quickfetch() calls rev-list to check whether the objects we are about to
fetch are already present in the repo (if so, we can skip the object fetch).
However, when there are many (~1000) refs to be fetched, the rev-list
command line grows larger than the maximum command line size on some systems
(32K in Windows). This causes rev-list to fail, making quickfetch() return
non-zero, which unnecessarily triggers the transport machinery. This somehow
causes fetch to fail with an exit code.
By using the --stdin option to rev-list (and feeding the object list to its
standard input), we prevent the overflow of the rev-list command line,
which causes quickfetch(), and subsequently the overall fetch, to succeed.
However, using rev-list --stdin is not entirely straightforward: rev-list
terminates immediately when encountering an unknown object, which can
trigger SIGPIPE if we are still writing object's to its standard input.
We therefore temporarily ignore SIGPIPE so that the fetch process is not
terminated.
The patch also contains a testcase to verify the fix (note that before
the patch, the testcase would only fail on msysGit).
Signed-off-by: Johan Herland <johan@herland.net>
Improved-by: Johannes Sixt <j6t@kdbg.org>
Improved-by: Alex Riesen <raa.lkml@gmail.com>
Tested-by: Peter Krefting <peter@softwolves.pp.se>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Back in e3c6f240fd Junio taught
git-fetch to avoid copying objects when we are fetching from
a repository that is already registered as an alternate object
database. In such a case there is no reason to copy any objects
as we can already obtain them through the alternate.
However we need to ensure the objects are all reachable, so we
run `git rev-list --objects $theirs --not --all` to verify this.
If any object is missing or unreadable then we need to fetch/copy
the objects from the remote. When a missing object is detected
the git-rev-list process will exit with a non-zero exit status,
making this condition quite easy to detect.
Although git-fetch is currently a builtin (and so is rev-list)
we cannot invoke the traverse_objects() API at this point in the
transport code. The object walker within traverse_objects() calls
die() as soon as it finds an object it cannot read. If that happens
we want to resume the fetch process by calling do_fetch_pack().
To get around this we spawn git-rev-list into a background process
to prevent a die() from killing the foreground fetch process,
thus allowing the fetch process to resume into do_fetch_pack()
if copying is necessary.
We aren't interested in the output of rev-list (a list of SHA-1
object names that are reachable) or its errors (a "spurious" error
about an object not being found as we need to copy it) so we redirect
both stdout and stderr to /dev/null.
We run this git-rev-list based check before any fetch as we may
already have the necessary objects local from a prior fetch. If we
don't then its very likely the first $theirs object listed on the
command line won't exist locally and git-rev-list will die very
quickly, allowing us to start the network transfer. This test even
on remote URLs may save bandwidth if someone runs `git pull origin`,
sees a merge conflict, resets out, then redoes the same pull just
a short time later. If the remote hasn't changed between the two
pulls and the local repository hasn't had git-gc run in it then
there is probably no need to perform network transfer as all of
the objects are local.
Documentation for the new quickfetch function was suggested and
written by Junio, based on his original comment in git-fetch.sh.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
This updates git-rev-list --objects to be a bit more careful
when listing a blob object to make sure the blob actually
exists, and uses it to make sure the quick-fetch optimization we
introduced earlier is not fooled by a previous incomplete fetch.
The quick-fetch optimization works by running this command:
git rev-list --objects <<commit-list>> --not --all
where <<commit-list>> is a list of commits that we are going to
fetch from the other side. If there is any object missing to
complete the <<commit-list>>, the rev-list would fail and die
(say, the commit was in our repository, but its tree wasn't --
then it will barf while trying to list the blobs the tree
contains because it cannot read that tree).
Usually we do not have the objects (otherwise why would we
fetching?), but in one important special case we do: when the
remote repository is used as an alternate object store
(i.e. pointed by .git/objects/info/alternates). We could check
.git/objects/info/alternates to see if the remote we are
interacting with is one of them (or is used as an alternate,
recursively, by one of them), but that check is more cumbersome
than it is worth.
The above check however did not catch missing blob, because
object listing code did not read nor check blob objects, knowing
that blobs do not contain any further references to other
objects. This commit fixes it with practically unmeasurable
overhead.
I've benched this with
git rev-list --objects --all >/dev/null
in the kernel repository, with three different implementations
of the "check-blob".
- Checking with has_sha1_file() has negligible (unmeasurable)
performance penalty.
- Checking with sha1_object_info() makes it somewhat slower,
perhaps by 5%.
- Checking with read_sha1_file() to cause a fully re-validation
is prohibitively expensive (about 4 times as much runtime).
In my original patch, I had this as a command line option, but
the overhead is small enough that it is not really worth it.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Johan Herland
Shawn O. Pearce
Junio C Hamano