Re: [PATCH v4 0/2] ls: use statx() when it's available
On 31/10/2019 11:04, Jeff Layton wrote: On Thu, 2019-10-24 at 10:16 +0100, Pádraig Brady wrote: On 09/10/2019 22:23, Pádraig Brady wrote: On 09/10/2019 11:14, Jeff Layton wrote: On Wed, 2019-10-09 at 10:19 +0100, Pádraig Brady wrote: On 19/09/19 16:59, Jeff Layton wrote: v4: - set appropriate STATX_* bits for time_type, sort_type and print_block_size v3: - syntax cleanups. make syntax-check now passes v2: - add wrappers for stat_for_ino and fstat_for_ino, don't factor out loop detection - style cleanups Sorry for the delay in reviewing. This looks good, except for the usage of AT_STATX_DONT_SYNC when retrieving inode info. Sure that generally doesn't change, but that would be file system dependent, and I have seen file systems that populate inode with counters etc. Anyway that sort of decision would be best done in the kernel I think, where it would have the info whether it needs to sync for STATX_INO or not. OK for me to push without the DONT_SYNC ? Sure, that seems reasonable. Let me know if you need me to resend. Pushed without AT_STATX_DONT_SYNC. Also added the new statx.h to noinst_HEADERS, and used our _GL_ATTRIBUTE_PURE define rather than the less portable __attribute__. https://git.sv.gnu.org/gitweb/?p=coreutils.git;a=commitdiff;h=a99ab26 FYI I see this is causing issues with docker images (possibly due to seccomp). https://bugzilla.redhat.com/show_bug.cgi?id=1760300 (There is a bug tracking the kernel issue but it's not publicly accessible.) From scanning the comments it seems that once statx() is called state is messed up, so that having ls fall back from statx() to stat() wouldn't help in this case. So the assumption that whatever libc providing statx() does appropriate fallback to stat() is probably still valid. Yep, this seems like it's probably a seccomp bug. We were going to put these patches into Fedora 31's coreutils, but reverted the patches until this is resolved. Hopefully they'll nail down the problem soob. Jeff I don't have access to the kernel bug tracking this issue: https://bugzilla.redhat.com/show_bug.cgi?id=1762578 Could you give a summary of the current state? Am I correct in thinking that if statx() returns 2, then it would be futile to notice that and fall back to stat() as state is messed up causing future stat() calls to fail? thanks, Pádraig
Re: [PATCH v4 0/2] ls: use statx() when it's available
On Thu, 2019-10-24 at 10:16 +0100, Pádraig Brady wrote: > On 09/10/2019 22:23, Pádraig Brady wrote: > > On 09/10/2019 11:14, Jeff Layton wrote: > > > On Wed, 2019-10-09 at 10:19 +0100, Pádraig Brady wrote: > > > > On 19/09/19 16:59, Jeff Layton wrote: > > > > > v4: > > > > > - set appropriate STATX_* bits for time_type, sort_type and > > > > > print_block_size > > > > > > > > > > v3: > > > > > - syntax cleanups. make syntax-check now passes > > > > > > > > > > v2: > > > > > - add wrappers for stat_for_ino and fstat_for_ino, don't factor out > > > > > loop > > > > > detection > > > > > - style cleanups > > > > > > > > Sorry for the delay in reviewing. > > > > > > > > This looks good, except for the usage > > > > of AT_STATX_DONT_SYNC when retrieving inode info. > > > > Sure that generally doesn't change, but that > > > > would be file system dependent, and I have seen > > > > file systems that populate inode with counters etc. > > > > Anyway that sort of decision would be best done > > > > in the kernel I think, where it would have the > > > > info whether it needs to sync for STATX_INO or not. > > > > > > > > OK for me to push without the DONT_SYNC ? > > > > > > > > > > Sure, that seems reasonable. Let me know if you need me to resend. > > > > Pushed without AT_STATX_DONT_SYNC. > > Also added the new statx.h to noinst_HEADERS, > > and used our _GL_ATTRIBUTE_PURE define rather than > > the less portable __attribute__. > > > > https://git.sv.gnu.org/gitweb/?p=coreutils.git;a=commitdiff;h=a99ab26 > > FYI I see this is causing issues with docker images (possibly due to seccomp). > https://bugzilla.redhat.com/show_bug.cgi?id=1760300 > (There is a bug tracking the kernel issue but it's not publicly accessible.) > > From scanning the comments it seems that once statx() is called > state is messed up, so that having ls fall back from statx() to stat() > wouldn't help in this case. So the assumption that whatever libc providing > statx() does appropriate fallback to stat() is probably still valid. > Yep, this seems like it's probably a seccomp bug. We were going to put these patches into Fedora 31's coreutils, but reverted the patches until this is resolved. Hopefully they'll nail down the problem soob. -- Jeff Layton
Re: [PATCH v4 0/2] ls: use statx() when it's available
On 09/10/2019 22:23, Pádraig Brady wrote: On 09/10/2019 11:14, Jeff Layton wrote: On Wed, 2019-10-09 at 10:19 +0100, Pádraig Brady wrote: On 19/09/19 16:59, Jeff Layton wrote: v4: - set appropriate STATX_* bits for time_type, sort_type and print_block_size v3: - syntax cleanups. make syntax-check now passes v2: - add wrappers for stat_for_ino and fstat_for_ino, don't factor out loop detection - style cleanups Sorry for the delay in reviewing. This looks good, except for the usage of AT_STATX_DONT_SYNC when retrieving inode info. Sure that generally doesn't change, but that would be file system dependent, and I have seen file systems that populate inode with counters etc. Anyway that sort of decision would be best done in the kernel I think, where it would have the info whether it needs to sync for STATX_INO or not. OK for me to push without the DONT_SYNC ? Sure, that seems reasonable. Let me know if you need me to resend. Pushed without AT_STATX_DONT_SYNC. Also added the new statx.h to noinst_HEADERS, and used our _GL_ATTRIBUTE_PURE define rather than the less portable __attribute__. https://git.sv.gnu.org/gitweb/?p=coreutils.git;a=commitdiff;h=a99ab26 FYI I see this is causing issues with docker images (possibly due to seccomp). https://bugzilla.redhat.com/show_bug.cgi?id=1760300 (There is a bug tracking the kernel issue but it's not publicly accessible.) From scanning the comments it seems that once statx() is called state is messed up, so that having ls fall back from statx() to stat() wouldn't help in this case. So the assumption that whatever libc providing statx() does appropriate fallback to stat() is probably still valid. cheers, Pádraig
Re: [PATCH v4 0/2] ls: use statx() when it's available
On 09/10/2019 11:14, Jeff Layton wrote: On Wed, 2019-10-09 at 10:19 +0100, Pádraig Brady wrote: On 19/09/19 16:59, Jeff Layton wrote: v4: - set appropriate STATX_* bits for time_type, sort_type and print_block_size v3: - syntax cleanups. make syntax-check now passes v2: - add wrappers for stat_for_ino and fstat_for_ino, don't factor out loop detection - style cleanups Sorry for the delay in reviewing. This looks good, except for the usage of AT_STATX_DONT_SYNC when retrieving inode info. Sure that generally doesn't change, but that would be file system dependent, and I have seen file systems that populate inode with counters etc. Anyway that sort of decision would be best done in the kernel I think, where it would have the info whether it needs to sync for STATX_INO or not. OK for me to push without the DONT_SYNC ? Sure, that seems reasonable. Let me know if you need me to resend. Pushed without AT_STATX_DONT_SYNC. Also added the new statx.h to noinst_HEADERS, and used our _GL_ATTRIBUTE_PURE define rather than the less portable __attribute__. https://git.sv.gnu.org/gitweb/?p=coreutils.git;a=commitdiff;h=a99ab26 thanks, Pádraig
Re: [PATCH v4 0/2] ls: use statx() when it's available
On Wed, 2019-10-09 at 10:19 +0100, Pádraig Brady wrote: > On 19/09/19 16:59, Jeff Layton wrote: > > v4: > > - set appropriate STATX_* bits for time_type, sort_type and > > print_block_size > > > > v3: > > - syntax cleanups. make syntax-check now passes > > > > v2: > > - add wrappers for stat_for_ino and fstat_for_ino, don't factor out loop > > detection > > - style cleanups > > Sorry for the delay in reviewing. > > This looks good, except for the usage > of AT_STATX_DONT_SYNC when retrieving inode info. > Sure that generally doesn't change, but that > would be file system dependent, and I have seen > file systems that populate inode with counters etc. > Anyway that sort of decision would be best done > in the kernel I think, where it would have the > info whether it needs to sync for STATX_INO or not. > > OK for me to push without the DONT_SYNC ? > Sure, that seems reasonable. Let me know if you need me to resend. Thanks, -- Jeff Layton
Re: [PATCH v4 0/2] ls: use statx() when it's available
On 19/09/19 16:59, Jeff Layton wrote: v4: - set appropriate STATX_* bits for time_type, sort_type and print_block_size v3: - syntax cleanups. make syntax-check now passes v2: - add wrappers for stat_for_ino and fstat_for_ino, don't factor out loop detection - style cleanups Sorry for the delay in reviewing. This looks good, except for the usage of AT_STATX_DONT_SYNC when retrieving inode info. Sure that generally doesn't change, but that would be file system dependent, and I have seen file systems that populate inode with counters etc. Anyway that sort of decision would be best done in the kernel I think, where it would have the info whether it needs to sync for STATX_INO or not. OK for me to push without the DONT_SYNC ? thanks, Pádraig
Re: [PATCH v4 0/2] ls: use statx() when it's available
On Fri, 2019-09-20 at 00:53 +0100, Pádraig Brady wrote: > This looks good upon quick review. > I will try to merge over the next couple of days. > > thanks! > Pádraig Ping? Any idea when you'll get a chance to review and merge this? Thanks, -- Jeff Layton
Re: [PATCH v4 0/2] ls: use statx() when it's available
This looks good upon quick review. I will try to merge over the next couple of days. thanks! Pádraig
