DAOS-14262 cart: add ability to select traffic class for SWIM context (…

…#14893) (#14917)

Add SWIM_TRAFFIC_CLASS env var (default is unspec)

Signed-off-by: Jerome Soumagne <jerome.soumagne@intel.com>

src/cart/README.env

@@ -1,13 +1,10 @@
 This file lists the environment variables used in CaRT.
 
  . D_PROVIDER (Deprecated: CRT_PHY_ADDR_STR)
-   It determines which mercury NA plugin to be used:
+   It determines which mercury NA plugin and transport to be used:
    - set it as "ofi+verbs;ofi_rxm" to use OFI verbs;ofi_rxm provider
-   - set it as "ofi+gni" to use OFI gni provider
    - set it as "sm" to use SM plugin which only works within single node
-   - set it as "ofi+tcp;ofi_rxm" to use OFI tcp;ofi_rxm provider.
-   - set it as "ofi+sockets" to use OFI sockets provider
-      NOTE: This provider is deprecated in favor of "ofi+tcp;ofi_rxm"
+   - set it as "ofi+tcp" to use OFI tcp provider.
    - by default (not set or set as any other value) it will use ofi tcp
       provider.
 
@@ -205,3 +202,8 @@ This file lists the environment variables used in CaRT.
    start copying data in an effort to release multi-recv buffers. Copy will occur when at
    most D_MRECV_BUF_COPY buffers remain.
 
+   SWIM_TRAFFIC_CLASS
+   (server only) Select a traffic class for the SWIM protocol to use and prevent potential
+   traffic congestion. Available options are: "unspec" (default), "best_effort",
+   "low_latency", "bulk_data".
+

src/cart/crt_hg.c

@@ -863,6 +863,9 @@ crt_hg_class_init(crt_provider_t provider, int ctx_idx, bool primary, int iface_
 	init_info.request_post_incr         = crt_gdata.cg_post_incr;
 	init_info.multi_recv_op_max         = crt_gdata.cg_mrecv_buf;
 	init_info.multi_recv_copy_threshold = crt_gdata.cg_mrecv_buf_copy;
+	/* Separate SWIM traffic in an effort to prevent potential congestion. */
+	if (crt_is_service() && ctx_idx == crt_gdata.cg_swim_crt_idx)
+		init_info.traffic_class = (enum na_traffic_class)crt_gdata.cg_swim_tc;
 
 	hg_class = HG_Init_opt2(info_string, crt_is_service(), HG_VERSION(2, 4), &init_info);
 	if (hg_class == NULL) {

src/cart/crt_init.c

@@ -18,6 +18,10 @@ static volatile int	gdata_init_flag;
 struct crt_plugin_gdata crt_plugin_gdata;
 static bool		g_prov_settings_applied[CRT_PROV_COUNT];
 
+#define X(a, b) b,
+static const char *const crt_tc_name[] = {CRT_TRAFFIC_CLASSES};
+#undef X
+
 static void
 crt_lib_init(void) __attribute__((__constructor__));
 
@@ -237,18 +241,30 @@ crt_gdata_dump(void)
 	DUMP_GDATA_FIELD("%d", cg_rpc_quota);
 }
 
+static enum crt_traffic_class
+crt_str_to_tc(const char *str)
+{
+	enum crt_traffic_class i = 0;
+
+	while (str != NULL && strcmp(crt_tc_name[i], str) != 0 && i < CRT_TC_UNKNOWN)
+		i++;
+
+	return i == CRT_TC_UNKNOWN ? CRT_TC_UNSPEC : i;
+}
+
 /* first step init - for initializing crt_gdata */
 static int data_init(int server, crt_init_options_t *opt)
 {
-	uint32_t        timeout = 0;
-	uint32_t	credits;
-	uint32_t	fi_univ_size = 0;
-	uint32_t	mem_pin_enable = 0;
-	uint32_t        is_secondary;
-	uint32_t        post_init = CRT_HG_POST_INIT, post_incr = CRT_HG_POST_INCR;
-	unsigned int    mrecv_buf      = CRT_HG_MRECV_BUF;
-	unsigned int    mrecv_buf_copy = 0; /* buf copy disabled by default */
-	int             rc             = 0;
+	uint32_t     timeout = 0;
+	uint32_t     credits;
+	uint32_t     fi_univ_size   = 0;
+	uint32_t     mem_pin_enable = 0;
+	uint32_t     is_secondary;
+	uint32_t     post_init = CRT_HG_POST_INIT, post_incr = CRT_HG_POST_INCR;
+	unsigned int mrecv_buf          = CRT_HG_MRECV_BUF;
+	unsigned int mrecv_buf_copy     = 0; /* buf copy disabled by default */
+	char        *swim_traffic_class = NULL;
+	int          rc                 = 0;
 
 	crt_env_dump();
 
@@ -261,6 +277,8 @@ static int data_init(int server, crt_init_options_t *opt)
 	crt_gdata.cg_mrecv_buf = mrecv_buf;
 	crt_env_get(D_MRECV_BUF_COPY, &mrecv_buf_copy);
 	crt_gdata.cg_mrecv_buf_copy = mrecv_buf_copy;
+	crt_env_get(SWIM_TRAFFIC_CLASS, &swim_traffic_class);
+	crt_gdata.cg_swim_tc = crt_str_to_tc(swim_traffic_class);
 
 	is_secondary = 0;
 	/* Apply CART-890 workaround for server side only */

src/cart/crt_internal_types.h

@@ -42,6 +42,17 @@ struct crt_na_config {
 	char            **noc_domain_str; /* Array of domains */
 };
 
+#define CRT_TRAFFIC_CLASSES                                                                        \
+	X(CRT_TC_UNSPEC, "unspec")           /* Leave it upon plugin to choose */                  \
+	X(CRT_TC_BEST_EFFORT, "best_effort") /* Best effort */                                     \
+	X(CRT_TC_LOW_LATENCY, "low_latency") /* Low latency */                                     \
+	X(CRT_TC_BULK_DATA, "bulk_data")     /* Bulk data */                                       \
+	X(CRT_TC_UNKNOWN, "unknown")         /* Unknown */
+
+#define X(a, b) a,
+enum crt_traffic_class { CRT_TRAFFIC_CLASSES };
+#undef X
+
 struct crt_prov_gdata {
 	/** NA plugin type */
 	int			cpg_provider;
@@ -105,6 +116,9 @@ struct crt_gdata {
 	/** global swim index for all servers */
 	int32_t			cg_swim_crt_idx;
 
+	/** traffic class used by SWIM */
+	enum crt_traffic_class   cg_swim_tc;
+
 	/** credits limitation for #in-flight RPCs per target EP CTX */
 	uint32_t		cg_credit_ep_ctx;
 
@@ -220,6 +234,7 @@ struct crt_event_cb_priv {
 	ENV(SWIM_PING_TIMEOUT)                                                                     \
 	ENV(SWIM_PROTOCOL_PERIOD_LEN)                                                              \
 	ENV(SWIM_SUSPECT_TIMEOUT)                                                                  \
+	ENV_STR(SWIM_TRAFFIC_CLASS)                                                                \
 	ENV_STR(UCX_IB_FORK_INIT)
 
 /* uint env */