On GNU/Linux, it isn't declared in stdio.h unless we define
_GNU_SOURCE, which we don't; using the implicit declaration risks
crashes on platforms where sizeof(pointer) != sizeof(int). On other
platforms it isn't guaranteed to exist at all.
Signed-off-by: Simon McVittie <smcv@debian.org>
Otherwise, if third-party plugins extend newenviron by more than
3 entries, we could overflow the array. It seems unlikely that any
third-party plugin manipulates newenviron in practice, so this
is mostly theoretical. Just in case, I have deliberately avoided
using "i" as the variable name, so that any third-party plugin
that was manipulating newenviron directly will now result in the
wrapper failing to compile.
I have not assumed that realloc(NULL, ...) works as an equivalent of
malloc(...), in case there are still operating systems where that
doesn't work.
We had a weird problem where, after moving to a new, faster server,
"git push" would sometimes fail like this:
Unpacking objects: 100% (3/3), done.
fatal: The remote end hung up unexpectedly
fatal: The remote end hung up unexpectedly
What turned out to be going on was that git-receive-pack was dying due
to an uncaught SIGPIPE. The SIGPIPE occurred when it tried to write to
the pre-receive hook's stdin. The pre-receive hook, in this case, was
able to do all the checks it needed to do without the input, and so did
exit(0) without consuming it.
Apparently that causes a race. Most of the time, git forks the hook,
writes output to the hook, and then the hook runs, ignores it, and exits.
But sometimes, on our new faster server, git forked the hook, and it
ran, and exited, before git got around to writing to it, resulting in
the SIGPIPE.
write(7, "c9f98c67d70a1cfeba382ec27d87644a"..., 100) = -1 EPIPE (Broken
pipe)
--- SIGPIPE (Broken pipe) @ 0 (0) ---
I think git should ignore SIGPIPE when writing to hooks. Otherwise,
hooks may have to go out of their way to consume all input, and as I've
seen, the races when they fail to do this can lurk undiscovered.
I have written to the git mailing list about this.
As a workaround, consume all stdin before exiting.
Avoid the generic "you are not allowed to change" message,
and instead allow check_canedit to propigate out useful error messages.
Went back to calling check_canedit in fatal mode, but added a parameter to
avoid calling the troublesome subs that might cause a login attempt.
Everywhere that REMOTE_ADDR was used, a session object is available, so
instead use its remote_addr method.
In IkiWiki::Receive, stop setting a dummy REMOTE_ADDR.
Note that it's possible for a session cookie to be obtained using one IP
address, and then used from another IP. In this case, the first IP will now
be used. I think that should be ok.
Many calls to file_prune were incorrectly calling it with 2 parameters.
In cases where the filename being checked is relative to the srcdir,
that is not needed.
Made absolute filenames be pruned. (This won't work for the 2 parameter call
style.)
Loading and use of IkiWiki::Receive can all be pushed into the git plugin,
rather than scattered around.
I had at first wanted to make a receive plugin and move it there,
but a plugin was not a good fit; you don't want users to have to manually
load it, and making the git plugin load the receive plugin at the right
times would need more, and ugly code.
Simon McVittie
Joey Hess
Peter Gammie
Joey Hess
Joey Hess