Re: [PATCH v1] fs/epoll: use a per-cpu counter for user's watches count

From: Randy Dunlap
Date: Wed Aug 04 2021 - 15:07:17 EST

On 8/4/21 8:22 AM, Guenter Roeck wrote:
On Mon, Aug 02, 2021 at 01:20:13PM +1000, Nicholas Piggin wrote:
This counter tracks the number of watches a user has, to compare against
the 'max_user_watches' limit. This causes a scalability bottleneck on
SPECjbb2015 on large systems as there is only one user. Changing to a
per-cpu counter increases throughput of the benchmark by about 30% on a
16-socket, > 1000 thread system.

Reported-by: Anton Blanchard <anton@xxxxxxxxxx>
Signed-off-by: Nicholas Piggin <npiggin@xxxxxxxxx>

With all tinyconfig builds (and all other builds with CONFIG_EPOLL=n),
this patch results in:

kernel/user.c: In function 'free_user':
kernel/user.c:141:35: error: 'struct user_struct' has no member named 'epoll_watches'
141 | percpu_counter_destroy(&up->epoll_watches);
| ^~
kernel/user.c: In function 'alloc_uid':
kernel/user.c:189:45: error: 'struct user_struct' has no member named 'epoll_watches'
189 | if (percpu_counter_init(&new->epoll_watches, 0, GFP_KERNEL)) {
| ^~
kernel/user.c:203:52: error: 'struct user_struct' has no member named 'epoll_watches'
203 | percpu_counter_destroy(&new->epoll_watches);
| ^~
kernel/user.c: In function 'uid_cache_init':
kernel/user.c:225:43: error: 'struct user_struct' has no member named 'epoll_watches'
225 | if (percpu_counter_init(&root_user.epoll_watches, 0, GFP_KERNEL))
| ^


Nick and I have also posted patches for this.