prov/shm: Add unmap_region function

This function is mainly for the niche case where on progress_connreq
a peer is added to the map with its region needing to be mapped, and
then after mapping it, it's discovered that the newly mapped peer's
process died. In this case we need to unmap them and free any resources
that were opened for communicating with them.

Remove lock from map_to_region and unmap_region functions and require
lock acquirement before calling those functions. This is necessary because
on av removal path, map will be double locked if the functions also process
locking the map. The map_to_region function is updated to mirror this
policy.

Signed-off-by: Zach Dworkin <[email protected]>

Author: zachdworkin

prov/shm/src/smr_av.c

@@ -69,9 +69,12 @@ static void smr_map_cleanup(struct smr_map *map)
 {
 	int64_t i;
 
-	for (i = 0; i < SMR_MAX_PEERS; i++)
-		smr_map_del(map, i);
+	for (i = 0; i < SMR_MAX_PEERS; i++) {
+		if (map->peers[i].peer.id < 0)
+			continue;
 
+		smr_map_del(map, i);
+	}
 	ofi_rbmap_cleanup(&map->rbmap);
 }
 
@@ -210,7 +213,6 @@ static int smr_av_remove(struct fid_av *av_fid, fi_addr_t *fi_addr, size_t count
 		dlist_foreach(&util_av->ep_list, av_entry) {
 			util_ep = container_of(av_entry, struct util_ep, av_entry);
 			smr_ep = container_of(util_ep, struct smr_ep, util_ep);
-			smr_unmap_from_endpoint(smr_ep->region, id);
 			if (smr_av->smr_map.num_peers > 0)
 				smr_ep->region->max_sar_buf_per_peer =
 					SMR_MAX_PEERS /

prov/shm/src/smr_ep.c

@@ -223,7 +223,9 @@ int64_t smr_verify_peer(struct smr_ep *ep, fi_addr_t fi_addr)
 		return id;
 
 	if (!ep->region->map->peers[id].region) {
+		ofi_spin_lock(&ep->region->map->lock);
 		ret = smr_map_to_region(&smr_prov, ep->region->map, id);
+		ofi_spin_unlock(&ep->region->map->lock);
 		if (ret)
 			return -1;
 	}

prov/shm/src/smr_progress.c

@@ -878,7 +878,9 @@ static void smr_progress_connreq(struct smr_ep *ep, struct smr_cmd *cmd)
 
 	peer_smr = smr_peer_region(ep->region, idx);
 	if (!peer_smr) {
+		ofi_spin_lock(&ep->region->map->lock);
 		ret = smr_map_to_region(&smr_prov, ep->region->map, idx);
+		ofi_spin_unlock(&ep->region->map->lock);
 		if (ret) {
 			FI_WARN(&smr_prov, FI_LOG_EP_CTRL,
 				"Could not map peer region\n");
@@ -891,14 +893,11 @@ static void smr_progress_connreq(struct smr_ep *ep, struct smr_cmd *cmd)
 	if (peer_smr->pid != (int) cmd->msg.hdr.data) {
 		/* TODO track and update/complete in error any transfers
 		 * to or from old mapping
-		 *
-		 * TODO create smr_unmap_region
-		 * this needs to close peer_smr->map->peers[idx].pid_fd
-		 * This case will also return an unmapped region because the idx
-		 * is valid but the region was unmapped
 		 */
-		munmap(peer_smr, peer_smr->total_size);
+		ofi_spin_lock(&ep->region->map->lock);
+		smr_unmap_region(&smr_prov, ep->region->map, idx, false);
 		smr_map_to_region(&smr_prov, ep->region->map, idx);
+		ofi_spin_unlock(&ep->region->map->lock);
 		peer_smr = smr_peer_region(ep->region, idx);
 	}
 

prov/shm/src/smr_util.c

@@ -367,16 +367,15 @@ int smr_map_to_region(const struct fi_provider *prov, struct smr_map *map,
 	}
 	pthread_mutex_unlock(&ep_list_lock);
 
-	ofi_spin_lock(&map->lock);
 	if (peer_buf->region)
-		goto unlock;
+		return FI_SUCCESS;
 
+	assert(ofi_spin_held(&map->lock));
 	fd = shm_open(name, O_RDWR, S_IRUSR | S_IWUSR);
 	if (fd < 0) {
-		ret = -errno;
 		FI_WARN_ONCE(prov, FI_LOG_AV,
 			     "shm_open error: name %s errno %d\n", name, errno);
-		goto unlock;
+		return -errno;
 	}
 
 	memset(tmp, 0, sizeof(tmp));
@@ -437,8 +436,6 @@ int smr_map_to_region(const struct fi_provider *prov, struct smr_map *map,
 
 out:
 	close(fd);
-unlock:
-	ofi_spin_unlock(&map->lock);
 	return ret;
 }
 
@@ -448,6 +445,7 @@ void smr_map_to_endpoint(struct smr_region *region, int64_t id)
 	struct smr_region *peer_smr;
 	struct smr_peer_data *local_peers;
 
+	assert(ofi_spin_held(&region->map->lock));
 	peer_smr = smr_peer_region(region, id);
 	if (region->map->peers[id].peer.id < 0 || !peer_smr)
 	    return;
@@ -479,32 +477,81 @@ void smr_map_to_endpoint(struct smr_region *region, int64_t id)
 	return;
 }
 
+void smr_unmap_region(const struct fi_provider *prov, struct smr_map *map,
+		     int64_t peer_id, bool local)
+{
+	struct smr_region *peer_region;
+	struct smr_peer *peer;
+	struct util_ep *util_ep;
+	struct smr_ep *smr_ep;
+	struct smr_av *av;
+	int ret = 0;
+
+	assert(ofi_spin_held(&map->lock));
+	peer_region = map->peers[peer_id].region;
+	if (!peer_region)
+		return;
+
+	peer = &map->peers[peer_id];
+	av = container_of(map, struct smr_av, smr_map);
+	dlist_foreach_container(&av->util_av.ep_list, struct util_ep, util_ep,
+				av_entry) {
+		smr_ep = container_of(util_ep, struct smr_ep, util_ep);
+		smr_unmap_from_endpoint(smr_ep->region, peer_id);
+	}
+
+	/* Don't unmap memory owned by this pid because the endpoint it belongs
+	 * to might still be active.
+	 */
+	if (local)
+		return;
+
+	if (map->flags & SMR_FLAG_HMEM_ENABLED) {
+		ret = ofi_hmem_host_unregister(peer_region);
+		if (ret)
+			FI_WARN(prov, FI_LOG_EP_CTRL,
+				"unable to unregister shm with iface\n");
+
+		if (peer->pid_fd != -1) {
+			close(peer->pid_fd);
+			peer->pid_fd = -1;
+		}
+	}
+
+	munmap(peer_region, peer_region->total_size);
+	peer->region = NULL;
+}
+
 void smr_unmap_from_endpoint(struct smr_region *region, int64_t id)
 {
 	struct smr_region *peer_smr;
 	struct smr_peer_data *local_peers, *peer_peers;
 	int64_t peer_id;
 
-	local_peers = smr_peer_data(region);
 	if (region->map->peers[id].peer.id < 0)
 		return;
 
 	peer_smr = smr_peer_region(region, id);
-	peer_id = smr_peer_data(region)[id].addr.id;
-
+	assert(peer_smr);
 	peer_peers = smr_peer_data(peer_smr);
+	peer_id = smr_peer_data(region)[id].addr.id;
 
 	peer_peers[peer_id].addr.id = -1;
 	peer_peers[peer_id].name_sent = 0;
 
+	local_peers = smr_peer_data(region);
 	ofi_xpmem_release(&local_peers[peer_id].xpmem);
 }
 
 void smr_exchange_all_peers(struct smr_region *region)
 {
 	int64_t i;
+
+	ofi_spin_lock(&region->map->lock);
 	for (i = 0; i < SMR_MAX_PEERS; i++)
 		smr_map_to_endpoint(region, i);
+
+	ofi_spin_unlock(&region->map->lock);
 }
 
 int smr_map_add(const struct fi_provider *prov, struct smr_map *map,
@@ -546,37 +593,24 @@ int smr_map_add(const struct fi_provider *prov, struct smr_map *map,
 
 void smr_map_del(struct smr_map *map, int64_t id)
 {
-	struct dlist_entry *entry;
+	struct smr_ep_name *name;
+	bool local = false;
 
 	assert(id >= 0 && id < SMR_MAX_PEERS);
-
 	pthread_mutex_lock(&ep_list_lock);
-	entry = dlist_find_first_match(&ep_name_list, smr_match_name,
-				       smr_no_prefix(map->peers[id].peer.name));
+	dlist_foreach_container(&ep_name_list, struct smr_ep_name, name, entry) {
+		if (strcmp(name->name, map->peers[id].peer.name)) {
+			local = true;
+			break;
+		}
+	}
 	pthread_mutex_unlock(&ep_list_lock);
-
 	ofi_spin_lock(&map->lock);
-	(void) ofi_rbmap_find_delete(&map->rbmap,
-				     (void *) map->peers[id].peer.name);
-
+	smr_unmap_region(&smr_prov, map, id, local);
 	map->peers[id].fiaddr = FI_ADDR_NOTAVAIL;
 	map->peers[id].peer.id = -1;
 	map->num_peers--;
-
-	if (!map->peers[id].region)
-		goto unlock;
-
-	if (!entry) {
-		if (map->flags & SMR_FLAG_HMEM_ENABLED) {
-			if (map->peers[id].pid_fd != -1)
-				close(map->peers[id].pid_fd);
-
-			(void) ofi_hmem_host_unregister(map->peers[id].region);
-		}
-		munmap(map->peers[id].region, map->peers[id].region->total_size);
-		map->peers[id].region = NULL;
-	}
-unlock:
+	ofi_rbmap_find_delete(&map->rbmap, map->peers[id].peer.name);
 	ofi_spin_unlock(&map->lock);
 }
 

prov/shm/src/smr_util.h

@@ -356,6 +356,8 @@ void	smr_cleanup(void);
 int	smr_map_to_region(const struct fi_provider *prov, struct smr_map *map,
 			  int64_t id);
 void	smr_map_to_endpoint(struct smr_region *region, int64_t id);
+void	smr_unmap_region(const struct fi_provider *prov, struct smr_map *map,
+			  int64_t id, bool found);
 void	smr_unmap_from_endpoint(struct smr_region *region, int64_t id);
 void	smr_exchange_all_peers(struct smr_region *region);
 int	smr_map_add(const struct fi_provider *prov, struct smr_map *map,