Skip to contentSkip to conversation

bfq: Check if bfqq is NULL in bfq_insert_request

PATCH In review Guenter Roeck 3 participants 10 messages Jul 22, 2019 – Jul 31, 2019

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 bfq: Check if bfqq is NULL in bfq_insert_request

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")
Cc Hsin-Yi Wang <[email protected]>
Cc Nicolas Boichat <[email protected]>
Cc Doug Anderson <[email protected]>
---
 block/bfq-iosched.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
block/bfq-iosched.c +1 -1
index 72860325245a..56f3f4227010 100644
@@ -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
Signature · 1 line
2.7.4
#1 lore raw
Quoted 8 lines · Guenter Roeck

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.

The above function should have been ioc_create_icq(), sorry.

Guenter

Quoted 63 lines

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")
Cc Hsin-Yi Wang <[email protected]>
Cc Nicolas Boichat <[email protected]>
Cc Doug Anderson <[email protected]>
---
 block/bfq-iosched.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
block/bfq-iosched.c +1 -1
index 72860325245a..56f3f4227010 100644
@@ -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
Signature · 1 line
2.7.4
#2 lore raw
Quoted 10 lines · Guenter Roeck
Quoted 8 lines · Guenter Roeck

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.

Quoted 41 lines

The above function should have been ioc_create_icq(), sorry.

Guenter

Quoted 37 lines

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.

What about other place use blk_init_rq()?
E.g in bfq_request_merged():
1897 struct bfq_queue *bfqq = bfq_init_rq(req);
1898 struct bfq_data *bfqd = bfqq->bfqd;

Which may have same Null-pointer bug?

-Bob

Quoted 26 lines
Quoted 26 lines
Fixes 18e5a57d79878 ("block, bfq: postpone rq preparation to insert or merge")
Cc Hsin-Yi Wang <[email protected]>
Cc Nicolas Boichat <[email protected]>
Cc Doug Anderson <[email protected]>
---
 block/bfq-iosched.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
block/bfq-iosched.c +1 -1
index 72860325245a..56f3f4227010 100644
@@ -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
Signature · 1 line
2.7.4
#3 lore raw
Quoted 54 lines · Bob Liu
Quoted 10 lines · Guenter Roeck
Quoted 8 lines · Guenter Roeck

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.

Quoted 41 lines

The above function should have been ioc_create_icq(), sorry.

Guenter

Quoted 37 lines

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.

Quoted 7 lines

What about other place use blk_init_rq()?
E.g in bfq_request_merged():
1897 struct bfq_queue *bfqq = bfq_init_rq(req);
1898 struct bfq_data *bfqd = bfqq->bfqd;

Which may have same Null-pointer bug?

Good question. Doug asked it as well. My understanding is that the code won't hit those places since bfq is essentially bypassed. Of course, I may be completely wrong, so we should wait for Paolo to confirm if I my understanding is correct (or wrong).

Thanks,
Guenter

Quoted 29 lines

-Bob

Quoted 26 lines
Quoted 26 lines
Fixes 18e5a57d79878 ("block, bfq: postpone rq preparation to insert or merge")
Cc Hsin-Yi Wang <[email protected]>
Cc Nicolas Boichat <[email protected]>
Cc Doug Anderson <[email protected]>
---
 block/bfq-iosched.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
block/bfq-iosched.c +1 -1
index 72860325245a..56f3f4227010 100644
@@ -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
Signature · 1 line
2.7.4
#4 lore raw
In reply to Guenter Roeck

ping ... just in case this patch got lost in Paolo's queue.

Guenter

Quoted 71 lines · Guenter Roeck

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")
Cc Hsin-Yi Wang <[email protected]>
Cc Nicolas Boichat <[email protected]>
Cc Doug Anderson <[email protected]>
---
 block/bfq-iosched.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
block/bfq-iosched.c +1 -1
index 72860325245a..56f3f4227010 100644
@@ -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
Signature · 1 line
2.7.4
#5 lore raw

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.

If my considerations above are correct, do you want to propose a complete fix yourself?

Thanks,
Paolo

Quoted 78 lines

Il giorno 28 lug 2019, alle ore 17:19, Guenter Roeck <[email protected]> ha scritto:

ping ... just in case this patch got lost in Paolo's queue.

Guenter

Quoted 71 lines · Guenter Roeck

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")
Cc Hsin-Yi Wang <[email protected]>
Cc Nicolas Boichat <[email protected]>
Cc Doug Anderson <[email protected]>

---
block/bfq-iosched.c | 2 +-

1 file changed, 1 insertion(+), 1 deletion(-)
block/bfq-iosched.c +1 -1
index 72860325245a..56f3f4227010 100644
@@ -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
Signature · 1 line
2.7.4
#6 lore raw
Quoted 6 lines · Paolo Valente

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.

Quoted 11 lines

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.

Quoted 3 lines

If my considerations above are correct, do you want to propose a complete fix yourself?

Sure, I'll send an updated patch.

Thanks,
Guenter

Quoted 83 lines

Thanks,
Paolo

Quoted 78 lines

Il giorno 28 lug 2019, alle ore 17:19, Guenter Roeck <[email protected]> ha scritto:

ping ... just in case this patch got lost in Paolo's queue.

Guenter

Quoted 71 lines · Guenter Roeck

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")
Cc Hsin-Yi Wang <[email protected]>
Cc Nicolas Boichat <[email protected]>
Cc Doug Anderson <[email protected]>

---
block/bfq-iosched.c | 2 +-

1 file changed, 1 insertion(+), 1 deletion(-)
block/bfq-iosched.c +1 -1
index 72860325245a..56f3f4227010 100644
@@ -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
Signature · 1 line
2.7.4
#7 lore raw
Quoted 19 lines

Il giorno 30 lug 2019, alle ore 15:35, Guenter Roeck <[email protected]> ha scritto:

Quoted 6 lines · Paolo Valente

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

Quoted 8 lines

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.

Quoted 4 lines

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.

I don't remember such a filtering. I had a look again, but didn't find anything relevant. However, more competent people see these emails. Maybe someone can give us better advice. Otherwise, to stay on the safe side, I'd propose to handle any possible NULL return.

And I'll manage it, as per your request.

Thanks,
Paolo

Quoted 90 lines
Quoted 3 lines

If my considerations above are correct, do you want to propose a complete fix yourself?

Sure, I'll send an updated patch.

Thanks,
Guenter

Quoted 80 lines

Thanks,
Paolo

Quoted 78 lines

Il giorno 28 lug 2019, alle ore 17:19, Guenter Roeck <[email protected]> ha scritto:

ping ... just in case this patch got lost in Paolo's queue.

Guenter

Quoted 71 lines · Guenter Roeck

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")
Cc Hsin-Yi Wang <[email protected]>
Cc Nicolas Boichat <[email protected]>
Cc Doug Anderson <[email protected]>

---
block/bfq-iosched.c | 2 +-

1 file changed, 1 insertion(+), 1 deletion(-)
block/bfq-iosched.c +1 -1
index 72860325245a..56f3f4227010 100644
@@ -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
Signature · 1 line
2.7.4
#8 lore raw
Quoted 19 lines · Paolo Valente
Quoted 19 lines

Il giorno 30 lug 2019, alle ore 15:35, Guenter Roeck <[email protected]> ha scritto:

Quoted 14 lines · Paolo Valente

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.

Quoted 7 lines
Quoted 4 lines

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.

I don't remember such a filtering. I had a look again, but didn't find anything relevant. However, more competent people see these

Me not either, when I had a closer look yesterday. My conclusion was that your analysis is correct.

Thanks,
Guenter

Quoted 100 lines

emails. Maybe someone can give us better advice. Otherwise, to stay on the safe side, I'd propose to handle any possible NULL return.

And I'll manage it, as per your request.

Thanks,
Paolo

Quoted 90 lines
Quoted 3 lines

If my considerations above are correct, do you want to propose a complete fix yourself?

Sure, I'll send an updated patch.

Thanks,
Guenter

Quoted 80 lines

Thanks,
Paolo

Quoted 78 lines

Il giorno 28 lug 2019, alle ore 17:19, Guenter Roeck <[email protected]> ha scritto:

ping ... just in case this patch got lost in Paolo's queue.

Guenter

Quoted 71 lines · Guenter Roeck

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")
Cc Hsin-Yi Wang <[email protected]>
Cc Nicolas Boichat <[email protected]>
Cc Doug Anderson <[email protected]>

---
block/bfq-iosched.c | 2 +-

1 file changed, 1 insertion(+), 1 deletion(-)
block/bfq-iosched.c +1 -1
index 72860325245a..56f3f4227010 100644
@@ -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
Signature · 1 line
2.7.4
#9 lore raw
In reply to Paolo Valente

Paolo,

Quoted 17 lines · Paolo Valente

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.

Quoted 3 lines

If my considerations above are correct, do you want to propose a complete fix yourself?

I had another look into the code. Unfortunately, both bfq_request_merged() and bfq_requests_merged() simply assume that bfq_init_rq() never returns NULL, and don't give me an idea for a path of action if it returns NULL after all. I'll have to pass the problem off to you for a fix.

Thanks,
Guenter