lfs: fix the stall and corruption issue when concurrently uploading blobs
We've avoided the issue up to this point by gating worker usage with an
experimental config. See 10e62d5efa73, and the thread linked there for some of
the initial diagnosis, but essentially some data was being read from the blob
before an error occurred and `keepalive` retried, but didn't rewind the file
pointer. So the leading data was lost from the blob on the server, and the
connection stalled, trying to send more data than available.
In trying to recreate this, I was unable to do so uploading from Windows to
CentOS 7. But it reproduced every time going from CentOS 7 to another CentOS 7
over https.
I found recent fixes in the FaceBook repo to address this[1][2]. The commit
message for the first is:
The KeepAlive HTTP implementation is bugged in it's retry logic, it supports
reading from a file pointer, but doesn't support rewinding of the seek cursor
when it performs a retry. So it can happen that an upload fails for whatever
reason and will then 'hang' on the retry event.
The sequence of events that get triggered are:
- Upload file A, goes OK. Keep-Alive caches connection.
- Upload file B, fails due to (for example) failing Keep-Alive, but LFS file
pointer has been consumed for the upload and fd has been closed.
- Retry for file B starts, sets the Content-Length properly to the expected
file size, but since file pointer has been consumed no data will be uploaded,
causing the server to wait for the uploaded data until either client or
server reaches a timeout, making it seem as our mercurial process hangs.
This is just a stop-gap measure to prevent this behavior from blocking Mercurial
(LFS has retry logic). A proper solutions need to be build on top of this
stop-gap measure: for upload from file pointers, we should support fseek() on
the interface. Since we expect to consume the whole file always anyways, this
should be safe. This way we can seek back to the beginning on a retry.
I ported those two patches, and it works. But I see that `url._sendfile()` does
a rewind on `httpsendfile` objects[3], so maybe it's better to keep this all in
one place and avoid a second seek. We may still want the first FaceBook patch
as extra protection for this problem in general. The other two uses of
`httpsendfile` are in the wire protocol to upload bundles, and to upload
largefiles. Neither of these appear to use a worker, and I'm not sure why
workers seem to trigger this, or if this could have happened without a worker.
Since `httpsendfile` already has a `close()` method, that is dropped. That
class also explicitly says there's no `__len__` attribute, so that is removed
too. The override for `read()` is necessary to avoid the progressbar usage per
file.
[1] https://github.com/facebookexperimental/eden/commit/c350d6536d90c044c837abdd3675185644481469
[2] https://github.com/facebookexperimental/eden/commit/77f0d3fd0415e81b63e317e457af9c55c46103ee
[3] https://www.mercurial-scm.org/repo/hg/file/5.2.2/mercurial/url.py#l176
Differential Revision: https://phab.mercurial-scm.org/D7962
$ echo "[extensions]" >> $HGRCPATH
$ echo "mq=" >> $HGRCPATH
$ hg init a
$ cd a
$ echo 'base' > base
$ hg ci -Ambase -d '1 0'
adding base
$ hg qnew -d '1 0' pa
$ hg qnew -d '1 0' pb
$ hg qnew -d '1 0' pc
$ hg qdel
abort: qdelete requires at least one revision or patch name
[255]
$ hg qdel pc
abort: cannot delete applied patch pc
[255]
$ hg qpop
popping pc
now at: pb
Delete the same patch twice in one command (issue2427)
$ hg qdel pc pc
$ hg qseries
pa
pb
$ ls .hg/patches
pa
pb
series
status
$ hg qpop
popping pb
now at: pa
$ hg qdel -k 1
$ ls .hg/patches
pa
pb
series
status
$ hg qdel -r pa
patch pa finalized without changeset message
$ hg qapplied
$ hg log --template '{rev} {desc}\n'
1 [mq]: pa
0 base
$ hg qnew pd
$ hg qnew pe
$ hg qnew pf
$ hg qdel -r pe
abort: cannot delete revision 3 above applied patches
[255]
$ hg qdel -r qbase:pe
patch pd finalized without changeset message
patch pe finalized without changeset message
$ hg qapplied
pf
$ hg log --template '{rev} {desc}\n'
4 [mq]: pf
3 [mq]: pe
2 [mq]: pd
1 [mq]: pa
0 base
$ cd ..
$ hg init b
$ cd b
$ echo 'base' > base
$ hg ci -Ambase -d '1 0'
adding base
$ hg qfinish
abort: no revisions specified
[255]
$ hg qfinish -a
no patches applied
$ hg qnew -d '1 0' pa
$ hg qnew -d '1 0' pb
$ hg qnew pc # XXX fails to apply by /usr/bin/patch if we put a date
$ hg qfinish 0
abort: revision 0 is not managed
[255]
$ hg qfinish pb
abort: cannot delete revision 2 above applied patches
[255]
$ hg qpop
popping pc
now at: pb
$ hg qfinish -a pc
abort: unknown revision 'pc'!
[255]
$ hg qpush
applying pc
patch pc is empty
now at: pc
$ hg qfinish qbase:pb
patch pa finalized without changeset message
patch pb finalized without changeset message
$ hg qapplied
pc
$ hg log --template '{rev} {desc}\n'
3 imported patch pc
2 [mq]: pb
1 [mq]: pa
0 base
$ hg qfinish -a pc
patch pc finalized without changeset message
$ hg qapplied
$ hg log --template '{rev} {desc}\n'
3 imported patch pc
2 [mq]: pb
1 [mq]: pa
0 base
$ ls .hg/patches
series
status
qdel -k X && hg qimp -e X used to trigger spurious output with versioned queues
$ hg init --mq
$ hg qimport -r 3
$ hg qpop
popping imported_patch_pc
patch queue now empty
$ hg qdel -k imported_patch_pc
$ hg qimp -e imported_patch_pc
adding imported_patch_pc to series file
$ hg qfinish -a
no patches applied
resilience to inconsistency: qfinish -a with applied patches not in series
$ hg qser
imported_patch_pc
$ hg qapplied
$ hg qpush
applying imported_patch_pc
patch imported_patch_pc is empty
now at: imported_patch_pc
$ echo next >> base
$ hg qrefresh -d '1 0'
$ echo > .hg/patches/series # remove 3.diff from series to confuse mq
$ hg qfinish -a
revision 47dfa8501675 refers to unknown patches: imported_patch_pc
more complex state 'both known and unknown patches
$ echo hip >> base
$ hg qnew -f -d '1 0' -m 4 4.diff
$ echo hop >> base
$ hg qnew -f -d '1 0' -m 5 5.diff
$ echo > .hg/patches/series # remove 4.diff and 5.diff from series to confuse mq
$ echo hup >> base
$ hg qnew -f -d '1 0' -m 6 6.diff
$ echo pup > base
$ hg qfinish -a
warning: uncommitted changes in the working directory
revision 2b1c98802260 refers to unknown patches: 5.diff
revision 33a6861311c0 refers to unknown patches: 4.diff
$ cd ..