Commit 1f3f1605 authored by Avinash M N's avatar Avinash M N Committed by Jim Harris
Browse files

nvmf: Add IP based load balancing 



This patch adds support for IP based load balancing. All the rdma connections
coming from a particular IP will be assigned the same core on target. The load
balancing policy - roundrobin/hostip can be configured through rpc during startup only.
Roundrobin is the default policy.

Change-Id: I7bb216cb4cb2639fa401e23090143c4cf1d7bb5a
Signed-off-by: default avatarAvinash M N <avinash.m.n@wdc.com>
Reviewed-on: https://review.gerrithub.io/422190


Tested-by: default avatarSPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: default avatarBen Walker <benjamin.walker@intel.com>
Reviewed-by: default avatarJim Harris <james.r.harris@intel.com>
Chandler-Test-Pool: SPDK Automated Test System <sys_sgsw@intel.com>
parent 94f87a2d

lib/event/subsystems/nvmf/conf.c

+1 −0
Original line number Diff line number Diff line
@@ -146,6 +146,7 @@ spdk_nvmf_parse_tgt_conf(void) 
	}



	conf->acceptor_poll_rate = ACCEPT_TIMEOUT_US;

	conf->conn_sched = DEFAULT_CONN_SCHED;



	sp = spdk_conf_find_section(NULL, "Nvmf");

	if (sp != NULL) {

lib/event/subsystems/nvmf/event_nvmf.h

+7 −0
Original line number Diff line number Diff line
@@ -43,9 +43,16 @@ 
#include "spdk_internal/log.h"



#define ACCEPT_TIMEOUT_US	10000 /* 10ms */

#define DEFAULT_CONN_SCHED CONNECT_SCHED_ROUND_ROBIN



enum spdk_nvmf_connect_sched {

	CONNECT_SCHED_ROUND_ROBIN = 0,

	CONNECT_SCHED_HOST_IP,

};



struct spdk_nvmf_tgt_conf {

	uint32_t acceptor_poll_rate;

	enum spdk_nvmf_connect_sched conn_sched;

};



extern struct spdk_nvmf_tgt_opts *g_spdk_nvmf_tgt_opts;

lib/event/subsystems/nvmf/nvmf_rpc.c

+18 −0
Original line number Diff line number Diff line
@@ -1618,8 +1618,25 @@ nvmf_rpc_subsystem_set_tgt_opts(struct spdk_jsonrpc_request *request, 
}

SPDK_RPC_REGISTER("set_nvmf_target_options", nvmf_rpc_subsystem_set_tgt_opts, SPDK_RPC_STARTUP)



static int decode_conn_sched(const struct spdk_json_val *val, void *out)

{

	enum spdk_nvmf_connect_sched *sched = out;



	if (spdk_json_strequal(val, "roundrobin") == true) {

		*sched = CONNECT_SCHED_ROUND_ROBIN;

	} else if (spdk_json_strequal(val, "hostip") == true) {

		*sched = CONNECT_SCHED_HOST_IP;

	} else {

		SPDK_ERRLOG("Invalid connection scheduling parameter\n");

		return -EINVAL;

	}



	return 0;

}



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},

};



static void
@@ -1645,6 +1662,7 @@ nvmf_rpc_subsystem_set_tgt_conf(struct spdk_jsonrpc_request *request, 
	}



	conf->acceptor_poll_rate = ACCEPT_TIMEOUT_US;

	conf->conn_sched = DEFAULT_CONN_SCHED;



	if (params != NULL) {

		if (spdk_json_decode_object(params, nvmf_rpc_subsystem_tgt_conf_decoder,

lib/event/subsystems/nvmf/nvmf_tgt.c

+96 −6
Original line number Diff line number Diff line
@@ -59,11 +59,23 @@ struct nvmf_tgt_poll_group { 
	struct spdk_nvmf_poll_group *group;

};



struct nvmf_tgt_host_trid {

	struct spdk_nvme_transport_id       host_trid;

	uint32_t                            core;

	uint32_t                            ref;

	TAILQ_ENTRY(nvmf_tgt_host_trid)     link;

};



/* List of host trids that are connected to the target */

static TAILQ_HEAD(, nvmf_tgt_host_trid) g_nvmf_tgt_host_trids =

	TAILQ_HEAD_INITIALIZER(g_nvmf_tgt_host_trids);



struct spdk_nvmf_tgt *g_spdk_nvmf_tgt = NULL;



static enum nvmf_tgt_state g_tgt_state;



static uint32_t g_tgt_core; /* Round-robin tracking of cores for qpair assignment */

/* Round-Robin/IP-based tracking of cores for qpair assignment */

static uint32_t g_tgt_core;



static struct nvmf_tgt_poll_group *g_poll_groups = NULL;

static size_t g_num_poll_groups = 0;
@@ -111,6 +123,68 @@ nvmf_tgt_poll_group_add(void *arg1, void *arg2) 
	spdk_nvmf_poll_group_add(pg->group, qpair);

}



/* Round robin selection of cores */

static uint32_t

spdk_nvmf_get_core_rr(void)

{

	uint32_t core;



	core = g_tgt_core;

	g_tgt_core = spdk_env_get_next_core(core);

	if (g_tgt_core == UINT32_MAX) {

		g_tgt_core = spdk_env_get_first_core();

	}



	return core;

}



static uint32_t

nvmf_tgt_get_qpair_core(struct spdk_nvmf_qpair *qpair)

{

	struct spdk_nvme_transport_id trid;

	struct nvmf_tgt_host_trid *tmp_trid = NULL, *new_trid = NULL;

	int ret;

	uint32_t core = 0;



	switch (g_spdk_nvmf_tgt_conf->conn_sched) {

	case CONNECT_SCHED_HOST_IP:

		ret = spdk_nvmf_qpair_get_peer_trid(qpair, &trid);

		if (ret) {

			SPDK_ERRLOG("Invalid host transport Id. Assigning to core %d\n", core);

			break;

		}



		TAILQ_FOREACH(tmp_trid, &g_nvmf_tgt_host_trids, link) {

			if (tmp_trid && !strncmp(tmp_trid->host_trid.traddr,

						 trid.traddr, SPDK_NVMF_TRADDR_MAX_LEN + 1)) {

				tmp_trid->ref++;

				core = tmp_trid->core;

				break;

			}

		}

		if (!tmp_trid) {

			new_trid = calloc(1, sizeof(*new_trid));

			if (!new_trid) {

				SPDK_ERRLOG("Insufficient memory. Assigning to core %d\n", core);

				break;

			}

			/* Get the next available core for the new host */

			core = spdk_nvmf_get_core_rr();

			new_trid->core = core;

			memcpy(new_trid->host_trid.traddr, trid.traddr,

			       SPDK_NVMF_TRADDR_MAX_LEN + 1);

			TAILQ_INSERT_TAIL(&g_nvmf_tgt_host_trids, new_trid, link);

		}

		break;

	case CONNECT_SCHED_ROUND_ROBIN:

	default:

		core = spdk_nvmf_get_core_rr();

		break;

	}



	return core;

}



static void

new_qpair(struct spdk_nvmf_qpair *qpair)

{
@@ -123,11 +197,7 @@ new_qpair(struct spdk_nvmf_qpair *qpair) 
		return;

	}



	core = g_tgt_core;

	g_tgt_core = spdk_env_get_next_core(core);

	if (g_tgt_core == UINT32_MAX) {

		g_tgt_core = spdk_env_get_first_core();

	}

	core = nvmf_tgt_get_qpair_core(qpair);



	pg = &g_poll_groups[core];

	assert(pg != NULL);
@@ -222,7 +292,15 @@ nvmf_tgt_subsystem_stopped(struct spdk_nvmf_subsystem *subsystem, 
static void

nvmf_tgt_destroy_done(void *ctx, int status)

{

	struct nvmf_tgt_host_trid *trid, *tmp_trid;



	g_tgt_state = NVMF_TGT_STOPPED;



	TAILQ_FOREACH_SAFE(trid, &g_nvmf_tgt_host_trids, link, tmp_trid) {

		TAILQ_REMOVE(&g_nvmf_tgt_host_trids, trid, link);

		free(trid);

	}



	free(g_spdk_nvmf_tgt_conf);

	nvmf_tgt_advance_state();

}
@@ -334,6 +412,16 @@ spdk_nvmf_subsystem_init(void) 
	nvmf_tgt_advance_state();

}



static char *

get_conn_sched_string(enum spdk_nvmf_connect_sched sched)

{

	if (sched == CONNECT_SCHED_HOST_IP) {

		return "hostip";

	} else {

		return "roundrobin";

	}

}



static void

spdk_nvmf_subsystem_write_config_json(struct spdk_json_write_ctx *w, struct spdk_event *done_ev)

{
@@ -344,6 +432,8 @@ spdk_nvmf_subsystem_write_config_json(struct spdk_json_write_ctx *w, struct spdk 


	spdk_json_write_named_object_begin(w, "params");

	spdk_json_write_named_uint32(w, "acceptor_poll_rate", g_spdk_nvmf_tgt_conf->acceptor_poll_rate);

	spdk_json_write_named_string(w, "conn_sched",

				     get_conn_sched_string(g_spdk_nvmf_tgt_conf->conn_sched));

	spdk_json_write_object_end(w);

	spdk_json_write_object_end(w);

scripts/rpc.py

+6 −1
Original line number Diff line number Diff line
@@ -1238,10 +1238,15 @@ Format: 'user:u1 secret:s1 muser:mu1 msecret:ms1,user:u2 secret:s2 muser:mu2 mse 
    @call_cmd

    def set_nvmf_target_config(args):

        rpc.nvmf.set_nvmf_target_config(args.client,

                                        acceptor_poll_rate=args.acceptor_poll_rate)

                                        acceptor_poll_rate=args.acceptor_poll_rate,

                                        conn_sched=args.conn_sched)



    p = subparsers.add_parser('set_nvmf_target_config', help='Set NVMf target config')

    p.add_argument('-r', '--acceptor-poll-rate', help='Polling interval of the acceptor for incoming connections (usec)', type=int)

    p.add_argument('-s', '--conn-sched', help="""'roundrobin' - Schedule the incoming connections from any host

    on the cores in a round robin manner (Default). 'hostip' - Schedule all the incoming connections from a

    specific host IP on to the same core. Connections from different IP will be assigned to cores in a round

    robin manner""")

    p.set_defaults(func=set_nvmf_target_config)



    @call_cmd