Sluggishness on a GEM system due to SELinux's sidtab_search_context() as well as get_unmapped_area()

Sluggishness on a GEM system due to SELinux's sidtab_search_context() as well as get_unmapped_area()

[r6144]

Hello,

I'm noticing significant sluggishness when switching into a workspace
where Evolution is running.

My kernel is Fedora 12's 2.6.32.16-150.fc12.x86_64 (I know it is
old...), and the open source radeon (r600) driver is used.  Oprofile
shows the follows:

CPU: Core 2, speed 2003 MHz (estimated)
Counted CPU_CLK_UNHALTED events (Clock cycles when not halted) with a
unit mask of 0x00 (Unhalted core cycles) count 100000
Counted INST_RETIRED_ANY_P events (number of instructions retired) with
a unit mask of 0x00 (No unit mask) count 100000
samples  %        samples  %        image name               app name
symbol name
126807   31.2201  38744    28.3866  vmlinux                  Xorg
(deleted)           find_vma
48661    11.9804  3474      2.5453  vmlinux                  Xorg
(deleted)           mls_compute_sid
41806    10.2927  5700      4.1762  vmlinux                  Xorg
(deleted)           sidtab_search_context
11888     2.9268  4234      3.1021  Xorg (deleted)           Xorg
(deleted)           /usr/bin/Xorg (deleted)
8660      2.1321  1589      1.1642  drm                      Xorg
(deleted)           /drm
6228      1.5333  6773      4.9624  radeon                   Xorg
(deleted)           /radeon
6214      1.5299  1832      1.3423  libc-2.11.2.so (deleted) Xorg
(deleted)           /lib64/libc-2.11.2.so (deleted)
5542      1.3644  4929      3.6113
libpixman-1.so.0.16.6.#prelink#.3I5wow (deleted) Xorg
(deleted)           /usr/lib64/libpixman-1.so.0.16.6.#prelink#.3I5wow
(deleted)
4149      1.0215  1776      1.3012  libexa.so                Xorg
(deleted)           /usr/lib64/xorg/modules/libexa.so
4140      1.0193  640       0.4689  vmlinux                  oprofiled
mls_compute_sid
3183      0.7837  847       0.6206  vmlinux                  Xorg
(deleted)           arch_get_unmapped_area_topdown
...

Although I haven't looked into it, presumably Evolution is asking the X
server to create and map a lot of TTM objects (pixmaps?).  The creation
of each object uses shmem_file_setup() and thus necessitates SELinux
calls, and since the X process already has a lot of mappings (one
mapping for each XRender glyph it seems, with "wc -l /proc/`pgrep
Xorg`/maps" returning 6766), the mapping part could also be slow.

I think the slowness of find_vma is related to
https://bugzilla.kernel.org/show_bug.cgi?id=17531 .  mls_compute_sid()
did a linear search over the range transition rules and was therefore
slow, but more recent kernels use a hash table so this problem has been
solved.  Another offender is sidtab_search_context() used to convert a
SELinux context back into an sid, which is also a linear search as there
is only a sid => context hash table, and I haven't seen any recent
changes in this area.

Please CC me as I'm not subscribed.

r6144

Re: Sluggishness on a GEM system due to SELinux's sidtab_search_context() as well as get_unmapped_area()

[Andi Kleen]

r6144 <rainy6144@gmail.com> writes:


> Hello,
>
> I'm noticing significant sluggishness when switching into a workspace
> where Evolution is running.


Interesting result. Adding some relevant ccs with full-quote.
I guess graphics really should do less mmaps, but the underlying
performance problems should be investigated too.

-Andi

>
> My kernel is Fedora 12's 2.6.32.16-150.fc12.x86_64 (I know it is
> old...), and the open source radeon (r600) driver is used.  Oprofile
> shows the follows:
>
> CPU: Core 2, speed 2003 MHz (estimated)
> Counted CPU_CLK_UNHALTED events (Clock cycles when not halted) with a
> unit mask of 0x00 (Unhalted core cycles) count 100000
> Counted INST_RETIRED_ANY_P events (number of instructions retired) with
> a unit mask of 0x00 (No unit mask) count 100000
> samples  %        samples  %        image name               app name
> symbol name
> 126807   31.2201  38744    28.3866  vmlinux                  Xorg
> (deleted)           find_vma
> 48661    11.9804  3474      2.5453  vmlinux                  Xorg
> (deleted)           mls_compute_sid
> 41806    10.2927  5700      4.1762  vmlinux                  Xorg
> (deleted)           sidtab_search_context
> 11888     2.9268  4234      3.1021  Xorg (deleted)           Xorg
> (deleted)           /usr/bin/Xorg (deleted)
> 8660      2.1321  1589      1.1642  drm                      Xorg
> (deleted)           /drm
> 6228      1.5333  6773      4.9624  radeon                   Xorg
> (deleted)           /radeon
> 6214      1.5299  1832      1.3423  libc-2.11.2.so (deleted) Xorg
> (deleted)           /lib64/libc-2.11.2.so (deleted)
> 5542      1.3644  4929      3.6113
> libpixman-1.so.0.16.6.#prelink#.3I5wow (deleted) Xorg
> (deleted)           /usr/lib64/libpixman-1.so.0.16.6.#prelink#.3I5wow
> (deleted)
> 4149      1.0215  1776      1.3012  libexa.so                Xorg
> (deleted)           /usr/lib64/xorg/modules/libexa.so
> 4140      1.0193  640       0.4689  vmlinux                  oprofiled
> mls_compute_sid
> 3183      0.7837  847       0.6206  vmlinux                  Xorg
> (deleted)           arch_get_unmapped_area_topdown
> ...
>
> Although I haven't looked into it, presumably Evolution is asking the X
> server to create and map a lot of TTM objects (pixmaps?).  The creation
> of each object uses shmem_file_setup() and thus necessitates SELinux
> calls, and since the X process already has a lot of mappings (one
> mapping for each XRender glyph it seems, with "wc -l /proc/`pgrep
> Xorg`/maps" returning 6766), the mapping part could also be slow.
>
> I think the slowness of find_vma is related to
> https://bugzilla.kernel.org/show_bug.cgi?id=17531 .  mls_compute_sid()
> did a linear search over the range transition rules and was therefore
> slow, but more recent kernels use a hash table so this problem has been
> solved.  Another offender is sidtab_search_context() used to convert a
> SELinux context back into an sid, which is also a linear search as there
> is only a sid => context hash table, and I haven't seen any recent
> changes in this area.
>
> Please CC me as I'm not subscribed.
>
> r6144
>

-- 
ak@linux.intel.com -- Speaking for myself only.