/* op_flags=0, FI_SELECTIVE_COMPLETION not set, FI_AV_TABLE */
static ssize_t
psmx2_tagged_recv_no_flag_av_table(struct fid_ep *ep, void *buf, size_t len,
void *desc, fi_addr_t src_addr,
uint64_t tag, uint64_t ignore,
void *context)
{
struct psmx2_fid_ep *ep_priv;
struct psmx2_fid_av *av;
psm2_epaddr_t psm2_epaddr;
uint8_t vlane;
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag, psm2_tagsel;
uint32_t tag32, tagsel32;
struct fi_context *fi_context;
size_t idx;
int err;
ep_priv = container_of(ep, struct psmx2_fid_ep, ep);
fi_context = context;
PSMX2_CTXT_TYPE(fi_context) = PSMX2_TRECV_CONTEXT;
PSMX2_CTXT_USER(fi_context) = buf;
PSMX2_CTXT_EP(fi_context) = ep_priv;
PSMX2_CTXT_SIZE(fi_context) = len;
if ((ep_priv->caps & FI_DIRECTED_RECV) && src_addr != FI_ADDR_UNSPEC) {
av = ep_priv->av;
if (av && PSMX2_SEP_ADDR_TEST(src_addr)) {
psm2_epaddr = psmx2_av_translate_sep(av, ep_priv->trx_ctxt, src_addr);
vlane = 0;
} else {
idx = (size_t)src_addr;
if (idx >= av->last)
return -FI_EINVAL;
psm2_epaddr = av->epaddrs[idx];
vlane = av->vlanes[idx];
}
tag32 = PSMX2_TAG32(0, vlane, ep_priv->vlane);
tagsel32 = ~PSMX2_IOV_BIT;
} else {
psm2_epaddr = 0;
tag32 = PSMX2_TAG32(0, 0, ep_priv->vlane);
tagsel32 = ~(PSMX2_IOV_BIT | PSMX2_SRC_BITS);
}
PSMX2_SET_TAG(psm2_tag, tag, tag32);
PSMX2_SET_TAG(psm2_tagsel, ~ignore, tagsel32);
err = psm2_mq_irecv2(ep_priv->trx_ctxt->psm2_mq, psm2_epaddr,
&psm2_tag, &psm2_tagsel, 0, buf, len,
(void *)fi_context, &psm2_req);
if (err != PSM2_OK)
return psmx2_errno(err);
PSMX2_CTXT_REQ(fi_context) = psm2_req;
return 0;
}
/* op_flags=0, FI_SELECTIVE_COMPLETION set, FI_AV_TABLE */
static ssize_t
psmx2_tagged_recv_no_event_av_table(struct fid_ep *ep, void *buf, size_t len,
void *desc, fi_addr_t src_addr,
uint64_t tag, uint64_t ignore,
void *context)
{
struct psmx2_fid_ep *ep_priv;
struct psmx2_fid_av *av;
psm2_epaddr_t psm2_epaddr;
uint8_t vlane;
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag, psm2_tagsel;
uint32_t tag32, tagsel32;
struct fi_context *fi_context;
size_t idx;
int err;
ep_priv = container_of(ep, struct psmx2_fid_ep, ep);
fi_context = psmx2_ep_get_op_context(ep_priv);
PSMX2_CTXT_TYPE(fi_context) = PSMX2_NOCOMP_RECV_CONTEXT_ALLOC;
PSMX2_CTXT_EP(fi_context) = ep_priv;
PSMX2_CTXT_USER(fi_context) = buf;
PSMX2_CTXT_SIZE(fi_context) = len;
if ((ep_priv->caps & FI_DIRECTED_RECV) && src_addr != FI_ADDR_UNSPEC) {
av = ep_priv->av;
idx = (size_t)src_addr;
if (idx >= av->last)
return -FI_EINVAL;
psm2_epaddr = av->epaddrs[idx];
vlane = av->vlanes[idx];
tag32 = PSMX2_TAG32(0, vlane, ep_priv->vlane);
tagsel32 = ~PSMX2_IOV_BIT;
} else {
psm2_epaddr = 0;
tag32 = PSMX2_TAG32(0, 0, ep_priv->vlane);
tagsel32 = ~(PSMX2_IOV_BIT | PSMX2_SRC_BITS);
}
PSMX2_SET_TAG(psm2_tag, tag, tag32);
PSMX2_SET_TAG(psm2_tagsel, ~ignore, tagsel32);
err = psm2_mq_irecv2(ep_priv->domain->psm2_mq, psm2_epaddr,
&psm2_tag, &psm2_tagsel, 0, buf, len,
(void *)fi_context, &psm2_req);
return psmx2_errno(err);
}
/* op_flags=0, FI_SELECTIVE_COMPLETION set, FI_AV_MAP */
static ssize_t
psmx2_tagged_recv_no_event_av_map(struct fid_ep *ep, void *buf, size_t len,
void *desc, fi_addr_t src_addr,
uint64_t tag, uint64_t ignore,
void *context)
{
struct psmx2_fid_ep *ep_priv;
struct psmx2_fid_av *av;
psm2_epaddr_t psm2_epaddr;
uint8_t vlane;
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag, psm2_tagsel;
uint32_t tag32, tagsel32;
struct fi_context *fi_context;
int err;
ep_priv = container_of(ep, struct psmx2_fid_ep, ep);
fi_context = psmx2_ep_get_op_context(ep_priv);
PSMX2_CTXT_TYPE(fi_context) = PSMX2_NOCOMP_RECV_CONTEXT_ALLOC;
PSMX2_CTXT_EP(fi_context) = ep_priv;
PSMX2_CTXT_USER(fi_context) = buf;
PSMX2_CTXT_SIZE(fi_context) = len;
if ((ep_priv->caps & FI_DIRECTED_RECV) && src_addr != FI_ADDR_UNSPEC) {
av = ep_priv->av;
if (av && PSMX2_SEP_ADDR_TEST(src_addr)) {
psm2_epaddr = psmx2_av_translate_sep(av, ep_priv->trx_ctxt, src_addr);
vlane = 0;
} else {
psm2_epaddr = PSMX2_ADDR_TO_EP(src_addr);
vlane = PSMX2_ADDR_TO_VL(src_addr);
}
tag32 = PSMX2_TAG32(0, vlane, ep_priv->vlane);
tagsel32 = ~PSMX2_IOV_BIT;
} else {
psm2_epaddr = 0;
tag32 = PSMX2_TAG32(0, 0, ep_priv->vlane);
tagsel32 = ~(PSMX2_IOV_BIT | PSMX2_SRC_BITS);
}
PSMX2_SET_TAG(psm2_tag, tag, tag32);
PSMX2_SET_TAG(psm2_tagsel, ~ignore, tagsel32);
err = psm2_mq_irecv2(ep_priv->trx_ctxt->psm2_mq, psm2_epaddr,
&psm2_tag, &psm2_tagsel, 0, buf, len,
(void *)fi_context, &psm2_req);
return psmx2_errno(err);
}
/* op_flags=0, FI_SELECTIVE_COMPLETION not set, FI_AV_TABLE */
static ssize_t
psmx2_tagged_send_no_flag_av_table(struct fid_ep *ep, const void *buf,
size_t len, void *desc,
fi_addr_t dest_addr, uint64_t tag,
void *context)
{
struct psmx2_fid_ep *ep_priv;
struct psmx2_fid_av *av;
psm2_epaddr_t psm2_epaddr;
uint8_t vlane;
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag;
uint32_t tag32;
struct fi_context *fi_context;
size_t idx;
int err;
ep_priv = container_of(ep, struct psmx2_fid_ep, ep);
av = ep_priv->av;
if (av && PSMX2_SEP_ADDR_TEST(dest_addr)) {
psm2_epaddr = psmx2_av_translate_sep(av, ep_priv->trx_ctxt, dest_addr);
vlane = 0;
} else {
idx = (size_t)dest_addr;
if (idx >= av->last)
return -FI_EINVAL;
psm2_epaddr = av->epaddrs[idx];
vlane = av->vlanes[idx];
}
tag32 = PSMX2_TAG32(0, ep_priv->vlane, vlane);
PSMX2_SET_TAG(psm2_tag, tag, tag32);
fi_context = context;
PSMX2_CTXT_TYPE(fi_context) = PSMX2_TSEND_CONTEXT;
PSMX2_CTXT_USER(fi_context) = (void *)buf;
PSMX2_CTXT_EP(fi_context) = ep_priv;
err = psm2_mq_isend2(ep_priv->trx_ctxt->psm2_mq, psm2_epaddr, 0,
&psm2_tag, buf, len, (void*)fi_context,
&psm2_req);
if (err != PSM2_OK)
return psmx2_errno(err);
PSMX2_CTXT_REQ(fi_context) = psm2_req;
return 0;
}
/* op_flags=0, FI_SELECTIVE_COMPLETION not set, FI_AV_MAP */
static ssize_t
psmx2_tagged_recv_no_flag_av_map(struct fid_ep *ep, void *buf, size_t len,
void *desc, fi_addr_t src_addr,
uint64_t tag, uint64_t ignore,
void *context)
{
struct psmx2_fid_ep *ep_priv;
psm2_epaddr_t psm2_epaddr;
uint8_t vlane;
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag, psm2_tagsel;
uint32_t tag32, tagsel32;
struct fi_context *fi_context;
int err;
ep_priv = container_of(ep, struct psmx2_fid_ep, ep);
fi_context = context;
PSMX2_CTXT_TYPE(fi_context) = PSMX2_TRECV_CONTEXT;
PSMX2_CTXT_USER(fi_context) = buf;
PSMX2_CTXT_EP(fi_context) = ep_priv;
PSMX2_CTXT_SIZE(fi_context) = len;
if ((ep_priv->caps & FI_DIRECTED_RECV) && src_addr != FI_ADDR_UNSPEC) {
psm2_epaddr = PSMX2_ADDR_TO_EP(src_addr);
vlane = PSMX2_ADDR_TO_VL(src_addr);
tag32 = PSMX2_TAG32(0, vlane, ep_priv->vlane);
tagsel32 = ~PSMX2_IOV_BIT;
} else {
psm2_epaddr = 0;
tag32 = PSMX2_TAG32(0, 0, ep_priv->vlane);
tagsel32 = ~(PSMX2_IOV_BIT | PSMX2_SRC_BITS);
}
PSMX2_SET_TAG(psm2_tag, tag, tag32);
PSMX2_SET_TAG(psm2_tagsel, ~ignore, tagsel32);
err = psm2_mq_irecv2(ep_priv->domain->psm2_mq, psm2_epaddr,
&psm2_tag, &psm2_tagsel, 0, buf, len,
(void *)fi_context, &psm2_req);
if (err != PSM2_OK)
return psmx2_errno(err);
PSMX2_CTXT_REQ(fi_context) = psm2_req;
return 0;
}
/* op_flags=0, FI_SELECTIVE_COMPLETION not set, FI_AV_MAP */
static ssize_t
psmx2_tagged_send_no_flag_av_map(struct fid_ep *ep, const void *buf,
size_t len, void *desc,
fi_addr_t dest_addr, uint64_t tag,
void *context)
{
struct psmx2_fid_ep *ep_priv;
psm2_epaddr_t psm2_epaddr;
uint8_t vlane;
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag;
uint32_t tag32;
struct fi_context *fi_context;
int err;
ep_priv = container_of(ep, struct psmx2_fid_ep, ep);
psm2_epaddr = PSMX2_ADDR_TO_EP(dest_addr);
vlane = PSMX2_ADDR_TO_VL(dest_addr);
tag32 = PSMX2_TAG32(0, ep_priv->vlane, vlane);
PSMX2_SET_TAG(psm2_tag, tag, tag32);
fi_context = context;
PSMX2_CTXT_TYPE(fi_context) = PSMX2_TSEND_CONTEXT;
PSMX2_CTXT_USER(fi_context) = (void *)buf;
PSMX2_CTXT_EP(fi_context) = ep_priv;
err = psm2_mq_isend2(ep_priv->domain->psm2_mq, psm2_epaddr, 0,
&psm2_tag, buf, len, (void*)fi_context,
&psm2_req);
if (err != PSM2_OK)
return psmx2_errno(err);
PSMX2_CTXT_REQ(fi_context) = psm2_req;
return 0;
}
ssize_t psmx2_read_generic(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 psmx2_fid_ep *ep_priv;
struct psmx2_fid_av *av;
struct psmx2_epaddr_context *epaddr_context;
struct psmx2_am_request *req;
psm2_amarg_t args[8];
int chunk_size;
size_t offset = 0;
psm2_epaddr_t psm2_epaddr;
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag, psm2_tagsel;
size_t idx;
int err;
ep_priv = container_of(ep, struct psmx2_fid_ep, ep);
if (flags & FI_TRIGGER)
return psmx2_trigger_queue_read(ep, buf, len, desc, src_addr,
addr, key, context, flags);
if (!buf)
return -FI_EINVAL;
av = ep_priv->av;
if (av && PSMX2_SEP_ADDR_TEST(src_addr)) {
psm2_epaddr = psmx2_av_translate_sep(av, ep_priv->tx, src_addr);
} else if (av && av->type == FI_AV_TABLE) {
idx = src_addr;
if ((err = psmx2_av_check_table_idx(av, ep_priv->tx, idx)))
return err;
psm2_epaddr = av->tables[ep_priv->tx->id].epaddrs[idx];
} else {
if (!src_addr)
return -FI_EINVAL;
psm2_epaddr = PSMX2_ADDR_TO_EP(src_addr);
}
epaddr_context = psm2_epaddr_getctxt((void *)psm2_epaddr);
if (epaddr_context->epid == ep_priv->tx->psm2_epid)
return psmx2_rma_self(PSMX2_AM_REQ_READ, ep_priv,
buf, len, desc, addr, key,
context, flags, 0);
req = psmx2_am_request_alloc(ep_priv->tx);
if (!req)
return -FI_ENOMEM;
req->op = PSMX2_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;
req->cq_flags = FI_READ | FI_RMA;
PSMX2_CTXT_TYPE(&req->fi_context) = PSMX2_READ_CONTEXT;
PSMX2_CTXT_USER(&req->fi_context) = context;
PSMX2_CTXT_EP(&req->fi_context) = ep_priv;
if (ep_priv->send_selective_completion && !(flags & FI_COMPLETION)) {
PSMX2_CTXT_TYPE(&req->fi_context) = PSMX2_NOCOMP_READ_CONTEXT;
req->no_event = 1;
}
chunk_size = ep_priv->tx->psm2_am_param.max_reply_short;
args[0].u32w0 = 0;
if (psmx2_env.tagged_rma && len > chunk_size) {
PSMX2_SET_TAG(psm2_tag, (uint64_t)req, 0, PSMX2_RMA_TYPE_READ);
PSMX2_SET_MASK(psm2_tagsel, PSMX2_MATCH_ALL, PSMX2_RMA_TYPE_MASK);
psm2_mq_irecv2(ep_priv->tx->psm2_mq, psm2_epaddr,
&psm2_tag, &psm2_tagsel, 0, buf, len,
(void *)&req->fi_context, &psm2_req);
PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_READ_LONG);
args[0].u32w1 = len;
args[1].u64 = (uint64_t)req;
args[2].u64 = addr;
args[3].u64 = key;
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER,
args, 4, NULL, 0, 0, NULL, NULL);
return 0;
}
PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_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;
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER,
args, 5, NULL, 0, 0, NULL, NULL);
addr += chunk_size;
len -= chunk_size;
offset += chunk_size;
}
PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_EOM);
args[0].u32w1 = len;
args[2].u64 = addr;
args[4].u64 = offset;
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER,
args, 5, NULL, 0, 0, NULL, NULL);
return 0;
}
ssize_t psmx2_writev_generic(struct fid_ep *ep, const struct iovec *iov,
void **desc, size_t count, fi_addr_t dest_addr,
uint64_t addr, uint64_t key, void *context,
uint64_t flags, uint64_t data)
{
struct psmx2_fid_ep *ep_priv;
struct psmx2_fid_av *av;
struct psmx2_epaddr_context *epaddr_context;
struct psmx2_am_request *req;
psm2_amarg_t args[8];
int nargs;
int am_flags = PSM2_AM_FLAG_ASYNC;
int chunk_size;
psm2_epaddr_t psm2_epaddr;
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag;
size_t idx;
void *psm2_context;
int no_event;
size_t total_len, len, len_sent;
uint8_t *buf, *p;
int i;
int err;
ep_priv = container_of(ep, struct psmx2_fid_ep, ep);
if (flags & FI_TRIGGER)
return psmx2_trigger_queue_writev(ep, iov, desc, count,
dest_addr, addr, key,
context, flags, data);
av = ep_priv->av;
if (av && PSMX2_SEP_ADDR_TEST(dest_addr)) {
psm2_epaddr = psmx2_av_translate_sep(av, ep_priv->tx, dest_addr);
} else if (av && av->type == FI_AV_TABLE) {
idx = dest_addr;
if ((err = psmx2_av_check_table_idx(av, ep_priv->tx, idx)))
return err;
psm2_epaddr = av->tables[ep_priv->tx->id].epaddrs[idx];
} else {
if (!dest_addr)
return -FI_EINVAL;
psm2_epaddr = PSMX2_ADDR_TO_EP(dest_addr);
}
epaddr_context = psm2_epaddr_getctxt((void *)psm2_epaddr);
if (epaddr_context->epid == ep_priv->tx->psm2_epid)
return psmx2_rma_self(PSMX2_AM_REQ_WRITEV, ep_priv,
(void *)iov, count, desc, addr,
key, context, flags, data);
no_event = (flags & PSMX2_NO_COMPLETION) ||
(ep_priv->send_selective_completion && !(flags & FI_COMPLETION));
total_len = 0;
for (i=0; i<count; i++)
total_len += iov[i].iov_len;
chunk_size = ep_priv->tx->psm2_am_param.max_request_short;
req = psmx2_am_request_alloc(ep_priv->tx);
if (!req)
return -FI_ENOMEM;
/* Case 1: fit into a AM message, then pack and send */
if (total_len <= chunk_size) {
req->tmpbuf = malloc(total_len);
if (!req->tmpbuf) {
psmx2_am_request_free(ep_priv->tx, req);
return -FI_ENOMEM;
}
p = req->tmpbuf;
for (i=0; i<count; i++) {
if (iov[i].iov_len) {
memcpy(p, iov[i].iov_base, iov[i].iov_len);
p += iov[i].iov_len;
}
}
buf = req->tmpbuf;
len = total_len;
req->no_event = no_event;
req->op = PSMX2_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;
PSMX2_CTXT_USER(&req->fi_context) = context;
PSMX2_CTXT_EP(&req->fi_context) = ep_priv;
args[0].u32w0 = 0;
PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_WRITE);
args[0].u32w1 = len;
args[1].u64 = (uint64_t)(uintptr_t)req;
args[2].u64 = addr;
args[3].u64 = key;
nargs = 4;
if (flags & FI_REMOTE_CQ_DATA) {
PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_DATA | PSMX2_AM_EOM);
args[4].u64 = data;
nargs++;
} else {
PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_EOM);
}
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, nargs,
(void *)buf, len, am_flags, NULL, NULL);
return 0;
}
if (flags & FI_INJECT) {
psmx2_am_request_free(ep_priv->tx, req);
return -FI_EMSGSIZE;
}
PSMX2_CTXT_TYPE(&req->fi_context) = no_event ?
PSMX2_NOCOMP_WRITE_CONTEXT :
PSMX2_WRITE_CONTEXT;
req->no_event = no_event;
req->op = PSMX2_AM_REQ_WRITE;
req->write.buf = (void *)iov[0].iov_base;
req->write.len = total_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;
PSMX2_CTXT_USER(&req->fi_context) = context;
PSMX2_CTXT_EP(&req->fi_context) = ep_priv;
/* Case 2: send iov in sequence */
args[0].u32w0 = 0;
len_sent = 0;
for (i=0; i<count; i++) {
if (!iov[i].iov_len)
continue;
/* Case 2.1: use long protocol for the last segment if it is large */
if (psmx2_env.tagged_rma && iov[i].iov_len > chunk_size &&
len_sent + iov[i].iov_len == total_len) {
PSMX2_SET_TAG(psm2_tag, (uint64_t)req, 0, PSMX2_RMA_TYPE_WRITE);
PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_WRITE_LONG);
args[0].u32w1 = iov[i].iov_len;
args[1].u64 = (uint64_t)req;
args[2].u64 = addr;
args[3].u64 = key;
nargs = 4;
if (flags & FI_REMOTE_CQ_DATA) {
PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_DATA);
args[4].u64 = data;
nargs++;
}
if (flags & FI_DELIVERY_COMPLETE) {
args[0].u32w0 |= PSMX2_AM_FORCE_ACK;
psm2_context = NULL;
} else {
psm2_context = (void *)&req->fi_context;
}
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args,
nargs, NULL, 0, am_flags, NULL, NULL);
psm2_mq_isend2(ep_priv->tx->psm2_mq, psm2_epaddr, 0,
&psm2_tag, iov[i].iov_base, iov[i].iov_len,
psm2_context, &psm2_req);
return 0;
}
/* Case 2.2: use short protocol all other segments */
PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_WRITE);
nargs = 4;
buf = iov[i].iov_base;
len = iov[i].iov_len;
while (len > chunk_size) {
args[0].u32w1 = chunk_size;
args[1].u64 = (uint64_t)(uintptr_t)req;
args[2].u64 = addr;
args[3].u64 = key;
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args,
nargs, (void *)buf, chunk_size, am_flags,
NULL, NULL);
buf += chunk_size;
addr += chunk_size;
len -= chunk_size;
len_sent += chunk_size;
}
args[0].u32w1 = len;
args[1].u64 = (uint64_t)(uintptr_t)req;
args[2].u64 = addr;
args[3].u64 = key;
if (len_sent + len == total_len) {
if (flags & FI_REMOTE_CQ_DATA) {
PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_DATA | PSMX2_AM_EOM);
args[4].u64 = data;
nargs++;
} else {
PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_EOM);
}
}
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, nargs,
(void *)buf, len, am_flags, NULL, NULL);
addr += len;
len_sent += len;
}
return 0;
}
int psmx2_ep_open(struct fid_domain *domain, struct fi_info *info,
struct fid_ep **ep, void *context)
{
struct psmx2_fid_domain *domain_priv;
struct psmx2_fid_ep *ep_priv;
uint8_t vlane;
uint64_t ep_cap;
int err = -FI_EINVAL;
if (info)
ep_cap = info->caps;
else
ep_cap = FI_TAGGED;
domain_priv = container_of(domain, struct psmx2_fid_domain, domain.fid);
if (!domain_priv)
goto errout;
err = psmx2_domain_check_features(domain_priv, ep_cap);
if (err)
goto errout;
err = psmx2_alloc_vlane(domain_priv, &vlane);
if (err)
goto errout;
ep_priv = (struct psmx2_fid_ep *) calloc(1, sizeof *ep_priv);
if (!ep_priv) {
err = -FI_ENOMEM;
goto errout_free_vlane;
}
ep_priv->ep.fid.fclass = FI_CLASS_EP;
ep_priv->ep.fid.context = context;
ep_priv->ep.fid.ops = &psmx2_fi_ops;
ep_priv->ep.ops = &psmx2_ep_ops;
ep_priv->ep.cm = &psmx2_cm_ops;
ep_priv->domain = domain_priv;
ep_priv->vlane = vlane;
PSMX2_CTXT_TYPE(&ep_priv->nocomp_send_context) = PSMX2_NOCOMP_SEND_CONTEXT;
PSMX2_CTXT_EP(&ep_priv->nocomp_send_context) = ep_priv;
PSMX2_CTXT_TYPE(&ep_priv->nocomp_recv_context) = PSMX2_NOCOMP_RECV_CONTEXT;
PSMX2_CTXT_EP(&ep_priv->nocomp_recv_context) = ep_priv;
if (ep_cap & FI_TAGGED)
ep_priv->ep.tagged = &psmx2_tagged_ops;
if (ep_cap & FI_MSG)
ep_priv->ep.msg = &psmx2_msg_ops;
if (ep_cap & FI_RMA)
ep_priv->ep.rma = &psmx2_rma_ops;
if (ep_cap & FI_ATOMICS)
ep_priv->ep.atomic = &psmx2_atomic_ops;
ep_priv->caps = ep_cap;
err = psmx2_domain_enable_ep(domain_priv, ep_priv);
if (err)
goto errout_free_ep;
psmx2_domain_acquire(domain_priv);
domain_priv->eps[ep_priv->vlane] = ep_priv;
if (info) {
if (info->tx_attr)
ep_priv->flags = info->tx_attr->op_flags;
if (info->rx_attr)
ep_priv->flags |= info->rx_attr->op_flags;
}
psmx2_ep_optimize_ops(ep_priv);
*ep = &ep_priv->ep;
return 0;
errout_free_ep:
free(ep_priv);
errout_free_vlane:
psmx2_free_vlane(domain_priv, vlane);
errout:
return err;
}
ssize_t psmx2_write_generic(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 psmx2_fid_ep *ep_priv;
struct psmx2_fid_av *av;
struct psmx2_epaddr_context *epaddr_context;
struct psmx2_am_request *req;
psm2_amarg_t args[8];
int nargs;
int am_flags = PSM2_AM_FLAG_ASYNC;
int chunk_size;
psm2_epaddr_t psm2_epaddr;
uint8_t vlane;
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag;
uint32_t tag32;
size_t idx;
void *psm2_context;
int no_event;
ep_priv = container_of(ep, struct psmx2_fid_ep, ep);
if (flags & FI_TRIGGER) {
struct psmx2_trigger *trigger;
struct fi_triggered_context *ctxt = context;
trigger = calloc(1, sizeof(*trigger));
if (!trigger)
return -FI_ENOMEM;
trigger->op = PSMX2_TRIGGERED_WRITE;
trigger->cntr = container_of(ctxt->trigger.threshold.cntr,
struct psmx2_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;
psmx2_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;
psm2_epaddr = av->epaddrs[idx];
vlane = av->vlanes[idx];
} else {
if (!dest_addr)
return -FI_EINVAL;
psm2_epaddr = PSMX2_ADDR_TO_EP(dest_addr);
vlane = PSMX2_ADDR_TO_VL(dest_addr);
}
epaddr_context = psm2_epaddr_getctxt((void *)psm2_epaddr);
if (epaddr_context->epid == ep_priv->domain->psm2_epid)
return psmx2_rma_self(PSMX2_AM_REQ_WRITE, ep_priv,
ep_priv->domain->eps[vlane],
(void *)buf, len, desc, addr,
key, context, flags, data);
no_event = (flags & PSMX2_NO_COMPLETION) ||
(ep_priv->send_selective_completion && !(flags & FI_COMPLETION));
if (flags & FI_INJECT) {
if (len > PSMX2_INJECT_SIZE)
return -FI_EMSGSIZE;
req = malloc(sizeof(*req) + len);
if (!req)
return -FI_ENOMEM;
memset(req, 0, sizeof(*req));
memcpy((uint8_t *)req + sizeof(*req), (void *)buf, len);
buf = (uint8_t *)req + sizeof(*req);
} else {
req = calloc(1, sizeof(*req));
if (!req)
return -FI_ENOMEM;
PSMX2_CTXT_TYPE(&req->fi_context) = no_event ?
PSMX2_NOCOMP_WRITE_CONTEXT :
PSMX2_WRITE_CONTEXT;
}
req->no_event = no_event;
req->op = PSMX2_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;
PSMX2_CTXT_USER(&req->fi_context) = context;
PSMX2_CTXT_EP(&req->fi_context) = ep_priv;
chunk_size = psmx2_am_param.max_request_short;
args[0].u32w0 = 0;
PSMX2_AM_SET_SRC(args[0].u32w0, ep_priv->vlane);
PSMX2_AM_SET_DST(args[0].u32w0, vlane);
if (psmx2_env.tagged_rma && len > chunk_size) {
tag32 = PSMX2_TAG32(PSMX2_RMA_BIT, ep_priv->vlane, vlane);
PSMX2_SET_TAG(psm2_tag, (uint64_t)req, tag32);
PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_WRITE_LONG);
args[0].u32w1 = len;
args[1].u64 = (uint64_t)req;
args[2].u64 = addr;
args[3].u64 = key;
nargs = 4;
if (flags & FI_REMOTE_CQ_DATA) {
PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_DATA);
args[4].u64 = data;
nargs++;
}
if (flags & FI_DELIVERY_COMPLETE) {
args[0].u32w0 |= PSMX2_AM_FORCE_ACK;
psm2_context = NULL;
} else {
psm2_context = (void *)&req->fi_context;
}
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args,
nargs, NULL, 0, am_flags, NULL, NULL);
psm2_mq_isend2(ep_priv->domain->psm2_mq, psm2_epaddr, 0,
&psm2_tag, buf, len, psm2_context, &psm2_req);
return 0;
}
PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_WRITE);
nargs = 4;
while (len > chunk_size) {
args[0].u32w1 = chunk_size;
args[1].u64 = (uint64_t)(uintptr_t)req;
args[2].u64 = addr;
args[3].u64 = key;
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args,
nargs, (void *)buf, chunk_size, am_flags,
NULL, NULL);
buf = (const uint8_t *)buf + chunk_size;
addr += chunk_size;
len -= chunk_size;
}
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) {
PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_DATA | PSMX2_AM_EOM);
args[4].u64 = data;
nargs++;
} else {
PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_EOM);
}
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, nargs,
(void *)buf, len, am_flags, NULL, NULL);
return 0;
}
ssize_t psmx2_readv_generic(struct fid_ep *ep, const struct iovec *iov,
void *desc, size_t count, fi_addr_t src_addr,
uint64_t addr, uint64_t key, void *context,
uint64_t flags)
{
struct psmx2_fid_ep *ep_priv;
struct psmx2_fid_av *av;
struct psmx2_epaddr_context *epaddr_context;
struct psmx2_am_request *req;
psm2_amarg_t args[8];
int chunk_size;
size_t offset = 0;
psm2_epaddr_t psm2_epaddr;
uint8_t vlane;
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag, psm2_tagsel;
uint32_t tag32;
size_t idx;
size_t total_len, long_len, short_len;
void *long_buf;
int i;
ep_priv = container_of(ep, struct psmx2_fid_ep, ep);
if (flags & FI_TRIGGER) {
struct psmx2_trigger *trigger;
struct fi_triggered_context *ctxt = context;
trigger = calloc(1, sizeof(*trigger));
if (!trigger)
return -FI_ENOMEM;
trigger->op = PSMX2_TRIGGERED_READV;
trigger->cntr = container_of(ctxt->trigger.threshold.cntr,
struct psmx2_fid_cntr, cntr);
trigger->threshold = ctxt->trigger.threshold.threshold;
trigger->readv.ep = ep;
trigger->readv.iov = iov;
trigger->readv.count = count;
trigger->readv.desc = desc;
trigger->readv.src_addr = src_addr;
trigger->readv.addr = addr;
trigger->readv.key = key;
trigger->readv.context = context;
trigger->readv.flags = flags & ~FI_TRIGGER;
psmx2_cntr_add_trigger(trigger->cntr, trigger);
return 0;
}
av = ep_priv->av;
if (av && av->type == FI_AV_TABLE) {
idx = src_addr;
if (idx >= av->last)
return -FI_EINVAL;
psm2_epaddr = av->epaddrs[idx];
vlane = av->vlanes[idx];
} else {
if (!src_addr)
return -FI_EINVAL;
psm2_epaddr = PSMX2_ADDR_TO_EP(src_addr);
vlane = PSMX2_ADDR_TO_VL(src_addr);
}
epaddr_context = psm2_epaddr_getctxt((void *)psm2_epaddr);
if (epaddr_context->epid == ep_priv->domain->psm2_epid)
return psmx2_rma_self(PSMX2_AM_REQ_READV, ep_priv,
ep_priv->domain->eps[vlane],
(void *)iov, count, desc, addr,
key, context, flags, 0);
total_len = 0;
for (i=0; i<count; i++)
total_len += iov[i].iov_len;
req = calloc(1, sizeof(*req) + count * sizeof(struct iovec));
if (!req)
return -FI_ENOMEM;
memcpy(req->iov, iov, count * sizeof(struct iovec));
req->op = PSMX2_AM_REQ_READV;
req->read.iov_count = count;
req->read.len = total_len;
req->read.addr = addr; /* needed? */
req->read.key = key; /* needed? */
req->read.context = context;
req->ep = ep_priv;
req->cq_flags = FI_READ | FI_RMA;
PSMX2_CTXT_TYPE(&req->fi_context) = PSMX2_READ_CONTEXT;
PSMX2_CTXT_USER(&req->fi_context) = context;
PSMX2_CTXT_EP(&req->fi_context) = ep_priv;
if (ep_priv->send_selective_completion && !(flags & FI_COMPLETION)) {
PSMX2_CTXT_TYPE(&req->fi_context) = PSMX2_NOCOMP_READ_CONTEXT;
req->no_event = 1;
}
chunk_size = psmx2_am_param.max_reply_short;
long_len = 0;
if (psmx2_env.tagged_rma) {
for (i=count-1; i>=0; i--) {
if (iov[i].iov_len > chunk_size) {
long_buf = iov[i].iov_base;
long_len = iov[i].iov_len;
break;
} else if (iov[i].iov_len) {
break;
}
}
}
short_len = total_len - long_len;
/* Use short protocol for all but the last segment (long_len) */
args[0].u32w0 = 0;
PSMX2_AM_SET_SRC(args[0].u32w0, ep_priv->vlane);
PSMX2_AM_SET_DST(args[0].u32w0, vlane);
PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_READ);
args[1].u64 = (uint64_t)(uintptr_t)req;
args[3].u64 = key;
while (short_len > chunk_size) {
args[0].u32w1 = chunk_size;
args[2].u64 = addr;
args[4].u64 = offset;
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER,
args, 5, NULL, 0, 0, NULL, NULL);
addr += chunk_size;
short_len -= chunk_size;
offset += chunk_size;
}
if (!long_len)
PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_EOM);
args[0].u32w1 = short_len;
args[2].u64 = addr;
args[4].u64 = offset;
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER,
args, 5, NULL, 0, 0, NULL, NULL);
/* Use the long protocol for the last segment */
if (long_len) {
tag32 = PSMX2_TAG32(PSMX2_RMA_BIT, vlane, ep_priv->vlane);
PSMX2_SET_TAG(psm2_tag, (uint64_t)req, tag32);
PSMX2_SET_TAG(psm2_tagsel, -1ULL, -1);
psm2_mq_irecv2(ep_priv->domain->psm2_mq, psm2_epaddr,
&psm2_tag, &psm2_tagsel, 0,
long_buf, long_len,
(void *)&req->fi_context, &psm2_req);
PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_READ_LONG);
args[0].u32w1 = long_len;
args[1].u64 = (uint64_t)req;
args[2].u64 = addr + short_len;
args[3].u64 = key;
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER,
args, 4, NULL, 0, 0, NULL, NULL);
}
return 0;
}
ssize_t psmx2_sendv_generic(struct fid_ep *ep, const struct iovec *iov,
void **desc, size_t count, fi_addr_t dest_addr,
void *context, uint64_t flags, uint64_t data)
{
struct psmx2_fid_ep *ep_priv;
struct psmx2_fid_av *av;
psm2_epaddr_t psm2_epaddr;
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag;
uint32_t msg_flags;
struct fi_context * fi_context;
int send_flag = 0;
int err;
int no_completion = 0;
struct psmx2_cq_event *event;
size_t real_count;
size_t len, total_len;
char *p;
uint32_t *q;
int i, j;
struct psmx2_sendv_request *req;
ep_priv = container_of(ep, struct psmx2_fid_ep, ep);
if (flags & FI_TRIGGER)
return psmx2_trigger_queue_sendv(ep, iov, desc, count,
dest_addr, context, flags,
data);
total_len = 0;
real_count = 0;
for (i=0; i<count; i++) {
if (iov[i].iov_len) {
total_len += iov[i].iov_len;
real_count++;
j = i;
}
}
if (real_count == 1)
return psmx2_send_generic(ep, iov[j].iov_base, iov[j].iov_len,
desc ? desc[j] : NULL, dest_addr,
context, flags, data);
req = malloc(sizeof(*req));
if (!req)
return -FI_ENOMEM;
if (total_len <= PSMX2_IOV_BUF_SIZE) {
req->iov_protocol = PSMX2_IOV_PROTO_PACK;
p = req->buf;
for (i=0; i<count; i++) {
if (iov[i].iov_len) {
memcpy(p, iov[i].iov_base, iov[i].iov_len);
p += iov[i].iov_len;
}
}
msg_flags = PSMX2_TYPE_MSG;
len = total_len;
} else {
req->iov_protocol = PSMX2_IOV_PROTO_MULTI;
req->iov_done = 0;
req->iov_info.seq_num = (++ep_priv->iov_seq_num) %
PSMX2_IOV_MAX_SEQ_NUM + 1;
req->iov_info.count = (uint32_t)real_count;
req->iov_info.total_len = (uint32_t)total_len;
q = req->iov_info.len;
for (i=0; i<count; i++) {
if (iov[i].iov_len)
*q++ = (uint32_t)iov[i].iov_len;
}
msg_flags = PSMX2_TYPE_MSG | PSMX2_IOV_BIT;
len = (3 + real_count) * sizeof(uint32_t);
}
av = ep_priv->av;
assert(av);
psm2_epaddr = psmx2_av_translate_addr(av, ep_priv->tx, dest_addr, av->type);
if (flags & FI_REMOTE_CQ_DATA)
msg_flags |= PSMX2_IMM_BIT;
PSMX2_SET_TAG(psm2_tag, 0ULL, data, msg_flags);
if ((flags & PSMX2_NO_COMPLETION) ||
(ep_priv->send_selective_completion && !(flags & FI_COMPLETION)))
no_completion = 1;
if (flags & FI_INJECT) {
if (len > psmx2_env.inject_size) {
free(req);
return -FI_EMSGSIZE;
}
err = psm2_mq_send2(ep_priv->tx->psm2_mq, psm2_epaddr,
send_flag, &psm2_tag, req->buf, len);
free(req);
if (err != PSM2_OK)
return psmx2_errno(err);
if (ep_priv->send_cntr)
psmx2_cntr_inc(ep_priv->send_cntr, 0);
if (ep_priv->send_cq && !no_completion) {
event = psmx2_cq_create_event(
ep_priv->send_cq,
context, NULL, flags, len,
(uint64_t) data,
0 /* tag */,
0 /* olen */,
0 /* err */);
if (event)
psmx2_cq_enqueue_event(ep_priv->send_cq, event);
else
return -FI_ENOMEM;
}
return 0;
}
req->no_completion = no_completion;
req->user_context = context;
req->comp_flag = FI_MSG;
fi_context = &req->fi_context;
PSMX2_CTXT_TYPE(fi_context) = PSMX2_SENDV_CONTEXT;
PSMX2_CTXT_USER(fi_context) = req;
PSMX2_CTXT_EP(fi_context) = ep_priv;
err = psm2_mq_isend2(ep_priv->tx->psm2_mq, psm2_epaddr,
send_flag, &psm2_tag, req->buf, len,
(void *)fi_context, &psm2_req);
if (err != PSM2_OK) {
free(req);
return psmx2_errno(err);
}
PSMX2_CTXT_REQ(fi_context) = psm2_req;
if (req->iov_protocol == PSMX2_IOV_PROTO_MULTI) {
fi_context = &req->fi_context_iov;
PSMX2_CTXT_TYPE(fi_context) = PSMX2_IOV_SEND_CONTEXT;
PSMX2_CTXT_USER(fi_context) = req;
PSMX2_CTXT_EP(fi_context) = ep_priv;
PSMX2_SET_TAG(psm2_tag, req->iov_info.seq_num, 0, PSMX2_TYPE_IOV_PAYLOAD);
for (i=0; i<count; i++) {
if (iov[i].iov_len) {
err = psm2_mq_isend2(ep_priv->tx->psm2_mq,
psm2_epaddr, send_flag, &psm2_tag,
iov[i].iov_base, iov[i].iov_len,
(void *)fi_context, &psm2_req);
if (err != PSM2_OK)
return psmx2_errno(err);
}
}
}
return 0;
}
ssize_t psmx2_recv_generic(struct fid_ep *ep, void *buf, size_t len,
void *desc, fi_addr_t src_addr, void *context,
uint64_t flags)
{
struct psmx2_fid_ep *ep_priv;
struct psmx2_fid_av *av;
psm2_epaddr_t psm2_epaddr;
uint8_t vlane;
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag, psm2_tagsel;
uint32_t tag32, tagsel32;
struct fi_context *fi_context;
int recv_flag = 0;
size_t idx;
int err;
ep_priv = container_of(ep, struct psmx2_fid_ep, ep);
if (flags & FI_TRIGGER) {
struct psmx2_trigger *trigger;
struct fi_triggered_context *ctxt = context;
trigger = calloc(1, sizeof(*trigger));
if (!trigger)
return -FI_ENOMEM;
trigger->op = PSMX2_TRIGGERED_RECV;
trigger->cntr = container_of(ctxt->trigger.threshold.cntr,
struct psmx2_fid_cntr, cntr);
trigger->threshold = ctxt->trigger.threshold.threshold;
trigger->recv.ep = ep;
trigger->recv.buf = buf;
trigger->recv.len = len;
trigger->recv.desc = desc;
trigger->recv.src_addr = src_addr;
trigger->recv.context = context;
trigger->recv.flags = flags & ~FI_TRIGGER;
psmx2_cntr_add_trigger(trigger->cntr, trigger);
return 0;
}
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;
psm2_epaddr = av->epaddrs[idx];
vlane = av->vlanes[idx];
} else {
psm2_epaddr = PSMX2_ADDR_TO_EP(src_addr);
vlane = PSMX2_ADDR_TO_VL(src_addr);
}
tag32 = PSMX2_TAG32(PSMX2_MSG_BIT, vlane, ep_priv->vlane);
tagsel32 = ~(PSMX2_IOV_BIT | PSMX2_IMM_BIT);
} else {
psm2_epaddr = 0;
tag32 = PSMX2_TAG32(PSMX2_MSG_BIT, 0, ep_priv->vlane);
tagsel32 = ~(PSMX2_IOV_BIT | PSMX2_IMM_BIT | PSMX2_SRC_BITS);
}
PSMX2_SET_TAG(psm2_tag, 0ULL, tag32);
PSMX2_SET_TAG(psm2_tagsel, 0ULL, tagsel32);
if (ep_priv->recv_selective_completion && !(flags & FI_COMPLETION)) {
fi_context = psmx2_ep_get_op_context(ep_priv);
PSMX2_CTXT_TYPE(fi_context) = PSMX2_NOCOMP_RECV_CONTEXT_ALLOC;
PSMX2_CTXT_EP(fi_context) = ep_priv;
PSMX2_CTXT_USER(fi_context) = buf;
PSMX2_CTXT_SIZE(fi_context) = len;
} else {
if (!context)
return -FI_EINVAL;
fi_context = context;
if (flags & FI_MULTI_RECV) {
struct psmx2_multi_recv *req;
req = calloc(1, sizeof(*req));
if (!req)
return -FI_ENOMEM;
req->src_addr = psm2_epaddr;
req->tag = psm2_tag;
req->tagsel = psm2_tagsel;
req->flag = recv_flag;
req->buf = buf;
req->len = len;
req->offset = 0;
req->min_buf_size = ep_priv->min_multi_recv;
req->context = fi_context;
PSMX2_CTXT_TYPE(fi_context) = PSMX2_MULTI_RECV_CONTEXT;
PSMX2_CTXT_USER(fi_context) = req;
if (len > PSMX2_MAX_MSG_SIZE)
len = PSMX2_MAX_MSG_SIZE;
} else {
PSMX2_CTXT_TYPE(fi_context) = PSMX2_RECV_CONTEXT;
PSMX2_CTXT_USER(fi_context) = buf;
}
PSMX2_CTXT_EP(fi_context) = ep_priv;
PSMX2_CTXT_SIZE(fi_context) = len;
}
err = psm2_mq_irecv2(ep_priv->domain->psm2_mq, psm2_epaddr,
&psm2_tag, &psm2_tagsel, recv_flag, buf, len,
(void *)fi_context, &psm2_req);
if (err != PSM2_OK)
return psmx2_errno(err);
if (fi_context == context)
PSMX2_CTXT_REQ(fi_context) = psm2_req;
return 0;
}
ssize_t psmx2_send_generic(struct fid_ep *ep, const void *buf, size_t len,
void *desc, fi_addr_t dest_addr, void *context,
uint64_t flags, uint64_t data)
{
struct psmx2_fid_ep *ep_priv;
struct psmx2_fid_av *av;
psm2_epaddr_t psm2_epaddr;
uint8_t vlane;
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag;
uint32_t tag32;
struct fi_context * fi_context;
int send_flag = 0;
int err;
size_t idx;
int no_completion = 0;
struct psmx2_cq_event *event;
ep_priv = container_of(ep, struct psmx2_fid_ep, ep);
if (flags & FI_TRIGGER) {
struct psmx2_trigger *trigger;
struct fi_triggered_context *ctxt = context;
trigger = calloc(1, sizeof(*trigger));
if (!trigger)
return -FI_ENOMEM;
trigger->op = PSMX2_TRIGGERED_SEND;
trigger->cntr = container_of(ctxt->trigger.threshold.cntr,
struct psmx2_fid_cntr, cntr);
trigger->threshold = ctxt->trigger.threshold.threshold;
trigger->send.ep = ep;
trigger->send.buf = buf;
trigger->send.len = len;
trigger->send.desc = desc;
trigger->send.dest_addr = dest_addr;
trigger->send.context = context;
trigger->send.flags = flags & ~FI_TRIGGER;
trigger->send.data = data;
psmx2_cntr_add_trigger(trigger->cntr, trigger);
return 0;
}
av = ep_priv->av;
if (av && av->type == FI_AV_TABLE) {
idx = (size_t)dest_addr;
if (idx >= av->last)
return -FI_EINVAL;
psm2_epaddr = av->epaddrs[idx];
vlane = av->vlanes[idx];
} else {
psm2_epaddr = PSMX2_ADDR_TO_EP(dest_addr);
vlane = PSMX2_ADDR_TO_VL(dest_addr);
}
tag32 = PSMX2_TAG32(PSMX2_MSG_BIT, ep_priv->vlane, vlane);
if (flags & FI_REMOTE_CQ_DATA)
tag32 |= PSMX2_IMM_BIT;
PSMX2_SET_TAG(psm2_tag, data, tag32);
if ((flags & PSMX2_NO_COMPLETION) ||
(ep_priv->send_selective_completion && !(flags & FI_COMPLETION)))
no_completion = 1;
if (flags & FI_INJECT) {
if (len > PSMX2_INJECT_SIZE)
return -FI_EMSGSIZE;
err = psm2_mq_send2(ep_priv->domain->psm2_mq, psm2_epaddr,
send_flag, &psm2_tag, buf, len);
if (err != PSM2_OK)
return psmx2_errno(err);
if (ep_priv->send_cntr)
psmx2_cntr_inc(ep_priv->send_cntr);
if (ep_priv->send_cq && !no_completion) {
event = psmx2_cq_create_event(
ep_priv->send_cq,
context, (void *)buf, flags, len,
(uint64_t) data,
0 /* tag */,
0 /* olen */,
0 /* err */);
if (event)
psmx2_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;
PSMX2_CTXT_TYPE(fi_context) = PSMX2_SEND_CONTEXT;
PSMX2_CTXT_USER(fi_context) = (void *)buf;
PSMX2_CTXT_EP(fi_context) = ep_priv;
}
err = psm2_mq_isend2(ep_priv->domain->psm2_mq, psm2_epaddr,
send_flag, &psm2_tag, buf, len,
(void *)fi_context, &psm2_req);
if (err != PSM2_OK)
return psmx2_errno(err);
if (fi_context == context)
PSMX2_CTXT_REQ(fi_context) = psm2_req;
return 0;
}
ssize_t psmx2_recv_generic(struct fid_ep *ep, void *buf, size_t len,
void *desc, fi_addr_t src_addr, void *context,
uint64_t flags)
{
struct psmx2_fid_ep *ep_priv;
struct psmx2_fid_av *av;
psm2_epaddr_t psm2_epaddr;
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag, psm2_tagsel;
struct fi_context *fi_context;
int recv_flag = 0;
int err;
int enable_completion;
ep_priv = container_of(ep, struct psmx2_fid_ep, ep);
if (flags & FI_TRIGGER)
return psmx2_trigger_queue_recv(ep, buf, len, desc, src_addr,
context, flags);
if ((ep_priv->caps & FI_DIRECTED_RECV) && src_addr != FI_ADDR_UNSPEC) {
av = ep_priv->av;
assert(av);
psm2_epaddr = psmx2_av_translate_addr(av, ep_priv->rx, src_addr, av->type);
} else {
psm2_epaddr = 0;
}
PSMX2_SET_TAG(psm2_tag, 0ULL, 0, PSMX2_TYPE_MSG);
PSMX2_SET_MASK(psm2_tagsel, PSMX2_MATCH_NONE, PSMX2_TYPE_MASK);
enable_completion = !ep_priv->recv_selective_completion ||
(flags & FI_COMPLETION);
if (enable_completion) {
assert(context);
fi_context = context;
if (flags & FI_MULTI_RECV) {
struct psmx2_multi_recv *req;
req = calloc(1, sizeof(*req));
if (!req)
return -FI_ENOMEM;
req->src_addr = psm2_epaddr;
req->tag = psm2_tag;
req->tagsel = psm2_tagsel;
req->flag = recv_flag;
req->buf = buf;
req->len = len;
req->offset = 0;
req->min_buf_size = ep_priv->min_multi_recv;
req->context = fi_context;
PSMX2_CTXT_TYPE(fi_context) = PSMX2_MULTI_RECV_CONTEXT;
PSMX2_CTXT_USER(fi_context) = req;
if (len > PSMX2_MAX_MSG_SIZE)
len = PSMX2_MAX_MSG_SIZE;
} else {
PSMX2_CTXT_TYPE(fi_context) = PSMX2_RECV_CONTEXT;
PSMX2_CTXT_USER(fi_context) = buf;
}
PSMX2_CTXT_EP(fi_context) = ep_priv;
PSMX2_CTXT_SIZE(fi_context) = len;
} else {
PSMX2_EP_GET_OP_CONTEXT(ep_priv, fi_context);
#if !PSMX2_USE_REQ_CONTEXT
PSMX2_CTXT_TYPE(fi_context) = PSMX2_NOCOMP_RECV_CONTEXT;
PSMX2_CTXT_EP(fi_context) = ep_priv;
PSMX2_CTXT_USER(fi_context) = buf;
PSMX2_CTXT_SIZE(fi_context) = len;
#endif
}
err = psm2_mq_irecv2(ep_priv->rx->psm2_mq, psm2_epaddr,
&psm2_tag, &psm2_tagsel, recv_flag, buf, len,
(void *)fi_context, &psm2_req);
if (OFI_UNLIKELY(err != PSM2_OK))
return psmx2_errno(err);
if (enable_completion) {
PSMX2_CTXT_REQ(fi_context) = psm2_req;
} else {
#if PSMX2_USE_REQ_CONTEXT
PSMX2_REQ_GET_OP_CONTEXT(psm2_req, fi_context);
PSMX2_CTXT_TYPE(fi_context) = PSMX2_NOCOMP_RECV_CONTEXT;
PSMX2_CTXT_EP(fi_context) = ep_priv;
PSMX2_CTXT_USER(fi_context) = buf;
PSMX2_CTXT_SIZE(fi_context) = len;
#endif
}
return 0;
}
ssize_t psmx2_tagged_recv_generic(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 psmx2_fid_ep *ep_priv;
struct psmx2_fid_av *av;
psm2_epaddr_t psm2_epaddr;
uint8_t vlane;
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag, psm2_tagsel;
uint32_t tag32, tagsel32;
struct fi_context *fi_context;
size_t idx;
int err;
ep_priv = container_of(ep, struct psmx2_fid_ep, ep);
if (flags & FI_PEEK)
return psmx2_tagged_peek_generic(ep, buf, len, desc,
src_addr, tag, ignore,
context, flags);
if (flags & FI_TRIGGER) {
struct psmx2_trigger *trigger;
struct fi_triggered_context *ctxt = context;
trigger = calloc(1, sizeof(*trigger));
if (!trigger)
return -FI_ENOMEM;
trigger->op = PSMX2_TRIGGERED_TRECV;
trigger->cntr = container_of(ctxt->trigger.threshold.cntr,
struct psmx2_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;
psmx2_cntr_add_trigger(trigger->cntr, trigger);
return 0;
}
if (flags & FI_CLAIM) {
if (!context)
return -FI_EINVAL;
if (flags & FI_DISCARD) {
psm2_mq_status2_t psm2_status;
struct psmx2_cq_event *event;
fi_context = context;
psm2_req = PSMX2_CTXT_REQ(fi_context);
err = psm2_mq_imrecv(ep_priv->trx_ctxt->psm2_mq, 0,
NULL, 0, context, &psm2_req);
if (err != PSM2_OK)
return psmx2_errno(err);
psm2_mq_wait2(&psm2_req, &psm2_status);
if (ep_priv->recv_cq &&
(!ep_priv->recv_selective_completion || (flags & FI_COMPLETION))) {
tag = PSMX2_GET_TAG64(psm2_status.msg_tag);
event = psmx2_cq_create_event(
ep_priv->recv_cq,
context, /* op_context */
NULL, /* buf */
flags|FI_RECV|FI_TAGGED,/* flags */
0, /* len */
0, /* data */
tag, /* tag */
0, /* olen */
0); /* err */
if (!event)
return -FI_ENOMEM;
vlane = PSMX2_TAG32_GET_SRC(psm2_status.msg_tag.tag2);
event->source_is_valid = 1;
event->source = PSMX2_EP_TO_ADDR(psm2_status.msg_peer, vlane);
event->source_av = ep_priv->av;
psmx2_cq_enqueue_event(ep_priv->recv_cq, event);
}
if (ep_priv->recv_cntr)
psmx2_cntr_inc(ep_priv->recv_cntr);
return 0;
}
fi_context = context;
psm2_req = PSMX2_CTXT_REQ(fi_context);
PSMX2_CTXT_TYPE(fi_context) = PSMX2_TRECV_CONTEXT;
PSMX2_CTXT_USER(fi_context) = buf;
PSMX2_CTXT_EP(fi_context) = ep_priv;
err = psm2_mq_imrecv(ep_priv->trx_ctxt->psm2_mq, 0,
buf, len, context, &psm2_req);
if (err != PSM2_OK)
return psmx2_errno(err);
PSMX2_CTXT_REQ(fi_context) = psm2_req;
return 0;
}
if (ep_priv->recv_selective_completion && !(flags & FI_COMPLETION)) {
fi_context = psmx2_ep_get_op_context(ep_priv);
PSMX2_CTXT_TYPE(fi_context) = PSMX2_NOCOMP_RECV_CONTEXT_ALLOC;
PSMX2_CTXT_EP(fi_context) = ep_priv;
PSMX2_CTXT_USER(fi_context) = buf;
PSMX2_CTXT_SIZE(fi_context) = len;
} else {
if (!context)
return -FI_EINVAL;
fi_context = context;
PSMX2_CTXT_TYPE(fi_context) = PSMX2_TRECV_CONTEXT;
PSMX2_CTXT_USER(fi_context) = buf;
PSMX2_CTXT_EP(fi_context) = ep_priv;
PSMX2_CTXT_SIZE(fi_context) = len;
}
if ((ep_priv->caps & FI_DIRECTED_RECV) && src_addr != FI_ADDR_UNSPEC) {
av = ep_priv->av;
if (av && PSMX2_SEP_ADDR_TEST(src_addr)) {
psm2_epaddr = psmx2_av_translate_sep(av, ep_priv->trx_ctxt, src_addr);
vlane = 0;
} else if (av && av->type == FI_AV_TABLE) {
idx = (size_t)src_addr;
if (idx >= av->last)
return -FI_EINVAL;
psm2_epaddr = av->epaddrs[idx];
vlane = av->vlanes[idx];
} else {
psm2_epaddr = PSMX2_ADDR_TO_EP(src_addr);
vlane = PSMX2_ADDR_TO_VL(src_addr);
}
tag32 = PSMX2_TAG32(0, vlane, ep_priv->vlane);
tagsel32 = ~PSMX2_IOV_BIT;
} else {
psm2_epaddr = 0;
tag32 = PSMX2_TAG32(0, 0, ep_priv->vlane);
tagsel32 = ~(PSMX2_IOV_BIT | PSMX2_SRC_BITS);
}
PSMX2_SET_TAG(psm2_tag, tag, tag32);
PSMX2_SET_TAG(psm2_tagsel, ~ignore, tagsel32);
err = psm2_mq_irecv2(ep_priv->trx_ctxt->psm2_mq, psm2_epaddr,
&psm2_tag, &psm2_tagsel, 0, buf, len,
(void *)fi_context, &psm2_req);
if (err != PSM2_OK)
return psmx2_errno(err);
if (fi_context == context)
PSMX2_CTXT_REQ(fi_context) = psm2_req;
return 0;
}
int psmx2_handle_sendv_req(struct psmx2_fid_ep *ep,
PSMX2_STATUS_TYPE *status,
int multi_recv)
{
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag, psm2_tagsel;
struct psmx2_sendv_reply *rep;
struct psmx2_multi_recv *recv_req;
struct fi_context *fi_context;
struct fi_context *recv_context;
int i, err;
uint8_t *recv_buf;
size_t recv_len, len;
if (PSMX2_STATUS_ERROR(status) != PSM2_OK)
return psmx2_errno(PSMX2_STATUS_ERROR(status));
rep = malloc(sizeof(*rep));
if (!rep) {
PSMX2_STATUS_ERROR(status) = PSM2_NO_MEMORY;
return -FI_ENOMEM;
}
recv_context = PSMX2_STATUS_CONTEXT(status);
if (multi_recv) {
recv_req = PSMX2_CTXT_USER(recv_context);
recv_buf = recv_req->buf + recv_req->offset;
recv_len = recv_req->len - recv_req->offset;
rep->multi_recv = 1;
} else {
recv_buf = PSMX2_CTXT_USER(recv_context);
recv_len = PSMX2_CTXT_SIZE(recv_context);
rep->multi_recv = 0;
}
/* assert(PSMX2_STATUS_RCVLEN(status) <= PSMX2_IOV_BUF_SIZE); */
memcpy(&rep->iov_info, recv_buf, PSMX2_STATUS_RCVLEN(status));
rep->user_context = PSMX2_STATUS_CONTEXT(status);
rep->tag = PSMX2_STATUS_TAG(status);
rep->buf = recv_buf;
rep->no_completion = 0;
rep->iov_done = 0;
rep->bytes_received = 0;
rep->msg_length = 0;
rep->error_code = PSM2_OK;
fi_context = &rep->fi_context;
PSMX2_CTXT_TYPE(fi_context) = PSMX2_IOV_RECV_CONTEXT;
PSMX2_CTXT_USER(fi_context) = rep;
PSMX2_CTXT_EP(fi_context) = ep;
rep->comp_flag = PSMX2_IS_MSG(PSMX2_GET_FLAGS(rep->tag)) ? FI_MSG : FI_TAGGED;
if (PSMX2_HAS_IMM(PSMX2_GET_FLAGS(rep->tag)))
rep->comp_flag |= FI_REMOTE_CQ_DATA;
/* IOV payload uses a sequence number in place of a tag. */
PSMX2_SET_TAG(psm2_tag, rep->iov_info.seq_num, 0, PSMX2_TYPE_IOV_PAYLOAD);
PSMX2_SET_MASK(psm2_tagsel, PSMX2_MATCH_ALL, PSMX2_TYPE_MASK);
for (i=0; i<rep->iov_info.count; i++) {
if (recv_len) {
len = MIN(recv_len, rep->iov_info.len[i]);
err = psm2_mq_irecv2(ep->rx->psm2_mq,
PSMX2_STATUS_PEER(status),
&psm2_tag, &psm2_tagsel,
0/*flag*/, recv_buf, len,
(void *)fi_context, &psm2_req);
if (err) {
PSMX2_STATUS_ERROR(status) = err;
return psmx2_errno(err);
}
recv_buf += len;
recv_len -= len;
} else {
/* recv buffer full, post empty recvs */
err = psm2_mq_irecv2(ep->rx->psm2_mq,
PSMX2_STATUS_PEER(status),
&psm2_tag, &psm2_tagsel,
0/*flag*/, NULL, 0,
(void *)fi_context, &psm2_req);
if (err) {
PSMX2_STATUS_ERROR(status) = err;
return psmx2_errno(err);
}
}
}
if (multi_recv && recv_len < recv_req->min_buf_size)
rep->comp_flag |= FI_MULTI_RECV;
return 0;
}
ssize_t psmx2_readv_generic(struct fid_ep *ep, const struct iovec *iov,
void *desc, size_t count, fi_addr_t src_addr,
uint64_t addr, uint64_t key, void *context,
uint64_t flags)
{
struct psmx2_fid_ep *ep_priv;
struct psmx2_fid_av *av;
struct psmx2_epaddr_context *epaddr_context;
struct psmx2_am_request *req;
psm2_amarg_t args[8];
int chunk_size;
size_t offset = 0;
psm2_epaddr_t psm2_epaddr;
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag, psm2_tagsel;
size_t idx;
size_t total_len, long_len = 0, short_len;
void *long_buf = NULL;
int i;
int err;
ep_priv = container_of(ep, struct psmx2_fid_ep, ep);
if (flags & FI_TRIGGER)
return psmx2_trigger_queue_readv(ep, iov, desc, count, src_addr,
addr, key, context, flags);
av = ep_priv->av;
if (av && PSMX2_SEP_ADDR_TEST(src_addr)) {
psm2_epaddr = psmx2_av_translate_sep(av, ep_priv->tx, src_addr);
} else if (av && av->type == FI_AV_TABLE) {
idx = src_addr;
if ((err = psmx2_av_check_table_idx(av, ep_priv->tx, idx)))
return err;
psm2_epaddr = av->tables[ep_priv->tx->id].epaddrs[idx];
} else {
if (!src_addr)
return -FI_EINVAL;
psm2_epaddr = PSMX2_ADDR_TO_EP(src_addr);
}
epaddr_context = psm2_epaddr_getctxt((void *)psm2_epaddr);
if (epaddr_context->epid == ep_priv->tx->psm2_epid)
return psmx2_rma_self(PSMX2_AM_REQ_READV, ep_priv,
(void *)iov, count, desc, addr,
key, context, flags, 0);
total_len = 0;
for (i=0; i<count; i++)
total_len += iov[i].iov_len;
req = psmx2_am_request_alloc(ep_priv->tx);
if (!req)
return -FI_ENOMEM;
req->tmpbuf = malloc(count * sizeof(struct iovec));
if (!req->tmpbuf) {
psmx2_am_request_free(ep_priv->tx, req);
return -FI_ENOMEM;
}
req->iov = req->tmpbuf;
memcpy(req->iov, iov, count * sizeof(struct iovec));
req->op = PSMX2_AM_REQ_READV;
req->read.iov_count = count;
req->read.len = total_len;
req->read.addr = addr; /* needed? */
req->read.key = key; /* needed? */
req->read.context = context;
req->ep = ep_priv;
req->cq_flags = FI_READ | FI_RMA;
PSMX2_CTXT_TYPE(&req->fi_context) = PSMX2_READ_CONTEXT;
PSMX2_CTXT_USER(&req->fi_context) = context;
PSMX2_CTXT_EP(&req->fi_context) = ep_priv;
if (ep_priv->send_selective_completion && !(flags & FI_COMPLETION)) {
PSMX2_CTXT_TYPE(&req->fi_context) = PSMX2_NOCOMP_READ_CONTEXT;
req->no_event = 1;
}
chunk_size = ep_priv->tx->psm2_am_param.max_reply_short;
if (psmx2_env.tagged_rma) {
for (i=count-1; i>=0; i--) {
if (iov[i].iov_len > chunk_size) {
long_buf = iov[i].iov_base;
long_len = iov[i].iov_len;
break;
} else if (iov[i].iov_len) {
break;
}
}
}
short_len = total_len - long_len;
/* Use short protocol for all but the last segment (long_len) */
args[0].u32w0 = 0;
PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_READ);
args[1].u64 = (uint64_t)(uintptr_t)req;
args[3].u64 = key;
while (short_len > chunk_size) {
args[0].u32w1 = chunk_size;
args[2].u64 = addr;
args[4].u64 = offset;
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER,
args, 5, NULL, 0, 0, NULL, NULL);
addr += chunk_size;
short_len -= chunk_size;
offset += chunk_size;
}
if (!long_len)
PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_EOM);
args[0].u32w1 = short_len;
args[2].u64 = addr;
args[4].u64 = offset;
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER,
args, 5, NULL, 0, 0, NULL, NULL);
/* Use the long protocol for the last segment */
if (long_len) {
PSMX2_SET_TAG(psm2_tag, (uint64_t)req, 0, PSMX2_RMA_TYPE_READ);
PSMX2_SET_MASK(psm2_tagsel, PSMX2_MATCH_ALL, PSMX2_RMA_TYPE_MASK);
psm2_mq_irecv2(ep_priv->tx->psm2_mq, psm2_epaddr,
&psm2_tag, &psm2_tagsel, 0,
long_buf, long_len,
(void *)&req->fi_context, &psm2_req);
PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_READ_LONG);
args[0].u32w1 = long_len;
args[1].u64 = (uint64_t)req;
args[2].u64 = addr + short_len;
args[3].u64 = key;
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER,
args, 4, NULL, 0, 0, NULL, NULL);
}
return 0;
}
int psmx2_ep_open(struct fid_domain *domain, struct fi_info *info,
struct fid_ep **ep, void *context)
{
struct psmx2_fid_domain *domain_priv;
struct psmx2_fid_ep *ep_priv;
struct psmx2_context *item;
uint8_t vlane;
uint64_t ep_cap;
int err = -FI_EINVAL;
int i;
if (info)
ep_cap = info->caps;
else
ep_cap = FI_TAGGED;
domain_priv = container_of(domain, struct psmx2_fid_domain, domain.fid);
if (!domain_priv)
goto errout;
err = psmx2_domain_check_features(domain_priv, ep_cap);
if (err)
goto errout;
err = psmx2_alloc_vlane(domain_priv, &vlane);
if (err)
goto errout;
ep_priv = (struct psmx2_fid_ep *) calloc(1, sizeof *ep_priv);
if (!ep_priv) {
err = -FI_ENOMEM;
goto errout_free_vlane;
}
ep_priv->ep.fid.fclass = FI_CLASS_EP;
ep_priv->ep.fid.context = context;
ep_priv->ep.fid.ops = &psmx2_fi_ops;
ep_priv->ep.ops = &psmx2_ep_ops;
ep_priv->ep.cm = &psmx2_cm_ops;
ep_priv->domain = domain_priv;
ep_priv->vlane = vlane;
PSMX2_CTXT_TYPE(&ep_priv->nocomp_send_context) = PSMX2_NOCOMP_SEND_CONTEXT;
PSMX2_CTXT_EP(&ep_priv->nocomp_send_context) = ep_priv;
PSMX2_CTXT_TYPE(&ep_priv->nocomp_recv_context) = PSMX2_NOCOMP_RECV_CONTEXT;
PSMX2_CTXT_EP(&ep_priv->nocomp_recv_context) = ep_priv;
if (ep_cap & FI_TAGGED)
ep_priv->ep.tagged = &psmx2_tagged_ops;
if (ep_cap & FI_MSG)
ep_priv->ep.msg = &psmx2_msg_ops;
if (ep_cap & FI_RMA)
ep_priv->ep.rma = &psmx2_rma_ops;
if (ep_cap & FI_ATOMICS)
ep_priv->ep.atomic = &psmx2_atomic_ops;
ep_priv->caps = ep_cap;
err = psmx2_domain_enable_ep(domain_priv, ep_priv);
if (err)
goto errout_free_ep;
psmx2_domain_acquire(domain_priv);
domain_priv->eps[ep_priv->vlane] = ep_priv;
if (info) {
if (info->tx_attr)
ep_priv->flags = info->tx_attr->op_flags;
if (info->rx_attr)
ep_priv->flags |= info->rx_attr->op_flags;
}
psmx2_ep_optimize_ops(ep_priv);
slist_init(&ep_priv->free_context_list);
fastlock_init(&ep_priv->context_lock);
#define PSMX2_FREE_CONTEXT_LIST_SIZE 64
for (i=0; i<PSMX2_FREE_CONTEXT_LIST_SIZE; i++) {
item = calloc(1, sizeof(*item));
if (!item) {
FI_WARN(&psmx2_prov, FI_LOG_EP_CTRL, "out of memory.\n");
exit(-1);
}
slist_insert_tail(&item->list_entry, &ep_priv->free_context_list);
}
*ep = &ep_priv->ep;
return 0;
errout_free_ep:
free(ep_priv);
errout_free_vlane:
psmx2_free_vlane(domain_priv, vlane);
errout:
return err;
}
ssize_t psmx2_write_generic(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 psmx2_fid_ep *ep_priv;
struct psmx2_fid_av *av;
struct psmx2_epaddr_context *epaddr_context;
struct psmx2_am_request *req;
psm2_amarg_t args[8];
int nargs;
int am_flags = PSM2_AM_FLAG_ASYNC;
int chunk_size;
psm2_epaddr_t psm2_epaddr;
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag;
size_t idx;
void *psm2_context;
int no_event;
int err;
ep_priv = container_of(ep, struct psmx2_fid_ep, ep);
if (flags & FI_TRIGGER)
return psmx2_trigger_queue_write(ep, buf, len, desc, dest_addr,
addr, key, context, flags,
data);
if (!buf)
return -FI_EINVAL;
av = ep_priv->av;
if (av && PSMX2_SEP_ADDR_TEST(dest_addr)) {
psm2_epaddr = psmx2_av_translate_sep(av, ep_priv->tx, dest_addr);
} else if (av && av->type == FI_AV_TABLE) {
idx = dest_addr;
if ((err = psmx2_av_check_table_idx(av, ep_priv->tx, idx)))
return err;
psm2_epaddr = av->tables[ep_priv->tx->id].epaddrs[idx];
} else {
if (!dest_addr)
return -FI_EINVAL;
psm2_epaddr = PSMX2_ADDR_TO_EP(dest_addr);
}
epaddr_context = psm2_epaddr_getctxt((void *)psm2_epaddr);
if (epaddr_context->epid == ep_priv->tx->psm2_epid)
return psmx2_rma_self(PSMX2_AM_REQ_WRITE, ep_priv,
(void *)buf, len, desc, addr,
key, context, flags, data);
no_event = (flags & PSMX2_NO_COMPLETION) ||
(ep_priv->send_selective_completion && !(flags & FI_COMPLETION));
req = psmx2_am_request_alloc(ep_priv->tx);
if (!req)
return -FI_ENOMEM;
if (flags & FI_INJECT) {
if (len > psmx2_env.inject_size) {
psmx2_am_request_free(ep_priv->tx, req);
return -FI_EMSGSIZE;
}
req->tmpbuf = malloc(len);
if (!req->tmpbuf) {
psmx2_am_request_free(ep_priv->tx, req);
return -FI_ENOMEM;
}
memcpy(req->tmpbuf, (void *)buf, len);
buf = req->tmpbuf;
} else {
PSMX2_CTXT_TYPE(&req->fi_context) = no_event ?
PSMX2_NOCOMP_WRITE_CONTEXT :
PSMX2_WRITE_CONTEXT;
}
req->no_event = no_event;
req->op = PSMX2_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;
PSMX2_CTXT_USER(&req->fi_context) = context;
PSMX2_CTXT_EP(&req->fi_context) = ep_priv;
chunk_size = ep_priv->tx->psm2_am_param.max_request_short;
args[0].u32w0 = 0;
if (psmx2_env.tagged_rma && len > chunk_size) {
PSMX2_SET_TAG(psm2_tag, (uint64_t)req, 0, PSMX2_RMA_TYPE_WRITE);
PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_WRITE_LONG);
args[0].u32w1 = len;
args[1].u64 = (uint64_t)req;
args[2].u64 = addr;
args[3].u64 = key;
nargs = 4;
if (flags & FI_REMOTE_CQ_DATA) {
PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_DATA);
args[4].u64 = data;
nargs++;
}
if (flags & FI_DELIVERY_COMPLETE) {
args[0].u32w0 |= PSMX2_AM_FORCE_ACK;
psm2_context = NULL;
} else {
psm2_context = (void *)&req->fi_context;
}
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args,
nargs, NULL, 0, am_flags, NULL, NULL);
psm2_mq_isend2(ep_priv->tx->psm2_mq, psm2_epaddr, 0,
&psm2_tag, buf, len, psm2_context, &psm2_req);
return 0;
}
PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_WRITE);
nargs = 4;
while (len > chunk_size) {
args[0].u32w1 = chunk_size;
args[1].u64 = (uint64_t)(uintptr_t)req;
args[2].u64 = addr;
args[3].u64 = key;
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args,
nargs, (void *)buf, chunk_size, am_flags,
NULL, NULL);
buf = (const uint8_t *)buf + chunk_size;
addr += chunk_size;
len -= chunk_size;
}
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) {
PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_DATA | PSMX2_AM_EOM);
args[4].u64 = data;
nargs++;
} else {
PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_EOM);
}
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, nargs,
(void *)buf, len, am_flags, NULL, NULL);
return 0;
}
ssize_t psmx2_tagged_recv_generic(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 psmx2_fid_ep *ep_priv;
struct psmx2_fid_av *av;
psm2_epaddr_t psm2_epaddr;
uint8_t vlane;
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag, psm2_tagsel;
uint32_t tag32, tagsel32;
struct fi_context *fi_context;
size_t idx;
int err;
ep_priv = container_of(ep, struct psmx2_fid_ep, ep);
if (flags & FI_PEEK)
return psmx2_tagged_peek_generic(ep, buf, len, desc,
src_addr, tag, ignore,
context, flags);
if (flags & FI_TRIGGER) {
struct psmx2_trigger *trigger;
struct fi_triggered_context *ctxt = context;
trigger = calloc(1, sizeof(*trigger));
if (!trigger)
return -FI_ENOMEM;
trigger->op = PSMX2_TRIGGERED_TRECV;
trigger->cntr = container_of(ctxt->trigger.threshold.cntr,
struct psmx2_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;
psmx2_cntr_add_trigger(trigger->cntr, trigger);
return 0;
}
if (flags & FI_CLAIM) {
if (!context)
return -FI_EINVAL;
/* TODO: handle FI_DISCARD */
fi_context = context;
psm2_req = PSMX2_CTXT_REQ(fi_context);
PSMX2_CTXT_TYPE(fi_context) = PSMX2_TRECV_CONTEXT;
PSMX2_CTXT_USER(fi_context) = buf;
PSMX2_CTXT_EP(fi_context) = ep_priv;
err = psm2_mq_imrecv(ep_priv->domain->psm2_mq, 0,
buf, len, context, &psm2_req);
if (err != PSM2_OK)
return psmx2_errno(err);
PSMX2_CTXT_REQ(fi_context) = psm2_req;
return 0;
}
if (ep_priv->recv_selective_completion && !(flags & FI_COMPLETION)) {
fi_context = psmx2_ep_get_op_context(ep_priv);
PSMX2_CTXT_TYPE(fi_context) = PSMX2_NOCOMP_RECV_CONTEXT_ALLOC;
PSMX2_CTXT_EP(fi_context) = ep_priv;
PSMX2_CTXT_USER(fi_context) = buf;
PSMX2_CTXT_SIZE(fi_context) = len;
} else {
if (!context)
return -FI_EINVAL;
fi_context = context;
PSMX2_CTXT_TYPE(fi_context) = PSMX2_TRECV_CONTEXT;
PSMX2_CTXT_USER(fi_context) = buf;
PSMX2_CTXT_EP(fi_context) = ep_priv;
PSMX2_CTXT_SIZE(fi_context) = len;
}
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;
psm2_epaddr = av->epaddrs[idx];
vlane = av->vlanes[idx];
} else {
psm2_epaddr = PSMX2_ADDR_TO_EP(src_addr);
vlane = PSMX2_ADDR_TO_VL(src_addr);
}
tag32 = PSMX2_TAG32(0, vlane, ep_priv->vlane);
tagsel32 = ~PSMX2_IOV_BIT;
} else {
psm2_epaddr = 0;
tag32 = PSMX2_TAG32(0, 0, ep_priv->vlane);
tagsel32 = ~(PSMX2_IOV_BIT | PSMX2_SRC_BITS);
}
PSMX2_SET_TAG(psm2_tag, tag, tag32);
PSMX2_SET_TAG(psm2_tagsel, ~ignore, tagsel32);
err = psm2_mq_irecv2(ep_priv->domain->psm2_mq, psm2_epaddr,
&psm2_tag, &psm2_tagsel, 0, buf, len,
(void *)fi_context, &psm2_req);
if (err != PSM2_OK)
return psmx2_errno(err);
if (fi_context == context)
PSMX2_CTXT_REQ(fi_context) = psm2_req;
return 0;
}
ssize_t psmx2_sendv_generic(struct fid_ep *ep, const struct iovec *iov,
void *desc, size_t count, fi_addr_t dest_addr,
void *context, uint64_t flags, uint64_t data)
{
struct psmx2_fid_ep *ep_priv;
struct psmx2_fid_av *av;
psm2_epaddr_t psm2_epaddr;
uint8_t vlane;
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag;
uint32_t tag32, tag32_base;
struct fi_context * fi_context;
int send_flag = 0;
int err;
size_t idx;
int no_completion = 0;
struct psmx2_cq_event *event;
size_t real_count;
size_t len, total_len;
char *p;
uint32_t *q;
int i;
struct psmx2_sendv_request *req;
ep_priv = container_of(ep, struct psmx2_fid_ep, ep);
if (flags & FI_TRIGGER) {
struct psmx2_trigger *trigger;
struct fi_triggered_context *ctxt = context;
trigger = calloc(1, sizeof(*trigger));
if (!trigger)
return -FI_ENOMEM;
trigger->op = PSMX2_TRIGGERED_SENDV;
trigger->cntr = container_of(ctxt->trigger.threshold.cntr,
struct psmx2_fid_cntr, cntr);
trigger->threshold = ctxt->trigger.threshold.threshold;
trigger->sendv.ep = ep;
trigger->sendv.iov = iov;
trigger->sendv.desc = desc;
trigger->sendv.count = count;
trigger->sendv.dest_addr = dest_addr;
trigger->sendv.context = context;
trigger->sendv.flags = flags & ~FI_TRIGGER;
trigger->sendv.data = data;
psmx2_cntr_add_trigger(trigger->cntr, trigger);
return 0;
}
total_len = 0;
real_count = 0;
for (i=0; i<count; i++) {
if (iov[i].iov_len) {
total_len += iov[i].iov_len;
real_count++;
}
}
req = malloc(sizeof(*req));
if (!req)
return -FI_ENOMEM;
if (total_len <= PSMX2_IOV_BUF_SIZE) {
req->iov_protocol = PSMX2_IOV_PROTO_PACK;
p = req->buf;
for (i=0; i<count; i++) {
if (iov[i].iov_len) {
memcpy(p, iov[i].iov_base, iov[i].iov_len);
p += iov[i].iov_len;
}
}
tag32_base = PSMX2_MSG_BIT;
len = total_len;
} else {
req->iov_protocol = PSMX2_IOV_PROTO_MULTI;
req->iov_done = 0;
req->iov_info.seq_num = (++ep_priv->iov_seq_num) %
PSMX2_IOV_MAX_SEQ_NUM + 1;
req->iov_info.count = (uint32_t)real_count;
req->iov_info.total_len = (uint32_t)total_len;
q = req->iov_info.len;
for (i=0; i<count; i++) {
if (iov[i].iov_len)
*q++ = (uint32_t)iov[i].iov_len;
}
tag32_base = PSMX2_MSG_BIT | PSMX2_IOV_BIT;
len = (3 + real_count) * sizeof(uint32_t);
}
av = ep_priv->av;
if (av && av->type == FI_AV_TABLE) {
idx = (size_t)dest_addr;
if (idx >= av->last) {
free(req);
return -FI_EINVAL;
}
psm2_epaddr = av->epaddrs[idx];
vlane = av->vlanes[idx];
} else {
psm2_epaddr = PSMX2_ADDR_TO_EP(dest_addr);
vlane = PSMX2_ADDR_TO_VL(dest_addr);
}
tag32 = PSMX2_TAG32(tag32_base, ep_priv->vlane, vlane);
if (flags & FI_REMOTE_CQ_DATA)
tag32 |= PSMX2_IMM_BIT;
PSMX2_SET_TAG(psm2_tag, data, tag32);
if ((flags & PSMX2_NO_COMPLETION) ||
(ep_priv->send_selective_completion && !(flags & FI_COMPLETION)))
no_completion = 1;
if (flags & FI_INJECT) {
if (len > PSMX2_INJECT_SIZE) {
free(req);
return -FI_EMSGSIZE;
}
err = psm2_mq_send2(ep_priv->domain->psm2_mq, psm2_epaddr,
send_flag, &psm2_tag, req->buf, len);
free(req);
if (err != PSM2_OK)
return psmx2_errno(err);
if (ep_priv->send_cntr)
psmx2_cntr_inc(ep_priv->send_cntr);
if (ep_priv->send_cq && !no_completion) {
event = psmx2_cq_create_event(
ep_priv->send_cq,
context, NULL, flags, len,
(uint64_t) data,
0 /* tag */,
0 /* olen */,
0 /* err */);
if (event)
psmx2_cq_enqueue_event(ep_priv->send_cq, event);
else
return -FI_ENOMEM;
}
return 0;
}
req->no_completion = no_completion;
req->user_context = context;
req->comp_flag = FI_MSG;
fi_context = &req->fi_context;
PSMX2_CTXT_TYPE(fi_context) = PSMX2_SENDV_CONTEXT;
PSMX2_CTXT_USER(fi_context) = req;
PSMX2_CTXT_EP(fi_context) = ep_priv;
err = psm2_mq_isend2(ep_priv->domain->psm2_mq, psm2_epaddr,
send_flag, &psm2_tag, req->buf, len,
(void *)fi_context, &psm2_req);
if (err != PSM2_OK) {
free(req);
return psmx2_errno(err);
}
PSMX2_CTXT_REQ(fi_context) = psm2_req;
if (req->iov_protocol == PSMX2_IOV_PROTO_MULTI) {
fi_context = &req->fi_context_iov;
PSMX2_CTXT_TYPE(fi_context) = PSMX2_IOV_SEND_CONTEXT;
PSMX2_CTXT_USER(fi_context) = req;
PSMX2_CTXT_EP(fi_context) = ep_priv;
tag32 &= ~PSMX2_IOV_BIT;
PSMX2_TAG32_SET_SEQ(tag32, req->iov_info.seq_num);
PSMX2_SET_TAG(psm2_tag, data, tag32);
for (i=0; i<count; i++) {
if (iov[i].iov_len) {
err = psm2_mq_isend2(ep_priv->domain->psm2_mq,
psm2_epaddr, send_flag, &psm2_tag,
iov[i].iov_base, iov[i].iov_len,
(void *)fi_context, &psm2_req);
if (err != PSM2_OK)
return psmx2_errno(err);
}
}
}
return 0;
}
ssize_t psmx2_writev_generic(struct fid_ep *ep, const struct iovec *iov,
void **desc, size_t count, fi_addr_t dest_addr,
uint64_t addr, uint64_t key, void *context,
uint64_t flags, uint64_t data)
{
struct psmx2_fid_ep *ep_priv;
struct psmx2_fid_av *av;
struct psmx2_epaddr_context *epaddr_context;
struct psmx2_am_request *req;
psm2_amarg_t args[8];
int nargs;
int am_flags = PSM2_AM_FLAG_ASYNC;
int chunk_size;
psm2_epaddr_t psm2_epaddr;
uint8_t vlane;
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag;
uint32_t tag32;
size_t idx;
void *psm2_context;
int no_event;
size_t total_len, len, len_sent;
uint8_t *buf, *p;
int i;
ep_priv = container_of(ep, struct psmx2_fid_ep, ep);
if (flags & FI_TRIGGER) {
struct psmx2_trigger *trigger;
struct fi_triggered_context *ctxt = context;
trigger = calloc(1, sizeof(*trigger));
if (!trigger)
return -FI_ENOMEM;
trigger->op = PSMX2_TRIGGERED_WRITEV;
trigger->cntr = container_of(ctxt->trigger.threshold.cntr,
struct psmx2_fid_cntr, cntr);
trigger->threshold = ctxt->trigger.threshold.threshold;
trigger->writev.ep = ep;
trigger->writev.iov = iov;
trigger->writev.count = count;
trigger->writev.desc = desc;
trigger->writev.dest_addr = dest_addr;
trigger->writev.addr = addr;
trigger->writev.key = key;
trigger->writev.context = context;
trigger->writev.flags = flags & ~FI_TRIGGER;
trigger->writev.data = data;
psmx2_cntr_add_trigger(trigger->cntr, trigger);
return 0;
}
av = ep_priv->av;
if (av && av->type == FI_AV_TABLE) {
idx = dest_addr;
if (idx >= av->last)
return -FI_EINVAL;
psm2_epaddr = av->epaddrs[idx];
vlane = av->vlanes[idx];
} else {
if (!dest_addr)
return -FI_EINVAL;
psm2_epaddr = PSMX2_ADDR_TO_EP(dest_addr);
vlane = PSMX2_ADDR_TO_VL(dest_addr);
}
epaddr_context = psm2_epaddr_getctxt((void *)psm2_epaddr);
if (epaddr_context->epid == ep_priv->domain->psm2_epid)
return psmx2_rma_self(PSMX2_AM_REQ_WRITEV, ep_priv,
ep_priv->domain->eps[vlane],
(void *)iov, count, desc, addr,
key, context, flags, data);
no_event = (flags & PSMX2_NO_COMPLETION) ||
(ep_priv->send_selective_completion && !(flags & FI_COMPLETION));
total_len = 0;
for (i=0; i<count; i++)
total_len += iov[i].iov_len;
chunk_size = psmx2_am_param.max_request_short;
/* Case 1: fit into a AM message, then pack and send */
if (total_len <= chunk_size) {
req = malloc(sizeof(*req) + total_len);
if (!req)
return -FI_ENOMEM;
memset(req, 0, sizeof(*req));
p = (uint8_t *)req + sizeof(*req);
for (i=0; i<count; i++) {
if (iov[i].iov_len) {
memcpy(p, iov[i].iov_base, iov[i].iov_len);
p += iov[i].iov_len;
}
}
buf = (uint8_t *)req + sizeof(*req);
len = total_len;
req->no_event = no_event;
req->op = PSMX2_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;
PSMX2_CTXT_USER(&req->fi_context) = context;
PSMX2_CTXT_EP(&req->fi_context) = ep_priv;
args[0].u32w0 = 0;
PSMX2_AM_SET_SRC(args[0].u32w0, ep_priv->vlane);
PSMX2_AM_SET_DST(args[0].u32w0, vlane);
PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_WRITE);
args[0].u32w1 = len;
args[1].u64 = (uint64_t)(uintptr_t)req;
args[2].u64 = addr;
args[3].u64 = key;
nargs = 4;
if (flags & FI_REMOTE_CQ_DATA) {
PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_DATA | PSMX2_AM_EOM);
args[4].u64 = data;
nargs++;
} else {
PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_EOM);
}
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, nargs,
(void *)buf, len, am_flags, NULL, NULL);
return 0;
}
if (flags & FI_INJECT)
return -FI_EMSGSIZE;
req = calloc(1, sizeof(*req));
if (!req)
return -FI_ENOMEM;
PSMX2_CTXT_TYPE(&req->fi_context) = no_event ?
PSMX2_NOCOMP_WRITE_CONTEXT :
PSMX2_WRITE_CONTEXT;
req->no_event = no_event;
req->op = PSMX2_AM_REQ_WRITE;
req->write.buf = (void *)iov[0].iov_base;
req->write.len = total_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;
PSMX2_CTXT_USER(&req->fi_context) = context;
PSMX2_CTXT_EP(&req->fi_context) = ep_priv;
/* Case 2: send iov in sequence */
args[0].u32w0 = 0;
PSMX2_AM_SET_SRC(args[0].u32w0, ep_priv->vlane);
PSMX2_AM_SET_DST(args[0].u32w0, vlane);
len_sent = 0;
for (i=0; i<count; i++) {
if (!iov[i].iov_len)
continue;
/* Case 2.1: use long protocol for the last segment if it is large */
if (psmx2_env.tagged_rma && iov[i].iov_len > chunk_size &&
len_sent + iov[i].iov_len == total_len) {
tag32 = PSMX2_TAG32(PSMX2_RMA_BIT, ep_priv->vlane, vlane);
PSMX2_SET_TAG(psm2_tag, (uint64_t)req, tag32);
PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_WRITE_LONG);
args[0].u32w1 = iov[i].iov_len;
args[1].u64 = (uint64_t)req;
args[2].u64 = addr;
args[3].u64 = key;
nargs = 4;
if (flags & FI_REMOTE_CQ_DATA) {
PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_DATA);
args[4].u64 = data;
nargs++;
}
if (flags & FI_DELIVERY_COMPLETE) {
args[0].u32w0 |= PSMX2_AM_FORCE_ACK;
psm2_context = NULL;
} else {
psm2_context = (void *)&req->fi_context;
}
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args,
nargs, NULL, 0, am_flags, NULL, NULL);
psm2_mq_isend2(ep_priv->domain->psm2_mq, psm2_epaddr, 0,
&psm2_tag, iov[i].iov_base, iov[i].iov_len,
psm2_context, &psm2_req);
return 0;
}
/* Case 2.2: use short protocol all other segments */
PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_WRITE);
nargs = 4;
buf = iov[i].iov_base;
len = iov[i].iov_len;
while (len > chunk_size) {
args[0].u32w1 = chunk_size;
args[1].u64 = (uint64_t)(uintptr_t)req;
args[2].u64 = addr;
args[3].u64 = key;
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args,
nargs, (void *)buf, chunk_size, am_flags,
NULL, NULL);
buf += chunk_size;
addr += chunk_size;
len -= chunk_size;
len_sent += chunk_size;
}
args[0].u32w1 = len;
args[1].u64 = (uint64_t)(uintptr_t)req;
args[2].u64 = addr;
args[3].u64 = key;
if (len_sent + len == total_len) {
if (flags & FI_REMOTE_CQ_DATA) {
PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_DATA | PSMX2_AM_EOM);
args[4].u64 = data;
nargs++;
} else {
PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_EOM);
}
}
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, nargs,
(void *)buf, len, am_flags, NULL, NULL);
addr += len;
len_sent += len;
}
return 0;
}
int psmx2_handle_sendv_req(struct psmx2_fid_ep *ep,
psm2_mq_status2_t *psm2_status,
int multi_recv)
{
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag, psm2_tagsel;
struct psmx2_sendv_reply *rep;
struct psmx2_multi_recv *recv_req;
struct fi_context *fi_context;
struct fi_context *recv_context;
int i, err;
uint8_t *recv_buf;
size_t recv_len, len;
if (psm2_status->error_code != PSM2_OK)
return psmx2_errno(psm2_status->error_code);
rep = malloc(sizeof(*rep));
if (!rep) {
psm2_status->error_code = PSM2_NO_MEMORY;
return -FI_ENOMEM;
}
recv_context = psm2_status->context;
if (multi_recv) {
recv_req = PSMX2_CTXT_USER(recv_context);
recv_buf = recv_req->buf + recv_req->offset;
recv_len = recv_req->len - recv_req->offset;
rep->multi_recv = 1;
} else {
recv_buf = PSMX2_CTXT_USER(recv_context);
recv_len = PSMX2_CTXT_SIZE(recv_context);
rep->multi_recv = 0;
}
/* assert(psm2_status->nbytes <= PSMX2_IOV_BUF_SIZE */
memcpy(&rep->iov_info, recv_buf, psm2_status->nbytes);
rep->user_context = psm2_status->context;
rep->buf = recv_buf;
rep->no_completion = 0;
rep->iov_done = 0;
rep->bytes_received = 0;
rep->msg_length = 0;
rep->error_code = PSM2_OK;
fi_context = &rep->fi_context;
PSMX2_CTXT_TYPE(fi_context) = PSMX2_IOV_RECV_CONTEXT;
PSMX2_CTXT_USER(fi_context) = rep;
PSMX2_CTXT_EP(fi_context) = ep;
/* use the same tag, with IOV bit cleared, and seq_num added */
psm2_tag = psm2_status->msg_tag;
psm2_tag.tag2 &= ~PSMX2_IOV_BIT;
PSMX2_TAG32_SET_SEQ(psm2_tag.tag2, rep->iov_info.seq_num);
rep->comp_flag = (psm2_tag.tag2 & PSMX2_MSG_BIT) ? FI_MSG : FI_TAGGED;
if (psm2_tag.tag2 & PSMX2_IMM_BIT)
rep->comp_flag |= FI_REMOTE_CQ_DATA;
/* match every bit of the tag */
PSMX2_SET_TAG(psm2_tagsel, -1UL, -1);
for (i=0; i<rep->iov_info.count; i++) {
if (recv_len) {
len = MIN(recv_len, rep->iov_info.len[i]);
err = psm2_mq_irecv2(ep->domain->psm2_mq,
psm2_status->msg_peer,
&psm2_tag, &psm2_tagsel,
0/*flag*/, recv_buf, len,
(void *)fi_context, &psm2_req);
if (err) {
psm2_status->error_code = err;
return psmx2_errno(psm2_status->error_code);
}
recv_buf += len;
recv_len -= len;
} else {
/* recv buffer full, pust empty recvs */
err = psm2_mq_irecv2(ep->domain->psm2_mq,
psm2_status->msg_peer,
&psm2_tag, &psm2_tagsel,
0/*flag*/, NULL, 0,
(void *)fi_context, &psm2_req);
if (err) {
psm2_status->error_code = err;
return psmx2_errno(psm2_status->error_code);
}
}
}
return 0;
}
ssize_t psmx2_read_generic(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 psmx2_fid_ep *ep_priv;
struct psmx2_fid_av *av;
struct psmx2_epaddr_context *epaddr_context;
struct psmx2_am_request *req;
psm2_amarg_t args[8];
int chunk_size;
size_t offset = 0;
psm2_epaddr_t psm2_epaddr;
uint8_t vlane;
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag, psm2_tagsel;
uint32_t tag32;
size_t idx;
ep_priv = container_of(ep, struct psmx2_fid_ep, ep);
if (flags & FI_TRIGGER) {
struct psmx2_trigger *trigger;
struct fi_triggered_context *ctxt = context;
trigger = calloc(1, sizeof(*trigger));
if (!trigger)
return -FI_ENOMEM;
trigger->op = PSMX2_TRIGGERED_READ;
trigger->cntr = container_of(ctxt->trigger.threshold.cntr,
struct psmx2_fid_cntr, cntr);
trigger->threshold = ctxt->trigger.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;
psmx2_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 = src_addr;
if (idx >= av->last)
return -FI_EINVAL;
psm2_epaddr = av->epaddrs[idx];
vlane = av->vlanes[idx];
} else {
if (!src_addr)
return -FI_EINVAL;
psm2_epaddr = PSMX2_ADDR_TO_EP(src_addr);
vlane = PSMX2_ADDR_TO_VL(src_addr);
}
epaddr_context = psm2_epaddr_getctxt((void *)psm2_epaddr);
if (epaddr_context->epid == ep_priv->domain->psm2_epid)
return psmx2_rma_self(PSMX2_AM_REQ_READ, ep_priv,
ep_priv->domain->eps[vlane],
buf, len, desc, addr, key,
context, flags, 0);
req = calloc(1, sizeof(*req));
if (!req)
return -FI_ENOMEM;
req->op = PSMX2_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;
req->cq_flags = FI_READ | FI_RMA;
PSMX2_CTXT_TYPE(&req->fi_context) = PSMX2_READ_CONTEXT;
PSMX2_CTXT_USER(&req->fi_context) = context;
PSMX2_CTXT_EP(&req->fi_context) = ep_priv;
if (ep_priv->send_selective_completion && !(flags & FI_COMPLETION)) {
PSMX2_CTXT_TYPE(&req->fi_context) = PSMX2_NOCOMP_READ_CONTEXT;
req->no_event = 1;
}
chunk_size = psmx2_am_param.max_reply_short;
args[0].u32w0 = 0;
PSMX2_AM_SET_SRC(args[0].u32w0, ep_priv->vlane);
PSMX2_AM_SET_DST(args[0].u32w0, vlane);
if (psmx2_env.tagged_rma && len > chunk_size) {
tag32 = PSMX2_TAG32(PSMX2_RMA_BIT, vlane, ep_priv->vlane);
PSMX2_SET_TAG(psm2_tag, (uint64_t)req, tag32);
PSMX2_SET_TAG(psm2_tagsel, -1ULL, -1);
psm2_mq_irecv2(ep_priv->domain->psm2_mq, psm2_epaddr,
&psm2_tag, &psm2_tagsel, 0, buf, len,
(void *)&req->fi_context, &psm2_req);
PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_READ_LONG);
args[0].u32w1 = len;
args[1].u64 = (uint64_t)req;
args[2].u64 = addr;
args[3].u64 = key;
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER,
args, 4, NULL, 0, 0, NULL, NULL);
return 0;
}
PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_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;
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER,
args, 5, NULL, 0, 0, NULL, NULL);
addr += chunk_size;
len -= chunk_size;
offset += chunk_size;
}
PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_EOM);
args[0].u32w1 = len;
args[2].u64 = addr;
args[4].u64 = offset;
psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER,
args, 5, NULL, 0, 0, NULL, NULL);
return 0;
}
int psmx2_ep_open_internal(struct psmx2_fid_domain *domain_priv,
struct fi_info *info,
struct psmx2_fid_ep **ep_out, void *context,
struct psmx2_trx_ctxt *trx_ctxt)
{
struct psmx2_fid_ep *ep_priv;
uint64_t ep_cap;
int err = -FI_EINVAL;
if (info)
ep_cap = info->caps;
else
ep_cap = FI_TAGGED;
if (info && info->ep_attr && info->ep_attr->auth_key) {
if (info->ep_attr->auth_key_size != sizeof(psm2_uuid_t)) {
FI_WARN(&psmx2_prov, FI_LOG_EP_CTRL,
"Invalid auth_key_len %"PRIu64
", should be %"PRIu64".\n",
info->ep_attr->auth_key_size,
sizeof(psm2_uuid_t));
goto errout;
}
if (memcmp(domain_priv->fabric->uuid, info->ep_attr->auth_key,
sizeof(psm2_uuid_t))) {
FI_WARN(&psmx2_prov, FI_LOG_EP_CTRL,
"Invalid auth_key: %s\n",
psmx2_uuid_to_string((void *)info->ep_attr->auth_key));
goto errout;
}
}
ep_priv = (struct psmx2_fid_ep *) calloc(1, sizeof *ep_priv);
if (!ep_priv) {
err = -FI_ENOMEM;
goto errout;
}
ep_priv->ep.fid.fclass = FI_CLASS_EP;
ep_priv->ep.fid.context = context;
ep_priv->ep.fid.ops = &psmx2_fi_ops;
ep_priv->ep.ops = &psmx2_ep_ops;
ep_priv->ep.cm = &psmx2_cm_ops;
ep_priv->domain = domain_priv;
ep_priv->rx = trx_ctxt;
if (!(info && info->ep_attr && info->ep_attr->tx_ctx_cnt == FI_SHARED_CONTEXT))
ep_priv->tx = trx_ctxt;
ofi_atomic_initialize32(&ep_priv->ref, 0);
PSMX2_CTXT_TYPE(&ep_priv->nocomp_send_context) = PSMX2_NOCOMP_SEND_CONTEXT;
PSMX2_CTXT_EP(&ep_priv->nocomp_send_context) = ep_priv;
PSMX2_CTXT_TYPE(&ep_priv->nocomp_tsend_context) = PSMX2_NOCOMP_TSEND_CONTEXT;
PSMX2_CTXT_EP(&ep_priv->nocomp_tsend_context) = ep_priv;
if (ep_cap & FI_TAGGED)
ep_priv->ep.tagged = &psmx2_tagged_ops;
if (ep_cap & FI_MSG)
ep_priv->ep.msg = &psmx2_msg_ops;
if (ep_cap & FI_RMA)
ep_priv->ep.rma = &psmx2_rma_ops;
if (ep_cap & FI_ATOMICS)
ep_priv->ep.atomic = &psmx2_atomic_ops;
ep_priv->caps = ep_cap;
err = psmx2_domain_enable_ep(domain_priv, ep_priv);
if (err)
goto errout_free_ep;
psmx2_domain_acquire(domain_priv);
if (info) {
if (info->tx_attr)
ep_priv->tx_flags = info->tx_attr->op_flags;
if (info->rx_attr)
ep_priv->rx_flags = info->rx_attr->op_flags;
}
psmx2_ep_optimize_ops(ep_priv);
PSMX2_EP_INIT_OP_CONTEXT(ep_priv);
*ep_out = ep_priv;
return 0;
errout_free_ep:
free(ep_priv);
errout:
return err;
}
ssize_t psmx2_send_generic(struct fid_ep *ep, const void *buf, size_t len,
void *desc, fi_addr_t dest_addr, void *context,
uint64_t flags, uint64_t data)
{
struct psmx2_fid_ep *ep_priv;
struct psmx2_fid_av *av;
psm2_epaddr_t psm2_epaddr;
psm2_mq_req_t psm2_req;
psm2_mq_tag_t psm2_tag;
struct fi_context * fi_context;
int send_flag = 0;
int err;
int no_completion = 0;
struct psmx2_cq_event *event;
int have_data = (flags & FI_REMOTE_CQ_DATA) > 0;
ep_priv = container_of(ep, struct psmx2_fid_ep, ep);
if (flags & FI_TRIGGER)
return psmx2_trigger_queue_send(ep, buf, len, desc, dest_addr,
context, flags, data);
av = ep_priv->av;
assert(av);
psm2_epaddr = psmx2_av_translate_addr(av, ep_priv->tx, dest_addr, av->type);
PSMX2_SET_TAG(psm2_tag, 0, data, PSMX2_TYPE_MSG | PSMX2_IMM_BIT_SET(have_data));
if ((flags & PSMX2_NO_COMPLETION) ||
(ep_priv->send_selective_completion && !(flags & FI_COMPLETION)))
no_completion = 1;
if (flags & FI_INJECT) {
if (len > psmx2_env.inject_size)
return -FI_EMSGSIZE;
err = psm2_mq_send2(ep_priv->tx->psm2_mq, psm2_epaddr,
send_flag, &psm2_tag, buf, len);
if (err != PSM2_OK)
return psmx2_errno(err);
if (ep_priv->send_cntr)
psmx2_cntr_inc(ep_priv->send_cntr, 0);
if (ep_priv->send_cq && !no_completion) {
event = psmx2_cq_create_event(
ep_priv->send_cq,
context, (void *)buf, flags, len,
(uint64_t) data,
0 /* tag */,
0 /* olen */,
0 /* err */);
if (event)
psmx2_cq_enqueue_event(ep_priv->send_cq, event);
else
return -FI_ENOMEM;
}
return 0;
}
if (no_completion) {
fi_context = &ep_priv->nocomp_send_context;
} else {
assert(context);
fi_context = context;
PSMX2_CTXT_TYPE(fi_context) = PSMX2_SEND_CONTEXT;
PSMX2_CTXT_USER(fi_context) = (void *)buf;
PSMX2_CTXT_EP(fi_context) = ep_priv;
}
err = psm2_mq_isend2(ep_priv->tx->psm2_mq, psm2_epaddr,
send_flag, &psm2_tag, buf, len,
(void *)fi_context, &psm2_req);
if (err != PSM2_OK)
return psmx2_errno(err);
if (fi_context == context)
PSMX2_CTXT_REQ(fi_context) = psm2_req;
return 0;
}