On Wed Nov 18, 2026 at 9:14 AM CET, Amir Goldstein wrote: > On Wed, Nov 18, 2026 at 2:09 AM koraynilay wrote: >> >> Hello and thanks for replying, >> >> On Wed Nov 18, 2026 at 12:53 AM CET, Jeff Layton wrote: >> > On Sun, 2026-11-15 at 19:03 +0100, koraynilay wrote: >> >> This is an initial implementation of a new timestamp to indicate the >> >> time when a file/directory was initially created. >> >> >> >> This patch series takes inspiration from the one by Sean Smith >> >> [1], but using >> >> file_setattr()/file_getattr() as suggested by Darrick J. Wong >> >> in [2]. >> >> >> >> The semantics of this new timestamp should be at least these: >> >> - it should be the same as btime on a newly created file >> >> - it's user-settable, so it can be preserved for backups, filesystem >> >> conversions, archives and any similar operation >> >> - it should be preserved every time mtime gets preserved >> >> >> >> and the correspondence with the other timestamps is as follows: >> >> >> >> ctime <-> btime >> >> mtime <-> ptime >> >> >> > >> > I'm not sure how to interpret the correspondence above. >> >> |-----\_-----------------|-----------| >> | \_ change/ | birth/ | >> | \_ modification| creation | >> | \------------|-----------| >> | inode time | ctime | btime | >> |------------|-----------|-----------| >> | file time | mtime | ptime | >> |------------|-----------|-----------| >> >> ("file" as in "the concept of this file", I'd say "file content" but >> that would be incorrect for symlinks) >> >> so ctime is "inode change time", btime is "inode birth/creation time", >> while mtime is "file content change/modification time" and ptime would >> be "file content creation time" (same as Window's or MacOS's creation >> time). >> >> Does this make more sense? >> >> > >> >> Examples: >> >> 1) a .zsh_history file created in 2019, then after a few years the user >> >> decides to switch to another filesystem, copying the .zsh_history to >> >> the new home, in that case btime would have the timestamp of the copy >> >> and ptime would have the 2019 timestamp (just like mtime would have >> >> been preserved) >> >> 2) a symlink created in 2020 from the home directory to a /path going >> >> through the same process as the .zsh_history, in this case it would >> >> also make sense to sync the ptime to the 2020 timestamp instead of >> >> keeping only the btime (so the inode creation time). >> >> >> >> The proof-of-concept code (for testing) for stat(1) and cp(1) is >> >> available at https://github.com/koraynilay/coreutils (master branch) >> >> (Note: these coreutils code changes are a prototype and won't be >> >> submitted upstream as-is, but they work as expected in simple stat and >> >> cp -a/--preserve=ptime tests) >> >> >> >> Lastly, the name "provenance" and "ptime" are not final, I personally >> >> don't have a strong opinion on any name, but while testing with stat(1) >> >> I noticed that "Provenance" is definitely too long for the normal >> >> output, so another option which would fit well is "Origin time", >> >> abbreviated as "ntime" (OrigiN time) and not as "otime", as that would >> >> conflict with btrfs' internal naming and create confusion. >> >> >> >> v2 changes: >> >> - fix permission bypass leading to ptime update on immutable and >> >> append-only files (found by sashiko[3]) >> >> - normalize ptime timestamp as done in fs/attr.c:notify_change() (found >> >> by sashiko[3]) >> >> - fix wording in vfs patch to be like the cover-letter >> >> >> >> v3 changes: >> >> - vfs: reject setting ptime with -EOPNOTSUPP on filesystems that do not >> >> support it instead of silently ignoring it (found by sashiko[4]) >> >> - vfs: fix not being able to clear immutable and append-only on files >> >> with ptime, and allow updating immutable/append-only atomically along >> >> with ptime (found by sashiko[4]) >> >> - btrfs: validate on-disk ptime nanoseconds and reset to otime if >> >> invalid (found by sashiko[4]) >> >> - btrfs: avoid re-assigning ptime on unrelated chattr calls (found by >> >> sashiko[4]) >> >> - add ntfs3 and exfat support as suggested by Amir Goldstein[5], but no >> >> vfat, smb, or nfs4 yet as they are a bit more complex >> >> - move the introduction of fileattr_has_fsx_mask() from the btrfs patch >> >> to a standalone patch in the series >> >> - btrfs: add fileattr_[gs]et to symlink and special file operations >> >> >> >> v4 changes: >> >> - vfs: implement fsx_xflags_mask as suggested by Amir Goldstein[6]; >> >> - add SMB (ksmbd/cifs) and nfs (nfsd/nfs) support as requested in >> >> previous threads; >> >> >> >> Note: >> >> to apply this patch 2 previous patches[7][8] from the ksmbd/cifs >> >> subsystem are needed. >> >> >> >> New Question: >> >> should I rename fileattr_has_fsx_mask() to something like >> >> fileattr_has_unsupported_fsx() and keep its code as-is, or should I add >> >> the fsx_xflags_mask & ~allowed check in it and keep the name as-is? (or >> >> both?) >> >> >> >> Thanks. >> >> >> >> [1]: https://lore.kernel.org/20260405195007.1306-2-DefendTheDisabled@gmail.com >> >> [2]: https://lore.kernel.org/20260407014129.GC6192@frogsfrogsfrogs >> >> [3]: https://sashiko.dev/#/patchset/20261027220532.736352-1-koray.fra%40gmail.com >> >> [4]: https://sashiko.dev/#/patchset/20261028134728.294127-1-koray.fra%40gmail.com >> >> [5]: https://lore.kernel.org/CAOQ4uxiuy0VwZj+3M-zRqQ49WTpHZpiaQ+6MZ8z8WdLjyQM_zA@mail.gmail.com/ >> >> [6]: https://lore.kernel.org/CAOQ4uxjtuTDT63vJdYaq+RHQEQFLTksp66e46OMN99WhJz3SbA@mail.gmail.com/ >> >> [7]: https://lore.kernel.org/linux-cifs/20261112111637.184045-2-koray.fra@gmail.com/ >> >> [8]: https://lore.kernel.org/linux-cifs/20261115155308.415734-1-koray.fra@gmail.com/ >> >> >> >> >> >> koraynilay (10): >> >> vfs: add provenance time (ptime) >> >> vfs: add fileattr_has_fsx_mask() >> >> ntfs3: make crtime user-settable using ptime >> >> exfat: make crtime user-settable using ptime >> >> btrfs: add provenance time (ptime) >> >> vfs: add __vfs_fileattr_set_locked() for inode_lock()'d callers >> >> nfsd: add provenance time (ptime) via FATTR4_TIME_CREATE >> >> nfs: add provenance time (ptime) via FATTR4_TIME_CREATE >> >> ksmbd: add provenance time (ptime) and map it to CreationTime >> >> cifs: add provenance time (ptime) and map it to SMB's CreationTime >> >> >> >> fs/btrfs/btrfs_inode.h | 4 + >> >> fs/btrfs/delayed-inode.c | 6 ++ >> >> fs/btrfs/inode.c | 29 +++++- >> >> fs/btrfs/ioctl.c | 24 ++++- >> >> fs/btrfs/print-tree.c | 1 + >> >> fs/btrfs/tree-log.c | 3 + >> >> fs/exfat/exfat_fs.h | 2 + >> >> fs/exfat/file.c | 42 ++++++++- >> >> fs/exfat/namei.c | 1 + >> >> fs/file_attr.c | 156 ++++++++++++++++++++++++++------ >> >> fs/nfs/inode.c | 27 +++++- >> >> fs/nfs/internal.h | 2 + >> >> fs/nfs/nfs4proc.c | 56 +++++++++++- >> >> fs/nfs/nfs4xdr.c | 19 ++-- >> >> fs/nfs/symlink.c | 1 + >> >> fs/nfsd/nfs4proc.c | 9 ++ >> >> fs/nfsd/nfs4xdr.c | 40 +++++--- >> >> fs/nfsd/vfs.c | 17 ++++ >> >> fs/nfsd/vfs.h | 4 +- >> >> fs/nfsd/xdr4.h | 3 + >> >> fs/ntfs3/file.c | 20 +++- >> >> fs/ntfs3/frecord.c | 8 +- >> >> fs/smb/client/cifsfs.c | 8 ++ >> >> fs/smb/client/inode.c | 7 +- >> >> fs/smb/client/ioctl.c | 109 ++++++++++++++++------ >> >> fs/smb/server/smb2pdu.c | 37 +++++--- >> >> fs/smb/server/vfs.c | 43 ++++++++- >> >> fs/smb/server/vfs.h | 1 + >> >> fs/stat.c | 2 + >> >> include/linux/fileattr.h | 29 +++++- >> >> include/linux/nfs_xdr.h | 3 + >> >> include/linux/stat.h | 1 + >> >> include/uapi/linux/btrfs_tree.h | 3 +- >> >> include/uapi/linux/fs.h | 9 +- >> >> include/uapi/linux/stat.h | 7 +- >> >> 35 files changed, 614 insertions(+), 119 deletions(-) >> >> >> >> >> >> base-commit: 3bc1fb9c3b20cd478c83496ab16addecb401770d >> >> prerequisite-patch-id: 049ae0eba275b5be35a238833107e0b251a55aa2 >> >> prerequisite-patch-id: 614f0bc9411391403d4e9702bafeb31b0ec04ec1 >> >> prerequisite-patch-id: ddf636b9d33d82ea206ce265d817371350bf8858 >> > >> > I think this is going to be very hard to do properly, because the >> > kernel can't really manage this on its own. Something in userland is >> > going to have to set this value if you want to update it to something >> > other than the original inode instantiation time (aka btime). >> > >> > It really sounds like what you want is the ability to make the btime >> > settable. Personally I think the decision to forbid that was wrong- >> > headed and that it should have been settable from userland from the >> > beginning. Maybe we should just revisit that decision. >> > > Frankly, I don't think this decision was made at all. > > There are only 2 types of filesystems in Linux today: > 1. btime is constant (xfs, etx4, btrfs,...) > 2. crtime is settable in principle but not via a standardized Linux > API (ntfs, cifs, ...) > > AFAIK, btrfs in this series is the only filesystem which is going to > have two deviated copies of btime and ptime. > For all the rest, the deviation is artificial in statx. Well, my objective with this series would be having ptime just as supported as a/c/m/btime, especially from userspace in the 90% of setups, > So the questions to ask are: > 1. Do btrfs developers prefer to store a new settable ptime or > making otime/btime settable with this new API? > 2. Will any other fs in the future want to store settable ptime along > side the legacy immutable btime? > > My take on these question is: > - Personally I would love to see ptime support in xfs in the future and would > willing to contribute to this work, because I have hard evidence from the > performance benefits that it brings to smb server workloads > - Judging by past experience with i_version bump interpretation on xfs, > I find it hard to believe that xfs developer would agree to making the > legacy btime settable so I guess for v5 or v6 I'll also draft a patch for at least ext4 and xfs. > Hence the distinction in statx API is needed, but for most filesystems > it is a pure semantic way of saying "I do/don't have settable birth time". > >> Originally I just wanted to have btime settable, yes, and if btime was >> settable I wouldn't have made this RFC PATCH either, but after reading >> the already existing discussions about it, I can understand having an >> "immutable inode creation time", like ctime exists, so even after making >> btime settable, a new timestamp would be necessary anyway (e.g. for the >> people who want btime to exist as a forensics timestamp). >> > > Right, but every filesystem (e.g. btrfs) can still make an independent > choice whether storing one or two copies is desired. > The API does not limit those choices. > >> > Failing that, why not just track this in an xattr of some sort, since >> > this is really just an opaque blob once you allow it to be settable. If >> > the xattr isn't present, then you can just use the value of the btime >> > as-is. FWIW, I think samba does this today, so you could try to sync >> > with its behavior here. >> >> (This is copied and adapted from [2]) >> >> I actually have been using fully userspace xattrs[1] for a few months, >> mostly for backups, and it "works"; but last month I wanted to merge 2 >> FSs (and have another merge I have to do), so I ran [1] on all files and >> then compared all metadata/data between source and destination and >> noticed there were quite a lot of files that didn't have the xattr set. >> So I looked online and learned that xattrs only work on regular >> files/directories, not on symlinks or anything else (and the missing >> ones were of symlinks). >> This is what prompted me to search more if anyone had already proposed a >> new timestamp (instead of making btime settable) and found the email >> thread from april linked in the cover letter, which gave me the excuse >> to finally work on a proper patch. >> > > I am tired of repeating my arguments (and finding my own quotes) > about why using xattrs is a poor replacement for in-inode ptime. > AI agents can easily find those quotes for those who are interested > so I won't bother, I just say that smbd already uses xattr to store > the creation time - it's just bad for performance and disk usage in > metadata intensive workloads. (I really need to remember to add the xattr explanation in the cover letter, it's been 3 version since I wanted to but forgot lol) >> > That seems a lot simpler than adding an entirely new inode timestamp >> > that is always going to be a copy of the btime, except in rare cases. >> > Having to patch all of the on-disk filesystems with a new per-inode >> > field is a lot more work. >> >> > This patch is quite simple, but changing btrfs, then ext4, xfs, etc is a >> > mess, and inevitably you can't get these timestamps until you upgrade your >> > filesystems. >> > > Jeff, this is standard "how to make progress". > I completely don't agree with this argument. > We cannot use this argument to block progress. > Is it already a long standing practice for userspace file servers (samba) > to probe the host filesystem for specific features and fall back to > alternative implementation if they are not supported. > > There will be nothing new in this case, that's the only way to > make forward progress - define the API, make support queryable, > implement the feature in some filesystems, implement consumers, repeat. > > Also for nfsd, that's the only way to ever implement settable > FATTR4_TIME_CREATE which is the protocol allows. > This is never going to be possible with custom xattrs. > So I must say that I don't understand the objection from you and Chuck Thanks, koraynilay > Thanks, > Amir.