harvey/sys/src/9/amd64/acore.c.old

#include "u.h"
#include "../port/lib.h"
#include "mem.h"
#include "dat.h"
#include "fns.h"

#include <tos.h>
#include <pool.h>
#include "amd64.h"
#include "ureg.h"
#include "io.h"

/*
 * BUG:
 * The AC must not accept interrupts while in the kernel,
 * or we must be prepared for nesting them, which we are not.
 * This is important for note handling, because postnote()
 * assumes that it's ok to send an IPI to an AC, no matter its
 * state.
 * 
 */

void
intrac(Proc *p)
{
	Mach *ac;

	ac = p->ac;
	if(ac == nil){
		DBG("intrac: Proc.ac is nil. no ipi sent.\n");
		return;
	}
	/*
	 * It's ok if the AC gets idle in the mean time.
	 */
	DBG("intrac: ipi to cpu%d\n", ac->machno);
	apicipi(ac->apicno);
}

/*
 * Functions starting with ac... are run in the application core.
 * All other functions are run by the time-sharing cores.
 */

typedef void (*APfunc)(void);
extern int notify(Ureg*);
extern void _acsysret(void);
extern void _actrapret(void);

static char *acnames[] = { "Ok", "Trap", "Syscall"};

void
acmmuswitch(void)
{
	cr3put(machp()->pml4->pa);
}
void xactouser(u64int);
void
actouser(void)
{
	uintptr sp;
	Ureg *u;

	memmove(&sp, m->icc->data, sizeof(sp));
	u = m->proc->dbgreg;
	DBG("cpu%d: touser usp = %#p entry %#p\n", machp()->machno, sp, u->ip);


	/*
	 * This code for updating tos is wrong. It shouldn't go here.
	 * It's the fact that we assing a process to a core what makes
	 * it run in that core, not the fact that we call actouser(),
	 * In the future, we might not even call actouser here. -Nemo.
	 */

	/* BUG: add a function, called here and kexit */
	m->load = 100;
	xactouser(sp);
	panic("actouser");
}

void
actrapret(void)
{
	/* done by actrap() */
}

/*
 * Entered in AP core context, upon traps and system calls.
 * using up->dbgreg means cores MUST be homogeneous.
 */
void
actrap(Ureg *u)
{
	/* print instead of DBG, so we see any trap by now */
	switch(u->type){
	case IdtIPI:
		m->intr++;
		print("actrap: cpu%d: IPI\n", machp()->machno);
		/*
		 * Beware: BUG: we can get now IPIs while in kernel mode,
		 * after declaring the end of the interrupt.
		 * The code is not prepared for that.
		 */
		apiceoi(IdtIPI);
		break;
	case IdtPF:
		m->pfault++;
		print("actrap: cpu%d: PF\n", machp()->machno);
		break;
	default:
		print("actrap: cpu%d: %ulld\n", machp()->machno, u->type);
	}
	m->icc->rc = ICCTRAP;
	m->cr2 = cr2get();
	memmove(m->proc->dbgreg, u, sizeof *u);
	m->icc->fn = nil;
	mfence();
	ready(m->proc);
	m->load = 0;
	while(*m->icc->fn == nil)
		;
	m->load = 100;
	if(m->icc->flushtlb)
		acmmuswitch();
	DBG("actrap: ret\n");
	if(m->icc->fn != actrapret)
		acsched();
	memmove(u, m->proc->dbgreg, sizeof *u);
}

void
acsyscall(void)
{
	/*
	 * If we saved the Ureg into m->proc->dbgregs,
	 * There's nothing else we have to do.
	 * Otherwise, we should m->proc->dbgregs = u;
	 */
	DBG("acsyscall: cpu%d\n", machp()->machno);
	m->syscall++;	/* would also count it in the TS core */
	m->icc->rc = ICCSYSCALL;
	m->cr2 = cr2get();
	m->icc->fn = nil;
	ready(m->proc);
	m->load = 0;
	/*
	 * The next call is probably going to make us jmp
	 * into user code, forgetting all our state in this
	 * stack, upon the next syscall.
	 * We don't nest calls in the current stack for too long.
	 */
	acsched();
}

/*
 * Called in AP core context, to return from system call.
 */
void
acsysret(void)
{
	DBG("acsysret\n");
	_acsysret();
}

void
dumpreg(void *u)
{
	print("reg is %p\n", u);
	ndnr();
}


/*
 * run an arbitrary function with arbitrary args on an ap core
 * first argument is always pml4 for process
 * make a field and a struct for the args cache line.
 *
 * Returns the return-code for the ICC or -1 if the process was
 * interrupted while issuing the ICC.
 */
int
runac(int core, APfunc func, int flushtlb, void *a, long n)
{
	Mach *mp;
	uchar *dpg, *spg;

	if (n > sizeof(mp->icc->data))
		panic("runac: args too long");

	if((mp = sys->machptr[core]) == nil || mp->online == 0)
		panic("Bad core");
	if(mp->proc != nil && mp->proc != up)
		panic("runapfunc: mach is busy with another proc?");

	memmove(mp->icc->data, a, n);
	if(flushtlb){
		dpg = UINT2PTR(mp->pml4->va);
		spg = UINT2PTR(machp()->pml4->va);
		/* We should copy only user space mappings:
		 *	memmove(dgp, spg, machp()->pml4->daddr * sizeof(PTE));
		 */
		memmove(dpg, spg, PTPGSZ);
	}
	mp->icc->flushtlb = flushtlb;
	mp->icc->rc = ICCOK;

	DBG("runac: exotic proc on cpu%d\n", mp->machno);
	if(waserror()){
		qunlock(&up->debug);
		nexterror();
	}
	qlock(&up->debug);
	up->ac = mp;
	up->nicc++;
	up->state = Exotic;
	up->psstate = 0;
	qunlock(&up->debug);
	poperror();
	mfence();
	mp->icc->fn = func;
	sched();
	return mp->icc->rc;
}

/*
 * Cleanup done by runacore to pretend we are going back to user space.
 * We won't return and won't do what syscall() would normally do.
 * Do it here instead.
 */
static void
fakeretfromsyscall(Ureg *ureg)
{
	int s;

	poperror();	/* as syscall() would do if we would return */
	if(up->procctl == Proc_tracesyscall){	/* Would this work? */
		up->procctl = Proc_stopme;
		s = splhi();
		procctl(up);
		splx(s);
	}

	up->insyscall = 0;
	/* if we delayed sched because we held a lock, sched now */
	if(up->delaysched){
		sched();
		splhi();
	}
	kexit(ureg);
}

static void
testproc(void *a)
{
	Proc *p;

	p = a;
	if(p == nil){
		print("no proc to intr\n");
		return;
	}
	tsleep(&up->sleep, return0, 0, 10000);
	print("testproc: sending ipi to proc\n");
	intrac(p);
	print("sent\n");
}

/*
 * Move the current process to an application core.
 * This is performed at the end of execac(), and
 * we pretend to be returning to user-space, but instead we
 * dispatch the process to another core.
 * 1. We do the final bookkeeping that syscall() would do after
 *    a return from sysexec(), because we are not returning.
 * 2. We dispatch the process to an AC using an ICC.
 *
 * This function won't return unless the process is reclaimed back
 * to the time-sharing core, and is the handler for the process
 * to deal with traps and system calls until the process dies.
 *
 * Remember that this function is the "line" between user and kernel
 * space, it's not expected to raise|handle any error.
 *
 * We install a safety error label, just in case we raise errors,
 * which we shouldn't. (noerrorsleft knows that for exotic processes
 * there is an error label pushed by us).
 */
void
runacore(int core, u64int ar0p)
{
	Ureg *ureg;
	void (*fn)(void);
	int rc, flush, becometimesharing;

	if(waserror())
		panic("runacore: error: %s\n", up->errstr);
	ureg = up->dbgreg;
	ureg->ax = ar0p;		/* see xactouser */
	fakeretfromsyscall(ureg);

/* IPI testing */
if(0)
kproc("testproc", testproc, up);

	rc = runac(core, actouser, 1, &ureg->sp, sizeof ureg->sp);
	becometimesharing = 0;
	for(;;){
		flush = 0;
		fn = nil;
		switch(rc){
		case ICCTRAP:
			m->cr2 = up->ac->cr2;
			DBG("runacore: trap %ulld cr2 %#ullx ureg %#p\n",
				ureg->type, m->cr2, ureg);
			if(ureg->type == IdtIPI){
				if(up->procctl || up->nnote)
					notify(up->dbgreg);
				kexit(up->dbgreg);
			}else
				trap(ureg);
			flush = 1;
			fn = actrapret;
			break;
		case ICCSYSCALL:
			DBG("runacore: syscall ax %#ullx ureg %#p\n",
				ureg->ax, ureg);
			syscall(ureg->ax, ureg);
			flush = 1;
			fn = acsysret;
			break;
		default:
			panic("runacore: unexpected rc = %d", rc);
		}
		if(becometimesharing)
			break;
		rc = runac(core, fn, flush, &ureg, sizeof ureg);
	}

	fakeretfromsyscall(up->dbgreg);
	/*
	 * dettach from the AC.
	 */
	up->ac->proc = nil;
	up->ac = nil;
	/* And we return to syscall, which would do nothing but
	 * returning, and we'd be back to the TS core.
	 */
}

void
acmodeset(int mode)
{
	switch(mode){
	case NIXAC:
	case NIXKC:
	case NIXTC:
		break;
	default:
		panic("apmodeset: bad mode %d", mode);
	}
	m->nixtype = mode;
}

void
stopac(Proc *p)
{
	Mach *mp;

	mp = p->ac;
	if(mp == nil)
		return;
	if(mp->proc != p)
		return;
	DBG("stopac: cpu%d\n", mp->machno);
	p->ac = nil;
	mp->proc = nil;
//	send sipi to p->ac, it would rerun squidboy(), and
//	wait for us to give it a function to run.
}

void
acinit(void)
{
	/*
	 * Lower the priority of the apic to 0,
	 * to accept interrupts.
	 * Raise it later if needed to disable them.
	 */
	apicpri(0);
}