LWN: Comments on "fallocate()"

fallocate() - ignoring POSIX

pjdc — Thu, 29 Mar 2007 06:55:08 +0000

Surely the C library can easily implement posix_fallocate() in terms of fallocate().

Incomplete patch?

giraffedata — Tue, 27 Mar 2007 15:50:29 +0000

I guess this might be a case where we assume that if a filesystem does not explicitly initialise the field, it is NULL, as the system call code does check for this field to be non-NULL. If it is NULL, it returns ->ENOSYS. So, I guess, you are right in saying that there are certain per-filesystem patches that need to be installed.

So you're saying that the assumption that if a filesystem driver does not explicitly initialize the field, it is NULL, is wrong? What makes you think that?

Incomplete patch?

shishir — Mon, 26 Mar 2007 17:44:15 +0000

i guess this might be a case where we assume that if a filesystem does not explicitly initialise the field, it is NULL, as the system call code does check for this field to be non-NULL. If it is NULL, it returns -ENOSYS. So, I guess, you are right in saying that there are certain per-filesystem patches that need to be installed.

fallocate() - ignoring POSIX

rwmj — Mon, 26 Mar 2007 14:28:31 +0000

I don't know if I'm reading this right, but are they proposing to ignore the POSIX standard?

Rich.

Incomplete patch?

giraffedata — Fri, 23 Mar 2007 20:46:58 +0000

Have you looked at GCC sources ? That's where this code is, after all...

No, it's not. It's in the filesystem drivers, in the statement that declares the inode_operations variables. Looking to GCC source code for the setting of this member to NULL is like looking at at GCC source code to see the setting of 'a' to 7 in a program that contains the line "int a = 7;"

Hint: what does C standard says about initialization of static and globar structures ?

An even better hint is that all the filesystem drivers (I hope) initialize the inode operations field by assigning from a static constant inode_operations variable. Because that's not obvious, and is essential to this patch working.

fallocate()

k8to — Fri, 23 Mar 2007 14:50:21 +0000

Just when almost all of the use cases had dried up ;-)

fallocate()

nix — Fri, 23 Mar 2007 14:37:25 +0000

Quite so! punch should really have been in POSIX from the start, except
that sparse files were never really in POSIX per se: they were just an
efficiency hack atop seeking, so nobody seems to have thought `hey, what
if we want to make an existing file sparser than it is?'

Incomplete patch?

khim — Fri, 23 Mar 2007 09:56:21 +0000

I can't see any code for initializing this field to any value

Have you looked at GCC sources ? That's where this code is, after all...

P.S. Hint: what does C standard says about initialization of static and globar structures ?

Incomplete patch?

ldo — Fri, 23 Mar 2007 05:21:34 +0000

It seems to me Amit Arora's patch is incomplete. It adds a new field to the inode_operations structure, but I can't see any code for initializing this field to any value. Presumably there are mandatory patches for all filesystems to at least set this field to NULL if they don't support this operation. Otherwise, an attempt to invoke this new system call could crash your system.

fallocate()

landley — Thu, 22 Mar 2007 21:48:42 +0000

Punch! Finally!

http://www.uwsg.iu.edu/hypermail/linux/kernel/0312.0/0889...

fallocate()

nix — Thu, 22 Mar 2007 11:50:04 +0000

FA_DEALLOCATE is probably easier to implement, at least for filesystems supporting sparse files at all.