prov/shm: add unexpected message buffering on demand when needed

In order to support unlimited unexpected messaging, add a flag
SMR_BUFFER_RECV for the sender to let the receiver know that
resources are limited and the whole message should get buffered
on the target. This allows the command to be immediately returned
to the sender so that the sender is never blocked due to unexpected
messages at the target.
Buffering unexpected messages hurts performance so the default is to
wait until only a single command is left before requesting buffering,
but an environment variable is also added to toggle this for either
debugging purposes or workarounds.

Signed-off-by: Alexia Ingerson <alexia.ingerson@intel.com>

Author: aingerson

man/fi_shm.7.md

@@ -176,6 +176,14 @@ The *shm* provider checks for the following environment variables:
     chunks.  This environment variable is provided to fine tune performance
     on different systems.  Default 262144
 
+*FI_SHM_BUFFER_THRESHOLD*
+ :  When to start requesting forced unexpected messaging buffering. When this
+    threshold is reached, the sender will notify the receiver to force buffering
+    of the entire message if it is unexpected. If the message is matched when
+    received, it has no effect. Requesting unexpected message buffering allows
+    shm to support unlimited unexpected messaging (memory permitting).
+    Default: 1
+
 # SEE ALSO
 
 [`fabric`(7)](fabric.7.html),

prov/shm/src/smr_ep.c

@@ -478,6 +478,10 @@ static ssize_t smr_do_inject(struct smr_ep *ep, struct smr_region *peer_smr,
 	smr_generic_format(cmd, tx_id, rx_id, op, tag, data, op_flags);
 	smr_format_inject(ep, cmd, pend);
 
+	if (smr_freestack_avail(smr_cmd_stack(ep->region)) <=
+	    smr_env.buffer_threshold)
+		cmd->hdr.op_flags |= SMR_BUFFER_RECV;
+
 	return FI_SUCCESS;
 }
 
@@ -499,6 +503,10 @@ static ssize_t smr_do_iov(struct smr_ep *ep, struct smr_region *peer_smr,
 	smr_generic_format(cmd, tx_id, rx_id, op, tag, data, op_flags);
 	smr_format_iov(cmd, pend);
 
+	if (smr_freestack_avail(smr_cmd_stack(ep->region)) <=
+	    smr_env.buffer_threshold)
+		cmd->hdr.op_flags |= SMR_BUFFER_RECV;
+
 	return FI_SUCCESS;
 }
 
@@ -566,6 +574,10 @@ static ssize_t smr_do_ipc(struct smr_ep *ep, struct smr_region *peer_smr,
 
 	smr_format_tx_pend(pend, cmd, context, desc, iov, iov_count, op_flags);
 
+	if (smr_freestack_avail(smr_cmd_stack(ep->region)) <=
+	    smr_env.buffer_threshold)
+		cmd->hdr.op_flags |= SMR_BUFFER_RECV;
+
 	return FI_SUCCESS;
 }
 

prov/shm/src/smr_init.c

@@ -42,6 +42,7 @@ struct smr_env smr_env = {
 	.use_dsa_sar = false,
 	.max_gdrcopy_size = 3072,
 	.use_xpmem = false,
+	.buffer_threshold = 1,
 };
 
 static void smr_init_env(void)
@@ -51,6 +52,8 @@ static void smr_init_env(void)
 	fi_param_get_bool(&smr_prov, "disable_cma", &smr_env.disable_cma);
 	fi_param_get_bool(&smr_prov, "use_dsa_sar", &smr_env.use_dsa_sar);
 	fi_param_get_bool(&smr_prov, "use_xpmem", &smr_env.use_xpmem);
+	fi_param_get_size_t(&smr_prov, "buffer_threshold",
+			    &smr_env.buffer_threshold);
 }
 
 static void smr_resolve_addr(const char *node, const char *service,
@@ -216,6 +219,9 @@ SHM_INI
 	fi_param_define(&smr_prov, "use_xpmem", FI_PARAM_BOOL,
 			"Enable XPMEM over CMA when possible "
 			"(default: false)");
+	fi_param_define(&smr_prov, "buffer_threshold", FI_PARAM_SIZE_T,
+			"When to start requesting forced unexpected messaging "
+			"buffering. (default: 1)");
 
 	smr_init_env();