LinuxLists.cc - [git pull] vfs pile 1

2014-01-27 14:25:26

Subject: [git pull] vfs pile 1

2014-01-27 23:05:31

by Stephen Rothwell

[permalink] [raw]

Subject: Re: [git pull] vfs pile 1

Hi Al,

On Mon, 27 Jan 2014 14:25:20 +0000 Al Viro <[email protected]> wrote:
>
> Assorted stuff; the biggest pile here is Christoph's ACL series.
> Plus assorted cleanups and fixes all over the place... There will be
> another pile later this week.
>
> There's a couple of conflicts - rcupdate.h and posix_acl.h; proposed resolution
> in vfs.git#conflict-resolution. Please, pull from the usual place -
> git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs.git for-linus

Yet again, none of this has been anywhere near linux-next, in fact there
has been nothing in the vfs tree since v3.13-rc1 :-(

Some of these patches are dated last December, yet none of them were
committed to that tree before last Friday.

--
Cheers,
Stephen Rothwell [email protected]

Attachments:

(No filename) (795.00 B)
(No filename) (836.00 B)
Download all attachments

2014-01-29 03:26:47

by Linus Torvalds

[permalink] [raw]

Subject: Re: [git pull] vfs pile 1

On Mon, Jan 27, 2014 at 6:25 AM, Al Viro <[email protected]> wrote:
> Assorted stuff; the biggest pile here is Christoph's ACL series.
> Plus assorted cleanups and fixes all over the place... There will be
> another pile later this week.

The posix_acl_chmod() code looks wrong.

Not that it looked right before either, but whatever. The code
basically looks like some variation of this in most setattr()
implementations:

if (ia_valid & ATTR_MODE)
rc = posix_acl_chmod(inode, inode->i_mode);

but the mode we're changing to (and what ATTR_MODE guards) is actually
attr->ia_mode, not inode->i_mode. And quite frankly, passing in
inode->i_mode looks stupid, since we're already passing in the inode
pointer, so that's just redundant and pointless information.

Anyway, I noticed this after doing the (untested, and still un-acked -
hint, hint) ceph conversion. In that, I made ceph use attr->ia_mode.
Maybe that was wrong, but at least it's not insane and stupid like the
other filesystem implementations are.

Comments?

Linus

2014-01-29 13:37:19

by Jan Kara

[permalink] [raw]

Subject: Re: [git pull] vfs pile 1

On Tue 28-01-14 19:26:08, Linus Torvalds wrote:
> On Mon, Jan 27, 2014 at 6:25 AM, Al Viro <[email protected]> wrote:
> > Assorted stuff; the biggest pile here is Christoph's ACL series.
> > Plus assorted cleanups and fixes all over the place... There will be
> > another pile later this week.
>
> The posix_acl_chmod() code looks wrong.
>
> Not that it looked right before either, but whatever. The code
> basically looks like some variation of this in most setattr()
> implementations:
>
> if (ia_valid & ATTR_MODE)
> rc = posix_acl_chmod(inode, inode->i_mode);
>
> but the mode we're changing to (and what ATTR_MODE guards) is actually
> attr->ia_mode, not inode->i_mode.
Yes, but posix_acl_chmod() is called after setattr_copy() was done so
inode->i_mode should be the same as attr->ia_mode. Whether i_mode or
ia_mode is mode logical depends on whether you view posix_acl_chmod() as
"sync current i_mode into acls" or "reflect this i_mode change in acls".
I agree the function name suggests more the latter semantics.

> And quite frankly, passing in inode->i_mode looks stupid, since we're
> already passing in the inode pointer, so that's just redundant and
> pointless information.
Yes, it looks stupid. We could almost drop that argument, except that f2fs
tries to play some tricks with i_mode and stores i_mode in a different
place when acls are enabled. Huh? Jaegeuk, can you explain why are you
doing that?

Honza
--
Jan Kara <[email protected]>
SUSE Labs, CR

2014-01-30 02:02:51

by Jaegeuk Kim

[permalink] [raw]

Subject: Re: [git pull] vfs pile 1

2014-01-29 Jan Kara <[email protected]>:
> On Tue 28-01-14 19:26:08, Linus Torvalds wrote:
>> On Mon, Jan 27, 2014 at 6:25 AM, Al Viro <[email protected]> wrote:
>> > Assorted stuff; the biggest pile here is Christoph's ACL series.
>> > Plus assorted cleanups and fixes all over the place... There will be
>> > another pile later this week.
>>
>> The posix_acl_chmod() code looks wrong.
>>
>> Not that it looked right before either, but whatever. The code
>> basically looks like some variation of this in most setattr()
>> implementations:
>>
>> if (ia_valid & ATTR_MODE)
>> rc = posix_acl_chmod(inode, inode->i_mode);
>>
>> but the mode we're changing to (and what ATTR_MODE guards) is actually
>> attr->ia_mode, not inode->i_mode.
> Yes, but posix_acl_chmod() is called after setattr_copy() was done so
> inode->i_mode should be the same as attr->ia_mode. Whether i_mode or
> ia_mode is mode logical depends on whether you view posix_acl_chmod() as
> "sync current i_mode into acls" or "reflect this i_mode change in acls".
> I agree the function name suggests more the latter semantics.
>
>> And quite frankly, passing in inode->i_mode looks stupid, since we're
>> already passing in the inode pointer, so that's just redundant and
>> pointless information.
> Yes, it looks stupid. We could almost drop that argument, except that f2fs
> tries to play some tricks with i_mode and stores i_mode in a different
> place when acls are enabled. Huh? Jaegeuk, can you explain why are you
> doing that?

As described to Christoph before, the reason is for acl consistency
between on-disk xattr->mode and on-disk inode->mode.

Previously, there are three i_modes managed by:
inode->mode on-disk xattr->mode on-disk->i_mode
f2fs_setattr [x] y y
[update_inode] x y [x]
[checkpoint] x [y] x
__f2fs_setxattr x [x] x

In this flow, f2fs is able to break the consistency between on-disk
xattr->mode and on-disk->i_mode after checkpoint followed by
sudden-power-off.

So, fi->i_mode was introduced to address the problem.
The new f2fs_setattr triggers:
inode->mode fi->i_mode on-disk xattr->mode on-disk->i_mode
f2fs_setattr y [x] y
y
[update_inode] y x y
y
[checkpoint] y x y
y
__f2fs_setxattr [x] x [x]
[x]

Finally, __f2fs_setxattr synchronizes inode->mode, on-disk xattr->mode,
and on-disk inode->i_mode all together.

Am I missing something?

Thanks,

>
> Honza
> --
> Jan Kara <[email protected]>
> SUSE Labs, CR
> --
> To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html

2014-01-30 13:07:14

by Christoph Hellwig

[permalink] [raw]

Subject: Re: [git pull] vfs pile 1

On Tue, Jan 28, 2014 at 07:26:08PM -0800, Linus Torvalds wrote:
> but the mode we're changing to (and what ATTR_MODE guards) is actually
> attr->ia_mode, not inode->i_mode. And quite frankly, passing in
> inode->i_mode looks stupid, since we're already passing in the inode
> pointer, so that's just redundant and pointless information.

At this point inode->i_mode has been updated to attra->ia_mode. Passing
in the mode allows fileystems to defer the i_mode update until after
the ACL has been modified, which at this point only f2fs does.

In a perfect world filesystems do both under the same lock for in-core
access and in an transaction for the on-disk change, in which case
the exact order doesn't matter.

2014-01-30 15:26:41

by Jan Kara

[permalink] [raw]

Subject: Re: [git pull] vfs pile 1

On Thu 30-01-14 11:02:49, Kim Jaegeuk wrote:
> 2014-01-29 Jan Kara <[email protected]>:
> > On Tue 28-01-14 19:26:08, Linus Torvalds wrote:
> >> On Mon, Jan 27, 2014 at 6:25 AM, Al Viro <[email protected]> wrote:
> >> > Assorted stuff; the biggest pile here is Christoph's ACL series.
> >> > Plus assorted cleanups and fixes all over the place... There will be
> >> > another pile later this week.
> >>
> >> The posix_acl_chmod() code looks wrong.
> >>
> >> Not that it looked right before either, but whatever. The code
> >> basically looks like some variation of this in most setattr()
> >> implementations:
> >>
> >> if (ia_valid & ATTR_MODE)
> >> rc = posix_acl_chmod(inode, inode->i_mode);
> >>
> >> but the mode we're changing to (and what ATTR_MODE guards) is actually
> >> attr->ia_mode, not inode->i_mode.
> > Yes, but posix_acl_chmod() is called after setattr_copy() was done so
> > inode->i_mode should be the same as attr->ia_mode. Whether i_mode or
> > ia_mode is mode logical depends on whether you view posix_acl_chmod() as
> > "sync current i_mode into acls" or "reflect this i_mode change in acls".
> > I agree the function name suggests more the latter semantics.
> >
> >> And quite frankly, passing in inode->i_mode looks stupid, since we're
> >> already passing in the inode pointer, so that's just redundant and
> >> pointless information.
> > Yes, it looks stupid. We could almost drop that argument, except that f2fs
> > tries to play some tricks with i_mode and stores i_mode in a different
> > place when acls are enabled. Huh? Jaegeuk, can you explain why are you
> > doing that?
>
> As described to Christoph before, the reason is for acl consistency
> between on-disk xattr->mode and on-disk inode->mode.
>
> Previously, there are three i_modes managed by:
> inode->mode on-disk xattr->mode on-disk->i_mode
> f2fs_setattr [x] y y
> [update_inode] x y [x]
> [checkpoint] x [y] x
> __f2fs_setxattr x [x] x
>
> In this flow, f2fs is able to break the consistency between on-disk
> xattr->mode and on-disk->i_mode after checkpoint followed by
> sudden-power-off.
>
> So, fi->i_mode was introduced to address the problem.
> The new f2fs_setattr triggers:
> inode->mode fi->i_mode on-disk xattr->mode on-disk->i_mode
> f2fs_setattr y [x] y
> y
> [update_inode] y x y
> y
> [checkpoint] y x y
> y
> __f2fs_setxattr [x] x [x]
> [x]
>
> Finally, __f2fs_setxattr synchronizes inode->mode, on-disk xattr->mode,
> and on-disk inode->i_mode all together.
>
> Am I missing something?
OK, I see. Thanks for explanation.

Honza
--
Jan Kara <[email protected]>
SUSE Labs, CR