static cycles_t send_caps(void)
{
errval_t err;
cycles_t time_taken = 0;
srand(bench_tsc()); // random starting seed
for (int i=0; i<CAPS_PER_CORE; i++) {
coreid_t to_core;
do {
to_core = rand() % num_cores;
} while(to_core == my_coreid);
do {
cycles_t start = bench_tsc();
err = bindings[to_core]->tx_vtbl.send_cap(
bindings[to_core], NOP_CONT, my_caps[i]);
if (i >= 20 && i <= (CAPS_PER_CORE - 20)) { // avoid warmup / cooldown
time_taken += (bench_tsc() - start);
}
} while(redo_message(err));
if (err_is_fail(err)) {
DEBUG_ERR(err, "xcorecap: cap send failed\n");
abort();
}
}
return time_taken / (CAPS_PER_CORE - 40);
}
void experiment(coreid_t idx)
{
timestamps = malloc(sizeof(struct timestamps) * MAX_COUNT);
assert(timestamps != NULL);
struct bench_ump_binding *bu = (struct bench_ump_binding*)array[idx];
struct flounder_ump_state *fus = &bu->ump_state;
struct ump_chan *chan = &fus->chan;
struct ump_chan_state *send = &chan->send_chan;
struct ump_chan_state *recv = &chan->endpoint.chan;
printf("Running latency between core %"PRIuCOREID" and core %"PRIuCOREID"\n", my_core_id, idx);
/* Run experiment */
for (int i = 0; i < MAX_COUNT; i++) {
volatile struct ump_message *msg;
struct ump_control ctrl;
timestamps[i].time0 = bench_tsc();
msg = ump_impl_get_next(send, &ctrl);
msg->header.control = ctrl;
while (!ump_impl_recv(recv));
timestamps[i].time1 = bench_tsc();
}
/* Print results */
for (int i = MAX_COUNT / 10; i < MAX_COUNT; i++) {
if (timestamps[i].time1 > timestamps[i].time0) {
printf("page %d took %"PRIuCYCLES"\n", i,
timestamps[i].time1 - bench_tscoverhead() -
timestamps[i].time0);
}
}
}
void timing_sync_bench(void)
{
static cycles_t timestamp[ITERATIONS];
for(int i = 0; i < ITERATIONS; i++) {
cycles_t start = bench_tsc();
errval_t err = timing_sync_timer();
assert(err_is_ok(err));
cycles_t end = bench_tsc();
timestamp[i] = end - start;
}
for(int i = 0; i < ITERATIONS; i++) {
printf("duration %d: %" PRIuCYCLES "\n", i, timestamp[i]);
}
int nthreads = 0;
for(int i = 0; i <= MAX_COREID; i++) {
struct intermon_binding *b = NULL;
errval_t err = intermon_binding_get(i, &b);
if(err_is_ok(err) && b != NULL) {
nthreads++;
}
}
printf("number of threads: %d\n", nthreads);
printf("client done.\n");
}
static void run_experiment(void)
{
for (int i = 0; i < MAX_COUNT; i++) {
timestamps[i].time0 = bench_tsc();
clock_get_timestamp();
timestamps[i].time1 = bench_tsc();
}
}
static void prepare_bomp(void)
{
debug_printf("prepare_bomp\n");
cycles_t tsc_start = bench_tsc();
bomp_bomp_init(nthreads);
cycles_t tsc_end = bench_tsc();
timer_xompinit = bench_time_diff(tsc_start, tsc_end);
}
/**
* \brief Measure overhead of taking timestamp
*/
static void measure_tsc(void)
{
uint64_t begin;
uint64_t end;
begin = bench_tsc();
for(int i = 0; i < 1000; i++) {
end = bench_tsc();
}
tsc_overhead = (end - begin) / 1000;
}
int main(int argc, char *argv[])
{
bench_init();
int k = 300;
while(k--) {
uint64_t start = bench_tsc();
for (volatile int i = 0; i < ITERATIONS; i++);
uint64_t end = bench_tsc();
printf("%"PRIu64"\n", end - start);
}
return EXIT_SUCCESS;
}
int main(int argc, char *argv[])
{
volatile uint64_t workcnt = 0;
int nthreads;
debug_printf("bomptest started.\n");
bench_init();
#if CONFIG_TRACE
errval_t err = trace_control(TRACE_EVENT(TRACE_SUBSYS_ROUTE,
TRACE_EVENT_ROUTE_BENCH_START, 0),
TRACE_EVENT(TRACE_SUBSYS_ROUTE,
TRACE_EVENT_ROUTE_BENCH_STOP, 0), 0);
assert(err_is_ok(err));
#endif
if(argc == 2) {
nthreads = atoi(argv[1]);
backend_span_domain(nthreads, STACK_SIZE);
bomp_custom_init(NULL);
omp_set_num_threads(nthreads);
} else {
assert(!"Specify number of threads");
}
trace_event(TRACE_SUBSYS_ROUTE, TRACE_EVENT_ROUTE_BENCH_START, 0);
uint64_t start = bench_tsc();
#pragma omp parallel
while(rdtsc() < start + 805000000ULL) {
workcnt++;
}
uint64_t end = bench_tsc();
trace_event(TRACE_SUBSYS_ROUTE, TRACE_EVENT_ROUTE_BENCH_STOP, 0);
printf("done. time taken: %" PRIu64 " cycles.\n", end - start);
#if CONFIG_TRACE
char *buf = malloc(4096*4096);
trace_dump(buf, 4096*4096, NULL);
printf("%s\n", buf);
#endif
for(;;);
return 0;
}
static void single_run(int32_t chunksize, int32_t repetitions)
{
errval_t err;
vfs_handle_t handle;
// create file
err = vfs_create(FILENAME, &handle);
assert(err_is_ok(err));
// create chunk containing arbitraty data
uint8_t *chunk = malloc(chunksize);
assert(chunk != NULL);
// start time
printf("Start run with chunksize: %" PRId32 ", repetitions: %" PRId32
"\n", chunksize, repetitions);
cycles_t start_cycles = bench_tsc();
size_t written = 0;
for (int32_t i = 0; i < repetitions; i++) {
err = vfs_write(handle, chunk, chunksize, &written);
assert(err_is_ok(err));
assert(written == chunksize);
}
err = vfs_close(handle);
assert(err_is_ok(err));
// end time
cycles_t end_cycles = bench_tsc();
// evaluation
cycles_t cycles = end_cycles - start_cycles;
uint64_t ms = bench_tsc_to_ms(cycles);
double sec = (double) ms / 1000.0;
int64_t bytes_written = chunksize * repetitions;
double kibibytes_written = (double) bytes_written / 1024.0;
double mebibytes_written = (double) bytes_written / (1024.0 * 1024.0);
double kibps = kibibytes_written / sec;
printf("%" PRId64 " bytes (%.1f KiB, %.1f MiB) written in %" PRIuCYCLES
" cycles (%" PRIu64 " ms, %.1f s) -> %.1f KiB/s\n", bytes_written,
kibibytes_written, mebibytes_written, cycles, ms, sec, kibps);
// cleanup
free(chunk);
err = vfs_remove(FILENAME);
assert(err_is_ok(err));
}
void mp_barrier(cycles_t *measurement)
{
coreid_t tid = get_core_id();
#ifdef QRM_DBG_ENABLED
++_num_barrier;
uint32_t _num_barrier_recv = _num_barrier;
#endif
debug_printfff(DBG__REDUCE, "barrier enter #%d\n", _num_barrier);
// Recution
// --------------------------------------------------
#ifdef QRM_DBG_ENABLED
uint32_t _tmp =
#endif
mp_reduce(_num_barrier);
#ifdef QRM_DBG_ENABLED
// Sanity check
if (tid==get_sequentializer()) {
assert (_tmp == get_num_threads()*_num_barrier);
}
if (measurement)
*measurement = bench_tsc();
#endif
// Broadcast
// --------------------------------------------------
if (tid == get_sequentializer()) {
mp_send_ab(_num_barrier);
} else {
#ifdef QRM_DBG_ENABLED
_num_barrier_recv =
#endif
mp_receive_forward(0);
}
#ifdef QRM_DBG_ENABLED
if (_num_barrier_recv != _num_barrier) {
debug_printf("ASSERTION fail %d != %d\n", _num_barrier_recv, _num_barrier);
}
assert (_num_barrier_recv == _num_barrier);
// Add a shared memory barrier to absolutely make sure that
// everybody finished the barrier before leaving - this simplifies
// debugging, as the programm will get stuck if barriers are
// broken, rather than some threads (wrongly) continuing and
// causing problems somewhere else
#if 0 // Enable separately
debug_printfff(DBG_REDUCE, "finished barrier .. waiting for others\n");
shl_barrier_shm(get_num_threads());
#endif
#endif
debug_printfff(DBG__REDUCE, "barrier complete #%d\n", _num_barrier);
}
static cycles_t retype_caps(void)
{
errval_t err;
cycles_t time_taken = 0;
for (int i=0; i<CAPS_PER_CORE; i++) {
cycles_t start = bench_tsc();
err = cap_retype(retyped_caps[i], my_caps[i], ObjType_Frame, CHILD_BITS);
if (i >= 20 && i <= (CAPS_PER_CORE - 20)) { // avoid warmup / cooldown
time_taken += (bench_tsc() - start);
}
if (err_is_fail(err)) {
DEBUG_ERR(err, "xcorecap: Retype to frame failed\n");
}
}
return time_taken / (CAPS_PER_CORE - 40);
}
void record_packet_transmit_to_bf(void){
the_stats.last_packet_transmit_to_bf_ts = bench_tsc();
//printf("TOBF %"PRIu64" - %"PRIu64" = %"PRIu64"\n",the_stats.last_packet_transmit_to_bf_ts, the_stats.last_packet_receive_net_ts,the_stats.last_packet_transmit_to_bf_ts - the_stats.last_packet_receive_net_ts );
if(the_stats.last_packet_transmit_to_bf_ts > the_stats.last_packet_receive_net_ts){
push_net_to_bf_diff(the_stats.last_packet_transmit_to_bf_ts - the_stats.last_packet_receive_net_ts );
} else {
//printf("TOBF Skipped packet because of wrong ts order\n");
}
}
// called when a message is received
static inline void message_received(void)
{
errval_t err;
// save timestamp
timestamps[i].time1 = bench_tsc();
// trace receive event
err = trace_event(TRACE_SUBSYS_MULTIHOP,
TRACE_EVENT_MULTIHOP_MESSAGE_RECEIVE, 0);
if (err_is_fail(err)) {
USER_PANIC_ERR(err, "trace_event failed");
}
reply_received = true;
}
void record_packet_receive_from_net(void){
the_stats.last_packet_receive_net_ts = bench_tsc();
}
// continue experiment
static void experiment_cont(void* arg)
{
errval_t err;
static bool flag = false;
static int message_type = 0;
// Experiment finished (with this message type)
if (i == MAX_COUNT - 1) {
#if CONFIG_TRACE
#else
// print measured times
for (int j = MAX_COUNT / 10; j < MAX_COUNT; j++) {
printf(
"page %d took %"PRIuCYCLES"\n",
j,
timestamps[j].time1 - bench_tscoverhead()
- timestamps[j].time0);
}
#endif
// go to next message type
message_type++;
flag = false;
i = 0;
if (message_type > 13) {
// stop tracing
err = trace_event(TRACE_SUBSYS_MULTIHOP,
TRACE_EVENT_MULTIHOP_BENCH_STOP, 0);
if (err_is_fail(err)) {
USER_PANIC_ERR(err, "trace_event failed");
}
#if CONFIG_TRACE
// dump trace
char *buf = malloc(50*4096*4096);
size_t length = trace_dump(buf, 20*4096*4096, NULL);
printf("%s\n", buf);
printf("length of buffer %lu\n", length);
#endif
printf("client done!\n");
return;
}
}
if (!flag) { // Start experiment
#if CONFIG_TRACE
#else
printf("Running latency test for message %s...\n",
get_message_name(message_type));
#endif
flag = true;
timestamps[i].time0 = bench_tsc();
} else { // Continue experiment
i++;
timestamps[i].time0 = bench_tsc();
}
// trace send event
err = trace_event(TRACE_SUBSYS_MULTIHOP, TRACE_EVENT_MULTIHOP_MESSAGE_SEND,
message_type);
if (err_is_fail(err)) {
USER_PANIC_ERR(err, "trace_event failed");
}
// send next message
switch (message_type) {
case 0:
err = binding->tx_vtbl.fsb_empty_request(binding, NOP_CONT);
break;
case 1:
err = binding->tx_vtbl.fsb_payload32_1_request(binding, NOP_CONT, 1);
break;
case 2:
err = binding->tx_vtbl.fsb_payload32_2_request(binding, NOP_CONT, 1, 2);
break;
case 3:
err = binding->tx_vtbl.fsb_payload32_4_request(binding, NOP_CONT, 1, 2,
3, 4);
break;
case 4:
err = binding->tx_vtbl.fsb_payload32_8_request(binding, NOP_CONT, 1, 2,
3, 4, 5, 6, 7, 8);
break;
case 5:
err = binding->tx_vtbl.fsb_payload32_16_request(binding, NOP_CONT, 1, 2,
3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16);
break;
case 6:
err = binding->tx_vtbl.fsb_payload64_1_request(binding, NOP_CONT, 1);
break;
case 7:
err = binding->tx_vtbl.fsb_payload64_2_request(binding, NOP_CONT, 1, 2);
break;
case 8:
err = binding->tx_vtbl.fsb_payload64_4_request(binding, NOP_CONT, 1, 2,
3, 4);
break;
case 9:
err = binding->tx_vtbl.fsb_payload64_8_request(binding, NOP_CONT, 1, 2,
3, 4, 5, 6, 7, 8);
break;
case 10:
err = binding->tx_vtbl.fsb_payload64_16_request(binding, NOP_CONT, 1, 2,
3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16);
break;
case 11:
err = binding->tx_vtbl.fsb_buffer_request(binding, NOP_CONT, &buffer,
1);
break;
case 12:
err = binding->tx_vtbl.fsb_buffer_request(binding, NOP_CONT, buffer2,
100);
break;
case 13:
err = binding->tx_vtbl.fsb_buffer_request(binding, NOP_CONT, buffer3,
1000);
break;
default:
printf("unknown message type\n");
abort();
break;
}
// make sure send was successful
if (err_is_fail(err)) {
USER_PANIC_ERR(err, "while running experiment\n");
}
// receive reply (by dispatching events from the
// waitset we use for the benchmark)
while (reply_received == false) {
event_dispatch(&signal_waitset);
}
experiment();
}
static void gw_req_memory_call_rx(struct xomp_binding *b,
uint64_t addr,
uint8_t type)
{
XWI_DEBUG("gw_req_memory_call_rx: addr:%lx, tyep: %u\n", addr, type);
#if XOMP_BENCH_WORKER_EN
cycles_t mem_timer = bench_tsc();
#endif
struct txq_msg_st *msg_st = txq_msg_st_alloc(&txq);
assert(msg_st != NULL);
struct capref frame;
if (type == XOMP_FRAME_TYPE_REPL_RW) {
type = XOMP_FRAME_TYPE_SHARED_RW;
}
assert(!(worker_id & XOMP_WID_GATEWAY_FLAG));
msg_st->send = gw_req_memory_response_tx;
msg_st->cleanup = NULL;
XWR_DEBUG("Requesting frame from gateway: [%016lx]\n", usrdata);
msg_st->err = xomp_gateway_get_memory(addr, &frame);
if (err_is_fail(msg_st->err)) {
txq_send(msg_st);
return;
}
vregion_flags_t map_flags;
switch ((xomp_frame_type_t) type) {
case XOMP_FRAME_TYPE_MSG:
map_flags = VREGION_FLAGS_READ_WRITE;
break;
case XOMP_FRAME_TYPE_SHARED_RW:
case XOMP_FRAME_TYPE_REPL_RW:
map_flags = VREGION_FLAGS_READ_WRITE;
break;
case XOMP_FRAME_TYPE_SHARED_RO:
map_flags = VREGION_FLAGS_READ;
break;
default:
USER_PANIC("unknown type: %u", type)
break;
}
struct frame_identity id;
msg_st->err = invoke_frame_identify(frame, &id);
if (err_is_fail(msg_st->err)) {
txq_send(msg_st);
return;
}
if (addr) {
msg_st->err = vspace_map_one_frame_fixed_attr(addr, (1UL << id.bits),
frame, map_flags, NULL, NULL);
} else {
void *map_addr;
msg_st->err = vspace_map_one_frame_attr(&map_addr, (1UL << id.bits),
frame, map_flags, NULL, NULL);
}
#if XOMP_BENCH_WORKER_EN
mem_timer = bench_tsc() - mem_timer;
debug_printf("%lx mem request %016lx took %lu cycles, %lu ms\n", worker_id,
addr, mem_timer, bench_tsc_to_ms(mem_timer));
#endif
txq_send(msg_st);
}
errval_t spawn_xcore_monitor(coreid_t coreid, int hwid,
enum cpu_type cpu_type,
const char *cmdline,
struct frame_identity urpc_frame_id,
struct capref kcb)
{
uint64_t start = 0;
const char *monitorname = NULL, *cpuname = NULL;
genpaddr_t arch_page_size;
errval_t err;
err = get_architecture_config(cpu_type, &arch_page_size,
&monitorname, &cpuname);
assert(err_is_ok(err));
DEBUG("loading kernel: %s\n", cpuname);
DEBUG("loading 1st app: %s\n", monitorname);
// compute size of frame needed and allocate it
DEBUG("%s:%s:%d: urpc_frame_id.base=%"PRIxGENPADDR"\n",
__FILE__, __FUNCTION__, __LINE__, urpc_frame_id.base);
DEBUG("%s:%s:%d: urpc_frame_id.size=%d\n",
__FILE__, __FUNCTION__, __LINE__, urpc_frame_id.bits);
if (benchmark_flag) {
start = bench_tsc();
}
static size_t cpu_binary_size;
static lvaddr_t cpu_binary = 0;
static genpaddr_t cpu_binary_phys;
static const char* cached_cpuname = NULL;
if (cpu_binary == 0) {
cached_cpuname = cpuname;
// XXX: Caching these for now, until we have unmap
err = lookup_module(cpuname, &cpu_binary, &cpu_binary_phys,
&cpu_binary_size);
if (err_is_fail(err)) {
DEBUG_ERR(err, "Can not lookup module");
return err;
}
}
// Ensure caching actually works and we're
// always loading same binary. If this starts to fail, get rid of caching.
assert (strcmp(cached_cpuname, cpuname) == 0);
static size_t monitor_binary_size;
static lvaddr_t monitor_binary = 0;
static genpaddr_t monitor_binary_phys;
static const char* cached_monitorname = NULL;
if (monitor_binary == 0) {
cached_monitorname = monitorname;
// XXX: Caching these for now, until we have unmap
err = lookup_module(monitorname, &monitor_binary,
&monitor_binary_phys, &monitor_binary_size);
if (err_is_fail(err)) {
DEBUG_ERR(err, "Can not lookup module");
return err;
}
}
// Again, ensure caching actually worked (see above)
assert (strcmp(cached_monitorname, monitorname) == 0);
if (benchmark_flag) {
bench_data->load = bench_tsc() - start;
start = bench_tsc();
}
struct capref cpu_memory_cap;
struct frame_identity frameid;
size_t cpu_memory;
err = allocate_kernel_memory(cpu_binary, arch_page_size,
&cpu_memory_cap, &cpu_memory, &frameid);
if (err_is_fail(err)) {
DEBUG_ERR(err, "Can not allocate space for new app kernel.");
return err;
}
err = cap_mark_remote(cpu_memory_cap);
if (err_is_fail(err)) {
DEBUG_ERR(err, "Can not mark cap remote.");
return err;
}
void *cpu_buf_memory;
err = vspace_map_one_frame(&cpu_buf_memory, cpu_memory, cpu_memory_cap,
NULL, NULL);
if (err_is_fail(err)) {
return err_push(err, LIB_ERR_VSPACE_MAP);
}
if (benchmark_flag) {
bench_data->alloc_cpu = bench_tsc() - start;
start = bench_tsc();
}
/* Chunk of memory to load monitor on the app core */
struct capref spawn_memory_cap;
struct frame_identity spawn_memory_identity;
err = frame_alloc_identify(&spawn_memory_cap,
X86_CORE_DATA_PAGES * arch_page_size,
NULL, &spawn_memory_identity);
if (err_is_fail(err)) {
return err_push(err, LIB_ERR_FRAME_ALLOC);
}
err = cap_mark_remote(spawn_memory_cap);
if (err_is_fail(err)) {
DEBUG_ERR(err, "Can not mark cap remote.");
return err;
}
if (benchmark_flag) {
bench_data->alloc_mon = bench_tsc() - start;
start = bench_tsc();
}
/* Load cpu */
struct elf_allocate_state state;
state.vbase = (char *)cpu_buf_memory + arch_page_size;
assert(sizeof(struct x86_core_data) <= arch_page_size);
state.elfbase = elf_virtual_base(cpu_binary);
struct Elf64_Ehdr *cpu_head = (struct Elf64_Ehdr *)cpu_binary;
genvaddr_t cpu_entry;
err = elf_load(cpu_head->e_machine, elfload_allocate, &state,
cpu_binary, cpu_binary_size, &cpu_entry);
if (err_is_fail(err)) {
return err;
}
if (benchmark_flag) {
bench_data->elf_load = bench_tsc() - start;
start = bench_tsc();
}
err = relocate_cpu_binary(cpu_binary, cpu_head, state, frameid, arch_page_size);
if (err_is_fail(err)) {
DEBUG_ERR(err, "Can not relocate new kernel.");
return err;
}
if (benchmark_flag) {
bench_data->elf_reloc = bench_tsc() - start;
}
genvaddr_t cpu_reloc_entry = cpu_entry - state.elfbase
+ frameid.base + arch_page_size;
/* Compute entry point in the foreign address space */
forvaddr_t foreign_cpu_reloc_entry = (forvaddr_t)cpu_reloc_entry;
/* Setup the core_data struct in the new kernel */
struct x86_core_data *core_data = (struct x86_core_data *)cpu_buf_memory;
switch (cpu_head->e_machine) {
case EM_X86_64:
case EM_K1OM:
core_data->elf.size = sizeof(struct Elf64_Shdr);
core_data->elf.addr = cpu_binary_phys + (uintptr_t)cpu_head->e_shoff;
core_data->elf.num = cpu_head->e_shnum;
break;
case EM_386:
core_data->elf.size = sizeof(struct Elf32_Shdr);
struct Elf32_Ehdr *head32 = (struct Elf32_Ehdr *)cpu_binary;
core_data->elf.addr = cpu_binary_phys + (uintptr_t)head32->e_shoff;
core_data->elf.num = head32->e_shnum;
break;
default:
return SPAWN_ERR_UNKNOWN_TARGET_ARCH;
}
core_data->module_start = cpu_binary_phys;
core_data->module_end = cpu_binary_phys + cpu_binary_size;
core_data->urpc_frame_base = urpc_frame_id.base;
core_data->urpc_frame_bits = urpc_frame_id.bits;
core_data->monitor_binary = monitor_binary_phys;
core_data->monitor_binary_size = monitor_binary_size;
core_data->memory_base_start = spawn_memory_identity.base;
core_data->memory_bits = spawn_memory_identity.bits;
core_data->src_core_id = disp_get_core_id();
core_data->src_arch_id = my_arch_id;
core_data->dst_core_id = coreid;
struct frame_identity fid;
err = invoke_frame_identify(kcb, &fid);
if (err_is_fail(err)) {
USER_PANIC_ERR(err, "Invoke frame identity for KCB failed. "
"Did you add the syscall handler for that architecture?");
}
DEBUG("%s:%s:%d: fid.base is 0x%"PRIxGENPADDR"\n",
__FILE__, __FUNCTION__, __LINE__, fid.base);
core_data->kcb = (genpaddr_t) fid.base;
#ifdef CONFIG_FLOUNDER_BACKEND_UMP_IPI
core_data->chan_id = chanid;
#endif
if (cmdline != NULL) {
// copy as much of command line as will fit
snprintf(core_data->kernel_cmdline, sizeof(core_data->kernel_cmdline),
"%s %s", cpuname, cmdline);
// ensure termination
core_data->kernel_cmdline[sizeof(core_data->kernel_cmdline) - 1] = '\0';
DEBUG("%s:%s:%d: %s\n", __FILE__, __FUNCTION__, __LINE__, core_data->kernel_cmdline);
}
/* Invoke kernel capability to boot new core */
if (cpu_type == CPU_X86_64 || cpu_type == CPU_K1OM) {
start_aps_x86_64_start(hwid, foreign_cpu_reloc_entry);
}
#ifndef __k1om__
else if (cpu_type == CPU_X86_32) {
start_aps_x86_32_start(hwid, foreign_cpu_reloc_entry);
}
#endif
/* Clean up */
// XXX: Should not delete the remote caps?
err = cap_destroy(spawn_memory_cap);
if (err_is_fail(err)) {
USER_PANIC_ERR(err, "cap_destroy failed");
}
err = vspace_unmap(cpu_buf_memory);
if (err_is_fail(err)) {
USER_PANIC_ERR(err, "vspace unmap CPU driver memory failed");
}
err = cap_destroy(cpu_memory_cap);
if (err_is_fail(err)) {
USER_PANIC_ERR(err, "cap_destroy failed");
}
return SYS_ERR_OK;
}
static errval_t replicate_frame(lvaddr_t addr, struct capref *frame)
{
errval_t err;
#if XOMP_BENCH_WORKER_EN
cycles_t repl_timer = bench_tsc();
#endif
struct frame_identity id;
err = invoke_frame_identify(*frame, &id);
if (err_is_fail(err)) {
return err;
}
XWR_DEBUG("Replicating frame: [%016lx]\n", id.base);
struct capref replicate;
err = frame_alloc(&replicate, (1UL << id.bits), NULL);
if (err_is_fail(err)) {
return err;
}
XWR_DEBUG("registering memory with DMA service\n");
#if XOMP_BENCH_WORKER_EN
cycles_t register_timer = bench_tsc();
#endif
err = dma_register_memory((struct dma_device *) dma_dev, *frame);
if (err_is_fail(err)) {
return err;
}
err = dma_register_memory((struct dma_device *) dma_dev, replicate);
if (err_is_fail(err)) {
return err;
}
#if XOMP_BENCH_WORKER_EN
cycles_t register_timer_end = bench_tsc();
#endif
struct dma_req_setup setup = {
.done_cb = dma_replication_cb,
.cb_arg = NULL,
.args = {
.memcpy = {
.src = id.base,
.bytes = (1UL << id.bits)
}
}
};
err = invoke_frame_identify(replicate, &id);
if (err_is_fail(err)) {
return err;
}
setup.args.memcpy.dst = id.base;
dma_replication_done = 0x0;
XWR_DEBUG("DMA request for replication\n");
err = dma_request_memcpy((struct dma_device *)dma_dev, &setup, NULL);
if (err_is_fail(err)) {
return err;
}
while (!dma_replication_done) {
messages_wait_and_handle_next();
}
XWR_DEBUG("Replication done.\n");
*frame = replicate;
#if XOMP_BENCH_WORKER_EN
cycles_t timer_end = bench_tsc();
debug_printf("%lx replication took %lu cycles, %lu ms\n", worker_id,
timer_end - repl_timer, bench_tsc_to_ms(timer_end - repl_timer));
debug_printf("%lx register mem took %lu cycles, %lu ms\n", worker_id,
register_timer_end - register_timer, bench_tsc_to_ms(register_timer_end - register_timer));
#endif
return SYS_ERR_OK;
}
static void add_memory_call_rx(struct xomp_binding *b,
struct capref frame,
uint64_t addr,
uint8_t type)
{
XWI_DEBUG("add_memory_call_rx: addr:%lx, tyep: %u\n", addr, type);
struct txq_msg_st *msg_st = txq_msg_st_alloc(&txq);
assert(msg_st != NULL);
msg_st->send = add_memory_response_tx;
msg_st->cleanup = NULL;
uint32_t map_flags = 0x0;
switch ((xomp_frame_type_t) type) {
case XOMP_FRAME_TYPE_MSG:
map_flags = VREGION_FLAGS_READ_WRITE;
break;
case XOMP_FRAME_TYPE_SHARED_RW:
map_flags = VREGION_FLAGS_READ_WRITE;
break;
case XOMP_FRAME_TYPE_SHARED_RO:
map_flags = VREGION_FLAGS_READ;
break;
default:
USER_PANIC("unknown type: %u", type)
break;
}
struct frame_identity id;
msg_st->err = invoke_frame_identify(frame, &id);
if(err_is_fail(msg_st->err)) {
txq_send(msg_st);
return;
}
#if XOMP_WORKER_ENABLE_DMA
if (0) {
// todo: replicate frame on the same node if needed..
replicate_frame(addr, &frame);
}
#endif
#if XOMP_BENCH_WORKER_EN
cycles_t map_start = bench_tsc();
#endif
if (addr) {
msg_st->err = vspace_map_one_frame_fixed_attr(addr, (1UL << id.bits),
frame, map_flags, NULL, NULL);
} else {
void *map_addr;
msg_st->err = vspace_map_one_frame_attr(&map_addr, (1UL << id.bits),
frame, map_flags, NULL, NULL);
}
#if XOMP_BENCH_WORKER_EN
cycles_t timer_end = bench_tsc();
debug_printf("%lx mem map %016lx took %lu cycles, %lu ms\n", worker_id, addr,
timer_end - map_start, bench_tsc_to_ms(timer_end - map_start));
#endif
txq_send(msg_st);
}
static void do_work_rx(struct xomp_binding *b,
uint64_t fn,
uint64_t arg,
uint64_t id,
uint64_t flags)
{
errval_t err;
XWP_DEBUG("do_work_rx: fn:%lx, id:%lx\n", fn, id);
#if XOMP_BENCH_WORKER_EN
cycles_t work_timer = bench_tsc();
#endif
struct txq_msg_st *msg_st = txq_msg_st_alloc(&txq);
assert(msg_st != NULL);
msg_st->err = SYS_ERR_OK;
struct bomp_work *work = tls;
XWP_DEBUG("do_work_rx: threadid = %u, nthreads = %u\n", work->thread_id,
work->num_threads);
g_bomp_state->num_threads = work->num_threads;
struct xomp_msg_st *st = (struct xomp_msg_st *) msg_st;
st->args.done_notify.id = id;
if (arg) {
msg_st->send = done_with_arg_tx;
st->args.done_notify.arg = arg;
} else {
msg_st->send = done_notify_tx;
}
if (fn & XOMP_FN_INDEX_FLAG) {
uint32_t idx = fn & ~XOMP_FN_INDEX_FLAG;
char *fn_name;
err = spawn_symval_lookup_idx(idx, &fn_name, &fn);
if (err_is_fail(err)) {
msg_st->err = err;
txq_send(msg_st);
return;
}
XWP_DEBUG("do_work_rx: function index %u -> %s\n", idx, fn_name);
}
xomp_worker_fn_t fnct = (xomp_worker_fn_t) fn;
XWP_DEBUG("do_work_rx: calling fnct %p with argument %p\n", fnct, work->data);
for (uint32_t i = 0; i < work->num_vtreads; ++i) {
fnct(work->data);
work->thread_id++;
}
#if XOMP_BENCH_WORKER_EN
work_timer = bench_tsc() - work_timer;
debug_printf("%lx work took %lu cycles, %lu ms\n", worker_id, work_timer,
bench_tsc_to_ms(work_timer));
#endif
txq_send(msg_st);
}
static int prepare_xomp(int argc,
char *argv[])
{
errval_t err;
xomp_wloc_t location = XOMP_WORKER_LOC_MIXED;
for (int i = 3; i < argc; ++i) {
if (!strncmp(argv[i], "--location=", 11)) {
char *p = strchr(argv[i], '=');
p++;
if (!strcmp(p, "local")) {
location = XOMP_WORKER_LOC_LOCAL;
}
}
}
if (location == XOMP_WORKER_LOC_MIXED) {
debug_printf("waiting for xeon phi to be ready\n");
err = xeon_phi_domain_blocking_lookup("xeon_phi.0.ready", NULL);
EXPECT_SUCCESS(err, "nameservice_blocking_lookup");
err = xeon_phi_domain_blocking_lookup("xeon_phi.1.ready", NULL);
EXPECT_SUCCESS(err, "nameservice_blocking_lookup");
#if XOMP_BENCH_ENABLED
xomp_master_bench_enable(BENCH_RUN_COUNT, nthreads,
XOMP_MASTER_BENCH_MEM_ADD);
#endif
}
struct xomp_spawn local_info = {
.argc = argc,
.argv = argv,
#ifdef __k1om__
.path = "/k1om/sbin/benchmarks/bomp_mm",
#else
.path = "/x86_64/sbin/benchmarks/bomp_mm",
#endif
};
struct xomp_spawn remote_info = {
.argc = argc,
.argv = argv,
.path = "/k1om/sbin/benchmarks/bomp_mm",
};
struct xomp_args xomp_arg = {
.type = XOMP_ARG_TYPE_DISTINCT,
.core_stride = 0, // use default
.args = {
.distinct = {
.nthreads = nthreads,
.worker_loc = location,
.nphi = 2,
.local = local_info,
.remote = remote_info
}
}
};
cycles_t tsc_start = bench_tsc();
if (bomp_xomp_init(&xomp_arg)) {
debug_printf("bomp init failed!\n");
exit(1);
}
cycles_t tsc_end = bench_tsc();
timer_xompinit = bench_time_diff(tsc_start, tsc_end);
return (location == XOMP_WORKER_LOC_LOCAL);
}
int main(int argc,
char *argv[])
{
errval_t err;
xomp_wid_t wid;
bench_init();
err = xomp_worker_parse_cmdline(argc, argv, &wid);
if (err_is_ok(err)) {
struct xomp_args xw_arg = {
.type = XOMP_ARG_TYPE_WORKER,
.args = {
.worker = {
.id = wid
}
}
};
bomp_xomp_init(&xw_arg);
}
if (argc < 4) {
debug_printf("Usage: %s <size> <numthreats>\n", argv[0]);
exit(1);
}
nthreads = strtoul(argv[1], NULL, 10);
if (nthreads == 0) {
debug_printf("num threads must be >0\n");
exit(1);
}
DEBUG("\n");
DEBUG("======================================================\n");
debug_printf("Num Threads: %u\n", nthreads);
uint8_t is_shared = 0;
for (int i = 2; i < argc; ++i) {
if (!strcmp(argv[i], "bomp")) {
prepare_bomp();
is_shared = 1;
} else if (!strcmp(argv[i], "xomp")) {
is_shared = prepare_xomp(argc, argv);
} else {
debug_printf("ignoring argument {%s}\n", argv[i]);
}
}
debug_printf("-------------------------------------\n");
debug_printf("init time: %lu\n", timer_xompinit);
debug_printf("-------------------------------------\n");
#if XOMP_BENCH_ENABLED
xomp_master_bench_print_results();
#endif
while (1)
;
}
int start_aps_x86_32_start(uint8_t core_id, genvaddr_t entry)
{
DEBUG("%s:%d: start_aps_x86_32_start\n", __FILE__, __LINE__);
// Copy the startup code to the real-mode address
uint8_t *real_src = (uint8_t *) &x86_32_start_ap;
uint8_t *real_end = (uint8_t *) &x86_32_start_ap_end;
struct capref bootcap;
struct acpi_rpc_client* acl = get_acpi_rpc_client();
errval_t error_code;
errval_t err = acl->vtbl.mm_realloc_range_proxy(acl, 16, 0x0,
&bootcap, &error_code);
if (err_is_fail(err)) {
USER_PANIC_ERR(err, "mm_alloc_range_proxy failed.");
}
if (err_is_fail(error_code)) {
USER_PANIC_ERR(error_code, "mm_alloc_range_proxy return failed.");
}
void* real_base;
err = vspace_map_one_frame(&real_base, 1<<16, bootcap, NULL, NULL);
uint8_t* real_dest = (uint8_t*)real_base + X86_32_REAL_MODE_LINEAR_OFFSET;
memcpy(real_dest, real_src, real_end - real_src);
/* Pointer to the entry point called from init_ap.S */
volatile uint64_t *absolute_entry_ptr = (volatile uint64_t *)
((
(lpaddr_t) &x86_32_init_ap_absolute_entry -
(lpaddr_t) &x86_32_start_ap
)
+
real_dest);
//copy the address of the function start (in boot.S) to the long-mode
//assembler code to be able to perform an absolute jump
*absolute_entry_ptr = entry;
// pointer to the shared global variable amongst all kernels
volatile uint64_t *ap_global = (volatile uint64_t *)
((
(lpaddr_t) &x86_32_init_ap_global -
(lpaddr_t) &x86_32_start_ap
)
+ real_dest);
genpaddr_t global;
struct monitor_blocking_rpc_client *mc =
get_monitor_blocking_rpc_client();
err = mc->vtbl.get_global_paddr(mc, &global);
if (err_is_fail(err)) {
DEBUG_ERR(err, "invoke spawn core");
return err_push(err, MON_ERR_SPAWN_CORE);
}
*ap_global = (uint64_t)(genpaddr_t)global;
// pointer to the pseudo-lock used to detect boot up of new core
volatile uint32_t *ap_wait = (volatile uint32_t *)
((lpaddr_t) &x86_32_init_ap_wait -
((lpaddr_t) &x86_32_start_ap) +
real_dest);
// Pointer to the lock variable in the realmode code
volatile uint8_t *ap_lock = (volatile uint8_t *)
((lpaddr_t) &x86_32_init_ap_lock -
((lpaddr_t) &x86_32_start_ap) +
real_dest);
*ap_wait = AP_STARTING_UP;
end = bench_tsc();
err = invoke_send_init_ipi(ipi_cap, core_id);
if (err_is_fail(err)) {
DEBUG_ERR(err, "invoke send init ipi");
return err;
}
err = invoke_send_start_ipi(ipi_cap, core_id, entry);
if (err_is_fail(err)) {
DEBUG_ERR(err, "invoke sipi");
return err;
}
//give the new core a bit time to start-up and set the lock
for (uint64_t i = 0; i < STARTUP_TIMEOUT; i++) {
if (*ap_lock != 0) {
break;
}
}
// If the lock is set, the core has been started, otherwise assume, that
// a core with this APIC ID doesn't exist.
if (*ap_lock != 0) {
while (*ap_wait != AP_STARTED);
trace_event(TRACE_SUBSYS_KERNEL, TRACE_EVENT_KERNEL_CORE_START_REQUEST_ACK, core_id);
*ap_lock = 0;
return 0;
}
assert(!"badness");
return -1;
}