[v3,2/4] block-backend: fix edge case in bdrv_next() where BDS associated to BB changes

The old_bs variable in bdrv_next() is currently determined by looking
at the old block backend. However, if the block graph changes before
the next bdrv_next() call, it might be that the associated BDS is not
the same that was referenced previously. In that case, the wrong BDS
is unreferenced, leading to an assertion failure later:

> bdrv_unref: Assertion `bs->refcnt > 0' failed.

In particular, this can happen in the context of bdrv_flush_all(),
when polling for bdrv_co_flush() in the generated co-wrapper leads to
a graph change (for example with a stream block job [0]).

A racy reproducer:

> #!/bin/bash
> rm -f /tmp/backing.qcow2
> rm -f /tmp/top.qcow2
> ./qemu-img create /tmp/backing.qcow2 -f qcow2 64M
> ./qemu-io -c "write -P42 0x0 0x1" /tmp/backing.qcow2
> ./qemu-img create /tmp/top.qcow2 -f qcow2 64M -b /tmp/backing.qcow2 -F qcow2
> ./qemu-system-x86_64 --qmp stdio \
> --blockdev qcow2,node-name=node0,file.driver=file,file.filename=/tmp/top.qcow2 \
> <<EOF
> {"execute": "qmp_capabilities"}
> {"execute": "block-stream", "arguments": { "job-id": "stream0", "device": "node0" } }
> {"execute": "quit"}
> EOF

[0]:

> #0  bdrv_replace_child_tran (child=..., new_bs=..., tran=...)
> #1  bdrv_replace_node_noperm (from=..., to=..., auto_skip=..., tran=..., errp=...)
> #2  bdrv_replace_node_common (from=..., to=..., auto_skip=..., detach_subchain=..., errp=...)
> #3  bdrv_drop_filter (bs=..., errp=...)
> #4  bdrv_cor_filter_drop (cor_filter_bs=...)
> #5  stream_prepare (job=...)
> #6  job_prepare_locked (job=...)
> #7  job_txn_apply_locked (fn=..., job=...)
> #8  job_do_finalize_locked (job=...)
> #9  job_exit (opaque=...)
> #10 aio_bh_poll (ctx=...)
> #11 aio_poll (ctx=..., blocking=...)
> #12 bdrv_poll_co (s=...)
> #13 bdrv_flush (bs=...)
> #14 bdrv_flush_all ()
> #15 do_vm_stop (state=..., send_stop=...)
> #16 vm_shutdown ()

Signed-off-by: Fiona Ebner <f.ebner@proxmox.com>
---

No changes in v3.
New in v2.

 block/block-backend.c | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

On Fri, Mar 22, 2024 at 10:50:07AM +0100, Fiona Ebner wrote:
> The old_bs variable in bdrv_next() is currently determined by looking
> at the old block backend. However, if the block graph changes before
> the next bdrv_next() call, it might be that the associated BDS is not
> the same that was referenced previously. In that case, the wrong BDS
> is unreferenced, leading to an assertion failure later:
> 
> > bdrv_unref: Assertion `bs->refcnt > 0' failed.
> 
> In particular, this can happen in the context of bdrv_flush_all(),
> when polling for bdrv_co_flush() in the generated co-wrapper leads to
> a graph change (for example with a stream block job [0]).
> 
> A racy reproducer:
> 
> > #!/bin/bash
> > rm -f /tmp/backing.qcow2
> > rm -f /tmp/top.qcow2
> > ./qemu-img create /tmp/backing.qcow2 -f qcow2 64M
> > ./qemu-io -c "write -P42 0x0 0x1" /tmp/backing.qcow2
> > ./qemu-img create /tmp/top.qcow2 -f qcow2 64M -b /tmp/backing.qcow2 -F qcow2
> > ./qemu-system-x86_64 --qmp stdio \
> > --blockdev qcow2,node-name=node0,file.driver=file,file.filename=/tmp/top.qcow2 \
> > <<EOF
> > {"execute": "qmp_capabilities"}
> > {"execute": "block-stream", "arguments": { "job-id": "stream0", "device": "node0" } }
> > {"execute": "quit"}
> > EOF
> 
> [0]:
> 
> > #0  bdrv_replace_child_tran (child=..., new_bs=..., tran=...)
> > #1  bdrv_replace_node_noperm (from=..., to=..., auto_skip=..., tran=..., errp=...)
> > #2  bdrv_replace_node_common (from=..., to=..., auto_skip=..., detach_subchain=..., errp=...)
> > #3  bdrv_drop_filter (bs=..., errp=...)
> > #4  bdrv_cor_filter_drop (cor_filter_bs=...)
> > #5  stream_prepare (job=...)
> > #6  job_prepare_locked (job=...)
> > #7  job_txn_apply_locked (fn=..., job=...)
> > #8  job_do_finalize_locked (job=...)
> > #9  job_exit (opaque=...)
> > #10 aio_bh_poll (ctx=...)
> > #11 aio_poll (ctx=..., blocking=...)
> > #12 bdrv_poll_co (s=...)
> > #13 bdrv_flush (bs=...)
> > #14 bdrv_flush_all ()
> > #15 do_vm_stop (state=..., send_stop=...)
> > #16 vm_shutdown ()
> 
> Signed-off-by: Fiona Ebner <f.ebner@proxmox.com>
> ---
> 
> No changes in v3.
> New in v2.
> 
>  block/block-backend.c | 7 +++----
>  1 file changed, 3 insertions(+), 4 deletions(-)

Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>

Am 22.03.2024 um 10:50 hat Fiona Ebner geschrieben:
> The old_bs variable in bdrv_next() is currently determined by looking
> at the old block backend. However, if the block graph changes before
> the next bdrv_next() call, it might be that the associated BDS is not
> the same that was referenced previously. In that case, the wrong BDS
> is unreferenced, leading to an assertion failure later:
> 
> > bdrv_unref: Assertion `bs->refcnt > 0' failed.

Your change makes sense, but in theory it shouldn't make a difference.
The real bug is in the caller, you can't allow graph modifications while
iterating the list of nodes. Even if it doesn't crash (like after your
patch), you don't have any guarantee that you will have seen every node
that exists that the end - and maybe not even that you don't get the
same node twice.

> In particular, this can happen in the context of bdrv_flush_all(),
> when polling for bdrv_co_flush() in the generated co-wrapper leads to
> a graph change (for example with a stream block job [0]).

The whole locking around this case is a bit tricky and would deserve
some cleanup.

The basic rule for bdrv_next() callers is that they need to hold the
graph reader lock as long as they are iterating the graph, otherwise
it's not safe. This implies that the ref/unref pairs in it should never
make a difference either - which is important, because at least
releasing the last reference is forbidden while holding the graph lock.
I intended to remove the ref/unref for bdrv_next(), but I didn't because
I realised that the callers need to be audited first that they really
obey the rules. You found one that would be problematic.

The thing that bdrv_flush_all() gets wrong is that it promises to follow
the graph lock rules with GRAPH_RDLOCK_GUARD_MAINLOOP(), but then calls
something that polls. The compiler can't catch this because bdrv_flush()
is a co_wrapper_mixed_bdrv_rdlock. The behaviour for these functions is:

- If called outside of coroutine context, they are GRAPH_UNLOCKED
- If called in coroutine context, they are GRAPH_RDLOCK

We should probably try harder to get rid of the mixed functions, because
a synchronous co_wrapper_bdrv_rdlock could actually be marked
GRAPH_UNLOCKED in the code and then the compiler could catch this case.

The fix for bdrv_flush_all() is probably to make it bdrv_co_flush_all()
with a coroutine wrapper so that the graph lock is held for the whole
function. Then calling bdrv_co_flush() while iterating the list is safe
and doesn't allow concurrent graph modifications.

Kevin

Hi Kevin,

Am 26.03.24 um 13:44 schrieb Kevin Wolf:
> Am 22.03.2024 um 10:50 hat Fiona Ebner geschrieben:
>> The old_bs variable in bdrv_next() is currently determined by looking
>> at the old block backend. However, if the block graph changes before
>> the next bdrv_next() call, it might be that the associated BDS is not
>> the same that was referenced previously. In that case, the wrong BDS
>> is unreferenced, leading to an assertion failure later:
>>
>>> bdrv_unref: Assertion `bs->refcnt > 0' failed.
> 
> Your change makes sense, but in theory it shouldn't make a difference.
> The real bug is in the caller, you can't allow graph modifications while
> iterating the list of nodes. Even if it doesn't crash (like after your
> patch), you don't have any guarantee that you will have seen every node
> that exists that the end - and maybe not even that you don't get the
> same node twice.
> 
>> In particular, this can happen in the context of bdrv_flush_all(),
>> when polling for bdrv_co_flush() in the generated co-wrapper leads to
>> a graph change (for example with a stream block job [0]).
> 
> The whole locking around this case is a bit tricky and would deserve
> some cleanup.
> 
> The basic rule for bdrv_next() callers is that they need to hold the
> graph reader lock as long as they are iterating the graph, otherwise
> it's not safe. This implies that the ref/unref pairs in it should never
> make a difference either - which is important, because at least
> releasing the last reference is forbidden while holding the graph lock.
> I intended to remove the ref/unref for bdrv_next(), but I didn't because
> I realised that the callers need to be audited first that they really
> obey the rules. You found one that would be problematic.
> 
> The thing that bdrv_flush_all() gets wrong is that it promises to follow
> the graph lock rules with GRAPH_RDLOCK_GUARD_MAINLOOP(), but then calls
> something that polls. The compiler can't catch this because bdrv_flush()
> is a co_wrapper_mixed_bdrv_rdlock. The behaviour for these functions is:
> 
> - If called outside of coroutine context, they are GRAPH_UNLOCKED
> - If called in coroutine context, they are GRAPH_RDLOCK
> 
> We should probably try harder to get rid of the mixed functions, because
> a synchronous co_wrapper_bdrv_rdlock could actually be marked
> GRAPH_UNLOCKED in the code and then the compiler could catch this case.
> 
> The fix for bdrv_flush_all() is probably to make it bdrv_co_flush_all()
> with a coroutine wrapper so that the graph lock is held for the whole
> function. Then calling bdrv_co_flush() while iterating the list is safe
> and doesn't allow concurrent graph modifications.

I attempted this now, but ran into two issues:

The first is that I had to add support for a function without arguments
to the block-coroutine-wrapper.py script. I can send this as an RFC in
any case if desired.

The second is that iotest 255 ran into an assertion failure upon QMP 'quit':

> ../block/graph-lock.c:113: bdrv_graph_wrlock: Assertion `!qemu_in_coroutine()' failed.

Looking at the backtrace:

> #5  0x0000762a90cc3eb2 in __GI___assert_fail
>     (assertion=0x5afb07991e7d "!qemu_in_coroutine()", file=0x5afb07991e00 "../block/graph-lock.c", line=113, function=0x5afb07991f20 <__PRETTY_FUNCTION__.4> "bdrv_graph_wrlock")
>     at ./assert/assert.c:101
> #6  0x00005afb07585311 in bdrv_graph_wrlock () at ../block/graph-lock.c:113
> #7  0x00005afb07573a36 in blk_remove_bs (blk=0x5afb0af99420) at ../block/block-backend.c:901
> #8  0x00005afb075729a7 in blk_delete (blk=0x5afb0af99420) at ../block/block-backend.c:487
> #9  0x00005afb07572d88 in blk_unref (blk=0x5afb0af99420) at ../block/block-backend.c:547
> #10 0x00005afb07572fe8 in bdrv_next (it=0x762a852fef00) at ../block/block-backend.c:618
> #11 0x00005afb0758cd65 in bdrv_co_flush_all () at ../block/io.c:2347
> #12 0x00005afb0753ba37 in bdrv_co_flush_all_entry (opaque=0x7ffff12c6050) at block/block-gen.c:1391
> #13 0x00005afb0773bf41 in coroutine_trampoline (i0=168365184, i1=23291)

So I guess calling bdrv_next() is not safe from a coroutine, because the
function doing the iteration could end up being the last thing to have a
reference for the BB.

Best Regards,
Fiona

Am 03.06.2024 um 16:17 hat Fiona Ebner geschrieben:
> Hi Kevin,
> 
> Am 26.03.24 um 13:44 schrieb Kevin Wolf:
> > Am 22.03.2024 um 10:50 hat Fiona Ebner geschrieben:
> >> The old_bs variable in bdrv_next() is currently determined by looking
> >> at the old block backend. However, if the block graph changes before
> >> the next bdrv_next() call, it might be that the associated BDS is not
> >> the same that was referenced previously. In that case, the wrong BDS
> >> is unreferenced, leading to an assertion failure later:
> >>
> >>> bdrv_unref: Assertion `bs->refcnt > 0' failed.
> > 
> > Your change makes sense, but in theory it shouldn't make a difference.
> > The real bug is in the caller, you can't allow graph modifications while
> > iterating the list of nodes. Even if it doesn't crash (like after your
> > patch), you don't have any guarantee that you will have seen every node
> > that exists that the end - and maybe not even that you don't get the
> > same node twice.
> > 
> >> In particular, this can happen in the context of bdrv_flush_all(),
> >> when polling for bdrv_co_flush() in the generated co-wrapper leads to
> >> a graph change (for example with a stream block job [0]).
> > 
> > The whole locking around this case is a bit tricky and would deserve
> > some cleanup.
> > 
> > The basic rule for bdrv_next() callers is that they need to hold the
> > graph reader lock as long as they are iterating the graph, otherwise
> > it's not safe. This implies that the ref/unref pairs in it should never
> > make a difference either - which is important, because at least
> > releasing the last reference is forbidden while holding the graph lock.
> > I intended to remove the ref/unref for bdrv_next(), but I didn't because
> > I realised that the callers need to be audited first that they really
> > obey the rules. You found one that would be problematic.
> > 
> > The thing that bdrv_flush_all() gets wrong is that it promises to follow
> > the graph lock rules with GRAPH_RDLOCK_GUARD_MAINLOOP(), but then calls
> > something that polls. The compiler can't catch this because bdrv_flush()
> > is a co_wrapper_mixed_bdrv_rdlock. The behaviour for these functions is:
> > 
> > - If called outside of coroutine context, they are GRAPH_UNLOCKED
> > - If called in coroutine context, they are GRAPH_RDLOCK
> > 
> > We should probably try harder to get rid of the mixed functions, because
> > a synchronous co_wrapper_bdrv_rdlock could actually be marked
> > GRAPH_UNLOCKED in the code and then the compiler could catch this case.
> > 
> > The fix for bdrv_flush_all() is probably to make it bdrv_co_flush_all()
> > with a coroutine wrapper so that the graph lock is held for the whole
> > function. Then calling bdrv_co_flush() while iterating the list is safe
> > and doesn't allow concurrent graph modifications.
> 
> I attempted this now, but ran into two issues:
> 
> The first is that I had to add support for a function without arguments
> to the block-coroutine-wrapper.py script. I can send this as an RFC in
> any case if desired.

I think I wouldn't necessarily merge it if we don't have code that makes
use of it, but having it on the list as an RFC can't hurt either way.
Then we can come back to it once we do need it for something.

> The second is that iotest 255 ran into an assertion failure upon QMP 'quit':
> 
> > ../block/graph-lock.c:113: bdrv_graph_wrlock: Assertion `!qemu_in_coroutine()' failed.
> 
> Looking at the backtrace:
> 
> > #5  0x0000762a90cc3eb2 in __GI___assert_fail
> >     (assertion=0x5afb07991e7d "!qemu_in_coroutine()", file=0x5afb07991e00 "../block/graph-lock.c", line=113, function=0x5afb07991f20 <__PRETTY_FUNCTION__.4> "bdrv_graph_wrlock")
> >     at ./assert/assert.c:101
> > #6  0x00005afb07585311 in bdrv_graph_wrlock () at ../block/graph-lock.c:113
> > #7  0x00005afb07573a36 in blk_remove_bs (blk=0x5afb0af99420) at ../block/block-backend.c:901
> > #8  0x00005afb075729a7 in blk_delete (blk=0x5afb0af99420) at ../block/block-backend.c:487
> > #9  0x00005afb07572d88 in blk_unref (blk=0x5afb0af99420) at ../block/block-backend.c:547
> > #10 0x00005afb07572fe8 in bdrv_next (it=0x762a852fef00) at ../block/block-backend.c:618
> > #11 0x00005afb0758cd65 in bdrv_co_flush_all () at ../block/io.c:2347
> > #12 0x00005afb0753ba37 in bdrv_co_flush_all_entry (opaque=0x7ffff12c6050) at block/block-gen.c:1391
> > #13 0x00005afb0773bf41 in coroutine_trampoline (i0=168365184, i1=23291)
> 
> So I guess calling bdrv_next() is not safe from a coroutine, because
> the function doing the iteration could end up being the last thing to
> have a reference for the BB.

Does your bdrv_co_flush_all() take the graph (reader) lock? If so, this
is surprising, because while we hold the graph lock, no reference should
be able to go away - you need the writer lock for that and you won't get
it as long as bdrv_co_flush_all() locks the graph. So whatever had a
reference before the bdrv_next() loop must still have it now. Do you
know where it gets dropped?

Kevin

Am 03.06.24 um 18:21 schrieb Kevin Wolf:
> Am 03.06.2024 um 16:17 hat Fiona Ebner geschrieben:
>> Am 26.03.24 um 13:44 schrieb Kevin Wolf:
>>>
>>> The fix for bdrv_flush_all() is probably to make it bdrv_co_flush_all()
>>> with a coroutine wrapper so that the graph lock is held for the whole
>>> function. Then calling bdrv_co_flush() while iterating the list is safe
>>> and doesn't allow concurrent graph modifications.
>>
>> The second is that iotest 255 ran into an assertion failure upon QMP 'quit':
>>
>>> ../block/graph-lock.c:113: bdrv_graph_wrlock: Assertion `!qemu_in_coroutine()' failed.
>>
>> Looking at the backtrace:
>>
>>> #5  0x0000762a90cc3eb2 in __GI___assert_fail
>>>     (assertion=0x5afb07991e7d "!qemu_in_coroutine()", file=0x5afb07991e00 "../block/graph-lock.c", line=113, function=0x5afb07991f20 <__PRETTY_FUNCTION__.4> "bdrv_graph_wrlock")
>>>     at ./assert/assert.c:101
>>> #6  0x00005afb07585311 in bdrv_graph_wrlock () at ../block/graph-lock.c:113
>>> #7  0x00005afb07573a36 in blk_remove_bs (blk=0x5afb0af99420) at ../block/block-backend.c:901
>>> #8  0x00005afb075729a7 in blk_delete (blk=0x5afb0af99420) at ../block/block-backend.c:487
>>> #9  0x00005afb07572d88 in blk_unref (blk=0x5afb0af99420) at ../block/block-backend.c:547
>>> #10 0x00005afb07572fe8 in bdrv_next (it=0x762a852fef00) at ../block/block-backend.c:618
>>> #11 0x00005afb0758cd65 in bdrv_co_flush_all () at ../block/io.c:2347
>>> #12 0x00005afb0753ba37 in bdrv_co_flush_all_entry (opaque=0x7ffff12c6050) at block/block-gen.c:1391
>>> #13 0x00005afb0773bf41 in coroutine_trampoline (i0=168365184, i1=23291)
>>
>> So I guess calling bdrv_next() is not safe from a coroutine, because
>> the function doing the iteration could end up being the last thing to
>> have a reference for the BB.
> 
> Does your bdrv_co_flush_all() take the graph (reader) lock? If so, this
> is surprising, because while we hold the graph lock, no reference should
> be able to go away - you need the writer lock for that and you won't get
> it as long as bdrv_co_flush_all() locks the graph. So whatever had a
> reference before the bdrv_next() loop must still have it now. Do you
> know where it gets dropped?
> 

AFAICT, yes, it does hold the graph reader lock. The generated code is:

> static void coroutine_fn bdrv_co_flush_all_entry(void *opaque)
> {
>     BdrvFlushAll *s = opaque;
> 
>     bdrv_graph_co_rdlock();
>     s->ret = bdrv_co_flush_all();
>     bdrv_graph_co_rdunlock();
>     s->poll_state.in_progress = false;
> 
>     aio_wait_kick();
> }

Apparently when the mirror job is aborted/exits, which can happen during
the polling for bdrv_co_flush_all_entry(), a reference can go away
without the write lock (at least my breakpoints didn't trigger) being held:

> #0  blk_unref (blk=0x5cdefe943d20) at ../block/block-backend.c:537
> #1  0x00005cdefb26697e in mirror_exit_common (job=0x5cdefeb53000) at ../block/mirror.c:710
> #2  0x00005cdefb263575 in mirror_abort (job=0x5cdefeb53000) at ../block/mirror.c:823
> #3  0x00005cdefb2248a6 in job_abort (job=0x5cdefeb53000) at ../job.c:825
> #4  0x00005cdefb2245f2 in job_finalize_single_locked (job=0x5cdefeb53000) at ../job.c:855
> #5  0x00005cdefb223852 in job_completed_txn_abort_locked (job=0x5cdefeb53000) at ../job.c:958
> #6  0x00005cdefb223714 in job_completed_locked (job=0x5cdefeb53000) at ../job.c:1065
> #7  0x00005cdefb224a8b in job_exit (opaque=0x5cdefeb53000) at ../job.c:1088
> #8  0x00005cdefb4134fc in aio_bh_call (bh=0x5cdefe7487c0) at ../util/async.c:171
> #9  0x00005cdefb4136ce in aio_bh_poll (ctx=0x5cdefd9cd750) at ../util/async.c:218
> #10 0x00005cdefb3efdfd in aio_poll (ctx=0x5cdefd9cd750, blocking=true) at ../util/aio-posix.c:722
> #11 0x00005cdefb20435e in bdrv_poll_co (s=0x7ffe491621d8) at ../block/block-gen.h:43
> #12 0x00005cdefb206a33 in bdrv_flush_all () at block/block-gen.c:1410
> #13 0x00005cdefae5c8ed in do_vm_stop (state=RUN_STATE_SHUTDOWN, send_stop=false)
>     at ../system/cpus.c:297
> #14 0x00005cdefae5c850 in vm_shutdown () at ../system/cpus.c:308
> #15 0x00005cdefae6d892 in qemu_cleanup (status=0) at ../system/runstate.c:871
> #16 0x00005cdefb1a7e78 in qemu_default_main () at ../system/main.c:38
> #17 0x00005cdefb1a7eb8 in main (argc=34, argv=0x7ffe491623a8) at ../system/main.c:48

Looking at the code in mirror_exit_common(), it doesn't seem to acquire
a write lock:

>     bdrv_graph_rdunlock_main_loop();
> 
>     /*
>      * Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
>      * inserting target_bs at s->to_replace, where we might not be able to get
>      * these permissions.
>      */
>     blk_unref(s->target);
>     s->target = NULL;

The write lock is taken in blk_remove_bs() when the refcount drops to 0
and the BB is actually removed:

>     bdrv_graph_wrlock();
>     bdrv_root_unref_child(root);
>     bdrv_graph_wrunlock();

Best Regards,
Fiona

Am 04.06.2024 um 09:58 hat Fiona Ebner geschrieben:
> Am 03.06.24 um 18:21 schrieb Kevin Wolf:
> > Am 03.06.2024 um 16:17 hat Fiona Ebner geschrieben:
> >> Am 26.03.24 um 13:44 schrieb Kevin Wolf:
> >>>
> >>> The fix for bdrv_flush_all() is probably to make it bdrv_co_flush_all()
> >>> with a coroutine wrapper so that the graph lock is held for the whole
> >>> function. Then calling bdrv_co_flush() while iterating the list is safe
> >>> and doesn't allow concurrent graph modifications.
> >>
> >> The second is that iotest 255 ran into an assertion failure upon QMP 'quit':
> >>
> >>> ../block/graph-lock.c:113: bdrv_graph_wrlock: Assertion `!qemu_in_coroutine()' failed.
> >>
> >> Looking at the backtrace:
> >>
> >>> #5  0x0000762a90cc3eb2 in __GI___assert_fail
> >>>     (assertion=0x5afb07991e7d "!qemu_in_coroutine()", file=0x5afb07991e00 "../block/graph-lock.c", line=113, function=0x5afb07991f20 <__PRETTY_FUNCTION__.4> "bdrv_graph_wrlock")
> >>>     at ./assert/assert.c:101
> >>> #6  0x00005afb07585311 in bdrv_graph_wrlock () at ../block/graph-lock.c:113
> >>> #7  0x00005afb07573a36 in blk_remove_bs (blk=0x5afb0af99420) at ../block/block-backend.c:901
> >>> #8  0x00005afb075729a7 in blk_delete (blk=0x5afb0af99420) at ../block/block-backend.c:487
> >>> #9  0x00005afb07572d88 in blk_unref (blk=0x5afb0af99420) at ../block/block-backend.c:547
> >>> #10 0x00005afb07572fe8 in bdrv_next (it=0x762a852fef00) at ../block/block-backend.c:618
> >>> #11 0x00005afb0758cd65 in bdrv_co_flush_all () at ../block/io.c:2347
> >>> #12 0x00005afb0753ba37 in bdrv_co_flush_all_entry (opaque=0x7ffff12c6050) at block/block-gen.c:1391
> >>> #13 0x00005afb0773bf41 in coroutine_trampoline (i0=168365184, i1=23291)
> >>
> >> So I guess calling bdrv_next() is not safe from a coroutine, because
> >> the function doing the iteration could end up being the last thing to
> >> have a reference for the BB.
> > 
> > Does your bdrv_co_flush_all() take the graph (reader) lock? If so, this
> > is surprising, because while we hold the graph lock, no reference should
> > be able to go away - you need the writer lock for that and you won't get
> > it as long as bdrv_co_flush_all() locks the graph. So whatever had a
> > reference before the bdrv_next() loop must still have it now. Do you
> > know where it gets dropped?
> > 
> 
> AFAICT, yes, it does hold the graph reader lock. The generated code is:
> 
> > static void coroutine_fn bdrv_co_flush_all_entry(void *opaque)
> > {
> >     BdrvFlushAll *s = opaque;
> > 
> >     bdrv_graph_co_rdlock();
> >     s->ret = bdrv_co_flush_all();
> >     bdrv_graph_co_rdunlock();
> >     s->poll_state.in_progress = false;
> > 
> >     aio_wait_kick();
> > }
> 
> Apparently when the mirror job is aborted/exits, which can happen during
> the polling for bdrv_co_flush_all_entry(), a reference can go away
> without the write lock (at least my breakpoints didn't trigger) being held:
> 
> > #0  blk_unref (blk=0x5cdefe943d20) at ../block/block-backend.c:537
> > #1  0x00005cdefb26697e in mirror_exit_common (job=0x5cdefeb53000) at ../block/mirror.c:710
> > #2  0x00005cdefb263575 in mirror_abort (job=0x5cdefeb53000) at ../block/mirror.c:823
> > #3  0x00005cdefb2248a6 in job_abort (job=0x5cdefeb53000) at ../job.c:825
> > #4  0x00005cdefb2245f2 in job_finalize_single_locked (job=0x5cdefeb53000) at ../job.c:855
> > #5  0x00005cdefb223852 in job_completed_txn_abort_locked (job=0x5cdefeb53000) at ../job.c:958
> > #6  0x00005cdefb223714 in job_completed_locked (job=0x5cdefeb53000) at ../job.c:1065
> > #7  0x00005cdefb224a8b in job_exit (opaque=0x5cdefeb53000) at ../job.c:1088
> > #8  0x00005cdefb4134fc in aio_bh_call (bh=0x5cdefe7487c0) at ../util/async.c:171
> > #9  0x00005cdefb4136ce in aio_bh_poll (ctx=0x5cdefd9cd750) at ../util/async.c:218
> > #10 0x00005cdefb3efdfd in aio_poll (ctx=0x5cdefd9cd750, blocking=true) at ../util/aio-posix.c:722
> > #11 0x00005cdefb20435e in bdrv_poll_co (s=0x7ffe491621d8) at ../block/block-gen.h:43
> > #12 0x00005cdefb206a33 in bdrv_flush_all () at block/block-gen.c:1410
> > #13 0x00005cdefae5c8ed in do_vm_stop (state=RUN_STATE_SHUTDOWN, send_stop=false)
> >     at ../system/cpus.c:297
> > #14 0x00005cdefae5c850 in vm_shutdown () at ../system/cpus.c:308
> > #15 0x00005cdefae6d892 in qemu_cleanup (status=0) at ../system/runstate.c:871
> > #16 0x00005cdefb1a7e78 in qemu_default_main () at ../system/main.c:38
> > #17 0x00005cdefb1a7eb8 in main (argc=34, argv=0x7ffe491623a8) at ../system/main.c:48
> 
> Looking at the code in mirror_exit_common(), it doesn't seem to acquire
> a write lock:
> 
> >     bdrv_graph_rdunlock_main_loop();
> > 
> >     /*
> >      * Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
> >      * inserting target_bs at s->to_replace, where we might not be able to get
> >      * these permissions.
> >      */
> >     blk_unref(s->target);
> >     s->target = NULL;
> 
> The write lock is taken in blk_remove_bs() when the refcount drops to 0
> and the BB is actually removed:
> 
> >     bdrv_graph_wrlock();
> >     bdrv_root_unref_child(root);
> >     bdrv_graph_wrunlock();

Ah, so it _would_ take the writer lock (and wait for bdrv_flush_all() to
finish) if it were the last reference, but because bdrv_next() took
another reference, it can just decrease the refcount without modifying
the graph.

Does this mean that if we just remove the whole blk/bdrv_ref/unref()
code from bdrv_next(), which is what the final state should look like
anyway, it would actually work?

Of course, we then need to audit the callers of bdrv_next() to make sure
that all of them really keep the graph lock and don't poll during the
whole iteration like they are supposed to (i.e. nobody else violates the
rules the same way bdrv_flush_all() does).

Kevin

Am 04.06.24 um 17:28 schrieb Kevin Wolf:
> Am 04.06.2024 um 09:58 hat Fiona Ebner geschrieben:
>> Am 03.06.24 um 18:21 schrieb Kevin Wolf:
>>> Am 03.06.2024 um 16:17 hat Fiona Ebner geschrieben:
>>>> Am 26.03.24 um 13:44 schrieb Kevin Wolf:
>>>>>
>>>>> The fix for bdrv_flush_all() is probably to make it bdrv_co_flush_all()
>>>>> with a coroutine wrapper so that the graph lock is held for the whole
>>>>> function. Then calling bdrv_co_flush() while iterating the list is safe
>>>>> and doesn't allow concurrent graph modifications.
>>>>
>>>> The second is that iotest 255 ran into an assertion failure upon QMP 'quit':
>>>>
>>>>> ../block/graph-lock.c:113: bdrv_graph_wrlock: Assertion `!qemu_in_coroutine()' failed.
>>>>
>>>> Looking at the backtrace:
>>>>
>>>>> #5  0x0000762a90cc3eb2 in __GI___assert_fail
>>>>>     (assertion=0x5afb07991e7d "!qemu_in_coroutine()", file=0x5afb07991e00 "../block/graph-lock.c", line=113, function=0x5afb07991f20 <__PRETTY_FUNCTION__.4> "bdrv_graph_wrlock")
>>>>>     at ./assert/assert.c:101
>>>>> #6  0x00005afb07585311 in bdrv_graph_wrlock () at ../block/graph-lock.c:113
>>>>> #7  0x00005afb07573a36 in blk_remove_bs (blk=0x5afb0af99420) at ../block/block-backend.c:901
>>>>> #8  0x00005afb075729a7 in blk_delete (blk=0x5afb0af99420) at ../block/block-backend.c:487
>>>>> #9  0x00005afb07572d88 in blk_unref (blk=0x5afb0af99420) at ../block/block-backend.c:547
>>>>> #10 0x00005afb07572fe8 in bdrv_next (it=0x762a852fef00) at ../block/block-backend.c:618
>>>>> #11 0x00005afb0758cd65 in bdrv_co_flush_all () at ../block/io.c:2347
>>>>> #12 0x00005afb0753ba37 in bdrv_co_flush_all_entry (opaque=0x7ffff12c6050) at block/block-gen.c:1391
>>>>> #13 0x00005afb0773bf41 in coroutine_trampoline (i0=168365184, i1=23291)
>>>>
>>>> So I guess calling bdrv_next() is not safe from a coroutine, because
>>>> the function doing the iteration could end up being the last thing to
>>>> have a reference for the BB.
>>>
>>> Does your bdrv_co_flush_all() take the graph (reader) lock? If so, this
>>> is surprising, because while we hold the graph lock, no reference should
>>> be able to go away - you need the writer lock for that and you won't get
>>> it as long as bdrv_co_flush_all() locks the graph. So whatever had a
>>> reference before the bdrv_next() loop must still have it now. Do you
>>> know where it gets dropped?
>>>
>>
>> AFAICT, yes, it does hold the graph reader lock. The generated code is:
>>
>>> static void coroutine_fn bdrv_co_flush_all_entry(void *opaque)
>>> {
>>>     BdrvFlushAll *s = opaque;
>>>
>>>     bdrv_graph_co_rdlock();
>>>     s->ret = bdrv_co_flush_all();
>>>     bdrv_graph_co_rdunlock();
>>>     s->poll_state.in_progress = false;
>>>
>>>     aio_wait_kick();
>>> }
>>
>> Apparently when the mirror job is aborted/exits, which can happen during
>> the polling for bdrv_co_flush_all_entry(), a reference can go away
>> without the write lock (at least my breakpoints didn't trigger) being held:
>>
>>> #0  blk_unref (blk=0x5cdefe943d20) at ../block/block-backend.c:537
>>> #1  0x00005cdefb26697e in mirror_exit_common (job=0x5cdefeb53000) at ../block/mirror.c:710
>>> #2  0x00005cdefb263575 in mirror_abort (job=0x5cdefeb53000) at ../block/mirror.c:823
>>> #3  0x00005cdefb2248a6 in job_abort (job=0x5cdefeb53000) at ../job.c:825
>>> #4  0x00005cdefb2245f2 in job_finalize_single_locked (job=0x5cdefeb53000) at ../job.c:855
>>> #5  0x00005cdefb223852 in job_completed_txn_abort_locked (job=0x5cdefeb53000) at ../job.c:958
>>> #6  0x00005cdefb223714 in job_completed_locked (job=0x5cdefeb53000) at ../job.c:1065
>>> #7  0x00005cdefb224a8b in job_exit (opaque=0x5cdefeb53000) at ../job.c:1088
>>> #8  0x00005cdefb4134fc in aio_bh_call (bh=0x5cdefe7487c0) at ../util/async.c:171
>>> #9  0x00005cdefb4136ce in aio_bh_poll (ctx=0x5cdefd9cd750) at ../util/async.c:218
>>> #10 0x00005cdefb3efdfd in aio_poll (ctx=0x5cdefd9cd750, blocking=true) at ../util/aio-posix.c:722
>>> #11 0x00005cdefb20435e in bdrv_poll_co (s=0x7ffe491621d8) at ../block/block-gen.h:43
>>> #12 0x00005cdefb206a33 in bdrv_flush_all () at block/block-gen.c:1410
>>> #13 0x00005cdefae5c8ed in do_vm_stop (state=RUN_STATE_SHUTDOWN, send_stop=false)
>>>     at ../system/cpus.c:297
>>> #14 0x00005cdefae5c850 in vm_shutdown () at ../system/cpus.c:308
>>> #15 0x00005cdefae6d892 in qemu_cleanup (status=0) at ../system/runstate.c:871
>>> #16 0x00005cdefb1a7e78 in qemu_default_main () at ../system/main.c:38
>>> #17 0x00005cdefb1a7eb8 in main (argc=34, argv=0x7ffe491623a8) at ../system/main.c:48
>>
>> Looking at the code in mirror_exit_common(), it doesn't seem to acquire
>> a write lock:
>>
>>>     bdrv_graph_rdunlock_main_loop();
>>>
>>>     /*
>>>      * Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
>>>      * inserting target_bs at s->to_replace, where we might not be able to get
>>>      * these permissions.
>>>      */
>>>     blk_unref(s->target);
>>>     s->target = NULL;
>>
>> The write lock is taken in blk_remove_bs() when the refcount drops to 0
>> and the BB is actually removed:
>>
>>>     bdrv_graph_wrlock();
>>>     bdrv_root_unref_child(root);
>>>     bdrv_graph_wrunlock();
> 
> Ah, so it _would_ take the writer lock (and wait for bdrv_flush_all() to
> finish) if it were the last reference, but because bdrv_next() took
> another reference, it can just decrease the refcount without modifying
> the graph.
> 
> Does this mean that if we just remove the whole blk/bdrv_ref/unref()
> code from bdrv_next(), which is what the final state should look like
> anyway, it would actually work?

I tried this and removed those calls. Unfortunately, I wasn't able to
run into the case where blk_unref() via job_exit() waits in
bdrv_graph_wrlock(), because there is a blk_drain() call before
blk_delete() and in my testing, it always waited there, in
bdrv_do_drained_begin()'s

>         BDRV_POLL_WHILE(bs, bdrv_drain_poll_top_level(bs, parent));

> Of course, we then need to audit the callers of bdrv_next() to make sure
> that all of them really keep the graph lock and don't poll during the
> whole iteration like they are supposed to (i.e. nobody else violates the
> rules the same way bdrv_flush_all() does).

Note that below, none of the functions I list at the top level (i.e.
after a number) are coroutines.

1. hmp_info_snapshots() in block/monitor/block-hmp-cmds.c
2. vm_completion() in migration/migration-hmp-cmds.c
both call bdrv_can_snapshot() which calls the coroutine wrapper
bdrv_is_inserted() and thus polls

3. bdrv_flush_all() in block/io.c:
calls the coroutine wrapper bdrv_flush() and thus polls, the one that
led us here ;)

4. bdrv_all_get_snapshot_devices() in block/snapshot.c:
marked as GRAPH_RDLOCK, only calls g_list_append() in the loop body, so
is fine :)

5. bdrv_activate_all() in block.c:
calls bdrv_activate() which can (e.g. after outgoing migration) call the
(manual) coroutine wrappers bdrv_invalidate_cache() and
bdrv_refresh_total_sectors() and thus poll

6. bdrv_inactivate_all() in block.c:
calls bdrv_inactivate_recurse() which for qcow2, calls
qcow2_inactivate() which calls qcow2_cache_flush() which calls the
coroutine wrapper bdrv_flush() and thus polls

My list of code paths leading to polling here is probably not even
complete, but it's already clear that there is some work left before no
callers of bdrv_next() violate the rules.

Best Regards,
Fiona