Re: [Qemu-devel] [PATCH] Fix block migration when the device size is not a multiple of 1 MB

[Pierre Riteau]

On 21 janv. 2011, at 14:59, Yoshiaki Tamura wrote:

> 2011/1/21 Pierre Riteau <address@hidden>:
>> On 21 janv. 2011, at 13:36, Yoshiaki Tamura wrote:
>> 
>>> 2011/1/21 Kevin Wolf <address@hidden>:
>>>> Am 21.01.2011 13:15, schrieb Yoshiaki Tamura:
>>>>> 2011/1/21 Pierre Riteau <address@hidden>:
>>>>>> Le 20 janv. 2011 à 17:18, Yoshiaki Tamura <address@hidden> a écrit :
>>>>>> 
>>>>>>> 2011/1/20 Pierre Riteau <address@hidden>:
>>>>>>>> On 20 janv. 2011, at 03:06, Yoshiaki Tamura wrote:
>>>>>>>> 
>>>>>>>>> 2011/1/19 Pierre Riteau <address@hidden>:
>>>>>>>>>> b02bea3a85cc939f09aa674a3f1e4f36d418c007 added a check on the return
>>>>>>>>>> value of bdrv_write and aborts migration when it fails. However, if 
>>>>>>>>>> the
>>>>>>>>>> size of the block device to migrate is not a multiple of BLOCK_SIZE
>>>>>>>>>> (currently 1 MB), the last bdrv_write will fail with -EIO.
>>>>>>>>>> 
>>>>>>>>>> Fixed by calling bdrv_write with the correct size of the last block.
>>>>>>>>>> ---
>>>>>>>>>>  block-migration.c |   16 +++++++++++++++-
>>>>>>>>>>  1 files changed, 15 insertions(+), 1 deletions(-)
>>>>>>>>>> 
>>>>>>>>>> diff --git a/block-migration.c b/block-migration.c
>>>>>>>>>> index 1475325..eeb9c62 100644
>>>>>>>>>> --- a/block-migration.c
>>>>>>>>>> +++ b/block-migration.c
>>>>>>>>>> @@ -635,6 +635,8 @@ static int block_load(QEMUFile *f, void *opaque, 
>>>>>>>>>> int version_id)
>>>>>>>>>>     int64_t addr;
>>>>>>>>>>     BlockDriverState *bs;
>>>>>>>>>>     uint8_t *buf;
>>>>>>>>>> +    int64_t total_sectors;
>>>>>>>>>> +    int nr_sectors;
>>>>>>>>>> 
>>>>>>>>>>     do {
>>>>>>>>>>         addr = qemu_get_be64(f);
>>>>>>>>>> @@ -656,10 +658,22 @@ static int block_load(QEMUFile *f, void 
>>>>>>>>>> *opaque, int version_id)
>>>>>>>>>>                 return -EINVAL;
>>>>>>>>>>             }
>>>>>>>>>> 
>>>>>>>>>> +            total_sectors = bdrv_getlength(bs) >> BDRV_SECTOR_BITS;
>>>>>>>>>> +            if (total_sectors <= 0) {
>>>>>>>>>> +                fprintf(stderr, "Error getting length of block 
>>>>>>>>>> device %s\n", device_name);
>>>>>>>>>> +                return -EINVAL;
>>>>>>>>>> +            }
>>>>>>>>>> +
>>>>>>>>>> +            if (total_sectors - addr < 
>>>>>>>>>> BDRV_SECTORS_PER_DIRTY_CHUNK) {
>>>>>>>>>> +                nr_sectors = total_sectors - addr;
>>>>>>>>>> +            } else {
>>>>>>>>>> +                nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
>>>>>>>>>> +            }
>>>>>>>>>> +
>>>>>>>>>>             buf = qemu_malloc(BLOCK_SIZE);
>>>>>>>>>> 
>>>>>>>>>>             qemu_get_buffer(f, buf, BLOCK_SIZE);
>>>>>>>>>> -            ret = bdrv_write(bs, addr, buf, 
>>>>>>>>>> BDRV_SECTORS_PER_DIRTY_CHUNK);
>>>>>>>>>> +            ret = bdrv_write(bs, addr, buf, nr_sectors);
>>>>>>>>>> 
>>>>>>>>>>             qemu_free(buf);
>>>>>>>>>>             if (ret < 0) {
>>>>>>>>>> --
>>>>>>>>>> 1.7.3.5
>>>>>>>>>> 
>>>>>>>>>> 
>>>>>>>>>> 
>>>>>>>>> 
>>>>>>>>> Hi Pierre,
>>>>>>>>> 
>>>>>>>>> I don't think the fix above is correct.  If you have a file which
>>>>>>>>> isn't aliened with BLOCK_SIZE, you won't get an error with the
>>>>>>>>> patch.  However, the receiver doesn't know how much sectors which
>>>>>>>>> the sender wants to be written, so the guest may fail after
>>>>>>>>> migration because some data may not be written.  IIUC, although
>>>>>>>>> changing bytestream should be prevented as much as possible, we
>>>>>>>>> should save/load total_sectors to check appropriate file is
>>>>>>>>> allocated on the receiver side.
>>>>>>>> 
>>>>>>>> Isn't the guest supposed to be started using a file with the correct 
>>>>>>>> size?
>>>>>>> 
>>>>>>> I personally don't like that; It's insisting too much to the user.
>>>>>>> Can't we expand the image on the fly?  We can just abort if expanding
>>>>>>> failed anyway.
>>>>>> 
>>>>>> At first I thought your expansion idea was best, but now I think there 
>>>>>> are valid scenarios where it fails.
>>>>>> 
>>>>>> Imagine both sides are not using a file but a disk partition as storage. 
>>>>>> If the partition size is not rounded to 1 MB, the last write will fail 
>>>>>> with the current code, and there is no way we can expand the partition.
>>>>>> 
>>>>> 
>>>>> Right.  But in case of partition doesn't the check in the patch below
>>>>> return error?  Does bdrv_getlength return the size correctly?
>>>> 
>>>> I'm pretty sure that it does. We would have problems in other places if
>>>> it didn't (e.g. we're checking if I/O requests are within the disk size).
>>> 
>>> Sorry for the noise.  I just learned it's returning the value of lseek
>>> in case of raw-posix.
>> 
>> 
>> And it does a ioctl call on other platforms than Linux.
> 
> Thanks.  Just a quick question regarding total_sectors.
> BlockDriverState seems to contain total_sectors.  Can we avoid
> calling bdrv_getlength() if bs->total_sectors were already there?

From a comment in bdrv_getlength():

Fixed size devices use the total_sectors value for speed instead of
issuing a length query (like lseek) on each call.  Also, legacy block
drivers don't provide a bdrv_getlength function and must use
total_sectors.

So using bdrv_getlength will protect against devices being resized during 
migration, but as far as I can see, the sender side doesn't support it: the 
value of total_sectors is cached for the whole block migration.

-- 
Pierre Riteau -- PhD student, Myriads team, IRISA, Rennes, France
http://perso.univ-rennes1.fr/pierre.riteau/