exit status = 1 from git fetch -t

exit status = 1 from git fetch -t

[Villeneuve]

Hi,

I'm getting an unexpected exit status of 1 from "git fetch
-t" on one of our largest repositories (by large, I mean 47
branches and 5442 tags, some of which might not be reachable
from branch heads).

The docs are not clear about the interpretation of git
fetch's exit status, so maybe I should not bother.

But assuming the exit status of 1 really indicates some
trouble, here are more details.

Next tests done with git over ssh (server 1.6.1.3, client
1.6.3.1) on RHEL4.

In the "REPO1" repository (exit status printed just before $
in prompt following command):

[REPO1] 0 $ git fetch -t -v -v
Server supports multi_ack
Server supports side-band-64k
Server supports ofs-delta
Marking 60e4d540748c5c3d368c888c4c248de0bdd684cc as complete
Marking 60e4d540748c5c3d368c888c4c248de0bdd684cc as complete
[REPO1] 1 $

while with another repository, "REPO2":

[REPO2] 0 $ git fetch -t -v -v
From ssh://user@machine/path/to/GIT/dir1/dir2/REPO2
 = [up to date]      v1.1 -> v1.1
 = [up to date]      v1.2 -> v1.2
 = [up to date]      v2.0 -> v2.0
... # lots of other similar lines
[REPO2] 0 $

I've compiled git in debug to step in the code, and the
source of the "error" exit status seems to come from
builtin-fetch-pack.c:everything_local, where *refs being
NULL leaves retval to 1 in the final loop.

I'm also puzzled that I don't get the "From ssh://..." in
the trace from "REPO1" as for the "REPO2" repository above.
I do get the "From ssh://..."  line if doing only "git fetch
-v -v" without the "-t" option in "REPO1".

Any help/advice appreciated,
--
Daniel Villeneuve
Kronos, AD OPT Division

Re: exit status = 1 from git fetch -t

[Michael J Gruber]

Villeneuve venit, vidit, dixit 28.05.2009 05:57:
> Hi,
> 
> I'm getting an unexpected exit status of 1 from "git fetch
> -t" on one of our largest repositories (by large, I mean 47
> branches and 5442 tags, some of which might not be reachable
> from branch heads).
> 
> The docs are not clear about the interpretation of git
> fetch's exit status, so maybe I should not bother.
> 
> But assuming the exit status of 1 really indicates some
> trouble, here are more details.
> 
> Next tests done with git over ssh (server 1.6.1.3, client
> 1.6.3.1) on RHEL4.
> 
> In the "REPO1" repository (exit status printed just before $
> in prompt following command):
> 
> [REPO1] 0 $ git fetch -t -v -v
> Server supports multi_ack
> Server supports side-band-64k
> Server supports ofs-delta
> Marking 60e4d540748c5c3d368c888c4c248de0bdd684cc as complete
> Marking 60e4d540748c5c3d368c888c4c248de0bdd684cc as complete
> [REPO1] 1 $
> 
> while with another repository, "REPO2":
> 
> [REPO2] 0 $ git fetch -t -v -v
> From ssh://user@machine/path/to/GIT/dir1/dir2/REPO2
>  = [up to date]      v1.1 -> v1.1
>  = [up to date]      v1.2 -> v1.2
>  = [up to date]      v2.0 -> v2.0
> ... # lots of other similar lines
> [REPO2] 0 $
> 
> I've compiled git in debug to step in the code, and the
> source of the "error" exit status seems to come from
> builtin-fetch-pack.c:everything_local, where *refs being
> NULL leaves retval to 1 in the final loop.
> 
> I'm also puzzled that I don't get the "From ssh://..." in
> the trace from "REPO1" as for the "REPO2" repository above.
> I do get the "From ssh://..."  line if doing only "git fetch
> -v -v" without the "-t" option in "REPO1".

Looking at that loop, I reckon everything_local() returns 0 if there is
at least one object we "want" from the remote and 1 if there is none
(i.e. if "everything is local"). So that seems intentional. (The
structure of REPO1 and your fetch refspecs lines should explain why we
don't want anything.)

About the "From": If fetch displays no note (i.e. "foo   ref -> ref") it
displays no "From" either, which makes sense. We're not getting anything
from anywhere, we're not even listing up to date info, so we're not
displaying the reference point.

Michael

RE: exit status = 1 from git fetch -t

[Villeneuve, Daniel]

>From: Michael J Gruber [mailto:git@drmicha.warpmail.net] 
>Villeneuve venit, vidit, dixit 28.05.2009 05:57:
>> I'm getting an unexpected exit status of 1 from "git fetch
>> -t" on one of our largest repositories (by large, I mean 47
>> branches and 5442 tags, some of which might not be reachable
>> from branch heads).
>... 
>> I've compiled git in debug to step in the code, and the
>> source of the "error" exit status seems to come from
>> builtin-fetch-pack.c:everything_local, where *refs being
>> NULL leaves retval to 1 in the final loop.
>> 
>> I'm also puzzled that I don't get the "From ssh://..." in
>> the trace from "REPO1" as for the "REPO2" repository above.
>> I do get the "From ssh://..."  line if doing only "git fetch
>> -v -v" without the "-t" option in "REPO1".
>
>Looking at that loop, I reckon everything_local() returns 0 if there is
>at least one object we "want" from the remote and 1 if there is none
>(i.e. if "everything is local"). So that seems intentional. (The
>structure of REPO1 and your fetch refspecs lines should explain why we
>don't want anything.)
>
>About the "From": If fetch displays no note (i.e. "foo   ref -> ref")
it
>displays no "From" either, which makes sense. We're not getting
anything
>from anywhere, we're not even listing up to date info, so we're not
>displaying the reference point.

I've cloned both repositories locally, so now everything runs
with git 1.6.3.1.  The config is set by clone (called with no
options) with just new tracking branches added via "git branch --track".

Both repositories are set in the same way.

Digging further, I've located the divergence point between REPO1 and
REPO2
in the return from quickfetch, called from builtin-fetch:fetch_refs.  On
the small repository (REPO2, 303 tags), quickfetch succeeds and returns
0.
On the large repository (REPO1, 5439 tags (maybe I lost 3 in the debug
setup steps)), quickfetch fails and returns 127.   After that point,
REPO1 goes on with transport_fetch_refs which causes a non-zero return,
while REPO2 continues with ret==0 and returns 0.

Using GIT_TRACE=1, we can see that the rev-list command built by
quickfetch
for REPO1 takes about 233930 bytes, which when tried directly from bash
yields
"Argument list too long".

In gdb, I also faked a return value of 1 from quickfetch for REPO2, and
it then
followed the same path as for REPO1, returning an exit status of 1.

So it seems that the exit status currently depends on which strategy
is used to identify some missing commits, and it's not clear that
a non-zero exit status indicates an error.

Maybe I should just grep stderr from "git fetch" for ".*failed.*"
instead of using the exit status (unless I'm missing something).
--
Daniel