[Cluster-devel] [PATCH] Fixing an invalid return error code, which leads to a panic in 64 bits.

[Mathieu Avila]

Hello GFS folks,

Sometimes, i get a panic at mount time, following messages saying it
cannot start the threads:
lock_gulm: ERROR Faild to send lock login. -111
lock_gulm: ERROR glq_startup failed. 111
lock_gulm: ERROR Got a 111 trying to start the threads.

A few lines later, a panic occurs in "do_kern_mount", because the
return value of "fill_super" is interpreted as a pointer and
dereferenced, instead of being interpreted as an error code.

This is due to the way "fs/super.c" interprets pointer return values,
using "ERR_PTR(ptr)", defined in include/linux/err.h
To detect that a pointer is in fact an error code, the return value
must be negative, at least on 64 bits.
returning a positive error code leads to considering the value as a
pointer, not an error code and the panic follows.

The first time the error is reported in GULM, it is
negative, then it is changed to a positive value. The patch doesn't
change it to a positive value.



diff src/gulm/gulm_lock_queue.c src-b/gulm/gulm_lock_queue.c
813c813
<               return -err;
---
>               return err;


Please tell me if this is the good way to fix this bug,

Thanks in advance,

--
Mathieu Avila