mirror of
https://github.com/facebook/sapling.git
synced 2024-10-11 09:17:30 +03:00
2c46c59ad6
Summary:
We have encountered cases where `eden health` reported
`"edenfs not healthy: edenfs not running"` even though the `edenfs` process is
still running. Because the existing implementation of `eden health` bases its
health check on the output of a `getStatus()` Thrift call, it will erroneously
report `"edenfs not running"` even if Eden is running but its Thrift server is
not running. This type of false negative could occur if `edenfs` has shutdown
the Thrift server, but not the rest of the process (quite possibly, its
shutdown is blocked on calls to `umount2()`).
This is further problematic because `eden daemon` checks `eden health`
before attempting to start the daemon. If it gets a false negative, then
`eden daemon` will forge ahead, trying to launch a new instance of the daemon,
but it will fail with a nasty error like the following:
```
I1017 11:59:25.188414 3064499 main.cpp:81] Starting edenfs. UID=5256, GID=100, PID=3064499
terminate called after throwing an instance of 'std::runtime_error'
what(): another instance of Eden appears to be running for /home/mbolin/local/.eden
*** Aborted at 1508266765 (Unix time, try 'date -d 1508266765') ***
*** Signal 6 (SIGABRT) (0x1488002ec2b3) received by PID 3064499 (pthread TID 0x7fd0d3787d40) (linux TID 3064499) (maybe from PID 30644
99, UID 5256), stack trace: ***
@ 000000000290d3cd folly::symbolizer::(anonymous namespace)::signalHandler(int, siginfo_t*, void*)
@ 00007fd0d133cacf (unknown)
@ 00007fd0d093e7c8 __GI_raise
@ 00007fd0d0940590 __GI_abort
@ 00007fd0d1dfeecc __gnu_cxx::__verbose_terminate_handler()
@ 00007fd0d1dfcdc5 __cxxabiv1::__terminate(void (*)())
@ 00007fd0d1dfce10 std::terminate()
@ 00007fd0d1dfd090 __cxa_throw
@ 00000000015fe8ca facebook::eden::EdenServer::acquireEdenLock()
@ 000000000160f27b facebook::eden::EdenServer::prepare()
@ 00000000016107d5 facebook::eden::EdenServer::run()
@ 000000000042c4ee main
@ 00007fd0d0929857 __libc_start_main
@ 0000000000548ad8 _start
Aborted
```
By providing more accurate information to `eden daemon`, if the user tries to
run it while the daemon is already running, they will get a more polite error
like the following:
```
error: edenfs is already running (pid 274205)
```
This revision addresses this issue by writing the PID of `edenfs` in the
lockfile. It updated the implementation of `eden health` to use the PID in the
lockfile to assess the health of Eden if the call to `getStatus()` fails. It
does this by running:
```
ps -p PID -o comm=
```
and applying some heuristics on the output to assess whether the command
associated with that process is the `edenfs` command. If it is, then
`eden health` reports the status as `STOPPED` whereas previously it would report
it as `DEAD`.
Reviewed By: wez
Differential Revision: D6086473
fbshipit-source-id: 825421a6818b56ddd7deea257a92c070c2232bdd
|
||
|---|---|---|
| .. | ||
| test | ||
| cmd_util.py | ||
| config.py | ||
| configinterpolator.py | ||
| debug.py | ||
| main.py | ||
| rage.py | ||
| stats_print.py | ||
| stats.py | ||
| TARGETS | ||
| util.py | ||