Add a high-level document that describes how multigrain timestamps work, rationale for them, and some info about implementation and tradeoffs. Signed-off-by: Jeff Layton <jlayton@xxxxxxxxxx> --- Documentation/filesystems/multigrain-ts.rst | 126 ++++++++++++++++++++++++++++ 1 file changed, 126 insertions(+) diff --git a/Documentation/filesystems/multigrain-ts.rst b/Documentation/filesystems/multigrain-ts.rst new file mode 100644 index 000000000000..beef7f79108c --- /dev/null +++ b/Documentation/filesystems/multigrain-ts.rst @@ -0,0 +1,126 @@ +.. SPDX-License-Identifier: GPL-2.0 + +===================== +Multigrain Timestamps +===================== + +Introduction +============ +Historically, the kernel has always used a coarse time values to stamp +inodes. This value is updated on every jiffy, so any change that happens +within that jiffy will end up with the same timestamp. + +When the kernel goes to stamp an inode (due to a read or write), it first gets +the current time and then compares it to the existing timestamp(s) to see +whether anything will change. If nothing changed, then it can avoid updating +the inode's metadata. + +Coarse timestamps are therefore good from a performance standpoint, since they +reduce the need for metadata updates, but bad from the standpoint of +determining whether anything has changed, since a lot of things can happen in a +jiffy. + +They are particularly troublesome with NFSv3, where unchanging timestamps can +make it difficult to tell whether to invalidate caches. NFSv4 provides a +dedicated change attribute that should always show a visible change, but not +all filesystems implement this properly, and many just populating this with +the ctime. + +Multigrain timestamps aim to remedy this by selectively using fine-grained +timestamps when a file has had its timestamps queried recently, and the current +coarse-grained time does not cause a change. + +Inode Timestamps +================ +There are currently 3 timestamps in the inode that are updated to the current +wallclock time on different activity: + +ctime: + The inode change time. This is stamped with the current time whenever + the inode's metadata is changed. Note that this value is not settable + from userland. + +mtime: + The inode modification time. This is stamped with the current time + any time a file's contents change. + +atime: + The inode access time. This is stamped whenever an inode's contents are + read. Widely considered to be a terrible mistake. Usually avoided with + options like noatime or relatime. + +Updating the mtime always implies a change to the ctime, but updating the +atime due to a read request does not. + +Multigrain timestamps are only tracked for the ctime and the mtime. atimes are +not affected and always use the coarse-grained value (subject to the floor). + +Inode Timestamp Ordering +======================== + +In addition just providing info about changes to individual files, file +timestamps also serve an important purpose in applications like "make". These +programs measure timestamps in order to determine whether source files might be +newer than cached objects. + +Userland applications like make can only determine ordering based on +operational boundaries. For a syscall those are the syscall entry and exit +points. For io_uring or nfsd operations, that's the request submission and +response. In the case of concurrent operations, userland can make no +determination about the order in which things will occur. + +For instance, if a single thread modifies one file, and then another file in +sequence, the second file must show an equal or later mtime than the first. The +same is true if two threads are issuing similar operations that do not overlap +in time. + +If however, two threads have racing syscalls that overlap in time, then there +is no such guarantee, and the second file may appear to have been modified +before, after or at the same time as the first, regardless of which one was +submitted first. + +Multigrain Timestamps +===================== +Multigrain timestamps are aimed at ensuring that changes to a single file are +always recognizeable, without violating the ordering guarantees when multiple +different files are modified. This affects the mtime and the ctime, but the +atime will always use coarse-grained timestamps. + +It uses the lowest-order bit in the timestamp as a flag that indicates whether +the mtime or ctime have been queried. If either or both have, then the kernel +takes special care to ensure the next timestamp update will display a visible +change. This ensures tight cache coherency for use-cases like NFS, without +sacrificing the benefits of reduced metadata updates when files aren't being +watched. + +The ctime Floor Value +===================== +It's not sufficient to simply use fine or coarse-grained timestamps based on +whether the mtime or ctime has been queried. A file could get a fine grained +timestamp, and then a second file modified later could get a coarse-grained one +that appears earlier than the first, which would break the kernel's timestamp +ordering guarantees. + +To mitigate this problem, we maintain a per-time_namespace floor value that +ensures that this can't happen. The two files in the above example may appear +to have been modified at the same time in such a case, but they will never show +the reverse order. + +Implementation Notes +==================== +Multigrain timestamps are intended for use by local filesystems that get +ctime values from the local clock. This is in contrast to network filesystems +and the like that just mirror timestamp values from a server. + +For most filesystems, it's sufficient to just set the FS_MGTIME flag in the +fstype->fs_flags in order to opt-in, providing the ctime is only ever set via +inode_set_ctime_current(). If the filesystem has a ->getattr routine that +doesn't call generic_fillattr, then you should have it call fill_mg_cmtime to +fill those values. + +Caveats +======= +The main sacrifice is the lowest bit in the ctime's field, since that's +where the flag is stored. Thus, timestamps presented by multigrain enabled +filesystems will always have an even tv_nsec value (since the lowest bit +is masked off). -- 2.45.2