Re: fio jobs die with sigsegv if --filesize=1tb

[Jens Axboe <jens.axboe@oracle.com>]

On Tue, Dec 22 2009, Case van Rij wrote:
> fairly basic random write test over nfs, 1 job, directio enabled,
> pre-existing 1TB file results in SIGSEGV,
> 
> tested on: CentOS 5.4 x86_64,  2.6.18-164.6.1.el5 kernel, fio from
> git, last change: Tue, 22 Dec 2009 08:06:43 +0000
> 
> fio --name=rndwrs --ioengine=libaio --iodepth=4 --rw=randwrite
> --bs=32k --direct=1 --size=1tb --numjobs=1 --filename=/mnt/nfs/1tb.vdb
> random-writers: (g=0): rw=randwrite, bs=32K-32K/32K-32K,
> ioengine=libaio, iodepth=4
> Starting 1 process
> fio: pid=24153, got signal=11
> 
> Run status group 0 (all jobs):
> fio: file hash not empty on exit
> 
> strace:
> [pid 23961] open("/sys/devices/system/cpu/cpu0/cache/index0/coherency_line_size",
> O_RDONLY) = 8
> [pid 23961] read(8, "64\n", 32)         = 3
> [pid 23961] close(8)                    = 0
> [pid 23961] getpriority(PRIO_PROCESS, 0) = 20
> [pid 23961] setpriority(PRIO_PROCESS, 0, 0) = 0
> [pid 23961] getpriority(PRIO_PROCESS, 0) = 20
> [pid 23961] io_setup(4, {47657422389248}) = 0
> [pid 23961] getrusage(RUSAGE_SELF, {ru_utime={0, 1999}, ru_stime={0,
> 0}, ...}) = 0
> [pid 23961] open("/mnt/nfs/1tb.vdb", O_RDWR|O_CREAT|O_DIRECT, 0600) = 8
> [pid 23961] fadvise64(8, 0, 1, POSIX_FADV_DONTNEED) = 0
> [pid 23961] fadvise64(8, 0, 1, POSIX_FADV_RANDOM) = 0
> [pid 23961] --- SIGSEGV (Segmentation fault) @ 0 (0) ---
> 
> but no core file, since reap_threads cleans up after the sigsegv (?)
> 
> with debug=all:
> io       26954 invalidate cache /mnt/nfs/1tb.vdb: 0/1
> file     26954 goodf=1, badf=2, ff=31
> file     26954 get_next_file_rr: 0x2abe73d38028
> file     26954 get_next_file: 0x2abe73d38028 [/mnt/nfs/1tb.vdb]
> file     26954 get file /mnt/nfs/1tb.vdb, ref=1
> random   26954 off rand 1425201762
> random   26954 free: b=12242389915983151104, idx=536870912, bit=0
> fio: pid=26954, got signal=11
> process  26952 pid=26954: runstate 4 -> 9
> process  26952 terminate group_id=-1
> process  26952 setting terminate on random-writers/26954
> diskutil    26952 update io ticks
> 
> the same test works if I replace --filesize=1tb with --filesize=1gb
> (but makes for a far less interesting test).

Looks like math overflow. Can you double check that ulimit -c is set
reasonably high (I usually just do ulimit -c1000000000), then remove the
-O2 from the fio makefile and recompile, then trigger the problem. That
should give you a clean core dump, invoke gdb with fio and that core
file so we can see exactly where it bombs.

-- 
Jens Axboe