Reduce contention around global locks used in kernfs.
The current kernfs design makes use of 3 global locks to synchronize
various operations. There are a few other global locks as well but
they are used for specific cases and hence don't cause severe contention.
The above mentioned 3 main global locks used in kernfs are:
1. A global mutex, kernfs_open_file_mutex, to protect the list of
kernfs_open_file instances correspondng to a sysfs attribute.
2. A global spinlock, kernfs_open_node_lock, to protect
kernfs_node->attr.open which points to kernfs_open_node instance
corresponding to a kernfs_node.
3. A global per-fs rw semaphore, root->kernfs_rwsem, to synchronize most
of the other operations like adding, removing, renaming etc. of a file
or directory.
Since these locks are global, they can cause contention when multiple
(for example few hundred) applications try to access sysfs (or other kernfs
based file system) files in parallel, even if the applications are
accessing different and unrelated files.
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 can see contention around above mentioned locks as follows:
- 54.07% 53.60% showgids [kernel.kallsyms] [k] osq_lock
- 53.60% __libc_start_main
- 32.29% __GI___libc_open
entry_SYSCALL_64_after_hwframe
do_syscall_64
sys_open
do_sys_open
do_filp_open
path_openat
vfs_open
do_dentry_open
kernfs_fop_open
mutex_lock
- __mutex_lock_slowpath
- 32.23% __mutex_lock.isra.5
osq_lock
- 21.31% __GI___libc_close
entry_SYSCALL_64_after_hwframe
do_syscall_64
exit_to_usermode_loop
task_work_run
____fput
__fput
kernfs_fop_release
kernfs_put_open_node.isra.8
mutex_lock
- __mutex_lock_slowpath
- 21.28% __mutex_lock.isra.5
osq_lock
- 10.49% 10.39% showgids [kernel.kallsyms] [k] down_read
10.39% __libc_start_main
__GI___libc_open
entry_SYSCALL_64_after_hwframe
do_syscall_64
sys_open
do_sys_open
do_filp_open
- path_openat
- 9.72% link_path_walk
- 5.21% inode_permission
- __inode_permission
- 5.21% kernfs_iop_permission
down_read
- 4.08% walk_component
lookup_fast
- d_revalidate.part.24
- 4.08% kernfs_dop_revalidate
- 7.48% 7.41% showgids [kernel.kallsyms] [k] up_read
7.41% __libc_start_main
__GI___libc_open
entry_SYSCALL_64_after_hwframe
do_syscall_64
sys_open
do_sys_open
do_filp_open
- path_openat
- 7.01% link_path_walk
- 4.12% inode_permission
- __inode_permission
- 4.12% kernfs_iop_permission
up_read
- 2.61% walk_component
lookup_fast
- d_revalidate.part.24
- 2.61% kernfs_dop_revalidate
Moreover this run of 200 application isntances takes 32-34 secs. to
complete.
This patch set is reducing the above mentioned contention by replacing
these global locks with hashed locks.
For example with the patched kernel and on the same test setup, we no
longer see contention around osq_lock (i.e kernfs_open_file_mutex) and also
contention around per-fs kernfs_rwsem has reduced significantly as well.
This can be seen in the following perf snippet:
- 1.66% 1.65% showgids [kernel.kallsyms] [k] down_read
1.65% __libc_start_main
__GI___libc_open
entry_SYSCALL_64_after_hwframe
do_syscall_64
sys_open
do_sys_open
do_filp_open
- path_openat
- 1.62% link_path_walk
- 0.98% inode_permission
- __inode_permission
+ 0.98% kernfs_iop_permission
- 0.52% walk_component
lookup_fast
- d_revalidate.part.24
- 0.52% kernfs_dop_revalidate
- 1.12% 1.11% showgids [kernel.kallsyms] [k] up_read
1.11% __libc_start_main
__GI___libc_open
entry_SYSCALL_64_after_hwframe
do_syscall_64
sys_open
do_sys_open
do_filp_open
- path_openat
- 1.11% link_path_walk
- 0.69% inode_permission
- __inode_permission
- 0.69% kernfs_iop_permission
up_read
Moreover the test execution time has reduced from 32-34 secs to 18-19 secs.
The patches of this patchset introduce following chanages:
PATCH-1: Make global kernfs_open_file_mutex and kernfs_open_node_lock
per-fs hashed locks, where address of a kernfs_node acts as hash key.
This results in kernfs_node objects, whose address give the
different hash value, using different kernfs_open_file_mutex
and kernfs_open_node_lock rather than all kernfs_node objects
using the same kernfs_open_file_mutex and kernfs_open_node_lock
as was the case earlier.
PATCH-2: Replace per-fs single rw semaphore with per-fs hashed semaphore,
where address of a kernfs_node acts as hash key to reduce contention
around single per-fs rw semaphore.
------------------------------------------------------------------
Changes since v3:
- Make open_file_mutex and open_node_lock per-fs.
- Replace per-fs rwsem with per-fs hashed rwsem.
(Suggested by Tejun Heo <[email protected]>)
Imran Khan (2):
kernfs: use hashed mutex and spinlock in place of global ones.
kernfs: Replace per-fs global rwsem with per-fs hashed rwsem.
fs/kernfs/dir.c | 273 +++++++++++++++++++++++++-----------
fs/kernfs/file.c | 67 ++++-----
fs/kernfs/inode.c | 18 ++-
fs/kernfs/kernfs-internal.h | 163 +++++++++++++++++++++
fs/kernfs/mount.c | 13 +-
fs/kernfs/symlink.c | 5 +-
include/linux/kernfs.h | 44 +++++-
7 files changed, 447 insertions(+), 136 deletions(-)
base-commit: a70bf4a85b43cb952bd39dd948b103b1b3eb2cf8
--
2.30.2