On 6/28/22 03:46, Amir Goldstein wrote:
[+linux-api]
On Tue, Jun 28, 2022 at 1:58 AM James Yonan <james@xxxxxxxxxxx> wrote:
RENAME_NEWER is a new userspace-visible flag for renameat2(), and
stands alongside existing flags such as RENAME_NOREPLACE,
RENAME_EXCHANGE, and RENAME_WHITEOUT.
RENAME_NEWER is a conditional variation on RENAME_NOREPLACE, and
indicates that if the target of the rename exists, the rename will
only succeed if the source file is newer than the target (i.e. source
mtime > target mtime). Otherwise, the rename will fail with -EEXIST
instead of replacing the target. When the target doesn't exist,
RENAME_NEWER does a plain rename like RENAME_NOREPLACE.
RENAME_NEWER is very useful in distributed systems that mirror a
directory structure, or use a directory as a key/value store, and need
to guarantee that files will only be overwritten by newer files, and
that all updates are atomic.
This feature sounds very cool.
For adding a new API it is always useful if you bring forward a userland
tool (rsync?) that intend to use it, preferably with a POC patch.
A concrete prospective user is always better than a hypothetical one.
Some people hold the opinion that only new APIs with real prospective
users should be merged.
Not sure that rsync would be the canonical user, though it might be a
reasonable POC. The problem that we are solving is essentially
near-real-time directory mirroring or replication of one source
directory to many target directories on follower nodes. Many writers,
many readers, filesystem-based, strong guarantees of eventual
convergence, infinitely scalable. You have a source directory that
could be an AWS S3 bucket. You have potentially thousands of follower
nodes that want to replicate the source directory on a local
filesystem. You have messages flying around the network (Kafka, AWS
SQS, etc.) representing file updates. These messages might be reordered
or duplicated but each contains the file content and a unique
nanosecond-scale timestamp. Because the file update throughput can be
in the thousands of files per second, you might have multiple threads on
each node, receiving updates, and moving them into the target
directory. The only way to guarantee that the state of the mirror
target directory on all nodes converges to the state of the source
directory is to have the last-step move operation be atomic and
conditional to the modification time (so that an earlier version of the
file doesn't overwrite a later version).
So I understand that there needs to be a strong case to extend the Linux
API, and I think the argument is that conditional atomic file operations
enable entirely new classes of applications. The fact that this can be
facilitated by essentially 5 lines of kernel code is remarkable.
While this patch may appear large at first glance, most of the changes
deal with renameat2() flags validation, and the core logic is only
5 lines in the do_renameat2() function in fs/namei.c:
if ((flags & RENAME_NEWER)
&& d_is_positive(new_dentry)
&& timespec64_compare(&d_backing_inode(old_dentry)->i_mtime,
&d_backing_inode(new_dentry)->i_mtime) <= 0)
goto exit5;
I have a few questions:
- Why mtime?
- Why not ctime?
- Shouldn't a feature like that protect from overwriting metadata changes
to the destination file?
In any case, I would be much more comfortable with comparing ctime
because when it comes to user settable times, what if rsync *wants*
to update the destination file's mtime to an earlier time that was set in
the rsync source?
If comparing ctime does not fit your use case and you can convince
the community that comparing mtime is a justified use case, we would
need to use a flag name to reflect that like RENAME_NEWER_MTIME
so we don't block future use case of RENAME_NEWER_CTIME.
I hope that we can agree that ctime is enough and that mtime will not
make sense so we can settle with RENAME_NEWER that means ctime.
So I actually think that mtime is the better timestamp to use because
ctime is modified by the rename operation itself, while mtime measures
the last modification time of the file content, which is what we care about.
It's pretty cool in a way that a new atomic file operation can even be
implemented in just 5 lines of code, and it's thanks to the existing
locking infrastructure around file rename/move that these operations
become almost trivial. Unfortunately, every fs must approve a new
renameat2() flag, so it bloats the patch a bit.
So one question to ask is could this functionality be implemented
in userspace without adding a new renameat2() flag? I think you
could attempt it with iterative RENAME_EXCHANGE, but it's hackish,
inefficient, and not atomic, because races could cause temporary
mtime backtracks. How about using file locking? Probably not,
because the problem we want to solve is maintaining file/directory
atomicity for readers by creating files out-of-directory, setting
their mtime, and atomically moving them into place. The strategy
to lock such an operation really requires more complex locking methods
than are generally exposed to userspace. And if you are using inotify
on the directory to notify readers of changes, it certainly makes
sense to reduce unnecessary churn by preventing a move operation
based on the mtime check.
While some people might question the utility of adding features to
filesystems to make them more like databases, there is real value
in the performance, atomicity, consistent VFS interface, multi-thread
safety, and async-notify capabilities of modern filesystems that
starts to blur the line, and actually make filesystem-based key-value
stores a win for many applications.
Like RENAME_NOREPLACE, the RENAME_NEWER implementation lives in
the VFS, however the individual fs implementations do strict flags
checking and will return -EINVAL for any flag they don't recognize.
For this reason, my general approach with flags is to accept
RENAME_NEWER wherever RENAME_NOREPLACE is also accepted, since
RENAME_NEWER is simply a conditional variant of RENAME_NOREPLACE.
You are not taking into account that mtime may be cached attribute that is not
uptodate in network filesystems (fuse too) so behavior can be a bit
unpredictable,
unless filesystem code compares also the cache coherency of the attributes
on both files and even then, without extending the network protocol this is
questionable behavior for client side.
So I think your filter of which filesystems to enable is way too wide.
So I'm new to the filesystem code, but my reading of do_renameat2() in
fs/namei.c seems to indicate that if d_is_positive(dentry) is true, it's
safe to access the inode struct via d_backing_inode(dentry) and
dereference mtime. But what you're saying is that the mtime value
cached in the inode struct might not be up-to-date for network filesystems?
How many of them did you test, I'll take a wild guess that not all of them.
Please do not enable RENAME_NEWER is any filesystem that you did
not test or that was not tested by some other fs developer using the
tests that you write.
So I'm mostly interested in implementing this on local filesystems,
because the application layer has already done the heavy lifting on the
networking side so that the filesystem layer can be local, fast, and
atomic. So yes, I haven't tested this yet on networked filesystems.
But I'm thinking that because all functionality is implemented at the
VFS layer, it should be portable to any fs that also supports
RENAME_NOREPLACE, with the caveat that it depends on the ability of the
VFS to get a current and accurate mtime attribute inside the critical
section between lock_rename() and unlock_rename().
I noticed only one fs driver (cifs) that treated RENAME_NOREPLACE
in a non-generic way, because no-replace is the natural behavior
for CIFS, and it therefore considers RENAME_NOREPLACE as a hint that
no replacement can occur. Aside from this special case, it seems
safe to assume that any fs that supports RENAME_NOREPLACE should
also be able to support RENAME_NEWER out of the box.
I did not notice a general self-test for renameat2() at the VFS
layer (outside of fs-specific tests), so I created one, though
at the moment it only exercises RENAME_NEWER. Build and run with:
make -C tools/testing/selftests TARGETS=renameat2 run_tests
Please make sure to also add test coverage in fstests and/or LTP.
See fstest generic/024 for RENAME_NOREPLACE and LTP test
renameat201 as examples.
renameat201 can and should be converted to newlib and run with
.all_filesystems = 1.
This is the easiest and fastest way for you to verify your patch
on the common fs that are installed on your system.
LTP maintainers can help you with that.
Thanks, that's good to know. And thanks for taking the time to write up
such a detailed response.
James
