svn commit: r330285 - head/sys/sys

[Konstantin Belousov]

On Sat, Mar 03, 2018 at 01:47:41PM +1100, Bruce Evans wrote:
> On Fri, 2 Mar 2018, Konstantin Belousov wrote:
> 
> > On Fri, Mar 02, 2018 at 12:43:34PM -0500, Pedro Giffuni wrote:
> >> ...
> >> I think use of _Nonnull attributes in the threading functions may also
> >> be a waste (I introduced them mostly to be compatible with Android).
> >> FWIW, Dragonfly sprinkled some restrict there recently:
> >>
> >> http://gitweb.dragonflybsd.org/dragonfly.git/commit/d33005aaee6af52c80428b59b52aee522c002492
> >>
> >> Just in case someone is considering more cleanups.
> >
> > This is not a cleanup for me, but a needed change. Right now x86
> > copyouts are implemented in asm, so whatever damage is done to the
> > prototypes, only effect is at the caller side. In my work, i386 copyouts
> > are done in C, so it starts matter.
> 
> That seems slow, especially for small sizes as are common for syscall args
> (in 1 of my versions, copyin() of args is optimized to fuword() in a loop,
> and fuword() is optimized to not use pcb_onfault, so it is not much more
> than 1 memory access.  However, in your i386 version this optimization
> would be negative since the slow part is switching the map, so fuword()
> should never be used to access multiple words).
Yes. I already explained it in private, the current choice for i386 is
either to be neglected very fast, or to get this change to still be a
Tier 1 32 bit platform. The change is to make 4/4g split for UVA/KVA.
In particular, the change ensures that it is possible to self-host i386
for forthcoming years, which is not practical for armv7 now and would be
less so with clang grow.

In other news, my system already boots single-user on SMP machine and
I have torture tests like setting invalid %ss segment by sigreturn(2),
work.  There is (much) more to come, but I am happy how the patch
progressed so far.

> However, copyinstr() and
> copystr() should never have been "optimized" by writing them in asm.  On
> x86, their asm is badly written so they are slower than simple C versions
> except on 8088's and maybe 8086's and maybe on the original i386.  (8088's
> were limited mainly by instruction bandwidth and the original i386 wasn't
> much better, so short CISC instructions like lodsb and stosb tended to be
> faster than larger separate instructions despite their large setup overheads.
Sure, copyinstr() is rewritten in C.  The current version of copyout stuff
is there:
https://kib.kiev.ua/git/gitweb.cgi?p=deviant2.git;a=blob;f=sys/i386/i386/copyout.c;h=9747c06a84d7d2b5faac946f5de57f6a34d96c8c;hb=refs/heads/pcid

> 
> > Also I looked at the dragonfly commit because I become curious what do you
> > mean by threading functions.  The first example was
> > int    pthread_attr_getguardsize(const pthread_attr_t * __restrict,
> > -                       size_t *);
> > +           size_t * __restrict);
> > POSIX agrees with the dragonfly change, but I do not understand it.
> > Aliasing rules already disallow the first and second arguments to point
> > to the same memory, because they have different types.
> 
> (1) thread_attr_t is opaque, so the types might be the same.
> (2) pthread_attr_t might be a pointer to a struct/union containing a size_t.
> (3) perhaps other reasons.  I'm not sure how 'restrict interacts with global
>      variables or even it it prevents the interaction in (2).  A previous
>      discussion showed that const doesn't make types different enough to
>      prevent aliasing.  Similarly for volatile.
> 
> Similarly for other pointers to {opaque, struct/union, or even integer} types.
> size_t can't be aliased to int, but it can be aliased to any unsigned type
> in C and to any unsigned type not smaller than uint16_t in POSIX (POSIX
> but not C requires u_char == uint8_t, so size_t can't be u_char in POSIX
> but it can be u_char in C).

I can only summarize it as 'there is no use to have restrict on the
pthread_attr_getguardsize() arguments'.