Re: [PATCH v2 3/4] backup: Make sure that source and target size match

[Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>]

06.05.2020 11:02, Kevin Wolf wrote:
> Am 06.05.2020 um 08:07 hat Vladimir Sementsov-Ogievskiy geschrieben:
>> 05.05.2020 13:03, Kevin Wolf wrote:
>>> Am 30.04.2020 um 20:21 hat Vladimir Sementsov-Ogievskiy geschrieben:
>>>> 30.04.2020 17:27, Kevin Wolf wrote:
>>>>> Since the introduction of a backup filter node in commit 00e30f05d, the
>>>>> backup block job crashes when the target image is smaller than the
>>>>> source image because it will try to write after the end of the target
>>>>> node without having BLK_PERM_RESIZE. (Previously, the BlockBackend layer
>>>>> would have caught this and errored out gracefully.)
>>>>>
>>>>> We can fix this and even do better than the old behaviour: Check that
>>>>> source and target have the same image size at the start of the block job
>>>>> and unshare BLK_PERM_RESIZE. (This permission was already unshared
>>>>> before the same commit 00e30f05d, but the BlockBackend that was used to
>>>>> make the restriction was removed without a replacement.) This will
>>>>> immediately error out when starting the job instead of only when writing
>>>>> to a block that doesn't exist in the target.
>>>>>
>>>>> Longer target than source would technically work because we would never
>>>>> write to blocks that don't exist, but semantically these are invalid,
>>>>> too, because a backup is supposed to create a copy, not just an image
>>>>> that starts with a copy.
>>>>>
>>>>> Fixes: 00e30f05de1d19586345ec373970ef4c192c6270
>>>>> Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1778593
>>>>> Cc: qemu-stable@nongnu.org
>>>>> Signed-off-by: Kevin Wolf <kwolf@redhat.com>
>>>>
>>>> I'm OK with it as is, as it fixes bug:
>>>>
>>>> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
>>>>
>>>> still, some notes below
>>>>
>>>>
>>>>> ---
>>>>>     block/backup-top.c | 14 +++++++++-----
>>>>>     block/backup.c     | 14 +++++++++++++-
>>>>>     2 files changed, 22 insertions(+), 6 deletions(-)
>>>>>
>>>>> diff --git a/block/backup-top.c b/block/backup-top.c
>>>>> index 3b50c06e2c..79b268e6dc 100644
>>>>> --- a/block/backup-top.c
>>>>> +++ b/block/backup-top.c
>>>>> @@ -148,8 +148,10 @@ static void backup_top_child_perm(BlockDriverState *bs, BdrvChild *c,
>>>>>              *
>>>>>              * Share write to target (child_file), to not interfere
>>>>>              * with guest writes to its disk which may be in target backing chain.
>>>>> +         * Can't resize during a backup block job because we check the size
>>>>> +         * only upfront.
>>>>>              */
>>>>> -        *nshared = BLK_PERM_ALL;
>>>>> +        *nshared = BLK_PERM_ALL & ~BLK_PERM_RESIZE;
>>>>>             *nperm = BLK_PERM_WRITE;
>>>>>         } else {
>>>>>             /* Source child */
>>>>> @@ -159,7 +161,7 @@ static void backup_top_child_perm(BlockDriverState *bs, BdrvChild *c,
>>>>>             if (perm & BLK_PERM_WRITE) {
>>>>>                 *nperm = *nperm | BLK_PERM_CONSISTENT_READ;
>>>>>             }
>>>>> -        *nshared &= ~BLK_PERM_WRITE;
>>>>> +        *nshared &= ~(BLK_PERM_WRITE | BLK_PERM_RESIZE);
>>>>>         }
>>>>>     }
>>>>> @@ -192,11 +194,13 @@ BlockDriverState *bdrv_backup_top_append(BlockDriverState *source,
>>>>>     {
>>>>>         Error *local_err = NULL;
>>>>>         BDRVBackupTopState *state;
>>>>> -    BlockDriverState *top = bdrv_new_open_driver(&bdrv_backup_top_filter,
>>>>> -                                                 filter_node_name,
>>>>> -                                                 BDRV_O_RDWR, errp);
>>>>> +    BlockDriverState *top;
>>>>>         bool appended = false;
>>>>> +    assert(source->total_sectors == target->total_sectors);
>>>>
>>>> May be better to error-out, just to keep backup-top independent. Still, now it's not
>>>> really needed, as we have only one caller. And this function have to be refactored
>>>> anyway, when publishing this filter (open() and close() should appear, so this code
>>>> will be rewritten anyway.)
>>>
>>> Yes, the whole function only works because it's used in this restricted
>>> context today. For example, we only know that total_sectors is up to
>>> date because the caller has called bdrv_getlength() just a moment ago.
>>>
>>> I think fixing this would be beyond the scope of this patch, but
>>> certainly a good idea anyway.
>>>
>>>> And the other thought: the permissions we declared above, will be activated only after
>>>> successful bdrv_child_refresh_perms(). I think some kind of race is possible, so that
>>>> size is changed actual permission activation. So, may be good to double check sizes after
>>>> bdrv_child_refresh_perms().. But it's a kind of paranoia.
>>>
>>> We're not in coroutine context, so we can't yield. I don't see who could
>>> change the size in parallel (apart from an external process, but an
>>> external process can mess up anything).
>>>
>>> When we make backup-top an independent driver, instead of double
>>> checking (what would you do on error?), maybe we could move the size
>>> initialisation (then with bdrv_getlength()) to after
>>> bdrv_child_refresh_perms().
>>>
>>>> Also, third thought: the restricted permissions doesn't save us from resizing
>>>> of the source through exactly this node, does it? Hmm, but your test works somehow. But
>>>> (I assume) it worked in a previous patch version without unsharing on source..
>>>>
>>>> Ha, but bdrv_co_truncate just can't work on backup-top, because it doesn't have file child.
>>>> But, if we fix bdrv_co_truncate to skip filters, we'll need to define .bdrv_co_truncate in
>>>> backup_top, which will return something like -EBUSY.. Or just -ENOTSUP, doesn't matter.
>>>
>>> Maybe this is a sign that bdrv_co_truncate shouldn't automatically skip
>>> filters because filters might depend on a fixed size?
>>>
>>> Or we could make the automatic skipping depend on having BLK_PERM_RESIZE
>>> for the child. If the filter doesn't have the permission, we must not
>>> call truncate for its child (it would crash). Then backup-top and
>>> similar filters must just be careful not to take RESIZE permissions.
>>>
>>
>> Hmm this should work.. Still it's a workaround, seems out of the
>> concept of permission system..
> 
> I'm not sure about this. I see the problem more with unconditionally
> skipping the filter without checking whether the operation is even
> allowed on the filtered child.

Agree. Blindly skipping the filter is wrong anyway.

> 
>> I think, that the problem is that .bdrv_top_child_perm can't return an
>> error.  The handler answers the question:
>>
>> - Hi, we are your owners and we want the following cumulative
>> permissions on you. Then, which permissions do you want on your child?
>>
>> And the handler can't answer: "Hi, you guys want too much, I refuse to
>> play by your rules"..
> 
> It can implement .bdrv_check_perm to do that. It's just a bit more work
> for each driver to do that.
> 

OK, so, it actually can.


-- 
Best regards,
Vladimir