[RFC PATCH 2/9] audit,io_uring,io-wq: add some basic audit support to io_uring

Jens Axboe axboe at kernel.dk
Wed May 26 17:15:24 UTC 2021

On 5/25/21 8:04 PM, Paul Moore wrote:
> On Tue, May 25, 2021 at 9:11 PM Jens Axboe <axboe at kernel.dk> 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!

>> aio never had any audit logging as far as I can tell. I think it'd make
>> a lot more sense to selectively enable audit logging only for opcodes
>> that we care about. File open/create/unlink/mkdir etc, that kind of
>> thing. File level operations that people would care about logging. Would
>> they care about logging a buffer registration or a polled read from a
>> device/file? I highly doubt it, and we don't do that for alternative
>> methods either. Doesn't really make sense for a lot of the other
>> operations, imho.
> We would need to check with the current security requirements (there
> are distro people on the linux-audit list that keep track of that
> stuff), but looking at the opcodes right now my gut feeling is that
> most of the opcodes would be considered "security relevant" so
> selective auditing might not be that useful in practice.  It would
> definitely clutter the code and increase the chances that new opcodes
> would not be properly audited when they are merged.

We don't audit read/write from aio, as mentioned. In the past two
decades, I take it that hasn't been a concern? I agree that some opcodes
should _definitely_ be audited. Things like opening a file, closing a
file, removing/creating a file, mount, etc. But normal read/write, I
think that's just utter noise and not useful at all. Auditing on a
per-opcode basis is trivial, io_uring already has provisions for
flagging opcode requirements and this would just be another one.

Jens Axboe

More information about the Linux-security-module-archive mailing list