prov/efa: Enable inject rdma write

Turns off emulated inject rdma write, and enables real inject rdma
write. Inject rdma write is special for two reasons:

1. The user gets to re-use their immediately after we return
2. RDMA operations require MR's, and the inject write interfaces do not
   provide a memory descritpor

In order to handle both cases, the users data is copied into a
pre-registered bounce buffer. The bounce buffer is allocated to be the
sizeof(struct efa_rdm_pke) + ep->mtu_size, where the wire data starts at
ep->mtu_size. The RDMA operations use part of the wire data for a RDMA
header struct efa_rdm_rma_context_pkt.  This change needs to lower the
maximum inject size to never be more than mtu_size - sizeof(struct
efa_rdm_rma_context_pkt). I created a github issue for libfabric 2.0 API
to seperate inject sizes for send/recv, atomic, and rdma operations
(issue 9510).

Signed-off-by: Seth Zegelstein <szegel@amazon.com>

prov/efa/src/rdm/efa_rdm_ope.c

@@ -1451,6 +1451,7 @@ int efa_rdm_ope_post_remote_write(struct efa_rdm_ope *ope)
 {
 	int err;
 	int iov_idx = 0, rma_iov_idx = 0;
+	ssize_t copied;
 	size_t iov_offset = 0, rma_iov_offset = 0;
 	size_t write_once_len, max_write_once_len;
 	struct efa_rdm_ep *ep;
@@ -1485,7 +1486,9 @@ int efa_rdm_ope_post_remote_write(struct efa_rdm_ope *ope)
 		return err;
 	}
 
-	efa_rdm_ope_try_fill_desc(ope, 0, FI_WRITE);
+	if (!(ope->fi_flags & FI_INJECT))
+		efa_rdm_ope_try_fill_desc(ope, 0, FI_WRITE);
+
 	assert(ope->bytes_write_submitted < ope->bytes_write_total_len);
 	max_write_once_len = MIN(efa_env.efa_write_segment_size, efa_rdm_ep_domain(ep)->device->max_rdma_size);
 
@@ -1517,7 +1520,7 @@ int efa_rdm_ope_post_remote_write(struct efa_rdm_ope *ope)
 		if (ep->efa_outstanding_tx_ops == ep->efa_max_outstanding_tx_ops)
 			return -FI_EAGAIN;
 
-		if (!ope->desc[iov_idx]) {
+		if (!ope->desc[iov_idx] && !(ope->fi_flags & FI_INJECT)) {
 			/* efa_rdm_ope_try_fill_desc() did not fill the desc,
 			 * which means memory registration failed.
 			 * return -FI_EAGAIN here will cause user to run progress
@@ -1531,6 +1534,16 @@ int efa_rdm_ope_post_remote_write(struct efa_rdm_ope *ope)
 		if (OFI_UNLIKELY(!pkt_entry))
 			return -FI_EAGAIN;
 
+		if (ope->fi_flags & FI_INJECT) {
+			assert(ope->iov_count == 1);
+			assert(ope->total_len <= ep->inject_size);
+			copied = ofi_copy_from_hmem_iov(pkt_entry->wiredata + sizeof(struct efa_rdm_rma_context_pkt),
+				ope->total_len, FI_HMEM_SYSTEM, 0, ope->iov, ope->iov_count, 0);
+			assert(copied == ope->total_len);
+			ope->desc[0] = fi_mr_desc(pkt_entry->mr);
+			ope->iov[0].iov_base = pkt_entry->wiredata + sizeof(struct efa_rdm_rma_context_pkt);
+		}
+
 		write_once_len = MIN(ope->iov[iov_idx].iov_len - iov_offset,
 				    ope->rma_iov[rma_iov_idx].len - rma_iov_offset);
 		write_once_len = MIN(write_once_len, max_write_once_len);

prov/efa/src/rdm/efa_rdm_pkt_type.c

@@ -4,6 +4,7 @@
 #include "efa_rdm_peer.h"
 #include "efa_rdm_protocol.h"
 #include "efa_rdm_pkt_type.h"
+#include "efa_rdm_pke_nonreq.h"
 
 struct efa_rdm_pkt_type_req_info EFA_RDM_PKT_TYPE_REQ_INFO_VEC[] = {
 	/* rtm header */
@@ -137,5 +138,8 @@ size_t efa_rdm_pkt_type_get_max_hdr_size(void)
 			pkt_type += 1;
 	}
 
+	/* Non-emulated (real) rdma inject write requires a header */
+	max_hdr_size = MAX(max_hdr_size, sizeof(struct efa_rdm_rma_context_pkt));
+
 	return max_hdr_size;
 }

prov/efa/src/rdm/efa_rdm_rma.c

@@ -317,14 +317,6 @@ ssize_t efa_rdm_rma_read(struct fid_ep *ep, void *buf, size_t len, void *desc,
 static inline
 bool efa_rdm_rma_should_write_using_rdma(struct efa_rdm_ep *ep, struct efa_rdm_ope *txe, struct efa_rdm_peer *peer)
 {
-	/*
-	 * RDMA_WRITE does not support FI_INJECT, because device may
-	 * need to re-send data and FI_INJECT allows user to re-use
-	 * these buffers immediately.
-	 */
-	if (txe->fi_flags & FI_INJECT)
-		return false;
-
 	/*
 	 * Because EFA is unordered and EFA iov descriptions can be more
 	 * expressive than the IBV sge's, we only implement