[[!comment format=mdwn
 username="joey"
 subject="""comment 7"""
 date="2016-05-02T14:19:43Z"
 content="""
Relevant parts of the strace:

	[pid 21392] read(0,  <unfinished ...>
	[pid 21393] <... futex resumed> )       = 0
	[pid 21392] <... read resumed> "\0\0009\214moov\0\0\0lmvhd\0\0\0\0\300\\\26\256\300\\\26\267\0\0\2X"..., 32752) = 32752
	[pid 21392] write(1, "/annex/objects/SHA256E-s2589913-"..., 102 <unfinished...>
	[pid 21389] <... read resumed> "/annex/objects/SHA256E-s2589913-"..., 2589913) = 102

21392 is git-annex smudge, and 21389 is apparently a thread belonging to git.
This shows git feeding 32752 bytes of content to git-annex smudge on stdin, and
git-annex smudge responding with the annex object.

Notice that git-annex calculates the actual file size to be 2589913 bytes, more
than git has sent it at the time it replies. It can tell this because it stats
and hashes the actual file.

	[pid 21392] +++ exited with 0 +++
	[pid 21390] write(9, "ml\3\23<\270\315\321s\322\306\341\301e\261\312[<\2\n\17\315\16a\v\245\215(q\207\314\23"..., 2495705) = -1 EPIPE (Broken pipe)
	[pid 21390] --- SIGPIPE {si_signo=SIGPIPE, si_code=SI_USER, si_pid=21389, si_uid=2001} ---
	[pid 21390] close(9)                    = 0
	[pid 21390] write(2, "error: cannot feed the input to "..., 76error: cannot feed the input to external filter git-annex smudge --clean %f

Now git-annex exits, and then git, tries to feed it another 2495705 bytes
of data after it's exited. 

So, it seems that the problem is git-annex is somehow doing a short read of its
stdin, and not waiting for git to feed it the full content.

And, I see this bug now in the code:

	b <- liftIO $ B.hGetContents stdin
	if isJust (parseLinkOrPointer b)

`parseLinkOrPointer` is careful to *not* look at the whole file content,
so nothing forces the lazy bytestring read to consume all of stdin.
So, if git's writes are broken up as happened in this strace, git-annex
will exit without consuming the whole stdin.

I've committed a fix along with this comment, so since you're building
from source, and able to reproduce this problem (which I have not
so far been able to reproduce myself)
you should be able to update and rebuild and verify the fix.
"""]]