XFS Preallocate using ALLOCSP

[Felix Blyakher]

On Jun 16, 2009, at 10:16 AM, Eric Sandeen wrote:

> Smit Shah wrote:
>
>> Even the man page of fallocate says that it allocates and initializes
>> to zero the disk space allocated
>
> Bleah, so it does:
>
>       FALLOC_FL_KEEP_SIZE
>              This flag allocates and initializes to zero the disk  
> space
>
> well, that's misleading and/or wrong.
>
>> but when i saw the code i did found out that it does not zero it out.
>> Hence was a kindof confused. So posix_allocate is similar to ALLOCSP
>> when falloc is not supported by the underlying filesystem  that is to
>> ftruncate the file  and zero it out. So all of them try to allocate
>> contiguous blocks but the only difference is when we use the
>> fallocate in ext4/xfs it does not zero out the preallocated space. Am
>> i right ?
>
> fallocate / sys_fallocate marks the region as uninitialized so that  
> you
> get back 0s when you read.  It's implemented on xfs, ext4, ocfs2,  
> and btrfs.
>
> posix_fallocate manages to reach sys_fallocate when all the stars  
> align:
> kernel,  glibc, and filesystgem.  Otherwise it writes 0s.
>
>> But  when i fallocate in ext4 i can see the write performance
>> improvement but not in xfs
>
> Testing how?
>
>> and reason i found out in one of your previous comments is  because
>> of the unwritten flag set in xfs. So how do we see if the unwritten
>> flag is set or not ? I did use xfs_info but it didnt show any such
>> information.
>
> ext4 & xfs are doing the same basic thing, they must maintain the
> unwritten state on the preallocated extents, and manage that as it
> changes when portions are written with real data.

Well, the difference in managing the unwritten state can
theoretically result in different performance. Not that I'd
expect ext4 being better than xfs in this respect.
More data is needed here.

> xfs_bmap -v -v -p on a file will show you extent state for xfs.
>
>> I guess i am not right here ftruncate simply does a lseek and wirtes
>
> ftruncate simply sets i_size, it does no data IO.

... and no block reservation/allocation either.

>> to it which might not be contiguous whereas fallocate tries to
>> allocate contiguous block so as to reduce fragmentation
>
> Actually fallocate's only official job is to reserve blocks so you  
> don't
> get ENOSPC later.  Because the request comes in all at once, you are
> very likely to get an optimal allocation, and that's a nice side  
> effect,
> but it's not actually required by the interface.
>
>> and hence i
>> thought to reduce fragmentation and for security reasons
>
> None of these normal interfaces poses any security risk.  If you build
> xfs without the unwritten extent feature

I don't think, it's possible. Not in any configurable way,
at least.

> you could allocate w/o flagging
> uninitialized and expose stale data, but that's not a normal mode of
> operation.

That's was possible with mount option unwritten=0, but
AFAIK, it's been recently completely removed from code.

>
>
>> its better
>> to use ALLOCSP rather than something like ftruncate /posix_fallocate
>> or RSEVSP which kindof performs bad for writes with unwritten flag
>> set and now there being a no direct way while creating the fs to
>> disable unwritten.
>
> In the end, there are only 2 ways to preallocate blocks: explicitly
> write 0s, or flag regions as unwritten (as xfs/ext4/... can do).

Exactly.
That's a trade-off between spending time on setup or at the
write time. And if explicit zeros are desirable for the former
approach, it can be driven from the user space (after
preallocation) rather than from the kernel with the exactly the
same outcome.

Just restating the same what Eric already said :)

Felix

>  (Ok,
> or a 3rd sorta-way, which is to reserve w/o flagging, maybe that's  
> what
> you're looking for, but that's deprecated or not really available at
> this point).
>
> Maybe I should ask what the end goal is here.  :)
>
> -Eric
>
>> Thanks, Smit
>
> _______________________________________________
> xfs mailing list
> xfs at oss.sgi.com
> http://oss.sgi.com/mailman/listinfo/xfs