Hello.
Took the time to re-run my performance tests on the current LSM patch
for io_uring_cmd. I'll explain what I did and then give the results:
How I ran it:
I took a version of the kernel with the patch a0d2212773d1 and then
compiled two versions: The first was the vanilla kernel and the other
was the same except for the LSM hook called from io_uring_cmd removed.
Same kernel configurations. For my tests I used one of the test files
from FIO called t/io_uring.c which is basically a READ test. I ran my
tests on both an nvme device and the null device (/dev/null). For the
first I did not change io_uring.c and for the second I replaced the
admin calls with dummy data that was not really needed for testing with
/dev/null. These are the arguments I used for the test
"./t/io_uring -b4096 -d128 -c32 -s32 -p0 -F1 -B0 -O0 -P1 -u1 -n1"
Finally, I'm taking the max of several samples.
Results:
+-------------------------------------+------------------------+-----------------------+
| Name | /dev/ng0n1 (BW: MiB/s) | /dev/null (BW: GiB/s) |
+-------------------------------------+------------------------+-----------------------+
| (A) for-next (vanilla) | 1341 | 30.16 |
| (B) for-next (no io_uring_cmd hook) | 1362 | 40.61 |
| [1-(A/B)] * 100 | 1.54185022 | 25.732578183 |
+-------------------------------------+------------------------+-----------------------+
So on a device (dev/ng0n1) there is a 1% performance difference on a
read. Whereas on the null device (dev/null) there is a 25% difference on
a read.
This difference is interesting and expected as there is a lot more stuff
happening when we go through the actual device.
Best
Joel
On Wed, Sep 07, 2022 at 10:17:29AM +0200, Joel Granados wrote:
> Hey Paul
>
> On Thu, Sep 01, 2022 at 05:30:38PM -0400, Paul Moore wrote:
> > On Thu, Sep 1, 2022 at 4:15 PM Joel Granados <j.granados@xxxxxxxxxxx> wrote:
> > > Hey Paul
> > >
> > > I realize that you have already sent this upstream but I wanted to share
> > > the Selinux part of the testing that we did to see if there is any
> > > feedback.
> > >
> > > With my tests I see that the selinux_uring_cmd hook is run and it
> > > results in a "avc : denied" when I run it with selinux in permissive
> > > mode with an unpriviledged user. I assume that this is the expected
> > > behavior. Here is how I tested
> > >
> > > *** With the patch:
> > > * I ran the io_uring_passthrough.c test on a char device with an
> > > unpriviledged user.
> > > * I took care of changing the permissions of /dev/ng0n1 to 666 prior
> > > to any testing.
> > > * made sure that Selinux was in permissive mode.
> > > * Made sure to have audit activated by passing "audit=1" to the kernel
> > > * After noticing that some audit messages where getting lost I upped the
> > > backlog limit to 256
> > > * Prior to executing the test, I also placed a breakpoint inside
> > > selinux_uring_cmd to make sure that it was executed.
> > > * This is the output of the audit when I executed the test:
> > >
> > > [ 136.615924] audit: type=1400 audit(1662043624.701:94): avc: denied { create } for pid=263 comm="io_uring_passth" anonclass=[io_uring] scontext=system_u:system_r:kernel_t tcontext=system_u:object_r:kernel_t tclass=anon_inode permissive=1
> > > [ 136.621036] audit: type=1300 audit(1662043624.701:94): arch=c000003e syscall=425 success=yes exit=3 a0=40 a1=7ffca29835a0 a2=7ffca29835a0 a3=561529be2300 items=0 ppid=252 pid=263 auid=1001 uid=1001 gid=1002 euid=1001 suid=1001 fsuid=1001 egid=1002 sgid=1002 fsgid=1002 tty=pts1 ses=3 comm="io_uring_passth" exe="/mnt/src/liburing/test/io_uring_passthrough.t" subj=system_u:system_r:kernel_t key=(null)
> > > [ 136.624812] audit: type=1327 audit(1662043624.701:94): proctitle=2F6D6E742F7372632F6C69627572696E672F746573742F696F5F7572696E675F706173737468726F7567682E74002F6465762F6E67306E31
> > > [ 136.626074] audit: type=1400 audit(1662043624.702:95): avc: denied { map } for pid=263 comm="io_uring_passth" path="anon_inode:[io_uring]" dev="anon_inodefs" ino=11715 scontext=system_u:system_r:kernel_t tcontext=system_u:object_r:kernel_t tclass=anon_inode permissive=1
> > > [ 136.628012] audit: type=1400 audit(1662043624.702:95): avc: denied { read write } for pid=263 comm="io_uring_passth" path="anon_inode:[io_uring]" dev="anon_inodefs" ino=11715 scontext=system_u:system_r:kernel_t tcontext=system_u:object_r:kernel_t tclass=anon_inode permissive=1
> > > [ 136.629873] audit: type=1300 audit(1662043624.702:95): arch=c000003e syscall=9 success=yes exit=140179765297152 a0=0 a1=1380 a2=3 a3=8001 items=0 ppid=252 pid=263 auid=1001 uid=1001 gid=1002 euid=1001 suid=1001 fsuid=1001 egid=1002 sgid=1002 fsgid=1002 tty=pts1 ses=3 comm="io_uring_passth" exe="/mnt/src/liburing/test/io_uring_passthrough.t" subj=system_u:system_r:kernel_t key=(null)
> > > [ 136.632415] audit: type=1327 audit(1662043624.702:95): proctitle=2F6D6E742F7372632F6C69627572696E672F746573742F696F5F7572696E675F706173737468726F7567682E74002F6465762F6E67306E31
> > > [ 136.633652] audit: type=1400 audit(1662043624.705:96): avc: denied { cmd } for pid=263 comm="io_uring_passth" path="/dev/ng0n1" dev="devtmpfs" ino=120 scontext=system_u:system_r:kernel_t tcontext=system_u:object_r:device_t tclass=io_uring permissive=1
> > > [ 136.635384] audit: type=1336 audit(1662043624.705:96): uring_op=46 items=0 ppid=252 pid=263 uid=1001 gid=1002 euid=1001 suid=1001 fsuid=1001 egid=1002 sgid=1002 fsgid=1002 subj=system_u:system_r:kernel_t key=(null)
> > > [ 136.636863] audit: type=1336 audit(1662043624.705:96): uring_op=46 items=0 ppid=252 pid=263 uid=1001 gid=1002 euid=1001 suid=1001 fsuid=1001 egid=1002 sgid=1002 fsgid=1002 subj=system_u:system_r:kernel_t key=(null)
> > >
> > > * From the output on time 136.633652 I see that the access should have
> > > been denied had selinux been enforcing.
> > > * I also saw that the breakpoint hit.
> > >
> > > *** Without the patch:
> > > * I ran the io_uring_passthrough.c test on a char device with an
> > > unpriviledged user.
> > > * I took care of changing the permissions of /dev/ng0n1 to 666 prior
> > > to any testing.
> > > * made sure that Selinux was in permissive mode.
> > > * Made sure to have audit activated by passing "audit=1" to the kernel
> > > * After noticing that some audit messages where getting lost I upped the
> > > backlog limit to 256
> > > * There were no audit messages when I executed the test.
> > >
> > > As with my smack tests I would really appreciate feecback on the
> > > approach I took to testing and it's validity.
> >
> > Hi Joel,
> >
> > Thanks for the additional testing and verification! Work like this is
> > always welcome, regardless if the patch has already been merged
> > upstream.
> np
>
> >
> > As far as you test approach is concerned, I think you are on the right
> > track, I might suggest resolving the other SELinux/AVC denials you are
> > seeing with your test application to help reduce the noise in the
> > logs. Are you familiar with the selinux-testsuite (link below)?
> >
> > * https://protect2.fireeye.com/v1/url?k=6f356c96-0ebe79ac-6f34e7d9-74fe4860008a-01002a6e4c92bb3e&q=1&e=46f33488-9311-49fa-9747-da210f2d147d&u=https%3A%2F%2Fgithub.com%2FSELinuxProject%2Fselinux-testsuite
> Thx. Could not figure out how to remove the AVC from a quick look at the
> page, but I'll probably figures something out :).
>
> ATM, I'm doing a performance test on the io_uring_passtrhough
> path to see how much, if any, perf we loose.
>
> Thx again
>
> Best
>
> Joel
>
> >
> > --
> > paul-moore.com
Attachment:
signature.asc
Description: PGP signature
