On 05.09.19 18:24, Kevin Wolf wrote:
> Am 12.08.2019 um 14:58 hat Max Reitz geschrieben:
>> On 10.08.19 17:36, Vladimir Sementsov-Ogievskiy wrote:
>>> 09.08.2019 19:13, Max Reitz wrote:
>>>> If the driver does not support .bdrv_co_flush() so bdrv_co_flush()
>>>> itself has to flush the children of the given node, it should not flush
>>>> just bs->file->bs, but in fact all children.
>>>>
>>>> In any case, the BLKDBG_EVENT() should be emitted on the primary child,
>>>> because that is where a blkdebug node would be if there is any.
>>>>
>>>> Suggested-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
>>>> Signed-off-by: Max Reitz <mreitz@redhat.com>
>>>> ---
>>>>   block/io.c | 23 +++++++++++++++++------
>>>>   1 file changed, 17 insertions(+), 6 deletions(-)
>>>>
>>>> diff --git a/block/io.c b/block/io.c
>>>> index c5a8e3e6a3..bcc770d336 100644
>>>> --- a/block/io.c
>>>> +++ b/block/io.c
>>>> @@ -2572,6 +2572,8 @@ static void coroutine_fn bdrv_flush_co_entry(void *opaque)
>>>>   
>>>>   int coroutine_fn bdrv_co_flush(BlockDriverState *bs)
>>>>   {
>>>> +    BdrvChild *primary_child = bdrv_primary_child(bs);
>>>> +    BdrvChild *child;
>>>>       int current_gen;
>>>>       int ret = 0;
>>>>   
>>>> @@ -2601,7 +2603,7 @@ int coroutine_fn bdrv_co_flush(BlockDriverState *bs)
>>>>       }
>>>>   
>>>>       /* Write back cached data to the OS even with cache=unsafe */
>>>> -    BLKDBG_EVENT(bs->file, BLKDBG_FLUSH_TO_OS);
>>>> +    BLKDBG_EVENT(primary_child, BLKDBG_FLUSH_TO_OS);
>>>>       if (bs->drv->bdrv_co_flush_to_os) {
>>>>           ret = bs->drv->bdrv_co_flush_to_os(bs);
>>>>           if (ret < 0) {
>>>> @@ -2611,15 +2613,15 @@ int coroutine_fn bdrv_co_flush(BlockDriverState *bs)
>>>>   
>>>>       /* But don't actually force it to the disk with cache=unsafe */
>>>>       if (bs->open_flags & BDRV_O_NO_FLUSH) {
>>>> -        goto flush_parent;
>>>> +        goto flush_children;
>>>>       }
>>>>   
>>>>       /* Check if we really need to flush anything */
>>>>       if (bs->flushed_gen == current_gen) {
>>>> -        goto flush_parent;
>>>> +        goto flush_children;
>>>>       }
>>>>   
>>>> -    BLKDBG_EVENT(bs->file, BLKDBG_FLUSH_TO_DISK);
>>>> +    BLKDBG_EVENT(primary_child, BLKDBG_FLUSH_TO_DISK);
>>>>       if (!bs->drv) {
>>>>           /* bs->drv->bdrv_co_flush() might have ejected the BDS
>>>>            * (even in case of apparent success) */
>>>> @@ -2663,8 +2665,17 @@ int coroutine_fn bdrv_co_flush(BlockDriverState *bs)
>>>>       /* Now flush the underlying protocol.  It will also have BDRV_O_NO_FLUSH
>>>>        * in the case of cache=unsafe, so there are no useless flushes.
>>>>        */
>>>> -flush_parent:
>>>> -    ret = bs->file ? bdrv_co_flush(bs->file->bs) : 0;
>>>> +flush_children:
>>>> +    ret = 0; > +    QLIST_FOREACH(child, &bs->children, next) {
>>>> +        int this_child_ret;
>>>> +
>>>> +        this_child_ret = bdrv_co_flush(child->bs);
>>>> +        if (!ret) {
>>>> +            ret = this_child_ret;
>>>> +        }
>>>> +    }
>>>
>>> Hmm, you said that we want to flush only children with write-access from parent..
>>
>> Good that you remember it, I must have overlooked it (when reading the
>> replies to the previous version). :-)
>>
>>> Shouldn't we check it? Or we assume that it's always safe to call bdrv_co_flush on
>>> a node?
>>
>> I think it’s always safe.  But checking it seems like a nice touch, yes.
> 
> I'm not sure why we would unconditionally flush all children anyway. The
> only drivers I can think of that really need to flush more than one
> child are blkverify and quorum, and both of them already implement this.
> blkverify implements .bdrv_co_flush, so it's not affected by the change
> anyway, but quorum children will be flushed twice now.
> 
> But more than this, I'm worried about the overhead of needlessly
> recursing through the whole backing chain and calling flush on every
> node there.  Maybe bs->write_gen saves us so that at least this doesn't
> result in an fdatasync() call for each, but still... Without a use case,
> I'd rather not do this.
> 
> Oh, well, after having written all of this, I see that qcow2 with an
> external data file is buggy... This could be fixed in the qcow2 driver,
> but maybe restricting the recursion to read-only is actually good enough
> then. Can you mention this case in the commit message and maybe build a
> test for it?

And I should thus probably drop vmdk’s .bdrv_co_flush_to_disk()
implementation.

I will indeed try to write a test, but to be completely honest, I feel
like this series is long enough.

Max