Re: [PATCH 1/6] fetch-pack: clear marks before everything_local()
On Tue, 5 Jun 2018 16:08:21 -0700 Jonathan Nieder wrote: > Hi, > > Jonathan Tan wrote: > > > If tag following is required when using a transport that does not > > support tag following, fetch_pack() will be invoked twice in the same > > process, necessitating a clearing of the object flags used by > > fetch_pack() sometime during the second invocation. This is currently > > done in find_common(), which means that the work done by > > everything_local() in marking complete remote refs as COMMON_REF is > > wasted. > > > > To avoid this wastage, move this clearing from find_common() to its > > parent function do_fetch_pack(), right before it calls > > everything_local(). > > I had to read this a few times and didn't end up understanding it. > > Is the idea that this will speed something up? Can you provide e.g. > "perf stat" output (or even a new perf test in t/perf) demonstrating > the improvement? Or is it a cleanup? Firstly, I don't know of a practical way to demonstrate this, because we don't have an implementation of a transport that does not support tag following. If we could demonstrate this, I think we can demonstrating it by constructing a negotiation scenario in which COMMON_REF would have been helpful, e.g. the following (untested) scenario: T C (T=tag, C=commit) |/ O We run "git fetch C" and on the second round (when we fetch T), if we wiped the flags *after* everything_local() (as is currently done), negotiation would send "have C" and "have O". But if we wiped the flags *before* everything_local(), then C would have the COMMON_REF flag and we will see that we only send "have C". So we have more efficient negotiation. > As an experiment, I tried applying the '-' part of the change without > the '+' part to get confidence that tests cover this well. To my > chagrin, all tests still passed. :/ Yes, because we don't have tests against a transport which doesn't support tag following. > In the preimage, we call clear_marks in find_common. This is right > before we start setting up the revision walk, e.g. by inserting > revisions for each ref. In the postimage, we call clear_marks in > do_fetch_pack. This is right before we call everything_local. > > I end up feeling that I don't understand the code well, neither before > nor after the patch. Ideas for making it clearer? One idea is to first separate everything_local() into its side effect parts and the "true" check that everything is local. I'll do that and send it as part of v2 of this patch series.
Re: [PATCH 1/6] fetch-pack: clear marks before everything_local()
Hi,
Jonathan Tan wrote:
> If tag following is required when using a transport that does not
> support tag following, fetch_pack() will be invoked twice in the same
> process, necessitating a clearing of the object flags used by
> fetch_pack() sometime during the second invocation. This is currently
> done in find_common(), which means that the work done by
> everything_local() in marking complete remote refs as COMMON_REF is
> wasted.
>
> To avoid this wastage, move this clearing from find_common() to its
> parent function do_fetch_pack(), right before it calls
> everything_local().
I had to read this a few times and didn't end up understanding it.
Is the idea that this will speed something up? Can you provide e.g.
"perf stat" output (or even a new perf test in t/perf) demonstrating
the improvement? Or is it a cleanup?
[...]
> --- a/fetch-pack.c
> +++ b/fetch-pack.c
> @@ -336,9 +336,6 @@ static int find_common(struct fetch_pack_args *args,
>
> if (args->stateless_rpc && multi_ack == 1)
> die(_("--stateless-rpc requires multi_ack_detailed"));
> - if (marked)
> - for_each_ref(clear_marks, NULL);
> - marked = 1;
>
> for_each_ref(rev_list_insert_ref_oid, NULL);
> for_each_cached_alternate(insert_one_alternate_object);
> @@ -1053,6 +1050,9 @@ static struct ref *do_fetch_pack(struct fetch_pack_args
> *args,
> if (!server_supports("deepen-relative") && args->deepen_relative)
> die(_("Server does not support --deepen"));
>
> + if (marked)
> + for_each_ref(clear_marks, NULL);
> + marked = 1;
> if (everything_local(args, &ref, sought, nr_sought)) {
As an experiment, I tried applying the '-' part of the change without
the '+' part to get confidence that tests cover this well. To my
chagrin, all tests still passed. :/
In the preimage, we call clear_marks in find_common. This is right
before we start setting up the revision walk, e.g. by inserting
revisions for each ref. In the postimage, we call clear_marks in
do_fetch_pack. This is right before we call everything_local.
I end up feeling that I don't understand the code well, neither before
nor after the patch. Ideas for making it clearer?
Thanks,
Jonathan
[PATCH 1/6] fetch-pack: clear marks before everything_local()
If tag following is required when using a transport that does not
support tag following, fetch_pack() will be invoked twice in the same
process, necessitating a clearing of the object flags used by
fetch_pack() sometime during the second invocation. This is currently
done in find_common(), which means that the work done by
everything_local() in marking complete remote refs as COMMON_REF is
wasted.
To avoid this wastage, move this clearing from find_common() to its
parent function do_fetch_pack(), right before it calls
everything_local().
This has been occurring since the commit that introduced the clearing of
marks, 420e9af498 ("Fix tag following", 2008-03-19).
The corresponding code for protocol v2 in do_fetch_pack_v2() does not
have this problem, as the clearing of flags is done before any marking
(whether by rev_list_insert_ref_oid() or everything_local()).
Signed-off-by: Jonathan Tan
---
fetch-pack.c | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)
diff --git a/fetch-pack.c b/fetch-pack.c
index a320ce987..1358973a4 100644
--- a/fetch-pack.c
+++ b/fetch-pack.c
@@ -336,9 +336,6 @@ static int find_common(struct fetch_pack_args *args,
if (args->stateless_rpc && multi_ack == 1)
die(_("--stateless-rpc requires multi_ack_detailed"));
- if (marked)
- for_each_ref(clear_marks, NULL);
- marked = 1;
for_each_ref(rev_list_insert_ref_oid, NULL);
for_each_cached_alternate(insert_one_alternate_object);
@@ -1053,6 +1050,9 @@ static struct ref *do_fetch_pack(struct fetch_pack_args
*args,
if (!server_supports("deepen-relative") && args->deepen_relative)
die(_("Server does not support --deepen"));
+ if (marked)
+ for_each_ref(clear_marks, NULL);
+ marked = 1;
if (everything_local(args, &ref, sought, nr_sought)) {
packet_flush(fd[1]);
goto all_done;
--
2.17.0.768.g1526ddbba1.dirty
