static void
adj_perf(cpumask_t xcpu_used, cpumask_t xcpu_pwrdom_used)
{
cpumask_t old_usched_used;
int cpu, inc;
/*
* Set cpus requiring performance to the userland process
* scheduler. Leave the rest of cpus unmapped.
*/
old_usched_used = usched_cpu_used;
usched_cpu_used = cpu_used;
if (CPUMASK_TESTZERO(usched_cpu_used))
CPUMASK_ORBIT(usched_cpu_used, 0);
if (CPUMASK_CMPMASKNEQ(usched_cpu_used, old_usched_used))
set_uschedcpus();
/*
* Adjust per-cpu performance.
*/
CPUMASK_XORMASK(xcpu_used, cpu_used);
while (CPUMASK_TESTNZERO(xcpu_used)) {
cpu = BSFCPUMASK(xcpu_used);
CPUMASK_NANDBIT(xcpu_used, cpu);
if (CPUMASK_TESTBIT(cpu_used, cpu)) {
/* Increase cpu performance */
inc = 1;
} else {
/* Decrease cpu performance */
inc = 0;
}
adj_cpu_perf(cpu, inc);
}
/*
* Adjust cpu power domain performance. This could affect
* a set of cpus.
*/
CPUMASK_XORMASK(xcpu_pwrdom_used, cpu_pwrdom_used);
while (CPUMASK_TESTNZERO(xcpu_pwrdom_used)) {
int dom;
dom = BSFCPUMASK(xcpu_pwrdom_used);
CPUMASK_NANDBIT(xcpu_pwrdom_used, dom);
if (CPUMASK_TESTBIT(cpu_pwrdom_used, dom)) {
/* Increase cpu power domain performance */
inc = 1;
} else {
/* Decrease cpu power domain performance */
inc = 0;
}
adj_cpu_pwrdom(dom, inc);
}
}
/*
* Get SMP fully working before we start initializing devices.
*/
static
void
ap_finish(void)
{
mp_finish = 1;
if (bootverbose)
kprintf("Finish MP startup\n");
/* build our map of 'other' CPUs */
mycpu->gd_other_cpus = smp_startup_mask;
CPUMASK_NANDBIT(mycpu->gd_other_cpus, mycpu->gd_cpuid);
/*
* Let the other cpu's finish initializing and build their map
* of 'other' CPUs.
*/
rel_mplock();
while (CPUMASK_CMPMASKNEQ(smp_active_mask,smp_startup_mask)) {
DELAY(100000);
cpu_lfence();
}
while (try_mplock() == 0)
DELAY(100000);
if (bootverbose)
kprintf("Active CPU Mask: %08lx\n",
(long)CPUMASK_LOWMASK(smp_active_mask));
}
static void
add_spare_cpus(const cpumask_t ocpu_used, int ncpu)
{
cpumask_t saved_pwrdom, xcpu_used;
int done = 0, cpu;
/*
* Find more cpus in the previous cpu set.
*/
xcpu_used = cpu_used;
CPUMASK_XORMASK(xcpu_used, ocpu_used);
while (CPUMASK_TESTNZERO(xcpu_used)) {
cpu = BSFCPUMASK(xcpu_used);
CPUMASK_NANDBIT(xcpu_used, cpu);
if (CPUMASK_TESTBIT(ocpu_used, cpu)) {
CPUMASK_ORBIT(cpu_pwrdom_used, cpu2pwrdom[cpu]);
CPUMASK_ORBIT(cpu_used, cpu);
--ncpu;
if (ncpu == 0)
return;
}
}
/*
* Find more cpus in the used cpu power domains.
*/
saved_pwrdom = cpu_pwrdom_used;
again:
while (CPUMASK_TESTNZERO(saved_pwrdom)) {
cpumask_t unused_cpumask;
int dom;
dom = BSFCPUMASK(saved_pwrdom);
CPUMASK_NANDBIT(saved_pwrdom, dom);
unused_cpumask = cpu_pwrdomain[dom]->dom_cpumask;
CPUMASK_NANDMASK(unused_cpumask, cpu_used);
while (CPUMASK_TESTNZERO(unused_cpumask)) {
cpu = BSFCPUMASK(unused_cpumask);
CPUMASK_NANDBIT(unused_cpumask, cpu);
CPUMASK_ORBIT(cpu_pwrdom_used, dom);
CPUMASK_ORBIT(cpu_used, cpu);
--ncpu;
if (ncpu == 0)
return;
}
}
if (!done) {
done = 1;
/*
* Find more cpus in unused cpu power domains
*/
saved_pwrdom = cpu_pwrdom_mask;
CPUMASK_NANDMASK(saved_pwrdom, cpu_pwrdom_used);
goto again;
}
if (DebugOpt)
printf("%d cpus not found\n", ncpu);
}
/*
* Figure out the cpu power domains.
*/
static int
acpi_get_cpupwrdom(void)
{
struct cpu_pwrdom *dom;
cpumask_t pwrdom_mask;
char buf[64];
char members[1024];
char *str;
size_t msize;
int n, i, ncpu = 0, dom_id;
memset(cpu2pwrdom, 0, sizeof(cpu2pwrdom));
memset(cpu_pwrdomain, 0, sizeof(cpu_pwrdomain));
CPUMASK_ASSZERO(cpu_pwrdom_mask);
for (i = 0; i < MAXDOM; ++i) {
snprintf(buf, sizeof(buf),
"hw.acpi.cpu.px_dom%d.available", i);
if (sysctlbyname(buf, NULL, NULL, NULL, 0) < 0)
continue;
dom = calloc(1, sizeof(*dom));
dom->dom_id = i;
if (cpu_pwrdomain[i] != NULL) {
fprintf(stderr, "cpu power domain %d exists\n", i);
exit(1);
}
cpu_pwrdomain[i] = dom;
CPUMASK_ORBIT(cpu_pwrdom_mask, i);
}
pwrdom_mask = cpu_pwrdom_mask;
while (CPUMASK_TESTNZERO(pwrdom_mask)) {
dom_id = BSFCPUMASK(pwrdom_mask);
CPUMASK_NANDBIT(pwrdom_mask, dom_id);
dom = cpu_pwrdomain[dom_id];
CPUMASK_ASSZERO(dom->dom_cpumask);
snprintf(buf, sizeof(buf),
"hw.acpi.cpu.px_dom%d.members", dom->dom_id);
msize = sizeof(members);
if (sysctlbyname(buf, members, &msize, NULL, 0) < 0) {
cpu_pwrdomain[dom_id] = NULL;
free(dom);
continue;
}
members[msize] = 0;
for (str = strtok(members, " "); str; str = strtok(NULL, " ")) {
n = -1;
sscanf(str, "cpu%d", &n);
if (n >= 0) {
++ncpu;
++dom->dom_ncpus;
CPUMASK_ORBIT(dom->dom_cpumask, n);
cpu2pwrdom[n] = dom->dom_id;
}
}
if (dom->dom_ncpus == 0) {
cpu_pwrdomain[dom_id] = NULL;
free(dom);
continue;
}
if (DebugOpt) {
printf("dom%d cpumask: ", dom->dom_id);
for (i = 0; i < (int)NELEM(dom->dom_cpumask.ary); ++i) {
printf("%jx ",
(uintmax_t)dom->dom_cpumask.ary[i]);
}
printf("\n");
}
}
if (ncpu != NCpus) {
if (DebugOpt)
printf("Found %d cpus, expecting %d\n", ncpu, NCpus);
pwrdom_mask = cpu_pwrdom_mask;
while (CPUMASK_TESTNZERO(pwrdom_mask)) {
dom_id = BSFCPUMASK(pwrdom_mask);
CPUMASK_NANDBIT(pwrdom_mask, dom_id);
dom = cpu_pwrdomain[dom_id];
if (dom != NULL)
free(dom);
}
return 0;
}
return 1;
}
int
main(int ac, char **av)
{
int ch;
int res;
char *sched = NULL;
char *cpustr = NULL;
char *sched_cpustr = NULL;
char *p = NULL;
cpumask_t cpumask;
int cpuid;
pid_t pid = getpid(); /* See usched_set(2) - BUGS */
CPUMASK_ASSZERO(cpumask);
while ((ch = getopt(ac, av, "d")) != -1) {
switch (ch) {
case 'd':
DebugOpt = 1;
break;
default:
usage();
/* NOTREACHED */
}
}
ac -= optind;
av += optind;
if (ac < 2) {
usage();
/* NOTREACHED */
}
sched_cpustr = strdup(av[0]);
sched = strsep(&sched_cpustr, ":");
if (strcmp(sched, "default") == 0)
fprintf(stderr, "Ignoring scheduler == \"default\": not implemented\n");
cpustr = strsep(&sched_cpustr, "");
if (strlen(sched) == 0 && cpustr == NULL) {
usage();
/* NOTREACHED */
}
/*
* XXX needs expanded support for > 64 cpus
*/
if (cpustr != NULL) {
uint64_t v;
v = (uint64_t)strtoull(cpustr, NULL, 0);
for (cpuid = 0; cpuid < (int)sizeof(v) * 8; ++cpuid) {
if (v & (1LU << cpuid))
CPUMASK_ORBIT(cpumask, cpuid);
}
}
if (strlen(sched) != 0) {
if (DebugOpt)
fprintf(stderr, "DEBUG: USCHED_SET_SCHEDULER: scheduler: %s\n", sched);
res = usched_set(pid, USCHED_SET_SCHEDULER, sched, strlen(sched));
if (res != 0) {
asprintf(&p, "usched_set(%d, USCHED_SET_SCHEDULER, \"%s\", %d)",
pid, sched, (int)strlen(sched));
perror(p);
exit(1);
}
}
if (CPUMASK_TESTNZERO(cpumask)) {
for (cpuid = 0; cpuid < (int)sizeof(cpumask) * 8; ++cpuid) {
if (CPUMASK_TESTBIT(cpumask, cpuid))
break;
}
if (DebugOpt) {
fprintf(stderr, "DEBUG: USCHED_SET_CPU: cpuid: %d\n",
cpuid);
}
res = usched_set(pid, USCHED_SET_CPU, &cpuid, sizeof(int));
if (res != 0) {
asprintf(&p, "usched_set(%d, USCHED_SET_CPU, &%d, %d)",
pid, cpuid, (int)sizeof(int));
perror(p);
exit(1);
}
CPUMASK_NANDBIT(cpumask, cpuid);
while (CPUMASK_TESTNZERO(cpumask)) {
++cpuid;
if (CPUMASK_TESTBIT(cpumask, cpuid) == 0)
continue;
CPUMASK_NANDBIT(cpumask, cpuid);
if (DebugOpt) {
fprintf(stderr,
"DEBUG: USCHED_ADD_CPU: cpuid: %d\n",
cpuid);
}
res = usched_set(pid, USCHED_ADD_CPU, &cpuid, sizeof(int));
if (res != 0) {
asprintf(&p, "usched_set(%d, USCHED_ADD_CPU, &%d, %d)",
pid, cpuid, (int)sizeof(int));
perror(p);
exit(1);
}
}
}
execvp(av[1], av + 1);
exit(1);
}
/*
* Called with a critical section held and interrupts enabled.
*/
int
pmap_inval_intr(cpumask_t *cpumaskp, int toolong)
{
globaldata_t gd = mycpu;
pmap_inval_info_t *info;
int loopme = 0;
int cpu;
cpumask_t cpumask;
/*
* Check all cpus for invalidations we may need to service.
*/
cpu_ccfence();
cpu = gd->gd_cpuid;
cpumask = *cpumaskp;
while (CPUMASK_TESTNZERO(cpumask)) {
int n = BSFCPUMASK(cpumask);
#ifdef LOOPRECOVER
KKASSERT(n >= 0 && n < MAXCPU);
#endif
CPUMASK_NANDBIT(cpumask, n);
info = &invinfo[n];
/*
* Due to interrupts/races we can catch a new operation
* in an older interrupt. A fence is needed once we detect
* the (not) done bit.
*/
if (!CPUMASK_TESTBIT(info->done, cpu))
continue;
cpu_lfence();
#ifdef LOOPRECOVER
if (toolong) {
kprintf("pminvl %d->%d %08jx %08jx mode=%d\n",
cpu, n, info->done.ary[0], info->mask.ary[0],
info->mode);
}
#endif
/*
* info->mask and info->done always contain the originating
* cpu until the originator is done. Targets may still be
* present in info->done after the originator is done (they
* will be finishing up their loops).
*
* Clear info->mask bits on other cpus to indicate that they
* have quiesced (entered the loop). Once the other mask bits
* are clear we can execute the operation on the original,
* then clear the mask and done bits on the originator. The
* targets will then finish up their side and clear their
* done bits.
*
* The command is considered 100% done when all done bits have
* been cleared.
*/
if (n != cpu) {
/*
* Command state machine for 'other' cpus.
*/
if (CPUMASK_TESTBIT(info->mask, cpu)) {
/*
* Other cpu indicate to originator that they
* are quiesced.
*/
ATOMIC_CPUMASK_NANDBIT(info->mask, cpu);
loopme = 1;
} else if (info->ptep &&
CPUMASK_TESTBIT(info->mask, n)) {
/*
* Other cpu must wait for the originator (n)
* to complete its command if ptep is not NULL.
*/
loopme = 1;
} else {
/*
* Other cpu detects that the originator has
* completed its command, or there was no
* command.
*
* Now that the page table entry has changed,
* we can follow up with our own invalidation.
*/
vm_offset_t va = info->va;
int npgs;
if (va == (vm_offset_t)-1 ||
info->npgs > MAX_INVAL_PAGES) {
cpu_invltlb();
} else {
for (npgs = info->npgs; npgs; --npgs) {
cpu_invlpg((void *)va);
va += PAGE_SIZE;
}
}
ATOMIC_CPUMASK_NANDBIT(info->done, cpu);
/* info invalid now */
/* loopme left alone */
}
} else if (CPUMASK_TESTBIT(info->mask, cpu)) {
/*
* Originator is waiting for other cpus
*/
if (CPUMASK_CMPMASKNEQ(info->mask, gd->gd_cpumask)) {
/*
* Originator waits for other cpus to enter
* their loop (aka quiesce).
*
* If this bugs out the IPI may have been lost,
* try to reissue by resetting our own
* reentrancy bit and clearing the smurf mask
* for the cpus that did not respond, then
* reissuing the IPI.
*/
loopme = 1;
#ifdef LOOPRECOVER
if (loopwdog(info)) {
info->failed = 1;
loopdebug("C", info);
/* XXX recover from possible bug */
mdcpu->gd_xinvaltlb = 0;
ATOMIC_CPUMASK_NANDMASK(smp_smurf_mask,
info->mask);
cpu_disable_intr();
smp_invlpg(&smp_active_mask);
/*
* Force outer-loop retest of Xinvltlb
* requests (see mp_machdep.c).
*/
mdcpu->gd_xinvaltlb = 2;
cpu_enable_intr();
}
#endif
} else {
/*
* Originator executes operation and clears
* mask to allow other cpus to finish.
*/
KKASSERT(info->mode != INVDONE);
if (info->mode == INVSTORE) {
if (info->ptep)
info->opte = atomic_swap_long(info->ptep, info->npte);
CHECKSIGMASK(info);
ATOMIC_CPUMASK_NANDBIT(info->mask, cpu);
CHECKSIGMASK(info);
} else {
if (atomic_cmpset_long(info->ptep,
info->opte, info->npte)) {
info->success = 1;
} else {
info->success = 0;
}
CHECKSIGMASK(info);
ATOMIC_CPUMASK_NANDBIT(info->mask, cpu);
CHECKSIGMASK(info);
}
loopme = 1;
}
} else {
/*
* Originator does not have to wait for the other
* cpus to finish. It clears its done bit. A new
* command will not be initiated by the originator
* until the other cpus have cleared their done bits
* (asynchronously).
*/
vm_offset_t va = info->va;
int npgs;
if (va == (vm_offset_t)-1 ||
info->npgs > MAX_INVAL_PAGES) {
cpu_invltlb();
} else {
for (npgs = info->npgs; npgs; --npgs) {
cpu_invlpg((void *)va);
va += PAGE_SIZE;
}
}
/* leave loopme alone */
/* other cpus may still be finishing up */
/* can't race originator since that's us */
info->mode = INVDONE;
ATOMIC_CPUMASK_NANDBIT(info->done, cpu);
}
}
return loopme;
}
