prov/cxi: Support FI_OPT_CUDA_API_PERMITTED

If GDRCopy is required by the application (ie. it has set FI_OPT_CUDA_API_PERMITTED), and is not available, return not supported, eliminating deadlocks due to calls to cudaMemcpy interacting with CUDA applications. Signed-off-by: Ian Ziemba <[email protected]>
HewlettPackard · Nov 20, 2024 · 3372122 · 3372122
1 parent 628dfa7
commit 3372122
Show file tree

Hide file tree

Showing 11 changed files with 140 additions and 34 deletions.
diff --git a/prov/cxi/include/cxip.h b/prov/cxi/include/cxip.h
@@ -1963,13 +1963,16 @@ struct cxip_rxc_rnr {
 };
 
 static inline void cxip_copy_to_md(struct cxip_md *md, void *dest,
-				   const void *src, size_t size)
+				   const void *src, size_t size,
+				   bool require_dev_reg_copy)
 {
 	ssize_t ret __attribute__((unused));
 	struct iovec iov;
+	bool dev_reg_copy = require_dev_reg_copy ||
+		(md->handle_valid && size <= cxip_env.safe_devmem_copy_threshold);
 
-	/* Favor CPU store access instead of relying on HMEM copy functions. */
-	if (md->handle_valid && size <= cxip_env.safe_devmem_copy_threshold) {
+	/* Favor dev reg access instead of relying on HMEM copy functions. */
+	if (dev_reg_copy) {
 		ret = ofi_hmem_dev_reg_copy_to_hmem(md->info.iface, md->handle,
 						    dest, src, size);
 		assert(ret == FI_SUCCESS);
@@ -1985,13 +1988,16 @@ static inline void cxip_copy_to_md(struct cxip_md *md, void *dest,
 }
 
 static inline void cxip_copy_from_md(struct cxip_md *md, void *dest,
-				     const void *src, size_t size)
+				     const void *src, size_t size,
+				     bool require_dev_reg_copy)
 {
 	ssize_t ret __attribute__((unused));
 	struct iovec iov;
+	bool dev_reg_copy = require_dev_reg_copy ||
+		(md->handle_valid && size <= cxip_env.safe_devmem_copy_threshold);
 
-	/* Favor CPU store access instead of relying on HMEM copy functions. */
-	if (md->handle_valid && size <= cxip_env.safe_devmem_copy_threshold) {
+	/* Favor dev reg access instead of relying on HMEM copy functions. */
+	if (dev_reg_copy) {
 		ret = ofi_hmem_dev_reg_copy_from_hmem(md->info.iface,
 						      md->handle,
 						      dest, src, size);
@@ -2438,6 +2444,9 @@ struct cxip_ep_obj {
 	struct fi_tx_attr tx_attr;
 	struct fi_rx_attr rx_attr;
 
+	/* Require memcpy's via the dev reg APIs. */
+	bool require_dev_reg_copy[OFI_HMEM_MAX];
+
 	/* Collectives support */
 	struct cxip_ep_coll_obj coll;
 	struct cxip_ep_zbcoll_obj zbcoll;
@@ -2448,6 +2457,25 @@ struct cxip_ep_obj {
 	struct cxip_portals_table *ptable;
 };
 
+int cxip_ep_obj_map(struct cxip_ep_obj *ep, const void *buf, unsigned long len,
+		    uint64_t flags, struct cxip_md **md);
+
+static inline void
+cxip_ep_obj_copy_to_md(struct cxip_ep_obj *ep, struct cxip_md *md, void *dest,
+		       const void *src, size_t size)
+{
+	cxip_copy_to_md(md, dest, src, size,
+			ep->require_dev_reg_copy[md->info.iface]);
+}
+
+static inline void
+cxip_ep_obj_copy_from_md(struct cxip_ep_obj *ep, struct cxip_md *md, void *dest,
+			 const void *src, size_t size)
+{
+	cxip_copy_from_md(md, dest, src, size,
+			  ep->require_dev_reg_copy[md->info.iface]);
+}
+
 static inline void cxip_txc_otx_reqs_inc(struct cxip_txc *txc)
 {
 	assert(ofi_genlock_held(&txc->ep_obj->lock) == 1);
@@ -3641,17 +3669,19 @@ cxip_txc_copy_from_hmem(struct cxip_txc *txc, struct cxip_md *hmem_md,
 	 */
 	if (!cxip_env.fork_safe_requested) {
 		if (!hmem_md) {
-			ret = cxip_map(domain, hmem_src, size, 0, &hmem_md);
+			ret = cxip_ep_obj_map(txc->ep_obj, hmem_src, size, 0,
+					      &hmem_md);
 			if (ret) {
-				TXC_WARN(txc, "cxip_map failed: %d:%s\n", ret,
-					 fi_strerror(-ret));
+				TXC_WARN(txc, "cxip_ep_obj_map failed: %d:%s\n",
+					 ret, fi_strerror(-ret));
 				return ret;
 			}
 
 			unmap_hmem_md = true;
 		}
 
-		cxip_copy_from_md(hmem_md, dest, hmem_src, size);
+		cxip_ep_obj_copy_from_md(txc->ep_obj, hmem_md, dest, hmem_src,
+					 size);
 		if (unmap_hmem_md)
 			cxip_unmap(hmem_md);
 

diff --git a/prov/cxi/src/cxip_atomic.c b/prov/cxi/src/cxip_atomic.c
@@ -612,8 +612,9 @@ static int cxip_amo_emit_idc(struct cxip_txc *txc,
 			if (result_mr) {
 				result_md = result_mr->md;
 			} else {
-				ret = cxip_map(dom, result, atomic_type_len, 0,
-					       &req->amo.result_md);
+				ret = cxip_ep_obj_map(txc->ep_obj, result,
+						      atomic_type_len, 0,
+						      &req->amo.result_md);
 				if (ret) {
 					TXC_WARN_RET(txc, ret,
 						     "Failed to map result buffer\n");
@@ -930,8 +931,9 @@ static int cxip_amo_emit_dma(struct cxip_txc *txc,
 		/* Optionally register result MR. */
 		if (result) {
 			if (!result_mr) {
-				ret = cxip_map(dom, result, atomic_type_len, 0,
-					       &req->amo.result_md);
+				ret = cxip_ep_obj_map(txc->ep_obj, result,
+						      atomic_type_len, 0,
+						      &req->amo.result_md);
 				if (ret) {
 					TXC_WARN(txc,
 						 "Failed to map result buffer: %d:%s\n",
@@ -1017,8 +1019,9 @@ static int cxip_amo_emit_dma(struct cxip_txc *txc,
 			buf_md = buf_mr->md;
 		} else {
 			/* Map user operand buffer for DMA command. */
-			ret = cxip_map(dom, buf, atomic_type_len, 0,
-				       &req->amo.oper1_md);
+			ret = cxip_ep_obj_map(txc->ep_obj, buf,
+					      atomic_type_len, 0,
+					      &req->amo.oper1_md);
 			if (ret) {
 				TXC_WARN(txc,
 					 "Failed to map operand buffer: %d:%s\n",

diff --git a/prov/cxi/src/cxip_coll.c b/prov/cxi/src/cxip_coll.c
@@ -1246,8 +1246,8 @@ static int _coll_add_buffers(struct cxip_coll_pte *coll_pte, size_t size,
 			ret = -FI_ENOMEM;
 			goto out;
 		}
-		ret = cxip_map(coll_pte->ep_obj->domain, (void *)buf->buffer,
-			       size, 0, &buf->cxi_md);
+		ret = cxip_ep_obj_map(coll_pte->ep_obj, (void *)buf->buffer,
+				      size, 0, &buf->cxi_md);
 		if (ret)
 			goto del_msg;
 		buf->bufsiz = size;

diff --git a/prov/cxi/src/cxip_ep.c b/prov/cxi/src/cxip_ep.c
@@ -1118,6 +1118,15 @@ int cxip_ep_getopt_priv(struct cxip_ep *ep, int level, int optname,
 		*optlen = sizeof(size_t);
 		break;
 
+	case FI_OPT_CUDA_API_PERMITTED:
+		if (!optval || !optlen)
+			return -FI_EINVAL;
+		if (*optlen < sizeof(bool))
+			return -FI_ETOOSMALL;
+
+		*(bool *)optval =
+			!ep->ep_obj->require_dev_reg_copy[FI_HMEM_CUDA];
+		break;
 	default:
 		return -FI_ENOPROTOOPT;
 	}
@@ -1140,6 +1149,7 @@ int cxip_ep_setopt_priv(struct cxip_ep *ep, int level, int optname,
 			const void *optval, size_t optlen)
 {
 	size_t min_multi_recv;
+	bool cuda_api_permitted;
 
 	if (level != FI_OPT_ENDPOINT)
 		return -FI_ENOPROTOOPT;
@@ -1158,6 +1168,28 @@ int cxip_ep_setopt_priv(struct cxip_ep *ep, int level, int optname,
 		}
 		ep->ep_obj->rxc->min_multi_recv = min_multi_recv;
 		break;
+	/*
+	 * If GDRCopy is required by the application (ie. it has set
+	 * FI_OPT_CUDA_API_PERMITTED), and is not available, return not
+	 * supported.
+	 */
+	case FI_OPT_CUDA_API_PERMITTED:
+		if (optlen != sizeof(bool))
+			return -FI_EINVAL;
+
+		if (!hmem_ops[FI_HMEM_CUDA].initialized) {
+			CXIP_WARN("FI_OPT_CUDA_API_PERMITTED cannot be set when CUDA library or CUDA device is not available\n");
+			return -FI_EOPNOTSUPP;
+		}
+
+		cuda_api_permitted = *(bool *)optval;
+
+		if (!cuda_api_permitted && !cuda_is_gdrcopy_enabled())
+			return -FI_EOPNOTSUPP;
+
+		ep->ep_obj->require_dev_reg_copy[FI_HMEM_CUDA] =
+			!cuda_api_permitted;
+		break;
 
 	default:
 		return -FI_ENOPROTOOPT;
@@ -1260,6 +1292,12 @@ int cxip_alloc_endpoint(struct cxip_domain *cxip_dom, struct fi_info *hints,
 	ep_obj->src_addr.pid = pid;
 	ep_obj->fi_addr = FI_ADDR_NOTAVAIL;
 
+	/* Default to allowing non-dev reg copy APIs unless the caller
+	 * disables it.
+	 */
+	for (i = 0; i < OFI_HMEM_MAX; i++)
+		ep_obj->require_dev_reg_copy[i] = false;
+
 	ofi_atomic_initialize32(&ep_obj->txq_ref, 0);
 	ofi_atomic_initialize32(&ep_obj->tgq_ref, 0);
 
@@ -1332,6 +1370,26 @@ int cxip_alloc_endpoint(struct cxip_domain *cxip_dom, struct fi_info *hints,
 	return ret;
 }
 
+int cxip_ep_obj_map(struct cxip_ep_obj *ep, const void *buf, unsigned long len,
+		    uint64_t flags, struct cxip_md **md)
+{
+	struct cxip_domain *dom = ep->domain;
+	int ret;
+
+	ret = cxip_map(dom, buf, len, flags, md);
+	if (ret != FI_SUCCESS)
+		return ret;
+
+	if (ep->require_dev_reg_copy[(*md)->info.iface] &&
+	    !((*md)->handle_valid)) {
+		CXIP_WARN("Required dev registration copy failed\n");
+		cxip_unmap(*md);
+		return -FI_EOPNOTSUPP;
+	}
+
+	return FI_SUCCESS;
+}
+
 /*
  * cxip_endpoint() - Provider fi_endpoint() implementation.
  */

diff --git a/prov/cxi/src/cxip_mr.c b/prov/cxi/src/cxip_mr.c
@@ -1283,6 +1283,15 @@ static int cxip_mr_bind(struct fid *fid, struct fid *bfid, uint64_t flags)
 			break;
 		}
 
+		/* Zero length MRs do not have MD. */
+		if (mr->md &&
+		    ep->ep_obj->require_dev_reg_copy[mr->md->info.iface] &&
+		    !mr->md->handle_valid) {
+			CXIP_WARN("Cannot bind to endpoint without required dev reg support\n");
+			ret = -FI_EOPNOTSUPP;
+			break;
+		}
+
 		mr->ep = ep;
 		ofi_atomic_inc32(&ep->ep_obj->ref);
 		break;
@@ -1439,6 +1448,10 @@ static int cxip_regattr(struct fid *fid, const struct fi_mr_attr *attr,
 		_mr->mr_fid.key = _mr->key;
 
 	if (_mr->len) {
+		/* Do not check whether cuda_api_permitted is set at this point,
+		 * because the mr is not bound to an endpoint.  Check instead in
+		 * cxip_mr_bind().
+		 */
 		ret = cxip_map(_mr->domain, (void *)_mr->buf, _mr->len, 0,
 			       &_mr->md);
 		if (ret) {

diff --git a/prov/cxi/src/cxip_msg.c b/prov/cxi/src/cxip_msg.c
@@ -60,7 +60,6 @@ int cxip_recv_req_alloc(struct cxip_rxc *rxc, void *buf, size_t len,
 			int (*recv_cb)(struct cxip_req *req,
 				       const union c_event *event))
 {
-	struct cxip_domain *dom = rxc->domain;
 	struct cxip_req *req;
 	struct cxip_md *recv_md = NULL;
 	int ret;
@@ -79,7 +78,8 @@ int cxip_recv_req_alloc(struct cxip_rxc *rxc, void *buf, size_t len,
 	if (len) {
 		/* If hybrid descriptor not passed, map for dma */
 		if (!md) {
-			ret = cxip_map(dom, (void *)buf, len, 0, &recv_md);
+			ret = cxip_ep_obj_map(rxc->ep_obj, (void *)buf, len, 0,
+					      &recv_md);
 			if (ret) {
 				RXC_WARN(rxc,
 					 "Map of recv buffer failed: %d, %s\n",
@@ -718,8 +718,8 @@ int cxip_send_buf_init(struct cxip_req *req)
 
 	/* Triggered operation always requires memory registration. */
 	if (req->triggered)
-		return cxip_map(txc->domain, req->send.buf, req->send.len, 0,
-			       &req->send.send_md);
+		return cxip_ep_obj_map(txc->ep_obj, req->send.buf,
+				       req->send.len, 0, &req->send.send_md);
 
 	/* FI_INJECT operations always require an internal bounce buffer. This
 	 * is needed to replay FI_INJECT operations which may experience flow
@@ -777,8 +777,8 @@ int cxip_send_buf_init(struct cxip_req *req)
 	}
 
 	/* Everything else requires memory registeration. */
-	return cxip_map(txc->domain, req->send.buf, req->send.len, 0,
-			&req->send.send_md);
+	return cxip_ep_obj_map(txc->ep_obj, req->send.buf, req->send.len, 0,
+			       &req->send.send_md);
 
 err_buf_fini:
 	cxip_send_buf_fini(req);

diff --git a/prov/cxi/src/cxip_msg_hpc.c b/prov/cxi/src/cxip_msg_hpc.c
@@ -629,8 +629,9 @@ static int cxip_ux_send(struct cxip_req *match_req, struct cxip_req *oflow_req,
 
 	/* Copy data out of overflow buffer. */
 	oflow_bytes = MIN(put_event->tgt_long.mlength, match_req->data_len);
-	cxip_copy_to_md(match_req->recv.recv_md, match_req->recv.recv_buf,
-			oflow_va, oflow_bytes);
+	cxip_ep_obj_copy_to_md(match_req->recv.rxc->ep_obj,
+			       match_req->recv.recv_md,
+			       match_req->recv.recv_buf, oflow_va, oflow_bytes);
 
 	if (oflow_req->type == CXIP_REQ_OFLOW)
 		oflow_req_put_bytes(oflow_req, put_event->tgt_long.mlength);

diff --git a/prov/cxi/src/cxip_msg_rnr.c b/prov/cxi/src/cxip_msg_rnr.c
@@ -1174,9 +1174,9 @@ cxip_send_common(struct cxip_txc *txc, uint32_t tclass, const void *buf,
 
 	if (send_req->send.len && !idc) {
 		if (!mr) {
-			ret = cxip_map(txc->domain, send_req->send.buf,
-				       send_req->send.len, 0,
-				       &send_req->send.send_md);
+			ret = cxip_ep_obj_map(txc->ep_obj, send_req->send.buf,
+					      send_req->send.len, 0,
+					      &send_req->send.send_md);
 			if (ret) {
 				TXC_WARN(txc,
 					 "Local buffer map failed: %d %s\n",

diff --git a/prov/cxi/src/cxip_ptelist_buf.c b/prov/cxi/src/cxip_ptelist_buf.c
@@ -132,8 +132,8 @@ cxip_ptelist_buf_alloc(struct cxip_ptelist_bufpool *pool)
 		}
 	}
 
-	ret = cxip_map(rxc->base.domain, buf->data, pool->attr.buf_size,
-		       OFI_MR_NOCACHE, &buf->md);
+	ret = cxip_ep_obj_map(rxc->base.ep_obj, buf->data, pool->attr.buf_size,
+			      OFI_MR_NOCACHE, &buf->md);
 	if (ret)
 		goto err_unreg_buf;
 

diff --git a/prov/cxi/src/cxip_rma.c b/prov/cxi/src/cxip_rma.c
@@ -269,7 +269,8 @@ static int cxip_rma_emit_dma(struct cxip_txc *txc, const void *buf, size_t len,
 		} else {
 			assert(req != NULL);
 
-			ret = cxip_map(dom, buf, len, 0, &req->rma.local_md);
+			ret = cxip_ep_obj_map(txc->ep_obj, buf, len, 0,
+					      &req->rma.local_md);
 			if (ret) {
 				TXC_WARN(txc, "Failed to map buffer: %d:%s\n",
 					ret, fi_strerror(-ret));

diff --git a/prov/cxi/src/cxip_txc.c b/prov/cxi/src/cxip_txc.c
@@ -63,8 +63,8 @@ int cxip_ibuf_chunk_init(struct ofi_bufpool_region *region)
 	struct cxip_md *md;
 	int ret;
 
-	ret = cxip_map(txc->domain, region->mem_region,
-		       region->pool->region_size, OFI_MR_NOCACHE, &md);
+	ret = cxip_ep_obj_map(txc->ep_obj, region->mem_region,
+			      region->pool->region_size, OFI_MR_NOCACHE, &md);
 	if (ret != FI_SUCCESS) {
 		CXIP_WARN("Failed to map inject buffer chunk\n");
 		return ret;