Re: [PATCH v4 2/6] core.fsyncobjectfiles: batched disk flushes

On Tue, Sep 21 2021, Neeraj Singh wrote:

On Tue, Sep 21, 2021 at 4:41 PM Ævar Arnfjörð Bjarmason
[off-list ref] wrote:

quoted

On Mon, Sep 20 2021, Neeraj Singh via GitGitGadget wrote:

quoted

When the new mode is enabled we do the following for new objects:

1. Create a tmp_obj_XXXX file and write the object data to it.
2. Issue a pagecache writeback request and wait for it to complete.
3. Record the tmp name and the final name in the bulk-checkin state for
   later rename.

At the end of the entire transaction we:
1. Issue a fsync against the lock file to flush the hardware writeback
   cache, which should by now have processed the tmp file writes.
2. Rename all of the temp files to their final names.
3. When updating the index and/or refs, we assume that Git will issue
   another fsync internal to that operation.

Perhaps note too that:

4. For loose objects, refs etc. we may or may not create directories,
   and most certainly will be updating metadata on the immediate
   directory containing the file, but none of that's fsync()'d.

quoted

On a filesystem with a singular journal that is updated during name
operations (e.g. create, link, rename, etc), such as NTFS and HFS+, we
would expect the fsync to trigger a journal writeout so that this
sequence is enough to ensure that the user's data is durable by the time
the git command returns.

This change also updates the macOS code to trigger a real hardware flush
via fnctl(fd, F_FULLFSYNC) when fsync_or_die is called. Previously, on
macOS there was no guarantee of durability since a simple fsync(2) call
does not flush any hardware caches.

There's no discussion of whether this is or isn't known to also work
some Linux FS's, and for these OS's where this does work is this only
for the object files themselves, or does metadata also "ride along"?

I unfortunately can't examine Linux kernel source code and the details
of metadata
consistency behavior across files is not something that anyone in that
group wants
to pin down. As far as I can tell, the only thing that's really
guaranteed is fsyncing
every single file you write down and its parent directory if you're
creating a new file
(which we always are).  As came up in conversation with Christoph
Hellwig elsewhere
on thread, Linux doesn't have any set of syscalls to make batch mode
safe.  It does look
like XFS would be safe if sync_file_ranges actually promised to wait
for all pagecache
writeback definitively, since it would do a "log force" to push all
the dirty metadata to
disk when we do our final fsync.

I really didn't want to say something definitive about what Linux can
or will do, since I'm
not in a position to really know or influence them.  Christoph did say
that he would be
interested in contributing a variant to this patch that would be
definitively safe on filesystems
that honor syncfs.

*nod*, it's fine if it's omitted. Just wondering if we knew but weren't
 saying etc.

quoted

quoted

_Performance numbers_:

Linux - Hyper-V VM running Kernel 5.11 (Ubuntu 20.04) on a fast SSD.
Mac - macOS 11.5.1 running on a Mac mini on a 1TB Apple SSD.
Windows - Same host as Linux, a preview version of Windows 11.
        This number is from a patch later in the series.

Adding 500 files to the repo with 'git add' Times reported in seconds.

core.fsyncObjectFiles | Linux | Mac   | Windows
----------------------|-------|-------|--------
                false | 0.06  |  0.35 | 0.61
                true  | 1.88  | 11.18 | 2.47
                batch | 0.15  |  0.41 | 1.53

Per my https://lore.kernel.org/git/87mtp5cwpn.fsf@evledraar.gmail.com (local)
and 6/6 in this series we've got perf tests for add/stash, but it would
be really interesting to see how this is impacted by
transfer.unpackLimit in cases where we may be writing packs or loose
objects.

I'm having trouble understanding how unpackLimit is related to 'git stash'
or 'git add'. From code inspection, it doesn't look like we're using
those settings
for adding objects except from across a transport.

Are you proposing that we have a similar setting for adding objects
via 'add' using
a packfile?  I think that would be a good goal, but it might be a bit
tricky since we've
likely done a lot of the work to buffer the input objects in order to
compute their OIDs,
before we know how many objects there are to add. If the policy were
to "always add to
a packfile", it would be easier.

No, just that in the documentation that we should be explaining to the
reader that this mode that optimizes for loose object writing benefits
particular commands, but e.g. on the server-side that we'll probably
never write 500 objects, but stream them to one pack.

Which might also inform next steps for the commands this does help with,
i.e. can we make more things stream to packs? I think having this mode
is at worst a good transitory thing to have, but perhaps longer term
we'll want to simply write fewer individual loose objects.

In any case, pushing to a server with this configured and scaling that
by transfer.unpackLimit should nicely demonstrate the pack v.s. loose
object scenario at different fsck-settings.

quoted

quoted

[...]
 core.fsyncObjectFiles::
-     This boolean will enable 'fsync()' when writing object files.
-+
-This is a total waste of time and effort on a filesystem that orders
-data writes properly, but can be useful for filesystems that do not use
-journalling (traditional UNIX filesystems) or that only journal metadata
-and not file contents (OS X's HFS+, or Linux ext3 with "data=writeback").
+     A value indicating the level of effort Git will expend in
+     trying to make objects added to the repo durable in the event
+     of an unclean system shutdown. This setting currently only
+     controls the object store, so updates to any refs or the
+     index may not be equally durable.

All these mentions of "object" should really clarify that it's "loose
objects", i.e. we always fsync pack files.

quoted

+* `false` allows data to remain in file system caches according to
+  operating system policy, whence it may be lost if the system loses power
+  or crashes.

As noted in point #4 of
https://lore.kernel.org/git/87mtp5cwpn.fsf@evledraar.gmail.com/ (local) while
this direction is overall an improvement over the previously flippant
docs, they at least alluded to the context that the assumption behind
"false" is that you don't really care about loose objects, you care
about loose objects *and* the ref update or whatever.

As I think (this is from memory) we've covered already this may have
been all based on some old ext3 assumption, but it's probably worth
summarizing that here, i.e. if you've got an FS with global ordered
operations you can probably skip this, but probably not etc.

quoted

+* `true` triggers a data integrity flush for each object added to the
+  object store. This is the safest setting that is likely to ensure durability
+  across all operating systems and file systems that honor the 'fsync' system
+  call. However, this setting comes with a significant performance cost on
+  common hardware.

This is really overpromising things by omitting the fact that eve if
we're getting this feature you've hacked up right, we're still not
fsyncing dir entries etc (also noted above).

So something that describes the narrow scope here, along with "loose
objects" etc....

quoted

+* `batch` enables an experimental mode that uses interfaces available in some
+  operating systems to write object data with a minimal set of FLUSH CACHE
+  (or equivalent) commands sent to the storage controller. If the operating
+  system interfaces are not available, this mode behaves the same as `true`.
+  This mode is expected to be safe on macOS for repos stored on HFS+ or APFS
+  filesystems and on Windows for repos stored on NTFS or ReFS.

Again, even if it's called "core.fsyncObjectFiles" if we're going to say
"safe" we really need to say safe in what sense. Having written and
fsync()'d the file is helping nobody if the metadata never arrives....

My concern with your feedback here is that this is user-facing documentation.
I'd assume that people who are not intimately familiar with both their
filesystem
and Git's internals would just be completely mystified by a long commentary on
the specifics in the Config documentation. I think over time Git should focus on
making this setting really guarantee durability in a meaningful way
across the entire
repository.

Yeah, this setting though is probably going to be tweaked only by fairly
expert-level users of git.

I think it's fine if it just explicitly punts and says something like
'this is what it does, this may or may not work on your FS' etc., my
main issue with the current docs is that they give off this vibe of
knowing a lot more than they're telling you.

quoted

quoted

+static void do_sync_and_rename(struct string_list *fsync_state, struct lock_file *lock_file)
+{
+     if (fsync_state->nr) {

I think less indentation here would be nice:

    if (!fsync_state->nr)
        return;
    /* rest of unindented body */

Will fix.

quoted

Or better yet do this check in unplug_bulk_checkin(), then here:

    fsync_or_die();
    for_each_string_list_item() { ...}
    string_list_clear(....);

I'd prefer to put it in the callee for reasons of
separation-of-concerns.  I don't want
to have the caller and callee partially implement the contract. The
compiler should
do a good enough job, since it's only one caller and will probably get
totally inilined.

*nod*

For what it's worth I meant the "inlined" just in terms of avoiding the
indirection for human readers, it won't matter to the machine,
especially since this is all I/O bound...