static int cos_net_tcp_recv(struct intern_connection *ic, void *data, int sz)
{
int xfer_amnt = 0;
assert(ic->conn_type == TCP);
/* If there is data available, get it */
if (ic->incoming_size > 0) {
struct packet_queue *pq;
struct tcp_pcb *tp;
char *data_start;
int data_left;
pq = ic->incoming;
assert(pq);
data_start = ((char*)pq->data) + ic->incoming_offset;
data_left = pq->len - ic->incoming_offset;
assert(data_left > 0 && (u32_t)data_left <= pq->len);
/* Consume all of first packet? */
if (data_left <= sz) {
ic->incoming = pq->next;
if (ic->incoming_last == pq) {
assert(NULL == ic->incoming);
ic->incoming_last = NULL;
}
memcpy(data, data_start, data_left);
xfer_amnt = data_left;
ic->incoming_offset = 0;
#ifdef TEST_TIMING
ic->ts_start = timing_record(APP_RECV, pq->ts_start);
#endif
free(pq);
}
/* Consume part of first packet */
else {
memcpy(data, data_start, sz);
xfer_amnt = sz;
ic->incoming_offset += sz;
assert(ic->incoming_offset >= 0 && (u32_t)ic->incoming_offset < pq->len);
}
ic->incoming_size -= xfer_amnt;
tp = ic->conn.tp;
tcp_recved(tp, xfer_amnt);
}
return xfer_amnt;
}
static err_t cos_net_stack_send(struct netif *ni, struct pbuf *p, struct ip_addr *ip)
{
int tot_len = 0, sz;
char *buff;
cbuf_t cb;
/* assuming the net lock is taken here */
assert(p && p->ref == 1);
assert(p->type == PBUF_RAM);
buff = cbuf_alloc(MTU, &cb);
assert(buff);
while (p) {
if (p->len + tot_len > MTU) BUG();
memcpy(buff + tot_len, p->payload, p->len);
tot_len += p->len;
#ifdef TCP_SEND_COPY
#ifdef TEST_TIMING
if ((p->type == PBUF_REF || p->type == PBUF_ROM)) {
struct packet_queue *pq;
pq = net_packet_pq(p->payload);
timing_record(SEND, pq->ts_start);
}
#endif
#endif
assert(p->type != PBUF_POOL);
assert(p->ref == 1);
p = p->next;
}
sz = parent_twrite(cos_spd_id(), ip_td, cb, tot_len);
if (sz <= 0) {
printc("<<transmit returns %d -> %d>>\n", sz, tot_len);
}
tcp_twrite_cnt++;
assert(sz > 0);
cbuf_free(buff);
/* cannot deallocate packets here as we might need to
* retransmit them. */
return ERR_OK;
}
void timing_basic_alltoall_nelements( int DIM1, int procs, int loop, char* testname, MPI_Comm local_communicator) {
float* send_array;
float* recv_array;
int myrank;
int base, typesize, bytes, i;
char method[50];
send_array = malloc( DIM1 * procs * sizeof(float));
recv_array = malloc( DIM1 * procs * sizeof(float));
MPI_Comm_rank( local_communicator, &myrank );
base = myrank * DIM1 + 1;
utilities_fill_unique_array_1D_float( &send_array[0], DIM1, base );
if ( myrank == 0 ) {
snprintf(method, 50, "reference");
MPI_Type_size( MPI_FLOAT, &typesize );
bytes = typesize * DIM1 * procs;
timing_init( testname, &method[0], bytes );
}
for( i=0 ; i<loop ; i++ ) {
MPI_Alltoall(&send_array[0], DIM1, MPI_FLOAT, &recv_array[0], DIM1, MPI_FLOAT, local_communicator );
MPI_Alltoall(&recv_array[0], DIM1, MPI_FLOAT, &send_array[0], DIM1, MPI_FLOAT, local_communicator );
if ( myrank == 0 ) {
timing_record(3);
}
}
if ( myrank == 0 ) {
timing_print( 1 );
}
free(send_array);
free(recv_array);
}
void timing_basic_ping_pong_nelements( int DIM1, int loop, char* testname, MPI_Comm local_communicator) {
float* array;
int myrank;
int base, typesize, bytes, i;
char method[50];
array = malloc( DIM1 * sizeof(float));
MPI_Comm_rank( local_communicator, &myrank );
base = myrank * DIM1 + 1;
utilities_fill_unique_array_1D_float( &array[0], DIM1, base );
if ( myrank == 0 ) {
snprintf(&method[0], 50, "reference");
MPI_Type_size( MPI_FLOAT, &typesize );
bytes = typesize * DIM1;
timing_init( testname, &method[0], bytes );
}
for( i=0 ; i<loop ; i++ ){
if ( myrank == 0 ) {
MPI_Send( &array[0], DIM1, MPI_FLOAT, 1, itag, local_communicator );
MPI_Recv( &array[0], DIM1, MPI_FLOAT, 1, itag, local_communicator, MPI_STATUS_IGNORE );
timing_record(3);
} else {
MPI_Recv( &array[0], DIM1, MPI_FLOAT, 0, itag, local_communicator, MPI_STATUS_IGNORE );
MPI_Send( &array[0], DIM1, MPI_FLOAT, 0, itag, local_communicator );
}
}
if ( myrank == 0 ) {
timing_print( 1 );
}
free(array);
}
static err_t cos_net_lwip_tcp_recv(void *arg, struct tcp_pcb *tp, struct pbuf *p, err_t err)
{
struct intern_connection *ic;
struct packet_queue *pq, *last;
void *headers;
struct pbuf *first;
ic = (struct intern_connection*)arg;
assert(NULL != ic);
assert(TCP == ic->conn_type);
if (NULL == p) {
assert(ic->conn.tp == tp);
/*
* This should call our registered error function
* above with ERR_ABRT, which will make progress
* towards closing the connection.
*
* Later, when the app calls some function in the API,
* TCP_CLOSED will be seen and the internal connection
* will be deallocated, and the application notified.
*/
tcp_abort(tp);
assert(ic->conn_type == TCP_CLOSED && NULL == ic->conn.tp);
/* tcp_close(tp); // Jiguo: aggressive close */
return ERR_CLSD;
}
first = p;
while (p) {
struct pbuf *q;
if (p->ref != 1) printc("pbuf with len %d, totlen %d and refcnt %d", p->len, p->tot_len, p->ref);
assert(p->len > 0);
assert(p->type == PBUF_ROM || p->type == PBUF_REF);
headers = cos_net_header_start(p, TCP);
assert (NULL != headers);
pq = net_packet_pq(headers);
pq->data = p->payload;
pq->len = p->len;
pq->next = NULL;
#ifdef TEST_TIMING
pq->ts_start = timing_record(RECV, pq->ts_start);
#endif
assert((NULL == ic->incoming) == (NULL == ic->incoming_last));
/* Is the queue empty? */
if (NULL == ic->incoming) {
assert(NULL == ic->incoming_last);
ic->incoming = ic->incoming_last = pq;
} else {
last = ic->incoming_last;
last->next = pq;
ic->incoming_last = pq;
}
ic->incoming_size += p->len;
//assert(1 == p->ref);
q = p->next;
p->payload = p->alloc_track = NULL;
assert(NULL != q || p->len == p->tot_len);
assert(p->ref == 1);
p = q;
}
/* Just make sure lwip is doing what we think its doing */
assert(first->ref == 1);
/* This should deallocate the entire chain */
pbuf_free(first);
/* printc("thd in %ld tcp_recv call trigger evt id %d\n", cos_get_thd_id(), ic->data); */
if (-1 != ic->data && evt_trigger(cos_spd_id(), ic->data)) BUG();
tcp_recv_cnt++;
/* /\* If the thread blocked waiting for a packet, wake it up *\/ */
/* if (RECVING == ic->thd_status) { */
/* ic->thd_status = ACTIVE; */
/* assert(ic->thd_status == ACTIVE); /\* Detect races *\/ */
/* if (sched_wakeup(cos_spd_id(), ic->tid)) BUG(); */
/* } */
return ERR_OK;
}
int net_send(spdid_t spdid, net_connection_t nc, void *data, int sz)
{
struct intern_connection *ic;
u16_t tid = cos_get_thd_id();
int ret = sz;
// if (!cos_argreg_buff_intern(data, sz)) return -EFAULT;
if (!net_conn_valid(nc)) return -EINVAL;
if (sz > MAX_SEND) return -EMSGSIZE;
// NET_LOCK_TAKE();
ic = net_conn_get_internal(nc);
if (NULL == ic) {
ret = -EINVAL;
goto err;
}
if (tid != ic->tid) {
ret = -EPERM;
goto err;
}
switch (ic->conn_type) {
case UDP:
{
struct udp_pcb *up;
struct pbuf *p;
/* There's no blocking in the UDP case, so this is simple */
up = ic->conn.up;
p = pbuf_alloc(PBUF_TRANSPORT, sz, PBUF_ROM);
if (NULL == p) {
ret = -ENOMEM;
goto err;
}
p->payload = data;
if (ERR_OK != udp_send(up, p)) {
pbuf_free(p);
/* IP/port must not be set */
ret = -ENOTCONN;
goto err;
}
pbuf_free(p);
break;
}
case TCP:
{
struct tcp_pcb *tp;
#define TCP_SEND_COPY
#ifdef TCP_SEND_COPY
void *d;
struct packet_queue *pq;
#endif
tp = ic->conn.tp;
if (tcp_sndbuf(tp) < sz) {
ret = 0;
break;
}
#ifdef TCP_SEND_COPY
pq = malloc(sizeof(struct packet_queue) + sz);
if (unlikely(NULL == pq)) {
ret = -ENOMEM;
goto err;
}
#ifdef TEST_TIMING
pq->ts_start = timing_record(APP_PROC, ic->ts_start);
#endif
pq->headers = NULL;
d = net_packet_data(pq);
memcpy(d, data, sz);
if (ERR_OK != (ret = tcp_write(tp, d, sz, 0))) {
#else
if (ERR_OK != (ret = tcp_write(tp, data, sz, TCP_WRITE_FLAG_COPY))) {
#endif
free(pq);
printc("tcp_write returned %d (sz %d, tcp_sndbuf %d, ERR_MEM: %d)",
ret, sz, tcp_sndbuf(tp), ERR_MEM);
BUG();
}
/* No implementation of nagle's algorithm yet. Send
* out the packet immediately if possible. */
if (ERR_OK != (ret = tcp_output(tp))) {
printc("tcp_output returned %d, ERR_MEM: %d", ret, ERR_MEM);
BUG();
}
ret = sz;
break;
}
case TCP_CLOSED:
ret = -EPIPE;
break;
default:
BUG();
}
err:
// NET_LOCK_RELEASE();
return ret;
}
/************************ LWIP integration: **************************/
struct ip_addr ip, mask, gw;
struct netif cos_if;
static void cos_net_interrupt(char *packet, int sz)
{
void *d;
int len;
struct pbuf *p;
struct ip_hdr *ih;
struct packet_queue *pq;
#ifdef TEST_TIMING
unsigned long long ts;
#endif
// printc(">>> %d\n", net_lock.lock_id);
NET_LOCK_TAKE();
// printc("<<< %d\n", net_lock.lock_id);
assert(packet);
ih = (struct ip_hdr*)packet;
if (unlikely(4 != IPH_V(ih))) goto done;
len = ntohs(IPH_LEN(ih));
if (unlikely(len != sz || len > MTU)) {
printc("len %d != %d or > %d", len, sz, MTU);
goto done;
}
p = pbuf_alloc(PBUF_IP, len, PBUF_ROM);
if (unlikely(!p)) {
prints("OOM in interrupt: allocation of pbuf failed.\n");
goto done;
}
/* For now, we're going to do an additional copy. Currently,
* packets should be small, so this shouldn't hurt that badly.
* This is done because 1) we are freeing the packet
* elsewhere, 2) we want to malloc some (small) packets to
* save space and free up the ring buffers, 3) it is difficult
* to know in (1) which deallocation method (free or return to
* ring buff) to use */
pq = malloc(len + sizeof(struct packet_queue));
if (unlikely(NULL == pq)) {
printc("OOM in interrupt: allocation of packet data (%d bytes) failed.\n", len);
pbuf_free(p);
goto done;
}
pq->headers = d = net_packet_data(pq);
#ifdef TEST_TIMING
#ifdef TCP_SEND_COPY
ts = pq->ts_start = timing_timestamp();
#endif
#endif
memcpy(d, packet, len);
p->payload = p->alloc_track = d;
/* hand off packet ownership here... */
if (ERR_OK != cos_if.input(p, &cos_if)) {
prints("net: failure in IP input.");
pbuf_free(p);
goto done;
}
#ifdef TEST_TIMING
timing_record(UPCALL_PROC, ts);
#endif
done:
NET_LOCK_RELEASE();
return;
}
int
main (int argc, char **argv)
{
char *ch;
if (argc != 2) {
fputs("Invalid number of arguments\n", stderr);
fputs("usage: hspwrap EXEFILE\n", stderr);
exit(EXIT_FAILURE);
}
ch = getenv("HSP_BCAST_CHUNK_SIZE");
if (ch) {
sscanf(ch, "%zu", &bcast_chunk_size);
} else {
bcast_chunk_size = 4L << 20;
}
ch = getenv("HSP_INPUT_FORMAT");
if (!ch || ch[0] == '\0' || ch[0] == 'l') {
info("Input format: Lines\n");
input_fmt = 'l';
} else if (ch[0] == 'f') {
info("Input format: FASTA\n");
input_fmt = 'f';
} else {
fputs("Invalid input format specified\n", stderr);
exit(EXIT_FAILURE);
}
// Pre-fork process pool (even on master)
#ifndef TIMING_MODE
sleep(1);
pool_ctl = process_pool_fork();
trace("Process pool created.\n");
sleep(1);
#endif
// Initialize MPI
int rank, ranks;
if (MPI_Init(NULL, NULL) != MPI_SUCCESS) {
fprintf(stderr, "Error initialize MPI.\n");
return EXIT_FAILURE;
}
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &ranks);
trace("MPI Initialized.\n");
// Initialize our state
timing_init(&timing);
if (rank) {
slave_init(rank, ranks-1, NUM_PROCS);
} else {
print_banner_slant(stderr);
master_init();
}
// Broadcast binary files first
if (rank) {
slave_broadcast_work_file("exefile");
} else {
master_broadcast_file(getenv("HSP_EXEFILE"));
}
// Distribute DB files
MPI_Barrier(MPI_COMM_WORLD);
timing_record(&timing.db_start);
char *dbdir = getenv("HSP_DBDIR");
char *dbfiles = strdup(getenv("HSP_DBFILES"));
char *fn, path[PATH_MAX];
for (fn = strtok(dbfiles, ":"); fn; fn = strtok(NULL, ":")) {
snprintf(path, sizeof(path), "%s/%s", dbdir, fn);
if (rank) {
timing.db_kbytes += slave_broadcast_shared_file(path)/1024;
} else {
timing.db_kbytes += master_broadcast_file(path)/1024;
}
}
free(dbfiles);
MPI_Barrier(MPI_COMM_WORLD);
timing_record(&timing.db_end);
#ifdef TIMING_MODE
if (!rank) {
timing_print(&timing);
}
MPI_Finalize();
return 0;
#endif
// FIXME: The order of things is generally wrong. Should be:
// Fork Forker, MPI_Init, PS Ctl, EXE/DB distro, forking, main loop
#if 0
// Now print some stats
if (rank) {
MPI_Barrier(MPI_COMM_WORLD);
printf("Rank %d Processes: %d", rank, ps_ctl->nprocesses);
printf(" Process ID: %d", getpid());
printf(" Files: %d (", ps_ctl->ft.nfiles);
for (i=0; i<ps_ctl->ft.nfiles; ++i) {
printf("%s, ", ps_ctl->ft.file[i].name);
}
puts(")");
} else {
printf("Ranks: %d\n\n", ranks);
MPI_Barrier(MPI_COMM_WORLD);
}
#endif
if (rank) {
slave_main(argv[1]);
} else {
master_main(ranks-1);
timing_print(&timing);
}
return 0;
}