Re: [Qemu-devel] [PATCH] blockdev: avoid acquiring AioContext lock twice at do_drive_backup()

[Max Reitz]

On 13.09.19 12:15, Sergio Lopez wrote:
> 
> Max Reitz <address@hidden> writes:
> 
>> On 13.09.19 11:37, Sergio Lopez wrote:
>>>
>>> Max Reitz <address@hidden> writes:
>>>
>>>> On 12.09.19 18:16, Sergio Lopez wrote:
>>>>> do_drive_backup() acquires the AioContext lock of the corresponding
>>>>> BlockDriverState. This is not a problem when it's called from
>>>>> qmp_drive_backup(), but drive_backup_prepare() also acquires the lock
>>>>> before calling it.
>>>>>
>>>>> This change adds a BlockDriverState argument to do_drive_backup(),
>>>>> which is used to signal that the context lock is already acquired and
>>>>> to save a couple of redundant calls.
>>>>
>>>> But those redundant calls don’t really hurt (it’s just bdrv_lookup_bs(),
>>>> as far as I can tell).  Wouldn’t it be simpler to just release the
>>>> context lock in drive_backup_prepare() before calling do_drive_backup()?
>>>>  The BDS is drained anyway.
>>>
>>> Redundant calls rarely hurt, they're just redundant ;-)
>>
>> If they’re expensive and in a hot path, they hurt.
>>
>>>> On top of that, do_backup_common() calls bdrv_try_set_aio_context() to
>>>> bring the target into the source’s AioContext.  However, this function
>>>> must be called with the old AioContext held, and the new context not held.
>>>
>>> Is this documented somewhere? I see nothing in the function declaration
>>> nor definition.
>>>
>>> I'm starting to get the feeling that the block layer is riddled with
>>> unwritten rules and assumptions that makes every change a lot harder
>>> than it should be.
>>
>> It is written, it’s just that it’s written in
>> bdrv_set_aio_context_ignore()’s definition.
>>
>> Yes, we should document it directly for bdrv_try_set_aio_context(), too,
>> because that’s what external callers are much more likely to use.
>>
>>>> Currently, it’s called exactly the other way around: With the new
>>>> context held, but the old one not held.
>>>>
>>>> So I think it indeed actually makes more sense to release the AioContext
>>>> before calling do_drive_backup(), and to move the
>>>> bdrv_try_set_aio_context() call for target_bs to the callers of
>>>> do_backup_common() (where they have not yet taken the AioContext lock).
>>>
>>> OK. I see this also happens in external_snapshot_prepare() and
>>> qmp_drive_mirror() too. I guess we should fix these too.
>>>
>>> In qmp_drive_mirror(), would it be safe to delay the acquisition of any
>>> context until just before the blockdev_mirror_common()?
>>
>> From mirror’s perspective I think so, but I don’t think it’s safe to
>> access any of a BDS’s fields without having acquired its AioContext.
>> (In fact, I wonder whether we should acquire the context even before
>> bdrv_op_is_blocked()...)
> 
> In that case, I wonder if we can safely release the context to honor
> bdrv_try_set_aio_context() requirements, knowing we aren't in a drained
> section.

Hm.  I suppose it depends.  From the main context, it is OK to access
all fields that only the main context accesses.  So
bdrv_try_set_aio_context() should be safe because they do that.

I suppose the same goes for bdrv_op_is_blocked(), because op blockers
are only set up or removed in the main context.  So it should be safe as
it is after all.

Max

signature.asc
Description: OpenPGP digital signature