Simplfying copy_siginfo_to_user

[Eric W. Biederman]

ebiederm@xxxxxxxxxxxx (Eric W. Biederman) writes:

> Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> writes:
>
>> On Tue, Jul 18, 2017 at 7:06 AM, Eric W. Biederman
>> <ebiederm@xxxxxxxxxxxx> wrote:
>>> struct siginfo is a union and the kernel since 2.4 has been hiding a union
>>> tag in the high 16bits of si_code using the values:
>>> __SI_KILL
>>> __SI_TIMER
>>> __SI_POLL
>>> __SI_FAULT
>>> __SI_CHLD
>>> __SI_RT
>>> __SI_MESGQ
>>> __SI_SYS
>>>
>>> While this looks plausible on the surface, in practice this situation has
>>> not worked well.
>>
>> So on the whole I think we just need to do this, but the part I really
>> hate about this series is still this the siginfo_layout() part.
>>
>> I can well believe that it is needed for the compat case. siginfo is a
>> piece of crap crazy type, and re-ordering fields for compat is
>> something we are always going to have to do.
>>
>> But for the native case, the *only* reason we do not just copy the
>> siginfo as-is seems to be that it's just too big, due to other bad
>> design decisions in siginfo ("let's make sure it's big enough by
>> allocating 512 bytes for it).
>>
>> And afaik, absolutely nobody uses more than about 36 bytes of that
>> 512-byte _sifields union (and that one use is SIGILL with three
>> pointers and three integers and some padding.
>>
>> So why don't we just say "screw this idiotic layout crap, and just
>> unconditionally copy that much smaller maximum of bytes"?
>>
>> Leave that layout thing purely for compat handling.
>
> I completely agree.

So I just did some measurements to see what the performance impact is of
doing the simple and obvious thing of always copying the entire siginfo
around.  There is a fair amount of variation in my timings but for
the whole change I see about a 20ns increase in time taken to send
a signal with siginfo from the current process to the current process.
AKA timing kill(getpid(),...). 

I played with some clever changes such as limiting the copy to 48 bytes,
disabling the memset and the like but I could not get a strong enough
signal to say that any one change removed the extra or a clear part of
it 20ns.

Do we care about those 20ns for signal deliver?  I suspect from my
previous numbers that if Andy can get signal delivery to use sysret
it will more than make up for the small increase in cost here.

Eric