Re: [Qemu-devel] [PATCH v2 14/17] block: Resize bitmaps on bdrv_truncate

[John Snow <jsnow@redhat.com>]

On 03/03/2015 04:27 PM, Max Reitz wrote:
> On 2015-03-03 at 16:24, John Snow wrote:
>>
>>
>> On 03/03/2015 11:02 AM, Max Reitz wrote:
>>> On 2015-03-02 at 18:20, John Snow wrote:
>>>> Signed-off-by: John Snow <jsnow@redhat.com>
>>>> ---
>>>>   block.c                | 22 ++++++++++++++++++++
>>>>   include/block/block.h  |  1 +
>>>>   include/qemu/hbitmap.h | 10 +++++++++
>>>>   util/hbitmap.c         | 55
>>>> ++++++++++++++++++++++++++++++++++++++++++++++++++
>>>>   4 files changed, 88 insertions(+)
>>>>
>>>> diff --git a/block.c b/block.c
>>>> index e6b2696..5eaa874 100644
>>>> --- a/block.c
>>>> +++ b/block.c
>>>> @@ -3543,6 +3543,7 @@ int bdrv_truncate(BlockDriverState *bs, int64_t
>>>> offset)
>>>>       ret = drv->bdrv_truncate(bs, offset);
>>>>       if (ret == 0) {
>>>>           ret = refresh_total_sectors(bs, offset >> BDRV_SECTOR_BITS);
>>>> +        bdrv_dirty_bitmap_truncate(bs);
>>>>           if (bs->blk) {
>>>>               blk_dev_resize_cb(bs->blk);
>>>>           }
>>>> @@ -5562,6 +5563,27 @@ BdrvDirtyBitmap
>>>> *bdrv_reclaim_dirty_bitmap(BlockDriverState *bs,
>>>>       return parent;
>>>>   }
>>>> +static void dirty_bitmap_truncate(BdrvDirtyBitmap *bitmap, uint64_t
>>>> size)
>>>> +{
>>>> +    /* Should only be frozen during a block backup job, which should
>>>> have
>>>> +     * blocked any resize actions. */
>>>> +    assert(!bdrv_dirty_bitmap_frozen(bitmap));
>>>> +    hbitmap_truncate(bitmap->bitmap, size);
>>>> +}
>>>> +
>>>> +void bdrv_dirty_bitmap_truncate(BlockDriverState *bs)
>>>> +{
>>>> +    BdrvDirtyBitmap *bitmap;
>>>> +    uint64_t size = bdrv_nb_sectors(bs);
>>>> +
>>>> +    QLIST_FOREACH(bitmap, &bs->dirty_bitmaps, list) {
>>>> +        if (bdrv_dirty_bitmap_frozen(bitmap)) {
>>>> +            continue;
>>>> +        }
>>>
>>> Hm!
>>>
>>> The assert() above doesn't do anything because this condition will just
>>> skip over frozen bitmaps. Maybe we should drop it?
>>>
>>> I guess my R-b stands since bdrv_dirty_bitmap_frozen(bitmap) will always
>>> be false anyway (hence the assertion), but it would stand with this
>>> block dropped, too.
>>>
>>> Max
>>>
>>
>> It's there for parity with the other bdrv helpers that perform a
>> similar check. I am guarding against, in the future, any potential
>> users from deciding to call this function directly for whichever reason.
>>
>> Requesting an action on a drive that contains bitmaps that are not
>> ready for the action is OK. Requesting an action on a bitmap that is
>> not ready for the action should never happen, hence the 'documentation
>> assert.'
>>
>> My default action will be to leave it in.
>
> Technically that's fine (which is why I said my R-b stands), my only
> problem with that is that if something really strange happens and we
> actually end up with a frozen bitmap here, the action we're taking is
> wrong. The bitmap will have the wrong size after the BDS was truncated,
> and that's not right.
>

If something really strange happens I would rather prefer the person 
making the really strange patch to know sooner rather than later that 
they're violating some assumptions made for incremental backups.

The only way the bitmap can be frozen is currently during a backup, 
which should prevent any user-facing resize actions from the outset.

If that changes, it's important that we know about it. truncating a 
frozen bitmap is not a circumstance I want to entertain being possible.

> I'd find just calling dirty_bitmap_truncate() regardless of the frozen
> state more correct, the idea being "We want to resize all bitmaps,
> because all bitmaps attached to this BDS need to be resized", and then,
> in dirty_bitmap_truncate(), we notice that we cannot resize a frozen
> bitmap. But we need to. So the assertion fails.
>

This is a good point, though. By the above reasoning, it's better to 
force this whole command to fail exquisitely. Since we already do a soft 
user check that should suffice up at qmp_drive_resize, we should already 
be avoiding the assert() under normal operating conditions.

I'd rather not re-spin for this one item, though ... If it's fine by 
everyone else I'll tidy this up in the transaction patch that's going 
out shortly.

> Max
>