nvmf: accepts new parameter to set poll groups mask

Parses and verifies hexadecimal cpu bit mask specified by the user.
Added verification to check for cpu cores range, making sure poll groups cores
assigned within the range of cpu cores allocated for the application.

RPC nvmf_set_config now takes an argument to configure ‘poll groups’,
a new parameter for NVMf subsystem. This parameter sets a CPU mask
to spawn threads which run an event loop for a ‘poll group’.

Change-Id: Ied9081c2213715ec94de00a8b37153730b8ac2ed
Signed-off-by: Yuri <yuriy.kirichok@hpe.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/5611
Community-CI: Mellanox Build Bot
Reviewed-by: Matt Dumm <matt.dumm@hpe.com>
Reviewed-by: Aleksey Marchuk <alexeymar@mellanox.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>

CHANGELOG.md

@@ -116,6 +116,10 @@ is associated with a fabrics controller.
 Added `min_cntlid` and `max_cntlid` to `nvmf_create_subsystem` to limit the controller ID range.
 Added `spdk_nvmf_subsystem_get_min_cntlid` and `spdk_nvmf_subsystem_get_max_cntlid` to request those values.
 
+A new parameter, `poll_groups_mask` was added to the nvmf_set_config RPC that allows specifying
+a subset of cores for the nvmf poll groups. This helps to avoid imbalances when some cores are
+busy with periodic timer tasks that run very frequently.
+
 `spdk_nvmf_request_get_buffers_multi` API is removed.
 
 Added the `nvmf_set_crdt` RPC for setting command retry delay times.

doc/jsonrpc.md

@@ -6783,6 +6783,7 @@ Name                    | Optional | Type        | Description
 ----------------------- | -------- | ----------- | -----------
 acceptor_poll_rate      | Optional | number      | Polling interval of the acceptor for incoming connections (microseconds)
 admin_cmd_passthru      | Optional | object      | Admin command passthru configuration
+poll_groups_mask        | Optional | string      | Set cpumask for NVMf poll groups
 
 ### admin_cmd_passthru {#spdk_nvmf_admin_passthru_conf}
 

module/event/subsystems/nvmf/nvmf_rpc.c

@@ -35,6 +35,7 @@
 
 #include "spdk/rpc.h"
 #include "spdk/util.h"
+#include "spdk/cpuset.h"
 
 static const struct spdk_json_object_decoder nvmf_rpc_subsystem_tgt_opts_decoder[] = {
 	{"max_subsystems", 0, spdk_json_decode_uint32, true}
@@ -98,10 +99,57 @@ static int decode_admin_passthru(const struct spdk_json_val *val, void *out)
 	return 0;
 }
 
+static int
+nvmf_is_subset_of_env_core_mask(const struct spdk_cpuset *set)
+{
+	uint32_t i, tmp_counter = 0;
+
+	SPDK_ENV_FOREACH_CORE(i) {
+		if (spdk_cpuset_get_cpu(set, i)) {
+			++tmp_counter;
+		}
+	}
+	return spdk_cpuset_count(set) - tmp_counter;
+}
+
+static int
+nvmf_decode_poll_groups_mask(const struct spdk_json_val *val, void *out)
+{
+	char *mask = spdk_json_strdup(val);
+	int ret = -1;
+
+	if (mask == NULL) {
+		return -1;
+	}
+
+	if (!(g_poll_groups_mask = spdk_cpuset_alloc())) {
+		SPDK_ERRLOG("Unable to allocate a poll groups mask object in nvmf_decode_poll_groups_mask.\n");
+		free(mask);
+		return -1;
+	}
+
+	ret = spdk_cpuset_parse(g_poll_groups_mask, mask);
+	free(mask);
+	if (ret == 0) {
+		if (nvmf_is_subset_of_env_core_mask(g_poll_groups_mask) == 0) {
+			return 0;
+		} else {
+			SPDK_ERRLOG("Poll groups cpumask 0x%s is out of range\n", spdk_cpuset_fmt(g_poll_groups_mask));
+		}
+	} else {
+		SPDK_ERRLOG("Invalid cpumask\n");
+	}
+
+	spdk_cpuset_free(g_poll_groups_mask);
+	g_poll_groups_mask = NULL;
+	return -1;
+}
+
 static const struct spdk_json_object_decoder nvmf_rpc_subsystem_tgt_conf_decoder[] = {
 	{"acceptor_poll_rate", offsetof(struct spdk_nvmf_tgt_conf, acceptor_poll_rate), spdk_json_decode_uint32, true},
 	{"conn_sched", offsetof(struct spdk_nvmf_tgt_conf, conn_sched), decode_conn_sched, true},
-	{"admin_cmd_passthru", offsetof(struct spdk_nvmf_tgt_conf, admin_passthru), decode_admin_passthru, true}
+	{"admin_cmd_passthru", offsetof(struct spdk_nvmf_tgt_conf, admin_passthru), decode_admin_passthru, true},
+	{"poll_groups_mask", 0, nvmf_decode_poll_groups_mask, true}
 };
 
 static void

module/event/subsystems/nvmf/nvmf_tgt.c

@@ -488,6 +488,9 @@ nvmf_subsystem_write_config_json(struct spdk_json_write_ctx *w)
 	spdk_json_write_named_bool(w, "identify_ctrlr",
 				   g_spdk_nvmf_tgt_conf.admin_passthru.identify_ctrlr);
 	spdk_json_write_object_end(w);
+	if (g_poll_groups_mask) {
+		spdk_json_write_named_string(w, "poll_groups_mask", spdk_cpuset_fmt(g_poll_groups_mask));
+	}
 	spdk_json_write_object_end(w);
 	spdk_json_write_object_end(w);
 

scripts/rpc.py

@@ -1895,7 +1895,8 @@ Format: 'user:u1 secret:s1 muser:mu1 msecret:ms1,user:u2 secret:s2 muser:mu2 mse
         rpc.nvmf.nvmf_set_config(args.client,
                                  acceptor_poll_rate=args.acceptor_poll_rate,
                                  conn_sched=args.conn_sched,
-                                 passthru_identify_ctrlr=args.passthru_identify_ctrlr)
+                                 passthru_identify_ctrlr=args.passthru_identify_ctrlr,
+                                 poll_groups_mask=args.poll_groups_mask)
 
     p = subparsers.add_parser('nvmf_set_config', aliases=['set_nvmf_target_config'],
                               help='Set NVMf target config')
@@ -1903,6 +1904,7 @@ Format: 'user:u1 secret:s1 muser:mu1 msecret:ms1,user:u2 secret:s2 muser:mu2 mse
     p.add_argument('-s', '--conn-sched', help='(Deprecated). Ignored.')
     p.add_argument('-i', '--passthru-identify-ctrlr', help="""Passthrough fields like serial number and model number
     when the controller has a single namespace that is an NVMe bdev""", action='store_true')
+    p.add_argument('-m', '--poll-groups-mask', help='Set cpumask for NVMf poll groups (optional)', type=str)
     p.set_defaults(func=nvmf_set_config)
 
     def nvmf_create_transport(args):

scripts/rpc/nvmf.py

@@ -23,7 +23,8 @@ def nvmf_set_max_subsystems(client,
 def nvmf_set_config(client,
                     acceptor_poll_rate=None,
                     conn_sched=None,
-                    passthru_identify_ctrlr=None):
+                    passthru_identify_ctrlr=None,
+                    poll_groups_mask=None):
     """Set NVMe-oF target subsystem configuration.
 
     Args:
@@ -43,6 +44,8 @@ def nvmf_set_config(client,
         admin_cmd_passthru = {}
         admin_cmd_passthru['identify_ctrlr'] = passthru_identify_ctrlr
         params['admin_cmd_passthru'] = admin_cmd_passthru
+    if poll_groups_mask:
+        params['poll_groups_mask'] = poll_groups_mask
 
     return client.call('nvmf_set_config', params)