git/list[1] front-page[2] threads[3] people[4] search[5] about
 

t5570-git-daemon fails with SIGPIPE on OSX

From
SZEDER Gábor <szeder.dev@gmail.com>
Date
Aug 6, 2018, 15:11 UTC
Message-ID
<CAM0VKj=MCS+cmOgzf_XyPeb+qZrFmuMH52-PV_NDMZA9X+rRoA@mail.gmail.com>

Travis CI changed its default OSX image to use XCode 9.4 on 2018-07-31 [1]. Since then OSX build jobs fail rather frequently because of a SIGPIPE in the tests 'fetch notices corrupt pack' or 'fetch notices corrupt idx' in 't5570-git-daemon.sh' [2]. I think this is a symptom a real bug in Git affecting other platforms as well, but these tests are too lax to catch it.

What it boils down to is this sequence:
  - The test first prepares a repository containing a corrupt pack,
    ready to be server via 'git daemon'.
  - Then the test runs 'test_must_fail git fetch ....', which connects
    to 'git daemon', which forks 'git upload-pack', which then
    advertises refs (only HEAD) and capabilities.  So far so good.
  - 'git fetch' eventually calls fetch-pack.c:find_common().  The
    first half of this function assembles a request consisting of a
    want and a flush pkt-line, and sends it via a send_request() call.
    At this point the scheduling becomes important: let's suppose that
    fetch is slow and upload-pack is fast.
  - 'git upload-pack' receives the request, parses the want line,
    notices the corrupt pack, responds with an 'ERR upload-pack: not
    our ref' pkt-line, and die()s right away.
  - 'git fetch' finally approaches the end of the function, where it
    attempts to send a done pkt-line via another send_request() call
    through the now closing TCP socket.
  - What happens now seems to depend on the platform:
    - On Linux, both on my machine and on Travis CI, it shows textbook
      example behaviour: write() returns with error and sets errno to
      ECONNRESET.  Since it happens in write_or_die(), 'git fetch'
      die()s with 'fatal: write error: Connection reset by peer', and
      doesn't show the error send by 'git upload-pack'; how could it,
      it doesn't even get as far to receive upload-pack's ERR
      pkt-line.
      The test only checks that 'git fetch' fails, but it doesn't
      check whether it failed with the right error message, so the
      test still succeeds.  Had it checked the error message as well,
      we most likely had noticed this issue already, it doesn't happen
      all that rarely.
    - On the new OSX images with XCode 9.4 on Travis CI the write()
      triggers SIGPIPE right away, and 'test_must_fail' notices it and
      fails the test.  I couldn't see any sign of an ECONNRESET or any
      other error that we could act upon to avoid the SIGPIPE.
    - On OSX with XCode 9.2 on Travis CI there is neither SIGPIPE, nor
      ECONNRESET, but sending the request actually succeeds even
      though there is no process on the other end of the socket
      anymore.  'git fetch' then simply continues execution, reads and
      parses the ERR pkt-line, and then dies()s with 'fatal: remote
      error: upload-pack: not our ref'.  So, on the face of it, it
      shows the desired behaviour, but I have no idea how that write()
      could succeed instead of returning error.

I don't know what happens on a real Mac as I don't have access to one; I figured out all the above by enabling packet tracing, adding a couple of well placed tracing printf() and sleep() calls, running a bunch of builds on Travis CI, and looking through their logs. But without access to a debugger and netstat and what not I can't really go any further. So I would now happily pass the baton to those who have a Mac and know a thing or two about its porting issues to first check whether OSX on a real Mac shows the same behaviour as it does in Travis CI's virtualized(?) environment. And then they can pass the baton to those who know all the intricacies of the pack protocol and its implementation to decide what to do with this issue.

For a mostly reliable reproduction recipe you might want to fetch this branch:

  https://github.com/szeder/git t5570-git-daemon-sigpipe
and then run 'make && cd t && ./t5570-git-daemon.sh -v -x'
Have fun! ;)
1 - https://blog.travis-ci.com/2018-07-19-xcode9-4-default-announce
2 - On git.git's master:
    https://travis-ci.org/git/git/jobs/411517552#L2717
Next: SZEDER Gábor
Message 1 of 21 in “t5570-git-daemon fails with SIGPIPE on OSX”
  1. SZEDER GáborAug 6, 2018
  2. SZEDER GáborAug 6, 2018
  3. Johannes SchindelinFeb 8, 2019
  4. SZEDER GáborFeb 8, 2019
  5. Jeff KingAug 14, 2018
  6. Jeff KingAug 14, 2018
  7. Johannes SchindelinFeb 8, 2019
  8. Johannes SchindelinFeb 8, 2019
  9. Jeff KingFeb 8, 2019
  10. Johannes SchindelinMar 1, 2019
  11. Jeff KingMar 1, 2019
  12. Johannes SchindelinMar 2, 2019
  13. Jeff KingMar 3, 2019
  14. 1/2 fetch: avoid calling write_or_die()Jeff King, Mar 3, 2019
  15. Duy NguyenMar 4, 2019
  16. Jeff KingMar 5, 2019
  17. 2/2 fetch: ignore SIGPIPE during network operationJeff King, Mar 3, 2019
  18. Junio C HamanoMar 4, 2019
  19. Jeff KingMar 5, 2019
  20. Junio C HamanoMar 3, 2019
  21. Johannes SchindelinMar 3, 2019

Read the whole thread, see it on lore, or plain text.

$ cat FOOTERMessages come from the public archive at lore.kernel.org/git, fetched every hour. The front page is chosen and written each morning by an AI editor and can be wrong; the threads themselves are the record. About and API. For agents: an MCP server at https://gitlist.dev/mcp, and any thread, story or person page as Markdown by adding .md to its URL (or sending Accept: text/markdown). Details in /llms.txt.