| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
Code cleanup to limit memory consumption and tighten protocol
message parsing.
* jk/reduce-malloc-in-v2-servers:
ls-refs: reject unknown arguments
serve: reject commands used as capabilities
serve: reject bogus v2 "command=ls-refs=foo"
docs/protocol-v2: clarify some ls-refs ref-prefix details
ls-refs: ignore very long ref-prefix counts
serve: drop "keys" strvec
serve: provide "receive" function for session-id capability
serve: provide "receive" function for object-format capability
serve: add "receive" method for v2 capabilities table
serve: return capability "value" from get_capability()
serve: rename is_command() to parse_command()
|
|
Code clean-up around "git serve".
* ab/serve-cleanup:
upload-pack: document and rename --advertise-refs
serve.[ch]: remove "serve_options", split up --advertise-refs code
{upload,receive}-pack tests: add --advertise-refs tests
serve.c: move version line to advertise_capabilities()
serve: move transfer.advertiseSID check into session_id_advertise()
serve.[ch]: don't pass "struct strvec *keys" to commands
serve: use designated initializers
transport: use designated initializers
transport: rename "fetch" in transport_vtable to "fetch_refs"
serve: mark has_capability() as static
|
|
Reduce number of write(2) system calls while sending the
ref advertisement.
* jv/pkt-line-batch:
upload-pack: use stdio in send_ref callbacks
pkt-line: add stdio packet write functions
|
|
The v2 ls-refs command may receive extra arguments from the client, one
per pkt-line. The spec is pretty clear that the arguments must come from
a specified set, but we silently ignore any unknown entries. For a
well-behaved client this doesn't matter, but it makes testing and
debugging more confusing. Let's tighten this up to match the spec.
In theory this liberal behavior _could_ be useful for extending the
protocol. But:
- every other part of the protocol requires that the server first
indicate that it supports the argument; this includes the fetch and
object-info commands, plus the "unborn" capability added to ls-refs
itself
- it's not a very good extension mechanism anyway; without the server
advertising support, clients would have no idea if the argument was
silently ignored, or accepted and simply had no effect
So we're not really losing anything by tightening this.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Because each "ref-prefix" capability from the client comes in its own
pkt-line, there's no limit to the number of them that a misbehaving
client may send. We read them all into a strvec, which means the client
can waste arbitrary amounts of our memory by just sending us "ref-prefix
foo" over and over.
One possible solution is to just drop the connection when the limit is
reached. If we set it high enough, then only misbehaving or malicious
clients would hit it. But "high enough" is vague, and it's unfriendly if
we guess wrong and a legitimate client hits this.
But we can do better. Since supporting the ref-prefix capability is
optional anyway, the client has to further cull the response based on
their own patterns. So we can simply ignore the patterns once we cross a
certain threshold. Note that we have to ignore _all_ patterns, not just
the ones past our limit (since otherwise we'd send too little data).
The limit here is fairly arbitrary, and probably much higher than anyone
would need in practice. It might be worth limiting it further, if only
because we check it linearly (so with "m" local refs and "n" patterns,
we do "m * n" string comparisons). But if we care about optimizing this,
an even better solution may be a more advanced data structure anyway.
I didn't bother making the limit configurable, since it's so high and
since Git should behave correctly in either case. It wouldn't be too
hard to do, but it makes both the code and documentation more complex.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
* ab/serve-cleanup:
upload-pack: document and rename --advertise-refs
serve.[ch]: remove "serve_options", split up --advertise-refs code
{upload,receive}-pack tests: add --advertise-refs tests
serve.c: move version line to advertise_capabilities()
serve: move transfer.advertiseSID check into session_id_advertise()
serve.[ch]: don't pass "struct strvec *keys" to commands
serve: use designated initializers
transport: use designated initializers
transport: rename "fetch" in transport_vtable to "fetch_refs"
serve: mark has_capability() as static
|
|
In both protocol v0 and v2, upload-pack writes one pktline packet per
advertised ref to stdout. That means one or two write(2) syscalls per
ref. This is problematic if these writes become network sends with
high overhead.
This commit changes both send_ref callbacks to use buffered IO using
stdio.
To give an example of the impact: I set up a single-threaded loop that
calls ls-remote (with HTTP and protocol v2) on a local GitLab
instance, on a repository with 11K refs. When I switch from Git
v2.32.0 to this patch, I see a 40% reduction in CPU time for Git, and
65% for Gitaly (GitLab's Git RPC service).
So using buffered IO not only saves syscalls in upload-pack, it also
saves time in things that consume upload-pack's output.
Helped-by: Jeff King <peff@peff.net>
Signed-off-by: Jacob Vosmaer <jacob@gitlab.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
In the initial reference advertisement, the Git server will first
announce all of its references to the client. The logic is handled in
`send_ref()`, which will allocate a new buffer for each refline it is
about to send. This is quite wasteful: instead of allocating a new
buffer each time, we can just reuse a buffer.
Improve this by passing in a buffer via the `ls_refs_data` struct which
is then reused on each reference.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
The serve.c API added in ed10cb952d3 (serve: introduce git-serve,
2018-03-15) was passing in the raw capabilities "keys", but nothing
downstream of it ever used them.
Let's remove that code because it's not needed. If we do end up
needing to pass information about the advertisement in the future
it'll make more sense to have serve.c parse the capabilities keys and
pass the result of its parsing, rather than expecting expecting its
API users to parse the same keys again.
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
"git clone" tries to locally check out the branch pointed at by
HEAD of the remote repository after it is done, but the protocol
did not convey the information necessary to do so when copying an
empty repository. The protocol v2 learned how to do so.
* jt/clone-unborn-head:
clone: respect remote unborn HEAD
connect, transport: encapsulate arg in struct
ls-refs: report unborn targets of symrefs
|
|
The ls-refs protocol operation has been optimized to narrow the
sub-hierarchy of refs/ it walks to produce response.
* tb/ls-refs-optim:
ls-refs.c: traverse prefixes of disjoint "ref-prefix" sets
ls-refs.c: initialize 'prefixes' before using it
refs: expose 'for_each_fullref_in_prefixes'
|
|
When cloning, we choose the default branch based on the remote HEAD.
But if there is no remote HEAD reported (which could happen if the
target of the remote HEAD is unborn), we'll fall back to using our local
init.defaultBranch. Traditionally this hasn't been a big deal, because
most repos used "master" as the default. But these days it is likely to
cause confusion if the server and client implementations choose
different values (e.g., if the remote started with "main", we may choose
"master" locally, create commits there, and then the user is surprised
when they push to "master" and not "main").
To solve this, the remote needs to communicate the target of the HEAD
symref, even if it is unborn, and "git clone" needs to use this
information.
Currently, symrefs that have unborn targets (such as in this case) are
not communicated by the protocol. Teach Git to advertise and support the
"unborn" feature in "ls-refs" (by default, this is advertised, but
server administrators may turn this off through the lsrefs.unborn
config). This feature indicates that "ls-refs" supports the "unborn"
argument; when it is specified, "ls-refs" will send the HEAD symref with
the name of its unborn target.
This change is only for protocol v2. A similar change for protocol v0
would require independent protocol design (there being no analogous
position to signal support for "unborn") and client-side plumbing of the
data required, so the scope of this patch set is limited to protocol v2.
The client side will be updated to use this in a subsequent commit.
Signed-off-by: Jonathan Tan <jonathantanmy@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
ls-refs performs a single revision walk over the whole ref namespace,
and sends ones that match with one of the given ref prefixes down to the
user.
This can be expensive if there are many refs overall, but the portion of
them covered by the given prefixes is small by comparison.
To attempt to reduce the difference between the number of refs
traversed, and the number of refs sent, only traverse references which
are in the longest common prefix of the given prefixes. This is very
reminiscent of the approach taken in b31e2680c4 (ref-filter.c: find
disjoint pattern prefixes, 2019-06-26) which does an analogous thing for
multi-patterned 'git for-each-ref' invocations.
The callback 'send_ref' is resilient to ignore extra patterns by
discarding any arguments which do not begin with at least one of the
specified prefixes.
Similarly, the code introduced in b31e2680c4 is resilient to stop early
at metacharacters, but we only pass strict prefixes here. At worst we
would return too many results, but the double checking done by send_ref
will throw away anything that doesn't start with something in the prefix
list.
Finally, if no prefixes were provided, then implicitly add the empty
string (which will match all references) since this matches the existing
behavior (see the "no restrictions" comment in "ls-refs.c:ref_match()").
Original-patch-by: Jacob Vosmaer <jacob@gitlab.com>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Correctly initialize the "prefixes" strvec using strvec_init() instead
of simply zeroing it via the earlier memset().
There's no way to trigger a crash, since the first 'ref-prefix' command
will initialize the strvec via the 'ALLOC_GROW' in 'strvec_push_nodup()'
(the alloc and nr variables are already zero'd, so the call to
ALLOC_GROW is valid).
If no "ref-prefix" command was given, then the call to
'ls-refs.c:ref_match()' will abort early after it reads the zero in
'prefixes->nr'. Likewise, strvec_clear() will only call free() on the
array, which is NULL, so we're safe there, too.
But, all of this is dangerous and requires more reasoning than it would
if we simply called 'strvec_init()', so do that.
Signed-off-by: Jacob Vosmaer <jacob@gitlab.com>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
The peel_ref() interface is confusing and error-prone:
- it's typically used by ref iteration callbacks that have both a
refname and oid. But since they pass only the refname, we may load
the ref value from the filesystem again. This is inefficient, but
also means we are open to a race if somebody simultaneously updates
the ref. E.g., this:
int some_ref_cb(const char *refname, const struct object_id *oid, ...)
{
if (!peel_ref(refname, &peeled))
printf("%s peels to %s",
oid_to_hex(oid), oid_to_hex(&peeled);
}
could print nonsense. It is correct to say "refname peels to..."
(you may see the "before" value or the "after" value, either of
which is consistent), but mentioning both oids may be mixing
before/after values.
Worse, whether this is possible depends on whether the optimization
to read from the current iterator value kicks in. So it is actually
not possible with:
for_each_ref(some_ref_cb);
but it _is_ possible with:
head_ref(some_ref_cb);
which does not use the iterator mechanism (though in practice, HEAD
should never peel to anything, so this may not be triggerable).
- it must take a fully-qualified refname for the read_ref_full() code
path to work. Yet we routinely pass it partial refnames from
callbacks to for_each_tag_ref(), etc. This happens to work when
iterating because there we do not call read_ref_full() at all, and
only use the passed refname to check if it is the same as the
iterator. But the requirements for the function parameters are quite
unclear.
Instead of taking a refname, let's instead take an oid. That fixes both
problems. It's a little funny for a "ref" function not to involve refs
at all. The key thing is that it's optimizing under the hood based on
having access to the ref iterator. So let's change the name to make it
clear why you'd want this function versus just peel_object().
There are two other directions I considered but rejected:
- we could pass the peel information into the each_ref_fn callback.
However, we don't know if the caller actually wants it or not. For
packed-refs, providing it is essentially free. But for loose refs,
we actually have to peel the object, which would be wasteful in most
cases. We could likewise pass in a flag to the callback indicating
whether the peeled information is known, but that complicates those
callbacks, as they then have to decide whether to manually peel
themselves. Plus it requires changing the interface of every
callback, whether they care about peeling or not, and there are many
of them.
- we could make a function to return the peeled value of the current
iterated ref (computing it if necessary), and BUG() otherwise. I.e.:
int peel_current_iterated_ref(struct object_id *out);
Each of the current callers is an each_ref_fn callback, so they'd
mostly be happy. But:
- we use those callbacks with functions like head_ref(), which do
not use the iteration code. So we'd need to handle the fallback
case there, anyway.
- it's possible that a caller would want to call into generic code
that sometimes is used during iteration and sometimes not. This
encapsulates the logic to do the fast thing when possible, and
fallback when necessary.
The implementation is mostly obvious, but I want to call out a few
things in the patch:
- the test-tool coverage for peel_ref() is now meaningless, as it all
collapses to a single peel_object() call (arguably they were pretty
uninteresting before; the tricky part of that function is the
fast-path we see during iteration, but these calls didn't trigger
that). I've just dropped it entirely, though note that some other
tests relied on the tags we created; I've moved that creation to the
tests where it matters.
- we no longer need to take a ref_store parameter, since we'd never
look up a ref now. We do still rely on a global "current iterator"
variable which _could_ be kept per-ref-store. But in practice this
is only useful if there are multiple recursive iterations, at which
point the more appropriate solution is probably a stack of
iterators. No caller used the actual ref-store parameter anyway
(they all call the wrapper that passes the_repository).
- the original only kicked in the optimization when the "refname"
pointer matched (i.e., not string comparison). We do likewise with
the "oid" parameter here, but fall back to doing an actual oideq()
call. This in theory lets us kick in the optimization more often,
though in practice no current caller cares. It should never be
wrong, though (peeling is a property of an object, so two refs
pointing to the same object would peel identically).
- the original took care not to touch the peeled out-parameter unless
we found something to put in it. But no caller cares about this, and
anyway, it is enforced by peel_object() itself (and even in the
optimized iterator case, that's where we eventually end up). We can
shorten the code and avoid an extra copy by just passing the
out-parameter through the stack.
Signed-off-by: Jeff King <peff@peff.net>
Reviewed-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
The "argc" and "argv" names made sense when the struct was argv_array,
but now they're just confusing. Let's rename them to "nr" (which we use
for counts elsewhere) and "v" (which is rather terse, but reads well
when combined with typical variable names like "args.v").
Note that we have to update all of the callers immediately. Playing
tricks with the preprocessor is hard here, because we wouldn't want to
rewrite unrelated tokens.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
We eventually want to drop the argv_array name and just use strvec
consistently. There's no particular reason we have to do it all at once,
or care about interactions between converted and unconverted bits.
Because of our preprocessor compat layer, the names are interchangeable
to the compiler (so even a definition and declaration using different
names is OK).
This patch converts remaining files from the first half of the alphabet,
to keep the diff to a manageable size.
The conversion was done purely mechanically with:
git ls-files '*.c' '*.h' |
xargs perl -i -pe '
s/ARGV_ARRAY/STRVEC/g;
s/argv_array/strvec/g;
'
and then selectively staging files with "git add '[abcdefghjkl]*'".
We'll deal with any indentation/style fallouts separately.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
This requires updating #include lines across the code-base, but that's
all fairly mechanical, and was done with:
git ls-files '*.c' '*.h' |
xargs perl -i -pe 's/argv-array.h/strvec.h/'
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
When processing the arguments list for a v2 ls-refs or fetch command, we
loop like this:
while (packet_reader_read(request) != PACKET_READ_FLUSH) {
const char *arg = request->line;
...handle arg...
}
to read and handle packets until we see a flush. The hidden assumption
here is that anything except PACKET_READ_FLUSH will give us valid packet
data to read. But that's not true; PACKET_READ_DELIM or PACKET_READ_EOF
will leave packet->line as NULL, and we'll segfault trying to look at
it.
Instead, we should follow the more careful model demonstrated on the
client side (e.g., in process_capabilities_v2): keep looping as long
as we get normal packets, and then make sure that we broke out of the
loop due to a real flush. That fixes the segfault and correctly
diagnoses any unexpected input from the client.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Since 7171d8c15f (upload-pack: send symbolic ref information as
capability, 2013-09-17), we've sent cloning and fetching clients special
information about which branch HEAD is pointing to, so that they don't
have to guess based on matching up commit ids.
However, this feature has never worked properly with the GIT_NAMESPACE
feature. Because upload-pack uses head_ref_namespaced(find_symref), we
do find and report on refs/namespaces/foo/HEAD instead of the actual
HEAD of the repo. This makes sense, since the branch pointed to by the
top-level HEAD may not be advertised at all. But we do two things wrong:
1. We report the full name refs/namespaces/foo/HEAD, instead of just
HEAD. Meaning no client is going to bother doing anything with that
symref, since we're not otherwise advertising it.
2. We report the symref destination using its full name (e.g.,
refs/namespaces/foo/refs/heads/master). That's similarly useless to
the client, who only saw "refs/heads/master" in the advertisement.
We should be stripping the namespace prefix off of both places (which
this patch fixes).
Likely nobody noticed because we tend to do the right thing anyway. Bug
(1) means that we said nothing about HEAD (just refs/namespace/foo/HEAD).
And so the client half of the code, from a45b5f0552 (connect: annotate
refs with their symref information in get_remote_head(), 2013-09-17),
does not annotate HEAD, and we use the fallback in guess_remote_head(),
matching refs by object id. Which is usually right. It only falls down
in ambiguous cases, like the one laid out in the included test.
This also means that we don't have to worry about breaking anybody who
was putting pre-stripped names into their namespace symrefs when we fix
bug (2). Because of bug (1), nobody would have been using the symref we
advertised in the first place (not to mention that those symrefs would
have appeared broken for any non-namespaced access).
Note that we have separate fixes here for the v0 and v2 protocols. The
symref advertisement moved in v2 to be a part of the ls-refs command.
This actually gets part (1) right, since the symref annotation
piggy-backs on the existing ref advertisement, which is properly
stripped. But it still needs a fix for part (2). The included tests
cover both protocols.
Reported-by: Bryan Turner <bturner@atlassian.com>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Fix namespace support in protocol v2.
* jt/namespaced-ls-refs-fix:
ls-refs: filter refs using namespace-stripped name
|
|
If a user fetches refs/heads/master from a repo with namespace "ns", the
remote is expected to (1) not send the real refs/heads/master, and (2)
send refs/namespaces/ns/refs/heads/master with the name
refs/heads/master. (1) indeed happens now, but not (2) - Git only sends
refs that have the user-given prefix, but it checks them against the
full name of the ref (the one starting with refs/namespaces), and not
the namespace-stripped one.
This is demonstrated by the patch in the test. Currently, it results in
"fatal: couldn't find remote ref refs/heads/master" despite both
unnamespaced and namespaced master being present. With the code change,
it produces the expected result.
Check the ref prefixes against the namespace-stripped name.
This bug was discovered through applying patches [1] that override
protocol.version to 2 in repositories when running tests, allowing us to
notice differences in behavior across different protocol versions.
[1] https://public-inbox.org/git/cover.1547677183.git.jonathantanmy@google.com/
Signed-off-by: Jonathan Tan <jonathantanmy@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
In the v2 protocol, upload-pack's advertisement has been moved to the
"ls-refs" command. That command does not respect hidden-ref config (like
transfer.hiderefs) at all, and advertises everything.
While there are some features that are not supported in v2 (e.g., v2
always allows fetching any sha1 without respect to advertisements), the
lack of this feature is not documented and is likely just a bug. Let's
make it work, as otherwise upgrading a server to a v2-capable git will
start exposing these refs that the repository admin has asked to remain
hidden.
Note that we assume we're operating on behalf of a fetch here, since
that's the only thing implemented in v2 at this point. See the in-code
comment. We'll have to figure out how this works when the v2 push
protocol is designed (both here in ls-refs, but also rejecting updates
to hidden refs).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Introduce the ls-refs server command. In protocol v2, the ls-refs
command is used to request the ref advertisement from the server. Since
it is a command which can be requested (as opposed to mandatory in v1),
a client can sent a number of parameters in its request to limit the ref
advertisement based on provided ref-prefixes.
Signed-off-by: Brandon Williams <bmwill@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
Junio C Hamano
Jeff King
Jacob Vosmaer
Patrick Steinhardt
Ævar Arnfjörð Bjarmason
Jonathan Tan
Taylor Blau
Brandon Williams