ssize_t sock_comm_send(struct sock_pe_entry *pe_entry,
const void *buf, size_t len)
{
ssize_t ret, used;
if (len > pe_entry->cache_sz) {
used = ofi_rbused(&pe_entry->comm_buf);
if (used == sock_comm_flush(pe_entry)) {
return sock_comm_send_socket(pe_entry->conn, buf, len);
} else {
return 0;
}
}
if (ofi_rbavail(&pe_entry->comm_buf) < len) {
ret = sock_comm_flush(pe_entry);
if (ret <= 0)
return 0;
}
ret = MIN(ofi_rbavail(&pe_entry->comm_buf), len);
ofi_rbwrite(&pe_entry->comm_buf, buf, ret);
ofi_rbcommit(&pe_entry->comm_buf);
SOCK_LOG_DBG("buffered %lu\n", ret);
return ret;
}
ssize_t sock_comm_flush(struct sock_pe_entry *pe_entry)
{
ssize_t ret1, ret2 = 0;
size_t endlen, len, xfer_len;
len = ofi_rbused(&pe_entry->comm_buf);
endlen = pe_entry->comm_buf.size -
(pe_entry->comm_buf.rcnt & pe_entry->comm_buf.size_mask);
xfer_len = MIN(len, endlen);
ret1 = sock_comm_send_socket(pe_entry->conn, (char*)pe_entry->comm_buf.buf +
(pe_entry->comm_buf.rcnt & pe_entry->comm_buf.size_mask),
xfer_len);
if (ret1 > 0)
pe_entry->comm_buf.rcnt += ret1;
if (ret1 == xfer_len && xfer_len < len) {
ret2 = sock_comm_send_socket(pe_entry->conn, (char*)pe_entry->comm_buf.buf +
(pe_entry->comm_buf.rcnt & pe_entry->comm_buf.size_mask),
len - xfer_len);
if (ret2 > 0)
pe_entry->comm_buf.rcnt += ret2;
else
ret2 = 0;
}
return (ret1 > 0) ? ret1 + ret2 : 0;
}
ssize_t sock_comm_recv(struct sock_pe_entry *pe_entry, void *buf, size_t len)
{
ssize_t read_len;
if (ofi_rbempty(&pe_entry->comm_buf)) {
if (len <= pe_entry->cache_sz) {
sock_comm_recv_buffer(pe_entry);
} else {
return sock_comm_recv_socket(pe_entry->conn, buf, len);
}
}
read_len = MIN(len, ofi_rbused(&pe_entry->comm_buf));
ofi_rbread(&pe_entry->comm_buf, buf, read_len);
SOCK_LOG_DBG("read from buffer: %lu\n", read_len);
return read_len;
}
static ssize_t sock_cq_readerr(struct fid_cq *cq, struct fi_cq_err_entry *buf,
uint64_t flags)
{
struct sock_cq *sock_cq;
ssize_t ret;
struct fi_cq_err_entry entry;
uint32_t api_version;
size_t err_data_size = 0;
void *err_data = NULL;
sock_cq = container_of(cq, struct sock_cq, cq_fid);
if (sock_cq->domain->progress_mode == FI_PROGRESS_MANUAL)
sock_cq_progress(sock_cq);
fastlock_acquire(&sock_cq->lock);
if (ofi_rbused(&sock_cq->cqerr_rb) >= sizeof(struct fi_cq_err_entry)) {
api_version = sock_cq->domain->fab->fab_fid.api_version;
ofi_rbread(&sock_cq->cqerr_rb, &entry, sizeof(entry));
if ((FI_VERSION_GE(api_version, FI_VERSION(1, 5)))
&& buf->err_data && buf->err_data_size) {
err_data = buf->err_data;
err_data_size = buf->err_data_size;
*buf = entry;
buf->err_data = err_data;
/* Fill provided user's buffer */
buf->err_data_size = MIN(entry.err_data_size, err_data_size);
memcpy(buf->err_data, entry.err_data, buf->err_data_size);
} else {
*buf = entry;
}
ret = 1;
} else {
ret = -FI_EAGAIN;
}
fastlock_release(&sock_cq->lock);
return ret;
}
static ssize_t sock_cq_sreadfrom(struct fid_cq *cq, void *buf, size_t count,
fi_addr_t *src_addr, const void *cond, int timeout)
{
int ret = 0;
size_t threshold;
struct sock_cq *sock_cq;
uint64_t start_ms = 0, end_ms = 0;
ssize_t cq_entry_len, avail;
sock_cq = container_of(cq, struct sock_cq, cq_fid);
if (ofi_rbused(&sock_cq->cqerr_rb))
return -FI_EAVAIL;
cq_entry_len = sock_cq->cq_entry_size;
if (sock_cq->attr.wait_cond == FI_CQ_COND_THRESHOLD)
threshold = MIN((uintptr_t) cond, count);
else
threshold = count;
if (timeout >= 0) {
start_ms = fi_gettime_ms();
end_ms = start_ms + timeout;
}
if (sock_cq->domain->progress_mode == FI_PROGRESS_MANUAL) {
do {
sock_cq_progress(sock_cq);
fastlock_acquire(&sock_cq->lock);
avail = ofi_rbfdused(&sock_cq->cq_rbfd);
if (avail)
ret = sock_cq_rbuf_read(sock_cq, buf,
MIN(threshold, (size_t)(avail / cq_entry_len)),
src_addr, cq_entry_len);
fastlock_release(&sock_cq->lock);
if (ret == 0 && timeout >= 0) {
if (fi_gettime_ms() >= end_ms)
return -FI_EAGAIN;
}
} while (ret == 0);
} else {
do {
ret = ofi_rbfdwait(&sock_cq->cq_rbfd, timeout);
if (ret <= 0)
break;
fastlock_acquire(&sock_cq->lock);
ret = 0;
avail = ofi_rbfdused(&sock_cq->cq_rbfd);
if (avail)
ret = sock_cq_rbuf_read(sock_cq, buf,
MIN(threshold, (size_t)(avail / cq_entry_len)),
src_addr, cq_entry_len);
else /* No CQ entry available, read the fd */
ofi_rbfdreset(&sock_cq->cq_rbfd);
fastlock_release(&sock_cq->lock);
if ((ret == -FI_EAGAIN || ret == 0) && timeout >= 0) {
timeout = end_ms - fi_gettime_ms();
if (timeout <= 0)
break;
}
} while (ret == 0 || ret == -FI_EAGAIN);
}
return (ret == 0 || ret == -FI_ETIMEDOUT) ? -FI_EAGAIN : ret;
}
static ssize_t sock_cq_sreadfrom(struct fid_cq *cq, void *buf, size_t count,
fi_addr_t *src_addr, const void *cond, int timeout)
{
int ret = 0;
size_t threshold;
struct sock_cq *sock_cq;
uint64_t start_ms;
ssize_t cq_entry_len, avail;
sock_cq = container_of(cq, struct sock_cq, cq_fid);
if (ofi_rbused(&sock_cq->cqerr_rb))
return -FI_EAVAIL;
cq_entry_len = sock_cq->cq_entry_size;
if (sock_cq->attr.wait_cond == FI_CQ_COND_THRESHOLD)
threshold = MIN((uintptr_t) cond, count);
else
threshold = count;
start_ms = (timeout >= 0) ? fi_gettime_ms() : 0;
if (sock_cq->domain->progress_mode == FI_PROGRESS_MANUAL) {
while (1) {
sock_cq_progress(sock_cq);
fastlock_acquire(&sock_cq->lock);
avail = ofi_rbfdused(&sock_cq->cq_rbfd);
if (avail) {
ret = sock_cq_rbuf_read(sock_cq, buf,
MIN(threshold, (size_t)(avail / cq_entry_len)),
src_addr, cq_entry_len);
}
fastlock_release(&sock_cq->lock);
if (ret)
return ret;
if (timeout >= 0) {
timeout -= (int) (fi_gettime_ms() - start_ms);
if (timeout <= 0)
return -FI_EAGAIN;
}
if (ofi_atomic_get32(&sock_cq->signaled)) {
ofi_atomic_set32(&sock_cq->signaled, 0);
return -FI_ECANCELED;
}
};
} else {
do {
fastlock_acquire(&sock_cq->lock);
ret = 0;
avail = ofi_rbfdused(&sock_cq->cq_rbfd);
if (avail) {
ret = sock_cq_rbuf_read(sock_cq, buf,
MIN(threshold, (size_t)(avail / cq_entry_len)),
src_addr, cq_entry_len);
} else {
ofi_rbfdreset(&sock_cq->cq_rbfd);
}
fastlock_release(&sock_cq->lock);
if (ret && ret != -FI_EAGAIN)
return ret;
if (timeout >= 0) {
timeout -= (int) (fi_gettime_ms() - start_ms);
if (timeout <= 0)
return -FI_EAGAIN;
}
if (ofi_atomic_get32(&sock_cq->signaled)) {
ofi_atomic_set32(&sock_cq->signaled, 0);
return -FI_ECANCELED;
}
ret = ofi_rbfdwait(&sock_cq->cq_rbfd, timeout);
} while (ret > 0);
}
return (ret == 0 || ret == -FI_ETIMEDOUT) ? -FI_EAGAIN : ret;
}