, false);
 }
-- 
2.31.1

Re: [PATCH] blk-mq: fix kernel panic during iterating over flush request

[Christoph Hellwig]

Looks good,

Reviewed-by: Christoph Hellwig <hch@lst.de>

Re: [PATCH] blk-mq: fix kernel panic during iterating over flush request

[John Garry]

On 11/08/2021 15:26, Ming Lei wrote:
> For fixing use-after-free during iterating over requests, we grabbed
> request's refcount before calling ->fn in commit 2e315dc07df0 ("blk-mq:
> grab rq->refcount before calling ->fn in blk_mq_tagset_busy_iter").
> Turns out this way may cause kernel panic when iterating over one flush
> request:
> 
> 1) old flush request's tag is just released, and this tag is reused by
> one new request, but ->rqs[] isn't updated yet
> 
> 2) the flush request can be re-used for submitting one new flush command,
> so blk_rq_init() is called at the same time
> 
> 3) meantime blk_mq_queue_tag_busy_iter() is called, and old flush request
> is retrieved from ->rqs[tag]; when blk_mq_put_rq_ref() is called,
> flush_rq->end_io may not be updated yet, so NULL pointer dereference
> is triggered in blk_mq_put_rq_ref().
> 
> Fix the issue by calling refcount_set(&flush_rq->ref, 1) after
> flush_rq->end_io is set. So far the only other caller of blk_rq_init() is
> scsi_ioctl_reset() in which the request doesn't enter block IO stack and
> the request reference count isn't used, so the change is safe.
> 
> Fixes: 2e315dc07df0 ("blk-mq: grab rq->refcount before calling ->fn in
> blk_mq_tagset_busy_iter")
> Reported-by: "Blank-Burian, Markus, Dr." <blankburian@uni-muenster.de>
> Tested-by: "Blank-Burian, Markus, Dr." <blankburian@uni-muenster.de>
> Signed-off-by: Ming Lei <ming.lei@redhat.com>

it looks ok:

Reviewed-by: John Garry <john.garry@huawei.com>

> ---
>   block/blk-core.c  | 1 -
>   block/blk-flush.c | 8 ++++++++
>   2 files changed, 8 insertions(+), 1 deletion(-)
> 
> diff --git a/block/blk-core.c b/block/blk-core.c
> index 0874bc2fcdb4..b5098739f72a 100644
> --- a/block/blk-core.c
> +++ b/block/blk-core.c
> @@ -121,7 +121,6 @@ void blk_rq_init(struct request_queue *q, struct request *rq)
>   	rq->internal_tag = BLK_MQ_NO_TAG;
>   	rq->start_time_ns = ktime_get_ns();
>   	rq->part = NULL;
> -	refcount_set(&rq->ref, 1);
>   	blk_crypto_rq_set_defaults(rq);
>   }
>   EXPORT_SYMBOL(blk_rq_init);
> diff --git a/block/blk-flush.c b/block/blk-flush.c
> index 1002f6c58181..4912c8dbb1d8 100644
> --- a/block/blk-flush.c
> +++ b/block/blk-flush.c
> @@ -329,6 +329,14 @@ static void blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq,
>   	flush_rq->rq_flags |= RQF_FLUSH_SEQ;
>   	flush_rq->rq_disk = first_rq->rq_disk;
>   	flush_rq->end_io = flush_end_io;
> +	/*
> +	 * Order WRITE ->end_io and WRITE rq->ref, and its pair is the one
> +	 * implied in refcount_inc_not_zero() called from
> +	 * blk_mq_find_and_get_req(), which orders WRITE/READ flush_rq->ref
> +	 * and READ flush_rq->end_io
> +	 */
> +	smp_wmb();
> +	refcount_set(&flush_rq->ref, 1);

Re: [PATCH] blk-mq: fix kernel panic during iterating over flush request

[Xiaoguang Wang]

hi,


> For fixing use-after-free during iterating over requests, we grabbed
> request's refcount before calling ->fn in commit 2e315dc07df0 ("blk-mq:
> grab rq->refcount before calling ->fn in blk_mq_tagset_busy_iter").
> Turns out this way may cause kernel panic when iterating over one flush
> request:
>
> 1) old flush request's tag is just released, and this tag is reused by
> one new request, but ->rqs[] isn't updated yet
>
> 2) the flush request can be re-used for submitting one new flush command,
> so blk_rq_init() is called at the same time
>
> 3) meantime blk_mq_queue_tag_busy_iter() is called, and old flush request
> is retrieved from ->rqs[tag]; when blk_mq_put_rq_ref() is called,
> flush_rq->end_io may not be updated yet, so NULL pointer dereference
> is triggered in blk_mq_put_rq_ref().
>
> Fix the issue by calling refcount_set(&flush_rq->ref, 1) after
> flush_rq->end_io is set. So far the only other caller of blk_rq_init() is
> scsi_ioctl_reset() in which the request doesn't enter block IO stack and
> the request reference count isn't used, so the change is safe.
>
> Fixes: 2e315dc07df0 ("blk-mq: grab rq->refcount before calling ->fn in
> blk_mq_tagset_busy_iter")
> Reported-by: "Blank-Burian, Markus, Dr." <blankburian@uni-muenster.de>
> Tested-by: "Blank-Burian, Markus, Dr." <blankburian@uni-muenster.de>
> Signed-off-by: Ming Lei <ming.lei@redhat.com>
> ---
>   block/blk-core.c  | 1 -
>   block/blk-flush.c | 8 ++++++++
>   2 files changed, 8 insertions(+), 1 deletion(-)
>
> diff --git a/block/blk-core.c b/block/blk-core.c
> index 0874bc2fcdb4..b5098739f72a 100644
> --- a/block/blk-core.c
> +++ b/block/blk-core.c
> @@ -121,7 +121,6 @@ void blk_rq_init(struct request_queue *q, struct request *rq)
>   	rq->internal_tag = BLK_MQ_NO_TAG;
>   	rq->start_time_ns = ktime_get_ns();
>   	rq->part = NULL;
> -	refcount_set(&rq->ref, 1);
>   	blk_crypto_rq_set_defaults(rq);
>   }
>   EXPORT_SYMBOL(blk_rq_init);
> diff --git a/block/blk-flush.c b/block/blk-flush.c
> index 1002f6c58181..4912c8dbb1d8 100644
> --- a/block/blk-flush.c
> +++ b/block/blk-flush.c
> @@ -329,6 +329,14 @@ static void blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq,
>   	flush_rq->rq_flags |= RQF_FLUSH_SEQ;
>   	flush_rq->rq_disk = first_rq->rq_disk;
>   	flush_rq->end_io = flush_end_io;
> +	/*
> +	 * Order WRITE ->end_io and WRITE rq->ref, and its pair is the one
> +	 * implied in refcount_inc_not_zero() called from
> +	 * blk_mq_find_and_get_req(), which orders WRITE/READ flush_rq->ref
> +	 * and READ flush_rq->end_io

Recently we run into similar panic which is a NULL dereference on 
rq->mq_hctx in is_flush_rq(), we also

guess there is race bug just what you have fixed.

But I have one question here, for a blk-mq device, before issuing the 
first flush req, flush_rq->end_io

is NULL, and for following flush reqs on this blk-mq device, 
flush_rq->end_io won't be NULL. I searched

the codes and don't find any place that sets flush_rq->end_io to be NULL 
once flush_rq has been completed.


Regards,

Xiaoguang Wang

> +	 */
> +	smp_wmb();
> +	refcount_set(&flush_rq->ref, 1);
>   
>   	blk_flush_queue_rq(flush_rq, false);
>   }

Re: [PATCH] blk-mq: fix kernel panic during iterating over flush request

[Ming Lei]

On Thu, Aug 12, 2021 at 07:47:55PM +0800, Xiaoguang Wang wrote:
> hi,
> 
> 
> > For fixing use-after-free during iterating over requests, we grabbed
> > request's refcount before calling ->fn in commit 2e315dc07df0 ("blk-mq:
> > grab rq->refcount before calling ->fn in blk_mq_tagset_busy_iter").
> > Turns out this way may cause kernel panic when iterating over one flush
> > request:
> > 
> > 1) old flush request's tag is just released, and this tag is reused by
> > one new request, but ->rqs[] isn't updated yet
> > 
> > 2) the flush request can be re-used for submitting one new flush command,
> > so blk_rq_init() is called at the same time
> > 
> > 3) meantime blk_mq_queue_tag_busy_iter() is called, and old flush request
> > is retrieved from ->rqs[tag]; when blk_mq_put_rq_ref() is called,
> > flush_rq->end_io may not be updated yet, so NULL pointer dereference
> > is triggered in blk_mq_put_rq_ref().
> > 
> > Fix the issue by calling refcount_set(&flush_rq->ref, 1) after
> > flush_rq->end_io is set. So far the only other caller of blk_rq_init() is
> > scsi_ioctl_reset() in which the request doesn't enter block IO stack and
> > the request reference count isn't used, so the change is safe.
> > 
> > Fixes: 2e315dc07df0 ("blk-mq: grab rq->refcount before calling ->fn in
> > blk_mq_tagset_busy_iter")
> > Reported-by: "Blank-Burian, Markus, Dr." <blankburian@uni-muenster.de>
> > Tested-by: "Blank-Burian, Markus, Dr." <blankburian@uni-muenster.de>
> > Signed-off-by: Ming Lei <ming.lei@redhat.com>
> > ---
> >   block/blk-core.c  | 1 -
> >   block/blk-flush.c | 8 ++++++++
> >   2 files changed, 8 insertions(+), 1 deletion(-)
> > 
> > diff --git a/block/blk-core.c b/block/blk-core.c
> > index 0874bc2fcdb4..b5098739f72a 100644
> > --- a/block/blk-core.c
> > +++ b/block/blk-core.c
> > @@ -121,7 +121,6 @@ void blk_rq_init(struct request_queue *q, struct request *rq)
> >   	rq->internal_tag = BLK_MQ_NO_TAG;
> >   	rq->start_time_ns = ktime_get_ns();
> >   	rq->part = NULL;
> > -	refcount_set(&rq->ref, 1);
> >   	blk_crypto_rq_set_defaults(rq);
> >   }
> >   EXPORT_SYMBOL(blk_rq_init);
> > diff --git a/block/blk-flush.c b/block/blk-flush.c
> > index 1002f6c58181..4912c8dbb1d8 100644
> > --- a/block/blk-flush.c
> > +++ b/block/blk-flush.c
> > @@ -329,6 +329,14 @@ static void blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq,
> >   	flush_rq->rq_flags |= RQF_FLUSH_SEQ;
> >   	flush_rq->rq_disk = first_rq->rq_disk;
> >   	flush_rq->end_io = flush_end_io;
> > +	/*
> > +	 * Order WRITE ->end_io and WRITE rq->ref, and its pair is the one
> > +	 * implied in refcount_inc_not_zero() called from
> > +	 * blk_mq_find_and_get_req(), which orders WRITE/READ flush_rq->ref
> > +	 * and READ flush_rq->end_io
> 
> Recently we run into similar panic which is a NULL dereference on
> rq->mq_hctx in is_flush_rq(), we also
> 
> guess there is race bug just what you have fixed.
> 
> But I have one question here, for a blk-mq device, before issuing the first
> flush req, flush_rq->end_io
> 
> is NULL, and for following flush reqs on this blk-mq device,
> flush_rq->end_io won't be NULL. I searched
> 
> the codes and don't find any place that sets flush_rq->end_io to be NULL
> once flush_rq has been completed.

blk_kick_flush():
	blk_rq_init
		memset(rq, 0, sizeof(*rq));


Thanks,
Ming

Re: [PATCH] blk-mq: fix kernel panic during iterating over flush request

[Xiaoguang Wang]

hi,


> On Thu, Aug 12, 2021 at 07:47:55PM +0800, Xiaoguang Wang wrote:
>> hi,
>>
>>
>>> For fixing use-after-free during iterating over requests, we grabbed
>>> request's refcount before calling ->fn in commit 2e315dc07df0 ("blk-mq:
>>> grab rq->refcount before calling ->fn in blk_mq_tagset_busy_iter").
>>> Turns out this way may cause kernel panic when iterating over one flush
>>> request:
>>>
>>> 1) old flush request's tag is just released, and this tag is reused by
>>> one new request, but ->rqs[] isn't updated yet
>>>
>>> 2) the flush request can be re-used for submitting one new flush command,
>>> so blk_rq_init() is called at the same time
>>>
>>> 3) meantime blk_mq_queue_tag_busy_iter() is called, and old flush request
>>> is retrieved from ->rqs[tag]; when blk_mq_put_rq_ref() is called,
>>> flush_rq->end_io may not be updated yet, so NULL pointer dereference
>>> is triggered in blk_mq_put_rq_ref().
>>>
>>> Fix the issue by calling refcount_set(&flush_rq->ref, 1) after
>>> flush_rq->end_io is set. So far the only other caller of blk_rq_init() is
>>> scsi_ioctl_reset() in which the request doesn't enter block IO stack and
>>> the request reference count isn't used, so the change is safe.
>>>
>>> Fixes: 2e315dc07df0 ("blk-mq: grab rq->refcount before calling ->fn in
>>> blk_mq_tagset_busy_iter")
>>> Reported-by: "Blank-Burian, Markus, Dr." <blankburian@uni-muenster.de>
>>> Tested-by: "Blank-Burian, Markus, Dr." <blankburian@uni-muenster.de>
>>> Signed-off-by: Ming Lei <ming.lei@redhat.com>
>>> ---
>>>    block/blk-core.c  | 1 -
>>>    block/blk-flush.c | 8 ++++++++
>>>    2 files changed, 8 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/block/blk-core.c b/block/blk-core.c
>>> index 0874bc2fcdb4..b5098739f72a 100644
>>> --- a/block/blk-core.c
>>> +++ b/block/blk-core.c
>>> @@ -121,7 +121,6 @@ void blk_rq_init(struct request_queue *q, struct request *rq)
>>>    	rq->internal_tag = BLK_MQ_NO_TAG;
>>>    	rq->start_time_ns = ktime_get_ns();
>>>    	rq->part = NULL;
>>> -	refcount_set(&rq->ref, 1);
>>>    	blk_crypto_rq_set_defaults(rq);
>>>    }
>>>    EXPORT_SYMBOL(blk_rq_init);
>>> diff --git a/block/blk-flush.c b/block/blk-flush.c
>>> index 1002f6c58181..4912c8dbb1d8 100644
>>> --- a/block/blk-flush.c
>>> +++ b/block/blk-flush.c
>>> @@ -329,6 +329,14 @@ static void blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq,
>>>    	flush_rq->rq_flags |= RQF_FLUSH_SEQ;
>>>    	flush_rq->rq_disk = first_rq->rq_disk;
>>>    	flush_rq->end_io = flush_end_io;
>>> +	/*
>>> +	 * Order WRITE ->end_io and WRITE rq->ref, and its pair is the one
>>> +	 * implied in refcount_inc_not_zero() called from
>>> +	 * blk_mq_find_and_get_req(), which orders WRITE/READ flush_rq->ref
>>> +	 * and READ flush_rq->end_io
>> Recently we run into similar panic which is a NULL dereference on
>> rq->mq_hctx in is_flush_rq(), we also
>>
>> guess there is race bug just what you have fixed.
>>
>> But I have one question here, for a blk-mq device, before issuing the first
>> flush req, flush_rq->end_io
>>
>> is NULL, and for following flush reqs on this blk-mq device,
>> flush_rq->end_io won't be NULL. I searched
>>
>> the codes and don't find any place that sets flush_rq->end_io to be NULL
>> once flush_rq has been completed.
> blk_kick_flush():
> 	blk_rq_init
> 		memset(rq, 0, sizeof(*rq));

Oh, thanks.

I searched codes for quite a while...


Regards,

Xiaoguang Wang

>
>
> Thanks,
> Ming

Re: [PATCH] blk-mq: fix kernel panic during iterating over flush request

[Ming Lei]

On Wed, Aug 11, 2021 at 10:26:24PM +0800, Ming Lei wrote:
> For fixing use-after-free during iterating over requests, we grabbed
> request's refcount before calling ->fn in commit 2e315dc07df0 ("blk-mq:
> grab rq->refcount before calling ->fn in blk_mq_tagset_busy_iter").
> Turns out this way may cause kernel panic when iterating over one flush
> request:
> 
> 1) old flush request's tag is just released, and this tag is reused by
> one new request, but ->rqs[] isn't updated yet
> 
> 2) the flush request can be re-used for submitting one new flush command,
> so blk_rq_init() is called at the same time
> 
> 3) meantime blk_mq_queue_tag_busy_iter() is called, and old flush request
> is retrieved from ->rqs[tag]; when blk_mq_put_rq_ref() is called,
> flush_rq->end_io may not be updated yet, so NULL pointer dereference
> is triggered in blk_mq_put_rq_ref().
> 
> Fix the issue by calling refcount_set(&flush_rq->ref, 1) after
> flush_rq->end_io is set. So far the only other caller of blk_rq_init() is
> scsi_ioctl_reset() in which the request doesn't enter block IO stack and
> the request reference count isn't used, so the change is safe.
> 
> Fixes: 2e315dc07df0 ("blk-mq: grab rq->refcount before calling ->fn in
> blk_mq_tagset_busy_iter")
> Reported-by: "Blank-Burian, Markus, Dr." <blankburian@uni-muenster.de>
> Tested-by: "Blank-Burian, Markus, Dr." <blankburian@uni-muenster.de>
> Signed-off-by: Ming Lei <ming.lei@redhat.com>

Hello Jens,

Ping...

Thanks,
Ming

Re: [PATCH] blk-mq: fix kernel panic during iterating over flush request

[Jens Axboe]

On 8/11/21 8:26 AM, Ming Lei wrote:
> For fixing use-after-free during iterating over requests, we grabbed
> request's refcount before calling ->fn in commit 2e315dc07df0 ("blk-mq:
> grab rq->refcount before calling ->fn in blk_mq_tagset_busy_iter").
> Turns out this way may cause kernel panic when iterating over one flush
> request:
> 
> 1) old flush request's tag is just released, and this tag is reused by
> one new request, but ->rqs[] isn't updated yet
> 
> 2) the flush request can be re-used for submitting one new flush command,
> so blk_rq_init() is called at the same time
> 
> 3) meantime blk_mq_queue_tag_busy_iter() is called, and old flush request
> is retrieved from ->rqs[tag]; when blk_mq_put_rq_ref() is called,
> flush_rq->end_io may not be updated yet, so NULL pointer dereference
> is triggered in blk_mq_put_rq_ref().
> 
> Fix the issue by calling refcount_set(&flush_rq->ref, 1) after
> flush_rq->end_io is set. So far the only other caller of blk_rq_init() is
> scsi_ioctl_reset() in which the request doesn't enter block IO stack and
> the request reference count isn't used, so the change is safe.

Applied, thanks.

-- 
Jens Axboe