On 21.08.20 17:50, Kevin Wolf wrote:
> Am 25.06.2020 um 17:22 hat Max Reitz geschrieben:
>> We have to perform an active commit whenever the top node has a parent
>> that has taken the WRITE permission on it.
>>
>> Signed-off-by: Max Reitz <mreitz@redhat.com>
>> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
>> ---
>>  blockdev.c | 24 +++++++++++++++++++++---
>>  1 file changed, 21 insertions(+), 3 deletions(-)
>>
>> diff --git a/blockdev.c b/blockdev.c
>> index 402f1d1df1..237fffbe53 100644
>> --- a/blockdev.c
>> +++ b/blockdev.c
>> @@ -2589,6 +2589,7 @@ void qmp_block_commit(bool has_job_id, const char *job_id, const char *device,
>>      AioContext *aio_context;
>>      Error *local_err = NULL;
>>      int job_flags = JOB_DEFAULT;
>> +    uint64_t top_perm, top_shared;
>>  
>>      if (!has_speed) {
>>          speed = 0;
>> @@ -2704,14 +2705,31 @@ void qmp_block_commit(bool has_job_id, const char *job_id, const char *device,
>>          goto out;
>>      }
>>  
>> -    if (top_bs == bs) {
>> +    /*
>> +     * Active commit is required if and only if someone has taken a
>> +     * WRITE permission on the top node.
> 
> ...or if someone wants to take a WRITE permission while the job is
> running.
> 
> Future intentions of the user is something that we can't know, so maybe
> this should become an option in the future (not in this series, of
> course).
> 
>>                                            Historically, we have always
>> +     * used active commit for top nodes, so continue that practice.
>> +     * (Active commit is never really wrong.)
>> +     */
> 
> Changing the practice would break compatibility with clients that start
> an active commit job and then attach it to a read-write device, so we
> must continue the practice. I think the comment should be clearer about
> this, it sounds more like "no reason, but why not".

I think that’s what I meant by “historically”.  Is “legacily” a word?

But sure, I can make it more explicit.

> This is even more problematic because the commit job doesn't unshare
> BLK_PERM_WRITE yet, so it would lead to silent corruption rather than an
> error.
> 
>> +    bdrv_get_cumulative_perm(top_bs, &top_perm, &top_shared);
>> +    if (top_perm & BLK_PERM_WRITE ||
>> +        bdrv_skip_filters(top_bs) == bdrv_skip_filters(bs))
>> +    {
>>          if (has_backing_file) {
>>              error_setg(errp, "'backing-file' specified,"
>>                               " but 'top' is the active layer");
> 
> Hm, this error message isn't accurate any more.
> 
> In fact, the implementation isn't consistent with the QAPI documentation
> any more, because backing-file is only an error for the top level.

Hm.  I wanted to agree, and then I wanted to come up with a QAPI
documentation that fits the new behavior (because I think it makes more
sense to change the QAPI documentation along with the behavior change,
rather than to force us to allow backing-file for anything that isn’t on
the top layer).

But in the process of coming up with a better description, I noticed
that this doesn’t say “is a root node”, it says “is the active layer”.
I would say a node in the active layer is a node that has some parent
that has taken a WRITE permission on it.  So actually I think that the
documentation is right, and this code only now fits.

Though I do think this wants for some clarification.  Perhaps “If 'top'
is the active layer (i.e., is a node that may be written to), specifying
a backing [...]”?

There’s more wrong with the specification, namely the whole part under
@backing-file past the “(Since 2.1)”, starting with “If top == base”.  I
think all of that should go to the top level.  (And “If top == active”
should be changed to “If top is active (i.e., may be written to)”.)

Max