plan9fox/sys/src/9/pc/mp.c
cinap_lenrek 1d93a5628a pc, pc64, xen: rewrite interrupt handling code
This implements proper intrdisable() support for all
interrupt controllers.

For enable, (*arch->intrassign)(Vctl*) fills in the
Vctl.enable and Vctl.disable pointers with the
appropriate routines and returns the assigned
vector number.

Once the Vctl struct has been linked to its vector
chain, Vctl.enable(Vctl*, shared) gets called with a
flag if the vector has been already enabled (shared).

This order is important here as enabling the interrupt
on the controller before we have linked the chain can
cause spurious interrupts, expecially on mp system
where the interrupt can target a different cpu than
the caller of intrenable().

The intrdisable() case is the other way around.
We first disable the interrupt on the controller
and after that unlink the Vctl from the chain.
On a multiprocessor, the xfree() of the Vctl struct
is delayed to avoid freeing it while it is still
in use by another cpu.

The xen port now also uses pc/irq.c which has been
made generic enougth to handle xen's irq scheme.
Also, archgeneric is now a separate file to avoid
pulling in dependencies from the 8259 interrupt
controller code.
2020-11-29 17:43:22 +01:00

618 lines
13 KiB
C

#include "u.h"
#include "../port/lib.h"
#include "mem.h"
#include "dat.h"
#include "fns.h"
#include "io.h"
#include "../port/pci.h"
#include "ureg.h"
#include "mp.h"
#include "apbootstrap.i"
extern void i8259init(void);
/* filled in by pcmpinit or acpiinit */
Bus* mpbus;
Bus* mpbuslast;
int mpisabus = -1;
int mpeisabus = -1;
Apic *mpioapic[MaxAPICNO+1];
Apic *mpapic[MaxAPICNO+1];
int
mpintrinit(Bus* bus, PCMPintr* intr, int vno, int /*irq*/)
{
int el, po, v;
/*
* Parse an I/O or Local APIC interrupt table entry and
* return the encoded vector.
*/
v = vno;
po = intr->flags & PcmpPOMASK;
el = intr->flags & PcmpELMASK;
switch(intr->intr){
default: /* PcmpINT */
v |= ApicFIXED; /* no-op */
break;
case PcmpNMI:
v |= ApicNMI;
po = PcmpHIGH;
el = PcmpEDGE;
break;
case PcmpSMI:
v |= ApicSMI;
break;
case PcmpExtINT:
v |= ApicExtINT;
/*
* The AMI Goliath doesn't boot successfully with it's LINTR0
* entry which decodes to low+level. The PPro manual says ExtINT
* should be level, whereas the Pentium is edge. Setting the
* Goliath to edge+high seems to cure the problem. Other PPro
* MP tables (e.g. ASUS P/I-P65UP5 have a entry which decodes
* to edge+high, so who knows.
* Perhaps it would be best just to not set an ExtINT entry at
* all, it shouldn't be needed for SMP mode.
*/
po = PcmpHIGH;
el = PcmpEDGE;
break;
}
/*
*/
if(bus->type == BusEISA && !po && !el /*&& !(i8259elcr & (1<<irq))*/){
po = PcmpHIGH;
el = PcmpEDGE;
}
if(!po)
po = bus->po;
if(po == PcmpLOW)
v |= ApicLOW;
else if(po != PcmpHIGH){
print("mpintrinit: bad polarity 0x%uX\n", po);
return ApicIMASK;
}
if(!el)
el = bus->el;
if(el == PcmpLEVEL)
v |= ApicLEVEL;
else if(el != PcmpEDGE){
print("mpintrinit: bad trigger 0x%uX\n", el);
return ApicIMASK;
}
return v;
}
uvlong
tscticks(uvlong *hz)
{
if(hz != nil)
*hz = m->cpuhz;
cycles(&m->tscticks); /* Uses the rdtsc instruction */
return m->tscticks;
}
void
syncclock(void)
{
uvlong x;
if(arch->fastclock != tscticks)
return;
if(m->machno == 0){
wrmsr(0x10, 0);
m->tscticks = 0;
} else {
x = MACHP(0)->tscticks;
while(x == MACHP(0)->tscticks)
;
wrmsr(0x10, MACHP(0)->tscticks);
cycles(&m->tscticks);
}
}
void
mpinit(void)
{
int ncpu, i;
Apic *apic;
char *cp;
i8259init();
syncclock();
if(getconf("*apicdebug")){
Bus *b;
Aintr *ai;
PCMPintr *pi;
for(i=0; i<=MaxAPICNO; i++){
if(apic = mpapic[i])
print("LAPIC%d: pa=%lux va=%#p flags=%x\n",
i, apic->paddr, apic->addr, apic->flags);
if(apic = mpioapic[i])
print("IOAPIC%d: pa=%lux va=%#p flags=%x gsibase=%d mre=%d\n",
i, apic->paddr, apic->addr, apic->flags, apic->gsibase, apic->mre);
}
for(b = mpbus; b; b = b->next){
print("BUS%d type=%d flags=%x\n", b->busno, b->type, b->po|b->el);
for(ai = b->aintr; ai; ai = ai->next){
if(pi = ai->intr)
print("\ttype=%d irq=%d (%d [%c]) apic=%d intin=%d flags=%x\n",
pi->type, pi->irq, pi->irq>>2, "ABCD"[pi->irq&3],
pi->apicno, pi->intin, pi->flags);
}
}
}
apic = nil;
for(i=0; i<=MaxAPICNO; i++){
if(mpapic[i] == nil)
continue;
if(mpapic[i]->flags & PcmpBP){
apic = mpapic[i];
break;
}
}
if(apic == nil){
panic("mpinit: no bootstrap processor");
return;
}
apic->online = 1;
lapicinit(apic);
/*
* These interrupts are local to the processor
* and do not appear in the I/O APIC so it is OK
* to set them now.
*/
intrenable(IrqTIMER, lapicclock, 0, BUSUNKNOWN, "clock");
intrenable(IrqERROR, lapicerror, 0, BUSUNKNOWN, "lapicerror");
intrenable(IrqSPURIOUS, lapicspurious, 0, BUSUNKNOWN, "lapicspurious");
lapiconline();
/*
* Initialise the application processors.
*/
if(cp = getconf("*ncpu")){
ncpu = strtol(cp, 0, 0);
if(ncpu < 1)
ncpu = 1;
else if(ncpu > MAXMACH)
ncpu = MAXMACH;
}
else
ncpu = MAXMACH;
memmove((void*)APBOOTSTRAP, apbootstrap, sizeof(apbootstrap));
for(i=0; i<nelem(mpapic); i++){
if((apic = mpapic[i]) == nil)
continue;
if(apic->machno >= MAXMACH)
continue;
if(ncpu <= 1)
break;
if((apic->flags & (PcmpBP|PcmpEN)) == PcmpEN){
mpstartap(apic);
conf.nmach++;
ncpu--;
}
}
/*
* we don't really know the number of processors till
* here.
*
* set conf.copymode here if nmach > 1.
* Should look for an ExtINT line and enable it.
*/
if(m->cpuidfamily == 3 || conf.nmach > 1)
conf.copymode = 1;
}
static int
mpintrcpu(void)
{
static Lock physidlock;
static int physid;
int i;
/*
* The bulk of this code was written ~1995, when there was
* one architecture and one generation of hardware, the number
* of CPUs was up to 4(8) and the choices for interrupt routing
* were physical, or flat logical (optionally with lowest
* priority interrupt). Logical mode hasn't scaled well with
* the increasing number of packages/cores/threads, so the
* fall-back is to physical mode, which works across all processor
* generations, both AMD and Intel, using the APIC and xAPIC.
*
* Interrupt routing policy can be set here.
* Currently, just assign each interrupt to a different CPU on
* a round-robin basis. Some idea of the packages/cores/thread
* topology would be useful here, e.g. to not assign interrupts
* to more than one thread in a core, or to use a "noise" core.
* But, as usual, Intel make that an onerous task.
*/
lock(&physidlock);
for(;;){
i = physid++;
if(physid >= nelem(mpapic))
physid = 0;
if(mpapic[i] == nil)
continue;
if(mpapic[i]->online)
break;
}
unlock(&physidlock);
return mpapic[i]->apicno;
}
/*
* With the APIC a unique vector can be assigned to each
* request to enable an interrupt. There are two reasons this
* is a good idea:
* 1) to prevent lost interrupts, no more than 2 interrupts
* should be assigned per block of 16 vectors (there is an
* in-service entry and a holding entry for each priority
* level and there is one priority level per block of 16
* interrupts).
* 2) each input pin on the IOAPIC will receive a different
* vector regardless of whether the devices on that pin use
* the same IRQ as devices on another pin.
*/
static int
allocvector(void)
{
static int round = 0, num = 0;
static Lock l;
int vno;
lock(&l);
vno = VectorAPIC + num;
if(vno < MaxVectorAPIC-7)
num += 8;
else
num = ++round % 8;
unlock(&l);
return vno;
}
static int
ioapicirqenable(Vctl *v, int shared)
{
Aintr *aintr = v->aux;
int lo, hi;
if(shared)
return 0;
hi = v->cpu<<24;
lo = mpintrinit(aintr->bus, aintr->intr, v->vno, v->irq);
lo |= ApicPHYSICAL; /* no-op */
ioapicrdtw(aintr->apic, aintr->intr->intin, hi, lo);
return 0;
}
static int
ioapicirqdisable(Vctl *v, int shared)
{
Aintr *aintr = v->aux;
int lo, hi;
if(shared)
return 0;
hi = 0;
lo = ApicIMASK;
ioapicrdtw(aintr->apic, aintr->intr->intin, hi, lo);
return 0;
}
static int
mpintrassignx(Vctl* v, int tbdf)
{
Bus *bus;
Pcidev *pci;
Aintr *aintr;
int bno, dno, pin, irq, type, lo, hi, n;
type = BUSTYPE(tbdf);
bno = BUSBNO(tbdf);
dno = BUSDNO(tbdf);
pin = 0;
pci = nil;
if(type == BusPCI){
if((pci = pcimatchtbdf(tbdf)) != nil)
pin = pcicfgr8(pci, PciINTP);
} else if(type == BusISA)
bno = mpisabus;
Findbus:
for(bus = mpbus; bus != nil; bus = bus->next){
if(bus->type != type)
continue;
if(bus->busno == bno)
break;
}
if(bus == nil){
/*
* if the PCI device is behind a PCI-PCI bridge thats not described
* by the MP or ACPI tables then walk up the bus translating interrupt
* pin to parent bus.
*/
if(pci != nil && pci->parent != nil && pin > 0){
pin = ((dno+(pin-1))%4)+1;
pci = pci->parent;
bno = BUSBNO(pci->tbdf);
dno = BUSDNO(pci->tbdf);
goto Findbus;
}
print("mpintrassign: can't find bus type %d, number %d\n", type, bno);
return -1;
}
/*
* For PCI devices the interrupt pin (INT[ABCD]) and device
* number are encoded into the entry irq field, so create something
* to match on.
*/
if(bus->type == BusPCI){
if(pin > 0)
irq = (dno<<2)|(pin-1);
else
irq = -1;
} else
irq = v->irq;
/*
* Find a matching interrupt entry from the list of interrupts
* attached to this bus.
*/
for(aintr = bus->aintr; aintr != nil; aintr = aintr->next){
if(aintr->intr->irq != irq)
continue;
/*
* Check if already enabled. Multifunction devices may share
* INT[A-D]# so, if already enabled, check the polarity matches
* and the trigger is level.
*/
ioapicrdtr(aintr->apic, aintr->intr->intin, &hi, &lo);
if(lo & ApicIMASK){
v->vno = allocvector();
v->cpu = mpintrcpu();
lo = mpintrinit(aintr->bus, aintr->intr, v->vno, v->irq);
lo |= ApicPHYSICAL; /* no-op */
if(lo & ApicIMASK){
print("mpintrassign: disabled irq %d, tbdf %uX, lo %8.8uX, hi %8.8uX\n",
v->irq, v->tbdf, lo, hi);
break;
}
} else {
v->vno = lo & 0xFF;
v->cpu = hi >> 24;
lo &= ~(ApicRemoteIRR|ApicDELIVS);
n = mpintrinit(aintr->bus, aintr->intr, v->vno, v->irq);
n |= ApicPHYSICAL; /* no-op */
if(lo != n){
print("mpintrassign: multiple botch irq %d, tbdf %uX, lo %8.8uX, n %8.8uX\n",
v->irq, v->tbdf, lo, n);
break;
}
}
v->isr = lapicisr;
v->eoi = lapiceoi;
if((aintr->apic->flags & PcmpEN) && aintr->apic->type == PcmpIOAPIC){
v->aux = aintr;
v->enable = ioapicirqenable;
v->disable = ioapicirqdisable;
}
return v->vno;
}
return -1;
}
enum {
HTMSIMapping = 0xA8,
HTMSIFlags = 0x02,
HTMSIFlagsEn = 0x01,
};
static int
htmsicapenable(Pcidev *p)
{
int cap, flags;
if((cap = pcihtcap(p, HTMSIMapping)) <= 0)
return -1;
flags = pcicfgr8(p, cap + HTMSIFlags);
if((flags & HTMSIFlagsEn) == 0)
pcicfgw8(p, cap + HTMSIFlags, flags | HTMSIFlagsEn);
return 0;
}
static int
htmsienable(Pcidev *pdev)
{
Pcidev *p;
p = nil;
while((p = pcimatch(p, 0x1022, 0)) != nil)
if(p->did == 0x1103 || p->did == 0x1203)
break;
if(p == nil)
return 0; /* not hypertransport platform */
p = nil;
while((p = pcimatch(p, 0x10de, 0)) != nil){
switch(p->did){
case 0x02f0: /* NVIDIA NFORCE C51 MEMC0 */
case 0x02f1: /* NVIDIA NFORCE C51 MEMC1 */
case 0x02f2: /* NVIDIA NFORCE C51 MEMC2 */
case 0x02f3: /* NVIDIA NFORCE C51 MEMC3 */
case 0x02f4: /* NVIDIA NFORCE C51 MEMC4 */
case 0x02f5: /* NVIDIA NFORCE C51 MEMC5 */
case 0x02f6: /* NVIDIA NFORCE C51 MEMC6 */
case 0x02f7: /* NVIDIA NFORCE C51 MEMC7 */
case 0x0369: /* NVIDIA NFORCE MCP55 MEMC */
htmsicapenable(p);
break;
}
}
if(htmsicapenable(pdev) == 0)
return 0;
for(p = pdev->parent; p != nil; p = p->parent)
if(htmsicapenable(p) == 0)
return 0;
return -1;
}
static int
msiirqenable(Vctl *v, int)
{
Pcidev *pci = v->aux;
return pcimsienable(pci, 0xFEE00000ULL | (v->cpu << 12), v->vno | (1<<14));
}
static int
msiirqdisable(Vctl *v, int)
{
Pcidev *pci = v->aux;
return pcimsidisable(pci);
}
static int
msiintrenable(Vctl *v)
{
Pcidev *pci;
int tbdf;
if(getconf("*nomsi") != nil)
return -1;
tbdf = v->tbdf;
if(tbdf == BUSUNKNOWN || BUSTYPE(tbdf) != BusPCI)
return -1;
pci = pcimatchtbdf(tbdf);
if(pci == nil) {
print("msiintrenable: could not find Pcidev for tbdf %uX\n", tbdf);
return -1;
}
if(htmsienable(pci) < 0)
return -1;
if(pcimsidisable(pci) < 0)
return -1;
v->vno = allocvector();
v->cpu = mpintrcpu();
v->isr = lapicisr;
v->eoi = lapiceoi;
v->aux = pci;
v->enable = msiirqenable;
v->disable = msiirqdisable;
return v->vno;
}
int
mpintrassign(Vctl* v)
{
int irq, tbdf, vno;
vno = msiintrenable(v);
if(vno != -1)
return vno;
/*
* If the bus is known, try it.
* BUSUNKNOWN is given both by [E]ISA devices and by
* interrupts local to the processor (local APIC, coprocessor
* breakpoint and page-fault).
*/
tbdf = v->tbdf;
if(tbdf != BUSUNKNOWN && (vno = mpintrassignx(v, tbdf)) != -1)
return vno;
irq = v->irq;
if(irq >= IrqLINT0 && irq <= MaxIrqLAPIC){
v->local = 1;
if(irq != IrqSPURIOUS)
v->isr = lapiceoi;
return VectorPIC+irq;
}
if(irq < 0 || irq > MaxIrqPIC){
print("mpintrassign: irq %d out of range\n", irq);
return -1;
}
/*
* Either didn't find it or have to try the default buses
* (ISA and EISA). This hack is due to either over-zealousness
* or laziness on the part of some manufacturers.
*
* The MP configuration table on some older systems
* (e.g. ASUS PCI/E-P54NP4) has an entry for the EISA bus
* but none for ISA. It also has the interrupt type and
* polarity set to 'default for this bus' which wouldn't
* be compatible with ISA.
*/
if(mpeisabus != -1){
vno = mpintrassignx(v, MKBUS(BusEISA, 0, 0, 0));
if(vno != -1)
return vno;
}
if(mpisabus != -1){
vno = mpintrassignx(v, MKBUS(BusISA, 0, 0, 0));
if(vno != -1)
return vno;
}
print("mpintrassign: out of choices eisa %d isa %d tbdf %uX irq %d\n",
mpeisabus, mpisabus, v->tbdf, v->irq);
return -1;
}
void
mpshutdown(void)
{
/*
* Park application processors.
*/
if(m->machno != 0){
splhi();
arch->introff();
for(;;) idle();
}
delay(1000);
splhi();
/*
* INIT all excluding self.
*/
lapicicrw(0, 0x000C0000|ApicINIT);
pcireset();
}