mirror of
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git
synced 2024-09-03 17:40:04 +00:00
![]() Right now a global per-fs rwsem (kernfs_rwsem) synchronizes multiple kernfs operations. On a large system with few hundred CPUs and few hundred applications simultaneoulsy trying to access sysfs, this results in multiple sys_open(s) contending on kernfs_rwsem via kernfs_iop_permission and kernfs_dop_revalidate. For example on a system with 384 cores, if I run 200 instances of an application which is mostly executing the following loop: for (int loop = 0; loop <100 ; loop++) { for (int port_num = 1; port_num < 2; port_num++) { for (int gid_index = 0; gid_index < 254; gid_index++ ) { char ret_buf[64], ret_buf_lo[64]; char gid_file_path[1024]; int ret_len; int ret_fd; ssize_t ret_rd; ub4 i, saved_errno; memset(ret_buf, 0, sizeof(ret_buf)); memset(gid_file_path, 0, sizeof(gid_file_path)); ret_len = snprintf(gid_file_path, sizeof(gid_file_path), "/sys/class/infiniband/%s/ports/%d/gids/%d", dev_name, port_num, gid_index); ret_fd = open(gid_file_path, O_RDONLY | O_CLOEXEC); if (ret_fd < 0) { printf("Failed to open %s\n", gid_file_path); continue; } /* Read the GID */ ret_rd = read(ret_fd, ret_buf, 40); if (ret_rd == -1) { printf("Failed to read from file %s, errno: %u\n", gid_file_path, saved_errno); continue; } close(ret_fd); } } I see contention around kernfs_rwsem as follows: path_openat | |----link_path_walk.part.0.constprop.0 | | | |--49.92%--inode_permission | | | | | --48.69%--kernfs_iop_permission | | | | | |--18.16%--down_read | | | | | |--15.38%--up_read | | | | | --14.58%--_raw_spin_lock | | | | | ----- | | | |--29.08%--walk_component | | | | | --29.02%--lookup_fast | | | | | |--24.26%--kernfs_dop_revalidate | | | | | | | |--14.97%--down_read | | | | | | | --9.01%--up_read | | | | | --4.74%--__d_lookup | | | | | --4.64%--_raw_spin_lock | | | | | ---- Having a separate per-fs rwsem to protect kernfs inode attributes, will avoid the above mentioned contention and result in better performance as can bee seen below: path_openat | |----link_path_walk.part.0.constprop.0 | | | | | |--27.06%--inode_permission | | | | | --25.84%--kernfs_iop_permission | | | | | |--9.29%--up_read | | | | | |--8.19%--down_read | | | | | --7.89%--_raw_spin_lock | | | | | ---- | | | |--22.42%--walk_component | | | | | --22.36%--lookup_fast | | | | | |--16.07%--__d_lookup | | | | | | | --16.01%--_raw_spin_lock | | | | | | | ---- | | | | | --6.28%--kernfs_dop_revalidate | | | | | |--3.76%--down_read | | | | | --2.26%--up_read As can be seen from the above data the overhead due to both kerfs_iop_permission and kernfs_dop_revalidate have gone down and this also reduces overall run time of the earlier mentioned loop. Signed-off-by: Imran Khan <imran.f.khan@oracle.com> Link: https://lore.kernel.org/r/20230309110932.2889010-2-imran.f.khan@oracle.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org> |
||
---|---|---|
.. | ||
dir.c | ||
file.c | ||
inode.c | ||
Kconfig | ||
kernfs-internal.h | ||
Makefile | ||
mount.c | ||
symlink.c |