new lsub nix patch: noknixcalls
Fco. J. Ballesteros
kernel support for nix calls is gone.
it was an experiment. moving forward now.
files:
/sys/src/nix/k10/k8cpufs k8cpufs
/sys/src/nix/k10/mkfile mkfile
/sys/src/nix/k10/k8cpu k8cpu
/sys/src/nix/port/portfns.h portfns.h
/sys/src/nix/port/proc.c proc.c
/sys/src/nix/port/sysproc.c sysproc.c
/sys/src/nix/port/syssem.c syssem.c
removed:
/sys/src/nix/port/devtube.c
/sys/src/nix/port/nixcall.c
/sys/src/nix/port/tube.h
removed: /sys/src/nix/port/devtube.c
removed: /sys/src/nix/port/nixcall.c
removed: /sys/src/nix/port/tube.h
/sys/src/nix/k10/k8cpufs:
k8cpufs.orig:180 d k8cpufs:179
< nixcall
/sys/src/nix/k10/mkfile:
mkfile.orig:94,95 d mkfile:93
< devtube.$O: ../port/tube.h
< syssem.$O: ../port/tube.h
/sys/src/nix/k10/k8cpu:
k8cpu.orig:15 c k8cpu:15
< # pmc pmcio
---
> pmc pmcio
k8cpu.orig:18 d k8cpu:17
< tube
k8cpu.orig:103 d k8cpu:101
< devtube 'T'
k8cpu.orig:112 d k8cpu:109
< nixcall 'n'
k8cpu.orig:165 d k8cpu:161
< nixcall
k8cpu.orig:184 a k8cpu:181
> # tcklock
/sys/src/nix/port/portfns.h:
portfns.h.orig:350 d portfns.h:349
< void stopnixproc(void);
/sys/src/nix/port/proc.c:
proc.c.orig:440 a proc.c:441
>
proc.c.orig:1420 d proc.c:1420
< stopnixproc();
/sys/src/nix/port/sysproc.c:
sysproc.c.orig:704 a sysproc.c:705,710
> void
> sysnixsyscall(Ar0* , va_list )
> {
> print("nixsyscall() called. recompile your binary\n");
> }
>
/sys/src/nix/port/syssem.c:
syssem.c.orig:8,9 d syssem.c:7
< #include "../port/tube.h"
<
syssem.c.orig:310,454 d syssem.c:307
< }
<
< /*
< * Tube entry points as seen by the kernel.
< * They use Sem*, and not int*
< */
<
< enum{Doblock, Dontblock, Already}; /* xsend() nb argument */
<
< /*
< * Assuming t points to valid memory,
< * create a KTube for the Tube pointed to by t.
< * This is suitable for using following functions to operate
< * on user tubes from the kernel.
< */
< KTube*
< u2ktube(Tube *t)
< {
< KTube *kt;
< Segment *sg;
<
< kt = smalloc(sizeof *kt);
< kt->t = t;
< if((sg = seg(up, PTR2UINT(&t->nhole), 0)) == nil)
< panic("u2ktube: was the tube verified?");
< kt->hole = segmksem(sg, &t->nhole);
< if(seg(up, PTR2UINT(&t->nmsg), 0) != sg)
< panic("u2ktube: tricked user tube?");
< kt->msg = segmksem(sg, &t->nmsg);
< return kt;
< }
<
< static int
< xsend(KTube *kt, void *p, int nb)
< {
< int n;
< uchar *c;
< Tube *t;
<
< t = kt->t;
< assert(t != nil && p != nil);
< if(nb != Already && downsem(kt->hole, nb) < 0)
< return -1;
< n = ainc(&t->tl) - 1;
< n %= t->tsz;
< c = (uchar*)&t[1];
< c += (1+t->msz) * n;
< memmove(c+1, p, t->msz);
< coherence();
< *c = 1;
< upsem(kt->msg);
< return 0;
< }
<
< static int
< xrecv(KTube *kt, void *p, int nb)
< {
< int n;
< uchar *c;
< Tube *t;
<
< t = kt->t;
< assert(t != nil && p != nil);
< if(nb != Already && downsem(kt->msg, nb) < 0)
< return -1;
< n = ainc(&t->hd) - 1;
< n %= t->tsz;
< c = (uchar*)&t[1];
< c += (1+t->msz) * n;
< while(*c == 0)
< yield();
< memmove(p, c+1, t->msz);
< coherence();
< *c = 0;
< upsem(kt->hole);
< return 0;
< }
<
< void
< tsend(KTube *t, void *p)
< {
< xsend(t, p, Doblock);
< }
<
< void
< trecv(KTube *t, void *p)
< {
< xrecv(t, p, Doblock);
< }
<
< int
< nbtsend(KTube *t, void *p)
< {
< return xsend(t, p, Dontblock);
< }
<
< int
< nbtrecv(KTube *t, void *p)
< {
< return xrecv(t, p, Dontblock);
< }
<
< int
< talt(KTalt a[], int na)
< {
< int i, n;
< Sem **ss;
<
< assert(a != nil && na > 0);
< ss = smalloc(sizeof(Sem*) * na);
< n = 0;
< for(i = 0; i < na; i++)
< switch(a[i].op){
< case TSND:
< ss[n++] = a[i].kt->hole;
< break;
< case TRCV:
< ss[n++] = a[i].kt->msg;
< break;
< case TNBSND:
< if(nbtsend(a[i].kt, a[i].m) != -1)
< return i;
< break;
< case TNBRCV:
< if(nbtrecv(a[i].kt, a[i].m) != -1)
< return i;
< break;
< }
< if(n == 0)
< return -1;
< i = altsems(ss, n);
< free(ss);
< if(i < 0)
< return -1;
< switch(a[i].op){
< case TSND:
< xsend(a[i].kt, a[i].m, Already);
< break;
< case TRCV:
< xrecv(a[i].kt, a[i].m, Already);
< break;
< default:
< panic("talt");
< }
< return i;
erik quanstrom
could you explain what the thinking is in a bit more detail?
(i'm not arguing. i just want to have this clear in my head.)
by the way, can't we do asynchronous syscalls to some extent by
providing the proper device?
- erik
Nemo
yet because I'm still thinking on it and changing it in the whiteboard.
But I'm sure now nixcalls are not the way to go and I don't think there's
any point in keeping them.
The thinking of the code I want to remove is to use channels to the kernel
to send messages to issue system calls. That's too complex and it's even
slower than synchronous calls when used from TCs.
And yes, I think the answer to your question is yes.
But I think there might be a better way.
Nemo
;; patch/Apply nonixcalls
applying to /n/src/nix...
;; patch/Apply noknixcalls
applying to /n/src/nix...
On Apr 15, 2012, at 6:54 PM, erik quanstrom wrote:
erik quanstrom
> But I think there might be a better way.
without knowing what you're better way might be,
i'll just note that a device has the advantage (or disadvantage)
of being mountable from another machine.
- erik
Charles Forsyth
A system call is a protected procedure (function) call. An asynchronous function is a process!