Date:   Tue, 09 Aug 2022 14:28:39 -0400
From:   Jeff Layton <jlayton@...nel.org>
To:     "Darrick J. Wong" <djwong@...nel.org>
Cc:     Dave Chinner <david@...morbit.com>, linux-fsdevel@...r.kernel.org,
        dhowells@...hat.com, lczerner@...hat.com, bxue@...hat.com,
        ceph-devel@...r.kernel.org, linux-nfs@...r.kernel.org,
        linux-afs@...ts.infradead.org, linux-ext4@...r.kernel.org,
        linux-xfs@...r.kernel.org, linux-btrfs@...r.kernel.org,
        fstests <fstests@...r.kernel.org>
Subject: Re: [RFC PATCH 1/4] vfs: report change attribute in statx for
 IS_I_VERSION inodes

On Tue, 2022-08-09 at 14:04 -0400, Jeff Layton wrote:
> On Tue, 2022-08-09 at 08:33 -0700, Darrick J. Wong wrote:
> > On Mon, Aug 08, 2022 at 09:19:05AM -0400, Jeff Layton wrote:
> > > On Fri, 2022-08-05 at 18:06 -0400, Jeff Layton wrote:
> > > > On Sat, 2022-08-06 at 08:01 +1000, Dave Chinner wrote:
> > > > > On Fri, Aug 05, 2022 at 02:35:40PM -0400, Jeff Layton wrote:
> > > > > > From: Jeff Layton <jlayton@...hat.com>
> > > > > > 
> > > > > > Claim one of the spare fields in struct statx to hold a 64-bit change
> > > > > > attribute. When statx requests this attribute, do an
> > > > > > inode_query_iversion and fill the result in the field.
> > > > > > 
> > > > > > Also update the test-statx.c program to fetch the change attribute as
> > > > > > well.
> > > > > > 
> > > > > > Signed-off-by: Jeff Layton <jlayton@...nel.org>
> > > > > > ---
> > > > > >  fs/stat.c                 | 7 +++++++
> > > > > >  include/linux/stat.h      | 1 +
> > > > > >  include/uapi/linux/stat.h | 3 ++-
> > > > > >  samples/vfs/test-statx.c  | 4 +++-
> > > > > >  4 files changed, 13 insertions(+), 2 deletions(-)
> > > > > > 
> > > > > > diff --git a/fs/stat.c b/fs/stat.c
> > > > > > index 9ced8860e0f3..976e0a59ab23 100644
> > > > > > --- a/fs/stat.c
> > > > > > +++ b/fs/stat.c
> > > > > > @@ -17,6 +17,7 @@
> > > > > >  #include <linux/syscalls.h>
> > > > > >  #include <linux/pagemap.h>
> > > > > >  #include <linux/compat.h>
> > > > > > +#include <linux/iversion.h>
> > > > > >  
> > > > > >  #include <linux/uaccess.h>
> > > > > >  #include <asm/unistd.h>
> > > > > > @@ -118,6 +119,11 @@ int vfs_getattr_nosec(const struct path *path, struct kstat *stat,
> > > > > >  	stat->attributes_mask |= (STATX_ATTR_AUTOMOUNT |
> > > > > >  				  STATX_ATTR_DAX);
> > > > > >  
> > > > > > +	if ((request_mask & STATX_CHGATTR) && IS_I_VERSION(inode)) {
> > > > > > +		stat->result_mask |= STATX_CHGATTR;
> > > > > > +		stat->chgattr = inode_query_iversion(inode);
> > > > > > +	}
> > > > > 
> > > > > If you're going to add generic support for it, shouldn't there be a
> > > > > generic test in fstests that ensures that filesystems that advertise
> > > > > STATX_CHGATTR support actually behave correctly? Including across
> > > > > mounts, and most importantly, that it is made properly stable by
> > > > > fsync?
> > > > > 
> > > > > i.e. what good is this if different filesystems have random quirks
> > > > > that mean it can't be relied on by userspace to tell it changes have
> > > > > occurred?
> > > > 
> > > > Absolutely. Being able to better test the i_version field for consistent
> > > > behavior is a primary goal. I haven't yet written any yet, but we'd
> > > > definitely want something in xfstests if we decide this is worthwhile.
> > > 
> > > I started writing some tests for this today, and hit a bit of a chicken-
> > > and-egg problem:
> > > 
> > > I'd prefer to use xfs_io for easier maintainability, but the STATX_*
> > > constants are defined via UAPI header. Older kernels don't have them and
> > > old xfs_io programs don't understand or report this value.
> > > 
> > > Should I just write a one-off binary program for xfstests to fetch this
> > > value for now, or are we better off merging the patchset first, and then
> > > fix xfs_io and then write the tests using the updated xfs_io program?
> > 
> > What we've done in the past to support new APIs until they land in
> > kernel headers is:
> > 
> > Add an autoconf macro to decide if the system header files are recent
> > enough to support whatever functionality is needed by xfs_io;
> > 
> > Modify the build system to #define OVERRIDE_FUBAR if the system headers
> > aren't new enough to have FUBAR; and
> > 
> > Modify (or create) the relevant header file to override the system
> > header definitions as needed to support building the relevant pieces of
> > code.  A year or so after the functionality lands, we can then remove
> > the overrides, or just leave them in place until the next time we need
> > it.
> > 
> > For example, Eric Biggers wanted to teach the fscrypt commands to use a
> > new feature he was adding to an existing API, so he AC_DEFUN'd a macro
> > that checks to see if the system linux/fs.h *does not* define a
> > structure containing the desired field.  If this is the case, it sets
> > need_internal_fscrypt_add_key_arg=yes.
> > 
> > AC_DEFUN([AC_NEED_INTERNAL_FSCRYPT_ADD_KEY_ARG],
> >   [
> >     AC_CHECK_TYPE(struct fscrypt_add_key_arg,
> >       [
> >         AC_CHECK_MEMBER(struct fscrypt_add_key_arg.key_id,
> >           ,
> >           need_internal_fscrypt_add_key_arg=yes,
> >           [#include <linux/fs.h>]
> >         )
> >       ],,
> >       [#include <linux/fs.h>]
> >     )
> >     AC_SUBST(need_internal_fscrypt_add_key_arg)
> >   ])
> > 
> > This macro is called from configure.ac.
> > 
> > Next, include/builddefs.in was modified to include the selected value in
> > the make variables:
> > 
> > NEED_INTERNAL_FSCRYPT_ADD_KEY_ARG = @need_internal_fscrypt_add_key_arg@
> > 
> > And then the shouty variable is used in the same file to set a compiler
> > define:
> > 
> > ifeq ($(NEED_INTERNAL_FSCRYPT_ADD_KEY_ARG),yes)
> > PCFLAGS+= -DOVERRIDE_SYSTEM_FSCRYPT_ADD_KEY_ARG
> > endif
> > 
> > Then io/encrypt.c does the following to move the system's definition of
> > struct fscrypt_add_key_arg out of the way...
> > 
> > #ifdef OVERRIDE_SYSTEM_FSCRYPT_ADD_KEY_ARG
> > #  define fscrypt_add_key_arg sys_fscrypt_add_key_arg
> > #endif
> > #include <linux/fs.h>  /* via io.h -> xfs.h -> xfs/linux.h */
> > 
> > ...so that the file can provide its own definition further down:
> > 
> > /*
> >  * Since the key_id field was added later than struct
> >  * fscrypt_add_key_arg itself, we may need to override the system
> >  * definition to get that field.
> >  */
> > #if !defined(FS_IOC_ADD_ENCRYPTION_KEY) || \
> > 	defined(OVERRIDE_SYSTEM_FSCRYPT_ADD_KEY_ARG)
> > #undef fscrypt_add_key_arg
> > struct fscrypt_add_key_arg {
> > 	struct fscrypt_key_specifier key_spec;
> > 	__u32 raw_size;
> > 	__u32 key_id;
> > 	__u32 __reserved[8];
> > 	__u8 raw[];
> > };
> > #endif
> > 
> 
> Darrick, thanks for the detailed instructions. They've been very
> helpful! My approach is slightly different since xfsprogs already has a
> statx.h. I'm just updating that to allow for overriding. I think I have
> the autoconf part worked out.
> 
> I'm having a problem with the above though. I have this in statx.h:
> 
> #undef statx_timestamp
> struct statx_timestamp {
>         __s64   tv_sec;
>         __s32   tv_nsec;
>         __s32   __reserved;
> };
> 
> ...but when I go to build, I get this:
> 
> In file included from stat.c:11:
> statx.h:60:8: error: redefinition of ‘struct statx_timestamp’
>    60 | struct statx_timestamp {
>       |        ^~~~~~~~~~~~~~~
> 
> ...it seems like the "#undef statx_timestamp" isn't doing the right
> thing. Is my syntax wrong?

I think I got it sorted. I had to rename the structs and then use the
preprocessor to override the old struct names in the code. It's a little
ugly, but it works:

struct statx_timestamp_internal {
        __s64   tv_sec;
        __s32   tv_nsec;
        __s32   __reserved;
};
#define statx_timestamp statx_timestamp_internal

I just sent a patch to linux-xfs that seems to do the right thing.
-- 
Jeff Layton <jlayton@...nel.org>

Powered by blists - more mailing lists