static mca_mtl_base_module_t*
ompi_mtl_ofi_component_init(bool enable_progress_threads,
bool enable_mpi_threads)
{
int ret, fi_version;
struct fi_info *hints;
struct fi_info *providers = NULL, *prov = NULL;
struct fi_cq_attr cq_attr = {0};
struct fi_av_attr av_attr = {0};
char ep_name[FI_NAME_MAX] = {0};
size_t namelen;
/**
* Hints to filter providers
* See man fi_getinfo for a list of all filters
* mode: Select capabilities MTL is prepared to support.
* In this case, MTL will pass in context into communication calls
* ep_type: reliable datagram operation
* caps: Capabilities required from the provider.
* Tag matching is specified to implement MPI semantics.
* msg_order: Guarantee that messages with same tag are ordered.
*/
hints = fi_allocinfo();
if (!hints) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: Could not allocate fi_info\n",
__FILE__, __LINE__);
goto error;
}
hints->mode = FI_CONTEXT;
hints->ep_attr->type = FI_EP_RDM; /* Reliable datagram */
hints->caps = FI_TAGGED; /* Tag matching interface */
hints->tx_attr->msg_order = FI_ORDER_SAS;
hints->rx_attr->msg_order = FI_ORDER_SAS;
hints->domain_attr->threading = FI_THREAD_UNSPEC;
hints->domain_attr->control_progress = FI_PROGRESS_MANUAL;
/**
* FI_VERSION provides binary backward and forward compatibility support
* Specify the version of OFI is coded to, the provider will select struct
* layouts that are compatible with this version.
*/
fi_version = FI_VERSION(1, 0);
/**
* fi_getinfo: returns information about fabric services for reaching a
* remote node or service. this does not necessarily allocate resources.
* Pass NULL for name/service because we want a list of providers supported.
*/
ret = fi_getinfo(fi_version, /* OFI version requested */
NULL, /* Optional name or fabric to resolve */
NULL, /* Optional service name or port to request */
0ULL, /* Optional flag */
hints, /* In: Hints to filter providers */
&providers); /* Out: List of matching providers */
if (0 != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: fi_getinfo failed: %s\n",
__FILE__, __LINE__, fi_strerror(-ret));
goto error;
}
/**
* Select a provider from the list returned by fi_getinfo().
*/
prov = select_ofi_provider(providers);
if (!prov) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: select_ofi_provider: no provider found\n",
__FILE__, __LINE__);
goto error;
}
/**
* Open fabric
* The getinfo struct returns a fabric attribute struct that can be used to
* instantiate the virtual or physical network. This opens a "fabric
* provider". See man fi_fabric for details.
*/
ret = fi_fabric(prov->fabric_attr, /* In: Fabric attributes */
&ompi_mtl_ofi.fabric, /* Out: Fabric handle */
NULL); /* Optional context for fabric events */
if (0 != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: fi_fabric failed: %s\n",
__FILE__, __LINE__, fi_strerror(-ret));
goto error;
}
/**
* Create the access domain, which is the physical or virtual network or
* hardware port/collection of ports. Returns a domain object that can be
* used to create endpoints. See man fi_domain for details.
*/
ret = fi_domain(ompi_mtl_ofi.fabric, /* In: Fabric object */
prov, /* In: Provider */
&ompi_mtl_ofi.domain, /* Out: Domain oject */
NULL); /* Optional context for domain events */
if (0 != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: fi_domain failed: %s\n",
__FILE__, __LINE__, fi_strerror(-ret));
goto error;
}
/**
* Create a transport level communication endpoint. To use the endpoint,
* it must be bound to completion counters or event queues and enabled,
* and the resources consumed by it, such as address vectors, counters,
* completion queues, etc.
* see man fi_endpoint for more details.
*/
ret = fi_endpoint(ompi_mtl_ofi.domain, /* In: Domain object */
prov, /* In: Provider */
&ompi_mtl_ofi.ep, /* Out: Endpoint object */
NULL); /* Optional context */
if (0 != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: fi_endpoint failed: %s\n",
__FILE__, __LINE__, fi_strerror(-ret));
goto error;
}
/**
* Save the maximum inject size.
*/
ompi_mtl_ofi.max_inject_size = prov->tx_attr->inject_size;
/**
* Create the objects that will be bound to the endpoint.
* The objects include:
* - completion queue for events
* - address vector of other endpoint addresses
* - dynamic memory-spanning memory region
*/
cq_attr.format = FI_CQ_FORMAT_TAGGED;
ret = fi_cq_open(ompi_mtl_ofi.domain, &cq_attr, &ompi_mtl_ofi.cq, NULL);
if (ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: fi_cq_open failed: %s\n",
__FILE__, __LINE__, fi_strerror(-ret));
goto error;
}
/**
* The remote fi_addr will be stored in the ofi_endpoint struct.
* So, we use the AV in "map" mode.
*/
av_attr.type = FI_AV_MAP;
ret = fi_av_open(ompi_mtl_ofi.domain, &av_attr, &ompi_mtl_ofi.av, NULL);
if (ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: fi_av_open failed: %s\n",
__FILE__, __LINE__, fi_strerror(-ret));
goto error;
}
/**
* Bind the CQ and AV to the endpoint object.
*/
ret = fi_ep_bind(ompi_mtl_ofi.ep,
(fid_t)ompi_mtl_ofi.cq,
FI_SEND | FI_RECV);
if (0 != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: fi_bind CQ-EP failed: %s\n",
__FILE__, __LINE__, fi_strerror(-ret));
goto error;
}
ret = fi_ep_bind(ompi_mtl_ofi.ep,
(fid_t)ompi_mtl_ofi.av,
0);
if (0 != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: fi_bind AV-EP failed: %s\n",
__FILE__, __LINE__, fi_strerror(-ret));
goto error;
}
/**
* Enable the endpoint for communication
* This commits the bind operations.
*/
ret = fi_enable(ompi_mtl_ofi.ep);
if (0 != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: fi_enable failed: %s\n",
__FILE__, __LINE__, fi_strerror(-ret));
goto error;
}
/**
* Free providers info since it's not needed anymore.
*/
fi_freeinfo(hints);
hints = NULL;
fi_freeinfo(providers);
providers = NULL;
/**
* Get our address and publish it with modex.
*/
namelen = sizeof(ep_name);
ret = fi_getname((fid_t)ompi_mtl_ofi.ep, &ep_name[0], &namelen);
if (ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: fi_getname failed: %s\n",
__FILE__, __LINE__, fi_strerror(-ret));
goto error;
}
OFI_COMPAT_MODEX_SEND(ret,
&mca_mtl_ofi_component.super.mtl_version,
&ep_name,
namelen);
if (OMPI_SUCCESS != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: modex_send failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
ompi_mtl_ofi.epnamelen = namelen;
/**
* Set the ANY_SRC address.
*/
ompi_mtl_ofi.any_addr = FI_ADDR_UNSPEC;
/**
* Activate progress callback.
*/
ret = opal_progress_register(ompi_mtl_ofi_progress_no_inline);
if (OMPI_SUCCESS != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: opal_progress_register failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
return &ompi_mtl_ofi.base;
error:
if (providers) {
(void) fi_freeinfo(providers);
}
if (hints) {
(void) fi_freeinfo(hints);
}
if (ompi_mtl_ofi.av) {
(void) fi_close((fid_t)ompi_mtl_ofi.av);
}
if (ompi_mtl_ofi.cq) {
(void) fi_close((fid_t)ompi_mtl_ofi.cq);
}
if (ompi_mtl_ofi.ep) {
(void) fi_close((fid_t)ompi_mtl_ofi.ep);
}
if (ompi_mtl_ofi.domain) {
(void) fi_close((fid_t)ompi_mtl_ofi.domain);
}
if (ompi_mtl_ofi.fabric) {
(void) fi_close((fid_t)ompi_mtl_ofi.fabric);
}
return NULL;
}
static int common_setup(void)
{
struct fi_info *fi;
int ret;
ret = getaddr(src_addr, port, (struct sockaddr **) &hints.src_addr,
(socklen_t *) &hints.src_addrlen);
if (ret)
printf("source address error %s\n", gai_strerror(ret));
ret = fi_getinfo(FI_VERSION(1, 0), dst_addr, port, 0, &hints, &fi);
if (ret) {
printf("fi_getinfo %s\n", strerror(-ret));
goto err0;
}
if (fi->ep_attr->max_msg_size) {
max_msg_size = fi->ep_attr->max_msg_size;
}
ret = fi_fabric(fi->fabric_attr, &fab, NULL);
if (ret) {
printf("fi_fabric %s\n", fi_strerror(-ret));
goto err1;
}
if (fi->mode & FI_MSG_PREFIX) {
prefix_len = fi->ep_attr->msg_prefix_size;
}
ret = fi_domain(fab, fi, &dom, NULL);
if (ret) {
printf("fi_fdomain %s %s\n", fi_strerror(-ret),
fi->domain_attr->name);
goto err2;
}
if (fi->src_addr != NULL) {
((struct sockaddr_in *)fi->src_addr)->sin_port =
((struct sockaddr_in *)hints.src_addr)->sin_port;
if (dst_addr == NULL) {
printf("Local address %s:%d\n",
inet_ntoa(((struct sockaddr_in *)fi->src_addr)->sin_addr),
ntohs(((struct sockaddr_in *)fi->src_addr)->sin_port));
}
}
ret = fi_endpoint(dom, fi, &ep, NULL);
if (ret) {
printf("fi_endpoint %s\n", fi_strerror(-ret));
goto err3;
}
ret = alloc_ep_res(fi);
if (ret) {
printf("alloc_ep_res %s\n", fi_strerror(-ret));
goto err4;
}
ret = bind_ep_res();
if (ret) {
printf("bind_ep_res %s\n", fi_strerror(-ret));
goto err5;
}
if (hints.src_addr)
free(hints.src_addr);
fi_freeinfo(fi);
return 0;
err5:
free_ep_res();
err4:
fi_close(&ep->fid);
err3:
fi_close(&dom->fid);
err2:
fi_close(&fab->fid);
err1:
fi_freeinfo(fi);
err0:
if (hints.src_addr)
free(hints.src_addr);
return ret;
}
int do_test(void)
{
struct fi_cq_msg_entry comp;
int len = msg_len * post_depth;
int msg_cnt = num_msgs;
int tx_bufs_sent = 0;
int ret;
char *mp;
u64 time_elap;
#if SREAD == 0
int eagain_cnt = EAGAIN_TRIES;
#endif
print_trace("in\n");
if (!ctx.buf) {
ctx.buf = kmalloc(len, GFP_KERNEL);
if (!ctx.buf) {
print_err("kalloc failed!\n");
return -ENOMEM;
}
ret = fi_mr_reg(ctx.domain, ctx.buf, len, 0, 0, 0, 0,
&ctx.mr, NULL);
if (ret) {
print_err("fi_mr_reg returned %d\n", ret);
kfree(ctx.buf);
ctx.buf = ERR_PTR(-EFAULT);
return ret;
}
} else if (IS_ERR(ctx.buf))
return 0;
print_msg("post_depth %d num_msgs %d msg_len %d SREAD[%d]\n",
post_depth, num_msgs, msg_len, SREAD);
print_dbg("ctx.buf %p '%s' len %ld msg_len %d\n",
ctx.buf, ctx.buf, strlen(ctx.buf)+1, msg_len);
time_elap = get_jiffies_64();
for (mp = ctx.buf; msg_cnt > 0 && !kthread_should_stop(); ) {
int post_cnt, cnt;
post_cnt = (msg_cnt > post_depth ? post_depth : msg_cnt);
for (cnt = 0, mp = ctx.buf; cnt < post_cnt;
cnt++, mp += msg_len) {
if (verify) {
sprintf(mp, TEST_MESSAGE, tx_bufs_sent);
tx_bufs_sent++;
}
ret = fi_send(ctx.ep, mp, msg_len, fi_mr_desc(ctx.mr),
0, mp);
if (ret) {
print_err("fi_send returned %d '%s'\n",
ret, fi_strerror(ret));
return ret;
}
if (kthread_should_stop())
return -EINTR;
}
/* reap completions */
for (cnt = 0; cnt < post_cnt; cnt++) {
#if SREAD
ret = fi_cq_sread(ctx.scq, &comp, 1, 0, TIMEOUT);
if (ret == -ETIMEDOUT) {
print_msg("%s(ETIMEDOUT) cnt %d post_cnt %d "
"msg_cnt %d\n", "fi_cq_sread", cnt,
post_cnt, msg_cnt);
}
if (kthread_should_stop())
return -EINTR;
#else
do {
ret = fi_cq_read(ctx.scq, &comp, 1);
if (ret == 0 || ret == -EAGAIN) {
if (--eagain_cnt <= 0) {
dprint(DEBUG_HIGH,
"%s(resched %d) cnt "
"%d post_cnt %d\n",
"fi_cq_read", ret, cnt,
post_cnt);
eagain_cnt = EAGAIN_TRIES;
schedule();
}
}
if (kthread_should_stop())
return -EINTR;
} while (ret == 0 || ret == -EAGAIN);
#endif
if (ret < 0) {
struct fi_cq_err_entry cqe = { 0 };
int rc;
rc = fi_cq_readerr(ctx.scq, &cqe, 0);
print_err("fi_cq_read returned %d '%s'\n",
ret, fi_strerror(ret));
if (rc) {
char buf[64];
print_err("fi_cq_readerr() err '%s'(%d)"
"\n", fi_strerror(cqe.err),
cqe.err);
print_err("fi_cq_readerr() prov_err "
"'%s'(%d)\n",
fi_cq_strerror(ctx.scq,
cqe.prov_errno,
cqe.err_data, buf,
sizeof(buf)),
cqe.prov_errno);
}
return ret;
}
if (!ret)
print_err("fi_cq_sread no completion? ret %d\n",
ret);
#if 0
if ((char *)comp.op_context < (char *)ctx.buf ||
(char *)comp.op_context >= (char *)
&ctx.buf[msg_len*post_depth]) {
print_err("cq.op_context(%p) not in range "
"[ctx.buf(%p) ... &ctx.buf[%d](%p)]\n",
(void *)comp.op_context,
(void *)ctx.buf,
msg_len,
(void *)&ctx.buf[msg_len]);
}
#endif
if (verify)
print_msg("Tx '%s'\n",
(char *) comp.op_context);
}
msg_cnt -= post_cnt;
}
time_elap = get_jiffies_64() - time_elap;
#define AGIG (1024UL*1024UL*1024UL)
#define AMEG (1024UL*1024UL)
#define AKILO (1024UL)
{
struct timeval tv;
ulong rate, rate_mod, bytes, units_of;
char units;
jiffies_to_timeval(time_elap, &tv);
bytes = (ulong) num_msgs * (ulong) msg_len;
if (bytes >= AKILO && tv.tv_sec > 0) {
rate = bytes / tv.tv_sec;
rate_mod = bytes % tv.tv_sec;
if (rate >= AGIG) {
units = 'G';
units_of = AGIG;
} else if (rate >= AMEG) {
units = 'M';
units_of = AMEG;
} else {
units = 'K';
units_of = AKILO;
}
rate /= units_of;
} else {
rate = rate_mod = 0UL;
units = ' ';
units_of = 1UL;
}
print_info("Tx %d msgs (%lu.%lu%cB) @ ~%lu.%lu %cB/sec (%ld sec %ld "
"usec)\n",
num_msgs, (bytes/units_of), (bytes % units_of),
units, rate, rate_mod, units,
tv.tv_sec, tv.tv_usec);
}
return 0;
}
static void setup_ofi_active(struct fi_info *info,
struct fid_ep **ep)
{
// Make an EQ
int ret;
ret = fi_endpoint(fidev.domain, info, ep, NULL);
if (0 != ret) {
error("fi_endpoint failed");
}
#if WANT_FDS
// Add the EQ FD to the epoll fd
static struct epoll_event edt;
memset(&edt, 0, sizeof(edt));
edt.events = EPOLLIN;
edt.data.u32 = 2222;
ret = epoll_ctl(epoll_fd, EPOLL_CTL_ADD, fidev.eq_fd, &edt);
if (ret < 0) {
error("epoll_ctl failed");
}
#endif
// Bind the EP to the EQ
ret = fi_ep_bind(*ep, &fidev.eq->fid, 0);
if (0 != ret) {
error("fi_ep_bind(eq) failed");
}
// Make a CQ
struct fi_cq_attr cq_attr;
memset(&cq_attr, 0, sizeof(cq_attr));
cq_attr.format = FI_CQ_FORMAT_CONTEXT;
cq_attr.wait_obj = FI_WAIT_FD;
cq_attr.size = 32; // JMS POC
ret = fi_cq_open(fidev.domain, &cq_attr, &ficonn.cq, NULL);
if (ret != 0) {
error("fi_cq_open failed");
}
// Bind the CQ TX and RX queues to the EQ
ret = fi_ep_bind(*ep, &ficonn.cq->fid, FI_TRANSMIT);
if (0 != ret) {
error("fi_ep_bind(cq tx) failed");
}
ret = fi_ep_bind(*ep, &ficonn.cq->fid, FI_RECV);
if (0 != ret) {
error("fi_ep_bind(cq rx) failed");
}
#if WANT_FDS
// Get the fd associated with this CQ
ret = fi_control(&(ficonn.cq->fid), FI_GETWAIT, &ficonn.cq_fd);
if (ret != 0) {
error("fi_control to get cq fq failed");
}
#endif
// Enable the EP!
ret = fi_enable(*ep);
if (0 != ret) {
error("fi_enable failed");
}
// Register the buffers (must use different keys for each)
ret = fi_mr_reg(fidev.domain, send_buffer, sizeof(send_buffer),
FI_SEND, 0, (uintptr_t) send_buffer,
0, &ficonn.send_mr, NULL);
if (ret != 0) {
error("fi_mr_reg(send) failed\n");
}
ret = fi_mr_reg(fidev.domain, recv_buffer, sizeof(recv_buffer),
FI_RECV, 0, (uintptr_t) recv_buffer,
0, &ficonn.recv_mr, NULL);
if (ret != 0) {
printf("ERROR: ret=%d, %s\n", ret, fi_strerror(-ret));
error("fi_mr_reg(recv) failed\n");
}
}
static void test_connect_with_accept_blocking_on_eq_fq_CLIENT(void)
{
int ret;
printf("CLIENT running\n");
// Get the server's node (IP addr) and service (port)
MPI_Recv(ofi_node, sizeof(ofi_node) - 1, MPI_CHAR,
0, 101, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
MPI_Recv(ofi_service, sizeof(ofi_service) - 1, MPI_CHAR,
0, 102, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
printf("CLIENT received via MPI: %s / %s\n", ofi_node, ofi_service);
//setup_ofi(ofi_node, ofi_service);
setup_ofi(NULL, NULL, 0);
memset(&sin, 0, sizeof(sin));
sin.sin_family = AF_INET;
inet_aton(ofi_node, &sin.sin_addr);
sin.sin_port = htons(atoi(ofi_service));
printf("CLIENT translated: %s\n", addrstr(&sin));
setup_ofi_active(fidev.info, &ficonn.ep);
// Print server addr
printf("CLIENT connecting to %s\n", addrstr(&sin));
// Connect!
printf("Client connecting...\n");
ret = fi_connect(ficonn.ep,
//fidev.info->dest_addr,
&sin,
(void*) client_data, sizeof(client_data));
if (ret < 0) {
error("fi_connect failed");
}
#if WANT_FDS
// Now wait for the listen to complete
int nevents;
#define NEVENTS 32
struct epoll_event events[NEVENTS];
int timeout = 10000;
while (1) {
printf("CLIENT blocking on epoll\n");
nevents = epoll_wait(epoll_fd, events, NEVENTS, timeout);
if (nevents < 0) {
if (errno != EINTR) {
error("client epoll wait failed");
} else {
continue;
}
} else {
printf("CLIENT successfully woke up from epoll! %d events\n", nevents);
for (int i = 0; i < nevents; ++i) {
if (events[i].data.u32 != 2222) {
error("CLIENT unexpected epoll return type");
}
}
// If we got the expected event, then go read from the EQ
break;
}
}
#endif
// Wait for FI_CONNECTED event
uint32_t event;
uint8_t *entry_buffer;
size_t expected_len = sizeof(struct fi_eq_cm_entry) +
sizeof(client_data);
entry_buffer = (uint8_t*) calloc(1, expected_len);
if (NULL == entry_buffer) {
error("calloc failed");
}
struct fi_eq_cm_entry *entry = (struct fi_eq_cm_entry*) entry_buffer;
while (1) {
printf("CLIENT waiting for FI_CONNECTED\n");
#if WANT_FDS
ret = fi_eq_read(fidev.eq, &event, entry, expected_len, 0);
#else
ret = fi_eq_sread(fidev.eq, &event, entry, expected_len, -1, 0);
#endif
if (-FI_EAVAIL == ret) {
fprintf(stderr, "client fi_eq_sread failed because there's something in the error queue\n");
char buffer[2048];
struct fi_eq_err_entry *err_entry = (struct fi_eq_err_entry*) buffer;
ret = fi_eq_readerr(fidev.eq, err_entry, 0);
fprintf(stderr, "error code: %d (%s), prov err code: %d (%s)\n", err_entry->err, fi_strerror(err_entry->err), err_entry->prov_errno, fi_strerror(err_entry->prov_errno));
error("sad panda");
} else if (ret == -EAGAIN) {
fprintf(stderr, "CLIENT fi_eq_sread fail got -EAGAIN... trying again...\n");
sleep(1);
continue;
} else if (ret < 0) {
fprintf(stderr, "SERVER fi_eq_sread fail: %s, ret = %d)\n", fi_strerror(-ret), ret);
error("client fi_eq_sread failed for some random reason");
} else if (event != FI_CONNECTED) {
error("client got some unexpected event");
} else if (ret != expected_len) {
error("client got wrong length back from fi_eq_sread");
}
uint32_t *d = (uint32_t*) entry->data;
for (int i = 0; i < (sizeof(server_data) / sizeof(uint32_t)); ++i) {
if (d[i] != server_data[i]) {
printf("CLIENT got wrong CM client data: d[%d]=%d, should be %d\n",
i, d[i], server_data[i]);
}
}
printf("client got FI_CONNECTED, correct size, and correct data -- yay!\n");
break;
}
printf("CLIENT connecting -- waiting for server before sending\n");
MPI_Barrier(MPI_COMM_WORLD);
sleep(1);
int msg[4] = { 99, 100, 101, 102 };
int len = sizeof(msg);
printf("CLIENT sending len of %d\n", len);
struct fid_mr no_mr;
struct fid_mr *mr;
void *send_context = (void*) 0x42;
#if 0
fi_mr_reg(fidev.domain, msg, len, FI_SEND | FI_RECV,
0, (uint64_t)(uintptr_t) msg, 0, &mr, NULL);
#else
// Try using no mr, like fi_msg_pingpong...
memset(&no_mr, 0, sizeof(no_mr));
mr = &no_mr;
#endif
ret = fi_send(ficonn.ep, msg, len,
fi_mr_desc(mr), 0, send_context);
if (ret < 0) {
printf("fi_Send failed! %d, %s\n", ret, fi_strerror(-ret));
MPI_Abort(MPI_COMM_WORLD, 37);
}
// Wait for send completion
struct fi_cq_entry cqe;
while (1) {
ret = fi_cq_sread(ficonn.cq, &cqe, 1, 0, -1);
if (cqe.op_context == send_context) {
printf("CLIENT send completed\n");
break;
} else {
printf("CLIENT got some other completion... continuing\n");
}
}
printf("CLIENT sent -- waiting for server before teardown\n");
MPI_Barrier(MPI_COMM_WORLD);
printf("CLIENT tearing down\n");
fi_close(&(mr->fid));
teardown_ofi();
}
/*
* Set up the progress thread
*/
static void progress_thread(void *args) {
struct progress_thread_info* pti = args;
const int id = pti->id;
const int num_rbufs = 2;
struct iovec iov[num_rbufs];
struct fi_msg msg[num_rbufs];
struct ofi_am_info* dst_buf[num_rbufs];
const int rbuf_len = 10;
const size_t rbuf_size = rbuf_len*sizeof(dst_buf[0][0]);
const int num_cqes = rbuf_len;
struct fi_cq_data_entry cqes[num_cqes];
int num_read;
int i;
for (i = 0; i < num_rbufs; i++) {
dst_buf[i] = chpl_mem_allocMany(rbuf_len, sizeof(dst_buf[i][0]),
CHPL_RT_MD_COMM_PER_LOC_INFO, 0, 0);
iov[i].iov_base = dst_buf[i];
iov[i].iov_len = rbuf_size;
msg[i].msg_iov = &iov[i];
msg[i].desc = (void **) fi_mr_desc(ofi.mr);
msg[i].iov_count = 1;
msg[i].addr = FI_ADDR_UNSPEC;
msg[i].context = (void *) (uint64_t) i;
msg[i].data = 0x0;
OFICHKERR(fi_recvmsg(ofi.am_rx_ep[id], &msg[i], FI_MULTI_RECV));
}
// Count this progress thread as running. The creator thread wants to
// be released as soon as at least one progress thread is running, so
// if we're the first, do that.
if (atomic_fetch_add_uint_least32_t(&progress_thread_count, 1) == 0) {
CALL_CHECK_ZERO(pthread_mutex_lock(&progress_thread_entEx_cond_mutex));
CALL_CHECK_ZERO(pthread_cond_signal(&progress_thread_enter_cond));
CALL_CHECK_ZERO(pthread_mutex_unlock(&progress_thread_entEx_cond_mutex));
}
// Wait for events
while (!atomic_load_bool(&progress_threads_please_exit)) {
num_read = fi_cq_read(ofi.am_rx_cq[id], cqes, num_cqes);
if (num_read > 0) {
for (i = 0; i < num_read; i++) {
chpl_comm_ofi_am_handler(&cqes[i]);
// send ack
}
} else {
if (num_read != -FI_EAGAIN) {
chpl_internal_error(fi_strerror(-num_read));
}
}
}
// Un-count this progress thread. Whoever told us to exit wants to
// be released once all the progress threads are done, so if we're
// the last, do that.
if (atomic_fetch_sub_uint_least32_t(&progress_thread_count, 1) == 1) {
CALL_CHECK_ZERO(pthread_mutex_lock(&progress_thread_entEx_cond_mutex));
CALL_CHECK_ZERO(pthread_cond_signal(&progress_thread_exit_cond));
CALL_CHECK_ZERO(pthread_mutex_unlock(&progress_thread_entEx_cond_mutex));
}
}
static int client_connect(void)
{
struct fi_eq_cm_entry entry;
uint32_t event;
struct fi_info *fi;
ssize_t rd;
int ret;
if (src_addr) {
ret = getaddr(src_addr, NULL, (struct sockaddr **) &hints.src_addr,
(socklen_t *) &hints.src_addrlen);
if (ret)
printf("source address error %s\n", gai_strerror(ret));
}
ret = fi_getinfo(FI_VERSION(1, 0), dst_addr, port, 0, &hints, &fi);
if (ret) {
printf("fi_getinfo %s\n", strerror(-ret));
goto err0;
}
ret = fi_fabric(fi->fabric_attr, &fab, NULL);
if (ret) {
printf("fi_fabric %s\n", fi_strerror(-ret));
goto err1;
}
ret = fi_domain(fab, fi, &dom, NULL);
if (ret) {
printf("fi_domain %s %s\n", fi_strerror(-ret),
fi->domain_attr->name);
goto err2;
}
ret = fi_endpoint(dom, fi, &ep, NULL);
if (ret) {
printf("fi_endpoint %s\n", fi_strerror(-ret));
goto err3;
}
ret = alloc_ep_res(fi);
if (ret)
goto err4;
ret = bind_ep_res();
if (ret)
goto err5;
ret = fi_connect(ep, fi->dest_addr, NULL, 0);
if (ret) {
printf("fi_connect %s\n", fi_strerror(-ret));
goto err5;
}
rd = fi_eq_sread(cmeq, &event, &entry, sizeof entry, -1, 0);
if (rd != sizeof entry) {
printf("fi_eq_sread %zd %s\n", rd, fi_strerror((int) -rd));
return (int) rd;
}
if (event != FI_COMPLETE || entry.fid != &ep->fid) {
printf("Unexpected CM event %d fid %p (ep %p)\n",
event, entry.fid, ep);
ret = -FI_EOTHER;
goto err1;
}
if (hints.src_addr)
free(hints.src_addr);
fi_freeinfo(fi);
return 0;
err5:
free_ep_res();
err4:
fi_close(&ep->fid);
err3:
fi_close(&dom->fid);
err2:
fi_close(&fab->fid);
err1:
fi_freeinfo(fi);
err0:
if (hints.src_addr)
free(hints.src_addr);
return ret;
}
/* mca_btl_ofi_context_alloc_scalable()
*
* This function allocate communication contexts and return the pointer
* to the first btl context. It also take care of all the bindings needed.
* USE WITH SCALABLE ENDPOINT ONLY */
mca_btl_ofi_context_t *mca_btl_ofi_context_alloc_scalable(struct fi_info *info,
struct fid_domain *domain,
struct fid_ep *sep,
struct fid_av *av,
size_t num_contexts)
{
BTL_VERBOSE(("creating %zu contexts", num_contexts));
int rc;
size_t i;
char *linux_device_name = info->domain_attr->name;
struct fi_cq_attr cq_attr = {0};
struct fi_tx_attr tx_attr = {0};
struct fi_rx_attr rx_attr = {0};
mca_btl_ofi_context_t *contexts;
tx_attr.op_flags = FI_DELIVERY_COMPLETE;
contexts = (mca_btl_ofi_context_t*) calloc(num_contexts, sizeof(*contexts));
if (NULL == contexts) {
BTL_VERBOSE(("cannot allocate communication contexts."));
return NULL;
}
/* Don't really need to check, just avoiding compiler warning because
* BTL_VERBOSE is a no op in performance build and the compiler will
* complain about unused variable. */
if (NULL == linux_device_name) {
BTL_VERBOSE(("linux device name is NULL. This shouldn't happen."));
goto scalable_fail;
}
/* bind AV to endpoint */
rc = fi_scalable_ep_bind(sep, (fid_t)av, 0);
if (0 != rc) {
BTL_VERBOSE(("%s failed fi_scalable_ep_bind with err=%s",
linux_device_name,
fi_strerror(-rc)
));
goto scalable_fail;
}
for (i=0; i < num_contexts; i++) {
rc = fi_tx_context(sep, i, &tx_attr, &contexts[i].tx_ctx, NULL);
if (0 != rc) {
BTL_VERBOSE(("%s failed fi_tx_context with err=%s",
linux_device_name,
fi_strerror(-rc)
));
goto scalable_fail;
}
/* We don't actually need a receiving context as we only do one-sided.
* However, sockets provider will hang if we dont have one. It is
* also nice to have equal number of tx/rx context. */
rc = fi_rx_context(sep, i, &rx_attr, &contexts[i].rx_ctx, NULL);
if (0 != rc) {
BTL_VERBOSE(("%s failed fi_rx_context with err=%s",
linux_device_name,
fi_strerror(-rc)
));
goto scalable_fail;
}
/* create CQ */
cq_attr.format = FI_CQ_FORMAT_CONTEXT;
cq_attr.wait_obj = FI_WAIT_NONE;
rc = fi_cq_open(domain, &cq_attr, &contexts[i].cq, NULL);
if (0 != rc) {
BTL_VERBOSE(("%s failed fi_cq_open with err=%s",
linux_device_name,
fi_strerror(-rc)
));
goto scalable_fail;
}
/* bind cq to transmit context */
uint32_t cq_flags = (FI_TRANSMIT);
rc = fi_ep_bind(contexts[i].tx_ctx, (fid_t)contexts[i].cq, cq_flags);
if (0 != rc) {
BTL_VERBOSE(("%s failed fi_ep_bind with err=%s",
linux_device_name,
fi_strerror(-rc)
));
goto scalable_fail;
}
/* enable the context. */
rc = fi_enable(contexts[i].tx_ctx);
if (0 != rc) {
BTL_VERBOSE(("%s failed fi_enable with err=%s",
linux_device_name,
fi_strerror(-rc)
));
goto scalable_fail;
}
rc = fi_enable(contexts[i].rx_ctx);
if (0 != rc) {
BTL_VERBOSE(("%s failed fi_enable with err=%s",
linux_device_name,
fi_strerror(-rc)
));
goto scalable_fail;
}
/* initialize completion freelist. */
rc = ofi_comp_list_init(&contexts[i].comp_list);
if (rc != OPAL_SUCCESS) {
goto scalable_fail;
}
/* assign the id */
contexts[i].context_id = i;
}
return contexts;
scalable_fail:
/* close and free */
for(i=0; i < num_contexts; i++) {
mca_btl_ofi_context_finalize(&contexts[i], true);
}
free(contexts);
return NULL;
}
/* mca_btl_ofi_context_alloc_normal()
*
* This function will allocate an ofi_context, map the endpoint to tx/rx context,
* bind CQ,AV to the endpoint and initialize all the structure.
* USE WITH NORMAL ENDPOINT ONLY */
mca_btl_ofi_context_t *mca_btl_ofi_context_alloc_normal(struct fi_info *info,
struct fid_domain *domain,
struct fid_ep *ep,
struct fid_av *av)
{
int rc;
uint32_t cq_flags = FI_TRANSMIT;
char *linux_device_name = info->domain_attr->name;
struct fi_cq_attr cq_attr = {0};
mca_btl_ofi_context_t *context;
context = (mca_btl_ofi_context_t*) calloc(1, sizeof(*context));
if (NULL == context) {
BTL_VERBOSE(("cannot allocate context"));
return NULL;
}
/* Don't really need to check, just avoiding compiler warning because
* BTL_VERBOSE is a no op in performance build and the compiler will
* complain about unused variable. */
if (NULL == linux_device_name) {
BTL_VERBOSE(("linux device name is NULL. This shouldn't happen."));
goto single_fail;
}
cq_attr.format = FI_CQ_FORMAT_CONTEXT;
cq_attr.wait_obj = FI_WAIT_NONE;
rc = fi_cq_open(domain, &cq_attr, &context->cq, NULL);
if (0 != rc) {
BTL_VERBOSE(("%s failed fi_cq_open with err=%s",
linux_device_name,
fi_strerror(-rc)
));
goto single_fail;
}
rc = fi_ep_bind(ep, (fid_t)av, 0);
if (0 != rc) {
BTL_VERBOSE(("%s failed fi_ep_bind with err=%s",
linux_device_name,
fi_strerror(-rc)
));
goto single_fail;
}
rc = fi_ep_bind(ep, (fid_t)context->cq, cq_flags);
if (0 != rc) {
BTL_VERBOSE(("%s failed fi_scalable_ep_bind with err=%s",
linux_device_name,
fi_strerror(-rc)
));
goto single_fail;
}
rc = ofi_comp_list_init(&context->comp_list);
if (rc != OPAL_SUCCESS) {
goto single_fail;
}
context->tx_ctx = ep;
context->rx_ctx = ep;
context->context_id = 0;
return context;
single_fail:
mca_btl_ofi_context_finalize(context, false);
return NULL;
}
int _gnix_cm_nic_alloc(struct gnix_fid_domain *domain,
struct fi_info *info,
uint32_t cdm_id,
struct gnix_auth_key *auth_key,
struct gnix_cm_nic **cm_nic_ptr)
{
int ret = FI_SUCCESS;
struct gnix_cm_nic *cm_nic = NULL;
gnix_hashtable_attr_t gnix_ht_attr = {0};
uint32_t name_type = GNIX_EPN_TYPE_UNBOUND;
struct gnix_nic_attr nic_attr = {0};
struct gnix_ep_name ep_name;
struct gnix_dgram_hndl_attr dgram_hndl_attr = {0};
struct gnix_dgram_hndl_attr *dgram_hndl_attr_ptr = NULL;
GNIX_TRACE(FI_LOG_EP_CTRL, "\n");
*cm_nic_ptr = NULL;
/*
* if app has specified a src_addr in the info
* argument and length matches that for gnix_ep_name
* we must allocate a cm_nic, otherwise we first
* check to see if there is a cm_nic already for this domain
* and just use it.
*/
if (info->src_addr) {
/*TODO (optimization): strchr to name_type and strtol */
_gnix_get_ep_name(info->src_addr, 0, &ep_name, domain);
name_type = ep_name.name_type;
}
GNIX_INFO(FI_LOG_EP_CTRL, "creating cm_nic for %u/0x%x/%u\n",
auth_key->ptag, auth_key->cookie, cdm_id);
cm_nic = (struct gnix_cm_nic *)calloc(1, sizeof(*cm_nic));
if (cm_nic == NULL) {
ret = -FI_ENOMEM;
goto err;
}
/*
* we have to force allocation of a new nic since we want
* an a particulard cdm id
*/
nic_attr.must_alloc = true;
nic_attr.use_cdm_id = true;
nic_attr.cdm_id = cdm_id;
nic_attr.auth_key = auth_key;
ret = gnix_nic_alloc(domain, &nic_attr, &cm_nic->nic);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"gnix_nic_alloc returned %s\n",
fi_strerror(-ret));
goto err;
}
cm_nic->my_name.gnix_addr.cdm_id = cdm_id;
cm_nic->ptag = auth_key->ptag;
cm_nic->my_name.cookie = auth_key->cookie;
cm_nic->my_name.gnix_addr.device_addr = cm_nic->nic->device_addr;
cm_nic->domain = domain;
cm_nic->ctrl_progress = domain->control_progress;
cm_nic->my_name.name_type = name_type;
cm_nic->poll_cnt = 0;
fastlock_init(&cm_nic->wq_lock);
dlist_init(&cm_nic->cm_nic_wq);
/*
* prep the cm nic's dgram component
*/
if (domain->control_progress == FI_PROGRESS_AUTO) {
dgram_hndl_attr.timeout_needed = __gnix_cm_nic_timeout_needed;
dgram_hndl_attr.timeout_progress = __gnix_cm_nic_timeout_progress;
dgram_hndl_attr.timeout_data = (void *)cm_nic;
dgram_hndl_attr.timeout = domain->params.dgram_progress_timeout;
dgram_hndl_attr_ptr = &dgram_hndl_attr;
};
ret = _gnix_dgram_hndl_alloc(cm_nic,
domain->control_progress,
dgram_hndl_attr_ptr,
&cm_nic->dgram_hndl);
if (ret != FI_SUCCESS)
goto err;
/*
* allocate hash table for translating ep addresses
* to ep's.
* This table will not be large - how many FI_EP_RDM ep's
* will an app create using one domain?, nor in the critical path
* so just use defaults.
*/
cm_nic->addr_to_ep_ht = calloc(1, sizeof(struct gnix_hashtable));
if (cm_nic->addr_to_ep_ht == NULL)
goto err;
gnix_ht_attr.ht_initial_size = 64;
gnix_ht_attr.ht_maximum_size = 1024;
gnix_ht_attr.ht_increase_step = 2;
gnix_ht_attr.ht_increase_type = GNIX_HT_INCREASE_MULT;
gnix_ht_attr.ht_collision_thresh = 500;
gnix_ht_attr.ht_hash_seed = 0xdeadbeefbeefdead;
gnix_ht_attr.ht_internal_locking = 1;
gnix_ht_attr.destructor = NULL;
ret = _gnix_ht_init(cm_nic->addr_to_ep_ht, &gnix_ht_attr);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"gnix_ht_init returned %s\n",
fi_strerror(-ret));
goto err;
}
_gnix_ref_init(&cm_nic->ref_cnt, 1, __cm_nic_destruct);
*cm_nic_ptr = cm_nic;
pthread_mutex_lock(&gnix_cm_nic_list_lock);
dlist_insert_tail(&cm_nic->cm_nic_list, &gnix_cm_nic_list);
pthread_mutex_unlock(&gnix_cm_nic_list_lock);
return ret;
err:
if (cm_nic->dgram_hndl)
_gnix_dgram_hndl_free(cm_nic->dgram_hndl);
if (cm_nic->nic)
_gnix_nic_free(cm_nic->nic);
if (cm_nic->addr_to_ep_ht) {
_gnix_ht_destroy(cm_nic->addr_to_ep_ht);
free(cm_nic->addr_to_ep_ht);
}
if (cm_nic != NULL)
free(cm_nic);
return ret;
}
static int __process_datagram(struct gnix_datagram *dgram,
struct gnix_address peer_address,
gni_post_state_t state)
{
int ret = FI_SUCCESS;
struct gnix_cm_nic *cm_nic = NULL;
uint8_t in_tag = 0, out_tag = 0;
char rcv_buf[GNIX_CM_NIC_MAX_MSG_SIZE];
GNIX_TRACE(FI_LOG_EP_CTRL, "\n");
cm_nic = (struct gnix_cm_nic *)dgram->cache;
if (cm_nic == NULL) {
GNIX_WARN(FI_LOG_EP_CTRL,
"process_datagram, null cache\n");
goto err;
}
if (state != GNI_POST_COMPLETED) {
ret = __process_dgram_w_error(cm_nic,
dgram,
peer_address,
state);
GNIX_WARN(FI_LOG_EP_CTRL,
"process_datagram bad post state %d\n", state);
goto err;
}
__dgram_get_in_tag(dgram, &in_tag);
if ((in_tag != GNIX_CM_NIC_BND_TAG) &&
(in_tag != GNIX_CM_NIC_WC_TAG)) {
GNIX_WARN(FI_LOG_EP_CTRL,
"datagram with unknown in tag %d\n", in_tag);
goto err;
}
__dgram_unpack_out_tag(dgram, &out_tag);
if ((out_tag != GNIX_CM_NIC_BND_TAG) &&
(out_tag != GNIX_CM_NIC_WC_TAG)) {
GNIX_WARN(FI_LOG_EP_CTRL,
"datagram with unknown out tag %d\n", out_tag);
goto err;
}
/*
* if out buf actually has data, call consumer's
* receive callback
*/
if (out_tag == GNIX_CM_NIC_BND_TAG) {
_gnix_dgram_unpack_buf(dgram,
GNIX_DGRAM_OUT_BUF,
rcv_buf,
GNIX_CM_NIC_MAX_MSG_SIZE);
ret = cm_nic->rcv_cb_fn(cm_nic,
rcv_buf,
peer_address);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"cm_nic->rcv_cb_fn returned %s\n",
fi_strerror(-ret));
goto err;
}
ret = _gnix_cm_nic_progress(cm_nic);
if (ret != FI_SUCCESS)
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_cm_nic_progress returned %s\n",
fi_strerror(-ret));
}
/*
* if we are processing a WC datagram, repost, otherwise
* just put back on the freelist.
*/
if (in_tag == GNIX_CM_NIC_WC_TAG) {
dgram->callback_fn = __process_datagram;
dgram->cache = cm_nic;
__dgram_set_tag(dgram, in_tag);
ret = _gnix_dgram_wc_post(dgram);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_dgram_wc_post returned %s\n",
fi_strerror(-ret));
goto err;
}
} else {
ret = _gnix_dgram_free(dgram);
if (ret != FI_SUCCESS)
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_dgram_free returned %s\n",
fi_strerror(-ret));
}
return ret;
err:
if (in_tag == GNIX_CM_NIC_BND_TAG)
_gnix_dgram_free(dgram);
return ret;
}
int _gnix_cm_nic_send(struct gnix_cm_nic *cm_nic,
char *sbuf, size_t len,
struct gnix_address target_addr)
{
int ret = FI_SUCCESS;
struct gnix_datagram *dgram = NULL;
ssize_t __attribute__((unused)) plen;
uint8_t tag;
struct gnix_work_req *work_req;
GNIX_TRACE(FI_LOG_EP_CTRL, "\n");
if ((cm_nic == NULL) || (sbuf == NULL))
return -FI_EINVAL;
if (len > GNI_DATAGRAM_MAXSIZE)
return -FI_ENOSPC;
ret = _gnix_dgram_alloc(cm_nic->dgram_hndl,
GNIX_DGRAM_BND,
&dgram);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_dgram_alloc returned %s\n",
fi_strerror(-ret));
goto exit;
}
dgram->target_addr = target_addr;
dgram->callback_fn = __process_datagram;
dgram->cache = cm_nic;
tag = GNIX_CM_NIC_BND_TAG;
__dgram_set_tag(dgram, tag);
plen = _gnix_dgram_pack_buf(dgram, GNIX_DGRAM_IN_BUF,
sbuf, len);
assert (plen == len);
/* If connecting with the same CM NIC, skip datagram exchange. The
* caller could be holding an endpoint lock, so schedule connection
* completion for later. */
if (GNIX_ADDR_EQUAL(target_addr, cm_nic->my_name.gnix_addr)) {
char tmp_buf[GNIX_CM_NIC_MAX_MSG_SIZE];
/* Pack output buffer with input data. */
_gnix_dgram_unpack_buf(dgram, GNIX_DGRAM_IN_BUF, tmp_buf,
GNIX_CM_NIC_MAX_MSG_SIZE);
_gnix_dgram_pack_buf(dgram, GNIX_DGRAM_OUT_BUF, tmp_buf,
GNIX_CM_NIC_MAX_MSG_SIZE);
work_req = calloc(1, sizeof(*work_req));
if (work_req == NULL) {
_gnix_dgram_free(dgram);
return -FI_ENOMEM;
}
work_req->progress_fn = __gnix_cm_nic_intra_progress_fn;
work_req->data = dgram;
work_req->completer_fn = NULL;
fastlock_acquire(&cm_nic->wq_lock);
dlist_insert_before(&work_req->list, &cm_nic->cm_nic_wq);
fastlock_release(&cm_nic->wq_lock);
GNIX_INFO(FI_LOG_EP_CTRL, "Initiated intra-CM NIC connect\n");
} else {
ret = _gnix_dgram_bnd_post(dgram);
if (ret == -FI_EBUSY) {
ret = -FI_EAGAIN;
_gnix_dgram_free(dgram);
}
}
exit:
return ret;
}
int _gnix_cm_nic_progress(void *arg)
{
struct gnix_cm_nic *cm_nic = (struct gnix_cm_nic *)arg;
int ret = FI_SUCCESS;
int complete;
struct gnix_work_req *p = NULL;
/*
* if we're doing FI_PROGRESS_MANUAL,
* see what's going on inside kgni's datagram
* box...
*/
if (cm_nic->ctrl_progress == FI_PROGRESS_MANUAL) {
++cm_nic->poll_cnt;
if (((cm_nic->poll_cnt % 512) == 0) ||
!dlist_empty(&cm_nic->cm_nic_wq)) {
ret = _gnix_dgram_poll(cm_nic->dgram_hndl,
GNIX_DGRAM_NOBLOCK);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_dgram_poll returned %s\n",
fi_strerror(-ret));
goto err;
}
}
}
/*
* do a quick check if queue doesn't have anything yet,
* don't need this to be atomic
*/
check_again:
if (dlist_empty(&cm_nic->cm_nic_wq))
return ret;
/*
* okay, stuff to do, lock work queue,
* dequeue head, unlock, process work element,
* if it doesn't compete, put back at the tail
* of the queue.
*/
fastlock_acquire(&cm_nic->wq_lock);
p = dlist_first_entry(&cm_nic->cm_nic_wq, struct gnix_work_req,
list);
if (p == NULL) {
fastlock_release(&cm_nic->wq_lock);
return ret;
}
dlist_remove_init(&p->list);
fastlock_release(&cm_nic->wq_lock);
assert(p->progress_fn);
ret = p->progress_fn(p->data, &complete);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"dgram prog fn returned %s\n",
fi_strerror(-ret));
}
if (complete == 1) {
if (p->completer_fn) {
ret = p->completer_fn(p->completer_data);
free(p);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"dgram completer fn returned %s\n",
fi_strerror(-ret));
goto err;
}
} else {
free(p);
}
goto check_again;
} else {
fastlock_acquire(&cm_nic->wq_lock);
dlist_insert_before(&p->list, &cm_nic->cm_nic_wq);
fastlock_release(&cm_nic->wq_lock);
}
err:
return ret;
}
std::string fi_error_to_string(int err_code) {
return fi_strerror(err_code >= 0 ? err_code : -err_code);
}
static int __gnix_vc_hndl_conn_req(struct gnix_cm_nic *cm_nic,
char *msg_buffer,
struct gnix_address src_cm_nic_addr)
{
int ret = FI_SUCCESS;
gni_return_t __attribute__((unused)) status;
struct gnix_fid_ep *ep = NULL;
gnix_ht_key_t *key_ptr;
struct gnix_av_addr_entry entry;
struct gnix_address src_addr, target_addr;
struct gnix_vc *vc = NULL;
struct gnix_vc *vc_try = NULL;
struct gnix_work_req *work_req;
int src_vc_id;
gni_smsg_attr_t src_smsg_attr;
uint64_t src_vc_ptr;
struct wq_hndl_conn_req *data = NULL;
ssize_t __attribute__((unused)) len;
GNIX_TRACE(FI_LOG_EP_CTRL, "\n");
/*
* unpack the message
*/
__gnix_vc_unpack_conn_req(msg_buffer,
&target_addr,
&src_addr,
&src_vc_id,
&src_vc_ptr,
&src_smsg_attr);
GNIX_DEBUG(FI_LOG_EP_CTRL,
"conn req rx: (From Aries addr 0x%x Id %d to Aries 0x%x Id %d src vc 0x%lx )\n",
src_addr.device_addr,
src_addr.cdm_id,
target_addr.device_addr,
target_addr.cdm_id,
src_vc_ptr);
/*
* lookup the ep from the addr_to_ep_ht using the target_addr
* in the datagram
*/
key_ptr = (gnix_ht_key_t *)&target_addr;
ep = (struct gnix_fid_ep *)_gnix_ht_lookup(cm_nic->addr_to_ep_ht,
*key_ptr);
if (ep == NULL) {
GNIX_WARN(FI_LOG_EP_DATA,
"_gnix_ht_lookup addr_to_ep failed\n");
ret = -FI_ENOENT;
goto err;
}
/*
* look to see if there is a VC already for the
* address of the connecting EP.
*/
key_ptr = (gnix_ht_key_t *)&src_addr;
fastlock_acquire(&ep->vc_ht_lock);
vc = (struct gnix_vc *)_gnix_ht_lookup(ep->vc_ht,
*key_ptr);
/*
* if there is no corresponding vc in the hash,
* or there is an entry and its not in connecting state
* go down the conn req ack route.
*/
if ((vc == NULL) ||
(vc->conn_state == GNIX_VC_CONN_NONE)) {
if (vc == NULL) {
entry.gnix_addr = src_addr;
entry.cm_nic_cdm_id = src_cm_nic_addr.cdm_id;
ret = _gnix_vc_alloc(ep,
&entry,
&vc_try);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_vc_alloc returned %s\n",
fi_strerror(-ret));
goto err;
}
vc_try->conn_state = GNIX_VC_CONNECTING;
ret = _gnix_ht_insert(ep->vc_ht,
*key_ptr,
vc_try);
if (likely(ret == FI_SUCCESS)) {
vc = vc_try;
vc->modes |= GNIX_VC_MODE_IN_HT;
} else if (ret == -FI_ENOSPC) {
_gnix_vc_destroy(vc_try);
} else {
GNIX_WARN(FI_LOG_EP_DATA,
"_gnix_ht_insert returned %s\n",
fi_strerror(-ret));
goto err;
}
} else
vc->conn_state = GNIX_VC_CONNECTING;
/*
* prepare a work request to
* initiate an request response
*/
work_req = calloc(1, sizeof(*work_req));
if (work_req == NULL) {
ret = -FI_ENOMEM;
goto err;
}
data = calloc(1, sizeof(struct wq_hndl_conn_req));
if (data == NULL) {
ret = -FI_ENOMEM;
goto err;
}
memcpy(&data->src_smsg_attr,
&src_smsg_attr,
sizeof(src_smsg_attr));
data->vc = vc;
data->src_vc_id = src_vc_id;
data->src_vc_ptr = src_vc_ptr;
work_req->progress_fn = __gnix_vc_conn_ack_prog_fn;
work_req->data = data;
work_req->completer_fn = __gnix_vc_conn_ack_comp_fn;
work_req->completer_data = data;
/*
* add the work request to the tail of the
* cm_nic's work queue, progress the cm_nic.
*/
fastlock_acquire(&cm_nic->wq_lock);
dlist_insert_before(&work_req->list, &cm_nic->cm_nic_wq);
fastlock_release(&cm_nic->wq_lock);
fastlock_release(&ep->vc_ht_lock);
_gnix_vc_schedule(vc);
ret = _gnix_cm_nic_progress(cm_nic);
if (ret != FI_SUCCESS)
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_cm_nic_progress returned %s\n",
fi_strerror(-ret));
} else {
/*
* we can only be in connecting state if we
* reach here. We have all the informatinon,
* and the other side will get the information
* at some point, so go ahead and build SMSG connection.
*/
if (vc->conn_state != GNIX_VC_CONNECTING) {
GNIX_WARN(FI_LOG_EP_CTRL,
"vc %p not in connecting state nor in cm wq\n",
vc, vc->conn_state);
ret = -FI_EINVAL;
goto err;
}
ret = __gnix_vc_smsg_init(vc, src_vc_id,
&src_smsg_attr);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_vc_smsg_init returned %s\n",
fi_strerror(-ret));
goto err;
}
vc->conn_state = GNIX_VC_CONNECTED;
GNIX_DEBUG(FI_LOG_EP_CTRL, "moving vc %p state to connected\n",
vc);
fastlock_release(&ep->vc_ht_lock);
ret = _gnix_vc_schedule(vc);
ret = _gnix_cm_nic_progress(cm_nic);
if (ret != FI_SUCCESS)
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_cm_nic_progress returned %s\n",
fi_strerror(-ret));
}
err:
return ret;
}
int main(int argc, char **argv)
{
int op, ret, ndoms = 3, neps = 3, i, j;
hints = fi_allocinfo();
if (!hints) {
exit(1);
}
while ((op = getopt(argc, argv, "f:p:d:D:n:d:e:h")) != -1) {
switch (op) {
case 'f':
hints->fabric_attr->name = strdup(optarg);
break;
case 'p':
hints->fabric_attr->prov_name = strdup(optarg);
break;
case 'd':
ndoms = atoi(optarg);
break;
case 'e':
neps = atoi(optarg);
break;
case 'h':
default:
printf("usage: %s\n", argv[0]);
printf("\t[-f fabric_name]\n");
printf("\t[-p provider_name]\n");
printf("\t[-d ndomains]\n");
printf("\t[-e neps]\n");
exit(1);
}
}
hints->mode = ~0;
ret = fi_getinfo(FI_VERSION(1, 0), NULL, 0, 0, hints, &fi);
if (ret != 0) {
printf("fi_getinfo %s\n", fi_strerror(-ret));
exit(1);
}
ret = fi_fabric(fi->fabric_attr, &fabric, NULL);
if (ret != 0) {
printf("fi_fabric %s\n", fi_strerror(-ret));
exit(1);
}
domains = (struct fid_domain **)malloc(ndoms * sizeof(struct fid_domain *));
assert(domains);
eps = (struct fid_ep **)malloc(neps * ndoms * sizeof(struct fid_ep *));
assert(eps);
for (i = 0; i < ndoms; i++) {
ret = fi_domain(fabric, fi, &domains[i], NULL);
if (ret != 0) {
printf("fi_domain %s\n", fi_strerror(-ret));
exit(1);
}
for (j = 0; j < neps; j++) {
int idx = (i * neps) + j;
ret = fi_endpoint(domains[i], fi, &eps[idx], NULL);
if (ret != 0) {
printf("[%d:%d] ]fi_endpoint %s\n", i, j, fi_strerror(-ret));
exit(1);
}
}
}
for (i = 0; i < ndoms; i++) {
for (j = 0; j < neps; j++) {
int idx = (i * neps) + j;
ret = fi_close(&eps[idx]->fid);
if (ret != 0) {
printf("Error %d closing ep: %s\n", ret, fi_strerror(-ret));
exit(1);
}
}
ret = fi_close(&domains[i]->fid);
if (ret != 0) {
printf("Error %d closing domain: %s\n", ret, fi_strerror(-ret));
exit(1);
}
}
free(eps);
free(domains);
ret = fi_close(&fabric->fid);
if (ret != 0) {
printf("Error %d closing fabric: %s\n", ret, fi_strerror(-ret));
exit(1);
}
fi_freeinfo(fi);
fi_freeinfo(hints);
return ret;
}
static int __gnix_vc_conn_ack_prog_fn(void *data, int *complete_ptr)
{
int ret = FI_SUCCESS;
int complete = 0;
struct wq_hndl_conn_req *work_req_data;
struct gnix_vc *vc;
struct gnix_mbox *mbox = NULL;
gni_smsg_attr_t smsg_mbox_attr;
struct gnix_fid_ep *ep = NULL;
struct gnix_fid_domain *dom = NULL;
struct gnix_cm_nic *cm_nic = NULL;
char sbuf[GNIX_CM_NIC_MAX_MSG_SIZE] = {0};
GNIX_TRACE(FI_LOG_EP_CTRL, "\n");
work_req_data = (struct wq_hndl_conn_req *)data;
vc = work_req_data->vc;
if (vc == NULL)
return -FI_EINVAL;
ep = vc->ep;
if (ep == NULL)
return -FI_EINVAL;
dom = ep->domain;
if (dom == NULL)
return -FI_EINVAL;
cm_nic = ep->cm_nic;
if (cm_nic == NULL)
return -FI_EINVAL;
fastlock_acquire(&ep->vc_ht_lock);
/*
* we may have already been moved to connecting or
* connected, if so early exit.
*/
if(vc->conn_state == GNIX_VC_CONNECTED) {
complete = 1;
goto exit;
}
/*
* first see if we still need a mailbox
*/
if (vc->smsg_mbox == NULL) {
ret = _gnix_mbox_alloc(ep->nic->mbox_hndl,
&mbox);
if (ret == FI_SUCCESS)
vc->smsg_mbox = mbox;
else
goto exit;
}
mbox = vc->smsg_mbox;
/*
* prep the smsg_mbox_attr
¬ */
smsg_mbox_attr.msg_type = GNI_SMSG_TYPE_MBOX_AUTO_RETRANSMIT;
smsg_mbox_attr.msg_buffer = mbox->base;
smsg_mbox_attr.buff_size = ep->nic->mem_per_mbox;
smsg_mbox_attr.mem_hndl = *mbox->memory_handle;
smsg_mbox_attr.mbox_offset = (uint64_t)mbox->offset;
smsg_mbox_attr.mbox_maxcredit = dom->params.mbox_maxcredit;
smsg_mbox_attr.msg_maxsize = dom->params.mbox_msg_maxsize;
/*
* serialize the resp message in the buffer
*/
__gnix_vc_pack_conn_resp(sbuf,
work_req_data->src_vc_ptr,
(uint64_t)vc,
vc->vc_id,
&smsg_mbox_attr);
/*
* try to send the message, if it succeeds,
* initialize mailbox and move vc to connected
* state.
*/
ret = _gnix_cm_nic_send(cm_nic,
sbuf,
GNIX_CM_NIC_MAX_MSG_SIZE,
vc->peer_cm_nic_addr);
if (ret == FI_SUCCESS) {
ret = __gnix_vc_smsg_init(vc,
work_req_data->src_vc_id,
&work_req_data->src_smsg_attr);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_vc_smsg_init returned %s\n",
fi_strerror(-ret));
goto exit;
}
complete = 1;
vc->conn_state = GNIX_VC_CONNECTED;
GNIX_DEBUG(FI_LOG_EP_CTRL,
"moving vc %p to connected\n",vc);
} else if (ret == -FI_EAGAIN) {
ret = _gnix_vc_schedule(vc);
ret = FI_SUCCESS;
} else
assert(0);
exit:
fastlock_release(&ep->vc_ht_lock);
*complete_ptr = complete;
return ret;
}
static int
usdf_getinfo(uint32_t version, const char *node, const char *service,
uint64_t flags, struct fi_info *hints, struct fi_info **info)
{
struct usdf_usnic_info *dp;
struct usdf_dev_entry *dep;
struct usd_device_attrs *dap;
struct fi_info *fi_first;
struct fi_info *fi_last;
struct addrinfo *ai;
struct sockaddr_in *src;
struct sockaddr_in *dest;
enum fi_ep_type ep_type;
int metric;
int d;
int ret;
USDF_TRACE("\n");
fi_first = NULL;
fi_last = NULL;
ai = NULL;
src = NULL;
dest = NULL;
/*
* Get and cache usNIC device info
*/
if (__usdf_devinfo == NULL) {
ret = usdf_get_devinfo();
if (ret != 0) {
USDF_WARN("failed to usdf_get_devinfo, ret=%d (%s)\n",
ret, fi_strerror(-ret));
if (ret == -FI_ENODEV)
ret = -FI_ENODATA;
goto fail;
}
}
dp = __usdf_devinfo;
if (node != NULL || service != NULL) {
ret = getaddrinfo(node, service, NULL, &ai);
if (ret != 0) {
USDF_DBG("getaddrinfo failed, likely bad node/service specified (%s:%s)\n",
node, service);
ret = -errno;
goto fail;
}
if (flags & FI_SOURCE) {
src = (struct sockaddr_in *)ai->ai_addr;
} else {
dest = (struct sockaddr_in *)ai->ai_addr;
}
}
if (hints != NULL) {
if (dest == NULL && hints->dest_addr != NULL) {
dest = hints->dest_addr;
}
if (src == NULL && hints->src_addr != NULL) {
src = hints->src_addr;
}
}
for (d = 0; d < dp->uu_num_devs; ++d) {
dep = &dp->uu_info[d];
dap = &dep->ue_dattr;
/* skip this device if it has some problem */
if (!dep->ue_dev_ok) {
USDF_DBG("skipping %s/%s\n", dap->uda_devname,
dap->uda_ifname);
continue;
}
/* See if dest is reachable from this device */
if (dest != NULL && dest->sin_addr.s_addr != INADDR_ANY) {
ret = usdf_get_distance(dap,
dest->sin_addr.s_addr, &metric);
if (ret != 0) {
goto fail;
}
if (metric == -1) {
USDF_DBG("dest %s unreachable from %s/%s, skipping\n",
inet_ntoa(dest->sin_addr),
dap->uda_devname, dap->uda_ifname);
continue;
}
}
/* Does this device match requested attributes? */
if (hints != NULL) {
ret = usdf_validate_hints(hints, dap);
if (ret != 0) {
USDF_DBG("hints do not match for %s/%s, skipping\n",
dap->uda_devname, dap->uda_ifname);
continue;
}
ep_type = hints->ep_attr ? hints->ep_attr->type :
FI_EP_UNSPEC;
} else {
ep_type = FI_EP_UNSPEC;
}
if (ep_type == FI_EP_DGRAM || ep_type == FI_EP_UNSPEC) {
ret = usdf_fill_info_dgram(version, hints, src, dest,
dap, &fi_first, &fi_last);
if (ret != 0 && ret != -FI_ENODATA) {
goto fail;
}
}
if (ep_type == FI_EP_MSG || ep_type == FI_EP_UNSPEC) {
ret = usdf_fill_info_msg(hints, src, dest, dap,
&fi_first, &fi_last);
if (ret != 0 && ret != -FI_ENODATA) {
goto fail;
}
}
if (ep_type == FI_EP_RDM || ep_type == FI_EP_UNSPEC) {
ret = usdf_fill_info_rdm(hints, src, dest, dap,
&fi_first, &fi_last);
if (ret != 0 && ret != -FI_ENODATA) {
goto fail;
}
}
}
if (fi_first != NULL) {
*info = fi_first;
ret = 0;
} else {
ret = -FI_ENODATA;
}
fail:
if (ret != 0) {
fi_freeinfo(fi_first);
}
if (ai != NULL) {
freeaddrinfo(ai);
}
if (ret != 0) {
USDF_INFO("returning %d (%s)\n", ret, fi_strerror(-ret));
}
return ret;
}
static int server_connect(void)
{
struct fi_eq_cm_entry entry;
uint32_t event;
struct fi_info *info = NULL;
ssize_t rd;
int ret;
rd = fi_eq_sread(cmeq, &event, &entry, sizeof entry, -1, 0);
if (rd != sizeof entry) {
printf("fi_eq_sread %zd %s\n", rd, fi_strerror((int) -rd));
return (int) rd;
}
if (event != FI_CONNREQ) {
printf("Unexpected CM event %d\n", event);
ret = -FI_EOTHER;
goto err1;
}
info = entry.info;
ret = fi_domain(fab, info, &dom, NULL);
if (ret) {
printf("fi_domain %s\n", fi_strerror(-ret));
goto err1;
}
ret = fi_endpoint(dom, info, &ep, NULL);
if (ret) {
printf("fi_endpoint for req %s\n", fi_strerror(-ret));
goto err1;
}
ret = alloc_ep_res(info);
if (ret)
goto err2;
ret = bind_ep_res();
if (ret)
goto err3;
ret = fi_accept(ep, NULL, 0);
if (ret) {
printf("fi_accept %s\n", fi_strerror(-ret));
goto err3;
}
rd = fi_eq_sread(cmeq, &event, &entry, sizeof entry, -1, 0);
if (rd != sizeof entry) {
printf("fi_eq_sread %zd %s\n", rd, fi_strerror((int) -rd));
goto err3;
}
if (event != FI_COMPLETE || entry.fid != &ep->fid) {
printf("Unexpected CM event %d fid %p (ep %p)\n",
event, entry.fid, ep);
ret = -FI_EOTHER;
goto err3;
}
fi_freeinfo(info);
return 0;
err3:
free_ep_res();
err2:
fi_close(&ep->fid);
err1:
fi_reject(pep, info->connreq, NULL, 0);
fi_freeinfo(info);
return ret;
}
static int
usdf_fabric_open(struct fi_fabric_attr *fattrp, struct fid_fabric **fabric,
void *context)
{
struct fid_fabric *ff;
struct usdf_fabric *fp;
struct usdf_usnic_info *dp;
struct usdf_dev_entry *dep;
struct epoll_event ev;
struct sockaddr_in sin;
int ret;
int d;
USDF_TRACE("\n");
/* Make sure this fabric exists */
dp = __usdf_devinfo;
for (d = 0; d < dp->uu_num_devs; ++d) {
dep = &dp->uu_info[d];
if (dep->ue_dev_ok &&
strcmp(fattrp->name, dep->ue_dattr.uda_devname) == 0) {
break;
}
}
if (d >= dp->uu_num_devs) {
USDF_INFO("device \"%s\" does not exit, returning -FI_ENODEV\n",
fattrp->name);
return -FI_ENODEV;
}
fp = calloc(1, sizeof(*fp));
if (fp == NULL) {
USDF_INFO("unable to allocate memory for fabric\n");
return -FI_ENOMEM;
}
fp->fab_epollfd = -1;
fp->fab_arp_sockfd = -1;
LIST_INIT(&fp->fab_domain_list);
fp->fab_attr.fabric = fab_utof(fp);
fp->fab_attr.name = strdup(fattrp->name);
fp->fab_attr.prov_name = strdup(USDF_PROV_NAME);
fp->fab_attr.prov_version = USDF_PROV_VERSION;
if (fp->fab_attr.name == NULL ||
fp->fab_attr.prov_name == NULL) {
ret = -FI_ENOMEM;
goto fail;
}
fp->fab_fid.fid.fclass = FI_CLASS_FABRIC;
fp->fab_fid.fid.context = context;
fp->fab_fid.fid.ops = &usdf_fi_ops;
fp->fab_fid.ops = &usdf_ops_fabric;
fp->fab_dev_attrs = &dep->ue_dattr;
fp->fab_epollfd = epoll_create(1024);
if (fp->fab_epollfd == -1) {
ret = -errno;
USDF_INFO("unable to allocate epoll fd\n");
goto fail;
}
fp->fab_eventfd = eventfd(0, EFD_NONBLOCK | EFD_SEMAPHORE);
if (fp->fab_eventfd == -1) {
ret = -errno;
USDF_INFO("unable to allocate event fd\n");
goto fail;
}
fp->fab_poll_item.pi_rtn = usdf_fabric_progression_cb;
fp->fab_poll_item.pi_context = fp;
ev.events = EPOLLIN;
ev.data.ptr = &fp->fab_poll_item;
ret = epoll_ctl(fp->fab_epollfd, EPOLL_CTL_ADD, fp->fab_eventfd, &ev);
if (ret == -1) {
ret = -errno;
USDF_INFO("unable to EPOLL_CTL_ADD\n");
goto fail;
}
/* initialize timer subsystem */
ret = usdf_timer_init(fp);
if (ret != 0) {
USDF_INFO("unable to initialize timer\n");
goto fail;
}
ret = pthread_create(&fp->fab_thread, NULL,
usdf_fabric_progression_thread, fp);
if (ret != 0) {
ret = -ret;
USDF_INFO("unable to create progress thread\n");
goto fail;
}
/* create and bind socket for ARP resolution */
memset(&sin, 0, sizeof(sin));
sin.sin_family = AF_INET;
sin.sin_addr.s_addr = fp->fab_dev_attrs->uda_ipaddr_be;
fp->fab_arp_sockfd = socket(AF_INET, SOCK_DGRAM, 0);
if (fp->fab_arp_sockfd == -1) {
USDF_INFO("unable to create socket\n");
goto fail;
}
ret = bind(fp->fab_arp_sockfd, (struct sockaddr *) &sin, sizeof(sin));
if (ret == -1) {
ret = -errno;
goto fail;
}
atomic_initialize(&fp->fab_refcnt, 0);
fattrp->fabric = fab_utof(fp);
fattrp->prov_version = USDF_PROV_VERSION;
*fabric = fab_utof(fp);
USDF_INFO("successfully opened %s/%s\n", fattrp->name,
fp->fab_dev_attrs->uda_ifname);
return 0;
fail:
free(fp->fab_attr.name);
free(fp->fab_attr.prov_name);
ff = fab_utof(fp);
usdf_fabric_close(&ff->fid);
USDF_DBG("returning %d (%s)\n", ret, fi_strerror(-ret));
return ret;
}
static int rxd_ep_bind(struct fid *ep_fid, struct fid *bfid, uint64_t flags)
{
struct rxd_ep *ep;
struct rxd_av *av;
struct util_cq *cq;
struct util_cntr *cntr;
int ret = 0;
ep = container_of(ep_fid, struct rxd_ep, util_ep.ep_fid.fid);
switch (bfid->fclass) {
case FI_CLASS_AV:
av = container_of(bfid, struct rxd_av, util_av.av_fid.fid);
ret = ofi_ep_bind_av(&ep->util_ep, &av->util_av);
if (ret)
return ret;
ret = fi_ep_bind(ep->dg_ep, &av->dg_av->fid, flags);
if (ret)
return ret;
break;
case FI_CLASS_CQ:
cq = container_of(bfid, struct util_cq, cq_fid.fid);
ret = ofi_ep_bind_cq(&ep->util_ep, cq, flags);
if (ret)
return ret;
if (!ep->dg_cq) {
ret = rxd_dg_cq_open(ep, cq->wait ? FI_WAIT_FD : FI_WAIT_NONE);
if (ret)
return ret;
}
if (cq->wait)
ret = rxd_ep_wait_fd_add(ep, cq->wait);
break;
case FI_CLASS_EQ:
break;
case FI_CLASS_CNTR:
cntr = container_of(bfid, struct util_cntr, cntr_fid.fid);
ret = ofi_ep_bind_cntr(&ep->util_ep, cntr, flags);
if (ret)
return ret;
if (!ep->dg_cq) {
ret = rxd_dg_cq_open(ep, cntr->wait ? FI_WAIT_FD : FI_WAIT_NONE);
} else if (!ep->dg_cq_fd && cntr->wait) {
/* Reopen CQ with WAIT fd set */
ret = fi_close(&ep->dg_cq->fid);
if (ret) {
FI_WARN(&rxd_prov, FI_LOG_EP_CTRL,
"Unable to close dg CQ: %s\n",
fi_strerror(-ret));
return ret;
}
ep->dg_cq = NULL;
ret = rxd_dg_cq_open(ep, FI_WAIT_FD);
}
if (ret)
return ret;
if (cntr->wait)
ret = rxd_ep_wait_fd_add(ep, cntr->wait);
break;
default:
FI_WARN(&rxd_prov, FI_LOG_EP_CTRL,
"invalid fid class\n");
ret = -FI_EINVAL;
break;
}
return ret;
}
/*
* this function is intended to be invoked as an argument to pthread_create,
*/
static void *_gnix_dgram_prog_thread_fn(void *the_arg)
{
int ret = FI_SUCCESS, prev_state;
struct gnix_dgram_hndl *the_hndl = (struct gnix_dgram_hndl *)the_arg;
sigset_t sigmask;
GNIX_TRACE(FI_LOG_EP_CTRL, "\n");
/*
* temporarily disable cancelability while we set up
* some stuff
*/
pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &prev_state);
/*
* help out Cray core-spec, say we're not an app thread
* and can be run on core-spec cpus.
*/
ret = _gnix_task_is_not_app();
if (ret)
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_task_is_not_app call returned %d\n", ret);
/*
* block all signals, don't want this thread to catch
* signals that may be for app threads
*/
memset(&sigmask, 0, sizeof(sigset_t));
ret = sigfillset(&sigmask);
if (ret) {
GNIX_WARN(FI_LOG_EP_CTRL,
"sigfillset call returned %d\n", ret);
} else {
ret = pthread_sigmask(SIG_SETMASK,
&sigmask, NULL);
if (ret)
GNIX_WARN(FI_LOG_EP_CTRL,
"pthread_sigmask call returned %d\n", ret);
}
/*
* okay now we're ready to be cancelable.
*/
pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &prev_state);
pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, NULL);
retry:
ret = _gnix_dgram_poll(the_hndl, GNIX_DGRAM_BLOCK);
if ((ret == -FI_ETIMEDOUT) || (ret == FI_SUCCESS))
goto retry;
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_dgram_poll returned %s\n", fi_strerror(-ret));
/*
* TODO: need to be able to enqueue events on to the
* ep associated with the cm_nic.
*/
return NULL;
}
static void test_connect_with_accept_blocking_on_eq_fq_SERVER(void)
{
int ret;
printf("SERVER running\n");
setup_ofi(NULL, NULL, FI_SOURCE);
#if WANT_FDS
// Add the EQ FD to the epoll fd
static struct epoll_event edt;
memset(&edt, 0, sizeof(edt));
edt.events = EPOLLIN;
edt.data.u32 = 2222;
ret = epoll_ctl(epoll_fd, EPOLL_CTL_ADD, fidev.eq_fd, &edt);
if (ret < 0) {
error("server epoll_ctl failed");
}
#endif
// Make a PEP
ret = fi_passive_ep(fidev.fabric, fidev.info, &fidev.pep, NULL);
if (0 != ret) {
error("fi_passive_ep failed");
}
#if WANT_FIXED_PORT
size_t ss = sizeof(sin);
ret = fi_getname(&(fidev.pep->fid), &sin, &ss);
if (0 != ret) {
error("fi_setname failed");
}
sin.sin_port = htons(listen_port);
// Bind the PEP to listen on a specific port
ret = fi_setname(&(fidev.pep->fid), &sin, sizeof(sin));
if (0 != ret) {
error("fi_setname failed");
}
#endif
// Bind the EQ to the PEP
ret = fi_pep_bind(fidev.pep, &fidev.eq->fid, 0);
if (0 != ret) {
error("fi_pep_bind failed");
}
// Listen
ret = fi_listen(fidev.pep);
if (0 != ret) {
error("fi_listen failed");
}
// Get the actual address of this PEP
struct sockaddr_in sinout;
size_t s = sizeof(sinout);
ret = fi_getname(&(fidev.pep->fid), &sinout, &s);
if (0 != ret) {
error("fi_setname failed");
}
sin.sin_family = sinout.sin_family;
sin.sin_addr = sinout.sin_addr;
sin.sin_port = sinout.sin_port;
// Print server addr
printf("SERVER listening on %s\n", addrstr(&sin));
// Send our node (IP addr) and service (port) to the client
snprintf(ofi_node, sizeof(ofi_node) - 1, "%s",
inet_ntoa(sin.sin_addr));
snprintf(ofi_service, sizeof(ofi_service) - 1, "%d",
ntohs(sin.sin_port));
MPI_Send(ofi_node, sizeof(ofi_node) - 1, MPI_CHAR,
1, 101, MPI_COMM_WORLD);
MPI_Send(ofi_service, sizeof(ofi_service) - 1, MPI_CHAR,
1, 102, MPI_COMM_WORLD);
printf("SERVER sent via MPI to client: %s / %s\n", ofi_node, ofi_service);
#if WANT_FDS
// Now wait for the listen to complete
int nevents;
#define NEVENTS 32
struct epoll_event events[NEVENTS];
int timeout = 10000;
while (1) {
printf("SERVER blocking on epoll\n");
nevents = epoll_wait(epoll_fd, events, NEVENTS, timeout);
if (nevents < 0) {
if (errno != EINTR) {
error("server epoll wait failed");
} else {
continue;
}
} else {
printf("SERVER successfully woke up from epoll! %d events\n", nevents);
for (int i = 0; i < nevents; ++i) {
if (events[i].data.u32 != 2222) {
error("server unexpected epoll return type");
}
}
// If we got the expected event, then go read from the EQ
break;
}
}
#endif
// Wait for the FI_CONNREQ event
uint32_t event;
uint8_t *entry_buffer;
size_t expected_len = sizeof(struct fi_eq_cm_entry) +
sizeof(client_data);
entry_buffer = (uint8_t*) calloc(1, expected_len);
if (NULL == entry_buffer) {
error("calloc failed");
}
struct fi_eq_cm_entry *entry = (struct fi_eq_cm_entry*) entry_buffer;
while (1) {
printf("SERVER waiting for FI_CONNREQ\n");
#if WANT_FDS
ret = fi_eq_read(fidev.eq, &event, entry, expected_len, 0);
#else
ret = fi_eq_sread(fidev.eq, &event, entry, expected_len, -1, 0);
#endif
if (-FI_EAVAIL == ret) {
printf("server fi_eq_sread failed because there's something in the error queue\n");
char buffer[2048];
struct fi_eq_err_entry *err_entry = (struct fi_eq_err_entry*) buffer;
ret = fi_eq_readerr(fidev.eq, err_entry, 0);
printf("error code: %d (%s), prov err code: %d (%s)\n", err_entry->err, fi_strerror(err_entry->err), err_entry->prov_errno, fi_strerror(err_entry->prov_errno));
error("sad panda");
} else if (-EAGAIN == ret) {
fprintf(stderr, "SERVER fi_eq_sread fail got -EAGAIN... trying again...\n");
sleep(1);
continue;
} else if (ret < 0) {
fprintf(stderr, "SERVER fi_eq_sread fail: %s (FI_EAVAIL = %d, -ret = %d)\n", fi_strerror(-ret), FI_EAVAIL, -ret);
error("SERVER fi_eq_sread failed for some random reason");
} else if (event != FI_CONNREQ) {
error("SERVER got some unexpected event");
} else if (ret != expected_len) {
error("SERVER got wrong length back from fi_eq_sread");
}
uint32_t *d = (uint32_t*) entry->data;
for (int i = 0; i < (sizeof(client_data) / sizeof(uint32_t)); ++i) {
if (d[i] != client_data[i]) {
printf("SERVER got wrong CM client data: d[%d]=%d, should be %d\n",
i, d[i], client_data[i]);
}
}
printf("SERVER got FI_CONNREQ, correct size, and correct data -- yay!\n");
break;
}
// Silly logistics: setup_ofi_active adds the fd to the epoll set.
// But we already added it. So for simplicity, just remove it
// here so that setup_ofi_active() can re-add it.
#if WANT_FDS
// Remove the EQ FD from the epoll fd
ret = epoll_ctl(epoll_fd, EPOLL_CTL_DEL, fidev.eq_fd, &edt);
if (ret < 0) {
error("server epoll_ctl DEL failed");
}
#endif
// Make an active endpoint
setup_ofi_active(entry->info, &ficonn.ep);
// Accept the incoming connection
ret = fi_accept(ficonn.ep, (void*) server_data, sizeof(server_data));
if (ret != 0) {
printf("fi_accept: ret=%d, %s\n", ret, fi_strerror(-ret));
error("SERVER fi_accept failed\n");
}
// Need to read and get a FI_CONNECTED event
while (1) {
printf("SERVER waiting for FI_CONNECTED\n");
#if WANT_FDS
ret = fi_eq_read(fidev.eq, &event, entry, expected_len, 0);
#else
ret = fi_eq_sread(fidev.eq, &event, entry, expected_len, -1, 0);
#endif
if (-FI_EAVAIL == ret) {
printf("server fi_eq_sread failed because there's something in the error queue\n");
char buffer[2048];
struct fi_eq_err_entry *err_entry = (struct fi_eq_err_entry*) buffer;
ret = fi_eq_readerr(fidev.eq, err_entry, 0);
printf("error code: %d (%s), prov err code: %d (%s)\n", err_entry->err, fi_strerror(err_entry->err), err_entry->prov_errno, fi_strerror(err_entry->prov_errno));
error("sad panda");
} else if (-EAGAIN == ret) {
fprintf(stderr, "SERVER fi_eq_sread fail got -EAGAIN... trying again...\n");
sleep(1);
continue;
} else if (ret < 0) {
fprintf(stderr, "SERVER fi_eq_sread fail: %s (FI_EAVAIL = %d, -ret = %d)\n", fi_strerror(-ret), FI_EAVAIL, -ret);
error("SERVER fi_eq_sread failed for some random reason");
} else if (event != FI_CONNECTED) {
error("SERVER got some unexpected event");
}
printf("SERVER got FI_CONNECTED -- yay!\n");
break;
}
// Post a recv buffer for the client to send
int msg[4] = { 0 };
int len = sizeof(msg);
printf("SERVER receiving len of %d\n", len);
struct fid_mr no_mr;
struct fid_mr *mr;
void *recv_context = (void*) 0x17;
#if 0
fi_mr_reg(fidev.domain, msg, len, FI_SEND | FI_RECV,
0, (uint64_t)(uintptr_t) msg, 0, &mr, NULL);
#else
// Try using no mr, like fi_msg_pingpong...
memset(&no_mr, 0, sizeof(no_mr));
mr = &no_mr;
#endif
ret = fi_recv(ficonn.ep, msg, len,
fi_mr_desc(mr), 0, recv_context);
if (ret < 0) {
printf("fi_recv failed! %d, %s\n", ret, fi_strerror(-ret));
MPI_Abort(MPI_COMM_WORLD, 37);
}
sleep(1);
printf("SERVER posted receive -- waiting for client to send\n");
MPI_Barrier(MPI_COMM_WORLD);
// Wait for receive completion
struct fi_cq_entry cqe;
while (1) {
ret = fi_cq_sread(ficonn.cq, &cqe, 1, 0, -1);
if (cqe.op_context == recv_context) {
printf("SERVER receive completed\n");
break;
} else {
printf("SERVER got some other completion... continuing\n");
}
}
printf("SERVER finished -- waiting for client before teardown\n");
MPI_Barrier(MPI_COMM_WORLD);
printf("SERVER tearing down\n");
fi_close(&(mr->fid));
teardown_ofi();
}
/* Destroy an unconnected VC. More Support is needed to shutdown and destroy
* an active VC. */
int _gnix_vc_destroy(struct gnix_vc *vc)
{
int ret = FI_SUCCESS;
struct gnix_nic *nic = NULL;
gni_return_t status;
GNIX_TRACE(FI_LOG_EP_CTRL, "\n");
if (vc->ep == NULL) {
GNIX_WARN(FI_LOG_EP_CTRL, "ep null\n");
return -FI_EINVAL;
}
nic = vc->ep->nic;
if (nic == NULL) {
GNIX_WARN(FI_LOG_EP_CTRL, "ep nic null for vc %p\n", vc);
return -FI_EINVAL;
}
/*
* move vc state to terminating
*/
vc->conn_state = GNIX_VC_CONN_TERMINATING;
/*
* try to unbind the gni_ep if non-NULL.
* If there are SMSG or PostFMA/RDMA outstanding
* wait here for them to complete
*/
if (vc->gni_ep != NULL) {
while (status == GNI_RC_NOT_DONE) {
fastlock_acquire(&nic->lock);
status = GNI_EpUnbind(vc->gni_ep);
fastlock_release(&nic->lock);
if ((status != GNI_RC_NOT_DONE) &&
(status != GNI_RC_SUCCESS)) {
GNIX_WARN(FI_LOG_EP_CTRL,
"GNI_EpUnBind returned %s\n",
gni_err_str[status]);
break;
}
if (status == GNI_RC_NOT_DONE)
_gnix_nic_progress(nic);
}
fastlock_acquire(&nic->lock);
status = GNI_EpDestroy(vc->gni_ep);
fastlock_release(&nic->lock);
if (status != GNI_RC_SUCCESS)
GNIX_WARN(FI_LOG_EP_CTRL,
"GNI_EpDestroy returned %s\n",
gni_err_str[status]);
}
/*
* if the vc is in a nic's work queue, remove it
*/
__gnix_vc_cancel(vc);
/*
* We may eventually want to check the state of the VC, if we
* implement true VC shutdown.
if ((vc->conn_state != GNIX_VC_CONN_NONE)
&& (vc->conn_state != GNIX_VC_CONN_TERMINATED)) {
GNIX_WARN(FI_LOG_EP_CTRL,
"vc conn state %d\n",
vc->conn_state);
GNIX_WARN(FI_LOG_EP_CTRL, "vc conn state error\n");
return -FI_EBUSY;
}
*/
/*
* if send_q not empty, return -FI_EBUSY
* Note for FI_EP_MSG type eps, this behavior
* may not be correct for handling fi_shutdown.
*/
if (!slist_empty(&vc->tx_queue)) {
GNIX_WARN(FI_LOG_EP_CTRL, "vc sendqueue not empty\n");
return -FI_EBUSY;
}
fastlock_destroy(&vc->tx_queue_lock);
if (vc->smsg_mbox != NULL) {
ret = _gnix_mbox_free(vc->smsg_mbox);
if (ret != FI_SUCCESS)
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_mbox_free returned %s\n",
fi_strerror(-ret));
vc->smsg_mbox = NULL;
}
if (vc->dgram != NULL) {
ret = _gnix_dgram_free(vc->dgram);
if (ret != FI_SUCCESS)
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_dgram_free returned %s\n",
fi_strerror(-ret));
vc->dgram = NULL;
}
ret = _gnix_nic_free_rem_id(nic, vc->vc_id);
if (ret != FI_SUCCESS)
GNIX_WARN(FI_LOG_EP_CTRL,
"__gnix_vc_free_id returned %s\n",
fi_strerror(-ret));
_gnix_free_bitmap(&vc->flags);
free(vc);
return ret;
}
int ADD_SUFFIX(MPID_nem_ofi_iprobe_impl)(struct MPIDI_VC *vc,
int source,
int tag,
MPIR_Comm * comm,
int context_offset,
int *flag, MPI_Status * status, MPIR_Request ** rreq_ptr)
{
int ret, mpi_errno = MPI_SUCCESS;
fi_addr_t remote_proc = 0;
uint64_t match_bits, mask_bits;
size_t len;
MPIR_Request rreq_s, *rreq;
BEGIN_FUNC(FCNAME);
if (rreq_ptr) {
MPIDI_CH3I_NM_OFI_RC(MPID_nem_ofi_create_req(&rreq, 1));
rreq->kind = MPIR_REQUEST_KIND__RECV;
*rreq_ptr = rreq;
rreq->comm = comm;
rreq->dev.match.parts.rank = source;
rreq->dev.match.parts.tag = tag;
rreq->dev.match.parts.context_id = comm->context_id;
MPIR_Comm_add_ref(comm);
}
else {
rreq = &rreq_s;
rreq->dev.OnDataAvail = NULL;
}
REQ_OFI(rreq)->pack_buffer = NULL;
REQ_OFI(rreq)->event_callback = ADD_SUFFIX(peek_callback);
REQ_OFI(rreq)->match_state = PEEK_INIT;
OFI_ADDR_INIT(source, vc, remote_proc);
#if API_SET == API_SET_1
match_bits = init_recvtag(&mask_bits, comm->context_id + context_offset, source, tag);
#elif API_SET == API_SET_2
match_bits = init_recvtag_2(&mask_bits, comm->context_id + context_offset, tag);
#endif
/* ------------------------------------------------------------------------- */
/* fi_recvmsg with FI_PEEK: */
/* Initiate a search for a match in the hardware or software queue. */
/* The search can complete immediately with -ENOMSG. */
/* I successful, libfabric will enqueue a context entry into the completion */
/* queue to make the search nonblocking. This code will poll until the */
/* entry is enqueued. */
/* ------------------------------------------------------------------------- */
msg_tagged_t msg;
uint64_t msgflags = FI_PEEK;
msg.msg_iov = NULL;
msg.desc = NULL;
msg.iov_count = 0;
msg.addr = remote_proc;
msg.tag = match_bits;
msg.ignore = mask_bits;
msg.context = (void *) &(REQ_OFI(rreq)->ofi_context);
msg.data = 0;
if(*flag == CLAIM_PEEK)
msgflags|=FI_CLAIM;
ret = fi_trecvmsg(gl_data.endpoint,&msg,msgflags);
if(ret == -ENOMSG) {
if (rreq_ptr) {
MPIR_Request_free(rreq);
*rreq_ptr = NULL;
*flag = 0;
}
MPID_nem_ofi_poll(MPID_NONBLOCKING_POLL);
goto fn_exit;
}
MPIR_ERR_CHKANDJUMP4((ret < 0), mpi_errno, MPI_ERR_OTHER,
"**ofi_peek", "**ofi_peek %s %d %s %s",
__SHORT_FILE__, __LINE__, FCNAME, fi_strerror(-ret));
while (PEEK_INIT == REQ_OFI(rreq)->match_state)
MPID_nem_ofi_poll(MPID_BLOCKING_POLL);
if (PEEK_NOT_FOUND == REQ_OFI(rreq)->match_state) {
if (rreq_ptr) {
MPIR_Request_free(rreq);
*rreq_ptr = NULL;
*flag = 0;
}
MPID_nem_ofi_poll(MPID_NONBLOCKING_POLL);
goto fn_exit;
}
if (status != MPI_STATUS_IGNORE)
*status = rreq->status;
if (rreq_ptr)
MPIR_Request_add_ref(rreq);
*flag = 1;
END_FUNC_RC(FCNAME);
}
/*
* connect to self, since we use a lock here
* the only case we need to deal with is one
* vc connect request with the peer vc not yet
* being set up
*/
static int __gnix_vc_connect_to_same_cm_nic(struct gnix_vc *vc)
{
int ret = FI_SUCCESS;
struct gnix_fid_domain *dom = NULL;
struct gnix_fid_ep *ep = NULL;
struct gnix_fid_ep *ep_peer = NULL;
struct gnix_cm_nic *cm_nic = NULL;
struct gnix_mbox *mbox = NULL, *mbox_peer = NULL;
struct gnix_vc *vc_peer;
gni_smsg_attr_t smsg_mbox_attr;
gni_smsg_attr_t smsg_mbox_attr_peer;
gnix_ht_key_t *key_ptr;
struct gnix_av_addr_entry entry;
GNIX_TRACE(FI_LOG_EP_CTRL, "\n");
ep = vc->ep;
if (ep == NULL)
return -FI_EINVAL;
cm_nic = ep->cm_nic;
if (cm_nic == NULL) {
ret = -FI_EINVAL;
goto exit;
}
dom = ep->domain;
if (dom == NULL) {
ret = -FI_EINVAL;
goto exit;
}
fastlock_acquire(&ep->vc_ht_lock);
if ((vc->conn_state == GNIX_VC_CONNECTING) ||
(vc->conn_state == GNIX_VC_CONNECTED)) {
fastlock_release(&ep->vc_ht_lock);
return FI_SUCCESS;
} else
vc->conn_state = GNIX_VC_CONNECTING;
fastlock_release(&ep->vc_ht_lock);
GNIX_DEBUG(FI_LOG_EP_CTRL, "moving vc %p state to connecting\n", vc);
if (vc->smsg_mbox == NULL) {
ret = _gnix_mbox_alloc(vc->ep->nic->mbox_hndl,
&mbox);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_DATA,
"_gnix_mbox_alloc returned %s\n",
fi_strerror(-ret));
goto exit;
}
vc->smsg_mbox = mbox;
} else
mbox = vc->smsg_mbox;
smsg_mbox_attr.msg_type = GNI_SMSG_TYPE_MBOX_AUTO_RETRANSMIT;
smsg_mbox_attr.msg_buffer = mbox->base;
smsg_mbox_attr.buff_size = vc->ep->nic->mem_per_mbox;
smsg_mbox_attr.mem_hndl = *mbox->memory_handle;
smsg_mbox_attr.mbox_offset = (uint64_t)mbox->offset;
smsg_mbox_attr.mbox_maxcredit = dom->params.mbox_maxcredit;
smsg_mbox_attr.msg_maxsize = dom->params.mbox_msg_maxsize;
key_ptr = (gnix_ht_key_t *)&vc->peer_addr;
ep_peer = (struct gnix_fid_ep *)_gnix_ht_lookup(cm_nic->addr_to_ep_ht,
*key_ptr);
if (ep_peer == NULL) {
GNIX_WARN(FI_LOG_EP_DATA,
"_gnix_ht_lookup addr_to_ep failed\n");
ret = -FI_ENOENT;
goto exit;
}
key_ptr = (gnix_ht_key_t *)&ep->my_name.gnix_addr;
fastlock_acquire(&ep_peer->vc_ht_lock);
vc_peer = (struct gnix_vc *)_gnix_ht_lookup(ep_peer->vc_ht,
*key_ptr);
/*
* handle the special case of connecting to self
*/
if (vc_peer == vc) {
ret = __gnix_vc_smsg_init(vc, vc->vc_id, &smsg_mbox_attr);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_DATA,
"_gnix_vc_smsg_init returned %s\n",
fi_strerror(-ret));
goto exit_w_lock;
}
vc->conn_state = GNIX_VC_CONNECTED;
GNIX_DEBUG(FI_LOG_EP_CTRL, "moving vc %p state to connected\n",
vc);
goto exit_w_lock;
}
if ((vc_peer != NULL) &&
(vc_peer->conn_state != GNIX_VC_CONN_NONE)) {
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_vc_connect self, vc_peer in inconsistent state\n");
ret = -FI_ENOSPC;
goto exit_w_lock;
}
if (vc_peer == NULL) {
entry.gnix_addr = ep->my_name.gnix_addr;
entry.cm_nic_cdm_id = ep->my_name.cm_nic_cdm_id;
ret = _gnix_vc_alloc(ep_peer,
&entry,
&vc_peer);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_vc_alloc returned %s\n",
fi_strerror(-ret));
goto exit_w_lock;
}
ret = _gnix_ht_insert(ep_peer->vc_ht,
*key_ptr,
vc_peer);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_ht_insert returned %s\n",
fi_strerror(-ret));
goto exit_w_lock;
}
vc_peer->modes |= GNIX_VC_MODE_IN_HT;
}
vc_peer->conn_state = GNIX_VC_CONNECTING;
GNIX_DEBUG(FI_LOG_EP_CTRL, "moving vc %p state to connecting\n",
vc_peer);
if (vc_peer->smsg_mbox == NULL) {
ret = _gnix_mbox_alloc(vc_peer->ep->nic->mbox_hndl,
&mbox_peer);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_DATA,
"_gnix_mbox_alloc returned %s\n",
fi_strerror(-ret));
goto exit_w_lock;
}
vc_peer->smsg_mbox = mbox_peer;
} else
mbox_peer = vc_peer->smsg_mbox;
smsg_mbox_attr_peer.msg_type = GNI_SMSG_TYPE_MBOX_AUTO_RETRANSMIT;
smsg_mbox_attr_peer.msg_buffer = mbox_peer->base;
smsg_mbox_attr_peer.buff_size = vc_peer->ep->nic->mem_per_mbox;
smsg_mbox_attr_peer.mem_hndl = *mbox_peer->memory_handle;
smsg_mbox_attr_peer.mbox_offset = (uint64_t)mbox_peer->offset;
smsg_mbox_attr_peer.mbox_maxcredit = dom->params.mbox_maxcredit;
smsg_mbox_attr_peer.msg_maxsize = dom->params.mbox_msg_maxsize;
ret = __gnix_vc_smsg_init(vc, vc_peer->vc_id, &smsg_mbox_attr_peer);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_DATA,
"_gnix_vc_smsg_init returned %s\n",
fi_strerror(-ret));
goto exit_w_lock;
}
ret = __gnix_vc_smsg_init(vc_peer, vc->vc_id, &smsg_mbox_attr);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_DATA,
"_gnix_vc_smsg_init returned %s\n",
fi_strerror(-ret));
goto exit_w_lock;
}
vc->conn_state = GNIX_VC_CONNECTED;
GNIX_DEBUG(FI_LOG_EP_CTRL, "moving vc %p state to connected\n",
vc);
vc_peer->conn_state = GNIX_VC_CONNECTED;
GNIX_DEBUG(FI_LOG_EP_CTRL, "moving vc %p state to connected\n",
vc_peer);
exit_w_lock:
fastlock_release(&ep_peer->vc_ht_lock);
exit:
return ret;
}
static int run(void)
{
char *node, *service;
uint64_t flags;
int ret;
size_t i;
ret = ft_read_addr_opts(&node, &service, hints, &flags, &opts);
if (ret)
return ret;
ret = opts.dst_addr ? client_setup() : server_setup();
if (ret) {
fprintf(stderr, "error: %s\n", fi_strerror(-ret));
return ret;
}
/* Leave extra space for invalid size test */
cm_data = calloc(1, cm_data_size + 1);
if (!cm_data)
return -FI_ENOMEM;
entry = calloc(1, sizeof(*entry) + cm_data_size);
if (!entry)
return -FI_ENOMEM;
if (opts.dst_addr) {
ret = ft_sock_connect(opts.dst_addr, sock_service);
if (ret)
goto err2;
} else {
ret = ft_sock_listen(sock_service);
if (ret)
goto err2;
ret = ft_sock_accept();
if (ret)
goto err1;
}
for (i = 1; i <= cm_data_size; i <<= 1) {
printf("trying with data size: %zu\n", i);
if (opts.dst_addr)
ret = client(i);
else
ret = server(i);
if (ret)
goto err1;
ret = ft_sock_sync(0);
if (ret)
goto err1;
}
/* Despite server not being setup to handle this, the client should fail
* with -FI_EINVAL since this exceeds its max data size.
*/
if (opts.dst_addr) {
printf("trying with data size exceeding maximum: %zu\n",
cm_data_size + 1);
/* Don't call client since it produces an error message. */
ret = client_connect(cm_data_size + 1);
if (ret != -FI_EINVAL) {
FT_ERR("expected -FI_EINVAL, got: [%d]:%s\n", ret,
fi_strerror(-ret));
} else {
ret = FI_SUCCESS;
}
}
err1:
ft_sock_shutdown(sock);
err2:
free(entry);
return ret;
}
static int __gnix_vc_hndl_conn_resp(struct gnix_cm_nic *cm_nic,
char *msg_buffer,
struct gnix_address src_cm_nic_addr)
{
int ret = FI_SUCCESS;
int peer_id;
struct gnix_vc *vc = NULL;
uint64_t peer_vc_addr;
struct gnix_fid_ep *ep;
gni_smsg_attr_t peer_smsg_attr;
GNIX_TRACE(FI_LOG_EP_CTRL, "\n");
/*
* unpack the message
*/
__gnix_vc_unpack_resp(msg_buffer,
(uint64_t *)&vc,
&peer_vc_addr,
&peer_id,
&peer_smsg_attr);
GNIX_DEBUG(FI_LOG_EP_CTRL,
"resp rx: (From Aries 0x%x Id %d src vc %p peer vc addr 0x%lx)\n",
src_cm_nic_addr.device_addr,
src_cm_nic_addr.cdm_id,
vc,
peer_vc_addr);
ep = vc->ep;
assert(ep != NULL);
fastlock_acquire(&ep->vc_ht_lock);
/*
* at this point vc should be in connecting state
*/
if (vc->conn_state != GNIX_VC_CONNECTING) {
GNIX_WARN(FI_LOG_EP_CTRL,
"vc %p not in connecting state, rather %d\n",
vc, vc->conn_state);
ret = -FI_EINVAL;
goto err;
}
/*
* build the SMSG connection
*/
ret = __gnix_vc_smsg_init(vc, peer_id, &peer_smsg_attr);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"__gnix_vc_smsg_init returned %s\n",
fi_strerror(-ret));
goto err;
}
/*
* transition the VC to connected
* put in to the nic's work queue for
* further processing
*/
vc->conn_state = GNIX_VC_CONNECTED;
GNIX_DEBUG(FI_LOG_EP_CTRL,
" moving vc %p to state connected\n",vc);
fastlock_release(&ep->vc_ht_lock);
ret = _gnix_vc_schedule(vc);
return ret;
err:
vc->conn_state = GNIX_VC_CONN_ERROR;
fastlock_release(&ep->vc_ht_lock);
return ret;
}
int create_connection(void)
{
struct fi_info *prov;
struct fi_eq_cm_entry entry;
struct fi_info *info;
ssize_t n;
uint32_t event;
int ret = -1;
print_trace("in\n");
memset(&ctx, 0, sizeof(ctx));
if (match_provider(&prov))
goto err1;
if (client_connect(prov, &ctx))
goto err2;
dprint(DEBUG_CONNECT, "Waiting for Server to connect\n");
n = fi_eq_sread(ctx.eq, &event, &entry, sizeof(entry), CTIMEOUT, 0);
if (n < sizeof(entry)) {
struct fi_eq_err_entry eqe;
int rc;
print_err("fi_eq_sread '%s'(%d)\n", fi_strerror(n), (int) n);
rc = fi_eq_readerr(ctx.eq, &eqe, 0);
if (rc)
print_err("fi_eq_readerr() returns %d '%s'\n",
rc, fi_strerror(rc));
else {
char buf[64];
print_err("fi_eq_readerr() prov_err '%s'(%d)\n",
fi_eq_strerror(ctx.eq, eqe.prov_errno,
eqe.err_data, buf, sizeof(buf)),
eqe.prov_errno);
print_err("fi_eq_readerr() err '%s'(%d)\n",
fi_strerror(eqe.err), eqe.err);
}
return (int) n;
}
if (event != FI_CONNECTED) {
print_err("unexpected event %d\n", event);
return -FI_EOTHER;
}
/* same context specified in fi_endpoint()? */
if (entry.fid->context != CONTEXT) {
print_err("entry.fid->context %lx != %lx\n",
(ulong)entry.fid->context, (ulong)CONTEXT);
}
info = entry.info;
dprint(DEBUG_CONNECT, "*** Client Connected\n");
dprint(DEBUG_CONNECT, "Client private data(len %ld): '%s'\n",
(n - sizeof(entry)), entry.data);
return 0;
err2:
client_disconnect(&ctx);
fi_freeinfo(prov);
err1:
return ret;
}
static int rxd_av_insert(struct fid_av *av_fid, const void *addr, size_t count,
fi_addr_t *fi_addr, uint64_t flags, void *context)
{
struct rxd_av *av;
int i = 0, index, ret = 0, success_cnt = 0, lookup = 1;
uint64_t dg_fiaddr;
av = container_of(av_fid, struct rxd_av, util_av.av_fid);
fastlock_acquire(&av->util_av.lock);
if (!av->dg_addrlen) {
ret = rxd_av_set_addrlen(av, addr);
if (ret)
goto out;
/* Skip lookups if this is the first insertion call. */
lookup = 0;
}
for (; i < count; i++, addr = (uint8_t *) addr + av->dg_addrlen) {
ret = lookup ? rxd_av_dg_reverse_lookup(av, i, addr, &dg_fiaddr) :
-FI_ENODATA;
if (ret) {
ret = fi_av_insert(av->dg_av, addr, 1, &dg_fiaddr,
flags, context);
if (ret != 1)
break;
}
ret = ofi_av_insert_addr(&av->util_av, &dg_fiaddr, dg_fiaddr, &index);
if (ret)
break;
success_cnt++;
if (fi_addr)
fi_addr[i] = index;
}
if (ret) {
FI_WARN(&rxd_prov, FI_LOG_AV,
"failed to insert address %d: %d (%s)\n",
i, -ret, fi_strerror(-ret));
if (av->util_av.eq)
ofi_av_write_event(&av->util_av, i, -ret, context);
if (fi_addr)
fi_addr[i] = FI_ADDR_NOTAVAIL;
i++;
}
out:
av->dg_av_used += success_cnt;
fastlock_release(&av->util_av.lock);
for (; i < count; i++) {
if (av->util_av.eq)
ofi_av_write_event(&av->util_av, i, FI_ECANCELED, context);
if (fi_addr)
fi_addr[i] = FI_ADDR_NOTAVAIL;
}
if (av->util_av.eq) {
ofi_av_write_event(&av->util_av, success_cnt, 0, context);
return 0;
}
return success_cnt;
}
