/*
- * rdma engine
+ * RDMA I/O engine
*
- * RDMA IO engine using OFED library.
- * Support both RDMA memory semantic and channel semantic
- * in InfiniBand, RoCE and iWarp environment.
+ * RDMA I/O engine based on the IB verbs and RDMA/CM user space libraries.
+ * Supports both RDMA memory semantics and channel semantics
+ * for the InfiniBand, RoCE and iWARP protocols.
*
- * This is currently disabled. To enable it, execute:
- *
- * $ export EXTFLAGS="-DFIO_HAVE_RDMA"
- * $ export EXTLIBS="-libverbs -lrdmacm"
- *
- * before running make. You'll need the OFED as well:
+ * You will need the Linux RDMA software installed, either
+ * from your Linux distributor or directly from openfabrics.org:
*
* http://www.openfabrics.org/downloads/OFED/
*
+ * Exchanging steps of RDMA ioengine control messages:
+ * 1. client side sends test mode (RDMA_WRITE/RDMA_READ/SEND)
+ * to server side.
+ * 2. server side parses test mode, and sends back confirmation
+ * to client side. In RDMA WRITE/READ test, this confirmation
+ * includes memory information, such as rkey, address.
+ * 3. client side initiates test loop.
+ * 4. In RDMA WRITE/READ test, client side sends a completion
+ * notification to server side. Server side updates its
+ * td->done as true.
+ *
*/
#include <stdio.h>
#include <stdlib.h>
#include <sys/time.h>
#include <sys/resource.h>
-#include <byteswap.h>
#include <pthread.h>
#include <inttypes.h>
#include "../fio.h"
-
-#ifdef FIO_HAVE_RDMA
+#include "../hash.h"
#include <rdma/rdma_cma.h>
#include <infiniband/arch.h>
-#define FIO_RDMA_MAX_IO_DPETH 128
+#define FIO_RDMA_MAX_IO_DEPTH 512
enum rdma_io_mode {
FIO_RDMA_UNKNOWN = 0,
uint32_t nr; /* client: io depth
server: number of records for memory semantic
*/
- struct remote_u rmt_us[FIO_RDMA_MAX_IO_DPETH];
+ struct remote_u rmt_us[FIO_RDMA_MAX_IO_DEPTH];
};
struct rdma_io_u_data {
int io_u_flight_nr;
struct io_u **io_us_completed;
int io_u_completed_nr;
+
+ struct frand_state rand_state;
};
static int client_recv(struct thread_data *td, struct ibv_wc *wc)
{
struct rdmaio_data *rd = td->io_ops->data;
- if (wc->wr_id == FIO_RDMA_MAX_IO_DPETH) {
+ if (wc->wr_id == FIO_RDMA_MAX_IO_DEPTH) {
rd->rdma_protocol = ntohl(rd->recv_buf.mode);
/* CHANNEL semantic, do nothing */
else
server_recv(td, &wc);
- if (wc.wr_id == FIO_RDMA_MAX_IO_DPETH)
+ if (wc.wr_id == FIO_RDMA_MAX_IO_DEPTH)
break;
for (i = 0; i < rd->io_u_flight_nr; i++) {
}
}
if (i == rd->io_u_flight_nr)
- log_err("fio: recv wr %ld not found\n",
+ log_err("fio: recv wr %" PRId64 " not found\n",
wc.wr_id);
else {
/* put the last one into middle of the list */
case IBV_WC_SEND:
case IBV_WC_RDMA_WRITE:
case IBV_WC_RDMA_READ:
- if (wc.wr_id == FIO_RDMA_MAX_IO_DPETH)
+ if (wc.wr_id == FIO_RDMA_MAX_IO_DEPTH)
break;
for (i = 0; i < rd->io_u_flight_nr; i++) {
}
}
if (i == rd->io_u_flight_nr)
- log_err("fio: send wr %ld not found\n",
+ log_err("fio: send wr %" PRId64 " not found\n",
wc.wr_id);
else {
/* put the last one into middle of the list */
rd->pd = ibv_alloc_pd(rd->child_cm_id->verbs);
else
rd->pd = ibv_alloc_pd(rd->cm_id->verbs);
+
if (rd->pd == NULL) {
log_err("fio: ibv_alloc_pd fail\n");
return 1;
/* setup work request */
/* recv wq */
rd->recv_sgl.addr = (uint64_t) (unsigned long)&rd->recv_buf;
- rd->recv_sgl.length = sizeof rd->recv_buf;
+ rd->recv_sgl.length = sizeof(rd->recv_buf);
rd->recv_sgl.lkey = rd->recv_mr->lkey;
rd->rq_wr.sg_list = &rd->recv_sgl;
rd->rq_wr.num_sge = 1;
- rd->rq_wr.wr_id = FIO_RDMA_MAX_IO_DPETH;
+ rd->rq_wr.wr_id = FIO_RDMA_MAX_IO_DEPTH;
/* send wq */
rd->send_sgl.addr = (uint64_t) (unsigned long)&rd->send_buf;
- rd->send_sgl.length = sizeof rd->send_buf;
+ rd->send_sgl.length = sizeof(rd->send_buf);
rd->send_sgl.lkey = rd->send_mr->lkey;
rd->sq_wr.opcode = IBV_WR_SEND;
rd->sq_wr.send_flags = IBV_SEND_SIGNALED;
rd->sq_wr.sg_list = &rd->send_sgl;
rd->sq_wr.num_sge = 1;
- rd->sq_wr.wr_id = FIO_RDMA_MAX_IO_DPETH;
+ rd->sq_wr.wr_id = FIO_RDMA_MAX_IO_DEPTH;
return 0;
}
enum rdma_cm_event_type wait_event)
{
struct rdmaio_data *rd = td->io_ops->data;
-
- int ret;
struct rdma_cm_event *event;
+ int ret;
ret = rdma_get_cm_event(channel, &event);
if (ret) {
- log_err("fio: rdma_get_cm_event");
+ log_err("fio: rdma_get_cm_event: %d\n", ret);
return 1;
}
int i;
io_u = rd->io_us_completed[0];
- for (i = 0; i < rd->io_u_completed_nr - 1; i++) {
+ for (i = 0; i < rd->io_u_completed_nr - 1; i++)
rd->io_us_completed[i] = rd->io_us_completed[i + 1];
- }
+
rd->io_u_completed_nr--;
dprint_io_u(io_u, "fio_rdmaio_event");
unsigned int max, struct timespec *t)
{
struct rdmaio_data *rd = td->io_ops->data;
- int r;
enum ibv_wc_opcode comp_opcode;
- comp_opcode = IBV_WC_RDMA_WRITE;
struct ibv_cq *ev_cq;
void *ev_ctx;
- int ret;
-
- r = 0;
+ int ret, r = 0;
+ comp_opcode = IBV_WC_RDMA_WRITE;
switch (rd->rdma_protocol) {
case FIO_RDMA_MEM_WRITE:
{
struct rdmaio_data *rd = td->io_ops->data;
struct ibv_send_wr *bad_wr;
+#if 0
enum ibv_wc_opcode comp_opcode;
comp_opcode = IBV_WC_RDMA_WRITE;
- int i, index;
+#endif
+ int i;
+ long index;
struct rdma_io_u_data *r_io_u_d;
r_io_u_d = NULL;
case FIO_RDMA_MEM_WRITE:
/* compose work request */
r_io_u_d = io_us[i]->engine_data;
- index = rand() % rd->rmt_nr;
+ index = __rand(&rd->rand_state) % rd->rmt_nr;
r_io_u_d->sq_wr.opcode = IBV_WR_RDMA_WRITE;
r_io_u_d->sq_wr.wr.rdma.rkey = rd->rmt_us[index].rkey;
r_io_u_d->sq_wr.wr.rdma.remote_addr = \
case FIO_RDMA_MEM_READ:
/* compose work request */
r_io_u_d = io_us[i]->engine_data;
- index = rand() % rd->rmt_nr;
+ index = __rand(&rd->rand_state) % rd->rmt_nr;
r_io_u_d->sq_wr.opcode = IBV_WR_RDMA_READ;
r_io_u_d->sq_wr.wr.rdma.rkey = rd->rmt_us[index].rkey;
r_io_u_d->sq_wr.wr.rdma.remote_addr = \
rdma_poll_wait(td, IBV_WC_RECV);
dprint(FD_IO, "fio: recv FINISH message\n");
- exit(0);
+ td->done = 1;
+ return 0;
}
return i;
io_us = rd->io_us_queued;
do {
/* RDMA_WRITE or RDMA_READ */
- if (rd->is_client) {
+ if (rd->is_client)
ret = fio_rdmaio_send(td, io_us, rd->io_u_queued_nr);
- } else if (!rd->is_client) {
+ else if (!rd->is_client)
ret = fio_rdmaio_recv(td, io_us, rd->io_u_queued_nr);
- } else
+ else
ret = 0; /* must be a SYNC */
if (ret > 0) {
struct rdma_conn_param conn_param;
struct ibv_send_wr *bad_wr;
- memset(&conn_param, 0, sizeof conn_param);
+ memset(&conn_param, 0, sizeof(conn_param));
conn_param.responder_resources = 1;
conn_param.initiator_depth = 1;
conn_param.retry_count = 10;
/* wait for remote MR info from server side */
rdma_poll_wait(td, IBV_WC_RECV);
+ /* In SEND/RECV test, it's a good practice to setup the iodepth of
+ * of the RECV side deeper than that of the SEND side to
+ * avoid RNR (receiver not ready) error. The
+ * SEND side may send so many unsolicited message before
+ * RECV side commits sufficient recv buffers into recv queue.
+ * This may lead to RNR error. Here, SEND side pauses for a while
+ * during which RECV side commits sufficient recv buffers.
+ */
+ usleep(500000);
+
return 0;
}
struct ibv_send_wr *bad_wr;
/* rdma_accept() - then wait for accept success */
- memset(&conn_param, 0, sizeof conn_param);
+ memset(&conn_param, 0, sizeof(conn_param));
conn_param.responder_resources = 1;
conn_param.initiator_depth = 1;
return 1;
}
- dprint(FD_IO, "fio: close infomation sent success\n");
+ dprint(FD_IO, "fio: close information sent success\n");
rdma_poll_wait(td, IBV_WC_SEND);
}
rdma_disconnect(rd->cm_id);
else {
rdma_disconnect(rd->child_cm_id);
-/* rdma_disconnect(rd->cm_id); */
+#if 0
+ rdma_disconnect(rd->cm_id);
+#endif
}
-/* if (get_next_channel_event(td, rd->cm_channel, RDMA_CM_EVENT_DISCONNECTED) != 0)
- {
- log_err("fio: wait for RDMA_CM_EVENT_DISCONNECTED\n");
- return 1;
- }*/
+#if 0
+ if (get_next_channel_event(td, rd->cm_channel, RDMA_CM_EVENT_DISCONNECTED) != 0) {
+ log_err("fio: wait for RDMA_CM_EVENT_DISCONNECTED\n");
+ return 1;
+ }
+#endif
- ibv_destroy_qp(rd->qp);
ibv_destroy_cq(rd->cq);
+ ibv_destroy_qp(rd->qp);
if (rd->is_client == 1)
rdma_destroy_id(rd->cm_id);
{
struct rdmaio_data *rd = td->io_ops->data;
struct ibv_recv_wr *bad_wr;
+ int err;
rd->addr.sin_family = AF_INET;
rd->addr.sin_port = htons(port);
}
/* resolve route */
- if (rdma_resolve_addr(rd->cm_id, NULL,
- (struct sockaddr *)&rd->addr, 2000) != 0) {
- log_err("fio: rdma_resolve_addr");
+ err = rdma_resolve_addr(rd->cm_id, NULL, (struct sockaddr *)&rd->addr, 2000);
+ if (err != 0) {
+ log_err("fio: rdma_resolve_addr: %d\n", err);
return 1;
}
- if (get_next_channel_event
- (td, rd->cm_channel, RDMA_CM_EVENT_ADDR_RESOLVED)
- != 0) {
- log_err("fio: get_next_channel_event");
+ err = get_next_channel_event(td, rd->cm_channel, RDMA_CM_EVENT_ADDR_RESOLVED);
+ if (err != 0) {
+ log_err("fio: get_next_channel_event: %d\n", err);
return 1;
}
/* resolve route */
- if (rdma_resolve_route(rd->cm_id, 2000) != 0) {
- log_err("fio: rdma_resolve_route");
+ err = rdma_resolve_route(rd->cm_id, 2000);
+ if (err != 0) {
+ log_err("fio: rdma_resolve_route: %d\n", err);
return 1;
}
- if (get_next_channel_event
- (td, rd->cm_channel, RDMA_CM_EVENT_ROUTE_RESOLVED) != 0) {
- log_err("fio: get_next_channel_event");
+ err = get_next_channel_event(td, rd->cm_channel, RDMA_CM_EVENT_ROUTE_RESOLVED);
+ if (err != 0) {
+ log_err("fio: get_next_channel_event: %d\n", err);
return 1;
}
return 1;
/* post recv buf */
- if (ibv_post_recv(rd->qp, &rd->rq_wr, &bad_wr) != 0) {
- log_err("fio: ibv_post_recv fail\n");
+ err = ibv_post_recv(rd->qp, &rd->rq_wr, &bad_wr);
+ if (err != 0) {
+ log_err("fio: ibv_post_recv fail: %d\n", err);
return 1;
}
return 0;
}
-static int fio_rdmaio_init(struct thread_data *td)
+static int check_set_rlimits(struct thread_data *td)
{
- struct rdmaio_data *rd = td->io_ops->data;
- unsigned int port;
- char host[64], buf[128];
- char *sep, *portp, *modep;
- int ret;
+#ifdef CONFIG_RLIMIT_MEMLOCK
struct rlimit rl;
- if (td_rw(td)) {
- log_err("fio: rdma connections must be read OR write\n");
- return 1;
- }
- if (td_random(td)) {
- log_err("fio: RDMA network IO can't be random\n");
- return 1;
- }
-
/* check RLIMIT_MEMLOCK */
if (getrlimit(RLIMIT_MEMLOCK, &rl) != 0) {
log_err("fio: getrlimit fail: %d(%s)\n",
/* soft limit */
if ((rl.rlim_cur != RLIM_INFINITY)
&& (rl.rlim_cur < td->orig_buffer_size)) {
- log_err("fio: soft RLIMIT_MEMLOCK is: %ld\n", rl.rlim_cur);
- log_err("fio: total block size is: %ld\n",
+ log_err("fio: soft RLIMIT_MEMLOCK is: %" PRId64 "\n",
+ rl.rlim_cur);
+ log_err("fio: total block size is: %zd\n",
td->orig_buffer_size);
/* try to set larger RLIMIT_MEMLOCK */
rl.rlim_cur = rl.rlim_max;
return 1;
}
}
+#endif
+
+ return 0;
+}
+
+static int fio_rdmaio_init(struct thread_data *td)
+{
+ struct rdmaio_data *rd = td->io_ops->data;
+ unsigned int max_bs;
+ unsigned int port;
+ char host[64], buf[128];
+ char *sep, *portp, *modep;
+ int ret, i;
+
+ if (td_rw(td)) {
+ log_err("fio: rdma connections must be read OR write\n");
+ return 1;
+ }
+ if (td_random(td)) {
+ log_err("fio: RDMA network IO can't be random\n");
+ return 1;
+ }
+
+ if (check_set_rlimits(td))
+ return 1;
strcpy(buf, td->o.filename);
if ((rd->rdma_protocol == FIO_RDMA_MEM_WRITE) ||
(rd->rdma_protocol == FIO_RDMA_MEM_READ)) {
rd->rmt_us =
- malloc(FIO_RDMA_MAX_IO_DPETH * sizeof(struct remote_u));
+ malloc(FIO_RDMA_MAX_IO_DEPTH * sizeof(struct remote_u));
memset(rd->rmt_us, 0,
- FIO_RDMA_MAX_IO_DPETH * sizeof(struct remote_u));
+ FIO_RDMA_MAX_IO_DEPTH * sizeof(struct remote_u));
rd->rmt_nr = 0;
}
ret = fio_rdmaio_setup_connect(td, host, port);
}
- struct flist_head *entry;
- unsigned int max_bs;
max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
/* register each io_u in the free list */
- int i = 0;
- flist_for_each(entry, &td->io_u_freelist) {
- struct io_u *io_u = flist_entry(entry, struct io_u, list);
+ for (i = 0; i < td->io_u_freelist.nr; i++) {
+ struct io_u *io_u = td->io_u_freelist.io_us[i];
io_u->engine_data = malloc(sizeof(struct rdma_io_u_data));
memset(io_u->engine_data, 0, sizeof(struct rdma_io_u_data));
rd->send_buf.rmt_us[i].rkey = htonl(io_u->mr->rkey);
rd->send_buf.rmt_us[i].size = htonl(max_bs);
-/* log_info("fio: Send rkey %x addr %" PRIx64 " len %d to client\n",
- io_u->mr->rkey, io_u->buf, max_bs); */
- i++;
+#if 0
+ log_info("fio: Send rkey %x addr %" PRIx64 " len %d to client\n", io_u->mr->rkey, io_u->buf, max_bs); */
+#endif
}
rd->send_buf.nr = htonl(i);
{
struct rdmaio_data *rd = td->io_ops->data;
- if (rd) {
-/* if (nd->listenfd != -1)
- close(nd->listenfd);
- if (nd->pipes[0] != -1)
- close(nd->pipes[0]);
- if (nd->pipes[1] != -1)
- close(nd->pipes[1]);
-*/
+ if (rd)
free(rd);
- }
}
static int fio_rdmaio_setup(struct thread_data *td)
struct rdmaio_data *rd;
if (!td->io_ops->data) {
- rd = malloc(sizeof(*rd));;
+ rd = malloc(sizeof(*rd));
memset(rd, 0, sizeof(*rd));
+ init_rand_seed(&rd->rand_state, (unsigned int) GOLDEN_RATIO_PRIME);
td->io_ops->data = rd;
}
.flags = FIO_DISKLESSIO | FIO_UNIDIR | FIO_PIPEIO,
};
-#else /* FIO_HAVE_RDMA */
-
-static int fio_rdmaio_open_file(struct thread_data *td, struct fio_file *f)
-{
- return 0;
-}
-
-static int fio_rdmaio_close_file(struct thread_data *td, struct fio_file *f)
-{
- return 0;
-}
-
-static int fio_rdmaio_queue(struct thread_data *td, struct io_u *io_u)
-{
- return FIO_Q_COMPLETED;
-}
-
-static int fio_rdmaio_init(struct thread_data fio_unused * td)
-{
- log_err("fio: rdma(librdmacm libibverbs) not available\n");
- log_err(" You haven't compiled rdma ioengine into fio.\n");
- log_err(" If you want to try rdma ioengine,\n");
- log_err(" make sure OFED is installed,\n");
- log_err(" $ ofed_info\n");
- log_err(" then try to make fio as follows:\n");
- log_err(" $ export EXTFLAGS=\"-DFIO_HAVE_RDMA\"\n");
- log_err(" $ export EXTLIBS=\"-libverbs -lrdmacm\"\n");
- log_err(" $ make clean && make\n");
- return 1;
-}
-
-static struct ioengine_ops ioengine_rw = {
- .name = "rdma",
- .version = FIO_IOOPS_VERSION,
- .init = fio_rdmaio_init,
- .queue = fio_rdmaio_queue,
- .open_file = fio_rdmaio_open_file,
- .close_file = fio_rdmaio_close_file,
- .flags = FIO_SYNCIO | FIO_DISKLESSIO | FIO_UNIDIR | FIO_PIPEIO,
-};
-
-#endif
-
static void fio_init fio_rdmaio_register(void)
{
register_ioengine(&ioengine_rw);