On 26.08.20 08:49, Vladimir Sementsov-Ogievskiy wrote:
> 25.08.2020 18:11, Max Reitz wrote:
>> On 21.08.20 16:11, Vladimir Sementsov-Ogievskiy wrote:
>>> It's intended to be inserted between format and protocol nodes to
>>> preallocate additional space (expanding protocol file) on writes
>>> crossing EOF. It improves performance for file-systems with slow
>>> allocation.
>>>
>>> Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
>>> ---
>>>   docs/system/qemu-block-drivers.rst.inc |  26 +++
>>>   qapi/block-core.json                   |  20 +-
>>>   block/preallocate.c                    | 291 +++++++++++++++++++++++++
>>>   block/Makefile.objs                    |   1 +
>>>   4 files changed, 337 insertions(+), 1 deletion(-)
>>>   create mode 100644 block/preallocate.c

[...]

>>> diff --git a/block/preallocate.c b/block/preallocate.c
>>> new file mode 100644
>>> index 0000000000..bdf54dbd2f
>>> --- /dev/null
>>> +++ b/block/preallocate.c
>>> @@ -0,0 +1,291 @@
>>> +/*
>>> + * preallocate filter driver
>>> + *
>>> + * The driver performs preallocate operation: it is injected above
>>> + * some node, and before each write over EOF it does additional
>>> preallocating
>>> + * write-zeroes request.
>>> + *
>>> + * Copyright (c) 2020 Virtuozzo International GmbH.
>>> + *
>>> + * Author:
>>> + *  Sementsov-Ogievskiy Vladimir <vsementsov@virtuozzo.com>
>>> + *
>>> + * This program is free software; you can redistribute it and/or modify
>>> + * it under the terms of the GNU General Public License as published by
>>> + * the Free Software Foundation; either version 2 of the License, or
>>> + * (at your option) any later version.
>>> + *
>>> + * This program is distributed in the hope that it will be useful,
>>> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
>>> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
>>> + * GNU General Public License for more details.
>>> + *
>>> + * You should have received a copy of the GNU General Public License
>>> + * along with this program. If not, see <http://www.gnu.org/licenses/>.
>>> + */
>>> +
>>> +#include "qemu/osdep.h"
>>> +
>>> +#include "qapi/error.h"
>>> +#include "qemu/module.h"
>>> +#include "qemu/option.h"
>>> +#include "qemu/units.h"
>>> +#include "block/block_int.h"
>>> +
>>> +
>>> +typedef struct BDRVPreallocateState {
>>> +    int64_t prealloc_size;
>>> +    int64_t prealloc_align;
>>> +
>>> +    /*
>>> +     * Filter is started as not-active, so it doesn't do any
>>> preallocations nor
>>> +     * requires BLK_PERM_RESIZE on its child. This is needed to
>>> create filter
>>> +     * above another node-child and then do bdrv_replace_node to
>>> insert the
>>> +     * filter.
>>
>> A bit weird, but seems fair.  It’s basically just a cache for whether
>> this node has a writer on it or not.
>>
>> Apart from the weirdness, I don’t understand the “another node-child”.
>> Say you have “format” -> “proto”, and then you want to insert
>> “prealloc”.  Wouldn’t you blockdev-add prealloc,file=proto and then
>> blockdev-replace format.file=prealloc?
> 
> Yes something like this. Actually, I'm about inserting the filter
> automatically from block layer code, but your reasoning is about same
> thing and is better.
> 
>> So what is that “other node-child”?
> 
> Hmm, just my misleading wording. At least '-' in wrong place. Just
> "other node child", or "child of another node"..

OK.

>>> +     *
>>> +     * Filter becomes active the first time it detects that its
>>> parents have
>>> +     * BLK_PERM_RESIZE on it.
>>> +     * Filter becomes active forever: it doesn't lose active status
>>> if parents
>>> +     * lose BLK_PERM_RESIZE, otherwise we'll not be able to shrink
>>> the file on
>>> +     * filter close.
>>
>> Oh, the file is shrunk?  That wasn’t clear from the documentation.
>>
>> Hm.  Seems like useful behavior.  I just wonder if keeping the
>> permission around indefinitely makes sense.  Why not just truncate it
>> when the permission is revoked?
> 
> How? Parent is closed earlier, so on close we will not have the
> permission. So, we force-keep the permission up to our close().

I thought that preallocate_child_perm() would be invoked when the parent
is detached, so we could do the truncate there, before relinquishing
preallocate’s RESIZE permission.

[...]

>>> +static void preallocate_close(BlockDriverState *bs)
>>> +{
>>> +    BDRVPreallocateState *s = bs->opaque;
>>> +
>>> +    if (s->active && s->data_end >= 0 &&
>>> +        bdrv_getlength(bs->file->bs) > s->data_end)
>>> +    {
>>> +        bdrv_truncate(bs->file, s->data_end, true,
>>> PREALLOC_MODE_OFF, 0, NULL);
>>
>> Now that I think more about it...  What if there are other writers on
>> bs->file?  This may throw data away.
> 
> Good point. Actually, if bs->file has other writing parents, the logic
> of the filter
> around "data_end" is broken. So we must unshare WRITE and RESIZE
> permissions.

That’s certainly a heavy hammer, but it’d work.

>>  Maybe BDS.wr_highest_offset can
>> help to make a better call?
> 
> Anyway, we'll have to use wr_highest_offset of the filter not the child
> node
> (in the child wr_highest_offset will track preallocations as well),

That’s true.

> so we want to unshare WRITE/RESIZE permissions.

[...]

>>> +static int coroutine_fn
>>> +preallocate_co_truncate(BlockDriverState *bs, int64_t offset,
>>> +                        bool exact, PreallocMode prealloc,
>>> +                        BdrvRequestFlags flags, Error **errp)
>>> +{
>>> +    BDRVPreallocateState *s = bs->opaque;
>>> +    int ret = bdrv_co_truncate(bs->file, offset, exact, prealloc,
>>> flags, errp);
>>> +
>>> +    /* s->data_end may become negative here, which means unknown
>>> data end */
>>> +    s->data_end = bdrv_getlength(bs->file->bs);
>>
>> What would be the problem with just setting data_end = offset?  We only
>> use data_end to truncate down on close, so basically repeat the
>> bdrv_co_truncate() call here, which seems correct.
> 
> We can use offset only if ret >= 0 && exact is true..

Well, on close, we ignore any errors from truncate() anyway.  So even if
we used exact=false here, it shouldn’t matter.

> But simpler is
> just call
> bdrv_getlength() anyway.

Certainly simpler than thinking about potential edge cases, true.

Max