[redhat-lspp] LSPP kernel performance

[Steve Grubb]

On Sunday 26 March 2006 17:15, Steve Grubb wrote:
> I will probably make a new lspp test kernel with slab debug turned off to
> get that out of the way.

OK, re-ran the tests with the lspp.14 kernel. The results changed. Here's what 
I have for the noaudit case:

real    0m50.186s
user    0m1.516s
sys     0m48.639s

 12133 total                                      0.0051
  1720 __d_lookup                                 6.3469
  1509 __link_path_walk                           0.3995
  1443 avc_has_perm_noaudit                       1.6454
   548 _atomic_dec_and_lock                       6.5238
   530 do_path_lookup                             0.7105
   494 link_path_walk                             2.1572
   432 kmem_cache_free                            1.2781
   383 inode_has_perm                             3.8687
   382 dput                                       0.8843
   321 permission                                 1.8448
   320 generic_permission                         1.6000
   315 strncpy_from_user                          2.7155
   293 system_call                                2.2366
   283 do_lookup                                  0.7219
   269 kmem_cache_alloc                           1.3938
   263 memset                                     1.3698
   261 _raw_spin_lock                             1.1013
   245 sys_faccessat                              0.6171


But, with audit turned on:

real    2m3.903s
user    0m1.556s
sys     2m2.236s

 30514 total                                      0.0129
  2659 context_struct_to_string                   6.6809
  2341 vsnprintf                                  1.6234
  1894 __d_lookup                                 6.9889
  1564 __link_path_walk                           0.4141
  1501 kfree                                      3.8096
  1320 audit_filter_syscall                       7.7193
  1286 avc_has_perm_noaudit                       1.4664
  1047 _atomic_dec_and_lock                      12.4643
   922 dput                                       2.1343
   853 _raw_spin_lock                             3.5992
   696 do_path_lookup                             0.9330
   690 audit_syscall_exit                         0.6866
   634 __kmalloc                                  2.6979
   624 mls_sid_to_context                         1.0314
   604 strnlen                                   23.2308
   567 _raw_read_lock                             3.5886
   489 mls_compute_context_len                    1.3216
   468 sidtab_search                              6.6857
   456 selinux_inode_getsecurity                  6.8060
   456 audit_getname                              1.7538
   394 strcmp                                    15.1538
   375 memcmp                                    11.0294
   362 strncpy_from_user                          3.1207
   358 __audit_inode                              1.0848
   349 kmem_cache_alloc                           1.8083
   348 selinux_getsecurity                        3.7826
   342 inode_has_perm                             3.4545
   323 generic_permission                         1.6150
   322 kmem_cache_free                            0.9527
   320 audit_inode_context                        1.2598
   307 link_path_walk                             1.3406
   283 permission                                 1.6264
   280 spin_bug                                   1.2556
   274 sys_faccessat                              0.6902

That works out to a 146% performance hit. The culprit has not changed. The 
problem is the calls to context_struct_to_string. 

-Steve