ssize_t _psmx_readfrom(struct fid_ep *ep, void *buf, size_t len,
void *desc, fi_addr_t src_addr,
uint64_t addr, uint64_t key, void *context,
uint64_t flags)
{
struct psmx_fid_ep *ep_priv;
struct psmx_fid_av *av;
struct psmx_epaddr_context *epaddr_context;
struct psmx_am_request *req;
psm_amarg_t args[8];
int err;
int chunk_size;
size_t offset = 0;
uint64_t psm_tag;
psm_mq_req_t psm_req;
size_t idx;
if (flags & FI_TRIGGER) {
struct psmx_trigger *trigger;
struct fi_triggered_context *ctxt = context;
trigger = calloc(1, sizeof(*trigger));
if (!trigger)
return -ENOMEM;
trigger->op = PSMX_TRIGGERED_READ;
trigger->cntr = container_of(ctxt->threshold.cntr,
struct psmx_fid_cntr, cntr);
trigger->threshold = ctxt->threshold.threshold;
trigger->read.ep = ep;
trigger->read.buf = buf;
trigger->read.len = len;
trigger->read.desc = desc;
trigger->read.src_addr = src_addr;
trigger->read.addr = addr;
trigger->read.key = key;
trigger->read.context = context;
trigger->read.flags = flags & ~FI_TRIGGER;
psmx_cntr_add_trigger(trigger->cntr, trigger);
return 0;
}
ep_priv = container_of(ep, struct psmx_fid_ep, ep);
assert(ep_priv->domain);
if (!buf)
return -EINVAL;
av = ep_priv->av;
if (av && av->type == FI_AV_TABLE) {
idx = src_addr;
if (idx >= av->last)
return -EINVAL;
src_addr = (fi_addr_t) av->psm_epaddrs[idx];
}
else if (!src_addr) {
return -EINVAL;
}
epaddr_context = psm_epaddr_getctxt((void *)src_addr);
if (epaddr_context->epid == ep_priv->domain->psm_epid)
return psmx_rma_self(PSMX_AM_REQ_READ,
ep_priv, buf, len, desc,
addr, key, context, flags, 0);
req = calloc(1, sizeof(*req));
if (!req)
return -ENOMEM;
req->op = PSMX_AM_REQ_READ;
req->read.buf = buf;
req->read.len = len;
req->read.addr = addr; /* needed? */
req->read.key = key; /* needed? */
req->read.context = context;
req->ep = ep_priv;
PSMX_CTXT_TYPE(&req->fi_context) = PSMX_READ_CONTEXT;
PSMX_CTXT_USER(&req->fi_context) = context;
if (ep_priv->send_cq_event_flag && !(flags & FI_EVENT)) {
PSMX_CTXT_TYPE(&req->fi_context) = PSMX_NOCOMP_READ_CONTEXT;
req->no_event = 1;
}
chunk_size = MIN(PSMX_AM_CHUNK_SIZE, psmx_am_param.max_reply_short);
if (psmx_env.tagged_rma && len > chunk_size) {
psm_tag = PSMX_RMA_BIT | ep_priv->domain->psm_epid;
err = psm_mq_irecv(ep_priv->domain->psm_mq, psm_tag, -1ULL,
0, buf, len, (void *)&req->fi_context, &psm_req);
args[0].u32w0 = PSMX_AM_REQ_READ_LONG;
args[0].u32w1 = len;
args[1].u64 = (uint64_t)req;
args[2].u64 = addr;
args[3].u64 = key;
args[4].u64 = psm_tag;
err = psm_am_request_short((psm_epaddr_t) src_addr,
PSMX_AM_RMA_HANDLER, args, 5, NULL, 0,
PSM_AM_FLAG_NOREPLY, NULL, NULL);
return 0;
}
args[0].u32w0 = PSMX_AM_REQ_READ;
args[1].u64 = (uint64_t)(uintptr_t)req;
args[3].u64 = key;
while (len > chunk_size) {
args[0].u32w1 = chunk_size;
args[2].u64 = addr;
args[4].u64 = offset;
err = psm_am_request_short((psm_epaddr_t) src_addr,
PSMX_AM_RMA_HANDLER, args, 5, NULL, 0,
0, NULL, NULL);
addr += chunk_size;
len -= chunk_size;
offset += chunk_size;
}
args[0].u32w0 = PSMX_AM_REQ_READ | PSMX_AM_EOM;
args[0].u32w1 = len;
args[2].u64 = addr;
args[4].u64 = offset;
err = psm_am_request_short((psm_epaddr_t) src_addr,
PSMX_AM_RMA_HANDLER, args, 5, NULL, 0,
0, NULL, NULL);
return 0;
}
ssize_t _psmx_writeto(struct fid_ep *ep, const void *buf, size_t len,
void *desc, fi_addr_t dest_addr,
uint64_t addr, uint64_t key, void *context,
uint64_t flags, uint64_t data)
{
struct psmx_fid_ep *ep_priv;
struct psmx_fid_av *av;
struct psmx_epaddr_context *epaddr_context;
struct psmx_am_request *req;
psm_amarg_t args[8];
int nargs;
int am_flags = PSM_AM_FLAG_ASYNC;
int err;
int chunk_size;
psm_mq_req_t psm_req;
uint64_t psm_tag;
size_t idx;
if (flags & FI_TRIGGER) {
struct psmx_trigger *trigger;
struct fi_triggered_context *ctxt = context;
trigger = calloc(1, sizeof(*trigger));
if (!trigger)
return -ENOMEM;
trigger->op = PSMX_TRIGGERED_WRITE;
trigger->cntr = container_of(ctxt->threshold.cntr,
struct psmx_fid_cntr, cntr);
trigger->threshold = ctxt->threshold.threshold;
trigger->write.ep = ep;
trigger->write.buf = buf;
trigger->write.len = len;
trigger->write.desc = desc;
trigger->write.dest_addr = dest_addr;
trigger->write.addr = addr;
trigger->write.key = key;
trigger->write.context = context;
trigger->write.flags = flags & ~FI_TRIGGER;
trigger->write.data = data;
psmx_cntr_add_trigger(trigger->cntr, trigger);
return 0;
}
ep_priv = container_of(ep, struct psmx_fid_ep, ep);
assert(ep_priv->domain);
if (!buf)
return -EINVAL;
av = ep_priv->av;
if (av && av->type == FI_AV_TABLE) {
idx = dest_addr;
if (idx >= av->last)
return -EINVAL;
dest_addr = (fi_addr_t) av->psm_epaddrs[idx];
}
else if (!dest_addr) {
return -EINVAL;
}
epaddr_context = psm_epaddr_getctxt((void *)dest_addr);
if (epaddr_context->epid == ep_priv->domain->psm_epid)
return psmx_rma_self(PSMX_AM_REQ_WRITE,
ep_priv, (void *)buf, len, desc,
addr, key, context, flags, data);
if (flags & FI_INJECT) {
req = malloc(sizeof(*req) + len);
if (!req)
return -ENOMEM;
memset((void *)req, 0, sizeof(*req));
memcpy((void *)req + sizeof(*req), (void *)buf, len);
buf = (void *)req + sizeof(*req);
PSMX_CTXT_TYPE(&req->fi_context) = PSMX_INJECT_WRITE_CONTEXT;
req->no_event = 1;
}
else {
req = calloc(1, sizeof(*req));
if (!req)
return -ENOMEM;
if (ep_priv->send_cq_event_flag && !(flags & FI_EVENT)) {
PSMX_CTXT_TYPE(&req->fi_context) = PSMX_NOCOMP_WRITE_CONTEXT;
req->no_event = 1;
}
else {
PSMX_CTXT_TYPE(&req->fi_context) = PSMX_WRITE_CONTEXT;
}
}
req->op = PSMX_AM_REQ_WRITE;
req->write.buf = (void *)buf;
req->write.len = len;
req->write.addr = addr; /* needed? */
req->write.key = key; /* needed? */
req->write.context = context;
req->ep = ep_priv;
PSMX_CTXT_USER(&req->fi_context) = context;
chunk_size = MIN(PSMX_AM_CHUNK_SIZE, psmx_am_param.max_request_short);
if (psmx_env.tagged_rma && len > chunk_size) {
psm_tag = PSMX_RMA_BIT | ep_priv->domain->psm_epid;
args[0].u32w0 = PSMX_AM_REQ_WRITE_LONG;
args[0].u32w1 = len;
args[1].u64 = (uint64_t)req;
args[2].u64 = addr;
args[3].u64 = key;
args[4].u64 = psm_tag;
nargs = 5;
if (flags & FI_REMOTE_CQ_DATA) {
args[5].u64 = data;
args[0].u32w0 |= PSMX_AM_DATA;
nargs++;
}
err = psm_am_request_short((psm_epaddr_t) dest_addr,
PSMX_AM_RMA_HANDLER, args, nargs,
NULL, 0, am_flags | PSM_AM_FLAG_NOREPLY,
NULL, NULL);
psm_mq_isend(ep_priv->domain->psm_mq, (psm_epaddr_t) dest_addr,
0, psm_tag, buf, len, (void *)&req->fi_context, &psm_req);
return 0;
}
nargs = 4;
while (len > chunk_size) {
args[0].u32w0 = PSMX_AM_REQ_WRITE;
args[0].u32w1 = chunk_size;
args[1].u64 = (uint64_t)(uintptr_t)req;
args[2].u64 = addr;
args[3].u64 = key;
err = psm_am_request_short((psm_epaddr_t) dest_addr,
PSMX_AM_RMA_HANDLER, args, nargs,
(void *)buf, chunk_size,
am_flags | PSM_AM_FLAG_NOREPLY, NULL, NULL);
buf += chunk_size;
addr += chunk_size;
len -= chunk_size;
}
args[0].u32w0 = PSMX_AM_REQ_WRITE | PSMX_AM_EOM;
args[0].u32w1 = len;
args[1].u64 = (uint64_t)(uintptr_t)req;
args[2].u64 = addr;
args[3].u64 = key;
if (flags & FI_REMOTE_CQ_DATA) {
args[4].u64 = data;
args[0].u32w0 |= PSMX_AM_DATA;
nargs++;
}
err = psm_am_request_short((psm_epaddr_t) dest_addr,
PSMX_AM_RMA_HANDLER, args, nargs,
(void *)buf, len, am_flags, NULL, NULL);
return 0;
}
ssize_t _psmx_atomic_readwrite(struct fid_ep *ep,
const void *buf,
size_t count, void *desc,
void *result, void *result_desc,
fi_addr_t dest_addr,
uint64_t addr, uint64_t key,
enum fi_datatype datatype,
enum fi_op op, void *context,
uint64_t flags)
{
struct psmx_fid_ep *ep_priv;
struct psmx_fid_av *av;
struct psmx_epaddr_context *epaddr_context;
struct psmx_am_request *req;
psm_amarg_t args[8];
int am_flags = PSM_AM_FLAG_ASYNC;
int chunk_size, len;
size_t idx;
ep_priv = container_of(ep, struct psmx_fid_ep, ep);
if (flags & FI_TRIGGER) {
struct psmx_trigger *trigger;
struct fi_triggered_context *ctxt = context;
trigger = calloc(1, sizeof(*trigger));
if (!trigger)
return -FI_ENOMEM;
trigger->op = PSMX_TRIGGERED_ATOMIC_READWRITE;
trigger->cntr = container_of(ctxt->trigger.threshold.cntr,
struct psmx_fid_cntr, cntr);
trigger->threshold = ctxt->trigger.threshold.threshold;
trigger->atomic_readwrite.ep = ep;
trigger->atomic_readwrite.buf = buf;
trigger->atomic_readwrite.count = count;
trigger->atomic_readwrite.desc = desc;
trigger->atomic_readwrite.result = result;
trigger->atomic_readwrite.result_desc = result_desc;
trigger->atomic_readwrite.dest_addr = dest_addr;
trigger->atomic_readwrite.addr = addr;
trigger->atomic_readwrite.key = key;
trigger->atomic_readwrite.datatype = datatype;
trigger->atomic_readwrite.atomic_op = op;
trigger->atomic_readwrite.context = context;
trigger->atomic_readwrite.flags = flags & ~FI_TRIGGER;
psmx_cntr_add_trigger(trigger->cntr, trigger);
return 0;
}
if (!buf && op != FI_ATOMIC_READ)
return -FI_EINVAL;
if (datatype < 0 || datatype >= FI_DATATYPE_LAST)
return -FI_EINVAL;
if (op < 0 || op >= FI_ATOMIC_OP_LAST)
return -FI_EINVAL;
av = ep_priv->av;
if (av && av->type == FI_AV_TABLE) {
idx = dest_addr;
if (idx >= av->last)
return -FI_EINVAL;
dest_addr = (fi_addr_t) av->psm_epaddrs[idx];
}
else if (!dest_addr) {
return -FI_EINVAL;
}
epaddr_context = psm_epaddr_getctxt((void *)dest_addr);
if (epaddr_context->epid == ep_priv->domain->psm_epid)
return psmx_atomic_self(PSMX_AM_REQ_ATOMIC_READWRITE,
ep_priv, buf, count, desc,
NULL, NULL, result, result_desc,
addr, key, datatype, op,
context, flags);
chunk_size = MIN(PSMX_AM_CHUNK_SIZE, psmx_am_param.max_request_short);
len = fi_datatype_size(datatype) * count;
if (len > chunk_size)
return -FI_EMSGSIZE;
if ((flags & FI_INJECT) && op != FI_ATOMIC_READ) {
req = malloc(sizeof(*req) + len);
if (!req)
return -FI_ENOMEM;
memset((void *)req, 0, sizeof(*req));
memcpy((void *)req+sizeof(*req), (void *)buf, len);
buf = (void *)req + sizeof(*req);
}
else {
req = calloc(1, sizeof(*req));
if (!req)
return -FI_ENOMEM;
}
req->no_event = (flags & PSMX_NO_COMPLETION) ||
(ep_priv->send_selective_completion && !(flags & FI_COMPLETION));
req->op = PSMX_AM_REQ_ATOMIC_READWRITE;
req->atomic.buf = (void *)buf;
req->atomic.len = len;
req->atomic.addr = addr;
req->atomic.key = key;
req->atomic.context = context;
req->atomic.result = result;
req->ep = ep_priv;
if (op == FI_ATOMIC_READ)
req->cq_flags = FI_READ | FI_ATOMIC;
else
req->cq_flags = FI_WRITE | FI_ATOMIC;
args[0].u32w0 = PSMX_AM_REQ_ATOMIC_READWRITE;
args[0].u32w1 = count;
args[1].u64 = (uint64_t)(uintptr_t)req;
args[2].u64 = addr;
args[3].u64 = key;
args[4].u32w0 = datatype;
args[4].u32w1 = op;
psm_am_request_short((psm_epaddr_t) dest_addr,
PSMX_AM_ATOMIC_HANDLER, args, 5,
(void *)buf, (buf?len:0), am_flags, NULL, NULL);
return 0;
}
ssize_t _psmx_tagged_send(struct fid_ep *ep, const void *buf, size_t len,
void *desc, fi_addr_t dest_addr, uint64_t tag,
void *context, uint64_t flags)
#endif
{
struct psmx_fid_ep *ep_priv;
struct psmx_fid_av *av;
psm_epaddr_t psm_epaddr;
psm_mq_req_t psm_req;
uint64_t psm_tag;
#if (PSM_VERNO_MAJOR >= 2)
psm_mq_tag_t psm_tag2;
#endif
struct fi_context *fi_context;
int err;
size_t idx;
int no_completion = 0;
struct psmx_cq_event *event;
ep_priv = container_of(ep, struct psmx_fid_ep, ep);
if (flags & FI_TRIGGER) {
struct psmx_trigger *trigger;
struct fi_triggered_context *ctxt = context;
trigger = calloc(1, sizeof(*trigger));
if (!trigger)
return -FI_ENOMEM;
trigger->op = PSMX_TRIGGERED_TSEND;
trigger->cntr = container_of(ctxt->trigger.threshold.cntr,
struct psmx_fid_cntr, cntr);
trigger->threshold = ctxt->trigger.threshold.threshold;
trigger->tsend.ep = ep;
trigger->tsend.buf = buf;
trigger->tsend.len = len;
trigger->tsend.desc = desc;
trigger->tsend.dest_addr = dest_addr;
trigger->tsend.tag = tag;
trigger->tsend.context = context;
trigger->tsend.flags = flags & ~FI_TRIGGER;
#if (PSM_VERNO_MAJOR >= 2)
trigger->tsend.data = data;
#endif
psmx_cntr_add_trigger(trigger->cntr, trigger);
return 0;
}
if (tag & ep_priv->domain->reserved_tag_bits) {
FI_WARN(&psmx_prov, FI_LOG_EP_DATA,
"using reserved tag bits."
"tag=%lx. reserved_bits=%lx.\n", tag,
ep_priv->domain->reserved_tag_bits);
}
av = ep_priv->av;
if (av && av->type == FI_AV_TABLE) {
idx = (size_t)dest_addr;
if (idx >= av->last)
return -FI_EINVAL;
psm_epaddr = av->psm_epaddrs[idx];
}
else {
psm_epaddr = (psm_epaddr_t) dest_addr;
}
psm_tag = tag & (~ep_priv->domain->reserved_tag_bits);
#if (PSM_VERNO_MAJOR >= 2)
PSMX_SET_TAG(psm_tag2, psm_tag, data);
#endif
if ((flags & PSMX_NO_COMPLETION) ||
(ep_priv->send_selective_completion && !(flags & FI_COMPLETION)))
no_completion = 1;
if (flags & FI_INJECT) {
if (len > PSMX_INJECT_SIZE)
return -FI_EMSGSIZE;
#if (PSM_VERNO_MAJOR >= 2)
err = psm_mq_send2(ep_priv->domain->psm_mq, psm_epaddr, 0,
&psm_tag2, buf, len);
#else
err = psm_mq_send(ep_priv->domain->psm_mq, psm_epaddr, 0,
psm_tag, buf, len);
#endif
if (err != PSM_OK)
return psmx_errno(err);
if (ep_priv->send_cntr)
psmx_cntr_inc(ep_priv->send_cntr);
if (ep_priv->send_cq && !no_completion) {
event = psmx_cq_create_event(
ep_priv->send_cq,
context, (void *)buf, flags, len,
#if (PSM_VERNO_MAJOR >= 2)
(uint64_t) data, psm_tag,
#else
0 /* data */, psm_tag,
#endif
0 /* olen */,
0 /* err */);
if (event)
psmx_cq_enqueue_event(ep_priv->send_cq, event);
else
return -FI_ENOMEM;
}
return 0;
}
if (no_completion && !context) {
fi_context = &ep_priv->nocomp_send_context;
}
else {
if (!context)
return -FI_EINVAL;
fi_context = context;
PSMX_CTXT_TYPE(fi_context) = PSMX_TSEND_CONTEXT;
PSMX_CTXT_USER(fi_context) = (void *)buf;
PSMX_CTXT_EP(fi_context) = ep_priv;
}
#if (PSM_VERNO_MAJOR >= 2)
err = psm_mq_isend2(ep_priv->domain->psm_mq, psm_epaddr, 0,
&psm_tag2, buf, len, (void*)fi_context, &psm_req);
#else
err = psm_mq_isend(ep_priv->domain->psm_mq, psm_epaddr, 0,
psm_tag, buf, len, (void*)fi_context, &psm_req);
#endif
if (err != PSM_OK)
return psmx_errno(err);
if (fi_context == context)
PSMX_CTXT_REQ(fi_context) = psm_req;
return 0;
}
ssize_t _psmx_tagged_recv(struct fid_ep *ep, void *buf, size_t len,
void *desc, fi_addr_t src_addr,
uint64_t tag, uint64_t ignore,
void *context, uint64_t flags)
{
struct psmx_fid_ep *ep_priv;
psm_mq_req_t psm_req;
uint64_t psm_tag, psm_tagsel;
#if (PSM_VERNO_MAJOR >= 2)
psm_mq_tag_t psm_tag2, psm_tagsel2;
struct psmx_fid_av *av;
size_t idx;
#endif
struct fi_context *fi_context;
int err;
ep_priv = container_of(ep, struct psmx_fid_ep, ep);
if (flags & FI_PEEK)
return _psmx_tagged_peek(ep, buf, len, desc, src_addr,
tag, ignore, context, flags);
if (flags & FI_TRIGGER) {
struct psmx_trigger *trigger;
struct fi_triggered_context *ctxt = context;
trigger = calloc(1, sizeof(*trigger));
if (!trigger)
return -FI_ENOMEM;
trigger->op = PSMX_TRIGGERED_TRECV;
trigger->cntr = container_of(ctxt->trigger.threshold.cntr,
struct psmx_fid_cntr, cntr);
trigger->threshold = ctxt->trigger.threshold.threshold;
trigger->trecv.ep = ep;
trigger->trecv.buf = buf;
trigger->trecv.len = len;
trigger->trecv.desc = desc;
trigger->trecv.src_addr = src_addr;
trigger->trecv.tag = tag;
trigger->trecv.ignore = ignore;
trigger->trecv.context = context;
trigger->trecv.flags = flags & ~FI_TRIGGER;
psmx_cntr_add_trigger(trigger->cntr, trigger);
return 0;
}
#if (PSM_VERNO_MAJOR >= 2)
if (flags & FI_CLAIM) {
if (!context)
return -FI_EINVAL;
/* TODO: handle FI_DISCARD */
fi_context = context;
psm_req = PSMX_CTXT_REQ(fi_context);
PSMX_CTXT_TYPE(fi_context) = PSMX_TRECV_CONTEXT;
PSMX_CTXT_USER(fi_context) = buf;
PSMX_CTXT_EP(fi_context) = ep_priv;
err = psm_mq_imrecv(ep_priv->domain->psm_mq, 0, /*flags*/
buf, len, context, &psm_req);
if (err != PSM_OK)
return psmx_errno(err);
PSMX_CTXT_REQ(fi_context) = psm_req;
return 0;
}
#endif
if (tag & ep_priv->domain->reserved_tag_bits) {
FI_WARN(&psmx_prov, FI_LOG_EP_DATA,
"using reserved tag bits."
"tag=%lx. reserved_bits=%lx.\n", tag,
ep_priv->domain->reserved_tag_bits);
}
psm_tag = tag & (~ep_priv->domain->reserved_tag_bits);
psm_tagsel = (~ignore) | ep_priv->domain->reserved_tag_bits;
if (ep_priv->recv_selective_completion && !(flags & FI_COMPLETION) && !context) {
fi_context = &ep_priv->nocomp_recv_context;
}
else {
if (!context)
return -FI_EINVAL;
fi_context = context;
PSMX_CTXT_TYPE(fi_context) = PSMX_TRECV_CONTEXT;
PSMX_CTXT_USER(fi_context) = buf;
PSMX_CTXT_EP(fi_context) = ep_priv;
}
#if (PSM_VERNO_MAJOR >= 2)
if ((ep_priv->caps & FI_DIRECTED_RECV) && src_addr != FI_ADDR_UNSPEC) {
av = ep_priv->av;
if (av && av->type == FI_AV_TABLE) {
idx = (size_t)src_addr;
if (idx >= av->last)
return -FI_EINVAL;
src_addr = (fi_addr_t)av->psm_epaddrs[idx];
}
}
else {
src_addr = 0;
}
PSMX_SET_TAG(psm_tag2, psm_tag, 0);
PSMX_SET_TAG(psm_tagsel2, psm_tagsel, 0);
err = psm_mq_irecv2(ep_priv->domain->psm_mq,
(psm_epaddr_t)src_addr,
&psm_tag2, &psm_tagsel2, 0, /* flags */
buf, len, (void *)fi_context, &psm_req);
#else
err = psm_mq_irecv(ep_priv->domain->psm_mq,
psm_tag, psm_tagsel, 0, /* flags */
buf, len, (void *)fi_context, &psm_req);
#endif
if (err != PSM_OK)
return psmx_errno(err);
if (fi_context == context)
PSMX_CTXT_REQ(fi_context) = psm_req;
return 0;
}
ssize_t _psmx_tagged_recv(struct fid_ep *ep, void *buf, size_t len,
void *desc, fi_addr_t src_addr,
uint64_t tag, uint64_t ignore,
void *context, uint64_t flags)
{
struct psmx_fid_ep *ep_priv;
psm_mq_req_t psm_req;
uint64_t psm_tag, psm_tagsel;
struct fi_context *fi_context;
int err;
ep_priv = container_of(ep, struct psmx_fid_ep, ep);
if (flags & FI_PEEK)
return _psmx_tagged_peek(ep, buf, len, desc, src_addr,
tag, ignore, context, flags);
if (flags & FI_TRIGGER) {
struct psmx_trigger *trigger;
struct fi_triggered_context *ctxt = context;
trigger = calloc(1, sizeof(*trigger));
if (!trigger)
return -FI_ENOMEM;
trigger->op = PSMX_TRIGGERED_TRECV;
trigger->cntr = container_of(ctxt->trigger.threshold.cntr,
struct psmx_fid_cntr, cntr);
trigger->threshold = ctxt->trigger.threshold.threshold;
trigger->trecv.ep = ep;
trigger->trecv.buf = buf;
trigger->trecv.len = len;
trigger->trecv.desc = desc;
trigger->trecv.src_addr = src_addr;
trigger->trecv.tag = tag;
trigger->trecv.ignore = ignore;
trigger->trecv.context = context;
trigger->trecv.flags = flags & ~FI_TRIGGER;
psmx_cntr_add_trigger(trigger->cntr, trigger);
return 0;
}
if (tag & ep_priv->domain->reserved_tag_bits) {
FI_WARN(&psmx_prov, FI_LOG_EP_DATA,
"using reserved tag bits."
"tag=%lx. reserved_bits=%lx.\n", tag,
ep_priv->domain->reserved_tag_bits);
}
psm_tag = tag & (~ep_priv->domain->reserved_tag_bits);
psm_tagsel = (~ignore) | ep_priv->domain->reserved_tag_bits;
if (ep_priv->recv_selective_completion && !(flags & FI_COMPLETION) && !context) {
fi_context = &ep_priv->nocomp_recv_context;
}
else {
if (!context)
return -FI_EINVAL;
fi_context = context;
PSMX_CTXT_TYPE(fi_context) = PSMX_TRECV_CONTEXT;
PSMX_CTXT_USER(fi_context) = buf;
PSMX_CTXT_EP(fi_context) = ep_priv;
}
err = psm_mq_irecv(ep_priv->domain->psm_mq,
psm_tag, psm_tagsel, 0, /* flags */
buf, len, (void *)fi_context, &psm_req);
if (err != PSM_OK)
return psmx_errno(err);
if (fi_context == context)
PSMX_CTXT_REQ(fi_context) = psm_req;
return 0;
}
ssize_t _psmx_write(struct fid_ep *ep, const void *buf, size_t len,
void *desc, fi_addr_t dest_addr,
uint64_t addr, uint64_t key, void *context,
uint64_t flags, uint64_t data)
{
struct psmx_fid_ep *ep_priv;
struct psmx_fid_av *av;
struct psmx_epaddr_context *epaddr_context;
struct psmx_am_request *req;
psm_amarg_t args[8];
int nargs;
int am_flags = PSM_AM_FLAG_ASYNC;
int chunk_size;
psm_mq_req_t psm_req;
uint64_t psm_tag;
size_t idx;
void *psm_context;
int no_event;
ep_priv = container_of(ep, struct psmx_fid_ep, ep);
if (flags & FI_TRIGGER) {
struct psmx_trigger *trigger;
struct fi_triggered_context *ctxt = context;
trigger = calloc(1, sizeof(*trigger));
if (!trigger)
return -FI_ENOMEM;
trigger->op = PSMX_TRIGGERED_WRITE;
trigger->cntr = container_of(ctxt->trigger.threshold.cntr,
struct psmx_fid_cntr, cntr);
trigger->threshold = ctxt->trigger.threshold.threshold;
trigger->write.ep = ep;
trigger->write.buf = buf;
trigger->write.len = len;
trigger->write.desc = desc;
trigger->write.dest_addr = dest_addr;
trigger->write.addr = addr;
trigger->write.key = key;
trigger->write.context = context;
trigger->write.flags = flags & ~FI_TRIGGER;
trigger->write.data = data;
psmx_cntr_add_trigger(trigger->cntr, trigger);
return 0;
}
if (!buf)
return -FI_EINVAL;
av = ep_priv->av;
if (av && av->type == FI_AV_TABLE) {
idx = dest_addr;
if (idx >= av->last)
return -FI_EINVAL;
dest_addr = (fi_addr_t) av->psm_epaddrs[idx];
}
else if (!dest_addr) {
return -FI_EINVAL;
}
epaddr_context = psm_epaddr_getctxt((void *)dest_addr);
if (epaddr_context->epid == ep_priv->domain->psm_epid)
return psmx_rma_self(PSMX_AM_REQ_WRITE,
ep_priv, (void *)buf, len, desc,
addr, key, context, flags, data);
no_event = (flags & PSMX_NO_COMPLETION) ||
(ep_priv->send_selective_completion && !(flags & FI_COMPLETION));
if (flags & FI_INJECT) {
if (len > PSMX_INJECT_SIZE)
return -FI_EMSGSIZE;
req = malloc(sizeof(*req) + len);
if (!req)
return -FI_ENOMEM;
memset((void *)req, 0, sizeof(*req));
memcpy((void *)req + sizeof(*req), (void *)buf, len);
buf = (void *)req + sizeof(*req);
}
else {
req = calloc(1, sizeof(*req));
if (!req)
return -FI_ENOMEM;
PSMX_CTXT_TYPE(&req->fi_context) = no_event ?
PSMX_NOCOMP_WRITE_CONTEXT :
PSMX_WRITE_CONTEXT;
}
req->no_event = no_event;
req->op = PSMX_AM_REQ_WRITE;
req->write.buf = (void *)buf;
req->write.len = len;
req->write.addr = addr; /* needed? */
req->write.key = key; /* needed? */
req->write.context = context;
req->ep = ep_priv;
req->cq_flags = FI_WRITE | FI_RMA;
PSMX_CTXT_USER(&req->fi_context) = context;
PSMX_CTXT_EP(&req->fi_context) = ep_priv;
chunk_size = MIN(PSMX_AM_CHUNK_SIZE, psmx_am_param.max_request_short);
if (psmx_env.tagged_rma && len > chunk_size) {
void *payload = NULL;
int payload_len = 0;
psm_tag = PSMX_RMA_BIT | ep_priv->domain->psm_epid;
args[0].u32w0 = PSMX_AM_REQ_WRITE_LONG;
args[0].u32w1 = len;
args[1].u64 = (uint64_t)req;
args[2].u64 = addr;
args[3].u64 = key;
args[4].u64 = psm_tag;
nargs = 5;
if (flags & FI_REMOTE_CQ_DATA) {
args[0].u32w0 |= PSMX_AM_DATA;
payload = &data;
payload_len = sizeof(data);
am_flags = 0;
}
if (flags & FI_DELIVERY_COMPLETE) {
args[0].u32w0 |= PSMX_AM_FORCE_ACK;
psm_context = NULL;
}
else {
psm_context = (void *)&req->fi_context;
}
/* NOTE: if nargs is greater than 5, the following psm_mq_isend
* would hang if the destination is on the same node (i.e. going
* through the shared memory path). As the result, the immediate
* data is sent as payload instead of args[5].
*/
psm_am_request_short((psm_epaddr_t) dest_addr,
PSMX_AM_RMA_HANDLER, args, nargs,
payload, payload_len, am_flags,
NULL, NULL);
psm_mq_isend(ep_priv->domain->psm_mq, (psm_epaddr_t) dest_addr,
0, psm_tag, buf, len, psm_context, &psm_req);
return 0;
}
nargs = 4;
while (len > chunk_size) {
args[0].u32w0 = PSMX_AM_REQ_WRITE;
args[0].u32w1 = chunk_size;
args[1].u64 = (uint64_t)(uintptr_t)req;
args[2].u64 = addr;
args[3].u64 = key;
psm_am_request_short((psm_epaddr_t) dest_addr,
PSMX_AM_RMA_HANDLER, args, nargs,
(void *)buf, chunk_size,
am_flags, NULL, NULL);
buf += chunk_size;
addr += chunk_size;
len -= chunk_size;
}
args[0].u32w0 = PSMX_AM_REQ_WRITE | PSMX_AM_EOM;
args[0].u32w1 = len;
args[1].u64 = (uint64_t)(uintptr_t)req;
args[2].u64 = addr;
args[3].u64 = key;
if (flags & FI_REMOTE_CQ_DATA) {
args[4].u64 = data;
args[0].u32w0 |= PSMX_AM_DATA;
nargs++;
}
psm_am_request_short((psm_epaddr_t) dest_addr,
PSMX_AM_RMA_HANDLER, args, nargs,
(void *)buf, len, am_flags, NULL, NULL);
return 0;
}
