To repeat what was posted on LU-3319.
Test conf-sanity 28 fails 4% of time with the following error.
LustreError: 18595:0:(lproc_llite.c:367:ll_max_read_ahead_whole_mb_seq_write()) can't set max_read_ahead_whole_mb more than max_read_ahead_per_file_mb: 0
LustreError: 18595:0:(obd_config.c:1443:class_process_proc_seq_param()) writing proc entry max_read_ahead_whole_mb err -34
Looking at the code this means the variable ra_max_pages is being set to zero. Looking at llite_lib.c
we see that
pages = si.totalram - si.totalhigh;
...
sbi->ll_ra_info.ra_max_pages_per_file = min(pages / 32,
SBI_DEFAULT_READAHEAD_MAX);
so for some reason si_meminfo() is reporting no free memory.
Possible idea is for test 28 to set max_readahead_per_file_mb before the test start since this is the reason for the test fails.
meminfo can report 0 memory if memory is exhausted. Memory is a limited resource and the kernel keeps the last little bit protected. Are there any other signs of memory pressure on the system?
Is there any correlation between the test failures and running on SLES clients? I'm wondering if there is any chance that this could be related to seq_file changes somehow?
This is showing up with an earlier patch for 7290. Waiting to see if the new patch has this problem still.
This problems seems to have gone away. We can close this ticket if that is the case but can be reopen if the problem appears again.
Does this problem still happen? If not can we close this ticket.
llite proc bug no longer shows up. Peter you can close this ticket.
As you wish 