Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

nvmf/rdma: Add NUMA aware policy #12

Open
wants to merge 1 commit into
base: upstream_master
Choose a base branch
from
Open

nvmf/rdma: Add NUMA aware policy #12

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
2 changes: 2 additions & 0 deletions etc/spdk/nvmf.conf.in
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -111,6 +111,8 @@
# Set the maximum number outstanding I/O per shared receive queue. Relevant only for RDMA transport
#MaxSRQDepth 4096

# Set NumaAwarePolicy for ConnectionScheduler: Transport. Relevant only for RDMA transport.
#NumaAwarePolicy False
[Transport]
# Set TCP transport type.
Type TCP
Expand Down
1 change: 1 addition & 0 deletions include/spdk/nvmf.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -83,6 +83,7 @@ struct spdk_nvmf_transport_opts {
bool no_srq;
bool c2h_success;
bool dif_insert_or_strip;
bool numa_policy;
uint32_t sock_priority;
};

Expand Down
100 changes: 97 additions & 3 deletions lib/nvmf/rdma.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,6 +36,8 @@
#include <infiniband/verbs.h>
#include <rdma/rdma_cma.h>
#include <rdma/rdma_verbs.h>
#include <numa.h>
#include <numaif.h>

#include "spdk/config.h"
#include "spdk/thread.h"
Expand Down Expand Up @@ -463,16 +465,21 @@ struct spdk_nvmf_rdma_poll_group {
struct spdk_nvmf_rdma_poll_group_stat stat;
TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers;
TAILQ_ENTRY(spdk_nvmf_rdma_poll_group) link;

int numa_node;
/*
* buffers which are split across multiple RDMA
* memory regions cannot be used by this transport.
*/
STAILQ_HEAD(, spdk_nvmf_transport_pg_cache_buf) retired_bufs;
};

#define MAX_NUMA 8

struct spdk_nvmf_rdma_conn_sched {
struct spdk_nvmf_rdma_poll_group *next_admin_pg;
struct spdk_nvmf_rdma_poll_group *next_io_pg;
struct spdk_nvmf_rdma_poll_group *next_io_pgs[MAX_NUMA];
};

/* Assuming rdma_cm uses just one protection domain per ibv_context. */
Expand All @@ -484,6 +491,7 @@ struct spdk_nvmf_rdma_device {
struct ibv_pd *pd;

int num_srq;
int numa_node;

TAILQ_ENTRY(spdk_nvmf_rdma_device) link;
};
Expand Down Expand Up @@ -2315,6 +2323,7 @@ spdk_nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport,
#define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32
#define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false
#define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false
#define SPDK_NVMF_RDMA_NUMA_AWARE_POLICY false;

static void
spdk_nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts)
Expand All @@ -2330,6 +2339,7 @@ spdk_nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts)
opts->max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH;
opts->no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ;
opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP;
opts->numa_policy = SPDK_NVMF_RDMA_NUMA_AWARE_POLICY;
}

const struct spdk_mem_map_ops g_nvmf_rdma_map_ops = {
Expand All @@ -2339,6 +2349,32 @@ const struct spdk_mem_map_ops g_nvmf_rdma_map_ops = {

static int spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport);

static int
spdk_nvmf_rdma_get_numa_node_for_device(const char *device_name)
{
int node = 0;
#if defined(__linux__)
FILE *numa_file;
char path[PATH_MAX] = {};
int res;

snprintf(path, sizeof(path), "/sys/class/infiniband/%s/device/numa_node", device_name);

numa_file = fopen(path, "r");
if (!numa_file) {
SPDK_ERRLOG("file with numa node for device not exist\n");
return 0;
}
res = fscanf(numa_file, "%d", &node);
if (res <= 0) {
SPDK_ERRLOG("file with numa node for device is empty\n");
node = 0;
}
fclose(numa_file);
#endif
return node;
}

static struct spdk_nvmf_transport *
spdk_nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts)
{
Expand Down Expand Up @@ -2384,13 +2420,17 @@ spdk_nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts)
TAILQ_INIT(&rtransport->ports);
TAILQ_INIT(&rtransport->poll_groups);

#ifdef __aarch64__
rtransport->transport.opts.numa_policy = false;
#endif
rtransport->transport.ops = &spdk_nvmf_transport_rdma;

SPDK_INFOLOG(SPDK_LOG_RDMA, "*** RDMA Transport Init ***\n"
" Transport opts: max_ioq_depth=%d, max_io_size=%d,\n"
" max_qpairs_per_ctrlr=%d, io_unit_size=%d,\n"
" in_capsule_data_size=%d, max_aq_depth=%d,\n"
" num_shared_buffers=%d, max_srq_depth=%d, no_srq=%d\n",
" num_shared_buffers=%d, max_srq_depth=%d, no_srq=%d\n,"
" numa_policy=%d\n",
opts->max_queue_depth,
opts->max_io_size,
opts->max_qpairs_per_ctrlr,
Expand All @@ -2399,7 +2439,8 @@ spdk_nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts)
opts->max_aq_depth,
opts->num_shared_buffers,
opts->max_srq_depth,
opts->no_srq);
opts->no_srq,
opts->numa_policy);

/* I/O unit size cannot be larger than max I/O size */
if (opts->io_unit_size > opts->max_io_size) {
Expand Down Expand Up @@ -2539,6 +2580,8 @@ spdk_nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts)

assert(device->map != NULL);
assert(device->pd != NULL);

device->numa_node = spdk_nvmf_rdma_get_numa_node_for_device(ibv_get_device_name(device->context->device));
}
rdma_free_devices(contexts);

Expand Down Expand Up @@ -3328,7 +3371,11 @@ spdk_nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport)
if (!rgroup) {
return NULL;
}

#ifdef __aarch64__
rgroup->numa_node = 0;
#else
rgroup->numa_node = numa_node_of_cpu(sched_getcpu());
#endif
TAILQ_INIT(&rgroup->pollers);
STAILQ_INIT(&rgroup->retired_bufs);

Expand Down Expand Up @@ -3404,15 +3451,57 @@ spdk_nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport)
}

TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link);

if (rtransport->conn_sched.next_admin_pg == NULL) {
rtransport->conn_sched.next_admin_pg = rgroup;
rtransport->conn_sched.next_io_pg = rgroup;
}
rtransport->conn_sched.next_io_pgs[rgroup->numa_node] = rgroup;

pthread_mutex_unlock(&rtransport->lock);

return &rgroup->group;
}

static struct spdk_nvmf_transport_poll_group *
spdk_nvmf_rdma_get_optimal_numa_poll_group(struct spdk_nvmf_qpair *qpair)
{
struct spdk_nvmf_rdma_transport *rtransport;
struct spdk_nvmf_rdma_poll_group **pg;
struct spdk_nvmf_transport_poll_group *result;
struct spdk_nvmf_rdma_qpair *rqpair;

rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);
rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);

pthread_mutex_lock(&rtransport->lock);

if (TAILQ_EMPTY(&rtransport->poll_groups)) {
pthread_mutex_unlock(&rtransport->lock);
return NULL;
}

if (qpair->qid == 0) {
pg = &rtransport->conn_sched.next_admin_pg;
} else {
pg = &rtransport->conn_sched.next_io_pgs[rqpair->device->numa_node];
}
assert(*pg != NULL);
Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The user can specify reactors mask which will use only 1 numa node (e.g. 0xf - 4 cores on numa 0). But the new connection can be triggered on device attached to numa 1. In this case conn_sched.next_io_pgs[1] will be equal NULL and you'll hit this assert. I think that we should fall back to existing scheduling policy if we don't find a poll group for specific numa node


result = &(*pg)->group;

do {
*pg = TAILQ_NEXT(*pg, link);
if (*pg == NULL) {
*pg = TAILQ_FIRST(&rtransport->poll_groups);
}
} while ((*pg)->numa_node != rqpair->device->numa_node);

pthread_mutex_unlock(&rtransport->lock);

return result;
}

static struct spdk_nvmf_transport_poll_group *
spdk_nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair)
{
Expand All @@ -3422,6 +3511,10 @@ spdk_nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair)

rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);

if (rtransport->transport.opts.numa_policy) {
return spdk_nvmf_rdma_get_optimal_numa_poll_group(qpair);
}

pthread_mutex_lock(&rtransport->lock);

if (TAILQ_EMPTY(&rtransport->poll_groups)) {
Expand All @@ -3440,6 +3533,7 @@ spdk_nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair)
result = &(*pg)->group;

*pg = TAILQ_NEXT(*pg, link);

if (*pg == NULL) {
*pg = TAILQ_FIRST(&rtransport->poll_groups);
}
Expand Down
2 changes: 2 additions & 0 deletions module/event/subsystems/nvmf/conf.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -644,6 +644,8 @@ spdk_nvmf_parse_transport(struct spdk_nvmf_parse_transport_ctx *ctx)
}
bval = spdk_conf_section_get_boolval(ctx->sp, "NoSRQ", false);
opts.no_srq = bval;
bval = spdk_conf_section_get_boolval(ctx->sp, "NumaAwarePolicy", false);
opts.numa_policy = bval;
}

if (trtype == SPDK_NVME_TRANSPORT_TCP) {
Expand Down