static int
acpi_timer_test(void)
{
uint32_t last, this;
int min, max, max2, n, delta;
register_t s;
min = INT32_MAX;
max = max2 = 0;
/* Test the timer with interrupts disabled to get accurate results. */
#if defined(__i386__)
s = read_eflags();
#elif defined(__x86_64__)
s = read_rflags();
#else
#error "no read_eflags"
#endif
cpu_disable_intr();
AcpiGetTimer(&last);
for (n = 0; n < 2000; n++) {
AcpiGetTimer(&this);
delta = acpi_TimerDelta(this, last);
if (delta > max) {
max2 = max;
max = delta;
} else if (delta > max2) {
max2 = delta;
}
if (delta < min)
min = delta;
last = this;
}
#if defined(__i386__)
write_eflags(s);
#elif defined(__x86_64__)
write_rflags(s);
#else
#error "no read_eflags"
#endif
delta = max2 - min;
if ((max - min > 8 || delta > 3) && vmm_guest == VMM_GUEST_NONE)
n = 0;
else if (min < 0 || max == 0 || max2 == 0)
n = 0;
else
n = 1;
if (bootverbose) {
kprintf("ACPI timer looks %s min = %d, max = %d, width = %d\n",
n ? "GOOD" : "BAD ",
min, max, max - min);
}
return (n);
}
static void
ioapic_abi_intr_teardown(int intr)
{
const struct ioapic_irqmap *map;
int vector, select;
uint32_t value;
register_t ef;
KASSERT(intr >= 0 && intr < IOAPIC_HWI_VECTORS,
("ioapic teardown, invalid irq %d\n", intr));
map = &ioapic_irqmaps[mycpuid][intr];
KASSERT(IOAPIC_IMT_ISHWI(map),
("ioapic teardown, not hwi irq %d, type %d, cpu%d",
intr, map->im_type, mycpuid));
if (map->im_type != IOAPIC_IMT_LEGACY)
return;
KASSERT(ioapic_irqs[intr].io_addr != NULL,
("ioapic teardown, no GSI information, irq %d\n", intr));
ef = read_rflags();
cpu_disable_intr();
/*
* Teardown an interrupt vector. The vector should already be
* installed in the cpu's IDT, but make sure.
*/
IOAPIC_INTRDIS(intr);
vector = IDT_OFFSET + intr;
/*
* In order to avoid losing an EOI for a level interrupt, which
* is vector based, make sure that the IO APIC is programmed for
* edge-triggering first, then reprogrammed with the new vector.
* This should clear the IRR bit.
*/
imen_lock();
select = ioapic_irqs[intr].io_idx;
value = ioapic_read(ioapic_irqs[intr].io_addr, select);
ioapic_write(ioapic_irqs[intr].io_addr, select,
(value & ~APIC_TRIGMOD_MASK));
ioapic_write(ioapic_irqs[intr].io_addr, select,
(value & ~IOART_INTVEC) | vector);
imen_unlock();
write_rflags(ef);
}
static int
acpi_timer_test(void)
{
uint32_t last, this;
int min, max, n, delta;
register_t s;
min = 10000000;
max = 0;
/* Test the timer with interrupts disabled to get accurate results. */
#if defined(__i386__)
s = read_eflags();
#elif defined(__x86_64__)
s = read_rflags();
#else
#error "no read_eflags"
#endif
cpu_disable_intr();
last = acpi_timer_read();
for (n = 0; n < 2000; n++) {
this = acpi_timer_read();
delta = acpi_TimerDelta(this, last);
if (delta > max)
max = delta;
else if (delta < min)
min = delta;
last = this;
}
#if defined(__i386__)
write_eflags(s);
#elif defined(__x86_64__)
write_rflags(s);
#else
#error "no read_eflags"
#endif
if (max - min > 2)
n = 0;
else if (min < 0 || max == 0)
n = 0;
else
n = 1;
if (bootverbose) {
kprintf("ACPI timer looks %s min = %d, max = %d, width = %d\n",
n ? "GOOD" : "BAD ",
min, max, max - min);
}
return (n);
}
static void
ioapic_abi_intr_setup(int intr, int flags)
{
const struct ioapic_irqmap *map;
int vector, select;
uint32_t value;
register_t ef;
KASSERT(intr >= 0 && intr < IOAPIC_HWI_VECTORS,
("ioapic setup, invalid irq %d", intr));
map = &ioapic_irqmaps[mycpuid][intr];
KASSERT(IOAPIC_IMT_ISHWI(map),
("ioapic setup, not hwi irq %d, type %d, cpu%d",
intr, map->im_type, mycpuid));
if (map->im_type != IOAPIC_IMT_LEGACY)
return;
KASSERT(ioapic_irqs[intr].io_addr != NULL,
("ioapic setup, no GSI information, irq %d", intr));
ef = read_rflags();
cpu_disable_intr();
vector = IDT_OFFSET + intr;
/*
* Now reprogram the vector in the IO APIC. In order to avoid
* losing an EOI for a level interrupt, which is vector based,
* make sure that the IO APIC is programmed for edge-triggering
* first, then reprogrammed with the new vector. This should
* clear the IRR bit.
*/
imen_lock();
select = ioapic_irqs[intr].io_idx;
value = ioapic_read(ioapic_irqs[intr].io_addr, select);
value |= IOART_INTMSET;
ioapic_write(ioapic_irqs[intr].io_addr, select,
(value & ~APIC_TRIGMOD_MASK));
ioapic_write(ioapic_irqs[intr].io_addr, select,
(value & ~IOART_INTVEC) | vector);
imen_unlock();
IOAPIC_INTREN(intr);
write_rflags(ef);
}
void
icu_definit(void)
{
register_t ef;
KKASSERT(MachIntrABI.type == MACHINTR_ICU);
ef = read_rflags();
cpu_disable_intr();
/* Leave interrupts masked */
outb(IO_ICU1 + ICU_IMR_OFFSET, 0xff);
outb(IO_ICU2 + ICU_IMR_OFFSET, 0xff);
MachIntrABI.setdefault();
icu_init();
write_rflags(ef);
}
void
icu_reinit_noioapic(void)
{
register_t ef;
KKASSERT(MachIntrABI.type == MACHINTR_ICU);
KKASSERT(ioapic_enable == 0);
crit_enter();
ef = read_rflags();
cpu_disable_intr();
/* Leave interrupts masked */
outb(IO_ICU1 + ICU_IMR_OFFSET, 0xff);
outb(IO_ICU2 + ICU_IMR_OFFSET, 0xff);
icu_init();
MachIntrABI.stabilize();
write_rflags(ef);
MachIntrABI.cleanup();
crit_exit();
}
/*
* API function - invalidate the pte at (va) and replace *ptep with npte
* atomically only if *ptep equals opte, across the pmap's active cpus.
*
* Returns 1 on success, 0 on failure (caller typically retries).
*/
int
pmap_inval_smp_cmpset(pmap_t pmap, vm_offset_t va, pt_entry_t *ptep,
pt_entry_t opte, pt_entry_t npte)
{
globaldata_t gd = mycpu;
pmap_inval_info_t *info;
int success;
int cpu = gd->gd_cpuid;
cpumask_t tmpmask;
unsigned long rflags;
/*
* Initialize invalidation for pmap and enter critical section.
*/
if (pmap == NULL)
pmap = &kernel_pmap;
pmap_inval_init(pmap);
/*
* Shortcut single-cpu case if possible.
*/
if (CPUMASK_CMPMASKEQ(pmap->pm_active, gd->gd_cpumask)) {
if (atomic_cmpset_long(ptep, opte, npte)) {
if (va == (vm_offset_t)-1)
cpu_invltlb();
else
cpu_invlpg((void *)va);
pmap_inval_done(pmap);
return 1;
} else {
pmap_inval_done(pmap);
return 0;
}
}
/*
* We need a critical section to prevent getting preempted while
* we setup our command. A preemption might execute its own
* pmap_inval*() command and create confusion below.
*/
info = &invinfo[cpu];
info->tsc_target = rdtsc() + (tsc_frequency * LOOPRECOVER_TIMEOUT1);
/*
* We must wait for other cpus which may still be finishing
* up a prior operation.
*/
while (CPUMASK_TESTNZERO(info->done)) {
#ifdef LOOPRECOVER
if (loopwdog(info)) {
info->failed = 1;
loopdebug("B", info);
/* XXX recover from possible bug */
CPUMASK_ASSZERO(info->done);
}
#endif
cpu_pause();
}
KKASSERT(info->mode == INVDONE);
/*
* Must set our cpu in the invalidation scan mask before
* any possibility of [partial] execution (remember, XINVLTLB
* can interrupt a critical section).
*/
ATOMIC_CPUMASK_ORBIT(smp_invmask, cpu);
info->va = va;
info->npgs = 1; /* unused */
info->ptep = ptep;
info->npte = npte;
info->opte = opte;
#ifdef LOOPRECOVER
info->failed = 0;
#endif
info->mode = INVCMPSET;
info->success = 0;
tmpmask = pmap->pm_active; /* volatile */
cpu_ccfence();
CPUMASK_ANDMASK(tmpmask, smp_active_mask);
CPUMASK_ORBIT(tmpmask, cpu);
info->mask = tmpmask;
/*
* Command may start executing the moment 'done' is initialized,
* disable current cpu interrupt to prevent 'done' field from
* changing (other cpus can't clear done bits until the originating
* cpu clears its mask bit).
*/
#ifdef LOOPRECOVER
info->sigmask = tmpmask;
CHECKSIGMASK(info);
#endif
cpu_sfence();
rflags = read_rflags();
cpu_disable_intr();
ATOMIC_CPUMASK_COPY(info->done, tmpmask);
/*
* Pass our copy of the done bits (so they don't change out from
* under us) to generate the Xinvltlb interrupt on the targets.
*/
smp_invlpg(&tmpmask);
success = info->success;
KKASSERT(info->mode == INVDONE);
ATOMIC_CPUMASK_NANDBIT(smp_invmask, cpu);
write_rflags(rflags);
pmap_inval_done(pmap);
return success;
}
/*
* Invalidate the specified va across all cpus associated with the pmap.
* If va == (vm_offset_t)-1, we invltlb() instead of invlpg(). The operation
* will be done fully synchronously with storing npte into *ptep and returning
* opte.
*
* If ptep is NULL the operation will execute semi-synchronously.
* ptep must be NULL if npgs > 1
*/
pt_entry_t
pmap_inval_smp(pmap_t pmap, vm_offset_t va, int npgs,
pt_entry_t *ptep, pt_entry_t npte)
{
globaldata_t gd = mycpu;
pmap_inval_info_t *info;
pt_entry_t opte = 0;
int cpu = gd->gd_cpuid;
cpumask_t tmpmask;
unsigned long rflags;
/*
* Initialize invalidation for pmap and enter critical section.
*/
if (pmap == NULL)
pmap = &kernel_pmap;
pmap_inval_init(pmap);
/*
* Shortcut single-cpu case if possible.
*/
if (CPUMASK_CMPMASKEQ(pmap->pm_active, gd->gd_cpumask)) {
/*
* Convert to invltlb if there are too many pages to
* invlpg on.
*/
if (npgs > MAX_INVAL_PAGES) {
npgs = 0;
va = (vm_offset_t)-1;
}
/*
* Invalidate the specified pages, handle invltlb if requested.
*/
while (npgs) {
--npgs;
if (ptep) {
opte = atomic_swap_long(ptep, npte);
++ptep;
}
if (va == (vm_offset_t)-1)
break;
cpu_invlpg((void *)va);
va += PAGE_SIZE;
}
if (va == (vm_offset_t)-1)
cpu_invltlb();
pmap_inval_done(pmap);
return opte;
}
/*
* We need a critical section to prevent getting preempted while
* we setup our command. A preemption might execute its own
* pmap_inval*() command and create confusion below.
*
* tsc_target is our watchdog timeout that will attempt to recover
* from a lost IPI. Set to 1/16 second for now.
*/
info = &invinfo[cpu];
info->tsc_target = rdtsc() + (tsc_frequency * LOOPRECOVER_TIMEOUT1);
/*
* We must wait for other cpus which may still be finishing up a
* prior operation that we requested.
*
* We do not have to disable interrupts here. An Xinvltlb can occur
* at any time (even within a critical section), but it will not
* act on our command until we set our done bits.
*/
while (CPUMASK_TESTNZERO(info->done)) {
#ifdef LOOPRECOVER
if (loopwdog(info)) {
info->failed = 1;
loopdebug("A", info);
/* XXX recover from possible bug */
CPUMASK_ASSZERO(info->done);
}
#endif
cpu_pause();
}
KKASSERT(info->mode == INVDONE);
/*
* Must set our cpu in the invalidation scan mask before
* any possibility of [partial] execution (remember, XINVLTLB
* can interrupt a critical section).
*/
ATOMIC_CPUMASK_ORBIT(smp_invmask, cpu);
info->va = va;
info->npgs = npgs;
info->ptep = ptep;
info->npte = npte;
info->opte = 0;
#ifdef LOOPRECOVER
info->failed = 0;
#endif
info->mode = INVSTORE;
tmpmask = pmap->pm_active; /* volatile (bits may be cleared) */
cpu_ccfence();
CPUMASK_ANDMASK(tmpmask, smp_active_mask);
/*
* If ptep is NULL the operation can be semi-synchronous, which means
* we can improve performance by flagging and removing idle cpus
* (see the idleinvlclr function in mp_machdep.c).
*
* Typically kernel page table operation is semi-synchronous.
*/
if (ptep == NULL)
smp_smurf_idleinvlclr(&tmpmask);
CPUMASK_ORBIT(tmpmask, cpu);
info->mask = tmpmask;
/*
* Command may start executing the moment 'done' is initialized,
* disable current cpu interrupt to prevent 'done' field from
* changing (other cpus can't clear done bits until the originating
* cpu clears its mask bit, but other cpus CAN start clearing their
* mask bits).
*/
#ifdef LOOPRECOVER
info->sigmask = tmpmask;
CHECKSIGMASK(info);
#endif
cpu_sfence();
rflags = read_rflags();
cpu_disable_intr();
ATOMIC_CPUMASK_COPY(info->done, tmpmask);
/* execution can begin here due to races */
/*
* Pass our copy of the done bits (so they don't change out from
* under us) to generate the Xinvltlb interrupt on the targets.
*/
smp_invlpg(&tmpmask);
opte = info->opte;
KKASSERT(info->mode == INVDONE);
/*
* Target cpus will be in their loop exiting concurrently with our
* cleanup. They will not lose the bitmask they obtained before so
* we can safely clear this bit.
*/
ATOMIC_CPUMASK_NANDBIT(smp_invmask, cpu);
write_rflags(rflags);
pmap_inval_done(pmap);
return opte;
}