int spin_check(spin_t *s)
{
/* Check whether a timeout has taken place. Return TRUE if the caller
* should continue spinning, and FALSE if a timeout has occurred. The
* implementation assumes that it is okay to spin a little bit too long
* (up to a full clock tick extra).
*/
u64_t cur_tsc, tsc_delta;
clock_t now, micro_delta;
switch (s->s_state) {
case STATE_INIT:
s->s_state = STATE_BASE_TS;
break;
case STATE_BASE_TS:
s->s_state = STATE_TS;
read_tsc_64(&s->s_base_tsc);
break;
case STATE_TS:
read_tsc_64(&cur_tsc);
tsc_delta = sub64(cur_tsc, s->s_base_tsc);
micro_delta = tsc_64_to_micros(tsc_delta);
if (micro_delta >= s->s_usecs) {
s->s_timeout = TRUE;
return FALSE;
}
if (micro_delta >= TSC_SPIN) {
s->s_usecs -= micro_delta;
getticks(&s->s_base_uptime);
s->s_state = STATE_UPTIME;
}
break;
case STATE_UPTIME:
getticks(&now);
/* We assume that sys_hz() caches its return value. */
micro_delta = ((now - s->s_base_uptime) * 1000 / sys_hz()) *
1000;
if (micro_delta >= s->s_usecs) {
s->s_timeout = TRUE;
return FALSE;
}
break;
default:
panic("spin_check: invalid state %d", s->s_state);
}
return TRUE;
}
PUBLIC void busy_delay_ms(int ms)
{
u64_t cycles = ms_2_cpu_time(ms), tsc0, tsc, tsc1;
read_tsc_64(&tsc0);
tsc1 = tsc0 + cycles;
do { read_tsc_64(&tsc); } while(tsc < tsc1);
return;
}
/* returns 0 on exit */
int main(int argc, char *argv[]) {
/* argument 1 is the number of iterations to run */
/* argument 2 is the argument to nice() */
/* PI = 4 * (1/1 - 1/3 + 1/5 - 1/7 + 1/9 - 1/11 ...) */
int i;
double sum = 0;
int denom, numer;
time_t current;
time_t start = time(NULL);
int seconds = 0;
pid_t process_id = getpid();
int retval, nice_val, iters;
u64_t s,e,diff;
double elapsed;
unsigned long max;
if (argc < 3) {
printf("usage: cpu iterations tickets\n");
return 1;
}
iters = atoi(argv[1]);
nice_val = atoi(argv[2]);
retval = nice(nice_val);
if (retval == -1) {
printf("Error calling nice()\n");
return 1;
}
printf("Process %d at nice %d\n", process_id, nice_val);
read_tsc_64(&s);
for (i = 1; i<iters; ++i) {
denom = 2 * i - 1;
numer;
if (i % 2)
numer = -1;
else
numer = 1;
sum += ((double)numer / (double)denom);
current = time(NULL);
if (current - start > seconds) {
seconds = current - start;
printf("Process %d has been running for %d seconds.\n", process_id, seconds);
/*if (seconds == 10)
break;*/
}
}
read_tsc_64(&e);
diff = sub64(e, s);
max = -1;
elapsed = (double)diff.hi + (double)(diff.lo/100000) / (double)(max/100000);
printf("CPU process %d (nice %d) calculated pi as %f at %f time units\n", process_id, nice_val, -4 * sum, elapsed);
return 0;
}
/*===========================================================================*
* idle *
*===========================================================================*/
PRIVATE void idle()
{
/* This function is called whenever there is no work to do.
* Halt the CPU, and measure how many timestamp counter ticks are
* spent not doing anything. This allows test setups to measure
* the CPU utiliziation of certain workloads with high precision.
*/
#ifdef CONFIG_IDLE_TSC
u64_t idle_start;
read_tsc_64(&idle_start);
idle_active = 1;
#endif
halt_cpu();
#ifdef CONFIG_IDLE_TSC
if (idle_active) {
IDLE_STOP;
printf("Kernel: idle active after resuming CPU\n");
}
idle_tsc = add64(idle_tsc, sub64(idle_stop, idle_start));
#endif
}
void cycles_accounting_init(void)
{
read_tsc_64(get_cpu_var_ptr(cpu, tsc_ctr_switch));
make_zero64(get_cpu_var(cpu, cpu_last_tsc));
make_zero64(get_cpu_var(cpu, cpu_last_idle));
}
void init_scheduling(void)
{
u64_t r;
balance_timeout = BALANCE_TIMEOUT * sys_hz();
init_timer(&sched_timer);
set_timer(&sched_timer, balance_timeout, balance_queues, 0);
read_tsc_64(&r);
srandom((unsigned)r);
}
double getidle(void)
{
u64_t stop, idle2;
u64_t idelta, tdelta;
double ifp, tfp, rfp;
int r;
if (!running) {
if ((r = sys_getidletsc(&idle)) != OK)
return -1.0;
running = 1;
read_tsc_64(&start);
return 0.0;
}
else {
read_tsc_64(&stop);
running = 0;
if ((r = sys_getidletsc(&idle2)) != OK)
return -1.0;
idelta = sub64(idle2, idle);
tdelta = sub64(stop, start);
if (cmp64(idelta, tdelta) >= 0)
return 100.0;
ifp = make_double(idelta);
tfp = make_double(tdelta);
rfp = ifp / tfp * 100.0;
if (rfp < 0.0) rfp = 0.0;
else if (rfp > 100.0) rfp = 100.0;
return rfp;
}
running = !running;
}
void init_scheduling(void)
{
/*Lottery Scheduling*/
u64_t r;
balance_timeout = BALANCE_TIMEOUT * sys_hz();
init_timer(&sched_timer);
set_timer(&sched_timer, balance_timeout, balance_queues, 0);
/*Lottery Scheduling*/
read_tsc_64(&r);
srandom();
}
PRIVATE int do_lottery() {
struct schedproc *rmp;
int rv, proc_nr;
int total_tickets = 0;
u64_t tsc;
int winner;
/* count the total number of tickets in all processes */
/* we really should have a global to keep track of this total */
/* rather than computing it every time */
for (proc_nr = 0, rmp = schedproc; proc_nr < NR_PROCS; ++proc_nr, ++rmp)
if (rmp->priority == HOLDING_Q && rmp->flags == (IN_USE | USER_PROCESS)) /* winnable? */
total_tickets += rmp->tickets;
if (!total_tickets) /* there were no winnable processes */
return OK;
/* generate a "random" winning ticket */
/* lower bits of time stamp counter are random enough */
/* and much faster then random() */
read_tsc_64(&tsc);
winner = tsc.lo % total_tickets + 1;
/* now find the process with the winning ticket */
for (proc_nr = 0, rmp = schedproc; proc_nr < NR_PROCS; ++proc_nr, ++rmp) {
if (rmp->priority == HOLDING_Q && rmp->flags == (IN_USE | USER_PROCESS)) /* winnable? */
winner -= rmp->tickets;
if (winner <= 0)
break;
}
printf("Process %d won with %d(%d) of %d tickets\n", proc_nr, rmp->tickets, rmp->blocking, total_tickets);
/* schedule new winning process */
rmp->priority = WINNING_Q;
rmp->time_slice = USER_QUANTUM;
/*if (rmp->blocking)
rmp->time_slice = USER_QUANTUM / (rmp->blocking + 1); */
rmp->blocking = 0;
if ((rv = schedule_process(rmp)) != OK)
return rv;
return OK;
}
PRIVATE int calib_cpu_handler(irq_hook_t * UNUSED(hook))
{
u64_t tsc;
probe_ticks++;
read_tsc_64(&tsc);
if (probe_ticks == 1) {
tsc0 = tsc;
}
else if (probe_ticks == PROBE_TICKS) {
tsc1 = tsc;
}
/* just in case we are in an SMP single cpu fallback mode */
BKL_UNLOCK();
return 1;
}
void context_stop(struct proc * p)
{
u64_t tsc;
u32_t tsc_delta;
u64_t * __tsc_ctr_switch = get_cpulocal_var_ptr(tsc_ctr_switch);
read_tsc_64(&tsc);
assert(tsc >= *__tsc_ctr_switch);
tsc_delta = tsc - *__tsc_ctr_switch;
p->p_cycles += tsc_delta;
if(kbill_ipc) {
kbill_ipc->p_kipc_cycles =
add64(kbill_ipc->p_kipc_cycles, tsc_delta);
kbill_ipc = NULL;
}
if(kbill_kcall) {
kbill_kcall->p_kcall_cycles =
add64(kbill_kcall->p_kcall_cycles, tsc_delta);
kbill_kcall = NULL;
}
/*
* deduct the just consumed cpu cycles from the cpu time left for this
* process during its current quantum. Skip IDLE and other pseudo kernel
* tasks
*/
if (p->p_endpoint >= 0) {
#if DEBUG_RACE
p->p_cpu_time_left = 0;
#else
if (tsc_delta < p->p_cpu_time_left) {
p->p_cpu_time_left -= tsc_delta;
} else p->p_cpu_time_left = 0;
#endif
}
*__tsc_ctr_switch = tsc;
}
PUBLIC short cpu_load(void)
{
u64_t current_tsc, *current_idle;
u64_t tsc_delta, idle_delta, busy;
struct proc *idle;
short load;
#ifdef CONFIG_SMP
unsigned cpu = cpuid;
#endif
u64_t *last_tsc, *last_idle;
last_tsc = get_cpu_var_ptr(cpu, cpu_last_tsc);
last_idle = get_cpu_var_ptr(cpu, cpu_last_idle);
idle = get_cpu_var_ptr(cpu, idle_proc);;
read_tsc_64(¤t_tsc);
current_idle = &idle->p_cycles; /* ptr to idle proc */
/* calculate load since last cpu_load invocation */
if (!is_zero64(*last_tsc)) {
tsc_delta = sub64(current_tsc, *last_tsc);
idle_delta = sub64(*current_idle, *last_idle);
busy = sub64(tsc_delta, idle_delta);
busy = mul64(busy, make64(100, 0));
load = ex64lo(div64(busy, tsc_delta));
if (load > 100)
load = 100;
} else
load = 0;
*last_tsc = current_tsc;
*last_idle = *current_idle;
return load;
}
PUBLIC void context_stop(struct proc * p)
{
u64_t tsc, tsc_delta;
u64_t * __tsc_ctr_switch = get_cpulocal_var_ptr(tsc_ctr_switch);
#ifdef CONFIG_SMP
unsigned cpu = cpuid;
/*
* This function is called only if we switch from kernel to user or idle
* or back. Therefore this is a perfect location to place the big kernel
* lock which will hopefully disappear soon.
*
* If we stop accounting for KERNEL we must unlock the BKL. If account
* for IDLE we must not hold the lock
*/
if (p == proc_addr(KERNEL)) {
u64_t tmp;
read_tsc_64(&tsc);
tmp = sub64(tsc, *__tsc_ctr_switch);
kernel_ticks[cpu] = add64(kernel_ticks[cpu], tmp);
p->p_cycles = add64(p->p_cycles, tmp);
BKL_UNLOCK();
} else {
u64_t bkl_tsc;
atomic_t succ;
read_tsc_64(&bkl_tsc);
/* this only gives a good estimate */
succ = big_kernel_lock.val;
BKL_LOCK();
read_tsc_64(&tsc);
bkl_ticks[cpu] = add64(bkl_ticks[cpu], sub64(tsc, bkl_tsc));
bkl_tries[cpu]++;
bkl_succ[cpu] += !(!(succ == 0));
p->p_cycles = add64(p->p_cycles, sub64(tsc, *__tsc_ctr_switch));
#ifdef CONFIG_SMP
/*
* Since at the time we got a scheduling IPI we might have been
* waiting for BKL already, we may miss it due to a similar IPI to
* the cpu which is already waiting for us to handle its. This
* results in a live-lock of these two cpus.
*
* Therefore we always check if there is one pending and if so,
* we handle it straight away so the other cpu can continue and
* we do not deadlock.
*/
smp_sched_handler();
#endif
}
#else
read_tsc_64(&tsc);
p->p_cycles = add64(p->p_cycles, sub64(tsc, *__tsc_ctr_switch));
#endif
tsc_delta = sub64(tsc, *__tsc_ctr_switch);
if(kbill_ipc) {
kbill_ipc->p_kipc_cycles =
add64(kbill_ipc->p_kipc_cycles, tsc_delta);
kbill_ipc = NULL;
}
if(kbill_kcall) {
kbill_kcall->p_kcall_cycles =
add64(kbill_kcall->p_kcall_cycles, tsc_delta);
kbill_kcall = NULL;
}
/*
* deduct the just consumed cpu cycles from the cpu time left for this
* process during its current quantum. Skip IDLE and other pseudo kernel
* tasks
*/
if (p->p_endpoint >= 0) {
#if DEBUG_RACE
make_zero64(p->p_cpu_time_left);
#else
/* if (tsc_delta < p->p_cpu_time_left) in 64bit */
if (ex64hi(tsc_delta) < ex64hi(p->p_cpu_time_left) ||
(ex64hi(tsc_delta) == ex64hi(p->p_cpu_time_left) &&
ex64lo(tsc_delta) < ex64lo(p->p_cpu_time_left)))
p->p_cpu_time_left = sub64(p->p_cpu_time_left, tsc_delta);
else {
make_zero64(p->p_cpu_time_left);
}
#endif
}
*__tsc_ctr_switch = tsc;
}
PUBLIC int init_local_timer(unsigned freq)
{
#ifdef USE_APIC
/* if we know the address, lapic is enabled and we should use it */
if (lapic_addr) {
unsigned cpu = cpuid;
tsc_per_ms[cpu] = div64u(cpu_get_freq(cpu), 1000);
lapic_set_timer_one_shot(1000000/system_hz);
} else
{
BOOT_VERBOSE(printf("Initiating legacy i8253 timer\n"));
#else
{
#endif
init_8253A_timer(freq);
estimate_cpu_freq();
/* always only 1 cpu in the system */
tsc_per_ms[0] = div64u(cpu_get_freq(0), 1000);
}
return 0;
}
PUBLIC void stop_local_timer(void)
{
#ifdef USE_APIC
if (lapic_addr) {
lapic_stop_timer();
apic_eoi();
} else
#endif
{
stop_8253A_timer();
}
}
PUBLIC void restart_local_timer(void)
{
#ifdef USE_APIC
if (lapic_addr) {
lapic_restart_timer();
}
#endif
}
PUBLIC int register_local_timer_handler(const irq_handler_t handler)
{
#ifdef USE_APIC
if (lapic_addr) {
/* Using APIC, it is configured in apic_idt_init() */
BOOT_VERBOSE(printf("Using LAPIC timer as tick source\n"));
} else
#endif
{
/* Using PIC, Initialize the CLOCK's interrupt hook. */
pic_timer_hook.proc_nr_e = NONE;
pic_timer_hook.irq = CLOCK_IRQ;
put_irq_handler(&pic_timer_hook, CLOCK_IRQ, handler);
}
return 0;
}
PUBLIC void cycles_accounting_init(void)
{
#ifdef CONFIG_SMP
unsigned cpu = cpuid;
#endif
read_tsc_64(get_cpu_var_ptr(cpu, tsc_ctr_switch));
make_zero64(get_cpu_var(cpu, cpu_last_tsc));
make_zero64(get_cpu_var(cpu, cpu_last_idle));
}
PUBLIC void context_stop(struct proc * p)
{
u64_t tsc, tsc_delta;
u64_t * __tsc_ctr_switch = get_cpulocal_var_ptr(tsc_ctr_switch);
#ifdef CONFIG_SMP
unsigned cpu = cpuid;
/*
* This function is called only if we switch from kernel to user or idle
* or back. Therefore this is a perfect location to place the big kernel
* lock which will hopefully disappear soon.
*
* If we stop accounting for KERNEL we must unlock the BKL. If account
* for IDLE we must not hold the lock
*/
if (p == proc_addr(KERNEL)) {
u64_t tmp;
read_tsc_64(&tsc);
tmp = sub64(tsc, *__tsc_ctr_switch);
kernel_ticks[cpu] = add64(kernel_ticks[cpu], tmp);
p->p_cycles = add64(p->p_cycles, tmp);
BKL_UNLOCK();
} else {
u64_t bkl_tsc;
atomic_t succ;
read_tsc_64(&bkl_tsc);
/* this only gives a good estimate */
succ = big_kernel_lock.val;
BKL_LOCK();
read_tsc_64(&tsc);
bkl_ticks[cpu] = add64(bkl_ticks[cpu], sub64(tsc, bkl_tsc));
bkl_tries[cpu]++;
bkl_succ[cpu] += !(!(succ == 0));
p->p_cycles = add64(p->p_cycles, sub64(tsc, *__tsc_ctr_switch));
}
#else
read_tsc_64(&tsc);
p->p_cycles = add64(p->p_cycles, sub64(tsc, *__tsc_ctr_switch));
#endif
tsc_delta = sub64(tsc, *__tsc_ctr_switch);
if(kbill_ipc) {
kbill_ipc->p_kipc_cycles =
add64(kbill_ipc->p_kipc_cycles, tsc_delta);
kbill_ipc = NULL;
}
if(kbill_kcall) {
kbill_kcall->p_kcall_cycles =
add64(kbill_kcall->p_kcall_cycles, tsc_delta);
kbill_kcall = NULL;
}
/*
* deduct the just consumed cpu cycles from the cpu time left for this
* process during its current quantum. Skip IDLE and other pseudo kernel
* tasks
*/
if (p->p_endpoint >= 0) {
#if DEBUG_RACE
make_zero64(p->p_cpu_time_left);
#else
/* if (tsc_delta < p->p_cpu_time_left) in 64bit */
if (ex64hi(tsc_delta) < ex64hi(p->p_cpu_time_left) ||
(ex64hi(tsc_delta) == ex64hi(p->p_cpu_time_left) &&
ex64lo(tsc_delta) < ex64lo(p->p_cpu_time_left)))
p->p_cpu_time_left = sub64(p->p_cpu_time_left, tsc_delta);
else {
make_zero64(p->p_cpu_time_left);
}
#endif
}
*__tsc_ctr_switch = tsc;
}
/*===========================================================================*
* main *
*===========================================================================*/
int main(int argc, char **argv)
{
endpoint_t ep_self, ep_child;
size_t size = BUF_SIZE;
int i, r, pid;
int status;
u64_t start, end, diff;
double micros;
char nr_pages_str[10], is_map_str[2], is_write_str[2];
int nr_pages, is_map, is_write;
/* SEF local startup. */
env_setargs(argc, argv);
sef_local_startup();
/* Parse the command line. */
r = env_get_param("pages", nr_pages_str, sizeof(nr_pages_str));
errno = 0;
nr_pages = atoi(nr_pages_str);
if (r != OK || errno || nr_pages <=0) {
exit_usage();
}
if(nr_pages > TEST_PAGE_NUM) {
printf("REQUESTOR: too many pages. Max allowed: %d\n",
TEST_PAGE_NUM);
exit_usage();
}
r = env_get_param("map", is_map_str, sizeof(is_map_str));
errno = 0;
is_map = atoi(is_map_str);
if (r != OK || errno || (is_map!=0 && is_map!=1)) {
exit_usage();
}
r = env_get_param("write", is_write_str, sizeof(is_write_str));
errno = 0;
is_write = atoi(is_write_str);
if (r != OK || errno || (is_write!=0 && is_write!=1)) {
exit_usage();
}
printf("REQUESTOR: Running tests with pages=%d map=%d write=%d...\n",
nr_pages, is_map, is_write);
/* Prepare work. */
buf = (char*) CLICK_CEIL(buf_buf);
fid_get = open(FIFO_GRANTOR, O_RDONLY);
fid_send = open(FIFO_REQUESTOR, O_WRONLY);
if(fid_get < 0 || fid_send < 0) {
printf("REQUESTOR: can't open fifo files.\n");
return 1;
}
/* Send the endpoint to the granter, in order to let him to
* create the grant.
*/
ep_self = getprocnr();
write(fid_send, &ep_self, sizeof(ep_self));
dprint("REQUESTOR: sending my endpoint: %d\n", ep_self);
/* Get the granter's endpoint and gid. */
read(fid_get, &ep_granter, sizeof(ep_granter));
read(fid_get, &gid, sizeof(gid));
dprint("REQUESTOR: getting granter's endpoint %d and gid %d\n",
ep_granter, gid);
FIFO_WAIT(fid_get);
diff = make64(0, 0);
if(is_map) {
/* Test safemap. */
for(i=0;i<NR_TEST_ITERATIONS;i++) {
read_tsc_64(&start);
r = sys_safemap(ep_granter, gid, 0, (long)buf,
nr_pages*CLICK_SIZE, D, 1);
if(r != OK) {
printf("REQUESTOR: safemap error: %d\n", r);
return 1;
}
read_write_buff(buf, nr_pages*CLICK_SIZE, is_write);
read_tsc_64(&end);
diff = add64(diff, (sub64(end, start)));
r = sys_safeunmap(D, (long)buf);
if(r != OK) {
printf("REQUESTOR: safeunmap error: %d\n", r);
return 1;
}
}
micros = ((double)tsc_64_to_micros(diff))
/ (NR_TEST_ITERATIONS*nr_pages);
REPORT_TEST("REQUESTOR", "SAFEMAP", micros);
}
else {
/* Test safecopy. */
for(i=0;i<NR_TEST_ITERATIONS;i++) {
read_tsc_64(&start);
r = sys_safecopyfrom(ep_granter, gid, 0, (long)buf,
nr_pages*CLICK_SIZE, D);
if(r != OK) {
printf("REQUESTOR: safecopy error: %d\n", r);
return 1;
}
read_write_buff(buf, nr_pages*CLICK_SIZE, is_write);
read_tsc_64(&end);
diff = add64(diff, (sub64(end, start)));
}
micros = ((double)tsc_64_to_micros(diff))
/ (NR_TEST_ITERATIONS*nr_pages);
REPORT_TEST("REQUESTOR", "SAFECOPY", micros);
}
FIFO_NOTIFY(fid_send);
return 0;
}
static struct block *block_alloc(size_t size)
{
struct block *block;
u8_t *dataptr, *p, *ptr;
unsigned page_index, page_index_max;
size_t sizerem, totalsize;
u64_t tsc;
LOG(("block_alloc; size=0x%x\n", size));
assert(size > 0);
/* round size up to machine word size */
sizerem = size % sizeof(long);
if (sizerem)
size += sizeof(long) - sizerem;
/* initialize address range */
if (!ptr_min && !ptr_max) {
/* keep a safe distance from areas that are in use:
* - 4MB from the break (should not change if traditional
* malloc is not used so a small margin is sufficient
* - 256MB from the stack (big margin because memory beyond
* this may be allocated by mmap when the address space
* starts to fill up)
*/
ptr_min = page_round_up_ptr((u8_t *) sbrk(0) + 0x400000);
ptr_max = page_round_down_ptr((u8_t *) &size - 0x10000000);
}
assert(ptr_min);
assert(ptr_max);
assert(ptr_min < ptr_max);
/* select address at random */
read_tsc_64(&tsc);
totalsize = block_get_totalsize(size);
page_index_max = (ptr_max - ptr_min - totalsize) / PAGE_SIZE;
page_index = (page_index_max > 0) ? (tsc.lo % page_index_max) : 0;
ptr = ptr_min + page_index * PAGE_SIZE;
/* allocate block */
block = (struct block *) mmap(
ptr, /* addr */
totalsize, /* len */
PROT_READ|PROT_WRITE, /* prot */
MAP_PREALLOC, /* flags */
-1, /* fd */
0); /* offset */
if (block == MAP_FAILED) {
/* mmap call failed */
abort();
}
/* block may not be at the requested location if that is in use */
if (ptr_min > (u8_t *) block)
ptr_min = (u8_t *) block;
if (ptr_max < (u8_t *) block)
ptr_max = (u8_t *) block;
/* initialize block, including fillers */
block->size = size;
block->magic = block_compute_magic(block);
dataptr = block_get_dataptr(block);
for (p = (u8_t *) (block + 1); p < dataptr; p++)
*p = ((unsigned long) p & 0xff);
LOG(("block_alloc; block=0x%x\n", block));
return block;
}
