int main(int argc, char** argv)
{
struct timeval start_tv = {0};
struct timeval end_tv = {0};
long usec_diff;
long nr_ctx_switches;
void *join_ret;
if (argc > 1)
nr_switch_loops = strtol(argv[1], 0, 10);
printf("Making 2 threads of %d switches each\n", nr_switch_loops);
pthread_can_vcore_request(FALSE); /* 2LS won't manage vcores */
pthread_need_tls(FALSE);
pthread_lib_init(); /* gives us one vcore */
/* each is passed the other's pthread_t. th1 starts the switching. */
if (pthread_create(&th1, NULL, &switch_thread, &th2))
perror("pth_create 1 failed");
/* thread 2 is created, but not put on the runnable list */
if (__pthread_create(&th2, NULL, &switch_thread, &th1))
perror("pth_create 2 failed");
if (gettimeofday(&start_tv, 0))
perror("Start time error...");
ready = TRUE; /* signal to any spinning uthreads to start */
pthread_join(th1, &join_ret);
pthread_join(th2, &join_ret);
if (gettimeofday(&end_tv, 0))
perror("End time error...");
nr_ctx_switches = 2 * nr_switch_loops;
usec_diff = (end_tv.tv_sec - start_tv.tv_sec) * 1000000 +
(end_tv.tv_usec - start_tv.tv_usec);
printf("Done: %d loops\n", nr_switch_loops);
printf("Nr context switches: %ld\n", nr_ctx_switches);
printf("Time to run: %ld usec\n", usec_diff);
printf("Context switch latency: %d nsec\n",
(int)(1000LL*usec_diff / nr_ctx_switches));
printf("Context switches / sec: %d\n\n",
(int)(1000000LL*nr_ctx_switches / usec_diff));
}
int main(int argc, char** argv)
{
struct timeval start_tv = {0};
struct timeval end_tv = {0};
long usec_diff;
long nr_ctx_switches;
if (argc > 1)
nr_yield_threads = strtol(argv[1], 0, 10);
if (argc > 2)
nr_yield_loops = strtol(argv[2], 0, 10);
if (argc > 3)
nr_vcores = strtol(argv[3], 0, 10);
if (argc > 4)
amt_fake_work = strtol(argv[4], 0, 10);
nr_yield_threads = MIN(nr_yield_threads, MAX_NR_TEST_THREADS);
printf("Making %d threads of %d loops each, on %d vcore(s), %d work\n",
nr_yield_threads, nr_yield_loops, nr_vcores, amt_fake_work);
/* OS dependent prep work */
if (nr_vcores) {
/* Only do the vcore trickery if requested */
pthread_can_vcore_request(FALSE); /* 2LS won't manage vcores */
pthread_need_tls(FALSE);
pthread_lib_init(); /* gives us one vcore */
vcore_request(nr_vcores - 1); /* ghetto incremental interface */
for (int i = 0; i < nr_vcores; i++) {
printd("Vcore %d mapped to pcore %d\n", i,
__procinfo.vcoremap[i].pcoreid);
}
}
/* create and join on yield */
for (int i = 0; i < nr_yield_threads; i++) {
printf_safe("[A] About to create thread %d\n", i);
if (pthread_create(&my_threads[i], NULL, &yield_thread, NULL))
perror("pth_create failed");
}
if (gettimeofday(&start_tv, 0))
perror("Start time error...");
ready = TRUE; /* signal to any spinning uthreads to start */
for (int i = 0; i < nr_yield_threads; i++) {
printf_safe("[A] About to join on thread %d(%p)\n", i, my_threads[i]);
pthread_join(my_threads[i], &my_retvals[i]);
printf_safe("[A] Successfully joined on thread %d (retval: %p)\n", i,
my_retvals[i]);
}
if (gettimeofday(&end_tv, 0))
perror("End time error...");
nr_ctx_switches = nr_yield_threads * nr_yield_loops;
usec_diff = (end_tv.tv_sec - start_tv.tv_sec) * 1000000 +
(end_tv.tv_usec - start_tv.tv_usec);
printf("Done: %d uthreads, %d loops, %d vcores, %d work\n",
nr_yield_threads, nr_yield_loops, nr_vcores, amt_fake_work);
printf("Nr context switches: %d\n", nr_ctx_switches);
printf("Time to run: %d usec\n", usec_diff);
if (nr_vcores == 1)
printf("Context switch latency: %d nsec\n",
(int)(1000LL*usec_diff / nr_ctx_switches));
printf("Context switches / sec: %d\n\n",
(int)(1000000LL*nr_ctx_switches / usec_diff));
}
int main(int argc, char** argv)
{
struct timeval start_tv = {0};
struct timeval end_tv = {0};
long usec_diff;
long nr_ctx_switches;
if (argc > 1)
nr_yield_threads = strtol(argv[1], 0, 10);
if (argc > 2)
nr_yield_loops = strtol(argv[2], 0, 10);
if (argc > 3)
nr_vcores = strtol(argv[3], 0, 10);
if (argc > 4)
amt_fake_work = strtol(argv[4], 0, 10);
nr_yield_threads = MIN(nr_yield_threads, MAX_NR_TEST_THREADS);
printf("Making %d threads of %d loops each, on %d vcore(s), %d work\n",
nr_yield_threads, nr_yield_loops, nr_vcores, amt_fake_work);
/* OS dependent prep work */
#ifdef __ros__
if (nr_vcores) {
/* Only do the vcore trickery if requested */
parlib_never_yield = TRUE;
pthread_need_tls(FALSE);
pthread_mcp_init(); /* gives us one vcore */
vcore_request_total(nr_vcores);
parlib_never_vc_request = TRUE;
for (int i = 0; i < nr_vcores; i++) {
printf_safe("Vcore %d mapped to pcore %d\n", i,
__procinfo.vcoremap[i].pcoreid);
}
}
struct uth_join_request *join_reqs;
join_reqs = malloc(nr_yield_threads * sizeof(struct uth_join_request));
for (int i = 0; i < nr_yield_threads; i++)
join_reqs[i].retval_loc = &my_retvals[i];
assert(join_reqs);
#endif /* __ros__ */
pthread_barrier_init(&barrier, NULL, nr_yield_threads);
/* create and join on yield */
for (int i = 0; i < nr_yield_threads; i++) {
printf_safe("[A] About to create thread %d\n", i);
if (pthread_create(&my_threads[i], NULL, &yield_thread, NULL))
perror("pth_create failed");
}
if (gettimeofday(&start_tv, 0))
perror("Start time error...");
/* Akaros supports parallel join */
#ifdef __ros__
for (int i = 0; i < nr_yield_threads; i++)
join_reqs[i].uth = (struct uthread*)my_threads[i];
uthread_join_arr(join_reqs, nr_yield_threads);
#else
for (int i = 0; i < nr_yield_threads; i++) {
printf_safe("[A] About to join on thread %d(%p)\n",
i, my_threads[i]);
pthread_join(my_threads[i], &my_retvals[i]);
printf_safe("[A] Successful join on thread %d (retval: %p)\n",
i, my_retvals[i]);
}
#endif
if (gettimeofday(&end_tv, 0))
perror("End time error...");
nr_ctx_switches = nr_yield_threads * nr_yield_loops;
usec_diff = (end_tv.tv_sec - start_tv.tv_sec) * 1000000 +
(end_tv.tv_usec - start_tv.tv_usec);
printf("Done: %d uthreads, %d loops, %d vcores, %d work\n",
nr_yield_threads, nr_yield_loops, nr_vcores, amt_fake_work);
printf("Nr context switches: %ld\n", nr_ctx_switches);
printf("Time to run: %ld usec\n", usec_diff);
if (nr_vcores == 1)
printf("Context switch latency: %d nsec\n",
(int)(1000LL*usec_diff / nr_ctx_switches));
printf("Context switches / sec: %d\n\n",
(int)(1000000LL*nr_ctx_switches / usec_diff));
}
int main(int argc, char **argv)
{
int i, amt;
int nr_gpcs = 1;
uint64_t entry;
int fd = open("#c/sysctl", O_RDWR), ret;
int kfd = -1;
bool smallkernel = false;
void * x;
static char cmd[512];
if (fd < 0) {
perror("#c/sysctl");
exit(1);
}
argc--,argv++;
if (argc != 2) {
fprintf(stderr, "Usage: %s vmimage entrypoint\n", argv[0]);
exit(1);
}
entry = strtoull(argv[1], 0, 0);
kfd = open(argv[0], O_RDONLY);
if (kfd < 0) {
perror(argv[0]);
exit(1);
}
if (ros_syscall(SYS_setup_vmm, nr_gpcs, 0, 0, 0, 0, 0) != nr_gpcs) {
perror("Guest pcore setup failed");
exit(1);
}
my_threads = malloc(sizeof(pthread_t) * nr_threads);
my_retvals = malloc(sizeof(void*) * nr_threads);
if (!(my_retvals && my_threads))
perror("Init threads/malloc");
pthread_can_vcore_request(FALSE); /* 2LS won't manage vcores */
pthread_need_tls(FALSE);
pthread_mcp_init(); /* gives us one vcore */
vcore_request(nr_threads - 1); /* ghetto incremental interface */
for (int i = 0; i < nr_threads; i++) {
x = __procinfo.vcoremap;
printf("%p\n", __procinfo.vcoremap);
printf("Vcore %d mapped to pcore %d\n", i,
__procinfo.vcoremap[i].pcoreid);
}
if (pthread_create(&my_threads[0], NULL, &talk_thread, NULL))
perror("pth_create failed");
// if (pthread_create(&my_threads[1], NULL, &fail, NULL))
// perror("pth_create failed");
printf("threads started\n");
if (0) for (int i = 0; i < nr_threads-1; i++) {
int ret;
if (pthread_join(my_threads[i], &my_retvals[i]))
perror("pth_join failed");
printf("%d %d\n", i, ret);
}
ret = syscall(33, 1);
if (ret < 0) {
perror("vm setup");
exit(1);
}
/* blob that is faulted in from the EPT first. we need this to be in low
* memory (not above the normal mmap_break), so the EPT can look it up.
* Note that we won't get 4096. The min is 1MB now, and ld is there. */
mmap_blob = mmap((int*)(15*1048576), 16 * 1048576, PROT_EXEC | PROT_READ | PROT_WRITE,
MAP_ANONYMOUS, -1, 0);
if (mmap_blob == MAP_FAILED) {
perror("Unable to mmap");
exit(1);
}
memset(mmap_blob, 0, 16*1048576);
// read in the kernel.
x = mmap_blob + 0x1000000;
for(;;) {
amt = read(kfd, x, 1048576);
if (amt < 0) {
perror("read");
exit(1);
}
if (amt == 0) {
break;
}
x += amt;
}
fprintf(stderr, "Read in %d bytes\n", x-mmap_blob);
p512 = mmap_blob;
p1 = &p512[512];
p2m = &p512[1024];
// We had thought to enter the kernel at the high address. But
// there's just too much state the kernel has to clean up to
// make this really work -- consider all the segment
// descriptors that have to move, etc. So we will enter the
// kernel in the low part of the address space, and let it
// work up its page tables and the other good fun. Map the
// kernel address space at low virtual, for 1G. It's ok to
// map memory we have no access to.
#define _2MiB 0x200000
p512[0] = (unsigned long long)p1 | 7;
// if only we could guarantee 1G pages everywhere!
p1[0] = /*0x87; */(unsigned long long)p2m | 7;
for(i = 0; i < 16; i++) {
p2m[i] = 0x87 | i * _2MiB;
printf("pwm[%d] = 0x%llx\n", i, p2m[i]);
}
printf("p512 %p p512[0] is 0x%lx p1 %p p1[0] is 0x%x\n", p512, p512[0], p1, p1[0]);
sprintf(cmd, "V 0x%llx 0x%llx 0x%llx", entry, (unsigned long long) &stack[1024], (unsigned long long) p512);
printf("Writing command :%s:\n", cmd);
ret = write(fd, cmd, strlen(cmd));
if (ret != strlen(cmd)) {
perror(cmd);
}
sprintf(cmd, "V 0 0 0");
while (! done) {
char c[1];
printf("hit return\n"); read(0, c, 1);
if (debug)
fprintf(stderr, "RESUME\n");
ret = write(fd, cmd, strlen(cmd));
if (ret != strlen(cmd)) {
perror(cmd);
}
}
return 0;
}