[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-block] [PATCH 12/22] qcow2-bitmap: add IN_USE flag

From: Max Reitz
Subject: Re: [Qemu-block] [PATCH 12/22] qcow2-bitmap: add IN_USE flag
Date: Wed, 26 Oct 2016 17:28:31 +0200
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:45.0) Gecko/20100101 Thunderbird/45.4.0

On 26.10.2016 11:04, Vladimir Sementsov-Ogievskiy wrote:
> 25.10.2016 13:53, Vladimir Sementsov-Ogievskiy wrote:
>> 24.10.2016 20:18, Max Reitz wrote:
>>> On 24.10.2016 19:08, Max Reitz wrote:
>>>> On 24.10.2016 13:35, Vladimir Sementsov-Ogievskiy wrote:
>>>>> 24.10.2016 13:32, Vladimir Sementsov-Ogievskiy пишет:
>>>>>> 21.10.2016 22:58, Max Reitz пишет:
>>>>>>> On 21.10.2016 17:34, Vladimir Sementsov-Ogievskiy wrote:
>>>>>>>> 07.10.2016 22:44, Max Reitz пишет:
>>>>>>>>> On 30.09.2016 12:53, Vladimir Sementsov-Ogievskiy wrote:
>>>>>>>>>> This flag means that the bitmap is now in use by the software or
>>>>>>>>>> was not
>>>>>>>>>> successfully saved. In any way, with this flag set the bitmap
>>>>>>>>>> data
>>>>>>>>>> must
>>>>>>>>>> be considered inconsistent and should not be loaded.
>>>>>>>>>> With current implementation this flag is never set, as we just
>>>>>>>>>> remove
>>>>>>>>>> bitmaps from the image after loading. But it defined in qcow2
>>>>>>>>>> spec
>>>>>>>>>> and
>>>>>>>>>> must be handled. Also, it can be used in future, if async
>>>>>>>>>> schemes of
>>>>>>>>>> bitmap loading/saving are implemented.
>>>>>>>>>> We also remove in-use bitmaps from the image on open.
>>>>>>>>>> Signed-off-by: Vladimir Sementsov-Ogievskiy
>>>>>>>>>> <address@hidden>
>>>>>>>>>> ---
>>>>>>>>>>     block/qcow2-bitmap.c | 17 ++++++++++++++++-
>>>>>>>>>>     1 file changed, 16 insertions(+), 1 deletion(-)
>>>>>>>>> Don't you want to make use of this flag? It would appear useful to
>>>>>>>>> me if
>>>>>>>>> you just marked autoload bitmaps as in_use instead of deleting
>>>>>>>>> them
>>>>>>>>> from
>>>>>>>>> the image when it's opened and then overwrite them when the
>>>>>>>>> image is
>>>>>>>>> closed.
>>>>>>>> And what is the use of it?
>>>>>>> You don't need to free any bitmaps when opening the file, and you
>>>>>>> don't
>>>>>>> need to allocate any new bitmap space when closing it.
>>>>>> As bitmaps are sparce in file, I need to allocate new space when
>>>>>> closing. Or free it...
>>>>> Is it a real problem to reallocate space in qcow2? If so, to minimaze
>>>>> allocations, I will have to make a list of clusters of in_use
>>>>> bitmaps on
>>>>> close, and then save current bitmaps to these clusters (and allocated
>>>>> some additional clusters, or free extra clusters).
>>>> It's not a real problem, but it does take time, and I maintain that
>>>> it's
>>>> time it doesn't need to take because you can just use the in_use flag.
>>>> I wouldn't worry about reusing clusters of other bitmaps. Of course we
>>>> can do that later on in some optimization but not now.
>>>> I just mean the basic case of some auto-loaded bitmap that is not being
>>>> deleted while the VM is running and is just saved back to the image
>>>> file
>>>> once it is closed. I don't expect that users will always consume all of
>>>> the auto-loaded bitmaps while the VM is active...
>>>>> Also, if I don't free them on open, I'll have to free them on
>>>>> remove of
>>>>> bdrv dirty bitmap..
>>>> Yes, so? That takes time, but that is something the user will probably
>>>> expect. I wouldn't expect opening of the file to take that time.
>>>> Overall, it doesn't matter time-wise whether you free the bitmap data
>>>> when opening the image file or when the dirty bitmap is actually
>>>> deleted
>>>> by the user or when the image file is closed. Just setting the single
>>>> in_use flag for all of the auto-loaded bitmaps will basically not take
>>>> any time.
>>>> On the other hand, as soon as just a single auto-loaded bitmap survives
>>>> a VM (or qemu-img) invocation, you will very, very likely safe at least
>>>> some time because writing the bitmap to the disk can reuse at least
>>>> some
>>>> of the existing clusters.
>>> By the way, dealing with removal of bitmaps when they are deleted by the
>>> user is also positive when it comes to migration or read-only disks that
>>> are later reopened R/W: Currently, as far as I can see, you just keep
>>> the auto-load bitmaps in the image if the image is part of an incoming
>>> migration or opened read-only, but then you continue under the
>>> assumption that they are removed from the image. That's not very good.
>> You are right, I need to handle reopening more carefully. In my way, I
>> need to delete bitmaps when reopening R/W and in yours - set in_use bit.
>>> If you delete the bitmaps only when the user asks you to, then you can
>>> just return an error that the bitmap cannot be removed at this time
>>> because the image file is read-only or used in migration (and I don't
>>> think the latter can even happen when it's the user who has requested
>>> removal of the bitmap).
>>> Max
>> Ok, I'll rewrite this way, using 'in_use' bit and trying to avoid
>> allocate/free overhead.
> Trying to reuse clusters of in_use bitmaps (including contiguous
> allocations for bitmap tables) will complicate things a lot.. We can use
> extra clusters of one in_use bitmap to save another one, the same is for
> bitmap tables. Extra clusters of old bitmap table (in case of resize)
> can be used for saving other bitmap data, etc..

I feel like you're trying to optimize too much. When rewriting the
bitmap data, reuse only the clusters you have to write to anyway.

Say the bitmap table looks like this when you open the image:

[sparse 0, cluster 42, cluster 45, sparse 0, cluster 46]

("sparse 0" means no cluster is allocated and the bitmap table entry is
0, thus creating a sparse bitmap where that part is supposed to be read
as 0.)

Now during runtime, you set some bits, maybe you even clear the whole
bitmap at some point because you're doing a backup, and then you set
some other bits, so when you close the image, the bitmap table would
have to look like this:

[sparse 0, sparse 0, data, data, data]

Now, when you want to save that bitmap data, you just walk through the
existing bitmap table. The first entry is a sparse 0, and the bitmap is
sparse there, too, so that can stay as it is.

The second entry points to data but the bitmap is now sparse. You can
then free the cluster (cluster 42) and write "sparse 0" into the bitmap
table entry.

The third entry points to data and the bitmap contains data there, so
you just write the data to cluster 45.

The fourth entry is sparse but your bitmap contains data. So now you
need to allocate a cluster, maybe that will cluster 67, and write your
data there, and make the bitmap table entry point there.

The fifth entry finally is handled just like the third entry.

So afterwards, your bitmap table would look like this:

[sparse 0, sparse 0, cluster 45, cluster 67, cluster 46]

It sounds as if you're trying to reuse all the clusters, i.e. you'd
ideally get:

[sparse 0, sparse 0, cluster 45, cluster 42, cluster 46]

May be possible, but I really wouldn't worry about that. I'd consider it
much too difficult. Just reuse the existing bitmap table.


Attachment: signature.asc
Description: OpenPGP digital signature

reply via email to

[Prev in Thread] Current Thread [Next in Thread]