[Qemu-devel] Unaligned images with O_DIRECT

[Qemu-devel] Unaligned images with O_DIRECT

[Max Reitz]

[-- Attachment #1: Type: text/plain, Size: 950 bytes --]

Hi,

Unaligned images don’t work so well with O_DIRECT:

$ echo > foo
$ qemu-img map --image-opts driver=file,filename=foo,cache.direct=on
Offset          Length          Mapped to       File
qemu-img: block/io.c:2093: bdrv_co_block_status: Assertion `*pnum &&
QEMU_IS_ALIGNED(*pnum, align) && align > offset - aligned_offset' failed.
[1]    10954 abort (core dumped)  qemu-img map --image-opts
driver=file,filename=foo,cache.direct=on

(compare https://bugzilla.redhat.com/show_bug.cgi?id=1588356)

This is because the request_alignment is 512 (in my case), but the EOF
is not aligned accordingly, so raw_co_block_status() returns an aligned
*pnum.

I suppose having an unaligned tail is not so bad and maybe we can just
adjust the assertion accordingly.  On the other hand, this has been
broken for a while.  Does it even make sense to use O_DIRECT with
unaligned images?  Shouldn’t we just reject them outright?

Max


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

Re: [Qemu-devel] Unaligned images with O_DIRECT

[Eric Blake]

[-- Attachment #1: Type: text/plain, Size: 2497 bytes --]

On 5/14/19 10:06 AM, Max Reitz wrote:
> Hi,
> 
> Unaligned images don’t work so well with O_DIRECT:
> 
> $ echo > foo
> $ qemu-img map --image-opts driver=file,filename=foo,cache.direct=on
> Offset          Length          Mapped to       File
> qemu-img: block/io.c:2093: bdrv_co_block_status: Assertion `*pnum &&
> QEMU_IS_ALIGNED(*pnum, align) && align > offset - aligned_offset' failed.
> [1]    10954 abort (core dumped)  qemu-img map --image-opts
> driver=file,filename=foo,cache.direct=on
> 
> (compare https://bugzilla.redhat.com/show_bug.cgi?id=1588356)
> 
> This is because the request_alignment is 512 (in my case), but the EOF
> is not aligned accordingly, so raw_co_block_status() returns an aligned
> *pnum.

Uggh. Yet another reason why I want qemu to support byte-accurate
sizing, instead of rounding up. The rounding keeps raising its head in
more and more places. I have pending patches that are trying to improve
block status to round driver answers up to match request_alignment (when
the protocol layer has finer granularity than the format layer); but
this sounds like it is a bug in the file driver itself for returning an
answer that is not properly rounded according to its own
request_alignment boundary, and not one where my pending patches would help.

> 
> I suppose having an unaligned tail is not so bad and maybe we can just
> adjust the assertion accordingly.  On the other hand, this has been
> broken for a while.  Does it even make sense to use O_DIRECT with
> unaligned images?  Shouldn’t we just reject them outright?

The tail of an unaligned file is generally inaccessible to O_DIRECT,
where it is easier to use ftruncate() up to an aligned boundary if you
really must play with that region of the file, and then ftruncate() back
to the intended size after I/O. But that sounds hairy.  We could also
round down and silently ignore the tail of the file, but that is at odds
with our practice of rounding size up.  So for the short term, I'd be
happy with a patch that just rejects any attempt to use cache.direct=on
(O_DIRECT) with a file that is not already a multiple of the alignment
required thereby. (For reference, that's what qemu as NBD client
recently did when talking to a server that advertises a size
inconsistent with forced minimum block access: commit 3add3ab7)

-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3226
Virtualization:  qemu.org | libvirt.org


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

Re: [Qemu-devel] Unaligned images with O_DIRECT

[Max Reitz]

[-- Attachment #1: Type: text/plain, Size: 2578 bytes --]

On 14.05.19 17:45, Eric Blake wrote:
> On 5/14/19 10:06 AM, Max Reitz wrote:
>> Hi,
>>
>> Unaligned images don’t work so well with O_DIRECT:
>>
>> $ echo > foo
>> $ qemu-img map --image-opts driver=file,filename=foo,cache.direct=on
>> Offset          Length          Mapped to       File
>> qemu-img: block/io.c:2093: bdrv_co_block_status: Assertion `*pnum &&
>> QEMU_IS_ALIGNED(*pnum, align) && align > offset - aligned_offset' failed.
>> [1]    10954 abort (core dumped)  qemu-img map --image-opts
>> driver=file,filename=foo,cache.direct=on
>>
>> (compare https://bugzilla.redhat.com/show_bug.cgi?id=1588356)
>>
>> This is because the request_alignment is 512 (in my case), but the EOF
>> is not aligned accordingly, so raw_co_block_status() returns an aligned
>> *pnum.
> 
> Uggh. Yet another reason why I want qemu to support byte-accurate
> sizing, instead of rounding up. The rounding keeps raising its head in
> more and more places. I have pending patches that are trying to improve
> block status to round driver answers up to match request_alignment (when
> the protocol layer has finer granularity than the format layer); but
> this sounds like it is a bug in the file driver itself for returning an
> answer that is not properly rounded according to its own
> request_alignment boundary, and not one where my pending patches would help.

Yes, I think so, too.

>> I suppose having an unaligned tail is not so bad and maybe we can just
>> adjust the assertion accordingly.  On the other hand, this has been
>> broken for a while.  Does it even make sense to use O_DIRECT with
>> unaligned images?  Shouldn’t we just reject them outright?
> 
> The tail of an unaligned file is generally inaccessible to O_DIRECT,

Especially with this.

> where it is easier to use ftruncate() up to an aligned boundary if you
> really must play with that region of the file, and then ftruncate() back
> to the intended size after I/O. But that sounds hairy.  We could also
> round down and silently ignore the tail of the file, but that is at odds
> with our practice of rounding size up.  So for the short term, I'd be
> happy with a patch that just rejects any attempt to use cache.direct=on
> (O_DIRECT) with a file that is not already a multiple of the alignment
> required thereby. (For reference, that's what qemu as NBD client
> recently did when talking to a server that advertises a size
> inconsistent with forced minimum block access: commit 3add3ab7)

OK, I’ll send a patch.  Thanks for you explanation!

Max


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

Re: [Qemu-devel] Unaligned images with O_DIRECT

[Max Reitz]

[-- Attachment #1: Type: text/plain, Size: 3831 bytes --]

On 14.05.19 18:15, Max Reitz wrote:
> On 14.05.19 17:45, Eric Blake wrote:
>> On 5/14/19 10:06 AM, Max Reitz wrote:
>>> Hi,
>>>
>>> Unaligned images don’t work so well with O_DIRECT:
>>>
>>> $ echo > foo
>>> $ qemu-img map --image-opts driver=file,filename=foo,cache.direct=on
>>> Offset          Length          Mapped to       File
>>> qemu-img: block/io.c:2093: bdrv_co_block_status: Assertion `*pnum &&
>>> QEMU_IS_ALIGNED(*pnum, align) && align > offset - aligned_offset' failed.
>>> [1]    10954 abort (core dumped)  qemu-img map --image-opts
>>> driver=file,filename=foo,cache.direct=on
>>>
>>> (compare https://bugzilla.redhat.com/show_bug.cgi?id=1588356)
>>>
>>> This is because the request_alignment is 512 (in my case), but the EOF
>>> is not aligned accordingly, so raw_co_block_status() returns an aligned
>>> *pnum.
>>
>> Uggh. Yet another reason why I want qemu to support byte-accurate
>> sizing, instead of rounding up. The rounding keeps raising its head in
>> more and more places. I have pending patches that are trying to improve
>> block status to round driver answers up to match request_alignment (when
>> the protocol layer has finer granularity than the format layer); but
>> this sounds like it is a bug in the file driver itself for returning an
>> answer that is not properly rounded according to its own
>> request_alignment boundary, and not one where my pending patches would help.
> 
> Yes, I think so, too.
> 
>>> I suppose having an unaligned tail is not so bad and maybe we can just
>>> adjust the assertion accordingly.  On the other hand, this has been
>>> broken for a while.  Does it even make sense to use O_DIRECT with
>>> unaligned images?  Shouldn’t we just reject them outright?
>>
>> The tail of an unaligned file is generally inaccessible to O_DIRECT,
> 
> Especially with this.
> 
>> where it is easier to use ftruncate() up to an aligned boundary if you
>> really must play with that region of the file, and then ftruncate() back
>> to the intended size after I/O. But that sounds hairy.  We could also
>> round down and silently ignore the tail of the file, but that is at odds
>> with our practice of rounding size up.  So for the short term, I'd be
>> happy with a patch that just rejects any attempt to use cache.direct=on
>> (O_DIRECT) with a file that is not already a multiple of the alignment
>> required thereby. (For reference, that's what qemu as NBD client
>> recently did when talking to a server that advertises a size
>> inconsistent with forced minimum block access: commit 3add3ab7)
> 
> OK, I’ll send a patch.  Thanks for you explanation!
Well, or maybe not.

$ ./qemu-img create -f qcow2 foo.qcow2 64M
$ ./qemu-img map --image-opts \
    driver=qcow2,file.filename=foo.qcow2,cache.direct=on
qemu-img: Could not open
'driver=qcow2,file.filename=foo.qcow2,cache.direct=on': File length
(196616 bytes) is not a multiple of the O_DIRECT alignment (512 bytes)
Try cache.direct=off, or increasing the file size to match the alignment

That may be considered a bug in qcow2.  Maybe it should always fill all
clusters.  But even if we did so and fixed it now, we can’t disallow
qemu from opening such images.

Also, well, the tail is accessible, we just need to access it with the
proper alignment (and then we get a short read).  This seems to be
handled just fine.

So I think file-posix should just return a rounded result.  Well, or
bdrv_co_Block_status() could ignore it for the EOF, because it throws
away everything past the EOF anyway with:

    if (*pnum > bytes) {
        *pnum = bytes;
    }

On one hand, I agree that file-posix should return an aligned result.
On the other, it doesn’t make a difference, so I don’t think we need to
enforce it (at EOF).

Max


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

Re: [Qemu-devel] Unaligned images with O_DIRECT

[Eric Blake]

[-- Attachment #1: Type: text/plain, Size: 3795 bytes --]

On 5/14/19 12:28 PM, Max Reitz wrote:

>>>
>>> The tail of an unaligned file is generally inaccessible to O_DIRECT,
>>
>> Especially with this.
>>
>>> where it is easier to use ftruncate() up to an aligned boundary if you
>>> really must play with that region of the file, and then ftruncate() back
>>> to the intended size after I/O. But that sounds hairy.  We could also
>>> round down and silently ignore the tail of the file, but that is at odds
>>> with our practice of rounding size up.  So for the short term, I'd be
>>> happy with a patch that just rejects any attempt to use cache.direct=on
>>> (O_DIRECT) with a file that is not already a multiple of the alignment
>>> required thereby. (For reference, that's what qemu as NBD client
>>> recently did when talking to a server that advertises a size
>>> inconsistent with forced minimum block access: commit 3add3ab7)
>>
>> OK, I’ll send a patch.  Thanks for you explanation!
> Well, or maybe not.
> 
> $ ./qemu-img create -f qcow2 foo.qcow2 64M
> $ ./qemu-img map --image-opts \
>     driver=qcow2,file.filename=foo.qcow2,cache.direct=on
> qemu-img: Could not open
> 'driver=qcow2,file.filename=foo.qcow2,cache.direct=on': File length
> (196616 bytes) is not a multiple of the O_DIRECT alignment (512 bytes)
> Try cache.direct=off, or increasing the file size to match the alignment
> 
> That may be considered a bug in qcow2.  Maybe it should always fill all
> clusters.  But even if we did so and fixed it now, we can’t disallow
> qemu from opening such images.
> 
> Also, well, the tail is accessible, we just need to access it with the
> proper alignment (and then we get a short read).  This seems to be
> handled just fine.

Oh. Yeah, short reads with O_DIRECT are possible (short writes not so
much; for those, you have to write a full buffer then ftruncate back
down). But we DO want to support short reads because of pre-existing
images, whether or not we also improve qcow2 to always create aligned
image sizes. The qcow2 spec allows unaligned images, even if we quit
creating new ones.

> 
> So I think file-posix should just return a rounded result.  Well, or
> bdrv_co_Block_status() could ignore it for the EOF, because it throws
> away everything past the EOF anyway with:
> 
>     if (*pnum > bytes) {
>         *pnum = bytes;
>     }
> 
> On one hand, I agree that file-posix should return an aligned result.
> On the other, it doesn’t make a difference, so I don’t think we need to
> enforce it (at EOF).

My thoughts:

Right now, only io.c sets (or even reads) BDRV_BLOCK_EOF, and it is
documented as an internal flag for optimizations.  But it would be very
easy to amend the contract of driver's .bdrv_co_block_status to state
that a driver may set BDRV_BLOCK_EOF at the end of a file, and MUST set
that flag if the end of the file also happens to be unaligned with
respect to the driver's request_alignment. (Most drivers won't need to
care, but file-posix.c under O_DIRECT would have to start caring).  Then
fix io.c to relax the assertion - the result must either be aligned
(current condition) OR the driver must have reported BDRV_BLOCK_EOF (new
condition). At that point, the block layer can take care of rounding out
the block status for the unaligned tail beyond EOF up to the alignment
boundary (similar to the rounding I have proposed in my other patches).
If you don't get to that first, then it looks like I'll have to fold
that in to my v2 patches when I get back to addressing those block
status alignment problems.

Thanks again for testing, and forcing me to think about the issue.

-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3226
Virtualization:  qemu.org | libvirt.org


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 488 bytes --]