The following changes since commit 477eb7ef564116dcedea92c8d5ed76c2c6fe1647:
Add LICENSE file specyfing what I expect from people using fio (2012-10-23 20:58:35 +0200)
are available in the git repository at:
git://git.kernel.dk/fio.git master
Jens Axboe (4):
rdma: use private random state
rdma: fixup compile issue
rdma: cleanup and fixes
Add max_latency option
Yufei Ren (2):
Replace FIO_HAVE_RUSAGE_THREAD with RUSAGE_THREAD
rdma ioengine improvement
HOWTO | 3 +
backend.c | 2 +-
engines/rdma.c | 142 ++++++++++++++++++++++++++++++++------------------------
fio.1 | 4 ++
fio.h | 2 +
io_u.c | 7 +++
options.c | 6 ++
os/os-linux.h | 5 --
stat.c | 2 +-
9 files changed, 105 insertions(+), 68 deletions(-)
---
Diff of recent changes:
diff --git a/HOWTO b/HOWTO
index 8eda99d..a7bd1fb 100644
--- a/HOWTO
+++ b/HOWTO
@@ -780,6 +780,9 @@ rate_iops_min=int If fio doesn't meet this rate of IO, it will cause
the job to exit. The same format as rate is used for read vs
write seperation.
+max_latency=int If set, fio will exit the job if it exceeds this maximum
+ latency. It will exit with an ETIME error.
+
ratecycle=int Average bandwidth for 'rate' and 'ratemin' over this number
of milliseconds.
diff --git a/backend.c b/backend.c
index 1c944d6..fd73eda 100644
--- a/backend.c
+++ b/backend.c
@@ -591,7 +591,7 @@ static void do_io(struct thread_data *td)
int ret2, full;
enum fio_ddir ddir;
- if (td->terminate)
+ if (td->terminate || td->done)
break;
update_tv_cache(td);
diff --git a/engines/rdma.c b/engines/rdma.c
index 79d72d2..9b18301 100644
--- a/engines/rdma.c
+++ b/engines/rdma.c
@@ -7,14 +7,25 @@
*
* This I/O engine is disabled by default. To enable it, execute:
*
- * $ export EXTFLAGS="-DFIO_HAVE_RDMA"
- * $ export EXTLIBS="-libverbs -lrdmacm"
+ * $ export EXTFLAGS+=" -DFIO_HAVE_RDMA "
+ * $ export EXTLIBS+=" -libverbs -lrdmacm "
*
* before running make. You will need the Linux RDMA software as well, 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>
@@ -35,13 +46,14 @@
#include <inttypes.h>
#include "../fio.h"
+#include "../hash.h"
#ifdef FIO_HAVE_RDMA
#include <rdma/rdma_cma.h>
#include <infiniband/arch.h>
-#define FIO_RDMA_MAX_IO_DEPTH 128
+#define FIO_RDMA_MAX_IO_DEPTH 512
enum rdma_io_mode {
FIO_RDMA_UNKNOWN = 0,
@@ -108,6 +120,8 @@ struct rdmaio_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)
@@ -311,6 +325,7 @@ static int fio_rdmaio_setup_qp(struct thread_data *td)
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;
@@ -402,7 +417,7 @@ static int fio_rdmaio_setup_control_msg_buffers(struct thread_data *td)
/* 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;
@@ -410,7 +425,7 @@ static int fio_rdmaio_setup_control_msg_buffers(struct thread_data *td)
/* 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;
@@ -427,13 +442,12 @@ static int get_next_channel_event(struct thread_data *td,
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;
}
@@ -507,9 +521,9 @@ static struct io_u *fio_rdmaio_event(struct thread_data *td, int event)
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");
@@ -521,14 +535,11 @@ static int fio_rdmaio_getevents(struct thread_data *td, unsigned int min,
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;
switch (rd->rdma_protocol) {
case FIO_RDMA_MEM_WRITE:
@@ -591,7 +602,8 @@ static int fio_rdmaio_send(struct thread_data *td, struct io_u **io_us,
enum ibv_wc_opcode comp_opcode;
comp_opcode = IBV_WC_RDMA_WRITE;
#endif
- int i, index;
+ int i;
+ long index;
struct rdma_io_u_data *r_io_u_d;
r_io_u_d = NULL;
@@ -602,7 +614,7 @@ static int fio_rdmaio_send(struct thread_data *td, struct io_u **io_us,
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 = \
@@ -612,7 +624,7 @@ static int fio_rdmaio_send(struct thread_data *td, struct io_u **io_us,
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 = \
@@ -734,11 +746,11 @@ static int fio_rdmaio_commit(struct thread_data *td)
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) {
@@ -760,7 +772,7 @@ static int fio_rdmaio_connect(struct thread_data *td, struct fio_file *f)
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;
@@ -790,6 +802,16 @@ static int fio_rdmaio_connect(struct thread_data *td, struct fio_file *f)
/* 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;
}
@@ -800,7 +822,7 @@ static int fio_rdmaio_accept(struct thread_data *td, struct fio_file *f)
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;
@@ -863,17 +885,20 @@ static int fio_rdmaio_close_file(struct thread_data *td, struct fio_file *f)
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);
@@ -893,6 +918,7 @@ static int fio_rdmaio_setup_connect(struct thread_data *td, const char *host,
{
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);
@@ -910,28 +936,28 @@ static int fio_rdmaio_setup_connect(struct thread_data *td, const char *host,
}
/* 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;
}
@@ -943,8 +969,9 @@ static int fio_rdmaio_setup_connect(struct thread_data *td, const char *host,
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;
}
@@ -996,10 +1023,12 @@ static int fio_rdmaio_setup_listen(struct thread_data *td, short port)
static int fio_rdmaio_init(struct thread_data *td)
{
struct rdmaio_data *rd = td->io_ops->data;
+ struct flist_head *entry;
+ unsigned int max_bs;
unsigned int port;
char host[64], buf[128];
char *sep, *portp, *modep;
- int ret;
+ int ret, i = 0;
struct rlimit rl;
if (td_rw(td)) {
@@ -1119,11 +1148,8 @@ static int fio_rdmaio_init(struct thread_data *td)
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);
@@ -1145,8 +1171,9 @@ static int fio_rdmaio_init(struct thread_data *td)
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); */
+#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
i++;
}
@@ -1162,16 +1189,8 @@ static void fio_rdmaio_cleanup(struct thread_data *td)
{
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)
@@ -1179,9 +1198,10 @@ 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;
}
@@ -1229,8 +1249,8 @@ static int fio_rdmaio_init(struct thread_data fio_unused * td)
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(" $ export EXTFLAGS+=\" -DFIO_HAVE_RDMA \"\n");
+ log_err(" $ export EXTLIBS+=\" -libverbs -lrdmacm \"\n");
log_err(" $ make clean && make\n");
return 1;
}
diff --git a/fio.1 b/fio.1
index bf65551..08d6c0f 100644
--- a/fio.1
+++ b/fio.1
@@ -639,6 +639,10 @@ is used for read vs write seperation.
Average bandwidth for \fBrate\fR and \fBratemin\fR over this number of
milliseconds. Default: 1000ms.
.TP
+.BI max_latency \fR=\fPint
+If set, fio will exit the job if it exceeds this maximum latency. It will exit
+with an ETIME error.
+.TP
.BI cpumask \fR=\fPint
Set CPU affinity for this job. \fIint\fR is a bitmask of allowed CPUs the job
may run on. See \fBsched_setaffinity\fR\|(2).
diff --git a/fio.h b/fio.h
index 03e4da1..139b938 100644
--- a/fio.h
+++ b/fio.h
@@ -198,6 +198,8 @@ struct thread_options {
enum fio_memtype mem_type;
unsigned int mem_align;
+ unsigned int max_latency;
+
unsigned int stonewall;
unsigned int new_group;
unsigned int numjobs;
diff --git a/io_u.c b/io_u.c
index e047677..b049b61 100644
--- a/io_u.c
+++ b/io_u.c
@@ -1325,6 +1325,13 @@ static void account_io_completion(struct thread_data *td, struct io_u *io_u,
tusec = utime_since(&io_u->start_time, &icd->time);
add_lat_sample(td, idx, tusec, bytes);
+
+ if (td->o.max_latency && tusec > td->o.max_latency) {
+ if (!td->error)
+ log_err("fio: latency of %lu usec exceeds specified max (%u usec)\n", tusec, td->o.max_latency);
+ td_verror(td, ETIME, "max latency exceeded");
+ icd->error = ETIME;
+ }
}
if (!td->o.disable_clat) {
diff --git a/options.c b/options.c
index e0b7fec..380df36 100644
--- a/options.c
+++ b/options.c
@@ -2099,6 +2099,12 @@ static struct fio_option options[FIO_MAX_OPTS] = {
.parent = "rate",
},
{
+ .name = "max_latency",
+ .type = FIO_OPT_INT,
+ .off1 = td_var_offset(max_latency),
+ .help = "Maximum tolerated IO latency (usec)",
+ },
+ {
.name = "invalidate",
.type = FIO_OPT_BOOL,
.off1 = td_var_offset(invalidate_cache),
diff --git a/os/os-linux.h b/os/os-linux.h
index 2b35f34..9b7ff29 100644
--- a/os/os-linux.h
+++ b/os/os-linux.h
@@ -14,7 +14,6 @@
#include <linux/unistd.h>
#include <linux/raw.h>
#include <linux/major.h>
-#include <linux/version.h>
#include <endian.h>
#include "indirect.h"
@@ -63,10 +62,6 @@
#define FIO_HAVE_FALLOC_ENG
#endif
-#if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,26)
-#define FIO_HAVE_RUSAGE_THREAD
-#endif
-
#ifdef SYNC_FILE_RANGE_WAIT_BEFORE
#define FIO_HAVE_SYNC_FILE_RANGE
#endif
diff --git a/stat.c b/stat.c
index af6e1f2..522901a 100644
--- a/stat.c
+++ b/stat.c
@@ -16,7 +16,7 @@ void update_rusage_stat(struct thread_data *td)
{
struct thread_stat *ts = &td->ts;
-#ifdef FIO_HAVE_RUSAGE_THREAD
+#ifdef RUSAGE_THREAD
getrusage(RUSAGE_THREAD, &td->ru_end);
#else
getrusage(RUSAGE_SELF, &td->ru_end);
--
To unsubscribe from this list: send the line "unsubscribe fio" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
