Re: Deadlock between git-remote-http and git fetch-pack

[Jonathan Tan <jonathantanmy@google.com>]

On 01/27/2017 02:31 PM, tsuna wrote:
> Hi there,
> While investigating a hung job in our CI system today, I think I found
> a deadlock in git-remote-http
>
> Git version: 2.9.3
> Linux (amd64) kernel 4.9.0
>
> Excerpt from the process list:
>
> jenkins  27316  0.0  0.0  18508  6024 ?        S    19:30   0:00  |
>                \_ git -C ../../../arista fetch --unshallow
> jenkins  27317  0.0  0.0 169608 10916 ?        S    19:30   0:00  |
>                    \_ git-remote-http origin http://gerrit/arista
> jenkins  27319  0.0  0.0  24160  8260 ?        S    19:30   0:00  |
>                        \_ git fetch-pack --stateless-rpc --stdin
> --lock-pack --include-tag --thin --no-progress --depth=2147483647
> http://gerrit/arista/
>
> Here PID 27319 (git fetch-pack) is stuck reading on stdin, while its
> parent, PID 27317 (git-remote-http) is stuck reading on its child’s
> stdout.  Nothing has moved for like 2h, it’s deadlocked.
>
>> strace -fp 27319
> strace: Process 27319 attached
> read(0,
>
> Here FD 0 is a pipe:
>
> ~ @8a33a534e2f7> lsof -np 27319 | grep 0r
> git     27319 jenkins    0r  FIFO   0,10      0t0 354519158 pipe
>
> The writing end of which is owned by the parent process:
>
> ~ @8a33a534e2f7> lsof -n 2>/dev/null | fgrep 354519158
> git-remot 27317                jenkins    4w     FIFO   0,10      0t0
> 354519158 pipe
> git       27319                jenkins    0r     FIFO   0,10      0t0
> 354519158 pipe
>
> And the parent process (git-remote-http) is stuck reading from another FD:
>
>> strace -fp 27317
> strace: Process 27317 attached
> read(5,
>
> And here FD 5 is another pipe:
>
> ~ @8a33a534e2f7> lsof -np 27317 | grep 5r
> git-remot 27317 jenkins    5r  FIFO   0,10      0t0 354519159 pipe
>
> Which is the child’s stdout:
>
>> lsof -n 2>/dev/null | fgrep 354519159
> git-remot 27317                jenkins    5r     FIFO   0,10      0t0
> 354519159 pipe
> git       27319                jenkins    1w     FIFO   0,10      0t0
> 354519159 pipe
>
> Hence the deadlock.
>
> Stack trace in git-remote-http:
>
> (gdb) bt
> #0  0x00007f04f1e1363d in read () from target:/lib64/libpthread.so.0
> #1  0x0000562417472d73 in xread ()
> #2  0x0000562417472f2b in read_in_full ()
> #3  0x0000562417438a6e in get_packet_data ()
> #4  0x0000562417439129 in packet_read ()
> #5  0x00005624174245e0 in rpc_service ()
> #6  0x0000562417424f10 in fetch_git ()
> #7  0x00005624174233fd in main ()
>
> Stack trace in git fetch-pack:
>
> (gdb) bt
> #0  0x00007fb3ab478620 in __read_nocancel () from target:/lib64/libpthread.so.0
> #1  0x000055f688827283 in xread ()
> #2  0x000055f68882743b in read_in_full ()
> #3  0x000055f6887ce35e in get_packet_data ()
> #4  0x000055f6887cea19 in packet_read ()
> #5  0x000055f6887ceb90 in packet_read_line ()
> #6  0x000055f68879dd05 in get_ack ()
> #7  0x000055f68879f6b4 in fetch_pack ()
> #8  0x000055f688710619 in cmd_fetch_pack ()
> #9  0x000055f6886dff7b in handle_builtin ()
> #10 0x000055f6886df026 in main ()
>
> I looked at the diff between v2.9.3 and HEAD on fetch-pack.c and
> remote-curl.c and didn’t see anything noteworthy in that area of the
> code, so I presume the bug is still there in master.
>

I haven't looked into this in detail, but this might be related to 
something I discovered while writing my patch set. I noticed that 
upload-pack (the process on the "other side" of fetch-pack) can die 
without first writing any notification, causing fetch-pack to block 
forever on a read. A fix would probably look like that patch [1].

[1] 
<afe5d7d3f876893fdad318665805df1e056717c6.1485381677.git.jonathantanmy@google.com>