On 05/15/2015 18:39, Leonid Yegoshin wrote: > On 05/15/2015 02:53 PM, Ralf Baechle wrote: >> On Thu, May 14, 2015 at 06:34:43PM -0700, Leonid Yegoshin wrote: >> >> The order 1 allocation for the PGD are concerning me a little. On a >> system under even moderate memory pressure that might become a bit of >> a reliability or performance issue. >> >> With 4kB pages we already need order 1 or even 2 allocations for the >> allocation of the stack and some folks have reported that to be an issue >> so we may have to start using the PUD for very large VA spaces. >> >> Ralf > > I don't think it is an issue here - people, who wants to exercise 256 TERABAIT > of memory PER PROCESS may even doesn't note that they have PGD = 2 pages. It is > definitely not for systems with 4GB physmemory. > > I also recommend for low memory to look into CONFIG_COMPACTION, it may be a > great help for them here, look into mm/vmscan.c, in_reclaim_compaction(). > > Besides that, I defined this feature for 16KB and 64KB pages only, not for 4KB. There's something screwy with R14000 CPUs and 16kb/64kb pages. I haven't figured it out yet, but you get random SIGSEGV and SIGBUS errors running that PAGE_SIZE. I figure it's some unknown/undocumented bit that SGI added and never disclosed anywhere. --J
