On Wed, May 26, 2021 at 2:57 PM Pavel Begunkov <asml.silence@xxxxxxxxx> wrote:
> On 5/26/21 7:44 PM, Paul Moore wrote:
> > On Wed, May 26, 2021 at 2:01 PM Jens Axboe <axboe@xxxxxxxxx> wrote:
> >> On 5/26/21 11:54 AM, Jens Axboe wrote:
> >>> On 5/26/21 11:31 AM, Jens Axboe wrote:
> >>>> On 5/26/21 11:15 AM, Jens Axboe wrote:
> >>>>> On 5/25/21 8:04 PM, Paul Moore wrote:
> >>>>>> On Tue, May 25, 2021 at 9:11 PM Jens Axboe <axboe@xxxxxxxxx> wrote:
> >>>>>>> On 5/24/21 1:59 PM, Paul Moore wrote:
> >>>>>>>> That said, audit is not for everyone, and we have build time and
> >>>>>>>> runtime options to help make life easier. Beyond simply disabling
> >>>>>>>> audit at compile time a number of Linux distributions effectively
> >>>>>>>> shortcut audit at runtime by adding a "never" rule to the audit
> >>>>>>>> filter, for example:
> >>>>>>>>
> >>>>>>>> % auditctl -a task,never
> >>>>>>>
> >>>>>>> As has been brought up, the issue we're facing is that distros have
> >>>>>>> CONFIG_AUDIT=y and hence the above is the best real world case outside
> >>>>>>> of people doing custom kernels. My question would then be how much
> >>>>>>> overhead the above will add, considering it's an entry/exit call per op.
> >>>>>>> If auditctl is turned off, what is the expectation in turns of overhead?
> >>>>>>
> >>>>>> I commented on that case in my last email to Pavel, but I'll try to go
> >>>>>> over it again in a little more detail.
> >>>>>>
> >>>>>> As we discussed earlier in this thread, we can skip the req->opcode
> >>>>>> check before both the _entry and _exit calls, so we are left with just
> >>>>>> the bare audit calls in the io_uring code. As the _entry and _exit
> >>>>>> functions are small, I've copied them and their supporting functions
> >>>>>> below and I'll try to explain what would happen in CONFIG_AUDIT=y,
> >>>>>> "task,never" case.
> >>>>>>
> >>>>>> + static inline struct audit_context *audit_context(void)
> >>>>>> + {
> >>>>>> + return current->audit_context;
> >>>>>> + }
> >>>>>>
> >>>>>> + static inline bool audit_dummy_context(void)
> >>>>>> + {
> >>>>>> + void *p = audit_context();
> >>>>>> + return !p || *(int *)p;
> >>>>>> + }
> >>>>>>
> >>>>>> + static inline void audit_uring_entry(u8 op)
> >>>>>> + {
> >>>>>> + if (unlikely(audit_enabled && audit_context()))
> >>>>>> + __audit_uring_entry(op);
> >>>>>> + }
> >>>>>>
> >>>>>> We have one if statement where the conditional checks on two
> >>>>>> individual conditions. The first (audit_enabled) is simply a check to
> >>>>>> see if anyone has "turned on" auditing at runtime; historically this
> >>>>>> worked rather well, and still does in a number of places, but ever
> >>>>>> since systemd has taken to forcing audit on regardless of the admin's
> >>>>>> audit configuration it is less useful. The second (audit_context())
> >>>>>> is a check to see if an audit_context has been allocated for the
> >>>>>> current task. In the case of "task,never" current->audit_context will
> >>>>>> be NULL (see audit_alloc()) and the __audit_uring_entry() slowpath
> >>>>>> will never be called.
> >>>>>>
> >>>>>> Worst case here is checking the value of audit_enabled and
> >>>>>> current->audit_context. Depending on which you think is more likely
> >>>>>> we can change the order of the check so that the
> >>>>>> current->audit_context check is first if you feel that is more likely
> >>>>>> to be NULL than audit_enabled is to be false (it may be that way now).
> >>>>>>
> >>>>>> + static inline void audit_uring_exit(int success, long code)
> >>>>>> + {
> >>>>>> + if (unlikely(!audit_dummy_context()))
> >>>>>> + __audit_uring_exit(success, code);
> >>>>>> + }
> >>>>>>
> >>>>>> The exit call is very similar to the entry call, but in the
> >>>>>> "task,never" case it is very simple as the first check to be performed
> >>>>>> is the current->audit_context check which we know to be NULL. The
> >>>>>> __audit_uring_exit() slowpath will never be called.
> >>>>>
> >>>>> I actually ran some numbers this morning. The test base is 5.13+, and
> >>>>> CONFIG_AUDIT=y and CONFIG_AUDITSYSCALL=y is set for both the baseline
> >>>>> test and the test with this series applied. I used your git branch as of
> >>>>> this morning.
> >>>>>
> >>>>> The test case is my usual peak perf test, which is random reads at
> >>>>> QD=128 and using polled IO. It's a single core test, not threaded. I ran
> >>>>> two different tests - one was having a thread just do the IO, the other
> >>>>> is using SQPOLL to do the IO for us. The device is capable than more
> >>>>> IOPS than a single core can deliver, so we're CPU limited in this test.
> >>>>> Hence it's a good test case as it does actual work, and shows software
> >>>>> overhead quite nicely. Runs are very stable (less than 0.5% difference
> >>>>> between runs on the same base), yet I did average 4 runs.
> >>>>>
> >>>>> Kernel SQPOLL IOPS Perf diff
> >>>>> ---------------------------------------------------------
> >>>>> 5.13 0 3029872 0.0%
> >>>>> 5.13 1 3031056 0.0%
> >>>>> 5.13 + audit 0 2894160 -4.5%
> >>>>> 5.13 + audit 1 2886168 -4.8%
> >>>>>
> >>>>> That's an immediate drop in perf of almost 5%. Looking at a quick
> >>>>> profile of it (nothing fancy, just checking for 'audit' in the profile)
> >>>>> shows this:
> >>>>>
> >>>>> + 2.17% io_uring [kernel.vmlinux] [k] __audit_uring_entry
> >>>>> + 0.71% io_uring [kernel.vmlinux] [k] __audit_uring_exit
> >>>>> 0.07% io_uring [kernel.vmlinux] [k] __audit_syscall_entry
> >>>>> 0.02% io_uring [kernel.vmlinux] [k] __audit_syscall_exit
> >>>>>
> >>>>> Note that this is with _no_ rules!
> >>>>
> >>>> io_uring also supports a NOP command, which basically just measures
> >>>> reqs/sec through the interface. Ran that as well:
> >>>>
> >>>> Kernel SQPOLL IOPS Perf diff
> >>>> ---------------------------------------------------------
> >>>> 5.13 0 31.05M 0.0%
> >>>> 5.13 + audit 0 25.31M -18.5%
> >>>>
> >>>> and profile for the latter includes:
> >>>>
> >>>> + 5.19% io_uring [kernel.vmlinux] [k] __audit_uring_entry
> >>>> + 4.31% io_uring [kernel.vmlinux] [k] __audit_uring_exit
> >>>> 0.26% io_uring [kernel.vmlinux] [k] __audit_syscall_entry
> >>>> 0.08% io_uring [kernel.vmlinux] [k] __audit_syscall_exit
> >>>
> >>> As Pavel correctly pointed it, looks like auditing is enabled. And
> >>> indeed it was! Hence the above numbers is without having turned off
> >>> auditing. Running the NOPs after having turned off audit, we get 30.6M
> >>> IOPS, which is down about 1.5% from the baseline. The results for the
> >>> polled random read test above did _not_ change from this, they are still
> >>> down the same amount.
> >>>
> >>> Note, and I should have included this in the first email, this is not
> >>> any kind of argument for or against audit logging. It's purely meant to
> >>> be a set of numbers that show how the current series impacts
> >>> performance.
> >>
> >> And finally, just checking if we make it optional per opcode if we see
> >> any real impact, and the answer is no. Using the below patch which
> >> effectively bypasses audit calls unless the opcode has flagged the need
> >> to do so, I cannot measure any difference in perf (as expected).
> >>
> >> To turn this into something useful, my suggestion as a viable path
> >> forward would be:
> >>
> >> 1) Use something like the below patch and flag request types that we
> >> want to do audit logging for.
> >>
> >> 2) As Pavel suggested, eliminate the need for having both and entry/exit
> >> hook, turning it into just one. That effectively cuts the number of
> >> checks and calls in half.
> >
> > I suspect the updated working-io_uring branch with HEAD at
> > 1f25193a3f54 (updated a short time ago, see my last email in this
> > thread) will improve performance. Also, as has been mention several
>
> See the email you replied to, ~1.5% was basically an overhead of
> two `if (io_op_defs[req->opcode].audit)` in case of nops, where at
> least once req->opcode is cached. But to be completely fair, misses
> unlikely
Maybe. I remain skeptical that "io_op_defs[req->opcode].audit" has
the same cost as "unlikely(audit_context())".
--
paul moore
www.paul-moore.com
