gaspi_return_t
pgaspi_set_socket_affinity (const gaspi_uchar socket)
{
cpu_set_t sock_mask;
if (socket >= 4)
{
gaspi_print_error("GPI-2 only allows up to a maximum of 4 NUMA sockets");
return GASPI_ERROR;
}
if (gaspi_get_affinity_mask (socket, &sock_mask) < 0)
{
gaspi_print_error ("Failed to get affinity mask");
return GASPI_ERROR;
}
else
{
if (sched_setaffinity (0, sizeof (cpu_set_t), &sock_mask) != 0)
{
gaspi_print_error ("Failed to set affinity");
return GASPI_ERROR;
}
}
return GASPI_SUCCESS;
}
static inline int
_gaspi_sn_segment_register_command(const gaspi_rank_t rank, void * arg)
{
gaspi_segment_id_t segment_id = * (gaspi_segment_id_t *) arg;
gaspi_cd_header cdh;
memset(&cdh, 0, sizeof(gaspi_cd_header));
cdh.op_len = 0; /* in-place */
cdh.op = GASPI_SN_SEG_REGISTER;
cdh.rank = glb_gaspi_ctx.rank;
cdh.seg_id = segment_id;
cdh.rkey = glb_gaspi_ctx.rrmd[segment_id][glb_gaspi_ctx.rank].rkey;
cdh.addr = glb_gaspi_ctx.rrmd[segment_id][glb_gaspi_ctx.rank].addr;
cdh.size = glb_gaspi_ctx.rrmd[segment_id][glb_gaspi_ctx.rank].size;
#ifdef GPI2_CUDA
cdh.host_rkey = glb_gaspi_ctx.rrmd[segment_id][glb_gaspi_ctx.rank].host_rkey;
cdh.host_addr = glb_gaspi_ctx.rrmd[segment_id][glb_gaspi_ctx.rank].host_addr;
#endif
ssize_t ret = gaspi_sn_writen(glb_gaspi_ctx.sockfd[rank], &cdh, sizeof(gaspi_cd_header));
if(ret != sizeof(gaspi_cd_header))
{
gaspi_print_error("Failed to write to rank %u (args: %d %p %lu)",
rank,
glb_gaspi_ctx.sockfd[rank],
&cdh,
sizeof(gaspi_cd_header));
return -1;
}
int result = 1;
ssize_t rret = gaspi_sn_readn(glb_gaspi_ctx.sockfd[rank], &result, sizeof(int));
if( rret != sizeof(int) )
{
gaspi_print_error("Failed to read from rank %u (args: %d %p %lu)",
rank,
glb_gaspi_ctx.sockfd[rank],
&rret,
sizeof(int));
return -1;
}
/* Registration failed on the remote side */
if( result != 0)
return -1;
return 0;
}
/* TODO: deal with timeout */
int
gaspi_sn_broadcast_topology(gaspi_context *ctx, const gaspi_timeout_t timeout_ms)
{
int mask = 0x1;
int relative_rank;
int dst, src;
const int root = 0;
relative_rank = (ctx->rank >= root) ? ctx->rank - root : ctx->rank - root + ctx->tnc;
while(mask <= ctx->tnc)
{
if(relative_rank & mask)
{
src = ctx->rank - mask;
if(src < 0)
src += ctx->tnc;
if(gaspi_sn_recv_topology(ctx) != 0)
{
gaspi_print_error("Failed to receive topology.");
return -1;
}
break;
}
mask <<=1;
}
mask >>=1;
while (mask > 0)
{
if(relative_rank + mask < ctx->tnc)
{
dst = ctx->rank + mask;
if(dst >= ctx->tnc)
dst -= ctx->tnc;
if(gaspi_sn_send_topology(ctx, dst, timeout_ms) != 0)
{
gaspi_print_error("Failed to send topology to %d", dst);
return -1;
}
}
mask >>=1;
}
return 0;
}
static int
gaspi_sn_connect2port_intern(const char *hn, const unsigned short port)
{
int ret;
int sockfd = -1;
struct sockaddr_in Host;
struct hostent *serverData;
sockfd = socket ( AF_INET, SOCK_STREAM, 0 );
if( -1 == sockfd )
{
/* at least deal with open files limit */
int errsv = errno;
if(errsv == EMFILE)
{
if( 0 == _gaspi_check_ofile_limit() )
{
sockfd = socket(AF_INET,SOCK_STREAM,0);
if(sockfd == -1)
return -1;
}
else
return -2;
}
else
return -1;
}
Host.sin_family = AF_INET;
Host.sin_port = htons(port);
if((serverData = gethostbyname(hn)) == NULL)
{
close(sockfd);
return -1;
}
memcpy(&Host.sin_addr, serverData->h_addr, serverData->h_length);
/* TODO: we need to be able to distinguish between an initialization
connection attemp and a connection attempt during run-time where
the remote node is gone (FT) */
ret = connect( sockfd, (struct sockaddr *) &Host, sizeof(Host) );
if( 0 != ret )
{
close( sockfd );
return -1;
}
if( 0 != gaspi_sn_set_default_opts(sockfd) )
{
gaspi_print_error("Failed to set options on socket");
close(sockfd);
return -1;
}
return sockfd;
}
int
gaspi_sn_set_default_opts(int sockfd)
{
int opt = 1;
if(setsockopt(sockfd, SOL_SOCKET, SO_REUSEADDR, &opt, sizeof(opt)) < 0)
{
gaspi_print_error("Failed to set options on socket");
return -1;
}
if(setsockopt(sockfd, IPPROTO_TCP, TCP_NODELAY, &opt, sizeof(opt)) < 0)
{
gaspi_print_error("Failed to set options on socket");
return -1;
}
return 0;
}
gaspi_return_t
pgaspi_dev_wait (const gaspi_queue_id_t queue,
int * counter,
const gaspi_timeout_t timeout_ms)
{
int ne = 0, i;
struct ibv_wc wc;
const int nr = *counter;
const gaspi_cycles_t s0 = gaspi_get_cycles ();
for (i = 0; i < nr; i++)
{
do
{
ne = ibv_poll_cq (glb_gaspi_ctx_ib.scqC[queue], 1, &wc);
*counter -= ne;
if (ne == 0)
{
const gaspi_cycles_t s1 = gaspi_get_cycles ();
const gaspi_cycles_t tdelta = s1 - s0;
const float ms = (float) tdelta * glb_gaspi_ctx.cycles_to_msecs;
if (ms > timeout_ms)
{
return GASPI_TIMEOUT;
}
}
}
while (ne == 0);
if ((ne < 0) || (wc.status != IBV_WC_SUCCESS))
{
gaspi_print_error("Failed request to %lu. Queue %d might be broken %s",
wc.wr_id, queue, ibv_wc_status_str(wc.status) );
glb_gaspi_ctx.qp_state_vec[queue][wc.wr_id] = GASPI_STATE_CORRUPT;
return GASPI_ERROR;
}
}
#ifdef GPI2_CUDA
int j,k;
for(k = 0;k < glb_gaspi_ctx.gpu_count; k++)
{
for(j = 0; j < GASPI_CUDA_EVENTS; j++)
gpus[k].events[queue][j].ib_use = 0;
}
#endif
return GASPI_SUCCESS;
}
static inline int
_gaspi_sn_group_connect(const gaspi_rank_t rank, void *arg)
{
int i = (int) rank;
gaspi_group_t group = *(gaspi_group_t *) arg;
gaspi_group_ctx *group_to_commit = &(glb_gaspi_group_ctx[group]);
gaspi_cd_header cdh;
memset(&cdh, 0, sizeof(gaspi_cd_header));
cdh.op_len = sizeof(gaspi_rc_mseg);
cdh.op = GASPI_SN_GRP_CONNECT;
cdh.rank = glb_gaspi_ctx.rank;
cdh.ret = group;
ssize_t ret = gaspi_sn_writen(glb_gaspi_ctx.sockfd[i], &cdh, sizeof(gaspi_cd_header));
if( ret != sizeof(gaspi_cd_header) )
{
gaspi_print_error("Failed to write to %u (%ld %d %p %lu)",
i,
ret,
glb_gaspi_ctx.sockfd[i],
&cdh,
sizeof(gaspi_cd_header));
return -1;
}
ssize_t rret = gaspi_sn_readn(glb_gaspi_ctx.sockfd[i], &group_to_commit->rrcd[i], sizeof(gaspi_rc_mseg));
if( rret != sizeof(gaspi_rc_mseg) )
{
gaspi_print_error("Failed to read from %d (%ld %d %p %lu)",
i,
ret,
glb_gaspi_ctx.sockfd[i],
&group_to_commit->rrcd[i],
sizeof(gaspi_rc_mseg));
return -1;
}
return 0;
}
static inline int
_gaspi_sn_connect_command(const gaspi_rank_t rank)
{
const int i = (int) rank;
gaspi_cd_header cdh;
memset(&cdh, 0, sizeof(gaspi_cd_header));
const size_t rc_size = pgaspi_dev_get_sizeof_rc();
cdh.op_len = (int) rc_size;
cdh.op = GASPI_SN_CONNECT;
cdh.rank = glb_gaspi_ctx.rank;
/* if we have something to exchange */
if(rc_size > 0 )
{
ssize_t ret = gaspi_sn_writen(glb_gaspi_ctx.sockfd[i], &cdh, sizeof(gaspi_cd_header));
if(ret != sizeof(gaspi_cd_header))
{
gaspi_print_error("Failed to write to %d", i);
return -1;
}
ret = gaspi_sn_writen(glb_gaspi_ctx.sockfd[i], pgaspi_dev_get_lrcd(i), rc_size);
if(ret != (ssize_t) rc_size)
{
gaspi_print_error("Failed to write to %d", i);
return -1;
}
ssize_t rret = gaspi_sn_readn(glb_gaspi_ctx.sockfd[i], pgaspi_dev_get_rrcd(i), rc_size);
if( rret != (ssize_t) rc_size )
{
gaspi_print_error("Failed to read from %d", i);
return -1;
}
}
return 0;
}
gaspi_return_t
gaspi_number_of_GPUs(gaspi_gpu_num *gpus)
{
gaspi_verify_init("gaspi_number_of_GPUs");
gaspi_verify_null_ptr(gpus);
if( 0 == glb_gaspi_ctx.use_gpus )
{
gaspi_print_error("GPUs are not initialized.");
return GASPI_ERROR;
}
*gpus = glb_gaspi_ctx.gpu_count;
return GASPI_SUCCESS;
}
gaspi_return_t
gaspi_gpu_number(gaspi_number_t* num_gpus)
{
gaspi_verify_init("gaspi_gpu_number");
gaspi_verify_null_ptr(num_gpus);
gaspi_context_t const * const gctx = &glb_gaspi_ctx;
if( 0 == gctx->use_gpus )
{
gaspi_print_error("GPUs are not initialized.");
return GASPI_ERROR;
}
*num_gpus = gctx->gpu_count;
return GASPI_SUCCESS;
}
static int
_gaspi_find_GPU_numa_node(int cudevice)
{
CUresult cres;
int domain, bus, dev;
char path[128];
FILE *sysfile = NULL;
domain = 0;
#ifdef CU_DEVICE_ATTRIBUTE_PCI_DOMAIN_ID
cres = cuDeviceGetAttribute(&domain, CU_DEVICE_ATTRIBUTE_PCI_DOMAIN_ID, cudevice);
if( CUDA_SUCCESS != cres )
{
errno = ENOSYS;
return -1;
}
#endif
cres = cuDeviceGetAttribute(&bus, CU_DEVICE_ATTRIBUTE_PCI_BUS_ID, cudevice);
if( CUDA_SUCCESS != cres )
{
return -1;
}
cres = cuDeviceGetAttribute(&dev, CU_DEVICE_ATTRIBUTE_PCI_DEVICE_ID, cudevice);
if( CUDA_SUCCESS != cres )
{
return -1;
}
sprintf(path, "/sys/bus/pci/devices/%04x:%02x:%02x.0/numa_node", domain, bus, dev);
sysfile = fopen(path, "r");
if( !sysfile )
{
gaspi_print_error("Failed to open %s.", path);
return -1;
}
int numa_node = -1;
fscanf (sysfile, "%1d", &numa_node);
fclose(sysfile);
return numa_node;
}
/* TODO: Not clear to me why we need this function */
gaspi_return_t
gaspi_GPU_ids(gaspi_gpu_t *gpu_ids)
{
gaspi_verify_init("gaspi_GPU_ids");
gaspi_verify_null_ptr(gpu_ids);
if( 0 == glb_gaspi_ctx.use_gpus )
{
gaspi_print_error("GPUs are not initialized.");
return GASPI_ERROR;
}
int i;
for (i = 0; i < glb_gaspi_ctx.gpu_count; i++)
gpu_ids[i] = gpus[i].device_id;
return GASPI_SUCCESS;
}
gaspi_return_t
pgaspi_numa_socket(gaspi_uchar * const socket)
{
char * numaPtr = getenv ("GASPI_SET_NUMA_SOCKET");
if(numaPtr)
{
if(atoi(numaPtr) == 1)
{
*socket = (gaspi_uchar) glb_gaspi_ctx.localSocket;
return GASPI_SUCCESS;
}
}
gaspi_print_error("NUMA was not enabled (-N option of gaspi_run)");
return GASPI_ERR_ENV;
}
gaspi_return_t
pgaspi_proc_term (const gaspi_timeout_t timeout)
{
int i;
gaspi_verify_init("gaspi_proc_term");
if(lock_gaspi_tout (&glb_gaspi_ctx_lock, timeout))
return GASPI_TIMEOUT;
pthread_kill(glb_gaspi_ctx.snt, SIGSTKFLT);
if(glb_gaspi_ctx.sockfd != NULL)
{
for(i = 0;i < glb_gaspi_ctx.tnc; i++)
{
shutdown(glb_gaspi_ctx.sockfd[i],2);
if(glb_gaspi_ctx.sockfd[i] > 0)
close(glb_gaspi_ctx.sockfd[i]);
}
free(glb_gaspi_ctx.sockfd);
}
#ifdef GPI2_WITH_MPI
if(glb_gaspi_ctx.rank == 0)
{
if(remove(glb_gaspi_ctx.mfile) < 0)
{
gaspi_print_error("Failed to remove tmp file (%s)", glb_gaspi_ctx.mfile);
}
}
#endif
if(pgaspi_cleanup_core() != GASPI_SUCCESS)
goto errL;
unlock_gaspi (&glb_gaspi_ctx_lock);
return GASPI_SUCCESS;
errL:
unlock_gaspi (&glb_gaspi_ctx_lock);
return GASPI_ERROR;
}
gaspi_return_t
pgaspi_cpu_frequency (gaspi_float * const cpu_mhz)
{
gaspi_verify_null_ptr(cpu_mhz);
if (!glb_gaspi_init)
{
*cpu_mhz = gaspi_get_cpufreq ();
}
else
{
*cpu_mhz = glb_gaspi_ctx.mhz;
}
if (*cpu_mhz == 0.0f)
{
gaspi_print_error ("Failed to get CPU frequency");
return GASPI_ERROR;
}
return GASPI_SUCCESS;
}
static int
_gaspi_find_GPU_ib_numa_node()
{
char path[128];
int numa_node;
FILE *sysfile = NULL;
sprintf(path, "/sys/class/infiniband/%s/device/numa_node",
ibv_get_device_name(glb_gaspi_ctx_ib.ib_dev));
sysfile = fopen(path, "r");
if (!sysfile)
{
gaspi_print_error("Failed to open %s.", path);
return -1;
}
fscanf (sysfile, "%1d", &numa_node);
fclose(sysfile);
return numa_node;
}
gaspi_return_t
pgaspi_proc_kill (const gaspi_rank_t rank,const gaspi_timeout_t timeout_ms)
{
gaspi_return_t eret = GASPI_ERROR;
gaspi_verify_init("gaspi_proc_kill");
gaspi_verify_rank(rank);
if( rank == glb_gaspi_ctx.rank )
{
gaspi_print_error("Invalid rank to kill");
return GASPI_ERR_INV_RANK;
}
if(lock_gaspi_tout(&glb_gaspi_ctx_lock, timeout_ms))
return GASPI_TIMEOUT;
eret = gaspi_sn_command(GASPI_SN_PROC_KILL, rank, timeout_ms, NULL);
unlock_gaspi(&glb_gaspi_ctx_lock);
return eret;
}
/* TODO: Do we really need this function or at least make it part of
the GPU interface and allow clients to use it? */
gaspi_return_t
gaspi_gpu_ids(gaspi_gpu_id_t* gpu_ids)
{
gaspi_verify_init("gaspi_gpu_ids");
gaspi_verify_null_ptr(gpu_ids);
gaspi_context_t const * const gctx = &glb_gaspi_ctx;
if( 0 == gctx->use_gpus )
{
gaspi_print_error("GPUs are not found/initialized.");
return GASPI_ERROR;
}
int i;
for(i = 0; i < gctx->gpu_count; i++)
{
gpu_ids[i] = gpus[i].device_id;
}
return GASPI_SUCCESS;
}
static inline int
_gaspi_sn_single_command(const gaspi_rank_t rank, const enum gaspi_sn_ops op)
{
gaspi_cd_header cdh;
memset(&cdh, 0, sizeof(gaspi_cd_header));
cdh.op_len = 1;
cdh.op = op;
cdh.rank = rank;
cdh.tnc = glb_gaspi_ctx.tnc;
ssize_t ret = gaspi_sn_writen(glb_gaspi_ctx.sockfd[rank], &cdh, sizeof(gaspi_cd_header));
if( ret != sizeof(gaspi_cd_header) )
{
gaspi_print_error("Failed to write to %u (%d %p %lu)",
rank,
glb_gaspi_ctx.sockfd[rank], &cdh, sizeof(gaspi_cd_header));
return -1;
}
return 0;
}
gaspi_return_t
gaspi_sn_connect_to_rank(const gaspi_rank_t rank, gaspi_timeout_t timeout_ms)
{
struct timeb t0, t1;
ftime(&t0);
#ifdef DEBUG
if( strcmp(gaspi_get_hn(rank), "") == 0 )
{
gaspi_print_error("Failed to obtain hostname for rank %u", rank);
return GASPI_ERROR;
}
#endif
/* TODO: introduce backoff delay? */
while(glb_gaspi_ctx.sockfd[rank] == -1)
{
glb_gaspi_ctx.sockfd[rank] =
gaspi_sn_connect2port(gaspi_get_hn(rank),
glb_gaspi_cfg.sn_port + glb_gaspi_ctx.poff[rank],
timeout_ms);
if( -2 == glb_gaspi_ctx.sockfd[rank] )
return GASPI_ERR_EMFILE;
if( -1 == glb_gaspi_ctx.sockfd[rank] )
{
ftime(&t1);
const unsigned int delta_ms = (t1.time - t0.time) * 1000 + (t1.millitm - t0.millitm);
if(delta_ms > timeout_ms)
return GASPI_TIMEOUT;
}
}
return GASPI_SUCCESS;
}
gaspi_return_t
pgaspi_proc_init (const gaspi_timeout_t timeout_ms)
{
gaspi_return_t eret = GASPI_ERROR;
int i;
const int num_queues = (int) glb_gaspi_cfg.queue_num;
if(lock_gaspi_tout (&glb_gaspi_ctx_lock, timeout_ms))
return GASPI_TIMEOUT;
if(glb_gaspi_sn_init == 0)
{
glb_gaspi_ctx.lockPS.lock = 0;
glb_gaspi_ctx.lockPR.lock = 0;
for (i = 0; i < num_queues; i++)
glb_gaspi_ctx.lockC[i].lock = 0;
memset (&glb_gaspi_ctx, 0, sizeof (gaspi_context));
struct utsname mbuf;
if (uname (&mbuf) == 0)
{
snprintf (glb_gaspi_ctx.mtyp, 64, "%s", mbuf.machine);
}
//timing
glb_gaspi_ctx.mhz = gaspi_get_cpufreq ();
if (glb_gaspi_ctx.mhz == 0.0f)
{
gaspi_print_error ("Failed to get CPU frequency");
goto errL;
}
glb_gaspi_ctx.cycles_to_msecs = 1.0f / (glb_gaspi_ctx.mhz * 1000.0f);
//handle environment
if(gaspi_handle_env(&glb_gaspi_ctx))
{
gaspi_print_error("Failed to handle environment");
eret = GASPI_ERR_ENV;
goto errL;
}
//start sn_backend
if(pthread_create(&glb_gaspi_ctx.snt, NULL, gaspi_sn_backend, NULL) != 0)
{
gaspi_print_error("Failed to create SN thread");
goto errL;
}
glb_gaspi_sn_init = 1;
}//glb_gaspi_sn_init
if(glb_gaspi_ctx.procType == MASTER_PROC)
{
if(glb_gaspi_dev_init == 0)
{
if(access (glb_gaspi_ctx.mfile, R_OK) == -1)
{
gaspi_print_error ("Incorrect permissions of machinefile");
eret = GASPI_ERR_ENV;
goto errL;
}
//read hostnames
char *line = NULL;
size_t len = 0;
int read;
FILE *fp = fopen (glb_gaspi_ctx.mfile, "r");
if (fp == NULL)
{
gaspi_print_error("Failed to open machinefile");
eret = GASPI_ERR_ENV;
goto errL;
}
glb_gaspi_ctx.tnc = 0;
while ((read = getline (&line, &len, fp)) != -1)
{
//we assume a single hostname per line
if ((read < 2) || (read > 64))
continue;
glb_gaspi_ctx.tnc++;
if (glb_gaspi_ctx.tnc >= GASPI_MAX_NODES)
break;
}
rewind (fp);
free (glb_gaspi_ctx.hn_poff);
glb_gaspi_ctx.hn_poff = (char *) calloc (glb_gaspi_ctx.tnc, 65);
if(glb_gaspi_ctx.hn_poff == NULL)
{
gaspi_print_error("Debug: Failed to allocate memory");
goto errL;
}
glb_gaspi_ctx.poff = glb_gaspi_ctx.hn_poff + glb_gaspi_ctx.tnc * 64;
int id = 0;
while((read = getline (&line, &len, fp)) != -1)
{
//we assume a single hostname per line
if((read < 2) || (read >= 64)) continue;
int inList = 0;
for(i = 0; i < id; i++)
{
//already in list ?
//TODO: 64? 63? Magic numbers -> just get cacheline from system or define as such
const int hnlen = MAX (strlen (glb_gaspi_ctx.hn_poff + i * 64), MIN (strlen (line) - 1, 63));
if(strncmp (glb_gaspi_ctx.hn_poff + i * 64, line, hnlen) == 0)
{
inList++;
}
}
glb_gaspi_ctx.poff[id] = inList;
strncpy (glb_gaspi_ctx.hn_poff + id * 64, line, MIN (read - 1, 63));
id++;
if(id >= GASPI_MAX_NODES)
break;
}
fclose (fp);
free (line);
//master
glb_gaspi_ctx.rank = 0;
free(glb_gaspi_ctx.sockfd);
glb_gaspi_ctx.sockfd = (int *) malloc (glb_gaspi_ctx.tnc * sizeof (int));
if(glb_gaspi_ctx.sockfd == NULL)
{
gaspi_print_error("Failed to allocate memory");
eret = GASPI_ERR_MEMALLOC;
goto errL;
}
for(i = 0; i < glb_gaspi_ctx.tnc; i++)
glb_gaspi_ctx.sockfd[i] = -1;
}//glb_gaspi_dev_init
}//MASTER_PROC
else if(glb_gaspi_ctx.procType != WORKER_PROC)
{
gaspi_print_error ("Invalid node type (GASPI_TYPE)");
eret = GASPI_ERR_ENV;
goto errL;
}
if( 0 != gaspi_sn_broadcast_topology(&glb_gaspi_ctx, GASPI_BLOCK) )
{
gaspi_print_error("Failed topology broadcast");
eret = GASPI_ERROR;
goto errL;
}
if( (eret = pgaspi_init_core()) != GASPI_SUCCESS )
{
goto errL;
}
/* Unleash SN thread */
__sync_fetch_and_add( &gaspi_master_topo_data, 1);
gaspi_init_collectives();
glb_gaspi_init = 1;
unlock_gaspi (&glb_gaspi_ctx_lock);
if(glb_gaspi_cfg.build_infrastructure)
{
/* configuration tells us to pre-connect */
if( GASPI_TOPOLOGY_STATIC == glb_gaspi_cfg.build_infrastructure )
{
for(i = glb_gaspi_ctx.rank; i >= 0; i--)
{
if( (eret = pgaspi_connect((gaspi_rank_t) i, timeout_ms)) != GASPI_SUCCESS )
{
goto errL;
}
}
}
eret = pgaspi_group_all_local_create(timeout_ms);
if(eret == GASPI_SUCCESS)
{
eret = gaspi_barrier(GASPI_GROUP_ALL, timeout_ms);
}
else
{
gaspi_print_error("Failed to create GASPI_GROUP_ALL.");
}
}
else /* dont build_infrastructure */
{
/* just reserve GASPI_GROUP_ALL */
glb_gaspi_ctx.group_cnt = 1;
glb_gaspi_group_ctx[GASPI_GROUP_ALL].id = -2;//disable
eret = GASPI_SUCCESS;
}
#ifdef GPI2_CUDA
/* init GPU counts */
glb_gaspi_ctx.use_gpus = 0;
glb_gaspi_ctx.gpu_count = 0;
#endif
return eret;
errL:
unlock_gaspi (&glb_gaspi_ctx_lock);
return eret;
}
static gaspi_return_t
pgaspi_init_core()
{
int i;
if (glb_gaspi_dev_init)
return -1;
memset (&glb_gaspi_group_ctx, 0, GASPI_MAX_GROUPS * sizeof (gaspi_group_ctx));
for (i = 0; i < GASPI_MAX_GROUPS; i++)
{
GASPI_RESET_GROUP(glb_gaspi_group_ctx, i);
glb_gaspi_group_ctx[i].gl.lock = 0;
glb_gaspi_group_ctx[i].del.lock = 0;
}
/* change/override num of queues at large scale */
if (glb_gaspi_ctx.tnc > 1000 && glb_gaspi_cfg.queue_num > 1)
{
if(glb_gaspi_ctx.rank == 0)
gaspi_printf("Warning: setting number of queues to 1\n");
glb_gaspi_cfg.queue_num = 1;
}
/* Create internal memory space */
const unsigned int size = NOTIFY_OFFSET + sizeof(gaspi_atomic_value_t);
const long page_size = sysconf (_SC_PAGESIZE);
if(page_size < 0)
{
gaspi_print_error ("Failed to get system's page size.");
return GASPI_ERROR;
}
glb_gaspi_ctx.nsrc.size = size;
if(posix_memalign ((void **) &glb_gaspi_ctx.nsrc.ptr, page_size, size)!= 0)
{
gaspi_print_error ("Memory allocation (posix_memalign) failed");
return GASPI_ERR_MEMALLOC;
}
memset(glb_gaspi_ctx.nsrc.buf, 0, size);
for(i = 0; i < GASPI_MAX_MSEGS; i++)
{
glb_gaspi_ctx.rrmd[i] = NULL;
}
glb_gaspi_ctx.ep_conn = (gaspi_endpoint_conn_t *) calloc(glb_gaspi_ctx.tnc, sizeof(gaspi_endpoint_conn_t));
if (glb_gaspi_ctx.ep_conn == NULL)
return GASPI_ERR_MEMALLOC;
if(pgaspi_dev_init_core(&glb_gaspi_cfg) != 0)
return GASPI_ERR_DEVICE;
for(i = 0; i < GASPI_MAX_QP + 3; i++)
{
glb_gaspi_ctx.qp_state_vec[i] = (unsigned char *) calloc (glb_gaspi_ctx.tnc, sizeof(unsigned char));
if(!glb_gaspi_ctx.qp_state_vec[i])
{
return GASPI_ERR_MEMALLOC;
}
}
glb_gaspi_dev_init = 1;
return GASPI_SUCCESS;
}
gaspi_return_t
pgaspi_atomic_fetch_add (const gaspi_segment_id_t segment_id,
const gaspi_offset_t offset, const gaspi_rank_t rank,
const gaspi_atomic_value_t val_add,
gaspi_atomic_value_t * const val_old,
const gaspi_timeout_t timeout_ms)
{
#ifdef DEBUG
if (glb_gaspi_ctx_ib.rrmd[segment_id] == NULL)
{
gaspi_printf("Debug: Invalid segment (gaspi_atomic_fetch_add)\n");
return GASPI_ERROR;
}
if( rank >= glb_gaspi_ctx.tnc)
{
gaspi_printf("Debug: Invalid rank (gaspi_atomic_fetch_add)\n");
return GASPI_ERROR;
}
if( offset > glb_gaspi_ctx_ib.rrmd[segment_id][rank].size)
{
gaspi_printf("Debug: Invalid offsets (gaspi_atomic_fetch_add)\n");
return GASPI_ERROR;
}
if( val_old == NULL)
{
gaspi_printf("Debug: Invalid pointer in parameter val_old (gaspi_atomic_fetch_add)\n");
return GASPI_ERROR;
}
#endif
struct ibv_send_wr *bad_wr;
struct ibv_sge slist;
struct ibv_send_wr swr;
int i;
if (offset & 0x7)
{
gaspi_print_error("Unaligned offset");
return GASPI_ERROR;
}
lock_gaspi_tout (&glb_gaspi_group_ib[0].gl, timeout_ms);
slist.addr = (uintptr_t) (glb_gaspi_group_ib[0].buf + NEXT_OFFSET);
slist.length = 8;
slist.lkey = glb_gaspi_group_ib[0].mr->lkey;
swr.wr.atomic.remote_addr =
glb_gaspi_ctx_ib.rrmd[segment_id][rank].addr + NOTIFY_OFFSET + offset;
swr.wr.atomic.rkey = glb_gaspi_ctx_ib.rrmd[segment_id][rank].rkey;
swr.wr.atomic.compare_add = val_add;
swr.wr_id = rank;
swr.sg_list = &slist;
swr.num_sge = 1;
swr.opcode = IBV_WR_ATOMIC_FETCH_AND_ADD;
swr.send_flags = IBV_SEND_SIGNALED;
swr.next = NULL;
if (ibv_post_send (glb_gaspi_ctx_ib.qpGroups[rank], &swr, &bad_wr))
{
glb_gaspi_ctx.qp_state_vec[GASPI_COLL_QP][rank] = 1;
unlock_gaspi (&glb_gaspi_group_ib[0].gl);
return GASPI_ERROR;
}
glb_gaspi_ctx_ib.ne_count_grp++;
int ne = 0;
for (i = 0; i < glb_gaspi_ctx_ib.ne_count_grp; i++)
{
do
{
ne =
ibv_poll_cq (glb_gaspi_ctx_ib.scqGroups, 1,
glb_gaspi_ctx_ib.wc_grp_send);
}
while (ne == 0);
if ((ne < 0)
|| (glb_gaspi_ctx_ib.wc_grp_send[i].status != IBV_WC_SUCCESS))
{
glb_gaspi_ctx.
qp_state_vec[GASPI_COLL_QP][glb_gaspi_ctx_ib.wc_grp_send[i].
wr_id] = 1;
unlock_gaspi (&glb_gaspi_group_ib[0].gl);
return GASPI_ERROR;
}
}
glb_gaspi_ctx_ib.ne_count_grp = 0;
*val_old =
*((gaspi_atomic_value_t *) (glb_gaspi_group_ib[0].buf + NEXT_OFFSET));
unlock_gaspi (&glb_gaspi_group_ib[0].gl);
return GASPI_SUCCESS;
}
gaspi_return_t
gaspi_sn_command(const enum gaspi_sn_ops op, const gaspi_rank_t rank, const gaspi_timeout_t timeout_ms, void * arg)
{
int ret = -1;
gaspi_return_t eret = GASPI_ERROR;
const int i = (int) rank;
eret = gaspi_sn_connect_to_rank(rank, timeout_ms);
if(eret != GASPI_SUCCESS)
{
return eret;
}
eret = GASPI_ERROR;
switch(op)
{
case GASPI_SN_CONNECT:
{
ret = _gaspi_sn_connect_command(rank);
break;
}
case GASPI_SN_PROC_PING:
case GASPI_SN_PROC_KILL:
{
ret = _gaspi_sn_single_command(rank, op);
break;
}
case GASPI_SN_SEG_REGISTER:
{
ret = _gaspi_sn_segment_register_command(rank, arg);
break;
}
case GASPI_SN_GRP_CHECK:
{
ret = _gaspi_sn_group_check(rank, timeout_ms, arg);
break;
}
case GASPI_SN_GRP_CONNECT:
{
ret = _gaspi_sn_group_connect(rank, arg);
break;
}
default:
{
gaspi_print_error("Unknown SN op");
eret = GASPI_ERROR;
}
};
if( 0 == ret )
eret = GASPI_SUCCESS;
if( 1 == ret )
eret = GASPI_TIMEOUT;
if(gaspi_sn_close(glb_gaspi_ctx.sockfd[i]) != 0)
{
gaspi_print_error("Failed to close socket to %d", i);
}
glb_gaspi_ctx.sockfd[i] = -1;
return eret;
}
void *
gaspi_sn_thread (void *arg)
{
gaspi_sn_packet snp;
struct sockaddr_in cliAddr;
fd_set rfds;
int i, ret;
const int dsock = gaspi_setup_dg_socket ();
if (dsock == -1)
{
gaspi_print_error ("Failed to setup create SN thread socket");
return NULL;
}
if (__sync_fetch_and_add (&glb_gaspi_sn_init, 1) != 0)
gaspi_print_error ("Failed SN init");
int local_fd =
gaspi_listen2port (GASPI_INT_PORT + glb_gaspi_ctx.localSocket,
GASPI_BLOCK);
if (local_fd < 0)
{
gaspi_print_error ("Failed to initialize SN thread");
return NULL;
}
while (1)
{
FD_ZERO (&rfds);
FD_SET (dsock, &rfds);
const int selret = select (FD_SETSIZE, &rfds, NULL, NULL, NULL);
if (selret <= 0)
{
continue;
}
if (FD_ISSET (dsock, &rfds))
{
const int cliLen = sizeof (cliAddr);
const int rlen =
recvfrom (dsock, &snp, sizeof (gaspi_sn_packet), MSG_WAITALL,
(struct sockaddr *) &cliAddr, (socklen_t *) & cliLen);
if ((rlen != sizeof (gaspi_sn_packet))
|| (snp.magic != GASPI_SNP_MAGIC))
goto checkL;
char hn[128];
int hn_found = 0;
getnameinfo ((struct sockaddr *) &cliAddr, cliLen, hn, 128, NULL, 0,
NI_NOFQDN);
const char *fhn = strtok (hn, ".");
for (i = 0; i < glb_gaspi_ctx.tnc; i++)
{
if (strncmp ((glb_gaspi_ctx.hn + i * 64), fhn, 64) == 0)
{
hn_found = 1;
break;
}
if (strncmp ("localhost", fhn, 64) == 0)
{
hn_found = 1;
break;
}
}
if (!hn_found)
{
snp.ret = -1;
int ret =
sendto (dsock, &snp, sizeof (gaspi_sn_packet), MSG_WAITALL,
(struct sockaddr *) &cliAddr, sizeof (cliAddr));
if (ret != sizeof (gaspi_sn_packet))
{
gaspi_print_error ("Hostname not part of machinefile");
}
goto checkL;
}
if (snp.magic == GASPI_SNP_MAGIC)
{
switch (snp.cmd)
{
case 1:
snp.ret = 0;
ret =
sendto (dsock, &snp, sizeof (gaspi_sn_packet),
MSG_WAITALL, (struct sockaddr *) &cliAddr,
sizeof (cliAddr));
if (ret != sizeof (gaspi_sn_packet))
{
gaspi_print_error ("SN thread failed to send cmd 1");
}
return NULL;
break;
case 2:
snp.ret = 0;
ret =
sendto (dsock, &snp, sizeof (gaspi_sn_packet),
MSG_WAITALL, (struct sockaddr *) &cliAddr,
sizeof (cliAddr));
if (ret != sizeof (gaspi_sn_packet))
{
gaspi_print_error ("SN thread failed to send cmd 2");
}
break;
case 3:
snp.ret = 0;
ret =
sendto (dsock, &snp, sizeof (gaspi_sn_packet),
MSG_WAITALL, (struct sockaddr *) &cliAddr,
sizeof (cliAddr));
if (ret != sizeof (gaspi_sn_packet))
{
gaspi_print_error ("SN thread failed to send cmd 3");
}
exit (-1);
break;
case 4:
snp.ret = gaspi_seg_reg_sn (snp);
ret =
sendto (dsock, &snp, sizeof (gaspi_sn_packet),
MSG_WAITALL, (struct sockaddr *) &cliAddr,
sizeof (cliAddr));
if (ret != sizeof (gaspi_sn_packet))
{
gaspi_print_error ("SN thread failed to send cmd 4");
}
break;
default:
break;
}; //switch
} //if
} //if(dsock...
checkL:
continue;
} //while(1)
return NULL;
}
gaspi_return_t
gaspi_gpu_init(void)
{
gaspi_context_t * const gctx = &glb_gaspi_ctx;
int deviceCount;
cudaError_t cuda_error_id = cudaGetDeviceCount(&deviceCount);
if( cuda_error_id != cudaSuccess )
{
gaspi_print_error("Failed cudaGetDeviceCount." );
return GASPI_ERR_DEVICE;
}
if( deviceCount <= 0 )
{
gaspi_print_error("No CUDA capable devices found.");
return GASPI_ERR_DEVICE;
}
const int ib_numa_node = _gaspi_find_dev_numa_node();
int device_id = 0;
int gaspi_devices = 0;
int direct_devices[GPI2_GPU_MAX_DIRECT_DEVS];
struct cudaDeviceProp deviceProp;
for(device_id = 0; device_id < deviceCount; device_id++)
{
//TODO: possibly add functionality to show properties structure
cuda_error_id = cudaGetDeviceProperties(&deviceProp, device_id);
if( cuda_error_id != cudaSuccess)
{
return GASPI_ERR_DEVICE;
}
if( deviceProp.major >= 3 ) /* TODO: magic number */
{
cuda_error_id = cudaSetDevice(device_id);
if( cuda_error_id != cudaSuccess )
{
return GASPI_ERR_DEVICE;
}
if( ib_numa_node == _gaspi_find_GPU_numa_node(device_id) )
{
if( gaspi_devices < GPI2_GPU_MAX_DIRECT_DEVS - 1 )
{
direct_devices[gaspi_devices] = device_id;
gaspi_devices++;
}
}
}
}
if( 0 == gaspi_devices )
{
gaspi_print_error("No GPU Direct RDMA capable devices on the correct NUMA-socket were found.");
return GASPI_ERROR;
}
gpus = (gaspi_gpu_t*) malloc(sizeof(gaspi_gpu_t) * gaspi_devices);
if( gpus == NULL )
{
gaspi_print_error("Failed to allocate memory.");
return GASPI_ERR_MEMALLOC;
}
int i, j, k;
for(k = 0 ; k < gaspi_devices; k++)
{
cuda_error_id = cudaSetDevice(direct_devices[k]);
if( cuda_error_id != cudaSuccess )
{
return GASPI_ERR_DEVICE;
}
for(i = 0; i < GASPI_MAX_QP; i++)
{
cuda_error_id = cudaStreamCreate(&gpus[k].streams[i]);
if( cuda_error_id != cudaSuccess )
{
return GASPI_ERR_DEVICE;
}
for(j = 0; j < GASPI_CUDA_EVENTS; j++)
{
cuda_error_id = cudaEventCreateWithFlags(&gpus[k].events[i][j].event, cudaEventDisableTiming);
if( cuda_error_id != cudaSuccess )
{
return GASPI_ERR_DEVICE;
}
}
cuda_error_id = cudaStreamCreateWithFlags(&gpus[k].streams[i], cudaStreamNonBlocking);
if( cuda_error_id != cudaSuccess )
{
return GASPI_ERR_DEVICE;
}
}
gpus[k].device_id = direct_devices[k];
}
gctx->gpu_count = gaspi_devices;
gctx->use_gpus = 1;
return GASPI_SUCCESS;
}
static int
gaspi_sn_send_topology(gaspi_context *ctx, const int i, const gaspi_timeout_t timeout_ms)
{
if( (ctx->sockfd[i] =
gaspi_sn_connect2port(gaspi_get_hn(i),
(glb_gaspi_cfg.sn_port + 64 + ctx->poff[i]),
timeout_ms)) < 0)
{
gaspi_print_error("Failed to connect to %d", i);
return -1;
}
if( 0 != gaspi_sn_set_default_opts(ctx->sockfd[i]) )
{
gaspi_print_error("Failed to opts");
close(ctx->sockfd[i]);
return -1;
}
gaspi_cd_header cdh;
memset(&cdh, 0, sizeof(gaspi_cd_header));
cdh.op_len = ctx->tnc * 65; //TODO: 65 is magic
cdh.op = GASPI_SN_TOPOLOGY;
cdh.rank = i;
cdh.tnc = ctx->tnc;
int retval = 0;
size_t len = sizeof(gaspi_cd_header);
void * ptr = &cdh;
int sockfd = ctx->sockfd[i];
if (sockfd <= 0 )
{
gaspi_print_error("Wrong fd %d %d", i, ctx->sockfd[i] );
retval = -1;
goto endL;
}
if ( gaspi_sn_writen( sockfd, ptr, len) != len )
{
gaspi_print_error("Failed to send topology header to %d.", i);
retval = -1;
goto endL;
}
/* the de facto topology */
ptr = ctx->hn_poff;
len = ctx->tnc * 65;
if ( gaspi_sn_writen( sockfd, ptr, len) != len )
{
gaspi_print_error("Failed to send topology command to %d.", i);
retval = -1;
goto endL;
}
endL:
ctx->sockfd[i] = -1;
if(gaspi_sn_close( sockfd ) != 0)
retval = -1;
return retval;
}
gaspi_return_t
pgaspi_dev_atomic_compare_swap (const gaspi_segment_id_t segment_id,
const gaspi_offset_t offset,
const gaspi_rank_t rank,
const gaspi_atomic_value_t comparator,
const gaspi_atomic_value_t val_new)
{
struct ibv_send_wr *bad_wr;
struct ibv_sge slist;
struct ibv_send_wr swr;
int i;
slist.addr = (uintptr_t) (glb_gaspi_ctx.nsrc.buf + NOTIFY_OFFSET);
slist.length = sizeof(gaspi_atomic_value_t);
slist.lkey = ((struct ibv_mr *) glb_gaspi_ctx.nsrc.mr)->lkey;
swr.wr.atomic.remote_addr =
glb_gaspi_ctx.rrmd[segment_id][rank].addr + NOTIFY_OFFSET + offset;
swr.wr.atomic.rkey = glb_gaspi_ctx.rrmd[segment_id][rank].rkey;
swr.wr.atomic.compare_add = comparator;
swr.wr.atomic.swap = val_new;
swr.wr_id = rank;
swr.sg_list = &slist;
swr.num_sge = 1;
swr.opcode = IBV_WR_ATOMIC_CMP_AND_SWP;
swr.send_flags = IBV_SEND_SIGNALED;
swr.next = NULL;
if (ibv_post_send (glb_gaspi_ctx_ib.qpGroups[rank], &swr, &bad_wr))
{
glb_gaspi_ctx.qp_state_vec[GASPI_COLL_QP][rank] = GASPI_STATE_CORRUPT;
return GASPI_ERROR;
}
glb_gaspi_ctx.ne_count_grp++;
int ne = 0;
for (i = 0; i < glb_gaspi_ctx.ne_count_grp; i++)
{
do
{
ne = ibv_poll_cq (glb_gaspi_ctx_ib.scqGroups, 1,
glb_gaspi_ctx_ib.wc_grp_send);
}
while (ne == 0);
if ((ne < 0) || (glb_gaspi_ctx_ib.wc_grp_send[i].status != IBV_WC_SUCCESS))
{
glb_gaspi_ctx.qp_state_vec[GASPI_COLL_QP][glb_gaspi_ctx_ib.wc_grp_send[i].wr_id] = GASPI_STATE_CORRUPT;
gaspi_print_error("Failed request to %lu : %s",
glb_gaspi_ctx_ib.wc_grp_send[i].wr_id,
ibv_wc_status_str(glb_gaspi_ctx_ib.wc_grp_send[i].status));
return GASPI_ERROR;
}
}
glb_gaspi_ctx.ne_count_grp = 0;
return GASPI_SUCCESS;
}
void *gaspi_sn_backend(void *arg)
{
int esock, lsock, n, i;
struct epoll_event ev;
struct epoll_event *ret_ev;
gaspi_mgmt_header *ev_mgmt, *mgmt;
signal(SIGSTKFLT, gaspi_sn_cleanup);
signal(SIGPIPE, SIG_IGN);
while(gaspi_master_topo_data == 0)
gaspi_delay();
lsock = socket(AF_INET, SOCK_STREAM, 0);
if(lsock < 0)
{
gaspi_print_error("Failed to create socket");
gaspi_sn_status = GASPI_SN_STATE_ERROR;
gaspi_sn_err = GASPI_ERROR;
return NULL;
}
if( 0 != gaspi_sn_set_default_opts(lsock) )
{
gaspi_print_error("Failed to modify socket");
gaspi_sn_status = GASPI_SN_STATE_ERROR;
gaspi_sn_err = GASPI_ERROR;
close(lsock);
return NULL;
}
signal(SIGPIPE, SIG_IGN);
struct sockaddr_in listeningAddress;
listeningAddress.sin_family = AF_INET;
listeningAddress.sin_port = htons((glb_gaspi_cfg.sn_port + glb_gaspi_ctx.localSocket));
listeningAddress.sin_addr.s_addr = htonl(INADDR_ANY);
if(bind(lsock, (struct sockaddr*)(&listeningAddress), sizeof(listeningAddress)) < 0)
{
gaspi_print_error("Failed to bind socket (port %d)",
glb_gaspi_cfg.sn_port + glb_gaspi_ctx.localSocket);
gaspi_sn_status = GASPI_SN_STATE_ERROR;
gaspi_sn_err = GASPI_ERR_SN_PORT;
close(lsock);
return NULL;
}
if ( 0 != gaspi_sn_set_non_blocking(lsock) )
{
gaspi_print_error("Failed to set socket");
gaspi_sn_status = GASPI_SN_STATE_ERROR;
gaspi_sn_err = GASPI_ERROR;
close(lsock);
return NULL;
}
if(listen(lsock, SOMAXCONN) < 0)
{
gaspi_print_error("Failed to listen on socket");
gaspi_sn_status = GASPI_SN_STATE_ERROR;
gaspi_sn_err = GASPI_ERROR;
close(lsock);
return NULL;
}
esock = epoll_create(GASPI_EPOLL_CREATE);
if(esock < 0)
{
gaspi_print_error("Failed to create IO event facility");
gaspi_sn_status = GASPI_SN_STATE_ERROR;
gaspi_sn_err = GASPI_ERROR;
close(lsock);
return NULL;
}
/* add lsock to epoll instance */
ev.data.ptr = malloc( sizeof(gaspi_mgmt_header) );
if(ev.data.ptr == NULL)
{
gaspi_print_error("Failed to allocate memory");
gaspi_sn_status = GASPI_SN_STATE_ERROR;
gaspi_sn_err = GASPI_ERROR;
close(lsock);
return NULL;
}
ev_mgmt = ev.data.ptr;
ev_mgmt->fd = lsock;
ev.events = EPOLLIN;
if(epoll_ctl(esock, EPOLL_CTL_ADD, lsock, &ev) < 0)
{
gaspi_print_error("Failed to modify IO event facility");
gaspi_sn_status = GASPI_SN_STATE_ERROR;
gaspi_sn_err = GASPI_ERROR;
close(lsock);
return NULL;
}
ret_ev = calloc(GASPI_EPOLL_MAX_EVENTS, sizeof(ev));
if(ret_ev == NULL)
{
gaspi_print_error("Failed to allocate memory");
gaspi_sn_status = GASPI_SN_STATE_ERROR;
gaspi_sn_err = GASPI_ERROR;
close(lsock);
return NULL;
}
/* main events loop */
while(1)
{
n = epoll_wait(esock,ret_ev, GASPI_EPOLL_MAX_EVENTS, -1);
/* loop over all triggered events */
for( i = 0; i < n; i++ )
{
mgmt = ret_ev[i].data.ptr;
if( (ret_ev[i].events & EPOLLERR) || (ret_ev[i].events & EPOLLHUP) ||
!((ret_ev[i].events & EPOLLIN) || (ret_ev[i].events & EPOLLOUT )) )
{
/* an error has occured on this fd. close it => removed from event list. */
gaspi_print_error( "Erroneous event." );
shutdown(mgmt->fd, SHUT_RDWR);
close(mgmt->fd);
free(mgmt);
continue;
}
else if(mgmt->fd == lsock)
{
/* process all new connections */
struct sockaddr in_addr;
socklen_t in_len = sizeof(in_addr);
int nsock = accept( lsock, &in_addr, &in_len );
if(nsock < 0)
{
if( (errno == EAGAIN) || (errno == EWOULDBLOCK) )
{
/* we have processed incoming connection */
break;
}
else
{
/* at least check/fix open files limit */
int errsv = errno;
if(errsv == EMFILE)
{
if( 0 == _gaspi_check_ofile_limit() )
{
nsock = accept( lsock, &in_addr, &in_len );
}
}
/* still erroneous? => makes no sense to continue */
if(nsock < 0)
{
gaspi_print_error( "Failed to accept connection." );
gaspi_sn_status = GASPI_SN_STATE_ERROR;
gaspi_sn_err = GASPI_ERROR;
close(lsock);
return NULL;
}
}
}
/* new socket */
if( 0 != gaspi_sn_set_non_blocking( nsock ) )
{
gaspi_print_error( "Failed to set socket options." );
gaspi_sn_status = GASPI_SN_STATE_ERROR;
gaspi_sn_err = GASPI_ERROR;
close(nsock);
return NULL;
}
/* add nsock */
ev.data.ptr = malloc( sizeof(gaspi_mgmt_header) );
if(ev.data.ptr == NULL)
{
gaspi_print_error("Failed to allocate memory.");
gaspi_sn_status = GASPI_SN_STATE_ERROR;
gaspi_sn_err = GASPI_ERROR;
close(nsock);
return NULL;
}
ev_mgmt = ev.data.ptr;
ev_mgmt->fd = nsock;
ev_mgmt->blen = sizeof(gaspi_cd_header);
ev_mgmt->bdone = 0;
ev_mgmt->op = GASPI_SN_HEADER;
ev.events = EPOLLIN ; /* read only */
if(epoll_ctl( esock, EPOLL_CTL_ADD, nsock, &ev ) < 0)
{
gaspi_print_error("Failed to modify IO event facility");
gaspi_sn_status = GASPI_SN_STATE_ERROR;
gaspi_sn_err = GASPI_ERROR;
close(nsock);
return NULL;
}
continue;
}/* if new connection(s) */
else
{
/* read or write ops */
int io_err = 0;
if( ret_ev[i].events & EPOLLIN )
{
while( 1 )
{
int rcount = 0;
int rsize = mgmt->blen - mgmt->bdone;
char *ptr = NULL;
if( mgmt->op == GASPI_SN_HEADER )
{
/* TODO: is it valid? */
ptr = (char *) &mgmt->cdh;
rcount = read( mgmt->fd, ptr + mgmt->bdone, rsize );
}
else if( mgmt->op == GASPI_SN_CONNECT )
{
while( !glb_gaspi_dev_init )
gaspi_delay();
ptr = pgaspi_dev_get_rrcd(mgmt->cdh.rank);
rcount = read( mgmt->fd, ptr + mgmt->bdone, rsize );
}
/* errno==EAGAIN => we have read all data */
int errsv = errno;
if(rcount < 0)
{
if (errsv == ECONNRESET || errsv == ENOTCONN)
{
gaspi_print_error(" Failed to read (op %d)", mgmt->op);
}
if(errsv != EAGAIN || errsv != EWOULDBLOCK)
{
gaspi_print_error(" Failed to read (op %d).", mgmt->op);
io_err = 1;
}
break;
}
else if(rcount == 0) /* the remote side has closed the connection */
{
io_err = 1;
break;
}
else
{
mgmt->bdone += rcount;
/* read all data? */
if(mgmt->bdone == mgmt->blen)
{
/* we got header, what do we have to do ? */
if(mgmt->op == GASPI_SN_HEADER)
{
if(mgmt->cdh.op == GASPI_SN_PROC_KILL)
{
_exit(-1);
}
else if(mgmt->cdh.op == GASPI_SN_CONNECT)
{
GASPI_SN_RESET_EVENT( mgmt, mgmt->cdh.op_len, mgmt->cdh.op );
}
else if(mgmt->cdh.op == GASPI_SN_PROC_PING)
{
GASPI_SN_RESET_EVENT( mgmt, sizeof(gaspi_cd_header), GASPI_SN_HEADER );
}
else if(mgmt->cdh.op == GASPI_SN_GRP_CHECK)
{
struct{gaspi_group_t group;int tnc, cs, ret;} gb;
memset(&gb, 0, sizeof(gb));
gb.ret = -1;
gb.cs = 0;
const int group = mgmt->cdh.rank;
const int tnc = mgmt->cdh.tnc;
lock_gaspi_tout (&glb_gaspi_group_ctx[group].del, GASPI_BLOCK);
if(glb_gaspi_group_ctx[group].id >= 0)
{
if(glb_gaspi_group_ctx[group].tnc == tnc)
{
int i;
gb.ret = 0;
gb.tnc = tnc;
for(i = 0; i < tnc; i++)
{
if( NULL != glb_gaspi_group_ctx[group].rank_grp )
gb.cs ^= glb_gaspi_group_ctx[group].rank_grp[i];
}
}
}
unlock_gaspi (&glb_gaspi_group_ctx[group].del);
if(gaspi_sn_writen( mgmt->fd, &gb, sizeof(gb) ) < sizeof(gb) )
{
gaspi_print_error("Failed response to group check.");
io_err = 1;
break;
}
GASPI_SN_RESET_EVENT(mgmt, sizeof(gaspi_cd_header), GASPI_SN_HEADER );
}
else if(mgmt->cdh.op == GASPI_SN_GRP_CONNECT)
{
while( !glb_gaspi_dev_init ||
( glb_gaspi_group_ctx[mgmt->cdh.ret].id == -1) )
gaspi_delay();
/* TODO: check the pointer */
if(gaspi_sn_writen( mgmt->fd,
&glb_gaspi_group_ctx[mgmt->cdh.ret].rrcd[glb_gaspi_ctx.rank],
sizeof(gaspi_rc_mseg) ) < sizeof(gaspi_rc_mseg) )
{
gaspi_print_error("Failed to connect group.");
io_err = 1;
break;
}
GASPI_SN_RESET_EVENT( mgmt, sizeof(gaspi_cd_header), GASPI_SN_HEADER );
}
else if(mgmt->cdh.op == GASPI_SN_SEG_REGISTER)
{
int rret = gaspi_sn_segment_register(mgmt->cdh);
/* write back result of registration */
if(gaspi_sn_writen( mgmt->fd, &rret, sizeof(int) ) < sizeof(int) )
{
gaspi_print_error("Failed response to segment register.");
io_err = 1;
break;
}
GASPI_SN_RESET_EVENT(mgmt, sizeof(gaspi_cd_header), GASPI_SN_HEADER );
}
}/* !header */
else if(mgmt->op == GASPI_SN_CONNECT)
{
/* TODO: to remove */
while( !glb_gaspi_dev_init )
gaspi_delay();
const size_t len = pgaspi_dev_get_sizeof_rc();
char *ptr = NULL;
gaspi_return_t eret = pgaspi_create_endpoint_to(mgmt->cdh.rank, GASPI_BLOCK);
if( eret == GASPI_SUCCESS )
{
eret = pgaspi_connect_endpoint_to(mgmt->cdh.rank, GASPI_BLOCK);
if( eret == GASPI_SUCCESS)
{
ptr = pgaspi_dev_get_lrcd(mgmt->cdh.rank);
}
}
if( eret != GASPI_SUCCESS )
{
/* We set io_err, connection is closed and remote peer reads EOF */
io_err = 1;
}
else
{
if( NULL != ptr )
{
if( gaspi_sn_writen( mgmt->fd, ptr, len ) < sizeof(len) )
{
gaspi_print_error("Failed response to connection request from %u.", mgmt->cdh.rank);
io_err = 1;
}
}
}
GASPI_SN_RESET_EVENT( mgmt, sizeof(gaspi_cd_header), GASPI_SN_HEADER );
}
else
{
gaspi_print_error("Received unknown SN operation");
GASPI_SN_RESET_EVENT( mgmt, sizeof(gaspi_cd_header), GASPI_SN_HEADER );
}
break;
} /* if all data */
}/* else */
}/* while(1) read */
}/* read in */
if( io_err )
{
shutdown(mgmt->fd, SHUT_RDWR);
close(mgmt->fd);
free(mgmt);
}
}
} /* for each event */
}/* event loop while(1) */
return NULL;
}
static inline int
_gaspi_sn_group_check(const gaspi_rank_t rank, gaspi_timeout_t timeout_ms, void *arg)
{
struct group_desc *gb = (struct group_desc *) arg;
struct group_desc rem_gb;
int i = (int) rank;
struct timeb t0, t1;
ftime(&t0);
gaspi_cd_header cdh;
memset(&cdh, 0, sizeof(gaspi_cd_header));
cdh.op_len = sizeof (*gb);
cdh.op = GASPI_SN_GRP_CHECK;
cdh.rank = gb->group;
cdh.tnc = gb->tnc;
cdh.ret = gb->cs;
do
{
memset(&rem_gb, 0, sizeof(rem_gb));
ssize_t ret = gaspi_sn_writen(glb_gaspi_ctx.sockfd[i], &cdh, sizeof(gaspi_cd_header));
if(ret != sizeof(gaspi_cd_header) )
{
gaspi_print_error("Failed to write (%d %p %lu)",
glb_gaspi_ctx.sockfd[i], &cdh, sizeof(gaspi_cd_header));
return -1;
}
ssize_t rret = gaspi_sn_readn(glb_gaspi_ctx.sockfd[i], &rem_gb, sizeof(rem_gb));
if( rret != sizeof(rem_gb) )
{
gaspi_print_error("Failed to read (%d %p %lu)",
glb_gaspi_ctx.sockfd[i],&rem_gb,sizeof(rem_gb));
return -1;
}
if((rem_gb.ret < 0) || (gb->cs != rem_gb.cs))
{
ftime(&t1);
const unsigned int delta_ms = (t1.time - t0.time) * 1000 + (t1.millitm - t0.millitm);
if(delta_ms > timeout_ms)
{
return 1;
}
if(gaspi_thread_sleep(250) < 0)
{
gaspi_printf("gaspi_thread_sleep Error %d: (%s)\n",ret, (char*)strerror(errno));
}
//check if groups match
/* if(gb.cs != rem_gb.cs) */
/* { */
/* gaspi_print_error("Mismatch with rank %d: ranks in group dont match\n", */
/* group_to_commit->rank_grp[i]); */
/* eret = GASPI_ERROR; */
/* goto errL; */
/* } */
//usleep(250000);
//gaspi_delay();
}
else
{
break;
}
}while(1);
return 0;
}