While the actual commit object format is presently the same, for a
number of reasons we'd like to change it fairly radically. Among
other things, we need to drop our a{sv} types in objects, to protect
against GVariant changing format.
Since now gnome-ostree now longer uses related objects, and nothing
ever used metadata, just drop them both.
If pull is interrupted, we may have downloaded an arbitrary subset of
the requested objects. Previously, we handled this by scanning for
all objects each time.
However, there's an easy optimization - this patch creates a lock file
in the repo. If we don't see that file when starting a pull, we know
we don't need to stat() every file; presence of a dirtree object for
example implies the existence of everything it references.
When multiple threads need to uncompress an object, there was
a race condition where thread A could get EEXIST, unlink,
then thread B calls linkat(), then thread A tries to link() but
fails.
We can just loop in this case.
This seems to work around a likely Linux kernel VFS bug, where I
randomly see ENOENT on link() when we *definitely* called mkdir() at
an earlier point in time.
This is an incompatible change to archive-z, thus it is now renamed to
archive-z2 and ostree will no longer parse archive-z.
I noticed in perf that we were spending some time zlib-decompressing
file headers, which is just inefficient. Rather than do this, keep
the headers uncompressed, and just zlib-compress content.
This gives us something closer to the advantages of archive and
archive-z when using the latter. Concretely we get deduplication
among multiple checkouts, along with the "devino" hash table trick
during commits to avoid checksumming content again.
This is enabled by default.
For similar reasons as metadata, this avoids having the main thread
blocked in fdatasync(), and even better - we can achieve much higher
parallelism if we have multiple threads blocked on fdatasync().
This is where loose content objects are stored as one compressed file,
instead of the two separate ones for regular archive mode. This mode
would be suitable for HTTP servers, beause only one HTTP request is
necessary, and the result would be compressed.
Cleanly separate metadata/content APIs, rather than defaulting to
raw streams. This helps most use cases.
Also, drop support for staging content without knowing the total
length. This complicated the code, and for things like streaming
HTTP, we should be able to figure this out from Content-Length.
They're not a large efficiency win at the moment, because we don't
do any delta compression.
At the moment, they simply served to compress data, but we will change
the archive mode to do that by default.
I run builds on my laptop, but it also crashes about 1/4 of the time
while suspending. It's definitely undesrirable to get e.g. empty
.dirtree objects because they corrupt builds. Concretely, I was
getting empty contents committed for xorg-util-macros.
Now, we used to write out temporary files using g_file_replace() which
does a fsync() during close, but then switched to a more "manual"
g_file_append_to().
We could switch back to g_file_replace(), but the problem is, we don't
want to call fsync() on temporary files in the case where we already
have the object. Attempting to add an object we already have is a
*very* common case.
This is both the old and new code sequence for the case where an
object is already stored:
open(temp, O_WRONLY)
write() write() write()
close()
lstat(objects/3a/9fe332...) = 0
unlink(temp)
In the *new* code, here's the case where an object *isn't* stored:
open(temp, O_WRONLY)
write() write() write()
close()
lstat(objects/3a/9fe332...) = -1
open(temp, O_RDONLY)
fdatasync()
close()
rename(temp, objects/3a/9fe332)
Compare with the *old* code path for when an object isn't stored:
open(temp, O_WRONLY)
write() write() write()
close()
lstat(objects/3a/9fe332...) = -1
link(temp, objects/3a/9fe332)
unlink(temp)
The problem with this is we really need to fdatasync(). Also doing
just rename() instead of the weird link()/unlink() helps us express to
the filesystem that we want atomic semantics. For example, BTRFS has
special handling for rename().
We really don't have a sane story for private files. This is a
defensive step ensuring that with old versions of gnome-ostree,
components that mistakenly have un-world-readable files don't break
pulls.
As the manual page doesn't say, but the in-code kernel documentation
shows, hardlinking for normal users can fail for a variety of
reasons (including very common situations such as non regular file
or non writable file), if the owner of the file does not match
the user linking (e.g. when checking out a shadow repo with a root-
owned master).
If that happens, fail back silently to copying instead of aborting
the whole operation.
https://bugzilla.gnome.org/show_bug.cgi?id=682298
Colin Walters
Giovanni Campagna