amd64/169927: siginfo, si_code for fpe errors when error occurs
using the SSE math processor
brde at optusnet.com.au
Wed Jul 18 06:59:49 UTC 2012
On Wed, 18 Jul 2012, Konstantin Belousov wrote:
> On Wed, Jul 18, 2012 at 02:36:16PM +1000, Bruce Evans wrote:
>> On Wed, 18 Jul 2012, Konstantin Belousov wrote:
>>> The curpcb access shall be spelled as PCPU_GET(curpcb). Please note that
>>> compiler is not allowed to cache the accesses, since there is __asm
>>> __volatile expression to indirect through %gs-prefixed read.
>> Fix curpcb then. curthread was de-pessimized to use __pure2 and to not
>> use __volatile. I could never this to work to cache curthread when I
>> ried it, but it apparently works in -current.
> curpcb() could be implemented like this for amd64 only:
Similarly for i386?
> diff --git a/sys/amd64/include/pcb.h b/sys/amd64/include/pcb.h
> index 22cbbe2..3417c52 100644
> --- a/sys/amd64/include/pcb.h
> +++ b/sys/amd64/include/pcb.h
> @@ -144,6 +144,17 @@ void makectx(struct trapframe *, struct pcb *);
> int savectx(struct pcb *) __returns_twice;
> void resumectx(struct pcb *);
> +static __inline __pure2 struct pcb *
> + struct pcb *pcb;
> + __asm("movq %%gs:%1,%0" : "=r" (pcb)
> + : "i" (offsetof(struct pcpu, pc_curpcb)));
pc_curpcb is only a pointer, so there is no need for pcb.h to be used
(any more than proc.h is neaded for __curthread()).
I think this can go in machine/pcpu.h next to __curthread(). Or
machine/cpuinfo can provide the an alternative to PCPU_GET() without
__volatile and sys/pcpu.h can do this.
__curthread() avoids a dependency on sys/pcpu.h by hard-coding the
offset of pc_curthread as 0. This hack seems to be unnecessary,
Including machine/pcpu.h without going through sys/pcpu.h should be an
error. This error is only in a few low tier files:
- i386/xen/pcpu.h: this bogusly includes machine/pcpu.h after already
- xen/evtchn.h. xen code is horrible. In i386/include/xen/xen_os.h,
as part of this file's namespace pollution, it claims to include
sys/time.h for pcpu.h. But actually sys/time.h only provides moderate
namespace pollution that doesn't include pcpu.h. It is another style
bug to include sys/time.h directly. sys/time.h is standard namespace
pollution in sys/param.h in the _KERNEL case.
So amd64 can apparently safely put this function in machine/pcpu.h.
i386 might have to untangle the xen pollution. Other arches aren't
directly affected, since they don't have npx. I wondered if they
even had curpcb (if not then pc_curpcb should be in the MD fields).
Grep shows that they have considerable variation and messes. At
lease ia64 doesn't use the "MI" pc_curpcb. There is similar variation
and messes for pcpup.
> + return (pcb);
> +#define curpcb (__curpcb())
> #endif /* _AMD64_PCB_H_ */
> diff --git a/sys/sys/user.h b/sys/sys/user.h
> index accb7c3..ab1c7a7 100644
> --- a/sys/sys/user.h
> +++ b/sys/sys/user.h
> @@ -35,6 +35,7 @@
> #ifndef _SYS_USER_H_
> #define _SYS_USER_H_
> +#include <sys/pcpu.h>
Too much namespace pollution for me.
> #include <machine/pcb.h>
Putting the definiton in machine/pcpu.h should avoid changing the
prerequistes for machine/pcb.h.
> #ifndef _KERNEL
> /* stuff that *used* to be included by user.h, or is now needed */
> Please note the location in pcb.h an not in machine/pcpu.h, where it
> cannot work for technical reasons (struct pcpu is not defined yet).
Not applicable -- see above.
> This should be committed separetely, together with the pass over amd64/
> to convert PCPU_GET(curpcb) into curpcb().
> Do you agree with committing the PR fix as is and adding the curpcb() later ?
> And removing fnclex() later (it seems I convinced enough for its removal).
More information about the freebsd-amd64