Re: [PATCH 3/7] seccomp_filter: Enable ftrace-based system call filtering

[Will Drewry]

On Thu, Apr 28, 2011 at 11:13 AM, Frederic Weisbecker
<fweisbec@xxxxxxxxx> wrote:
> On Thu, Apr 28, 2011 at 10:29:11AM -0500, Will Drewry wrote:
>> On Thu, Apr 28, 2011 at 10:12 AM, Frederic Weisbecker
>> <fweisbec@xxxxxxxxx> wrote:
>> > Instead of having such multiline filter definition with syscall
>> > names prepended, it would be nicer to make the parsing simplier.
>> >
>> > You could have either:
>> >
>> >        prctl(PR_SET_SECCOMP, mode);
>> >        /* Works only if we are in mode 2 */
>> >        prctl(PR_SET_SECCOMP_FILTER, syscall_nr, filter);
>>
>> It'd need to be syscall_name instead of syscall_nr.  Otherwise we're
>> right back to where Adam's patch was 2+ years ago :)  Using the event
>> names from the syscalls infrastructure means the consumer of the
>> interface doesn't need to be confident of the syscall number.
>
> Is it really a problem?

I'd prefer using numbers myself since it gives more flexibility around
filtering entries that haven't yet made it into ftrace syscalls hooks.
However, I think there's some complexity in ensuring it matches the
host kernel.  (It would also allow compat_task support which doesn't
seem to be possible now.)

> There are libraries that can resolve that. Of course I can't recall their name.

asm-generic/unistd.h will provide the mapping, if available. It would
mean that any helper libraries would need to be matched to the locally
running kernel.

If syscall names and numbers are supported, then it would mean that
this change could do both what Adam's original bitmask patch did, by
default, and when CONFIG_FTRACE_SYSCALLS are enabled, it could do name
resolution and apply more complex filters based on the syscalls
events.

Would that make sense?  I wonder if the inconsistency between
sometimes using names and filters and sometimes using numbers and 0|1
would become onerous if userspace applications started using this.
Maybe something like
  prctl(PR_SECCOMP_HAS_FILTERS);
would be desirable?  I dunno.

>> > or:
>> >        /*
>> >         * If mode == 2, set the filter to syscall_nr
>> >         * Recall this for each syscall that need a filter.
>> >         * If a filter was previously set on the targeted syscall,
>> >         * it will be overwritten.
>> >         */
>> >        prctl(PR_SET_SECCOMP, mode, syscall_nr, filter);
>> >
>> > One can erase a previous filter by setting the new filter "1".
>> >
>> > Also, instead of having a bitmap of syscall to accept. You could
>> > simply set "0" as a filter to those you want to deactivate:
>> >
>> > prctl(PR_SET_SECCOMP, 2, 1, 0); <- deactivate the syscall_nr 1
>> >
>> > Hm?
>>
>> I like the simplicity in not needing to parse anything extra, but it
>> does add the need for extra state - either a bit or a new field - to
>> represent "enabled/enforcing".
>
> And by the way I'm really puzzled about these. I don't understand
> well why we need this.

Right now, if you are in seccomp mode !=0 , your system calls will be
hooked (if TIF_SECCOMP is set too). The current patch begins filtering
immediately.  And once filtering is enabled, the process is not
allowed to expand its access to the kernel system cals - only shrink
it.  An "enabled" bit would be needed to tell it that even though it
is running in mode=2, it should/shouldn't kill the process for system
call filter violations and it should allow filters to be added, not
just dropped.  It's why the current patch does it in one step: set the
filters and the mode.  While an enabled bit could be hidden behind
whether TIF_SECCOMP is applied, I'm not sure if that would just be
confusing.  There'd still be a need to APPLY_FILTERS to get the
TIF_SECCOMP flag set to start hooking the system calls.

> As for the enable_on_next_syscall. The documentation says
> it's useful if you want the filter to only apply to the
> child. So if fork immediately follows, you will be able to fork
> but if the child doesn't have the right to exec, it won't be able
> to do so. Same for the mmap() that involves...
>
> So I'm a bit confused about that.

Here's an example:
  http://static.dataspill.org/seccomp_filter/v1/seccomp_launcher.c

You'd use it by calling fork() ... prctl(..., ON_NEXT_SYSCALL) then
execve since you'll already have fork()d.  It seems to work, but maybe
I'm missing something :)

> But yeah if that's really needed, it looks to me better to
> reduce the parsing and cut it that way:
>
>        prctl(PR_SET_SECCOMP, 2, syscall_name_or_nr, filter);
>        prctl(PR_SECCOMP_APPLY_FILTERS, enable_on_next_syscall?)
>
> or something...

Cool - if there are any other flags, it could be something like:
  prctl(PR_SECCOMP_SET_FLAGS, SECCOMP_ENFORCE|SECCOMP_DELAYED_ENFORCEMENT);

But only if there are other flags worth considering. I had a few
others in mind, but now I've completely forgotten :/

>>
>> The only way to do it without a third mode would be to take a
>> blacklist model - where all syscalls are allowed by default and the
>> caller has to enumerate them all and drop them. That would definitely
>> not be the right approach :)
>>
>> If a new bit of state was added,  it could be used as:
>>   prctl(PR_SET_SECCOMP, 2);
>>   prctl(PR_SET_SECCOMP, 2, "sys_read", "fd == 1");  /* add a read filter */
>>   prctl(PR_SET_SECCOMP, 2, "sys_write", "fd == 0");  /* add a read filter */
>>   ...
>>   prctl(PR_SET_SECCOMP, 2, "sys_read", "0");  /* clear the sys_read
>> filters and block it */  (or NULL?)
>>   prctl(PR_SET_SECCOMP, 2, "enable");  /* Start enforcing */
>>   prctl(PR_SET_SECCOMP, 2, "sys_write", "0");  /* Reduce attack
>> surface on the fly */
>>
>>
>> As to the "0" filter instead of a bitmask, would it make sense to just
>> cut over to an hlist now and drop the bitmask?
>> It looks like perf
>> uses that model, and I'd hope it wouldn't incur too much additional
>> overhead.  (The linked list approach now is certainly not scalable for
>> a large number of filters!)
>
> The linked list certainly doesn't scale there. But either a hlist for
> everything, or a hlist + bitmap to check if the syscall is enabled,
> why not. May be start with a pure hlist for any filters and if performance
> issues that really matter are pinpointed, then move the "1" and "0"
> implementation to a bitmap.
>
> My guess is that doesn't really matter. If it's "1" then you can
> just have an empty set of filters for the syscall and it goes ahead
> quickly. If it's "0" then the app fails, which is not what I would
> call a fast path.

Sounds reasonable to me. I'll start to iterate on the suggested
changes and see what emerges.

>> If that interface seems sane, I can certainly start exploring it and
>> see if I hit any surprises (and put it in the next version of the
>> patch :).  I think it'll simplify a fair amount of the add/drop code!
>
> Yup.

Thanks!
will
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/