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

Re: [libvirt] [PATCH 3/4] vol-info: Check for NFS FS type



On 08/09/2014 12:28 AM, John Ferlan wrote:
>> Testing seems to indicate that posix_fallocate() either doesn't work as
>> expected on the target or using the target.path is incorrect...
>>
>> Before posix_fallocate
>> stat st_blocks=0 st_blksize=1048576 st_size=10485760
>> lseek end=10485760
>>
>> posix_fallocate of 10485760 bytes on /home/nfs_pool/target/test-vol1
>>
>> After posix_fallocate
>> stat st_blocks=88 st_blksize=1048576 st_size=10485760
>> lseek end=10485760
>>
>>
>> Hmm... would going at the target be correct in this instance?  Same test
>> but use the source path:

Using the source path would only work if the NFS export is on the localhost :)

>>
>> Before posix_fallocate
>> stat st_blocks=0 st_blksize=4096 st_size=10485760
>> lseek end=10485760
>>
>> posix_fallocate of 10485760 bytes on /home/nfs_pool/nfs-export/test-vol1
>>
>> After posix_fallocate
>> stat st_blocks=20480 st_blksize=4096 st_size=10485760
>> lseek end=10485760
>>
>> ...
>>
>> hmm.... 20480 * 512 = 10485760
>>
>> # df
>> ...
>> localhost:/home/nfs_pool/nfs-export 140979200 35521536  98273280  27%
>> /home/nfs_pool/target
>> #
>>
> 
> Well it's a tangled web that's being weaved...  The blksize of the
> target volume comes from the 'wsize' value in the mount:
> 
> localhost:/home/nfs_pool/nfs-export on /home/nfs_pool/target type nfs4
> (rw,relatime,vers=4.0,rsize=1048576,wsize=1048576,...)
> 
> Further testing shows if I change the wsize to 4096, then I get what I
> expect; however, starting at 8192 I'll get decreasingly smaller
> allocations. So this is a "math problem".
> 
> So going back to writing via posix_fallocate() to the
> "/home/nfs_pool/target/test-vol" the "issue" is the blksize of the
> "source" (nfs-export) is 4096 while the blksize of the "target" (target)
> is 1048576 (as a result of the nfs mount settings).  What "seems" to
> happen is the posix_fallocate() makes 11 "writes" - I assume because
> blksize*10 < desired_size (10485760) (instead of <=...).
> 
> Thus 11 * 4096 = 45056 (bytes) / 1024 = 44 KiB which was displayed.
> 
> Why all this happens I'm not sure.  Bug in posix_fallocate()? Bug in
> configuration?  I have to assume that when this code was first added NFS
> probably was still using smaller block sizes.

The code was introduced in 2013. Maybe it wasn't tested on NFS at all?
It looks like a posix_fallocate bug to me.

The other method we have (syscall(SYS_fallocate)) gives me EOPNOTSUPP -
Operation not supported on transport endpoint.

> Whether anyone has
> noticed or not beyond the virt-test which discovered the issue - I'm not
> sure.  In any case, does anyone have feedback/thoughts for next steps?
> I can put together something that avoids posix_fallocate() for the
> create-as and resize paths.
> 

I think reporting an error when preallocate is requested for a NFS pool makes
sense, but that might be pretty annoying if something supplies the preallocate
flag by default.

Jan

Attachment: signature.asc
Description: OpenPGP digital signature


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