2516fde1fa
When running will-it-scale[1] open2_process testcase, in a system with a large number of cores, a bottleneck in retrieving the current task secid was detected: 27.73% ima_file_check;do_open (inlined);path_openat;do_filp_open;do_sys_openat2;__x64_sys_openat;do_syscall_x64 (inlined);do_syscall_64;entry_SYSCALL_64_after_hwframe (inlined);__libc_open64 (inlined) 27.72% 0.01% [kernel.vmlinux] [k] security_current_getsecid_subj - - 27.71% security_current_getsecid_subj;ima_file_check;do_open (inlined);path_openat;do_filp_open;do_sys_openat2;__x64_sys_openat;do_syscall_x64 (inlined);do_syscall_64;entry_SYSCALL_64_after_hwframe (inlined);__libc_open64 (inlined) 27.71% 27.68% [kernel.vmlinux] [k] apparmor_current_getsecid_subj - - 19.94% __refcount_add (inlined);__refcount_inc (inlined);refcount_inc (inlined);kref_get (inlined);aa_get_label (inlined);aa_get_label (inlined);aa_get_current_label (inlined);apparmor_current_getsecid_subj;security_current_getsecid_subj;ima_file_check;do_open (inlined);path_openat;do_filp_open;do_sys_openat2;__x64_sys_openat;do_syscall_x64 (inlined);do_syscall_64;entry_SYSCALL_64_after_hwframe (inlined);__libc_open64 (inlined) 7.72% __refcount_sub_and_test (inlined);__refcount_dec_and_test (inlined);refcount_dec_and_test (inlined);kref_put (inlined);aa_put_label (inlined);aa_put_label (inlined);apparmor_current_getsecid_subj;security_current_getsecid_subj;ima_file_check;do_open (inlined);path_openat;do_filp_open;do_sys_openat2;__x64_sys_openat;do_syscall_x64 (inlined);do_syscall_64;entry_SYSCALL_64_after_hwframe (inlined);__libc_open64 (inlined) A large amount of time was spent in the refcount. The most common case is that the current task label is available, and no need to take references for that one. That is exactly what the critical section helpers do, make use of them. New perf output: 39.12% vfs_open;path_openat;do_filp_open;do_sys_openat2;__x64_sys_openat;do_syscall_64;entry_SYSCALL_64_after_hwframe;__libc_open64 (inlined) 39.07% 0.13% [kernel.vmlinux] [k] do_dentry_open - - 39.05% do_dentry_open;vfs_open;path_openat;do_filp_open;do_sys_openat2;__x64_sys_openat;do_syscall_64;entry_SYSCALL_64_after_hwframe;__libc_open64 (inlined) 38.71% 0.01% [kernel.vmlinux] [k] security_file_open - - 38.70% security_file_open;do_dentry_open;vfs_open;path_openat;do_filp_open;do_sys_openat2;__x64_sys_openat;do_syscall_64;entry_SYSCALL_64_after_hwframe;__libc_open64 (inlined) 38.65% 38.60% [kernel.vmlinux] [k] apparmor_file_open - - 38.65% apparmor_file_open;security_file_open;do_dentry_open;vfs_open;path_openat;do_filp_open;do_sys_openat2;__x64_sys_openat;do_syscall_64;entry_SYSCALL_64_after_hwframe;__libc_open64 (inlined) The result is a throughput improvement of around 20% across the board on the open2 testcase. On more realistic workloads the impact should be much less. [1] https://github.com/antonblanchard/will-it-scale Signed-off-by: Vinicius Costa Gomes <vinicius.gomes@intel.com> Signed-off-by: John Johansen <john.johansen@canonical.com> |
||
---|---|---|
.. | ||
apparmor | ||
bpf | ||
integrity | ||
keys | ||
landlock | ||
loadpin | ||
lockdown | ||
safesetid | ||
selinux | ||
smack | ||
tomoyo | ||
yama | ||
commoncap.c | ||
device_cgroup.c | ||
inode.c | ||
Kconfig | ||
Kconfig.hardening | ||
lsm_audit.c | ||
Makefile | ||
min_addr.c | ||
security.c |