And Junio C Hamano writes:
- Hmph. We do fprintf(stderr, "blah\r") to draw them. The
- standard says that "standard error stream is not fully
- buffered", but I guess it does not necessarily mean it is
- unbuffered, so we probably need to fflush(3) there. Would
- something like this help?
I suppose I should have mentioned that I tried flushing
stderr. Your more comprehensive flushing also does not
fix it, giving outputs like:
> Unpacking Total 3333 objects
> , written 33 (delta 1), reused 0 (delta 0)
The problem is that stderr from a child is not tied to any
stream of its parent. Generally, as far as I know, you
cannot make any assumptions about how pipes from separate
processes are interleaved in the output. Some standard may
say something, but I have no idea what or if anyone listens.
And this particular system is a busy SMP node, making the
problem worse.
Line-buffered streams like stdout tend to work, but not
unbuffered streams like stderr. We can't make stderr line-
buffered without breaking the status indicator...
If I add a third fd to all the pipes and dup it to stderr,
the tests work. I never read from that fd, so I never get
the status output... Progress needs to be part of the
protocol so front ends can handle it cleanly rather than
using stderr tricks.
So some possibilities:
1) Add the ability to pass options through the whole
connect system. Then pass -q in the tester.
2) Add a specific "quiet" command to the protocol for
just passing -q from git-fetch-pack. Pass -q in the
tester.
3) Add an option to pack-objects that dumps progress
output to stdout in a special packet format. Then
update everyone who talks through upload-pack to
expect another phase of informational messages after
negotiating object differences and before the pack
data.
The first two are cosmetic fixes only, and #2 is a cheap,
ugly, but easy hack.
This problem is (to me) low priority. It unfortunately
breaks a test case on AIX, but I can live with it for now.
If others here start to listen to the gospel of git, well,
I'll need to fix it. (But I once recommended Arch, and
people stopped listening after they tried it.)
Folks using moderately-loaded SMPs may experience similar
problems. But if they're fetching large packs, the problem
likely won't appear at all.
Jason
P.S. For the whole finding-a-function-name business, some of
us are using git on fixed-format Fortran. Every non-comment
line begins with whitespace... ;) And in free format, many
people don't add that first indentation within subroutines.
-
: send the line "unsubscribe git" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html