Revision 4df910620bebb5cfe234af16ac8f6474b60215fd authored by Feng Tang on 25 November 2020, 05:22:21 UTC, committed by Linus Torvalds on 26 November 2020, 17:35:49 UTC
0day reported one -22.7% regression for will-it-scale page_fault2 case [1] on a 4 sockets 144 CPU platform, and bisected to it to be caused by Waiman's optimization (commit bd0b230fe1) of saving one 'struct page_counter' space for 'struct mem_cgroup'. Initially we thought it was due to the cache alignment change introduced by the patch, but further debug shows that it is due to some hot data members ('vmstats_local', 'vmstats_percpu', 'vmstats') sit in 2 adjacent cacheline (2N and 2N+1 cacheline), and when adjacent cache line prefetch is enabled, it triggers an "extended level" of cache false sharing for 2 adjacent cache lines. So exchange the 2 member blocks, while keeping mostly the original cache alignment, which can restore and even enhance the performance, and save 64 bytes of space for 'struct mem_cgroup' (from 2880 to 2816, with 0day's default RHEL-8.3 kernel config) [1]. https://lore.kernel.org/lkml/20201102091543.GM31092@shao2-debian/ Fixes: bd0b230fe145 ("mm/memcg: unify swap and memsw page counters") Reported-by: kernel test robot <rong.a.chen@intel.com> Signed-off-by: Feng Tang <feng.tang@intel.com> Acked-by: Waiman Long <longman@redhat.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
1 parent fa02fcd
File | Mode | Size |
---|---|---|
hci | ||
nci | ||
Kconfig | -rw-r--r-- | 845 bytes |
Makefile | -rw-r--r-- | 374 bytes |
af_nfc.c | -rw-r--r-- | 1.8 KB |
core.c | -rw-r--r-- | 24.2 KB |
digital.h | -rw-r--r-- | 5.2 KB |
digital_core.c | -rw-r--r-- | 19.8 KB |
digital_dep.c | -rw-r--r-- | 34.2 KB |
digital_technology.c | -rw-r--r-- | 26.4 KB |
llcp.h | -rw-r--r-- | 6.7 KB |
llcp_commands.c | -rw-r--r-- | 16.2 KB |
llcp_core.c | -rw-r--r-- | 34.0 KB |
llcp_sock.c | -rw-r--r-- | 22.0 KB |
netlink.c | -rw-r--r-- | 39.4 KB |
nfc.h | -rw-r--r-- | 4.4 KB |
rawsock.c | -rw-r--r-- | 8.9 KB |
Computing file changes ...