Re: [Qemu-devel] [PATCH] linux-user: Pass through nanosecond timestamp components for stat syscalls

[Chen-Yu Tsai <wens@kernel.org>]

On Wed, May 22, 2019 at 5:08 PM Laurent Vivier <laurent@vivier.eu> wrote:
>
> On 14/05/2019 16:53, Chen-Yu Tsai wrote:
> > From: Chen-Yu Tsai <wens@csie.org>
> >
> > Since Linux 2.6 the stat syscalls have mostly supported nanosecond
> > components for each of the file-related timestamps.
> >
> > QEMU user mode emulation currently does not pass through the nanosecond
> > portion of the timestamp, even when the host system fills in the value.
> > This results in a mismatch when run on subsecond resolution filesystems
> > such as ext4 or XFS.
> >
> > An example of this leading to inconsistency is cross-debootstraping a
> > full desktop root filesystem of Debian Buster. Recent versions of
> > fontconfig store the full timestamp (instead of just the second portion)
> > of the directory in its per-directory cache file, and checks this against
> > the directory to see if the cache is up-to-date. With QEMU user mode
> > emulation, the timestamp stored is incorrect, and upon booting the rootfs
> > natively, fontconfig discovers the mismatch, and proceeds to rebuild the
> > cache on the comparatively slow machine (low-power ARM vs x86). This
> > stalls the first attempt to open whatever application that incorporates
> > fontconfig.
> >
> > This patch renames the "unused" padding trailing each timestamp element
> > to its nanosecond counterpart name if such an element exists in the
> > kernel sources for the given platform. Not all do. Then have the syscall
> > wrapper fill in the nanosecond portion if the host supports it, as
> > specified by the _POSIX_C_SOURCE and _XOPEN_SOURCE feature macros.
> >
> > Recent versions of glibc only use stat64 and newfstatat syscalls on
> > 32-bit and 64-bit platforms respectively. The changes in this patch
> > were tested by directly calling the stat, stat64 and newfstatat syscalls
> > directly, in addition to the glibc wrapper, on arm and aarch64 little
> > endian targets.
> >
> > Signed-off-by: Chen-Yu Tsai <wens@csie.org>
> >
> > ---
> >
> > This issue was found while integrating some software that uses newer
> > versions of fontconfig into Raspbian images. We found that the first
> > launch of said software always stalls with fontconfig regenerating its
> > font cache files. Upon closer examination I found the timestamps were
> > not matching. The rest is explained above. Currently we're just working
> > around the problem by patching the correct timestamps into the cache
> > files after the fact.
> >
> > Please consider this a drive-by scratch-my-own-itch contribution, but I
> > will stick around to deal with any comments raised during review. I'm
> > not on the mailing lists either, so please keep me in CC.
> >
> > checkpatch returns "ERROR: code indent should never use tabs" for
> > linux-user/syscall_defs.h, however as far as I can tell the whole file
> > is indented with tabs. I'm not sure what to make of this.
>
> Yes, the file is entirely indented with tabs, so you can let this as-is.
> Anyway, I plan to split the file in several ones so we will be able to
> swap the tabs with spaces.
>
> Reviewed-by: Laurent Vivier <laurent@vivier.eu>

Thanks. Unfortunately this patch has some issues. It fails to build for
targets that don't have the *_nsec fields, such as Alpha or M68K.

I'll spin a v2 with a new macro TARGET_STAT_HAS_NSEC defined for targets
that have the fields, added before each struct stat definition. The hunk
below will gain a check against said macro. This is pretty much how the
kernel deals with the difference as well, as I just found out.

> > @@ -8866,6 +8876,14 @@ static abi_long do_syscall1(void *cpu_env, int num, abi_long arg1,
> >                   __put_user(st.st_atime, &target_st->target_st_atime);
> >                   __put_user(st.st_mtime, &target_st->target_st_mtime);
> >                   __put_user(st.st_ctime, &target_st->target_st_ctime);
> > +#if _POSIX_C_SOURCE >= 200809L || _XOPEN_SOURCE >= 700
> > +                __put_user(st.st_atim.tv_nsec,
> > +                           &target_st->target_st_atime_nsec);
> > +                __put_user(st.st_mtim.tv_nsec,
> > +                           &target_st->target_st_mtime_nsec);
> > +                __put_user(st.st_ctim.tv_nsec,
> > +                           &target_st->target_st_ctime_nsec);
> > +#endif
> >                   unlock_user_struct(target_st, arg2, 1);
> >               }
> >           }

If that sounds good to you I'll keep your reviewed-by for v2.

Thanks
ChenYu