test_nanosleep is flaky under Valgrind

[pscollins]

Example:

make -C tests valgrind
make[1]: Entering directory `~/lkl-linux-master/tools/lkl/tests'
VALGRIND_CMD="valgrind --suppressions=valgrind.supp --leak-check=full --show-leak-kinds=all" ./boot.sh -t ext4
20480+0 records in
20480+0 records out
20971520 bytes (21 MB) copied, 0.0601454 s, 349 MB/s
mke2fs 1.42.9 (4-Feb-2014)
==51041== Memcheck, a memory error detector
==51041== Copyright (C) 2002-2013, and GNU GPL'd, by Julian Seward et al.
==51041== Using Valgrind-3.10.1 and LibVEX; rerun with -h for copyright info
==51041== Command: ./boot -d /tmp/tmp.WW4HVFKxn7 -t ext4 -n lkl_boot
==51041== 
disk_add             passed [3 0]
netdev_add           passed [0 4 0]
==51041== Thread 6:
...<snip>...
==51041== 
getpid               passed [1]
umask                passed [22 777]
creat                passed [0]
close                passed [0]
failopen             passed [-2]
open                 passed [0]
write                passed [5]
lseek                passed [0 ]
read                 passed [5 test]
fstat                passed [0 100721 5]
mkdir                passed [0]
stat                 passed [0 40721]
nanosleep            **failed** [97188672]
mount                passed [0]
chdir                passed [0]
opendir              passed [1]
getdents64           passed [1 lost+found .. . ]
umount               passed [0 0 0]
lo_ifup              passed [0]
netdev_ifup          passed [0 2]
==51041== 
==51041== HEAP SUMMARY:

Presumably because there's an upper bound on the time that passes while the test thread sleeps (under test_nanosleep in boot.c):

clock_gettime(CLOCK_MONOTONIC, &start);
ret = lkl_sys_nanosleep(&ts, NULL);
clock_gettime(CLOCK_MONOTONIC, &stop);

delta = 1e9*(stop.tv_sec - start.tv_sec) +
    (stop.tv_nsec - start.tv_nsec);

snprintf(str, len, "%ld", delta);

if (ret == 0 && delta > sleep_ns * 0.9 && delta < sleep_ns * 1.1)
    return 1;

return 0;

I'm not sure how important you consider this part of the test to be --- is there a very compelling reason to check that delta < sleep_ns * 1.1, or can that line be changed to "if (ret == 0 && delta > sleep_ns)" instead?

[ghost]

Good point, I agree that we should remove the upper bound check.