static inline void __api_cq_setup(uint32_t version, int mr_mode)
{
int ret, i, j;
struct fi_av_attr attr;
size_t addrlen = 0;
for (i = 0; i < NUMEPS; i++) {
hints[i] = fi_allocinfo();
cr_assert(hints[i], "fi_allocinfo");
hints[i]->domain_attr->cq_data_size = NUMEPS * 2;
hints[i]->domain_attr->data_progress = FI_PROGRESS_AUTO;
hints[i]->domain_attr->mr_mode = mr_mode;
hints[i]->mode = mode_bits;
hints[i]->fabric_attr->prov_name = strdup("gni");
}
/* Get info about fabric services with the provided hints */
for (i = 0; i < NUMEPS; i++) {
ret = fi_getinfo(version, NULL, 0, 0, hints[i],
&fi[i]);
cr_assert(!ret, "fi_getinfo");
}
memset(&attr, 0, sizeof(attr));
attr.type = FI_AV_MAP;
attr.count = NUMEPS;
cq_attr.format = FI_CQ_FORMAT_TAGGED;
cq_attr.size = 1024;
cq_attr.wait_obj = 0;
/* 3x BUF_SZ for multi recv testing */
target_base = malloc(GNIT_ALIGN_LEN(BUF_SZ * 3));
assert(target_base);
target = GNIT_ALIGN_BUFFER(char *, target_base);
source_base = malloc(GNIT_ALIGN_LEN(BUF_SZ));
assert(source_base);
source = GNIT_ALIGN_BUFFER(char *, source_base);
uc_target = malloc(BUF_SZ);
assert(uc_target);
uc_source = malloc(BUF_SZ);
assert(uc_source);
ret = fi_fabric(fi[0]->fabric_attr, &fab, NULL);
cr_assert(!ret, "fi_fabric");
for (i = 0; i < NUMEPS; i++) {
ret = fi_domain(fab, fi[i], dom + i, NULL);
cr_assert(!ret, "fi_domain");
ret = fi_open_ops(&dom[i]->fid, FI_GNI_DOMAIN_OPS_1,
0, (void **) (gni_domain_ops + i), NULL);
ret = fi_av_open(dom[i], &attr, av + i, NULL);
cr_assert(!ret, "fi_av_open");
ret = fi_endpoint(dom[i], fi[i], ep + i, NULL);
cr_assert(!ret, "fi_endpoint");
ret = fi_cq_open(dom[i], &cq_attr, msg_cq + i, 0);
cr_assert(!ret, "fi_cq_open");
ret = fi_getname(&ep[i]->fid, NULL, &addrlen);
cr_assert(addrlen > 0);
ep_name[i] = malloc(addrlen);
cr_assert(ep_name[i] != NULL);
ret = fi_getname(&ep[i]->fid, ep_name[i], &addrlen);
cr_assert(ret == FI_SUCCESS);
}
for (i = 0; i < NUMEPS; i++) {
/* Insert all gni addresses into each av */
for (j = 0; j < NUMEPS; j++) {
ret = fi_av_insert(av[i], ep_name[j], 1, &gni_addr[j],
0, NULL);
cr_assert(ret == 1);
}
ret = fi_ep_bind(ep[i], &av[i]->fid, 0);
cr_assert(!ret, "fi_ep_bind");
}
for (i = 0; i < NUMEPS; i++) {
int target_requested_key =
USING_SCALABLE(fi[i]) ? (i * 2) : 0;
int source_requested_key =
USING_SCALABLE(fi[i]) ? (i * 2) + 1 : 0;
ret = fi_mr_reg(dom[i],
target,
3 * BUF_SZ,
FI_REMOTE_WRITE,
0,
target_requested_key,
0,
rem_mr + i,
&target);
cr_assert_eq(ret, 0);
ret = fi_mr_reg(dom[i],
source,
BUF_SZ,
FI_REMOTE_WRITE,
0,
source_requested_key,
0,
loc_mr + i,
&source);
cr_assert_eq(ret, 0);
if (USING_SCALABLE(fi[i])) {
MR_ENABLE(rem_mr[i], target, 3 * BUF_SZ);
MR_ENABLE(loc_mr[i], source, BUF_SZ);
}
mr_key[i] = fi_mr_key(rem_mr[i]);
}
}
/*
* rpmem_fip_init_cq -- (internal) initialize completion queue(s)
*/
static int
rpmem_fip_init_cq(struct rpmem_fip *fip)
{
int ret;
struct fi_cq_attr cq_attr = {
.size = fip->cq_size,
.flags = 0,
.format = FI_CQ_FORMAT_MSG,
.wait_obj = FI_WAIT_UNSPEC,
.signaling_vector = 0,
.wait_cond = FI_CQ_COND_NONE,
.wait_set = NULL,
};
ret = fi_cq_open(fip->domain, &cq_attr, &fip->cq, NULL);
if (ret) {
RPMEM_FI_ERR(ret, "opening completion queue");
goto err_cq_open;
}
return 0;
err_cq_open:
return -1;
}
/*
* rpmem_fip_fini_cq -- (internal) deinitialize completion queue(s)
*/
static int
rpmem_fip_fini_cq(struct rpmem_fip *fip)
{
return RPMEM_FI_CLOSE(fip->cq, "closing completion queue");
}
/*
* rpmem_fip_init_ep -- (internal) initialize endpoint
*/
static int
rpmem_fip_init_ep(struct rpmem_fip *fip)
{
int ret;
/* create an endpoint */
ret = fi_endpoint(fip->domain, fip->fi, &fip->ep, NULL);
if (ret) {
RPMEM_FI_ERR(ret, "allocating endpoint");
goto err_endpoint;
}
/*
* Bind an event queue to an endpoint to get
* connection-related events for the endpoint.
*/
ret = fi_ep_bind(fip->ep, &fip->eq->fid, 0);
if (ret) {
RPMEM_FI_ERR(ret, "binding event queue to endpoint");
goto err_ep_bind_eq;
}
/*
* Bind a completion queue to an endpoint to get completion
* events of specified inbound/outbound operations.
*
* FI_SELECTIVE_COMPLETION means all inbound/outbound operations
* must explicitly specify if the completion event should be
* generated or not using FI_COMPLETION flag.
*
* The completion events received are highly related to the
* persistency method used and are configured in lanes
* initialization specified for persistency method utilized.
*/
ret = fi_ep_bind(fip->ep, &fip->cq->fid,
FI_RECV | FI_TRANSMIT | FI_SELECTIVE_COMPLETION);
if (ret) {
RPMEM_FI_ERR(ret, "binding completion queue to endpoint");
goto err_ep_bind_cq;
}
/*
* Enable endpoint so it is possible to post inbound/outbound
* operations if required.
*/
ret = fi_enable(fip->ep);
if (ret) {
RPMEM_FI_ERR(ret, "activating endpoint");
goto err_fi_enable;
}
return 0;
err_fi_enable:
err_ep_bind_cq:
err_ep_bind_eq:
err_endpoint:
return ret;
}
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;
if (MTL_OFI_PROG_AUTO == control_progress) {
hints->domain_attr->control_progress = FI_PROGRESS_AUTO;
} else {
hints->domain_attr->control_progress = FI_PROGRESS_MANUAL;
}
if (MTL_OFI_PROG_MANUAL == data_progress) {
hints->domain_attr->data_progress = FI_PROGRESS_MANUAL;
} else {
hints->domain_attr->data_progress = FI_PROGRESS_AUTO;
}
if (MTL_OFI_AV_TABLE == av_type) {
hints->domain_attr->av_type = FI_AV_TABLE;
} else {
hints->domain_attr->av_type = FI_AV_MAP;
}
hints->domain_attr->resource_mgmt = FI_RM_ENABLED;
/**
* 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.
*/
av_attr.type = (MTL_OFI_AV_TABLE == av_type) ? FI_AV_TABLE: 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;
}
void rdm_sr_setup_common_eps(void)
{
int ret = 0, i = 0, j = 0;
struct fi_av_attr attr;
size_t addrlen = 0;
attr.type = FI_AV_MAP;
attr.count = NUMEPS;
cq_attr.format = FI_CQ_FORMAT_TAGGED;
cq_attr.size = 1024;
cq_attr.wait_obj = 0;
target = malloc(BUF_SZ * 3); /* 3x BUF_SZ for multi recv testing */
assert(target);
source = malloc(BUF_SZ);
assert(source);
uc_target = malloc(BUF_SZ);
assert(uc_target);
uc_source = malloc(BUF_SZ);
assert(uc_source);
ret = fi_fabric(fi[0]->fabric_attr, &fab, NULL);
cr_assert(!ret, "fi_fabric");
for (; i < NUMEPS; i++) {
ret = fi_domain(fab, fi[i], dom + i, NULL);
cr_assert(!ret, "fi_domain");
ret = fi_open_ops(&dom[i]->fid, FI_GNI_DOMAIN_OPS_1,
0, (void **) (gni_domain_ops + i), NULL);
ret = fi_av_open(dom[i], &attr, av + i, NULL);
cr_assert(!ret, "fi_av_open");
ret = fi_endpoint(dom[i], fi[i], ep + i, NULL);
cr_assert(!ret, "fi_endpoint");
ret = fi_cq_open(dom[i], &cq_attr, msg_cq + i, 0);
cr_assert(!ret, "fi_cq_open");
ret = fi_ep_bind(ep[i], &msg_cq[i]->fid, FI_SEND | FI_RECV);
cr_assert(!ret, "fi_ep_bind");
ret = fi_getname(&ep[i]->fid, NULL, &addrlen);
cr_assert(addrlen > 0);
ep_name[i] = malloc(addrlen);
cr_assert(ep_name[i] != NULL);
ret = fi_getname(&ep[i]->fid, ep_name[i], &addrlen);
cr_assert(ret == FI_SUCCESS);
}
for (i = 0; i < NUMEPS; i++) {
/* Insert all gni addresses into each av */
for (j = 0; j < NUMEPS; j++) {
ret = fi_av_insert(av[i], ep_name[j], 1, &gni_addr[j],
0, NULL);
cr_assert(ret == 1);
}
ret = fi_ep_bind(ep[i], &av[i]->fid, 0);
cr_assert(!ret, "fi_ep_bind");
ret = fi_enable(ep[i]);
cr_assert(!ret, "fi_ep_enable");
ret = fi_cntr_open(dom[i], &cntr_attr, send_cntr + i, 0);
cr_assert(!ret, "fi_cntr_open");
ret = fi_ep_bind(ep[i], &send_cntr[i]->fid, FI_SEND);
cr_assert(!ret, "fi_ep_bind");
ret = fi_cntr_open(dom[i], &cntr_attr, recv_cntr + i, 0);
cr_assert(!ret, "fi_cntr_open");
ret = fi_ep_bind(ep[i], &recv_cntr[i]->fid, FI_RECV);
cr_assert(!ret, "fi_ep_bind");
}
}
static int alloc_ep_res(struct fi_info *fi)
{
struct fi_cq_attr cq_attr;
struct fi_av_attr av_attr;
int ret;
buffer_size = opts.user_options & FT_OPT_SIZE ?
opts.transfer_size : test_size[TEST_CNT - 1].size;
if (max_msg_size > 0 && buffer_size > max_msg_size) {
buffer_size = max_msg_size;
}
if (buffer_size < fi->src_addrlen) {
buffer_size = fi->src_addrlen;
}
buffer_size += prefix_len;
buf = malloc(buffer_size);
if (!buf) {
perror("malloc");
return -1;
}
buf_ptr = (char *)buf + prefix_len;
memset(&cq_attr, 0, sizeof cq_attr);
cq_attr.format = FI_CQ_FORMAT_CONTEXT;
cq_attr.wait_obj = FI_WAIT_NONE;
cq_attr.size = max_credits << 1;
ret = fi_cq_open(dom, &cq_attr, &scq, NULL);
if (ret) {
FT_PRINTERR("fi_cq_open", ret);
goto err1;
}
ret = fi_cq_open(dom, &cq_attr, &rcq, NULL);
if (ret) {
FT_PRINTERR("fi_cq_open", ret);
goto err2;
}
ret = fi_mr_reg(dom, buf, buffer_size, 0, 0, 0, 0, &mr, NULL);
if (ret) {
FT_PRINTERR("fi_mr_reg", ret);
goto err3;
}
memset(&av_attr, 0, sizeof(av_attr));
av_attr.type = fi->domain_attr->av_type ?
fi->domain_attr->av_type : FI_AV_MAP;
av_attr.name = NULL;
av_attr.flags = 0;
ret = fi_av_open(dom, &av_attr, &av, NULL);
if (ret) {
FT_PRINTERR("fi_av_open", ret);
goto err4;
}
ret = fi_endpoint(dom, fi, &ep, NULL);
if (ret) {
FT_PRINTERR("fi_endpoint", ret);
goto err5;
}
return 0;
err5:
fi_close(&av->fid);
err4:
fi_close(&mr->fid);
err3:
fi_close(&rcq->fid);
err2:
fi_close(&scq->fid);
err1:
free(buf);
return ret;
}
static int alloc_ep_res(struct fi_info *fi)
{
struct fi_cq_attr cq_attr;
struct epoll_event event;
int ret, fd;
buf = malloc(buffer_size);
if (!buf) {
perror("malloc");
return -1;
}
memset(&cq_attr, 0, sizeof cq_attr);
cq_attr.format = FI_CQ_FORMAT_CONTEXT;
cq_attr.wait_obj = FI_WAIT_FD;
cq_attr.size = rx_depth;
/* Open completion queue for send completions */
ret = fi_cq_open(dom, &cq_attr, &scq, NULL);
if (ret) {
FT_PRINTERR("fi_cq_open", ret);
goto err1;
}
/* Open completion queue for recv completions */
ret = fi_cq_open(dom, &cq_attr, &rcq, NULL);
if (ret) {
FT_PRINTERR("fi_cq_open", ret);
goto err2;
}
/* Create epoll set */
epfd = epoll_create1(0);
if (epfd < 0) {
ret = -errno;
FT_PRINTERR("epoll_create1", ret);
goto err3;
}
/* Retrieve receive queue wait object */
ret = fi_control (&rcq->fid, FI_GETWAIT, (void *) &fd);
if (ret) {
FT_PRINTERR("fi_control(FI_GETWAIT)", ret);
goto err4;
}
/* Add receive queue wait object to epoll set */
memset((void *)&event, 0, sizeof event);
event.events = EPOLLIN;
event.data.ptr = (void *)&rcq->fid;
ret = epoll_ctl(epfd, EPOLL_CTL_ADD, fd, &event);
if (ret) {
ret = -errno;
FT_PRINTERR("epoll_ctl", ret);
goto err4;
}
/* Retrieve send queue wait object */
ret = fi_control (&scq->fid, FI_GETWAIT, (void *) &fd);
if (ret) {
FT_PRINTERR("fi_control(FI_GETWAIT)", ret);
goto err4;
}
/* Add send queue wait object to epoll set */
memset((void *)&event, 0, sizeof event);
event.events = EPOLLIN;
event.data.ptr = (void *)&scq->fid;
ret = epoll_ctl(epfd, EPOLL_CTL_ADD, fd, &event);
if (ret) {
ret = -errno;
FT_PRINTERR("epoll_ctl", ret);
goto err4;
}
/* Register memory */
ret = fi_mr_reg(dom, buf, buffer_size, 0, 0, 0, 0, &mr, NULL);
if (ret) {
FT_PRINTERR("fi_mr_reg", ret);
goto err4;
}
ret = fi_endpoint(dom, fi, &ep, NULL);
if (ret) {
FT_PRINTERR("fi_endpoint", ret);
goto err5;
}
return 0;
err5:
fi_close(&mr->fid);
err4:
close(epfd);
err3:
fi_close(&rcq->fid);
err2:
fi_close(&scq->fid);
err1:
free(buf);
return ret;
}
void rdm_api_setup_ep(void)
{
int ret, i, j;
struct fi_av_attr attr;
size_t addrlen = 0;
/* Get info about fabric services with the provided hints */
for (i = 0; i < NUMEPS; i++) {
ret = fi_getinfo(FI_VERSION(1, 0), NULL, 0, 0, hints[i],
&fi[i]);
cr_assert(!ret, "fi_getinfo");
}
attr.type = FI_AV_MAP;
attr.count = NUMEPS;
cq_attr.format = FI_CQ_FORMAT_TAGGED;
cq_attr.size = 1024;
cq_attr.wait_obj = 0;
target = malloc(BUF_SZ * 3); /* 3x BUF_SZ for multi recv testing */
assert(target);
source = malloc(BUF_SZ);
assert(source);
uc_target = malloc(BUF_SZ);
assert(uc_target);
uc_source = malloc(BUF_SZ);
assert(uc_source);
ret = fi_fabric(fi[0]->fabric_attr, &fab, NULL);
cr_assert(!ret, "fi_fabric");
for (i = 0; i < NUMEPS; i++) {
ret = fi_domain(fab, fi[i], dom + i, NULL);
cr_assert(!ret, "fi_domain");
ret = fi_open_ops(&dom[i]->fid, FI_GNI_DOMAIN_OPS_1,
0, (void **) (gni_domain_ops + i), NULL);
ret = fi_av_open(dom[i], &attr, av + i, NULL);
cr_assert(!ret, "fi_av_open");
ret = fi_endpoint(dom[i], fi[i], ep + i, NULL);
cr_assert(!ret, "fi_endpoint");
ret = fi_cq_open(dom[i], &cq_attr, msg_cq + i, 0);
cr_assert(!ret, "fi_cq_open");
ret = fi_ep_bind(ep[i], &msg_cq[i]->fid, FI_SEND | FI_RECV);
cr_assert(!ret, "fi_ep_bind");
ret = fi_getname(&ep[i]->fid, NULL, &addrlen);
cr_assert(addrlen > 0);
ep_name[i] = malloc(addrlen);
cr_assert(ep_name[i] != NULL);
ret = fi_getname(&ep[i]->fid, ep_name[i], &addrlen);
cr_assert(ret == FI_SUCCESS);
}
for (i = 0; i < NUMEPS; i++) {
/* Insert all gni addresses into each av */
for (j = 0; j < NUMEPS; j++) {
ret = fi_av_insert(av[i], ep_name[j], 1, &gni_addr[j],
0, NULL);
cr_assert(ret == 1);
}
ret = fi_ep_bind(ep[i], &av[i]->fid, 0);
cr_assert(!ret, "fi_ep_bind");
ret = fi_enable(ep[i]);
cr_assert(!ret, "fi_ep_enable");
ret = fi_cntr_open(dom[i], &cntr_attr, send_cntr + i, 0);
cr_assert(!ret, "fi_cntr_open");
ret = fi_ep_bind(ep[i], &send_cntr[i]->fid, FI_SEND);
cr_assert(!ret, "fi_ep_bind");
ret = fi_cntr_open(dom[i], &cntr_attr, recv_cntr + i, 0);
cr_assert(!ret, "fi_cntr_open");
ret = fi_ep_bind(ep[i], &recv_cntr[i]->fid, FI_RECV);
cr_assert(!ret, "fi_ep_bind");
}
for (i = 0; i < NUMEPS; i++) {
ret = fi_mr_reg(dom[i], target, 3 * BUF_SZ,
FI_REMOTE_WRITE, 0, 0, 0, rem_mr + i, &target);
cr_assert_eq(ret, 0);
ret = fi_mr_reg(dom[i], source, BUF_SZ,
FI_REMOTE_WRITE, 0, 0, 0, loc_mr + i, &source);
cr_assert_eq(ret, 0);
mr_key[i] = fi_mr_key(rem_mr[i]);
}
}
void rdm_sr_setup(void)
{
int ret = 0;
struct fi_av_attr attr;
size_t addrlen = 0;
hints = fi_allocinfo();
cr_assert(hints, "fi_allocinfo");
hints->domain_attr->cq_data_size = 4;
hints->mode = ~0;
hints->fabric_attr->name = strdup("gni");
ret = fi_getinfo(FI_VERSION(1, 0), NULL, 0, 0, hints, &fi);
cr_assert(!ret, "fi_getinfo");
ret = fi_fabric(fi->fabric_attr, &fab, NULL);
cr_assert(!ret, "fi_fabric");
ret = fi_domain(fab, fi, &dom, NULL);
cr_assert(!ret, "fi_domain");
attr.type = FI_AV_MAP;
attr.count = 16;
ret = fi_av_open(dom, &attr, &av, NULL);
cr_assert(!ret, "fi_av_open");
ret = fi_endpoint(dom, fi, &ep[0], NULL);
cr_assert(!ret, "fi_endpoint");
cq_attr.format = FI_CQ_FORMAT_CONTEXT;
cq_attr.size = 1024;
cq_attr.wait_obj = 0;
ret = fi_cq_open(dom, &cq_attr, &msg_cq[0], 0);
cr_assert(!ret, "fi_cq_open");
ret = fi_cq_open(dom, &cq_attr, &msg_cq[1], 0);
cr_assert(!ret, "fi_cq_open");
ret = fi_ep_bind(ep[0], &msg_cq[0]->fid, FI_SEND | FI_RECV);
cr_assert(!ret, "fi_ep_bind");
ret = fi_getname(&ep[0]->fid, NULL, &addrlen);
cr_assert(addrlen > 0);
ep_name[0] = malloc(addrlen);
cr_assert(ep_name[0] != NULL);
ret = fi_getname(&ep[0]->fid, ep_name[0], &addrlen);
cr_assert(ret == FI_SUCCESS);
ret = fi_endpoint(dom, fi, &ep[1], NULL);
cr_assert(!ret, "fi_endpoint");
ret = fi_ep_bind(ep[1], &msg_cq[1]->fid, FI_SEND | FI_RECV);
cr_assert(!ret, "fi_ep_bind");
ep_name[1] = malloc(addrlen);
cr_assert(ep_name[1] != NULL);
ret = fi_getname(&ep[1]->fid, ep_name[1], &addrlen);
cr_assert(ret == FI_SUCCESS);
ret = fi_av_insert(av, ep_name[0], 1, &gni_addr[0], 0,
NULL);
cr_assert(ret == 1);
ret = fi_av_insert(av, ep_name[1], 1, &gni_addr[1], 0,
NULL);
cr_assert(ret == 1);
ret = fi_ep_bind(ep[0], &av->fid, 0);
cr_assert(!ret, "fi_ep_bind");
ret = fi_ep_bind(ep[1], &av->fid, 0);
cr_assert(!ret, "fi_ep_bind");
ret = fi_enable(ep[0]);
cr_assert(!ret, "fi_ep_enable");
ret = fi_enable(ep[1]);
cr_assert(!ret, "fi_ep_enable");
target = malloc(BUF_SZ);
assert(target);
source = malloc(BUF_SZ);
assert(source);
ret = fi_mr_reg(dom, target, BUF_SZ,
FI_REMOTE_WRITE, 0, 0, 0, &rem_mr, &target);
cr_assert_eq(ret, 0);
ret = fi_mr_reg(dom, source, BUF_SZ,
FI_REMOTE_WRITE, 0, 0, 0, &loc_mr, &source);
cr_assert_eq(ret, 0);
mr_key = fi_mr_key(rem_mr);
}
static int alloc_ep_res(struct fi_info *fi)
{
struct fi_cq_attr cq_attr;
uint64_t access_mode;
int ret;
buffer_size = opts.user_options & FT_OPT_SIZE ?
opts.transfer_size : test_size[TEST_CNT - 1].size;
buf = malloc(MAX(buffer_size, sizeof(uint64_t)));
if (!buf) {
perror("malloc");
return -1;
}
memset(&cq_attr, 0, sizeof cq_attr);
cq_attr.format = FI_CQ_FORMAT_DATA;
cq_attr.wait_obj = FI_WAIT_NONE;
cq_attr.size = max_credits << 1;
ret = fi_cq_open(dom, &cq_attr, &scq, NULL);
if (ret) {
FT_PRINTERR("fi_cq_open", ret);
goto err1;
}
ret = fi_cq_open(dom, &cq_attr, &rcq, NULL);
if (ret) {
FT_PRINTERR("fi_cq_open", ret);
goto err2;
}
switch (op_type) {
case FT_RMA_READ:
access_mode = FI_REMOTE_READ;
break;
case FT_RMA_WRITE:
case FT_RMA_WRITEDATA:
access_mode = FI_REMOTE_WRITE;
break;
default:
assert(0);
ret = -FI_EINVAL;
goto err3;
}
ret = fi_mr_reg(dom, buf, MAX(buffer_size, sizeof(uint64_t)),
access_mode, 0, 0, 0, &mr, NULL);
if (ret) {
FT_PRINTERR("fi_mr_reg", ret);
goto err3;
}
if (!cmeq) {
ret = alloc_cm_res();
if (ret)
goto err4;
}
return 0;
err4:
fi_close(&mr->fid);
err3:
fi_close(&rcq->fid);
err2:
fi_close(&scq->fid);
err1:
free(buf);
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 ft_alloc_ep_res(struct fi_info *fi)
{
int ret;
if (hints->caps & FI_RMA) {
ret = ft_set_rma_caps(fi, opts.rma_op);
if (ret)
return ret;
}
ret = ft_alloc_msgs();
if (ret)
return ret;
if (cq_attr.format == FI_CQ_FORMAT_UNSPEC) {
if (fi->caps & FI_TAGGED)
cq_attr.format = FI_CQ_FORMAT_TAGGED;
else
cq_attr.format = FI_CQ_FORMAT_CONTEXT;
}
if (opts.options & FT_OPT_TX_CQ) {
ft_cq_set_wait_attr();
cq_attr.size = fi->tx_attr->size;
ret = fi_cq_open(domain, &cq_attr, &txcq, &txcq);
if (ret) {
FT_PRINTERR("fi_cq_open", ret);
return ret;
}
}
if (opts.options & FT_OPT_TX_CNTR) {
ft_cntr_set_wait_attr();
ret = fi_cntr_open(domain, &cntr_attr, &txcntr, &txcntr);
if (ret) {
FT_PRINTERR("fi_cntr_open", ret);
return ret;
}
}
if (opts.options & FT_OPT_RX_CQ) {
ft_cq_set_wait_attr();
cq_attr.size = fi->rx_attr->size;
ret = fi_cq_open(domain, &cq_attr, &rxcq, &rxcq);
if (ret) {
FT_PRINTERR("fi_cq_open", ret);
return ret;
}
}
if (opts.options & FT_OPT_RX_CNTR) {
ft_cntr_set_wait_attr();
ret = fi_cntr_open(domain, &cntr_attr, &rxcntr, &rxcntr);
if (ret) {
FT_PRINTERR("fi_cntr_open", ret);
return ret;
}
}
if (fi->ep_attr->type == FI_EP_RDM || fi->ep_attr->type == FI_EP_DGRAM) {
if (fi->domain_attr->av_type != FI_AV_UNSPEC)
av_attr.type = fi->domain_attr->av_type;
if (opts.av_name) {
av_attr.name = opts.av_name;
}
ret = fi_av_open(domain, &av_attr, &av, NULL);
if (ret) {
FT_PRINTERR("fi_av_open", ret);
return ret;
}
}
return 0;
}
void sep_setup_common(int av_type)
{
int ret, i, j;
struct fi_av_attr av_attr = {0};
size_t addrlen = 0;
hints = fi_allocinfo();
cr_assert(hints, "fi_allocinfo");
hints->ep_attr->type = FI_EP_RDM;
hints->caps = FI_ATOMIC | FI_RMA | FI_MSG | FI_NAMED_RX_CTX;
hints->mode = FI_LOCAL_MR;
hints->domain_attr->cq_data_size = NUMEPS * 2;
hints->domain_attr->data_progress = FI_PROGRESS_AUTO;
hints->domain_attr->mr_mode = FI_MR_BASIC;
hints->fabric_attr->prov_name = strdup("gni");
hints->ep_attr->tx_ctx_cnt = ctx_cnt;
hints->ep_attr->rx_ctx_cnt = ctx_cnt;
for (i = 0; i < NUMEPS; i++) {
ret = fi_getinfo(FI_VERSION(1, 0), NULL, 0, 0, hints, &fi[i]);
cr_assert(!ret, "fi_getinfo");
tx_cq[i] = calloc(ctx_cnt, sizeof(*tx_cq));
rx_cq[i] = calloc(ctx_cnt, sizeof(*rx_cq));
tx_ep[i] = calloc(ctx_cnt, sizeof(*tx_ep));
rx_ep[i] = calloc(ctx_cnt, sizeof(*rx_ep));
if (!tx_cq[i] || !tx_cq[i] ||
!tx_ep[i] || !rx_ep[i]) {
cr_assert(0, "calloc");
}
}
ctx_cnt = MIN(ctx_cnt, fi[0]->domain_attr->rx_ctx_cnt);
ctx_cnt = MIN(ctx_cnt, fi[0]->domain_attr->tx_ctx_cnt);
cr_assert(ctx_cnt, "ctx_cnt is 0");
ret = fi_fabric(fi[0]->fabric_attr, &fab, NULL);
cr_assert(!ret, "fi_fabric");
rx_ctx_bits = 0;
while (ctx_cnt >> ++rx_ctx_bits);
av_attr.rx_ctx_bits = rx_ctx_bits;
av_attr.type = av_type;
av_attr.count = NUMEPS;
cq_attr.format = FI_CQ_FORMAT_TAGGED;
cq_attr.size = 1024;
cq_attr.wait_obj = FI_WAIT_NONE;
rx_addr = calloc(ctx_cnt, sizeof(*rx_addr));
target = calloc(BUF_SZ, 1);
source = calloc(BUF_SZ, 1);
iov_src_buf = malloc(BUF_SZ * IOV_CNT);
iov_dest_buf = malloc(BUF_SZ * IOV_CNT);
src_iov = malloc(sizeof(struct iovec) * IOV_CNT);
dest_iov = malloc(sizeof(struct iovec) * IOV_CNT);
if (!rx_addr || !target || !source || !iov_src_buf || !iov_dest_buf ||
!src_iov || !dest_iov) {
cr_assert(0, "allocation");
}
for (i = 0; i < IOV_CNT; i++) {
src_iov[i].iov_base = malloc(BUF_SZ);
assert(src_iov[i].iov_base != NULL);
dest_iov[i].iov_base = malloc(BUF_SZ * 3);
assert(dest_iov[i].iov_base != NULL);
}
for (i = 0; i < NUMEPS; i++) {
fi[i]->ep_attr->tx_ctx_cnt = ctx_cnt;
fi[i]->ep_attr->rx_ctx_cnt = ctx_cnt;
ret = fi_domain(fab, fi[i], &dom[i], NULL);
cr_assert(!ret, "fi_domain");
ret = fi_scalable_ep(dom[i], fi[i], &sep[i], NULL);
cr_assert(!ret, "fi_scalable_ep");
ret = fi_av_open(dom[i], &av_attr, &av[i], NULL);
cr_assert(!ret, "fi_av_open");
ret = fi_cntr_open(dom[i], &cntr_attr, &send_cntr[i], 0);
cr_assert(!ret, "fi_cntr_open");
ret = fi_cntr_open(dom[i], &cntr_attr, &recv_cntr[i], 0);
cr_assert(!ret, "fi_cntr_open");
for (j = 0; j < ctx_cnt; j++) {
ret = fi_tx_context(sep[i], j, NULL, &tx_ep[i][j],
NULL);
cr_assert(!ret, "fi_tx_context");
ret = fi_cq_open(dom[i], &cq_attr, &tx_cq[i][j],
NULL);
cr_assert(!ret, "fi_cq_open");
ret = fi_rx_context(sep[i], j, NULL, &rx_ep[i][j],
NULL);
cr_assert(!ret, "fi_rx_context");
ret = fi_cq_open(dom[i], &cq_attr, &rx_cq[i][j],
NULL);
cr_assert(!ret, "fi_cq_open");
}
ret = fi_scalable_ep_bind(sep[i], &av[i]->fid, 0);
cr_assert(!ret, "fi_scalable_ep_bind");
for (j = 0; j < ctx_cnt; j++) {
ret = fi_ep_bind(tx_ep[i][j], &tx_cq[i][j]->fid,
FI_TRANSMIT);
cr_assert(!ret, "fi_ep_bind");
ret = fi_ep_bind(tx_ep[i][j], &send_cntr[i]->fid,
FI_SEND | FI_WRITE);
cr_assert(!ret, "fi_ep_bind");
ret = fi_enable(tx_ep[i][j]);
cr_assert(!ret, "fi_enable");
ret = fi_ep_bind(rx_ep[i][j], &rx_cq[i][j]->fid,
FI_RECV);
cr_assert(!ret, "fi_ep_bind");
ret = fi_ep_bind(rx_ep[i][j], &recv_cntr[i]->fid,
FI_RECV | FI_READ);
cr_assert(!ret, "fi_ep_bind");
ret = fi_enable(rx_ep[i][j]);
cr_assert(!ret, "fi_enable");
}
}
for (i = 0; i < NUMEPS; i++) {
ret = fi_enable(sep[i]);
cr_assert(!ret, "fi_enable");
ret = fi_getname(&sep[i]->fid, NULL, &addrlen);
cr_assert(addrlen > 0);
ep_name[i] = malloc(addrlen);
cr_assert(ep_name[i] != NULL);
ret = fi_getname(&sep[i]->fid, ep_name[i], &addrlen);
cr_assert(ret == FI_SUCCESS);
ret = fi_mr_reg(dom[i], target, BUF_SZ, FI_REMOTE_WRITE,
0, 0, 0, &rem_mr[i], &target);
cr_assert_eq(ret, 0);
ret = fi_mr_reg(dom[i], source, BUF_SZ, FI_REMOTE_WRITE,
0, 0, 0, &loc_mr[i], &source);
cr_assert_eq(ret, 0);
mr_key[i] = fi_mr_key(rem_mr[i]);
ret = fi_mr_reg(dom[i], iov_dest_buf, IOV_CNT * BUF_SZ,
FI_REMOTE_WRITE, 0, 0, 0, iov_dest_buf_mr + i,
&iov_dest_buf);
cr_assert_eq(ret, 0);
ret = fi_mr_reg(dom[i], iov_src_buf, IOV_CNT * BUF_SZ,
FI_REMOTE_WRITE, 0, 0, 0, iov_src_buf_mr + i,
&iov_src_buf);
cr_assert_eq(ret, 0);
}
for (i = 0; i < NUMEPS; i++) {
for (j = 0; j < NUMEPS; j++) {
ret = fi_av_insert(av[i], ep_name[j], 1, &gni_addr[j],
0, NULL);
cr_assert(ret == 1);
}
}
for (i = 0; i < ctx_cnt; i++) {
rx_addr[i] = fi_rx_addr(gni_addr[1], i, rx_ctx_bits);
}
}
static void vc_setup_common(void)
{
int ret = 0;
struct fi_av_attr attr;
size_t addrlen = 0;
struct gnix_fid_av *gnix_av;
hints->fabric_attr->name = strdup("gni");
ret = fi_getinfo(FI_VERSION(1, 0), NULL, 0, 0, hints, &fi);
cr_assert(!ret, "fi_getinfo");
ret = fi_fabric(fi->fabric_attr, &fab, NULL);
cr_assert(!ret, "fi_fabric");
ret = fi_domain(fab, fi, &dom, NULL);
cr_assert(!ret, "fi_domain");
attr.type = FI_AV_MAP;
attr.count = 16;
ret = fi_av_open(dom, &attr, &av, NULL);
cr_assert(!ret, "fi_av_open");
gnix_av = container_of(av, struct gnix_fid_av, av_fid);
ret = fi_endpoint(dom, fi, &ep[0], NULL);
cr_assert(!ret, "fi_endpoint");
ret = fi_getname(&ep[0]->fid, NULL, &addrlen);
cr_assert(addrlen > 0);
ep_name[0] = malloc(addrlen);
cr_assert(ep_name[0] != NULL);
ep_name[1] = malloc(addrlen);
cr_assert(ep_name[1] != NULL);
ret = fi_getname(&ep[0]->fid, ep_name[0], &addrlen);
cr_assert(ret == FI_SUCCESS);
ret = fi_endpoint(dom, fi, &ep[1], NULL);
cr_assert(!ret, "fi_endpoint");
ret = fi_getname(&ep[1]->fid, ep_name[1], &addrlen);
cr_assert(ret == FI_SUCCESS);
ret = fi_av_insert(av, ep_name[0], 1, &gni_addr[0], 0,
NULL);
cr_assert(ret == 1);
ret = _gnix_av_lookup(gnix_av, gni_addr[0], &gnix_addr[0]);
cr_assert(ret == FI_SUCCESS);
ret = fi_av_insert(av, ep_name[1], 1, &gni_addr[1], 0,
NULL);
cr_assert(ret == 1);
ret = _gnix_av_lookup(gnix_av, gni_addr[1], &gnix_addr[1]);
cr_assert(ret == FI_SUCCESS);
ret = fi_ep_bind(ep[0], &av->fid, 0);
cr_assert(!ret, "fi_ep_bind");
cq_attr.format = FI_CQ_FORMAT_TAGGED;
cq_attr.size = 1024;
cq_attr.wait_obj = 0;
ret = fi_cq_open(dom, &cq_attr, &cq, 0);
cr_assert(!ret, "fi_cq_open");
ret = fi_ep_bind(ep[0], &cq->fid, FI_SEND | FI_RECV);
cr_assert(!ret, "fi_ep_bind");
ret = fi_enable(ep[0]);
cr_assert(!ret, "fi_enable");
ret = fi_ep_bind(ep[1], &cq->fid, FI_SEND | FI_RECV);
cr_assert(!ret, "fi_ep_bind");
ret = fi_ep_bind(ep[1], &av->fid, 0);
cr_assert(!ret, "fi_ep_bind");
ret = fi_enable(ep[1]);
cr_assert(!ret, "fi_ep_enable");
}
static void setup(void)
{
int i, j;
int ret = 0;
struct fi_av_attr attr;
size_t addrlen = 0;
struct fi_gni_ops_domain *gni_domain_ops;
uint32_t rx_cq_size;
hints = fi_allocinfo();
cr_assert(hints, "fi_allocinfo");
hints->domain_attr->cq_data_size = 4;
hints->domain_attr->data_progress = FI_PROGRESS_MANUAL;
hints->mode = ~0;
hints->fabric_attr->name = strdup("gni");
ret = fi_getinfo(FI_VERSION(1, 0), NULL, 0, 0, hints, &fi);
cr_assert(!ret, "fi_getinfo");
ret = fi_fabric(fi->fabric_attr, &fab, NULL);
cr_assert(!ret, "fi_fabric");
attr.type = FI_AV_TABLE;
attr.count = NUM_EPS;
cq_attr.format = FI_CQ_FORMAT_CONTEXT;
cq_attr.size = 1024;
cq_attr.wait_obj = 0;
for (i = 0; i < NUM_EPS; i++) {
ret = fi_domain(fab, fi, &dom[i], NULL);
cr_assert(!ret, "fi_domain");
ret = fi_open_ops(&dom[i]->fid, FI_GNI_DOMAIN_OPS_1, 0,
(void **) &gni_domain_ops, NULL);
cr_assert(ret == FI_SUCCESS, "fi_open_ops");
rx_cq_size = min_rx_cq_size;
ret = gni_domain_ops->set_val(&dom[i]->fid, GNI_RX_CQ_SIZE,
&rx_cq_size);
cr_assert(ret == FI_SUCCESS, "set_val");
ret = fi_av_open(dom[i], &attr, &av[i], NULL);
cr_assert(!ret, "fi_av_open");
ret = fi_endpoint(dom[i], fi, &ep[i], NULL);
cr_assert(!ret, "fi_endpoint");
cr_assert(ep[i]);
ret = fi_cq_open(dom[i], &cq_attr, &msg_cq[i], 0);
cr_assert(!ret, "fi_cq_open");
ret = fi_ep_bind(ep[i], &msg_cq[i]->fid, FI_SEND | FI_RECV);
cr_assert(!ret, "fi_ep_bind");
}
ret = fi_getname(&ep[0]->fid, NULL, &addrlen);
cr_assert_eq(ret, -FI_ETOOSMALL);
cr_assert(addrlen > 0);
for (i = 0; i < NUM_EPS; i++) {
ep_name[i] = malloc(addrlen);
cr_assert(ep_name[i] != NULL);
ret = fi_getname(&ep[i]->fid, ep_name[i], &addrlen);
cr_assert(ret == FI_SUCCESS);
for (j = 0; j < NUM_EPS; j++) {
ret = fi_av_insert(av[j], ep_name[i],
1, &gni_addr[i], 0, NULL);
cr_assert(ret == 1);
}
}
for (i = 0; i < NUM_EPS; i++) {
ret = fi_ep_bind(ep[i], &av[i]->fid, 0);
cr_assert(!ret, "fi_ep_bind");
ret = fi_enable(ep[i]);
cr_assert(!ret, "fi_ep_enable");
ret = fi_mr_reg(dom[i], target, NUM_EPS*sizeof(int),
FI_RECV, 0, 0, 0, &rem_mr[i], &target);
cr_assert_eq(ret, 0);
ret = fi_mr_reg(dom[i], source, NUM_EPS*sizeof(int),
FI_SEND, 0, 0, 0, &loc_mr[i], &source);
cr_assert_eq(ret, 0);
mr_key[i] = fi_mr_key(rem_mr[i]);
}
}
static int alloc_ep_res(struct fi_info *fi)
{
struct fi_cq_attr cq_attr;
struct fi_rx_attr rx_attr;
struct fi_tx_attr tx_attr;
struct fi_av_attr av_attr;
int i, ret = 0;
buffer_size = test_size[TEST_CNT - 1].size;
buf = malloc(buffer_size);
remote_fi_addr = (fi_addr_t *)malloc(sizeof(*remote_fi_addr) * ep_cnt);
if (!buf || !remote_fi_addr) {
perror("malloc");
goto err1;
}
memset(&cq_attr, 0, sizeof cq_attr);
cq_attr.format = FI_CQ_FORMAT_CONTEXT;
cq_attr.wait_obj = FI_WAIT_NONE;
cq_attr.size = rx_depth;
memset(&tx_attr, 0, sizeof tx_attr);
memset(&rx_attr, 0, sizeof rx_attr);
ret = fi_stx_context(dom, &tx_attr, &stx_ctx, NULL);
if (ret) {
FT_PRINTERR("fi_stx_context", ret);
goto err1;
}
ret = fi_cq_open(dom, &cq_attr, &scq, NULL);
if (ret) {
FT_PRINTERR("fi_cq_open", ret);
goto err2;
}
ret = fi_srx_context(dom, &rx_attr, &srx_ctx, NULL);
if (ret) {
FT_PRINTERR("fi_srx_context", ret);
goto err3;
}
ret = fi_cq_open(dom, &cq_attr, &rcq, NULL);
if (ret) {
FT_PRINTERR("fi_cq_open", ret);
goto err4;
}
ret = fi_mr_reg(dom, buf, buffer_size, 0, 0, 0, 0, &mr, NULL);
if (ret) {
FT_PRINTERR("fi_mr_reg", ret);
goto err5;
}
memset(&av_attr, 0, sizeof av_attr);
av_attr.type = fi->domain_attr->av_type ?
fi->domain_attr->av_type : FI_AV_MAP;
av_attr.count = ep_cnt;
ret = fi_av_open(dom, &av_attr, &av, NULL);
if (ret) {
FT_PRINTERR("fi_av_open", ret);
goto err6;
}
ep = calloc(ep_cnt, sizeof(*ep));
if (!ep) {
perror("malloc");
goto err7;
}
for (i = 0; i < ep_cnt; i++) {
ret = fi_endpoint(dom, fi, &ep[i], NULL);
if (ret) {
FT_PRINTERR("fi_endpoint", ret);
goto err8;
}
}
return 0;
err8:
FT_CLOSEV(ep, ep_cnt);
err7:
fi_close(&av->fid);
err6:
fi_close(&mr->fid);
err5:
fi_close(&rcq->fid);
err4:
fi_close(&srx_ctx->fid);
err3:
fi_close(&scq->fid);
err2:
fi_close(&stx_ctx->fid);
err1:
free(buf);
free(remote_fi_addr);
return ret;
}
static int alloc_ep_res(struct fi_info *fi)
{
struct fi_cq_attr cq_attr;
struct fi_av_attr av_attr;
int ret;
buf = malloc(buffer_size);
if (!buf) {
perror("malloc");
return -1;
}
memset(&cq_attr, 0, sizeof cq_attr);
cq_attr.format = FI_CQ_FORMAT_CONTEXT;
cq_attr.wait_obj = FI_WAIT_NONE;
cq_attr.size = rx_depth;
ret = fi_cq_open(dom, &cq_attr, &scq, NULL);
if (ret) {
FT_PRINTERR("fi_cq_open", ret);
goto err1;
}
ret = fi_cq_open(dom, &cq_attr, &rcq, NULL);
if (ret) {
FT_PRINTERR("fi_cq_open", ret);
goto err2;
}
ret = fi_mr_reg(dom, buf, buffer_size, 0, 0, 0, 0, &mr, NULL);
if (ret) {
FT_PRINTERR("fi_mr_reg", ret);
goto err3;
}
memset(&av_attr, 0, sizeof av_attr);
av_attr.type = fi->domain_attr->av_type ?
fi->domain_attr->av_type : FI_AV_MAP;
av_attr.count = 1;
av_attr.name = NULL;
ret = fi_av_open(dom, &av_attr, &av, NULL);
if (ret) {
FT_PRINTERR("fi_av_open", ret);
goto err4;
}
ret = fi_endpoint(dom, fi, &ep, NULL);
if (ret) {
FT_PRINTERR("fi_endpoint", ret);
goto err5;
}
return 0;
err5:
fi_close(&av->fid);
err4:
fi_close(&mr->fid);
err3:
fi_close(&rcq->fid);
err2:
fi_close(&scq->fid);
err1:
free(buf);
return ret;
}
static void libfabric_init() {
int i;
struct fi_info *info = NULL;
struct fi_info *hints = fi_allocinfo();
struct fi_av_attr av_attr = {0};
struct fi_cq_attr cq_attr = {0};
int max_tx_ctx, max_rx_ctx;
int comm_concurrency;
int rx_ctx_cnt;
int rx_ctx_bits = 0;
hints->mode = ~0;
hints->caps = FI_RMA
| FI_ATOMIC
| FI_SOURCE /* do we want this? */
| FI_READ
| FI_WRITE
| FI_REMOTE_READ
| FI_REMOTE_WRITE
| FI_MULTI_RECV
| FI_FENCE;
hints->addr_format = FI_FORMAT_UNSPEC;
#if defined(CHPL_COMM_SUBSTRATE_SOCKETS)
//
// fi_freeinfo(hints) will free() hints->fabric_attr->prov_name; this
// is documented, though poorly. So, get that space from malloc().
//
{
const char s[] = "sockets";
char* sDup = sys_malloc(sizeof(s));
strcpy(sDup, s);
hints->fabric_attr->prov_name = sDup;
}
#elif defined(CHPL_COMM_SUBSTRATE_GNI)
#error "Substrate GNI not supported"
#else
#error "Substrate type not supported"
#endif
/* connectionless reliable */
hints->ep_attr->type = FI_EP_RDM;
hints->domain_attr->threading = FI_THREAD_UNSPEC;
hints->domain_attr->control_progress = FI_PROGRESS_MANUAL;
hints->domain_attr->data_progress = FI_PROGRESS_MANUAL;
hints->domain_attr->av_type = FI_AV_TABLE;
hints->domain_attr->mr_mode = FI_MR_SCALABLE;
hints->domain_attr->resource_mgmt = FI_RM_ENABLED;
// hints->domain_attr->cq_data_size
hints->tx_attr->op_flags = FI_COMPLETION;
hints->rx_attr->op_flags = FI_COMPLETION;
OFICHKERR(fi_getinfo(FI_VERSION(1,0), NULL, NULL, 0, hints, &info));
if (info == NULL) {
chpl_internal_error("No fabrics detected.");
} else {
#ifdef PRINT_FI_GETINFO
struct fi_info *cur;
for (cur = info; cur; cur = cur->next) {
printf("---\n");
printf("%s", fi_tostr(cur, FI_TYPE_INFO));
}
printf("\n");
#endif
}
ofi.num_am_ctx = 1; // Would we ever want more?
max_tx_ctx = info->domain_attr->max_ep_tx_ctx;
max_rx_ctx = info->domain_attr->max_ep_rx_ctx;
comm_concurrency = get_comm_concurrency();
ofi.num_tx_ctx = comm_concurrency+ofi.num_am_ctx > max_tx_ctx ?
max_tx_ctx-ofi.num_am_ctx : comm_concurrency;
ofi.num_rx_ctx = comm_concurrency+ofi.num_am_ctx > max_rx_ctx ?
max_rx_ctx-ofi.num_am_ctx : comm_concurrency;
info->ep_attr->tx_ctx_cnt = ofi.num_tx_ctx + ofi.num_am_ctx;
info->ep_attr->rx_ctx_cnt = ofi.num_rx_ctx + ofi.num_am_ctx;
OFICHKERR(fi_fabric(info->fabric_attr, &ofi.fabric, NULL));
OFICHKERR(fi_domain(ofi.fabric, info, &ofi.domain, NULL));
rx_ctx_cnt = ofi.num_rx_ctx + ofi.num_am_ctx;
while (rx_ctx_cnt >> ++rx_ctx_bits);
av_attr.rx_ctx_bits = rx_ctx_bits;
av_attr.type = FI_AV_TABLE;
av_attr.count = chpl_numNodes;
OFICHKERR(fi_av_open(ofi.domain, &av_attr, &ofi.av, NULL));
OFICHKERR(fi_scalable_ep(ofi.domain, info, &ofi.ep, NULL));
OFICHKERR(fi_scalable_ep_bind(ofi.ep, &ofi.av->fid, 0));
/* set up tx and rx contexts */
cq_attr.format = FI_CQ_FORMAT_CONTEXT;
cq_attr.size = 1024; /* ??? */
cq_attr.wait_obj = FI_WAIT_UNSPEC;
ofi.tx_ep = (struct fid_ep **) chpl_mem_allocMany(ofi.num_tx_ctx,
sizeof(ofi.tx_ep[0]),
CHPL_RT_MD_COMM_PER_LOC_INFO,
0, 0);
ofi.tx_cq = (struct fid_cq **) chpl_mem_allocMany(ofi.num_tx_ctx,
sizeof(ofi.tx_cq[0]),
CHPL_RT_MD_COMM_PER_LOC_INFO,
0, 0);
for (i = 0; i < ofi.num_tx_ctx; i++) {
OFICHKERR(fi_tx_context(ofi.ep, i, NULL, &ofi.tx_ep[i], NULL));
OFICHKERR(fi_cq_open(ofi.domain, &cq_attr, &ofi.tx_cq[i], NULL));
OFICHKERR(fi_ep_bind(ofi.tx_ep[i], &ofi.tx_cq[i]->fid, FI_TRANSMIT));
OFICHKERR(fi_enable(ofi.tx_ep[i]));
}
ofi.rx_ep = (struct fid_ep **) chpl_mem_allocMany(ofi.num_rx_ctx,
sizeof(ofi.rx_ep[0]),
CHPL_RT_MD_COMM_PER_LOC_INFO,
0, 0);
ofi.rx_cq = (struct fid_cq **) chpl_mem_allocMany(ofi.num_rx_ctx,
sizeof(ofi.rx_cq[0]),
CHPL_RT_MD_COMM_PER_LOC_INFO,
0, 0);
for (i = 0; i < ofi.num_rx_ctx; i++) {
OFICHKERR(fi_rx_context(ofi.ep, i, NULL, &ofi.rx_ep[i], NULL));
OFICHKERR(fi_cq_open(ofi.domain, &cq_attr, &ofi.rx_cq[i], NULL));
OFICHKERR(fi_ep_bind(ofi.rx_ep[i], &ofi.rx_cq[i]->fid, FI_RECV));
OFICHKERR(fi_enable(ofi.rx_ep[i]));
}
ofi.am_tx_ep = (struct fid_ep **) chpl_mem_allocMany(ofi.num_am_ctx,
sizeof(ofi.am_tx_ep[0]),
CHPL_RT_MD_COMM_PER_LOC_INFO,
0, 0);
ofi.am_tx_cq = (struct fid_cq **) chpl_mem_allocMany(ofi.num_am_ctx,
sizeof(ofi.am_tx_cq[0]),
CHPL_RT_MD_COMM_PER_LOC_INFO,
0, 0);
/* set up AM contexts */
for (i = 0; i < ofi.num_am_ctx; i++) {
OFICHKERR(fi_tx_context(ofi.ep, i+ofi.num_tx_ctx, NULL, &ofi.am_tx_ep[i], NULL));
OFICHKERR(fi_cq_open(ofi.domain, &cq_attr, &ofi.am_tx_cq[i], NULL));
OFICHKERR(fi_ep_bind(ofi.am_tx_ep[i], &ofi.am_tx_cq[i]->fid, FI_TRANSMIT));
OFICHKERR(fi_enable(ofi.am_tx_ep[i]));
}
ofi.am_rx_ep = (struct fid_ep **) chpl_mem_allocMany(ofi.num_am_ctx,
sizeof(ofi.am_rx_ep[0]),
CHPL_RT_MD_COMM_PER_LOC_INFO,
0, 0);
ofi.am_rx_cq = (struct fid_cq **) chpl_mem_allocMany(ofi.num_am_ctx,
sizeof(ofi.am_rx_cq[0]),
CHPL_RT_MD_COMM_PER_LOC_INFO,
0, 0);
for (i = 0; i < ofi.num_am_ctx; i++) {
OFICHKERR(fi_rx_context(ofi.ep, i+ofi.num_rx_ctx, NULL, &ofi.am_rx_ep[i], NULL));
OFICHKERR(fi_cq_open(ofi.domain, &cq_attr, &ofi.am_rx_cq[i], NULL));
OFICHKERR(fi_ep_bind(ofi.am_rx_ep[i], &ofi.am_rx_cq[i]->fid, FI_RECV));
OFICHKERR(fi_enable(ofi.am_rx_ep[i]));
}
OFICHKERR(fi_enable(ofi.ep));
libfabric_init_addrvec(rx_ctx_cnt, rx_ctx_bits);
OFICHKERR(fi_mr_reg(ofi.domain, 0, SIZE_MAX,
FI_READ | FI_WRITE | FI_REMOTE_READ | FI_REMOTE_WRITE |
FI_SEND | FI_RECV, 0,
(uint64_t) chpl_nodeID, 0, &ofi.mr, NULL));
fi_freeinfo(info); /* No error returned */
fi_freeinfo(hints); /* No error returned */
chpl_msg(2, "%d: completed libfabric initialization\n", chpl_nodeID);
}
static void setup_ep(void)
{
int ret;
struct fi_av_attr attr;
size_t addrlen = 0;
attr.type = FI_AV_MAP;
attr.count = 16;
ret = fi_av_open(dom, &attr, &av, NULL);
cr_assert(!ret, "fi_av_open");
ret = fi_endpoint(dom, fi, &ep[0], NULL);
cr_assert(!ret, "fi_endpoint");
cq_attr.format = FI_CQ_FORMAT_TAGGED;
cq_attr.size = 1024;
cq_attr.wait_obj = 0;
ret = fi_cq_open(dom, &cq_attr, &msg_cq[0], 0);
cr_assert(!ret, "fi_cq_open");
ret = fi_cq_open(dom, &cq_attr, &msg_cq[1], 0);
cr_assert(!ret, "fi_cq_open");
ret = fi_ep_bind(ep[0], &msg_cq[0]->fid, FI_SEND | FI_RECV);
cr_assert(!ret, "fi_ep_bind");
ret = fi_getname(&ep[0]->fid, NULL, &addrlen);
cr_assert(addrlen > 0);
ep_name[0] = malloc(addrlen);
cr_assert(ep_name[0] != NULL);
ret = fi_getname(&ep[0]->fid, ep_name[0], &addrlen);
cr_assert(ret == FI_SUCCESS);
ret = fi_endpoint(dom, fi, &ep[1], NULL);
cr_assert(!ret, "fi_endpoint");
ret = fi_ep_bind(ep[1], &msg_cq[1]->fid, FI_SEND | FI_RECV);
cr_assert(!ret, "fi_ep_bind");
ep_name[1] = malloc(addrlen);
cr_assert(ep_name[1] != NULL);
ret = fi_getname(&ep[1]->fid, ep_name[1], &addrlen);
cr_assert(ret == FI_SUCCESS);
ret = fi_av_insert(av, ep_name[0], 1, &gni_addr[0], 0,
NULL);
cr_assert(ret == 1);
ret = fi_av_insert(av, ep_name[1], 1, &gni_addr[1], 0,
NULL);
cr_assert(ret == 1);
ret = fi_ep_bind(ep[0], &av->fid, 0);
cr_assert(!ret, "fi_ep_bind");
ret = fi_ep_bind(ep[1], &av->fid, 0);
cr_assert(!ret, "fi_ep_bind");
ret = fi_enable(ep[0]);
cr_assert(!ret, "fi_ep_enable");
ret = fi_enable(ep[1]);
cr_assert(!ret, "fi_ep_enable");
}
int rxd_cq_open(struct fid_domain *domain, struct fi_cq_attr *attr,
struct fid_cq **cq_fid, void *context)
{
int ret;
struct rxd_cq *cq;
struct rxd_domain *rxd_domain;
cq = calloc(1, sizeof(*cq));
if (!cq)
return -FI_ENOMEM;
ret = ofi_cq_init(&rxd_prov, domain, attr, &cq->util_cq,
&rxd_cq_progress, context);
if (ret)
goto err1;
switch (attr->format) {
case FI_CQ_FORMAT_UNSPEC:
case FI_CQ_FORMAT_CONTEXT:
cq->write_fn = cq->util_cq.wait ?
rxd_cq_write_ctx_signal : rxd_cq_write_ctx;
break;
case FI_CQ_FORMAT_MSG:
cq->write_fn = cq->util_cq.wait ?
rxd_cq_write_msg_signal : rxd_cq_write_msg;
break;
case FI_CQ_FORMAT_DATA:
cq->write_fn = cq->util_cq.wait ?
rxd_cq_write_data_signal : rxd_cq_write_data;
break;
case FI_CQ_FORMAT_TAGGED:
cq->write_fn = cq->util_cq.wait ?
rxd_cq_write_tagged_signal : rxd_cq_write_tagged;
break;
default:
ret = -FI_EINVAL;
goto err2;
}
rxd_domain = container_of(domain, struct rxd_domain, util_domain.domain_fid);
attr->format = FI_CQ_FORMAT_MSG;
ret = fi_cq_open(rxd_domain->dg_domain, attr, &cq->dg_cq, context);
if (ret)
goto err2;
cq->unexp_pool = util_buf_pool_create(
RXD_EP_MAX_UNEXP_PKT * sizeof (struct rxd_unexp_cq_entry),
RXD_BUF_POOL_ALIGNMENT, 0, RXD_EP_MAX_UNEXP_PKT);
if (!cq->unexp_pool) {
ret = -FI_ENOMEM;
goto err3;
}
dlist_init(&cq->dom_entry);
dlist_init(&cq->unexp_list);
fastlock_init(&cq->lock);
fastlock_acquire(&rxd_domain->lock);
dlist_insert_tail(&cq->dom_entry, &rxd_domain->cq_list);
fastlock_release(&rxd_domain->lock);
*cq_fid = &cq->util_cq.cq_fid;
(*cq_fid)->fid.ops = &rxd_cq_fi_ops;
(*cq_fid)->ops = &rxd_cq_ops;
cq->domain = rxd_domain;
return 0;
err3:
ofi_cq_cleanup(&cq->util_cq);
err2:
fi_close(&cq->dg_cq->fid);
err1:
free(cq);
return ret;
}
int client_connect(struct fi_info *prov, simple_context_t *ctx)
{
struct fi_eq_attr eq_attr = { 0 };
struct fi_cq_attr cq_attr = { 0 };
struct sockaddr_in addr = { 0 };
int ret;
print_trace("in\n");
connected = 0;
ret = fi_fabric(prov->fabric_attr, &ctx->fabric, NULL);
if (ret) {
print_err("fi_fabric returned %d\n", ret);
ctx->fabric = NULL;
return ret;
}
ret = fi_domain(ctx->fabric, prov, &ctx->domain, NULL);
if (ret) {
print_err("fi_fdomain returned %d\n", ret);
ctx->domain = NULL;
return ret;
}
/* set QP WR depth */
prov->ep_attr->tx_ctx_cnt = (size_t) (post_depth + 1);
prov->ep_attr->rx_ctx_cnt = (size_t) (post_depth + 1);
/* set ScatterGather max depth */
prov->tx_attr->iov_limit = 1;
prov->rx_attr->iov_limit = 1;
prov->tx_attr->inject_size = 0; /* no INLINE support */
ret = fi_endpoint(ctx->domain, prov, &ctx->ep, CONTEXT);
if (ret) {
print_err("fi_endpoint returned %d\n", ret);
ctx->ep = NULL;
return ret;
}
eq_attr.wait_obj = FI_WAIT_NONE;
ret = fi_eq_open(ctx->fabric, &eq_attr, &ctx->eq, NULL);
if (ret) {
print_err("fi_eq_open returned %d\n", ret);
ctx->eq = NULL;
return ret;
}
cq_attr.size = post_depth * 4;
cq_attr.flags = FI_SEND;
cq_attr.format = FI_CQ_FORMAT_MSG;
cq_attr.wait_obj = FI_WAIT_NONE;
cq_attr.wait_cond = FI_CQ_COND_NONE;
ret = fi_cq_open(ctx->domain, &cq_attr, &ctx->scq, NULL);
if (ret) {
print_err("fi_cq_open returned %d\n", ret);
ctx->scq = NULL;
return ret;
}
cq_attr.flags = FI_RECV;
ret = fi_cq_open(ctx->domain, &cq_attr, &ctx->rcq, NULL);
if (ret) {
print_err("fi_cq_open returned %d\n", ret);
ctx->rcq = NULL;
return ret;
}
ret = fi_ep_bind(ctx->ep, &ctx->eq->fid, 0);
if (ret) {
print_err("fi_ep_bind returned %d\n", ret);
return ret;
}
ret = fi_ep_bind(ctx->ep, &ctx->scq->fid, FI_SEND);
if (ret) {
print_err("fi_ep_bind returned %d\n", ret);
return ret;
}
ret = fi_ep_bind(ctx->ep, &ctx->rcq->fid, FI_RECV);
if (ret) {
print_err("fi_ep_bind returned %d\n", ret);
return ret;
}
ret = fi_enable(ctx->ep);
if (ret) {
print_err("fi_enable returned %d\n", ret);
return ret;
}
addr.sin_family = AF_INET;
addr.sin_port = htons(TEST_PORT);
ret = in4_pton(svr_ipaddr, strlen(svr_ipaddr),
(u8 *)&addr.sin_addr.s_addr, '\0', NULL);
if (ret != 1) {
print_err("Err converting target server IP address '%s'?\n",
svr_ipaddr);
return -EINVAL;
}
ret = fi_connect(ctx->ep, &addr, PRIVATE_DATA, sizeof(PRIVATE_DATA));
if (ret) {
print_err("fi_connect returned %d\n", ret);
return ret;
}
connected = 1;
return 0;
}
int ft_alloc_active_res(struct fi_info *fi)
{
int ret;
ret = ft_alloc_msgs();
if (ret)
return ret;
if (cq_attr.format == FI_CQ_FORMAT_UNSPEC) {
if (fi->caps & FI_TAGGED)
cq_attr.format = FI_CQ_FORMAT_TAGGED;
else
cq_attr.format = FI_CQ_FORMAT_CONTEXT;
}
if (opts.options & FT_OPT_TX_CQ) {
ft_cq_set_wait_attr();
cq_attr.size = fi->tx_attr->size;
ret = fi_cq_open(domain, &cq_attr, &txcq, &txcq);
if (ret) {
FT_PRINTERR("fi_cq_open", ret);
return ret;
}
if (opts.comp_method == FT_COMP_WAIT_FD) {
ret = fi_control(&txcq->fid, FI_GETWAIT, (void *) &tx_fd);
if (ret) {
FT_PRINTERR("fi_control(FI_GETWAIT)", ret);
return ret;
}
}
}
if (opts.options & FT_OPT_TX_CNTR) {
ft_cntr_set_wait_attr();
ret = fi_cntr_open(domain, &cntr_attr, &txcntr, &txcntr);
if (ret) {
FT_PRINTERR("fi_cntr_open", ret);
return ret;
}
}
if (opts.options & FT_OPT_RX_CQ) {
ft_cq_set_wait_attr();
cq_attr.size = fi->rx_attr->size;
ret = fi_cq_open(domain, &cq_attr, &rxcq, &rxcq);
if (ret) {
FT_PRINTERR("fi_cq_open", ret);
return ret;
}
if (opts.comp_method == FT_COMP_WAIT_FD) {
ret = fi_control(&rxcq->fid, FI_GETWAIT, (void *) &rx_fd);
if (ret) {
FT_PRINTERR("fi_control(FI_GETWAIT)", ret);
return ret;
}
}
}
if (opts.options & FT_OPT_RX_CNTR) {
ft_cntr_set_wait_attr();
ret = fi_cntr_open(domain, &cntr_attr, &rxcntr, &rxcntr);
if (ret) {
FT_PRINTERR("fi_cntr_open", ret);
return ret;
}
}
if (fi->ep_attr->type == FI_EP_RDM || fi->ep_attr->type == FI_EP_DGRAM) {
if (fi->domain_attr->av_type != FI_AV_UNSPEC)
av_attr.type = fi->domain_attr->av_type;
if (opts.av_name) {
av_attr.name = opts.av_name;
}
ret = fi_av_open(domain, &av_attr, &av, NULL);
if (ret) {
FT_PRINTERR("fi_av_open", ret);
return ret;
}
}
ret = fi_endpoint(domain, fi, &ep, NULL);
if (ret) {
FT_PRINTERR("fi_endpoint", ret);
return ret;
}
return 0;
}
/* returns 0 on success or a negative value that can be stringified with
* fi_strerror on error */
static int setup_ep_fixture(struct fid_ep **ep_o)
{
int ret;
struct fi_info *myfi;
struct fi_av_attr av_attr;
struct fi_cq_attr cq_attr;
assert(ep_o != NULL);
ret = 0;
myfi = fi_dupinfo(fi);
if (myfi == NULL) {
printf("fi_dupinfo returned NULL\n");
goto fail;
}
ret = fi_endpoint(domain, myfi, ep_o, NULL);
if (ret != 0) {
printf("fi_endpoint %s\n", fi_strerror(-ret));
goto fail;
}
memset(&cq_attr, 0, sizeof cq_attr);
cq_attr.format = FI_CQ_FORMAT_CONTEXT;
cq_attr.wait_obj = FI_WAIT_NONE;
cq_attr.size = TX_CQ_DEPTH;
ret = fi_cq_open(domain, &cq_attr, &wcq, /*context=*/NULL);
if (ret != 0) {
printf("fi_cq_open %s\n", fi_strerror(-ret));
goto fail;
}
memset(&cq_attr, 0, sizeof cq_attr);
cq_attr.format = FI_CQ_FORMAT_CONTEXT;
cq_attr.wait_obj = FI_WAIT_NONE;
cq_attr.size = RX_CQ_DEPTH;
ret = fi_cq_open(domain, &cq_attr, &rcq, /*context=*/NULL);
if (ret != 0) {
printf("fi_cq_open %s\n", fi_strerror(-ret));
goto fail;
}
memset(&av_attr, 0, sizeof av_attr);
av_attr.type = myfi->domain_attr->av_type ?
myfi->domain_attr->av_type : FI_AV_MAP;
av_attr.count = 1;
av_attr.name = NULL;
ret = fi_av_open(domain, &av_attr, &av, NULL);
if (ret != 0) {
printf("fi_av_open %s\n", fi_strerror(-ret));
goto fail;
}
ret = fi_ep_bind(*ep_o, &wcq->fid, FI_SEND);
if (ret != 0) {
printf("fi_ep_bind(wcq) %s\n", fi_strerror(-ret));
goto fail;
}
ret = fi_ep_bind(*ep_o, &rcq->fid, FI_RECV);
if (ret != 0) {
printf("fi_ep_bind(rcq) %s\n", fi_strerror(-ret));
goto fail;
}
ret = fi_ep_bind(*ep_o, &av->fid, 0);
if (ret != 0) {
printf("fi_ep_bind(av) %s\n", fi_strerror(-ret));
goto fail;
}
ret = fi_enable(*ep_o);
if (ret != 0) {
printf("fi_enable %s\n", fi_strerror(-ret));
goto fail;
}
if (myfi != NULL) {
fi_freeinfo(myfi);
}
return ret;
fail:
if (myfi != NULL) {
fi_freeinfo(myfi);
}
return teardown_ep_fixture(*ep_o);
}
static int init_node(struct cma_node *node, struct fi_info *info)
{
struct fi_cq_attr cq_attr;
int ret;
ret = fi_domain(fabric, info, &node->domain, NULL);
if (ret) {
FT_PRINTERR("fi_domain", ret);
goto out;
}
memset(&cq_attr, 0, sizeof cq_attr);
cq_attr.size = hints->tx_attr->size ? hints->tx_attr->size : 1;
cq_attr.format = FI_CQ_FORMAT_CONTEXT;
ret = fi_cq_open(node->domain, &cq_attr, &node->cq[SEND_CQ_INDEX], NULL);
if (ret) {
FT_PRINTERR("fi_cq_open", ret);
goto out;
}
ret = fi_cq_open(node->domain, &cq_attr, &node->cq[RECV_CQ_INDEX], NULL);
if (ret) {
FT_PRINTERR("fi_cq_open", ret);
goto out;
}
ret = fi_endpoint(node->domain, info, &node->ep, node);
if (ret) {
FT_PRINTERR("fi_endpoint", ret);
goto out;
}
ret = fi_ep_bind(node->ep, &node->cq[SEND_CQ_INDEX]->fid, FI_SEND);
if (ret) {
FT_PRINTERR("fi_ep_bind", ret);
goto out;
}
ret = fi_ep_bind(node->ep, &node->cq[RECV_CQ_INDEX]->fid, FI_RECV);
if (ret) {
FT_PRINTERR("fi_ep_bind", ret);
goto out;
}
ret = fi_ep_bind(node->ep, &eq->fid, 0);
if (ret) {
FT_PRINTERR("fi_ep_bind", ret);
goto out;
}
ret = fi_enable(node->ep);
if (ret) {
FT_PRINTERR("fi_enable", ret);
goto out;
}
ret = create_messages(node);
if (ret) {
printf("cmatose: failed to create messages: %d\n", ret);
goto out;
}
out:
return ret;
}
void rdm_str_addr_sr_setup_common(void)
{
int ret = 0, i = 0, j = 0;
struct fi_av_attr attr;
memset(&attr, 0, sizeof(attr));
attr.type = FI_AV_MAP;
attr.count = NUMEPS;
cq_attr.format = FI_CQ_FORMAT_TAGGED;
cq_attr.size = 1024;
cq_attr.wait_obj = 0;
target_base = malloc(GNIT_ALIGN_LEN(BUF_SZ));
assert(target_base);
target = GNIT_ALIGN_BUFFER(char *, target_base);
source_base = malloc(GNIT_ALIGN_LEN(BUF_SZ));
assert(source_base);
source = GNIT_ALIGN_BUFFER(char *, source_base);
ret = fi_fabric(fi[0]->fabric_attr, &fab, NULL);
cr_assert(!ret, "fi_fabric");
for (i = 0; i < NUMEPS; i++) {
ret = fi_domain(fab, fi[i], dom + i, NULL);
cr_assert(!ret, "fi_domain");
ret = fi_av_open(dom[i], &attr, av + i, NULL);
cr_assert(!ret, "fi_av_open");
ret = fi_endpoint(dom[i], fi[i], ep + i, NULL);
cr_assert(!ret, "fi_endpoint");
ret = fi_cq_open(dom[i], &cq_attr, msg_cq + i, 0);
cr_assert(!ret, "fi_cq_open");
ret = fi_ep_bind(ep[i], &msg_cq[i]->fid, FI_SEND | FI_RECV);
cr_assert(!ret, "fi_ep_bind");
ret = fi_getname(&ep[i]->fid, NULL, &addrlen);
cr_assert(addrlen > 0);
ep_name[i] = malloc(addrlen);
cr_assert(ep_name[i] != NULL);
ret = fi_getname(&ep[i]->fid, ep_name[i], &addrlen);
cr_assert(ret == FI_SUCCESS);
}
for (i = 0; i < NUMEPS; i++) {
/*
* To test API-1.1: Reporting of unknown source addresses --
* only insert addresses into the sender's av
*/
if (i < (NUMEPS / 2)) {
for (j = 0; j < NUMEPS; j++) {
dbg_printf("Only does src EP insertions\n");
ret = fi_av_insert(av[i], ep_name[j], 1,
&gni_addr[j],
0, NULL);
cr_assert(ret == 1);
}
}
ret = fi_ep_bind(ep[i], &av[i]->fid, 0);
cr_assert(!ret, "fi_ep_bind");
ret = fi_enable(ep[i]);
cr_assert(!ret, "fi_ep_enable");
}
}
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 int alloc_ep_res(struct fi_info *fi)
{
struct fi_cq_attr cq_attr;
struct fi_av_attr av_attr;
int ret;
buffer_size = opts.user_options & FT_OPT_SIZE ?
opts.transfer_size : test_size[TEST_CNT - 1].size;
buf = malloc(MAX(buffer_size, sizeof(uint64_t)));
if (!buf) {
perror("malloc");
return -1;
}
result = malloc(MAX(buffer_size, sizeof(uint64_t)));
if (!result) {
perror("malloc");
return -1;
}
compare = malloc(MAX(buffer_size, sizeof(uint64_t)));
if (!compare) {
perror("malloc");
return -1;
}
memset(&cq_attr, 0, sizeof cq_attr);
cq_attr.format = FI_CQ_FORMAT_CONTEXT;
cq_attr.wait_obj = FI_WAIT_NONE;
cq_attr.size = 128;
ret = fi_cq_open(dom, &cq_attr, &scq, NULL);
if (ret) {
FT_PRINTERR("fi_cq_open", ret);
goto err1;
}
ret = fi_cq_open(dom, &cq_attr, &rcq, NULL);
if (ret) {
FT_PRINTERR("fi_cq_open", ret);
goto err2;
}
// registers local data buffer buff that specifies
// the first operand of the atomic operation
ret = fi_mr_reg(dom, buf, MAX(buffer_size, sizeof(uint64_t)),
FI_REMOTE_READ | FI_REMOTE_WRITE, 0,
get_mr_key(), 0, &mr, NULL);
if (ret) {
FT_PRINTERR("fi_mr_reg", ret);
goto err3;
}
// registers local data buffer that stores initial value of
// the remote buffer
ret = fi_mr_reg(dom, result, MAX(buffer_size, sizeof(uint64_t)),
FI_REMOTE_READ | FI_REMOTE_WRITE, 0,
get_mr_key(), 0, &mr_result, NULL);
if (ret) {
FT_PRINTERR("fi_mr_reg", -ret);
goto err4;
}
// registers local data buffer that contains comparison data
ret = fi_mr_reg(dom, compare, MAX(buffer_size, sizeof(uint64_t)),
FI_REMOTE_READ | FI_REMOTE_WRITE, 0,
get_mr_key(), 0, &mr_compare, NULL);
if (ret) {
FT_PRINTERR("fi_mr_reg", ret);
goto err5;
}
memset(&av_attr, 0, sizeof av_attr);
av_attr.type = fi->domain_attr->av_type ?
fi->domain_attr->av_type : FI_AV_MAP;
av_attr.count = 1;
av_attr.name = NULL;
ret = fi_av_open(dom, &av_attr, &av, NULL);
if (ret) {
FT_PRINTERR("fi_av_open", ret);
goto err6;
}
ret = fi_endpoint(dom, fi, &ep, NULL);
if (ret) {
FT_PRINTERR("fi_endpoint", ret);
goto err7;
}
return 0;
err7:
fi_close(&av->fid);
err6:
fi_close(&mr_compare->fid);
err5:
fi_close(&mr_result->fid);
err4:
fi_close(&mr->fid);
err3:
fi_close(&rcq->fid);
err2:
fi_close(&scq->fid);
err1:
free(buf);
free(result);
free(compare);
return ret;
}
void rdm_rma_setup(void)
{
int ret = 0;
struct fi_av_attr attr;
size_t addrlen = 0;
hints = fi_allocinfo();
cr_assert(hints, "fi_allocinfo");
hints->domain_attr->cq_data_size = 4;
hints->mode = ~0;
hints->fabric_attr->name = strdup("gni");
ret = fi_getinfo(FI_VERSION(1, 0), NULL, 0, 0, hints, &fi);
cr_assert(!ret, "fi_getinfo");
ret = fi_fabric(fi->fabric_attr, &fab, NULL);
cr_assert(!ret, "fi_fabric");
ret = fi_domain(fab, fi, &dom, NULL);
cr_assert(!ret, "fi_domain");
ret = fi_open_ops(&dom->fid, FI_GNI_DOMAIN_OPS_1,
0, (void **) &gni_domain_ops, NULL);
attr.type = FI_AV_MAP;
attr.count = 16;
ret = fi_av_open(dom, &attr, &av, NULL);
cr_assert(!ret, "fi_av_open");
ret = fi_endpoint(dom, fi, &ep[0], NULL);
cr_assert(!ret, "fi_endpoint");
cq_attr.format = FI_CQ_FORMAT_TAGGED;
cq_attr.size = 1024;
cq_attr.wait_obj = 0;
ret = fi_cq_open(dom, &cq_attr, &send_cq, 0);
cr_assert(!ret, "fi_cq_open");
/*
* imitate shmem, etc. use FI_WRITE for bind
* flag
*/
ret = fi_ep_bind(ep[0], &send_cq->fid, FI_TRANSMIT);
cr_assert(!ret, "fi_ep_bind");
ret = fi_getname(&ep[0]->fid, NULL, &addrlen);
cr_assert(addrlen > 0);
ep_name[0] = malloc(addrlen);
cr_assert(ep_name[0] != NULL);
ep_name[1] = malloc(addrlen);
cr_assert(ep_name[1] != NULL);
ret = fi_getname(&ep[0]->fid, ep_name[0], &addrlen);
cr_assert(ret == FI_SUCCESS);
ret = fi_endpoint(dom, fi, &ep[1], NULL);
cr_assert(!ret, "fi_endpoint");
cq_attr.format = FI_CQ_FORMAT_TAGGED;
ret = fi_cq_open(dom, &cq_attr, &recv_cq, 0);
cr_assert(!ret, "fi_cq_open");
ret = fi_ep_bind(ep[1], &recv_cq->fid, FI_RECV);
cr_assert(!ret, "fi_ep_bind");
ret = fi_getname(&ep[1]->fid, ep_name[1], &addrlen);
cr_assert(ret == FI_SUCCESS);
ret = fi_av_insert(av, ep_name[0], 1, &gni_addr[0], 0,
NULL);
cr_assert(ret == 1);
ret = fi_av_insert(av, ep_name[1], 1, &gni_addr[1], 0,
NULL);
cr_assert(ret == 1);
ret = fi_ep_bind(ep[0], &av->fid, 0);
cr_assert(!ret, "fi_ep_bind");
ret = fi_ep_bind(ep[1], &av->fid, 0);
cr_assert(!ret, "fi_ep_bind");
ret = fi_enable(ep[0]);
cr_assert(!ret, "fi_ep_enable");
ret = fi_enable(ep[1]);
cr_assert(!ret, "fi_ep_enable");
target = malloc(BUF_SZ);
assert(target);
source = malloc(BUF_SZ);
assert(source);
ret = fi_mr_reg(dom, target, BUF_SZ,
FI_REMOTE_WRITE, 0, 0, 0, &rem_mr, &target);
cr_assert_eq(ret, 0);
ret = fi_mr_reg(dom, source, BUF_SZ,
FI_REMOTE_WRITE, 0, 0, 0, &loc_mr, &source);
cr_assert_eq(ret, 0);
uc_source = malloc(BUF_SZ);
assert(uc_source);
mr_key = fi_mr_key(rem_mr);
ret = fi_cntr_open(dom, &cntr_attr, &write_cntr, 0);
cr_assert(!ret, "fi_cntr_open");
ret = fi_ep_bind(ep[0], &write_cntr->fid, FI_WRITE);
cr_assert(!ret, "fi_ep_bind");
ret = fi_cntr_open(dom, &cntr_attr, &read_cntr, 0);
cr_assert(!ret, "fi_cntr_open");
ret = fi_ep_bind(ep[0], &read_cntr->fid, FI_READ);
cr_assert(!ret, "fi_ep_bind");
writes = reads = write_errs = read_errs = 0;
}
void cancel_setup(void)
{
int ret = 0;
struct fi_av_attr attr;
size_t addrlen = 0;
int rem_requested_key, loc_requested_key;
hints = fi_allocinfo();
cr_assert(hints, "fi_allocinfo");
hints->domain_attr->mr_mode = GNIX_DEFAULT_MR_MODE;
hints->domain_attr->cq_data_size = 4;
hints->mode = mode_bits;
hints->fabric_attr->prov_name = strdup("gni");
ret = fi_getinfo(fi_version(), NULL, 0, 0, hints, &fi);
cr_assert(!ret, "fi_getinfo");
ret = fi_fabric(fi->fabric_attr, &fab, NULL);
cr_assert(!ret, "fi_fabric");
ret = fi_domain(fab, fi, &dom, NULL);
cr_assert(!ret, "fi_domain");
memset(&attr, 0, sizeof(attr));
attr.type = FI_AV_MAP;
attr.count = 16;
ret = fi_av_open(dom, &attr, &av, NULL);
cr_assert(!ret, "fi_av_open");
ret = fi_endpoint(dom, fi, &ep[0], NULL);
cr_assert(!ret, "fi_endpoint");
cq_attr.format = FI_CQ_FORMAT_CONTEXT;
cq_attr.size = 1024;
cq_attr.wait_obj = 0;
ret = fi_cq_open(dom, &cq_attr, &msg_cq[0], 0);
cr_assert(!ret, "fi_cq_open");
ret = fi_cq_open(dom, &cq_attr, &msg_cq[1], 0);
cr_assert(!ret, "fi_cq_open");
ret = fi_ep_bind(ep[0], &msg_cq[0]->fid, FI_SEND | FI_RECV);
cr_assert(!ret, "fi_ep_bind");
ret = fi_getname(&ep[0]->fid, NULL, &addrlen);
cr_assert(addrlen > 0);
ep_name[0] = malloc(addrlen);
cr_assert(ep_name[0] != NULL);
ret = fi_getname(&ep[0]->fid, ep_name[0], &addrlen);
cr_assert(ret == FI_SUCCESS);
ret = fi_endpoint(dom, fi, &ep[1], NULL);
cr_assert(!ret, "fi_endpoint");
ret = fi_ep_bind(ep[1], &msg_cq[1]->fid, FI_SEND | FI_RECV);
cr_assert(!ret, "fi_ep_bind");
ep_name[1] = malloc(addrlen);
cr_assert(ep_name[1] != NULL);
ret = fi_getname(&ep[1]->fid, ep_name[1], &addrlen);
cr_assert(ret == FI_SUCCESS);
ret = fi_av_insert(av, ep_name[0], 1, &gni_addr[0], 0,
NULL);
cr_assert(ret == 1);
ret = fi_av_insert(av, ep_name[1], 1, &gni_addr[1], 0,
NULL);
cr_assert(ret == 1);
ret = fi_ep_bind(ep[0], &av->fid, 0);
cr_assert(!ret, "fi_ep_bind");
ret = fi_ep_bind(ep[1], &av->fid, 0);
cr_assert(!ret, "fi_ep_bind");
ret = fi_enable(ep[0]);
cr_assert(!ret, "fi_ep_enable");
ret = fi_enable(ep[1]);
cr_assert(!ret, "fi_ep_enable");
target_base = malloc(GNIT_ALIGN_LEN(BUF_SZ));
assert(target_base);
target = GNIT_ALIGN_BUFFER(char *, target_base);
source_base = malloc(GNIT_ALIGN_LEN(BUF_SZ));
assert(source_base);
source = GNIT_ALIGN_BUFFER(char *, source_base);
rem_requested_key = USING_SCALABLE(fi) ? 1 : 0;
loc_requested_key = USING_SCALABLE(fi) ? 2 : 0;
ret = fi_mr_reg(dom,
target,
BUF_SZ,
FI_REMOTE_WRITE,
0,
rem_requested_key,
0,
&rem_mr,
&target);
cr_assert_eq(ret, 0);
ret = fi_mr_reg(dom,
source,
BUF_SZ,
FI_REMOTE_WRITE,
0,
loc_requested_key,
0,
&loc_mr,
&source);
cr_assert_eq(ret, 0);
if (USING_SCALABLE(fi)) {
MR_ENABLE(rem_mr, target, BUF_SZ);
MR_ENABLE(loc_mr, source, BUF_SZ);
}
mr_key = fi_mr_key(rem_mr);
}
static int alloc_ep_res(struct fid_ep *sep)
{
struct fi_cq_attr cq_attr;
struct fi_rx_attr rx_attr;
struct fi_tx_attr tx_attr;
struct fi_av_attr av_attr;
int i, ret;
buffer_size = test_size[TEST_CNT - 1].size;
buf = malloc(buffer_size);
scq = calloc(ctx_cnt, sizeof *scq);
rcq = calloc(ctx_cnt, sizeof *rcq);
tx_ep = calloc(ctx_cnt, sizeof *tx_ep);
rx_ep = calloc(ctx_cnt, sizeof *rx_ep);
remote_rx_addr = calloc(ctx_cnt, sizeof *remote_rx_addr);
if (!buf || !scq || !rcq || !tx_ep || !rx_ep || !remote_rx_addr) {
perror("malloc");
return -1;
}
memset(&cq_attr, 0, sizeof cq_attr);
cq_attr.format = FI_CQ_FORMAT_CONTEXT;
cq_attr.wait_obj = FI_WAIT_NONE;
cq_attr.size = rx_depth;
for (i = 0; i < ctx_cnt; i++) {
/* Create TX contexts: tx_ep */
ret = fi_tx_context(sep, i, &tx_attr, &tx_ep[i], NULL);
if (ret) {
FT_PRINTERR("fi_tx_context", ret);
goto err1;
}
ret = fi_cq_open(dom, &cq_attr, &scq[i], NULL);
if (ret) {
FT_PRINTERR("fi_cq_open", ret);
goto err2;
}
}
for (i = 0; i < ctx_cnt; i++) {
/* Create RX contexts: rx_ep */
ret = fi_rx_context(sep, i, &rx_attr, &rx_ep[i], NULL);
if (ret) {
FT_PRINTERR("fi_tx_context", ret);
goto err3;
}
ret = fi_cq_open(dom, &cq_attr, &rcq[i], NULL);
if (ret) {
FT_PRINTERR("fi_cq_open", ret);
goto err4;
}
}
ret = fi_mr_reg(dom, buf, buffer_size, 0, 0, 0, 0, &mr, NULL);
if (ret) {
FT_PRINTERR("fi_mr_reg", ret);
goto err5;
}
/* Get number of bits needed to represent ctx_cnt */
while (ctx_cnt >> ++rx_ctx_bits)
;
memset(&av_attr, 0, sizeof av_attr);
av_attr.type = fi->domain_attr->av_type ?
fi->domain_attr->av_type : FI_AV_MAP;
av_attr.count = 1;
av_attr.rx_ctx_bits = rx_ctx_bits;
/* Open Address Vector */
ret = fi_av_open(dom, &av_attr, &av, NULL);
if (ret) {
FT_PRINTERR("fi_av_open", ret);
goto err6;
}
return 0;
err6:
fi_close(&mr->fid);
err5:
FT_CLOSEV(rcq, ctx_cnt);
err4:
FT_CLOSEV(rx_ep, ctx_cnt);
err3:
FT_CLOSEV(scq, ctx_cnt);
err2:
FT_CLOSEV(tx_ep, ctx_cnt);
err1:
free(buf);
free(rcq);
free(scq);
free(tx_ep);
free(rx_ep);
free(remote_rx_addr);
return ret;
}