| Message ID | 20200715223112.2018556-1-jonathantanmy@google.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | upload-pack: do not lazy-fetch "have" objects | expand |
Jonathan Tan <jonathantanmy@google.com> writes: > When upload-pack receives a request containing "have" hashes, it (among > other things) checks if the served repository has the corresponding > objects. However, it does not do so with the > OBJECT_INFO_SKIP_FETCH_OBJECT flag, so if serving a partial clone, a > lazy fetch will be triggered first. OK. Fixing issues hit by real users reactively is a necessary and good thing, but this is not the first time we patch callers of has_object_file() for this kind of "we are merely trying to determine the boundary of what we have, so that we know what we need to add to this repository" queries, I am afraid. Perhaps it is a good idea to sweep all the hits from "git grep -e has_object_file \*.c" and audit the codebase to see if there are other problematic ones? For example, list-objects.c::process_blob() tries to if the object exists when --exclude-promisor-objects is in effect so that it can return early if the object is missing and it is a promisor object. I would imagine that we would not want to lazy-fetch the object in this case. Thanks. Will queue.
On Wed, Jul 15, 2020 at 03:55:18PM -0700, Junio C Hamano wrote: > Fixing issues hit by real users reactively is a necessary and good > thing, but this is not the first time we patch callers of > has_object_file() for this kind of "we are merely trying to > determine the boundary of what we have, so that we know what we need > to add to this repository" queries, I am afraid. > > Perhaps it is a good idea to sweep all the hits from "git grep -e > has_object_file \*.c" and audit the codebase to see if there are > other problematic ones? Interestingly, the case fixed here was mentioned after the three-dashes in this patch: https://lore.kernel.org/git/20191127123211.GG22221@sigill.intra.peff.net/ That thread was about OBJECT_INFO_QUICK, but I think any "we are finding the boundary" hunt should consider whether that's appropriate, too. In that case they are really the same issue: we do not want to expend effort A on the unlikely chance P that we will find the object and save ourselves effort B. The magnitude of A and B, and the probability P may be different for the two optimizations, though. I think the instance fixed here _probably_ could be converted to QUICK as well. It saves extra refreshes of the pack directory when a client asks for an object we don't have. That's usually not too expensive, but can be in some cases (many alternates, slow NFS filesystems, client has many objects we don't). In the worst case, using QUICK would mean we can get fooled by a simultaneous repack into thinking we don't have an object we _do_ have, and end up negotiating a worse pack for the client. > Thanks. Will queue. That said, yeah, I think Jonathan's patch is a strict improvement by itself. -Peff
Jeff King <peff@peff.net> writes: > I think the instance fixed here _probably_ could be converted to QUICK > as well. It saves extra refreshes of the pack directory when a client > asks for an object we don't have. That's usually not too expensive, but > can be in some cases (many alternates, slow NFS filesystems, client has > many objects we don't). In the worst case, using QUICK would mean we can > get fooled by a simultaneous repack into thinking we don't have an > object we _do_ have, and end up negotiating a worse pack for the client. > >> Thanks. Will queue. > > That said, yeah, I think Jonathan's patch is a strict improvement by > itself. I'll see what Jonathan would say, hoping that he'd consider if QUICK is appropriate or not sufficiently enough to result in an updated patch, that either has QUICK in the code or the reason why QUICK is not used in the log message. Thanks.
diff --git a/t/t5616-partial-clone.sh b/t/t5616-partial-clone.sh index 8a27452a51..37de0afb02 100755 --- a/t/t5616-partial-clone.sh +++ b/t/t5616-partial-clone.sh @@ -422,6 +422,44 @@ test_expect_success 'single-branch tag following respects partial clone' ' test_must_fail git -C single rev-parse --verify refs/tags/C ' +test_expect_success 'fetch from a partial clone, protocol v0' ' + rm -rf server client trace && + + # Pretend that the server is a partial clone + git init server && + git -C server remote add a_remote "file://$(pwd)/" && + test_config -C server core.repositoryformatversion 1 && + test_config -C server extensions.partialclone a_remote && + test_config -C server protocol.version 0 && + test_commit -C server foo && + + # Fetch from the server + git init client && + test_config -C client protocol.version 0 && + test_commit -C client bar && + GIT_TRACE_PACKET="$(pwd)/trace" git -C client fetch "file://$(pwd)/server" && + ! grep "version 2" trace +' + +test_expect_success 'fetch from a partial clone, protocol v2' ' + rm -rf server client trace && + + # Pretend that the server is a partial clone + git init server && + git -C server remote add a_remote "file://$(pwd)/" && + test_config -C server core.repositoryformatversion 1 && + test_config -C server extensions.partialclone a_remote && + test_config -C server protocol.version 2 && + test_commit -C server foo && + + # Fetch from the server + git init client && + test_config -C client protocol.version 2 && + test_commit -C client bar && + GIT_TRACE_PACKET="$(pwd)/trace" git -C client fetch "file://$(pwd)/server" && + grep "version 2" trace +' + . "$TEST_DIRECTORY"/lib-httpd.sh start_httpd diff --git a/upload-pack.c b/upload-pack.c index 951a2b23aa..af9d621755 100644 --- a/upload-pack.c +++ b/upload-pack.c @@ -482,7 +482,7 @@ static int got_oid(struct upload_pack_data *data, { if (get_oid_hex(hex, oid)) die("git upload-pack: expected SHA1 object, got '%s'", hex); - if (!has_object_file(oid)) + if (!has_object_file_with_flags(oid, OBJECT_INFO_SKIP_FETCH_OBJECT)) return -1; return do_got_oid(data, oid); } @@ -1423,7 +1423,8 @@ static int process_haves(struct upload_pack_data *data, struct oid_array *common for (i = 0; i < data->haves.nr; i++) { const struct object_id *oid = &data->haves.oid[i]; - if (!has_object_file(oid)) + if (!has_object_file_with_flags(oid, + OBJECT_INFO_SKIP_FETCH_OBJECT)) continue; oid_array_append(common, oid);
When upload-pack receives a request containing "have" hashes, it (among other things) checks if the served repository has the corresponding objects. However, it does not do so with the OBJECT_INFO_SKIP_FETCH_OBJECT flag, so if serving a partial clone, a lazy fetch will be triggered first. This was discovered at $DAYJOB when a user fetched from a partial clone (into another partial clone - although this would also happen if the repo to be fetched into is not a partial clone). Therefore, whenever "have" hashes are checked for existence, pass the OBJECT_INFO_SKIP_FETCH_OBJECT flag. Signed-off-by: Jonathan Tan <jonathantanmy@google.com> --- There is also the greater issue that if a lazy fetch fails, the fetch is usually fatal (and possibly always fatal - I haven't checked all the code paths) when the calling code could just as easily continue without the object (which is the case for upload-pack when checking "have"s), but I haven't addressed that here. --- t/t5616-partial-clone.sh | 38 ++++++++++++++++++++++++++++++++++++++ upload-pack.c | 5 +++-- 2 files changed, 41 insertions(+), 2 deletions(-)