nfs-ganesha server crashes regularly

[Ponnuvel Palaniyappan]

nfs-ganesha server crashes regularly.

It doesn't happen all the time or easily reproducible. But when it does crash, the backtrace looks like:

(gdb) bt
#0 atomic_postclear_uint16_t_bits (bits=<optimized out>, var=<optimized out>) at ./ntirpc/misc/abstract_atomic.h:1812
#1 svc_rqst_epoll_event (sr_rec=sr_rec@entry=0x564dc852efd8, ev=0x7f6134002900) at ./src/svc_rqst.c:1416
#2 0x00007f620fa76565 in svc_rqst_epoll_events (n_events=2, sr_rec=0x564dc852efd8) at ./src/svc_rqst.c:1466
#3 svc_rqst_epoll_loop (wpe=0x564dc852efd8) at ./src/svc_rqst.c:1566
#4 0x00007f620fa816d6 in work_pool_thread (arg=0x7f6064002280) at ./src/work_pool.c:184
#5 0x00007f621029a6db in start_thread (arg=0x7f6097cfa700) at pthread_create.c:463
#6 0x00007f620fdbb61f in clone () at ../sysdeps/unix/sysv/linux/x86_64/clone.S:95

This was using 3.0.3 on Bionic (available via Ubuntu Cloud Archive packages).

Upstream nfs-ganesha developers suggested that a "number of fixes" related to libntirpc fixed what looks like a race condition.

libntirpc is a submodule used in nfs-ganesha and it's where the problem comes form:
https://github.com/nfs-ganesha/ntirpc

There were a number of commits that went in since 3.0 [0]. Given the crash isn't reproducible easily, it's not straightforward to find the commits that fixed the issue between 3.0.3 and 3.5 for a potential SRU.

In a user environment where the problem occurred, they were able to test nfs-ganesha 3.5 and confirmed that it didn't crash over several days load test whereas 3.0.3 crashed at least once a day under a similar load/test environment.

[0] https://github.com/nfs-ganesha/ntirpc
[1] https://github.com/nfs-ganesha/ntirpc/commit/1da6533431a23af7406b5961d4b16ef61045b6af

[Chris MacNaughton (chris.macnaughton)]

In Ubuntu, nfs-ganesha pulls ntirpc from the libntirpc-dev package, rather than using the version vendored into nfs-ganesha upstream