#This patch was developed both in-house and from outside. We plan to submit it
#upstream, but do not yet have a target date for doing so
#
# HG changeset patch
# Parent 1afa90e87b4b0fcc3ccef29f1029fa7d1628f389
diff -r 1afa90e87b4b Makefile.in
--- a/Makefile.in Tue Feb 23 22:30:57 2016 -0800
+++ b/Makefile.in Tue Feb 23 22:30:58 2016 -0800
@@ -4,18 +4,22 @@
mandir = $(DESTDIR)@mandir@
incdir = $(DESTDIR)@includedir@
+CC=gcc
+
all: all-programs
-CFLAGS = -O2 -Wall -Iinclude
-CPPFLAGS = -DDEBUG_EXE -DRDS_VERSION=\"@VERSION@\" -MD -MP -MF $(@D)/.$(basename $(@F)).d
+CFLAGS += -O2 -Wall -Iinclude
+CPPFLAGS += -D_XOPEN_SOURCE=600 -D__EXTENSIONS__ \
+ -DDEBUG_EXE -DRDS_VERSION=\"@VERSION@\" -MD -MP -MF $(@D)/.$(basename $(@F)).d
+LDFLAGS += -libverbs -lsocket -lnsl -llgrp
-COMMON_SOURCES = pfhack.c
+COMMON_SOURCES = pfhack.c rds-vendor.c
SOURCES = $(addsuffix .c,$(PROGRAMS)) $(COMMON_SOURCES)
CLEAN_OBJECTS = $(addsuffix .o,$(PROGRAMS)) $(subst .c,.o,$(COMMON_SOURCES))
# This is the default
-DYNAMIC_PF_RDS = true
+#DYNAMIC_PF_RDS = true
ifneq ($(DYNAMIC_PF_RDS),)
CPPFLAGS += -DDYNAMIC_PF_RDS
@@ -29,14 +33,14 @@
all-programs: $(PROGRAMS)
install: $(PROGRAMS)
- install -d $(bindir)
- install -m 555 -s $(PROGRAMS) $(bindir)
- install -d $(mandir)/man1
- install -d $(mandir)/man7
- install -m 644 *.1 $(mandir)/man1
- install -m 644 *.7 $(mandir)/man7
- install -d $(incdir)/net
- install -m 444 include/rds.h $(incdir)/net
+ $(INSTALL) -d $(bindir)
+ $(INSTALL) -m 755 -s $(PROGRAMS) $(bindir)
+ $(INSTALL) -d $(mandir)/man1
+ $(INSTALL) -d $(mandir)/man7
+ $(INSTALL) -m 644 *.1 $(mandir)/man1
+ $(INSTALL) -m 644 *.7 $(mandir)/man7
+ $(INSTALL) -d $(incdir)/net
+ $(INSTALL) -m 444 include/rds.h $(incdir)/net
clean:
rm -f $(PROGRAMS) $(CLEAN_OBJECTS)
@@ -47,7 +51,7 @@
$(PROGRAMS) : % : %.o $(COMMON_OBJECTS)
- gcc $(CFLAGS) $(LDFLAGS) -o $@ $^
+ $(CC) $(CFLAGS) $(CPPFLAGS) $(LDFLAGS) -o $@ $^
LOCAL_DFILES := $(wildcard .*.d)
ifneq ($(LOCAL_DFILES),)
@@ -72,8 +76,6 @@
configure \
README \
- stap/rds.stp \
- stap/README \
diff -r 1afa90e87b4b configure
--- a/configure Tue Feb 23 22:30:57 2016 -0800
+++ b/configure Tue Feb 23 22:30:58 2016 -0800
@@ -1215,7 +1215,7 @@
-VERSION=2.0.4
+VERSION=2.0.7
RELEASE=1
diff -r 1afa90e87b4b configure.in
--- a/configure.in Tue Feb 23 22:30:57 2016 -0800
+++ b/configure.in Tue Feb 23 22:30:58 2016 -0800
@@ -1,7 +1,7 @@
AC_PREREQ(2.55)
AC_INIT()
-VERSION=2.0.4
+VERSION=2.0.7
RELEASE=1
AC_SUBST(VERSION)
diff -r 1afa90e87b4b examples/Makefile
--- a/examples/Makefile Tue Feb 23 22:30:57 2016 -0800
+++ b/examples/Makefile Tue Feb 23 22:30:58 2016 -0800
@@ -1,6 +1,12 @@
+CC=gcc
+LIBS = -lsocket -lnsl
+CPPFLAGS += -D_XOPEN_SOURCE=600 -D__EXTENSIONS__
all: rds-sample
-rds-sample: rds-sample.o
+rds-sample: rds-sample.c
+ $(CC) $(CPPFLAGS) $(CFLAGS) -o rds-sample rds-sample.c $(LIBS)
-CFLAGS = -I ../include
+clean:
+ rm -rf rds-sample
+
diff -r 1afa90e87b4b examples/rds-sample.c
--- a/examples/rds-sample.c Tue Feb 23 22:30:57 2016 -0800
+++ b/examples/rds-sample.c Tue Feb 23 22:30:58 2016 -0800
@@ -26,6 +26,7 @@
#include <string.h>
#include <stdlib.h>
+#if !(defined(__SVR4) && defined(__sun))
/* FIXME - this is a hack to getaround RDS not exporting any header files.
* This is a local copy of the file found at net/rds/
*/
@@ -33,6 +34,9 @@
/* These are defined in rds.h....but that file is not happily included */
#define SOL_RDS 272
#define PF_RDS 28
+#else
+#include <sys/rds.h>
+#endif
#define TESTPORT 4000
@@ -107,12 +111,12 @@
cmsg->cmsg_type = RDS_CMSG_RDMA_ARGS;
cmsg->cmsg_len = CMSG_LEN(sizeof(struct rds_rdma_args));
- iov.addr = (uint64_t) buf;
+ iov.addr = (uint64_t)(uintptr_t)buf;
iov.bytes = sizeof(struct rdss_message);
args->remote_vec.addr = 0;
args->remote_vec.bytes = sizeof(struct rdss_message);
- args->local_vec_addr = (uint64_t) &iov;
+ args->local_vec_addr = (uint64_t)(uintptr_t)&iov;
args->nr_local = 1;
args->flags = remote_flags ? (RDS_RDMA_READWRITE | RDS_RDMA_FENCE) : 0;
args->flags |= RDS_RDMA_NOTIFY_ME;
@@ -244,9 +248,9 @@
void *ctlbuf;
struct iovec *iov;
- mr_args.vec.addr = (uint64_t) buf;
+ mr_args.vec.addr = (uint64_t)(uintptr_t)buf;
mr_args.vec.bytes = sizeof(struct rdss_message);
- mr_args.cookie_addr = (uint64_t) cookie;
+ mr_args.cookie_addr = (uint64_t)(uintptr_t)cookie;
mr_args.flags = RDS_RDMA_USE_ONCE;
ctlbuf = calloc(1, CMSG_SPACE(sizeof(mr_args)));
diff -r 1afa90e87b4b include/rds.h
--- a/include/rds.h Tue Feb 23 22:30:57 2016 -0800
+++ b/include/rds.h Tue Feb 23 22:30:58 2016 -0800
@@ -84,6 +84,8 @@
#define RDS_CMSG_CONG_UPDATE 5
#define RDS_CMSG_ATOMIC_FADD 6
#define RDS_CMSG_ATOMIC_CSWP 7
+#define RDS_CMSG_MASKED_ATOMIC_FADD 8
+#define RDS_CMSG_MASKED_ATOMIC_CSWP 9
#define RDS_INFO_FIRST 10000
#define RDS_INFO_COUNTERS 10000
@@ -252,8 +254,25 @@
rds_rdma_cookie_t cookie;
uint64_t local_addr;
uint64_t remote_addr;
- uint64_t swap_add;
- uint64_t compare;
+ union {
+ struct {
+ uint64_t compare;
+ uint64_t swap;
+ } cswp;
+ struct {
+ uint64_t add;
+ } fadd;
+ struct {
+ uint64_t compare;
+ uint64_t swap;
+ uint64_t compare_mask;
+ uint64_t swap_mask;
+ } m_cswp;
+ struct {
+ uint64_t add;
+ uint64_t nocarry_mask;
+ } m_fadd;
+ };
u_int64_t flags;
u_int64_t user_token;
};
@@ -278,5 +297,6 @@
#define RDS_RDMA_USE_ONCE 0x0008 /* free MR after use */
#define RDS_RDMA_DONTWAIT 0x0010 /* Don't wait in SET_BARRIER */
#define RDS_RDMA_NOTIFY_ME 0x0020 /* Notify when operation completes */
+#define RDS_RDMA_SILENT 0x0040 /* Do not interrupt remote */
#endif /* IB_RDS_H */
diff -r 1afa90e87b4b pfhack.h
--- a/pfhack.h Tue Feb 23 22:30:57 2016 -0800
+++ b/pfhack.h Tue Feb 23 22:30:58 2016 -0800
@@ -44,9 +44,11 @@
#ifndef __PF_HACK_H
#define __PF_HACK_H
+#if !((defined(__SVR4) && defined(__sun)))
#define PF_RDS 21
#define AF_RDS 21
#define SOL_RDS 276
+#endif
extern int discover_pf_rds();
extern int discover_sol_rds();
diff -r 1afa90e87b4b rds-info.1
--- a/rds-info.1 Tue Feb 23 22:30:57 2016 -0800
+++ b/rds-info.1 Tue Feb 23 22:30:58 2016 -0800
@@ -1,162 +1,160 @@
-.Dd October 30, 2006
-.Dt RDS-INFO 1
-.Sh NAME
-.Nm rds-info
-.Nd display information from the RDS kernel module
-.Sh SYNOPSIS
-.Nm rds-info
-.Op Fl v
-.Bk -words
-.Op Fl cknrstIT
+.TH RDS-INFO 1 "October 30, 2006"
+.SH "NAME"
+rds-info - display information from the RDS kernel module
+.SH SYNOPSIS
+.B rds-info [-cknrstIT]
-.Sh DESCRIPTION
-The
-utility presents various sources of information that
+.SH DESCRIPTION
+.PP
+The utility presents various sources of information that
the RDS kernel module maintains. When run without any optional arguments
will output all the information it knows of. When options are specified then
only the information associated with those options is displayed.
The options are as follows:
-.Bl -tag -width Ds
-.It Fl v
+.SH OPTIONS
+.PP
+.TP 7
+\fB\-v
Requests verbose output. When this option is given, some classes of information
will display additional data.
-.It Fl c
+.TP
+\fB\-c
Display global counters. Each counter increments as its event
occurs. The counters may not be reset. The set of supported counters
may change over time.
-.Bl -tag -width 4
-.It CounterName
+.IP CounterName
The name of the counter. These names come from the kernel and can change
depending on the capability of the kernel module.
-.It Value
+.IP Value
The number of times that the counter has been incremented since the kernel
module was loaded.
-.It Fl k
+.TP
+\fB\-k\fR
Display all the RDS sockets in the system. There will always be one socket
listed that is neither bound to nor connected to any addresses because
itself uses an unbound socket to collect information.
-.Bl -tag -width 4
-.It BoundAddr, BPort
+.IP BoundAddr, BPort
The IP address and port that the socket is bound to. 0.0.0.0 0 indicates that
the socket has not been bound.
-.It ConnAddr, CPort
+.IP ConnAddr, CPort
The IP address and port that the socket is connected to. 0.0.0.0 0 indicates
that the socket has not been connected.
-.It SndBuf, RcvBuf
+.IP SndBuf, RcvBuf
The number of bytes of message payload which can be queued for sending or
receiving on the socket, respectively.
-.It Inode
+.IP Inode
The number of the inode object associated with the socket. Can be used to
locate the process owning a given socket by searching /proc/*/fd for
open files referencing a socket with this inode number.
-.It Fl n
+.TP
+\fB\-n\fR
Display all RDS connections. RDS connections are maintained between
nodes by transports.
-.Bl -tag -width 4
-.It LocalAddr
+.IP LocalAddr
The IP address of this node. For connections that originate and terminate on
the same node the local address indicates which address initiated the
connection establishment.
-.It RemoteAddr
+.IP RemoteAddr
The IP address of the remote end of the connection.
-.It NextTX
+.IP Tos
+The type of service value for this connection.
+.IP NextTX
The sequence number that will be given to the next message that is sent
over the connection.
-.It NextRX
+.IP NextRX
The sequence number that is expected from the next message to arrive over
the connection. Any incoming messages with sequence numbers less than this
will be dropped.
-.It Flg
+.IP Flg
Flags which indicate the state of the connection.
-.Bl -tag -width 4
-.It s
-A process is currently sending a message down the connection.
-.It c
-The transport is attempting to connect to the remote address.
-.It C
-The connection to the remote host is connected and active.
-.It Fl r, Fl s, Fl t
+.IP s
+ A process is currently sending a message down
+ the connection.
+.IP c
+ The transport is attempting to connect to the
+ remote address.
+.IP C
+ The connection to the remote host is connected
+ and active.
+.IP E
+ The connection to the remote host is in error.
+
+.TP
+\fB\-r\fR, \fB\-s\fR, \fB\-t\fR
Display the messages in the receive, send, or retransmit queues respectively.
-.Bl -tag -width 4
-.It LocalAddr, LPort
+
+.IP LocalAddr, LPort
The local IP address and port on this node associated with the message. For
sent messages this is the source address, for receive messages it is the
destination address.
-.It RemoteAddr, RPort
+.IP RemoteAddr, RPort
The remote IP address and port associated with the message. For sent messages
this is the destination address, for receive messages it is the source address.
-.It Seq
+.IP Tos
+The type of service for this message.
+.IP Seq
The sequence number of the message.
-.It Bytes
+.IP Bytes
The number of bytes in the message payload.
+
+.PP
The following information sources are dependent on specific transports which
may not always be available.
-.It Fl I
+.TP 7
+\fB\-I\fR
Display the IB connections which the IB transport is using to provide
RDS connections.
-.Bl -tag -width 4
-.It LocalAddr
+.IP LocalAddr
The local IP address of this connection.
-.It RemoteAddr
+.IP RemoteAddr
The remote IP address of this connection.
-.It LocalDev
+.IP Tos
+The type of service value for this connection.
+.IP SL
+The QoS Service Level for this connection.
+.IP LocalDev
The local IB Global Identifier, printed in IPv6 address syntax.
-.It RemoteDev
+.IP RemoteDev
The remote IB Global Identifier, printed in IPv6 address syntax.
If verbose output is requested, per-connection settings such as the
maximum number of send and receive work requests will be displayed
in addition.
-.It Fl T
+.TP 7
+\fB\-T\fR
Display the TCP sockets which the TCP transport is using to provide
RDS connections.
-.Bl -tag -width 4
-.It LocalAddr, LPort
+.IP LocalAddr, LPort
The local IP address and port of this socket.
-.It RemoteAddr, RPort
+.IP RemoteAddr, RPort
The remote IP address and port that this socket is connected to.
-.It HdrRemain
+.IP HdrRemain
The number of bytes that must be read off the socket to complete the next
full RDS header.
-.It DataRemain
+.IP DataRemain
The number of bytes that must be read off the socket to complete the data
payload of the message which is being received.
-.It SentNxt
+.IP SentNxt
The TCP sequence number of the first byte of the last message that we sent
down the connection.
-.It ExpectedUna
+.IP ExpectedUna
The TCP sequence number of the byte past the last byte of the last message
that we sent down the connection. When we see that the remote side has
acked up to this byte then we know that the remote side has received all
our RDS messages.
-.It SeenUna
+.IP SeenUna
The TCP sequence number of the byte past the last byte which has been
acked by the remote host.
-
diff -r 1afa90e87b4b rds-info.c
--- a/rds-info.c Tue Feb 23 22:30:57 2016 -0800
+++ b/rds-info.c Tue Feb 23 22:30:58 2016 -0800
@@ -42,16 +42,28 @@
#include <sys/types.h>
#include <sys/socket.h>
#include <errno.h>
+#if defined(__SVR4) && defined(__sun)
+#include <strings.h>
+#else
#include <string.h>
+#endif
#include <inttypes.h>
#include <netinet/in.h>
#include <arpa/inet.h>
+#if defined(__SVR4) && defined(__sun)
+#include <sys/rds.h>
+#include <infiniband/ofa_solaris.h>
+#else
#include "rds.h"
+#endif
#include "pfhack.h"
+/* WHUPS changed the struct rds_info_connection definition b/w rds in 1.4 & 1.5. gotta support both
+ for now. TODO remove check of transport[15] once ofed pre-1.5 is extinct. */
#define rds_conn_flag(conn, flag, letter) \
- (conn.flags & RDS_INFO_CONNECTION_FLAG_##flag ? letter : '-')
+ (conn.flags & RDS_INFO_CONNECTION_FLAG_##flag \
+ || conn.transport[15] & RDS_INFO_CONNECTION_FLAG_##flag ? letter : '-')
#define min(a, b) (a < b ? a : b)
#define array_size(foo) (sizeof(foo) / sizeof(foo[0]))
@@ -76,6 +88,10 @@
char *progname = "rds-info";
+#if defined(__SVR4) && defined(__sun)
+int sol_ioctl(int, int, struct rds_info_arg *, socklen_t *, int *);
+#endif
+
/* Like inet_ntoa, but can be re-entered several times without clobbering
* the previously returned string. */
static const char *paddr(int af, const void *addrp)
@@ -134,18 +150,20 @@
{
struct rds_info_connection conn;
- printf("\nRDS Connections:\n%15s %15s %16s %16s %3s\n",
- "LocalAddr", "RemoteAddr", "NextTX", "NextRX", "Flg");
+ printf("\nRDS Connections:\n%15s %15s %4s %16s %16s %4s\n",
+ "LocalAddr", "RemoteAddr", "Tos", "NextTX", "NextRX", "Flgs");
for_each(conn, data, each, len) {
- printf("%15s %15s %16"PRIu64" %16"PRIu64" %c%c%c\n",
+ printf("%15s %15s %4u %16"PRIu64" %16"PRIu64" %c%c%c%c\n",
ipv4addr(conn.laddr),
ipv4addr(conn.faddr),
+ conn.tos,
rds_conn_flag(conn, SENDING, 's'),
rds_conn_flag(conn, CONNECTING, 'c'),
- rds_conn_flag(conn, CONNECTED, 'C'));
+ rds_conn_flag(conn, CONNECTED, 'C'),
+ rds_conn_flag(conn, ERROR, 'E'));
}
}
@@ -153,16 +171,17 @@
{
struct rds_info_message msg;
- printf("\n%s Message Queue:\n%15s %5s %15s %5s %16s %10s\n",
+ printf("\n%s Message Queue:\n%15s %5s %15s %5s %4s %16s %10s\n",
(char *)extra,
- "LocalAddr", "LPort", "RemoteAddr", "RPort", "Seq", "Bytes");
+ "LocalAddr", "LPort", "RemoteAddr", "RPort", "Tos","Seq", "Bytes");
for_each(msg, data, each, len) {
- printf("%15s %5u %15s %5u %16"PRIu64" %10u\n",
+ printf("%15s %5u %15s %5u %4u %16"PRIu64" %10u\n",
ipv4addr(msg.laddr),
ntohs(msg.lport),
ipv4addr(msg.faddr),
ntohs(msg.fport),
+ msg.tos,
}
}
@@ -191,13 +210,14 @@
{
struct rds_info_rdma_connection ic;
- printf("\nRDS IB Connections:\n%15s %15s %32s %32s\n",
- "LocalAddr", "RemoteAddr", "LocalDev", "RemoteDev");
+ printf("\nRDS IB Connections:\n%15s %15s %4s %3s %32s %32s\n",
+ "LocalAddr", "RemoteAddr", "Tos", "SL", "LocalDev", "RemoteDev");
for_each(ic, data, each, len) {
- printf("%15s %15s %32s %32s",
+ printf("%15s %15s %4u %3u %32s %32s",
ipv4addr(ic.src_addr),
ipv4addr(ic.dst_addr),
ipv6addr(ic.src_gid),
ipv6addr(ic.dst_gid));
@@ -234,8 +254,10 @@
print_msgs, "Send", 0 },
['t'] = { RDS_INFO_RETRANS_MESSAGES, "retransmit queue messages",
print_msgs, "Retransmit", 0 },
+#if !(defined(__SVR4) && defined(__sun))
['T'] = { RDS_INFO_TCP_SOCKETS, "TCP transport sockets",
print_tcp_socks, NULL, 0 },
+#endif
['I'] = { RDS_INFO_IB_CONNECTIONS, "IB transport connections",
print_ib_conns, NULL, 0 },
};
@@ -266,6 +288,9 @@
char optstring[258] = "v+";
int given_options = 0;
socklen_t len = 0;
+#if defined(__SVR4) && defined(__sun)
+ struct rds_info_arg arg;
+#endif
void *data = NULL;
int fd;
int each;
@@ -322,6 +347,7 @@
(given_options && !infos[i].option_given))
continue;
+#if !(defined(__SVR4) && defined(__sun))
/* read in the info until we get a full snapshot */
while ((each = getsockopt(fd, sol, infos[i].opt_val, data,
&len)) < 0) {
@@ -345,15 +371,47 @@
return 1;
}
}
+#else
+ int retcode;
+
+ retcode = sol_ioctl(fd, infos[i].opt_val, &arg, &len, &each);
+ if (retcode != 0) {
+ if (retcode == 1) {
+ fprintf(stderr, "%s: Unable get statistics: "
+ "%s\n", progname, strerror(errno));
+ return 1;
+ } else if (retcode == 2) {
+ fprintf(stderr, "%s: Unable to allocate memory"
+ " for %u bytes of info: %s\n",
+ progname, len, strerror(errno));
+ return 1;
+ } else {
+ fprintf(stderr, "%s: Unable to get statistics:"
+ " %s\n", progname, strerror(errno));
+ return 1;
+ }
+ }
+
+ /* No data at the driver */
+ if (len == 0)
+ invalid_opt = 1;;
+#endif
if (invalid_opt)
continue;
+#if !(defined(__SVR4) && defined(__sun))
infos[i].print(data, each, len, infos[i].extra);
-
+#else
+ infos[i].print((void *)(uintptr_t)arg.datap, each, len,
+ infos[i].extra);
+#endif
if (given_options && --given_options == 0)
break;
}
+#if defined(__SVR4) && defined(__sun)
+ free((void *)(uintptr_t)arg.datap);
+#endif
return 0;
}
diff -r 1afa90e87b4b rds-ping.1
--- a/rds-ping.1 Tue Feb 23 22:30:57 2016 -0800
+++ b/rds-ping.1 Tue Feb 23 22:30:58 2016 -0800
@@ -1,69 +1,63 @@
-.Dd Apr 22, 2008
-.Dt RDS-PING 1
-.Sh NAME
-.Nm rds-ping
-.Nd test reachability of remote node over RDS
-.Sh SYNOPSIS
-.Nm rds-ping
-.Bk -words
-.Op Fl c Ar count
-.Op Fl i Ar interval
-.Op Fl I Ar local_addr
-.Ar remote_addr
+.TH RDS-PING 1 "Apr 22, 2008"
+.SH NAME
+rds-ping - test reachability of remote node over RDS
-.Sh DESCRIPTION
-.Nm rds-ping
-is used to test whether a remote node is reachable over RDS.
-Its interface is designed to operate pretty much the standard
-.Xr ping 8
+.SH SYNOPSIS
+.HP
+.nf
+rds-ping [-c count] [-Q tos] [-i interval] [-I local_addr]
+ remote_addr
+.fi
+
+.SH DESCRIPTION
+.PP
+rds-ping is used to test whether a remote node is reachable over RDS.
+Its interface is designed to operate pretty much the standard ping(8)
utility, even though the way it works is pretty different.
-.Nm rds-ping
-opens several RDS sockets and sends packets to port 0 on
+.PP
+rds-ping opens several RDS sockets and sends packets to port 0 on
the indicated host. This is a special port number to which
no socket is bound; instead, the kernel processes incoming
packets and responds to them.
-.Sh OPTIONS
+.SH OPTIONS
The following options are available for use on the command line:
-.Bl -tag -width Ds
-.It Fl c Ar count
-Causes
-.Nm rds-ping
-to exit after sending (and receiving) the specified number of
+.PP
+.TP 7
+\fB\-c count
+Causes rds-ping to exit after sending (and receiving) the specified number of
packets.
-.It Fl I Ar address
-By default,
-.Nm rds-ping
-will pick the local source address for the RDS socket based
+.TP
+\fB\-Q tos
+By default, rds-ping sends the ping requests on base (tos = 0) RDS connection.
+With this option, the requests are sent on RDS connection with the specified tos
+value. Valid values are 0-255.
+.TP
+\fB\-I address
+By default, rds-ping will pick the local source address for the RDS socket based
on routing information for the destination address (i.e. if
packets to the given destination would be routed through interface
-.Nm ib0 ,
+ib0 ,
then it will use the IP address of
-.Nm ib0
+ib0
as source address).
Using the
.Fl I
option, you can override this choice.
-.It Fl i Ar timeout
-By default,
-.Nm rds-ping
-will wait for one second between sending packets. Use this option
+.TP
+\fB\-i timeout
+By default, rds-ping will wait for one second between sending packets. Use this option
to specified a different interval. The timeout value is given in
seconds, and can be a floating point number. Optionally, append
-.Nm msec
+msec
or
-.Nm usec
+usec
to specify a timeout in milliseconds or microseconds, respectively.
+.IP
Specifying a timeout considerably smaller than the packet round-trip
time will produce unexpected results.
-.Sh AUTHORS
-.Nm rds-ping
+
+.SH AUTHORS
+rds-ping
was written by Olaf Kirch <olaf.kirch@oracle.com>.
-.Sh SEE ALSO
-.Xr rds 7 ,
-.Xr rds-info 1 ,
-.Xr rds-stress 1 .
+.S~ SEE ALSO
+rds 7, rds-info 1, rds-stress 1.
diff -r 1afa90e87b4b rds-ping.c
--- a/rds-ping.c Tue Feb 23 22:30:57 2016 -0800
+++ b/rds-ping.c Tue Feb 23 22:30:58 2016 -0800
@@ -48,7 +48,12 @@
#include <sys/poll.h>
#include <fcntl.h>
#include <getopt.h>
+#include <sys/ioctl.h>
+#if defined(__SVR4) && defined(__sun)
+#include <sys/rds.h>
+#else
#include "rds.h"
+#endif
#include "pfhack.h"
@@ -67,6 +72,7 @@
static unsigned long opt_count;
static struct in_addr opt_srcaddr;
static struct in_addr opt_dstaddr;
+static uint8_t opt_tos = 0;
/* For reasons of simplicity, RDS ping does not use a packet
* payload that is being echoed, the way ICMP does.
@@ -91,13 +97,14 @@
static int parse_timeval(const char *, struct timeval *);
static int parse_long(const char *ptr, unsigned long *);
static int parse_addr(const char *ptr, struct in_addr *);
+static unsigned long long parse_ull(char *ptr, unsigned long long max);
int
main(int argc, char **argv)
{
int c;
- while ((c = getopt(argc, argv, "c:i:I:")) != -1) {
+ while ((c = getopt(argc, argv, "c:i:I:Q:")) != -1) {
switch (c) {
case 'c':
if (!parse_long(optarg, &opt_count))
@@ -114,6 +121,9 @@
die("Bad wait time <%s>\n", optarg);
break;
+ case 'Q':
+ opt_tos = parse_ull(optarg, 255);
+ break;
default:
usage("Unknown option");
}
@@ -142,6 +152,9 @@
struct timeval next_ts;
struct socket socket[NSOCKETS];
struct pollfd pfd[NSOCKETS];
+#if !(defined(__SVR4) && defined(__sun))
+ int pending[NSOCKETS];
+#endif
int i, next = 0;
for (i = 0; i < NSOCKETS; ++i) {
@@ -152,10 +165,18 @@
socket[i].fd = fd;
pfd[i].fd = fd;
pfd[i].events = POLLIN;
+#if !(defined(__SVR4) && defined(__sun))
+ pending[i] = 0;
+#endif
}
memset(&sin, 0, sizeof(sin));
+#if defined(__SVR4) && defined(__sun)
+ sin.sin_family = AF_INET_OFFLOAD;
+#else
sin.sin_family = AF_INET;
+#endif
+
sin.sin_addr = opt_dstaddr;
gettimeofday(&next_ts, NULL);
@@ -180,13 +201,31 @@
if (opt_count && sent >= opt_count)
break;
- timeradd(&next_ts, &opt_wait, &next_ts);
- if (sendto(sp->fd, NULL, 0, 0, (struct sockaddr *) &sin, sizeof(sin)))
- err = errno;
- sp->sent_id = ++sent;
- sp->sent_ts = now;
- sp->nreplies = 0;
- next = (next + 1) % NSOCKETS;
+ timeradd(&now, &opt_wait, &next_ts);
+#if !(defined(__SVR4) && defined(__sun))
+ if (!pending[next]) {
+#endif
+ memset(&sin, 0, sizeof(sin));
+#if defined(__SVR4) && defined(__sun)
+ sin.sin_family = AF_INET_OFFLOAD;
+#else
+ sin.sin_family = AF_INET;
+#endif
+ sin.sin_addr = opt_dstaddr;
+
+ if (sendto(sp->fd, NULL, 0, 0, (struct sockaddr *) &sin, sizeof(sin)) < 0)
+ err = errno;
+ sp->sent_id = ++sent;
+ sp->sent_ts = now;
+ sp->nreplies = 0;
+#if !(defined(__SVR4) && defined(__sun))
+ if (!err)
+ pending[next] = 1;
+#endif
+ next = (next + 1) % NSOCKETS;
+#if !(defined(__SVR4) && defined(__sun))
+ }
+#endif
if (err) {
static unsigned int nerrs = 0;
@@ -223,6 +262,9 @@
report_packet(sp, &now, NULL, errno);
} else {
report_packet(sp, &now, &from.sin_addr, 0);
+#if !(defined(__SVR4) && defined(__sun))
+ pending[i] = 0;
+#endif
recv++;
}
}
@@ -258,7 +300,11 @@
int pf;
memset(&sin, 0, sizeof(sin));
+#if defined(__SVR4) && defined(__sun)
+ sin.sin_family = AF_INET_OFFLOAD;
+#else
sin.sin_family = AF_INET;
+#endif
#ifdef DYNAMIC_PF_RDS
pf = discover_pf_rds();
@@ -278,6 +324,9 @@
if (ufd < 0)
die_errno("unable to create UDP socket");
sin.sin_addr = *dst;
+#if defined(__SVR4) && defined(__sun)
+ sin.sin_family = AF_INET;
+#endif
sin.sin_port = htons(1);
if (connect(ufd, (struct sockaddr *) &sin, sizeof(sin)) < 0)
die_errno("unable to connect to %s",
@@ -289,6 +338,9 @@
*src = sin.sin_addr;
close(ufd);
+#if defined(__SVR4) && defined(__sun)
+ sin.sin_family = AF_INET_OFFLOAD;
+#endif
}
sin.sin_addr = *src;
@@ -297,6 +349,9 @@
if (bind(fd, (struct sockaddr *) &sin, sizeof(sin)))
die_errno("bind() failed");
+ if (opt_tos && ioctl(fd, SIOCRDSSETTOS, &opt_tos))
+ die_errno("ERROR: failed to set TOS\n");
+
return fd;
}
@@ -309,7 +364,8 @@
"%s\nUsage: rds-ping [options] dst_addr\n"
"Options:\n"
" -c count limit packet count\n"
- " -I interface source IP address\n",
+ " -I interface source IP address\n"
+ " -Q tos type of service\n",
complaint);
exit(1);
}
@@ -384,3 +440,31 @@
return 0;
}
+static unsigned long long parse_ull(char *ptr, unsigned long long max)
+{
+ unsigned long long val;
+ char *endptr;
+
+ val = strtoull(ptr, &endptr, 0);
+ switch (*endptr) {
+ case 'k': case 'K':
+ val <<= 10;
+ endptr++;
+ break;
+
+ case 'm': case 'M':
+ val <<= 20;
+ endptr++;
+ break;
+
+ case 'g': case 'G':
+ val <<= 30;
+ endptr++;
+ break;
+ }
+
+ if (*ptr && !*endptr && val <= max)
+ return val;
+
+ die("invalid number '%s'\n", ptr);
+}
diff -r 1afa90e87b4b rds-stress.1
--- a/rds-stress.1 Tue Feb 23 22:30:57 2016 -0800
+++ b/rds-stress.1 Tue Feb 23 22:30:58 2016 -0800
@@ -1,119 +1,131 @@
-.Dd May 15, 2007
-.Dt RDS-STRESS 1
-.Sh NAME
-.Nm rds-stress
-.Nd send messages between processes over RDS sockets
-.Sh SYNOPSIS
-.Nm rds-stress
-.Bk -words
-.Op Fl p Ar port_number
-.Op Fl r Ar receive_address
-.Op Fl s Ar send_address
-.Op Fl a Ar ack_bytes
-.Op Fl q Ar request_bytes
-.Op Fl D Ar rdma_bytes
-.Op Fl d Ar queue_depth
-.Op Fl t Ar nr_tasks
-.Op Fl c
-.Op Fl R
-.Op Fl V
-.Op Fl v
+.TH RDS-STRESS 1 " May 15, 2007"
+.SH "NAME"
+rds-stress - send messages between processes over RDS sockets
+.PP
+.SH SYNOPSIS
+.HP
+.nf
+rds-stress [-p port_number] -r [receive_address] [-s send_address]
+ [-Q tos] [-a ack_bytes] [-q request_bytes] [-D rdma_bytes]
+ [-d queue_depth] [-t nr_tasks] [-T time] [-c] [-R] [-V] [-v]
+.fi
-.Sh DESCRIPTION
-.Nm rds-stress
+
+.SH DESCRIPTION
+.PP
+.Nm
+rds-stress
sends messages between groups tasks, usually running on seperate
machines.
+.PP
First a passive receiving instance is started.
-.Dl $ rds-stress
+.RS 12
+
+ $ rds-stress
+.RE
+.PP
Then an active sending instance is started, giving it
the address and port at which it will find a listening
passive receiver. In addition, it is given configuration options which
both instances will use.
-.Dl $ rds-stress -s recvhost -p 4000 -t 1 -d 1
+.PP
+.RS 12
+ $ rds-stress -s recvhost -p 4000 -t 1 -d 1
+.RE
+.PP
The active sender will parse the options, connect to the passive receiver, and
send the options over this connection. From this point on both instances
exhibit the exact same behaviour.
+.PP
They will create a number of child tasks as specified by the -t option. Once
the children are created the parent sleeps for a second at a time, printing a
summary of statistics at each interval.
+.PP
Each child will open an RDS socket, each binding to a port number in order
after the port number given on the command line. The first child would bind to
port 4001 in our example. Each child sets the send and receive buffers to
exactly fit the number of messages, requests and acks, that will be in flight
as determind by the command line arguments.
+.PP
The children then enter their loop. They will keep a number of sent messages
outstanding as specified by the -d option. When they reach this limit they
will wait to receive acks which will allow them to send again. As they receive
messages from their peers they immediately send acks.
+.PP
Every second, the parent process will display statistics of the ongoing
stress test. The output is described in section OUTPUT below.
+.PP
If the -T option is given, the test will terminate after the specified time,
and a summary is printed.
+.PP
Each child maintains outstanding messages to all other children of the other instance.
They do not send to their siblings.
-.Sh OPTIONS
+.SH OPTIONS
+.PP
The following options are available for use on the command line:
-.Bl -tag -width Ds
-.It Fl p Ar port_number
+
+.TP 7
+\fB\-p port_number
Each parent binds a TCP socket to this port number and their respective
address. They will trade the negotiated options over this socket. Each
child will bind an RDS socket to the range of ports immediately following
this port number, for as many children as there are.
-.It Fl s Ar send_address
+.TP
+\fB\-s send_address
A connection attempt is made to this address. Once its complete and the
options are sent over it then children will be created and work will proceed.
-.It Fl r Ar receive_address
+.TP
+\fB\-r receive_address
This specifies the address that messages will be sent from. If -s is not
specified then rds-stress waits for a connection on this address before
proceeding.
+
If this option is not given, rds-stress will choose an appropriate address.
The passive process will accept connections on all local interfaces, and
obtain the address once the control connection is established.
The active process will choose a local address based on the interface through
which it connects to the destination address.
-.It Fl a Ar ack_bytes
+.TP
+\fB\-Q tos
+Uses the RDS connection between IP addresses with the specified tos value. By
+default, the base (tos = 0) RDS connection is used. Valid values are 0-255.
+.TP
+\fB\-a ack_bytes
This specifies the size of the ack messages, in bytes. There is a minimum size
which depends on the format of the ack messages, which may change over time.
See section "Message Sizes" below.
-.It Fl q Ar request_bytes
+.TP
+\fB\-q request_bytes
This specifies the size of the request messages, in bytes.
It also has a minimum size which may change over time.
See section "Message Sizes" below.
-.It Fl D Ar rdma_bytes
+.TP
+\fB\-D rdma_bytes
RDSv3 is capable of transmitting part of a message via RDMA directly from
application buffer to application buffer. This option enables RDMA support
in rds-stress: request packets include parameters for an RDMA READ or WRITE
operation, which the receiving process executes at the time the ACK packet
is sent.
See section "Message Sizes" below.
-.It Fl d Ar queue_depth
+.TP
+\fB\-d queue_depth
Each child will try to maintain this many sent messages outstanding to each
of its peers on the remote address.
-.It Fl t Ar nr_tasks
+.TP
+\fB\-t nr_tasks
Each parent will create this many children tasks.
-.It Fl T Ar seconds
+.TP
+\fB\-T seconds
Specify the duration of the test run. After the specified number of seconds,
all processes on both ends of the connection will terminate, and the
active instance will print a summary. By default, rds-stress will keep
on sending and receiving messages.
-.It Fl z
+.TP
+\fB\-z
This flag can be used in conjunction with -T. It suppresses the ongoing
display of statistics, and prints a summary only.
-.It Fl c
+.TP
+\fB\-c
This causes rds-stress to create child tasks which just consume CPU cycles.
One task is created for each CPU in the system. First each child observes the
maximum rate at which it can consume cycles. This means that this option
@@ -121,54 +133,70 @@
use of the system by observing the lesser rate at which the children consume
cycles. This option is *not* shared between the active and passive instances.
It must be specified on each rds-stress command line.
-.It Fl R
+.TP
+\fB\-R
This tells the rds-stress parent process to run with SCHED_RR priority,
giving it precedence over the child processes. This is useful when running
with lots of tasks, where there is a risk of the child processes starving
the parent, and skewing the results.
-.It Fl v
+.TP
+\fB\-v
With this option enabled, packets are filled with a pattern that is
verified by the receiver. This check can help detect data corruption
occuring under high load.
+.TP
+.PP
-.Ss Message Sizes
+.SS Message Sizes
Options which set a message size (such as -a) specify a number of bytes
by default. By appending \fBK\fP, \fBM\fP, or \fBG\fP, you can specify the size
in kilobytes, megabytes or gigabytes, respectively. For instance,
the following will run rds-stress with a message and ACK size of 1024
bytes, and an RDMA message size of 1048576 bytes:
-.Dl rds-stress ... -q 1K -a 1K -D 1M
-.Sh OUTPUT
+.PP
+.RS 12
+rds-stress ... -q 1K -a 1K -D 1M
+.RE
+.PP
+.PP
+.SH OUTPUT
Each parent outputs columns of statistics at a regular interval:
-.Bl -tag -width Ds
-.It tsks
+.TP 8
+tsks
The number of child tasks which are running.
-.It tx/s
+.TP
+tx/s
The number of sendmsg() calls that all children are executing, per second.
-.It tx+rx K/s
+.TP
+rx/s
+The number of recvmsg() calls that all children are executing, per second.
+.TP
+tx+rx K/s
The total number of bytes that are flowing through sendmsg() and recvmsg() for all children.
This includes both request and ack messages.
-.It rw+rr K/s
-The total number of bytes that are being transferred via RDMA READs and
+.TP
+mbi K/s
+The total number of bytes that are being received via RDMA READs and
WRITEs for all children.
-.It tx us/c
+.TP
+mbo K/s
+The total number of bytes that are being transmited via RDMA READs and
+WRITEs for all children.
+.TP
+tx us/c
The average number of microseconds spent in sendmsg() calls.
-.It rtt us
+.TP
+rtt us
The average round trip time for a request and ack message pair. This measures
the total time between when a task sends a request and when it finally receives
the ack for that message. Because it includes the time it takes for the
receiver to wake up, receive the message, and send an ack, it can grow to be
quite large under load.
-.It cpu %
+.TP
+cpu %
This is the percentage of available CPU resources on this machine that are being
consumed since rds-stress started running. It will show -1.00 if -c is not
given. It is calculated based on the amount of CPU resources that CPU soaking
tasks are able to consume. This lets it measure CPU use by the system, say in
interrupt handlers, that task-based CPU accounting does not include.
For this to work rds-stress must be started with -c on an idle system.
diff -r 1afa90e87b4b rds-stress.c
--- a/rds-stress.c Tue Feb 23 22:30:57 2016 -0800
+++ b/rds-stress.c Tue Feb 23 22:30:58 2016 -0800
@@ -15,17 +15,38 @@
#include <sys/time.h>
#include <time.h>
#include <inttypes.h>
+#if defined(__SVR4) && defined(__sun)
+#include <sys/syscall.h>
+#include <signal.h>
+#include <sys/lgrp_user.h>
+#else
#include <syscall.h>
+#endif
#include <sys/stat.h>
#include <sys/poll.h>
#include <ctype.h>
#include <fcntl.h>
#include <sched.h>
#include <getopt.h>
+#include <sys/ioctl.h>
+#if !(defined(__SVR4) && defined(__sun))
#include <byteswap.h>
#include "rds.h"
-
+#else
+#include <sys/rds.h>
+#endif
#include "pfhack.h"
+#if defined(__SVR4) && defined(__sun)
+#include <infiniband/ofa_solaris.h>
+#endif
+
+#if defined(__SVR4) && defined(__sun)
+/*
+ * This definition is forward looking and is not yet present
+ * in Solaris rds.h file
+ */
+#define RDS_CMSG_RDMA_SEND_STATUS RDS_CMSG_RDMA_STATUS
+#endif
/*
*
@@ -45,8 +66,9 @@
M_RDMA_READ_ONLY,
M_RDMA_WRITE_ONLY
};
-
-struct options {
+#define VERSION_MAX_LEN 16
+
+struct options_2_0_6 {
uint32_t req_depth;
uint32_t req_size;
uint32_t ack_size;
@@ -76,8 +98,67 @@
uint32_t connect_retries;
} __attribute__((packed));
+struct options {
+ char version[VERSION_MAX_LEN];
+ uint32_t req_depth;
+ uint32_t req_size;
+ uint32_t ack_size;
+ uint32_t rdma_size;
+ uint32_t send_addr;
+ uint32_t receive_addr;
+ uint16_t starting_port;
+ uint16_t nr_tasks;
+ uint32_t run_time;
+ uint8_t summary_only;
+ uint8_t rtprio;
+ uint8_t tracing;
+ uint8_t verify;
+ uint8_t show_params;
+ uint8_t show_perfdata;
+ uint8_t use_cong_monitor;
+ uint8_t rdma_use_once;
+ uint8_t rdma_use_get_mr;
+ uint8_t rdma_use_fence;
+ uint8_t rdma_cache_mrs;
+ uint8_t rdma_key_o_meter;
+ uint8_t suppress_warnings;
+ uint8_t simplex;
+ uint8_t rw_mode;
+ uint32_t rdma_vector;
+ uint32_t rdma_alignment;
+ uint32_t connect_retries;
+ uint8_t tos;
+ uint8_t async;
+} __attribute__((packed));
+
+#define MAX_BUCKETS 16
+
static struct options opt;
static int control_fd;
+static uint64_t rtt_threshold;
+static int show_histogram;
+static int reset_connection;
+static char peer_version[VERSION_MAX_LEN];
+
+static int get_bucket(uint64_t rtt_time)
+{
+ int i;
+ uint64_t l_rtt_time = rtt_time;
+
+ if (!l_rtt_time)
+ i = 0;
+ else
+ {
+ i = -1;
+ while (l_rtt_time)
+ {
+ i++;
+ l_rtt_time = (l_rtt_time >> 1);
+ }
+ }
+
+ return i;
+}
struct counter {
uint64_t nr;
@@ -102,6 +183,10 @@
#define NR_STATS S__LAST
+#if defined(__SVR4) && defined(__sun)
+int sol_ioctl(int, int, struct rds_info_arg *, socklen_t *, int *);
+#endif
+
/*
* Parents share a mapped array of these with their children. Each child
* gets one. It's used to communicate between the child and the parent
@@ -110,9 +195,11 @@
struct child_control {
pid_t pid;
int ready;
+ int stopping;
struct timeval start;
struct counter cur[NR_STATS];
struct counter last[NR_STATS];
+ uint64_t latency_histogram[MAX_BUCKETS];
} __attribute__((aligned (256))); /* arbitrary */
struct soak_control {
@@ -132,6 +219,7 @@
*/
#define OP_REQ 1
#define OP_ACK 2
+#define OP_DUMP 3
#define RDMA_OP_READ 1
#define RDMA_OP_WRITE 2
@@ -148,7 +236,7 @@
uint16_t from_port;
uint16_t to_port;
uint16_t index;
- uint8_t op;
+ uint8_t op;
/* RDMA related.
* rdma_op must be the first field, because we
@@ -162,12 +250,21 @@
uint32_t rdma_size;
uint32_t rdma_vector;
- uint8_t data[0];
+ /* Async send related. */
+ uint8_t retry;
+ uint8_t rdma_remote_err;
+ uint8_t pending;
+
+ uint8_t data[0];
} __attribute__((packed));
#define MIN_MSG_BYTES (sizeof(struct header))
#define BASIC_HEADER_SIZE (size_t)(&((struct header *) 0)->rdma_op)
+#define print_outlier(...) do { \
+ fprintf(stderr, __VA_ARGS__); \
+} while (0)
+
#define die(fmt...) do { \
fprintf(stderr, fmt); \
exit(1); \
@@ -254,6 +351,19 @@
die("invalid host name or dotted quad '%s'\n", ptr);
}
+#if defined(__SVR4) && defined(__sun)
+static lgrp_id_t lgrp_id = -1;
+
+static void
+set_my_lgrp(void)
+{
+ if (lgrp_id != -1) {
+ lgrp_affinity_set(P_LWPID, P_MYID, lgrp_id,
+ LGRP_AFF_STRONG);
+ yield(); /* force a context switch */
+ }
+}
+#endif
static void usage(void)
{
@@ -273,6 +383,7 @@
" -d [depth, 1] request pipeline depth, nr outstanding\n"
" -t [nr, 1] number of child tasks\n"
" -T [seconds, 0] runtime of test, 0 means infinite\n"
+ " -Q [tos, 0] Type of Service\n"
" -D [bytes] RDMA: size\n"
" -I [iovecs, 1] RDMA: number of user buffers to target (max 512)\n"
" -M [nr, 0] RDMA: mode (0=readwrite,1=readonly,2=writeonly)\n"
@@ -281,6 +392,9 @@
" -c measure cpu use with per-cpu soak processes\n"
" -V trace execution\n"
" -z print a summary at end of test only\n"
+#if defined(__SVR4) && defined(__sun)
+ " -g [lgrpid] bind the process to the specified lgrp\n"
+#endif
"\n"
"Example:\n"
" recv$ rds-stress\n"
@@ -310,7 +424,7 @@
static void check_parent(pid_t pid)
{
if (pid != getppid())
- die("parent %u exited\n", pid);
+ die("parent %u exited\n", (int)pid);
}
/*
@@ -334,6 +448,7 @@
msg_pattern[i] = k;
}
+#if !(defined(__SVR4) && defined(__sun))
#if __BYTE_ORDER == __LITTLE_ENDIAN
#define htonll(x) bswap_64(x)
#define ntohll(x) bswap_64(x)
@@ -341,6 +456,7 @@
#define htonll(x) (x)
#define ntohll(x) (x)
#endif
+#endif /* Not sun */
static void encode_hdr(struct header *dst, const struct header *hdr)
{
@@ -361,6 +477,7 @@
dst->rdma_key = htonll(hdr->rdma_key);
dst->rdma_size = htonl(hdr->rdma_size);
dst->rdma_vector = htonl(hdr->rdma_vector);
+ dst->retry = hdr->retry;
}
static void decode_hdr(struct header *dst, const struct header *hdr)
@@ -382,6 +499,7 @@
dst->rdma_key = ntohll(hdr->rdma_key);
dst->rdma_size = ntohl(hdr->rdma_size);
dst->rdma_vector = ntohl(hdr->rdma_vector);
+ dst->retry = hdr->retry;
}
static void fill_hdr(void *message, uint32_t bytes, struct header *hdr)
@@ -412,11 +530,19 @@
* Compare incoming message header with expected header. All header fields
* are in host byte order except for address and port fields.
*/
-static int check_hdr(void *message, uint32_t bytes, const struct header *hdr)
+static int check_hdr(void *message, uint32_t bytes, struct header *hdr, struct options *opts)
{
struct header msghdr;
+ uint32_t inc_seq;
+ uint32_t my_seq;
decode_hdr(&msghdr, message);
+ inc_seq = msghdr.seq;
+ my_seq = hdr->seq;
+
+ if (msghdr.retry && (inc_seq < my_seq))
+ return -1;
+
if (memcmp(&msghdr, hdr, BASIC_HEADER_SIZE)) {
#define bleh(var, disp) \
disp(hdr->var), \
@@ -428,7 +554,7 @@
* with stdout() and we don't get things stomping on each
* other
*/
- printf( "An incoming message had a header which\n"
+ printf( "An incoming message had a %s header which\n"
"didn't contain the fields we expected:\n"
" member expected eq got\n"
" seq %15u %s %15u\n"
@@ -438,6 +564,7 @@
" to_port %15u %s %15u\n"
" index %15u %s %15u\n"
" op %15u %s %15u\n",
+ (msghdr.retry) ? "RETRY" : "",
bleh(seq, /**/),
bleh(from_addr, inet_ntoa_32),
bleh(from_port, ntohs),
@@ -569,6 +696,9 @@
fcntl(fd, F_SETFL, O_NONBLOCK);
+ if (opts->tos && ioctl(fd, SIOCRDSSETTOS, &opts->tos))
+ die_errno("ERROR: failed to set TOS\n");
+
return fd;
}
@@ -584,7 +714,11 @@
if (opts->receive_addr == 0)
return 1;
+#if defined(__SVR4) && defined(__sun)
+ sin.sin_family = AF_INET_OFFLOAD;
+#else
sin.sin_family = AF_INET;
+#endif
sin.sin_port = htons(opts->starting_port);
sin.sin_addr.s_addr = htonl(opts->receive_addr);
@@ -639,7 +773,7 @@
mr_args.flags = RDS_FREE_MR_ARGS_INVALIDATE;
#endif
if (setsockopt(fd, sol, RDS_FREE_MR, &mr_args, sizeof(mr_args)))
- die_errno("setsockopt(RDS_FREE_MR) failed");
+ return;
mrs_allocated--;
}
@@ -677,7 +811,11 @@
size = sizeof(struct rdma_key_o_meter)
+ 2 * nr_tasks * sizeof(*kt)
+ 2 * RDMA_MAX_TRACKED_KEYS * sizeof(*ks);
+#if defined(__SVR4) && defined(__sun)
+ base = mmap(NULL, size, PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_SHARED, -1, 0);
+#else
base = mmap(NULL, size, PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_SHARED, 0, 0);
+#endif
if (base == MAP_FAILED)
die_errno("alloc_rdma_buffers: mmap failed");
@@ -828,13 +966,20 @@
}
if (!failed)
- trace("compare pass pattern %Lx addr %p\n",
+ trace("compare pass pattern 0x%Lx addr %p\n",
(unsigned long long) pattern, addr);
}
+struct retry_entry {
+ uint32_t retries;
+ uint32_t seq;
+ int status;
+};
+
struct task {
unsigned int nr;
unsigned int pending;
+ int trace;
unsigned int unacked;
struct sockaddr_in src_addr; /* same for all tasks */
struct sockaddr_in dst_addr;
@@ -846,6 +991,13 @@
uint16_t recv_index;
struct timeval * send_time;
struct header * ack_header;
+ struct header * ack2_header;
+ struct header * req_header;
+ uint64_t * retry_token;
+ uint32_t retries;
+ uint32_t last_retry_seq;
+ uint32_t retry_index;
+
/* RDMA related stuff */
uint64_t ** local_buf;
@@ -865,7 +1017,11 @@
/* We use mmap here rather than malloc, because it is always
* page aligned. */
len = 2 * opts->nr_tasks * opts->req_depth * (opts->rdma_vector * opts->rdma_size) + sys_page_size;
+#if defined(__SVR4) && defined(__sun)
+ base = mmap(NULL, len, PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE, -1, 0);
+#else
base = mmap(NULL, len, PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE, 0, 0);
+#endif
if (base == MAP_FAILED)
die_errno("alloc_rdma_buffers: mmap failed");
memset(base, 0x2f, len);
@@ -915,17 +1071,16 @@
if (RDMA_OP_READ == hdr->rdma_op) {
if (opt.verify)
rds_fill_buffer(rdma_addr, rdma_size, hdr->rdma_pattern);
- trace("Requesting RDMA read for pattern %Lx "
- "local addr to rdma read %p\n",
- (unsigned long long) hdr->rdma_pattern,
+ trace("Requesting RDMA read for pattern 0x%Lx"
+ "local addr to rdma read 0x%p\n",
+ hdr->rdma_pattern,
rdma_addr);
} else {
if (opt.verify)
rds_fill_buffer(rdma_addr, rdma_size, 0);
- trace("Requesting RDMA write for pattern %Lx "
- "local addr to rdma write %p\n",
- (unsigned long long) hdr->rdma_pattern,
- rdma_addr);
+
+ trace("Requesting RDMA write for pattern 0x%Lx",
+ hdr->rdma_pattern);
}
}
@@ -947,7 +1102,7 @@
die("Unexpected RDMA op %u in request\n", in_hdr->rdma_op);
- trace("RDS received request to issue rdma %s len %lu rva %Lx key %Lx pattern %Lx\n",
+ trace("RDS received request to issue rdma %s len %lu rva 0x%Lx key 0x%Lx pattern 0x%Lx\n",
in_hdr->rdma_op == RDMA_OP_WRITE? "write to" : "read from",
rdma_size,
(unsigned long long) in_hdr->rdma_addr,
@@ -966,20 +1121,32 @@
hdr->rdma_vector = in_hdr->rdma_vector;
}
-static inline unsigned int rdma_user_token(struct task *t, unsigned int qindex)
+static inline uint64_t rdma_user_token(struct task *t, unsigned int qindex, unsigned int type, uint32_t seq)
{
- return t->nr * opt.req_depth + qindex;
+ uint64_t tmp = seq;
+ return (tmp << 32) | ((t->nr * opt.req_depth + qindex) << 2 | type);
}
-static void rdma_mark_completed(struct task *tasks, unsigned int token, int status)
+static void rdma_mark_completed(struct task *tasks, uint64_t token, int status, struct options *opts)
{
struct task *t;
unsigned int i;
-
- trace("RDS rdma completion for token %x\n", token);
-
- t = &tasks[token / opt.req_depth];
- i = token % opt.req_depth;
+ struct header *hdr = NULL;
+ uint32_t seq = token >> 32;
+ unsigned int type = token & 0x03;
+ unsigned int index = (token & 0xFFFFFFFF) >> 2;
+
+ trace("RDS rdma completion for token 0x%lx\n", token);
+
+ t = &tasks[index / opt.req_depth];
+ i = index % opt.req_depth;
+
+ if (opts->async) {
+ if (type == OP_REQ)
+ hdr = &t->req_header[i];
+ else
+ hdr = &t->ack2_header[i];
+ }
if (status) {
const char *errmsg;
@@ -987,26 +1154,59 @@
switch (status) {
case RDS_RDMA_REMOTE_ERROR:
errmsg = "remote error"; break;
- case RDS_RDMA_CANCELED:
+ case RDS_RDMA_SEND_DROPPED:
+ errmsg = "operation was dropped"; break;
+ case RDS_RDMA_SEND_CANCELED:
errmsg = "operation was cancelled"; break;
- case RDS_RDMA_DROPPED:
- errmsg = "operation was dropped"; break;
- case RDS_RDMA_OTHER_ERROR:
+ case RDS_RDMA_SEND_OTHER_ERROR:
errmsg = "other error"; break;
default:
errmsg = "unknown error"; break;
}
- printf("%s:%u: RDMA op %u failed: %s\n",
+ trace("%s:%u: %s failed: %s\n",
inet_ntoa(t->dst_addr.sin_addr),
ntohs(t->dst_addr.sin_port),
- i, errmsg);
+ type ? "SEND" : "RDMA",
+ errmsg);
+
+ if (hdr &&
+ (status == RDS_RDMA_SEND_DROPPED ||
+ status == RDS_RDMA_REMOTE_ERROR)) {
+
+ if (hdr->seq == seq) {
+ hdr->retry = 1;
+ if (hdr->seq > t->last_retry_seq) {
+ if (status == RDS_RDMA_REMOTE_ERROR)
+ hdr->rdma_remote_err = 1;
+ t->retry_token[t->retry_index] = token;
+ t->retry_index = (t->retry_index + 1) %
+ (2 * opts->req_depth);
+ t->retries += 1;
+ t->last_retry_seq = hdr->seq;
+ if (t->retries > 2 * opts->req_depth)
+ die("Exceeded MAX retry entries..\n");
+ }
+ } else
+ die("SEQ Out-Of-Sync: %u/%u\n", hdr->seq, seq);
+ } else if (hdr) {
+ hdr->pending = 0;
+ hdr->retry = 0;
+ hdr->rdma_remote_err = 0;
+ }
+ } else if (hdr) {
+ hdr->pending = 0;
+ hdr->retry = 0;
+ hdr->rdma_remote_err = 0;
}
t->rdma_inflight[i] = 0;
t->drain_rdmas = 0;
}
+#if defined(__SVR4) && defined(__sun)
+#undef MSG_MAXIOVLEN
+#endif
#define MSG_MAXIOVLEN 2
/*
@@ -1018,11 +1218,14 @@
static char ctlbuf[1024];
struct cmsghdr *cmsg;
- msg->msg_control = ctlbuf;
- msg->msg_controllen = CMSG_SPACE(size);
-
- cmsg = CMSG_FIRSTHDR(msg);
- cmsg->cmsg_level = sol;
+ if (!msg->msg_control) {
+ msg->msg_control = ctlbuf;
+ msg->msg_controllen = CMSG_SPACE(size);
+ cmsg = CMSG_FIRSTHDR(msg);
+ } else {
+ cmsg = (struct cmsghdr *)((char *)msg->msg_control + msg->msg_controllen);
+ msg->msg_controllen += CMSG_SPACE(size);
+ }cmsg->cmsg_level = sol;
cmsg->cmsg_type = type;
cmsg->cmsg_len = CMSG_LEN(size);
memcpy(CMSG_DATA(cmsg), ptr, size);
@@ -1034,7 +1237,7 @@
* the ACK packet.
*/
static void rdma_build_cmsg_xfer(struct msghdr *msg, const struct header *hdr,
- unsigned int user_token, void *local_buf)
+ uint64_t user_token, void *local_buf)
{
#define RDS_MAX_IOV 512 /* FIX_ME - put this into rds.h or use socket max ?*/
@@ -1048,7 +1251,7 @@
rdma_size = hdr->rdma_size;
rdma_vector = hdr->rdma_vector;
- trace("RDS issuing rdma for token %x key %Lx len %u local_buf %p vector %u\n",
+ trace("RDS issuing rdma for token 0x%lx key 0x%llx len %d local_buf %p vector %d\n",
user_token,
(unsigned long long) hdr->rdma_key,
rdma_size, local_buf,
@@ -1102,6 +1305,15 @@
rdma_put_cmsg(msg, RDS_CMSG_RDMA_ARGS, &args, sizeof(args));
}
+static void build_cmsg_async_send(struct msghdr *msg, uint64_t user_token)
+{
+ struct rds_asend_args args;
+
+ args.flags |= RDS_SEND_NOTIFY_ME;
+ args.user_token = user_token;
+ rdma_put_cmsg(msg, RDS_CMSG_ASYNC_SEND, &args, sizeof(args));
+}
+
static void rdma_build_cmsg_dest(struct msghdr *msg, rds_rdma_cookie_t rdma_dest)
{
rdma_put_cmsg(msg, RDS_CMSG_RDMA_DEST, &rdma_dest, sizeof(rdma_dest));
@@ -1174,19 +1386,17 @@
hdr->index = qindex;
}
-static int send_packet(int fd, struct task *t,
- struct header *hdr, unsigned int size)
+static int send_msg(int fd, struct task *t, struct header *hdr,
+ unsigned int size, struct options *opts,
+ struct child_control *ctl)
{
- unsigned char buf[size], *rdma_flight_recorder = NULL;
+ unsigned char buf[size];
+ uint8_t *rdma_flight_recorder = NULL;
rds_rdma_cookie_t cookie = 0;
struct msghdr msg;
struct iovec iov;
ssize_t ret;
- /* Make sure we always have the current sequence number.
- * When we send ACK packets, the seq that gets filled in is
- * stale. */
- hdr->seq = t->send_seq;
fill_hdr(buf, size, hdr);
memset(&msg, 0, sizeof(msg));
@@ -1198,27 +1408,10 @@
iov.iov_base = buf;
iov.iov_len = size;
- /* If this is a REQ packet in which we pass the MR to the
- * peer, extract the RDMA cookie and pass it on in the control
- * message for now. */
- if (hdr->op == OP_REQ && hdr->rdma_op != 0) {
- if (hdr->rdma_key != 0) {
- /* We used GET_MR to obtain a key */
- rdma_build_cmsg_dest(&msg, hdr->rdma_key);
- cookie = hdr->rdma_key;
- hdr->rdma_key = 0;
- } else {
- /* Use the RDMA_MAP cmsg to have sendmsg do the
- * mapping on the fly. */
- rdma_build_cmsg_map(&msg, hdr->rdma_addr,
- hdr->rdma_size * hdr->rdma_vector,
- &cookie);
- }
- }
/* If this is an ACK packet with RDMA, build the cmsg
- * header that goes with it. */
- if (hdr->op == OP_ACK && hdr->rdma_op != 0) {
+ * header that goes with it. */
+ if (hdr->op == OP_ACK && hdr->rdma_op != 0 && !hdr->rdma_remote_err) {
unsigned int qindex = hdr->index;
if (t->rdma_inflight[qindex] != 0) {
@@ -1230,14 +1423,33 @@
*
* We return one of the more obscure error messages,
* which we recognize and handle in the top loop. */
- trace("Drain RDMA 0x%x\n", rdma_user_token(t, qindex));
+ trace("Drain RDMA 0x%lx\n", rdma_user_token(t, qindex, 0, hdr->seq));
errno = EBADSLT;
return -1;
}
rdma_build_cmsg_xfer(&msg, hdr,
- rdma_user_token(t, qindex),
+ rdma_user_token(t, qindex, 0, hdr->seq),
t->local_buf[qindex]);
rdma_flight_recorder = &t->rdma_inflight[qindex];
+ } else if (opts->async) {
+ if (hdr->op == OP_REQ)
+ build_cmsg_async_send(&msg,
+ rdma_user_token(t, hdr->index, OP_REQ, hdr->seq));
+ else
+ build_cmsg_async_send(&msg,
+ rdma_user_token(t, hdr->index, OP_ACK, hdr->seq));
+ }
+
+ if (hdr->op == OP_REQ && hdr->rdma_op != 0) {
+ if (hdr->rdma_key != 0) {
+ rdma_build_cmsg_dest(&msg, hdr->rdma_key);
+ cookie = hdr->rdma_key;
+ hdr->rdma_key = 0;
+ } else {
+ rdma_build_cmsg_map(&msg, hdr->rdma_addr,
+ hdr->rdma_size * hdr->rdma_vector,
+ &cookie);
+ }
}
ret = sendmsg(fd, &msg, 0);
@@ -1256,22 +1468,57 @@
* lower 32bit of the cookie */
rdma_key_o_meter_add(cookie);
}
+
+ hdr->pending = 1;
+
+ return ret;
+}
+
+static int send_packet(int fd, struct task *t,
+ struct header *hdr, unsigned int size,
+ struct options *opts, struct child_control *ctl)
+{
+ ssize_t ret;
+
+ /* Make sure we always have the current sequence number.
+ * When we send ACK packets, the seq that gets filled in is
+ * stale. */
+ hdr->seq = t->send_seq;
+
+ ret = send_msg(fd, t, hdr, size, opts, ctl);
+ if (ret < 0) return ret;
+
t->send_seq++;
return ret;
}
+static int resend_packet(int fd, struct task *t,
+ struct header *hdr, unsigned int size,
+ struct options *opts, struct child_control *ctl)
+{
+ ssize_t ret;
+
+ ret = send_msg(fd, t, hdr, size, opts, ctl);
+
+ return ret;
+}
+
static int send_one(int fd, struct task *t,
struct options *opts,
struct child_control *ctl)
{
struct timeval start;
struct timeval stop;
- struct header hdr;
+ struct header *hdr = &t->req_header[t->send_index];
int ret;
- build_header(t, &hdr, OP_REQ, t->send_index);
+ if (opts->async && hdr->pending) {
+ return -1;
+ }
+
+ build_header(t, hdr, OP_REQ, t->send_index);
if (opts->rdma_size && t->send_seq > 10)
- rdma_build_req(fd, &hdr, t,
+ rdma_build_req(fd, hdr, t,
opts->rdma_size,
opts->req_depth,
opts->rw_mode,
@@ -1279,7 +1526,7 @@
gettimeofday(&start, NULL);
- ret = send_packet(fd, t, &hdr, opts->req_size);
+ ret = send_packet(fd, t, hdr, opts->req_size, opts, ctl);
gettimeofday(&stop, NULL);
if (ret < 0)
@@ -1302,10 +1549,15 @@
struct child_control *ctl)
{
struct header *hdr = &t->ack_header[qindex];
+ struct header *hdr2 = &t->ack2_header[qindex];
ssize_t ret;
+ if (opts->async && hdr2->pending) {
+ return -1;
+ }
+
/* send an ack in response to the req we just got */
- ret = send_packet(fd, t, hdr, opts->ack_size);
+ ret = send_packet(fd, t, hdr, opts->ack_size, opts, ctl);
if (ret < 0)
return ret;
if (ret != opts->ack_size)
@@ -1324,6 +1576,8 @@
break;
}
+ memcpy(hdr2, hdr, sizeof(struct header));
+
return ret;
}
@@ -1354,8 +1608,49 @@
struct child_control *ctl,
int can_send, int do_work)
{
+ struct header *hdr;
+ unsigned int index;
+ int req_size;
+ int num_retries = t->retries;
+ uint64_t token;
+ unsigned int type;
+ unsigned int index2;
+ unsigned int i;
+
+ while (opts->async && num_retries > 0) {
+ index = (t->retry_index - num_retries +
+ (2 * opts->req_depth)) % (2 * opts->req_depth);
+
+ token = t->retry_token[index];
+ type = token & 0x03;
+ index2 = (token & 0xFFFFFFFF) >> 2;
+ i = index2 % opts->req_depth;
+
+ if (type == OP_REQ)
+ hdr = &t->req_header[i];
+ else
+ hdr = &t->ack2_header[i];
+
+ if (!hdr->retry)
+ goto next;
+
+ if (hdr->op == OP_REQ)
+ req_size = opts->req_size;
+ else
+ req_size = opts->ack_size;
+
+ if (resend_packet(fd, t, hdr, req_size, opts, ctl) < 0) {
+ return -1;
+ }
+ hdr->retry = 0;
+next:
+ num_retries--;
+ }
+ t->last_retry_seq = t->retries = 0;
+
if (ack_anything(fd, t, opts, ctl, can_send) < 0)
return -1;
+
while (do_work && t->pending < opts->req_depth) {
if (!can_send)
goto eagain;
@@ -1375,7 +1670,8 @@
rds_rdma_cookie_t *cookie,
struct sockaddr_in *sin,
struct timeval *tstamp,
- struct task *tasks)
+ struct task *tasks,
+ struct options *opts)
{
struct cmsghdr *cmsg;
char cmsgbuf[256];
@@ -1398,15 +1694,16 @@
if (ret < 0)
return ret;
- if (ret && ret < sizeof(struct header))
+ if (ret && !strcmp(RDS_VERSION, peer_version) &&
+ ret < sizeof(struct header))
die("recvmsg() returned short data: %zd", ret);
- if (msg.msg_namelen < sizeof(struct sockaddr_in))
+ if (ret && msg.msg_namelen < sizeof(struct sockaddr_in))
die("socklen = %d < sizeof(sin) (%zu)\n",
msg.msg_namelen, sizeof(struct sockaddr_in));
/* See if the message comes with a RDMA destination */
for (cmsg = CMSG_FIRSTHDR(&msg); cmsg; cmsg = CMSG_NXTHDR(&msg, cmsg)) {
- struct rds_rdma_notify notify;
+ struct rds_rdma_send_notify notify;
if (cmsg->cmsg_level != sol)
continue;
@@ -1432,11 +1729,11 @@
memcpy(cookie, CMSG_DATA(cmsg), sizeof(*cookie));
break;
- case RDS_CMSG_RDMA_STATUS:
+ case RDS_CMSG_RDMA_SEND_STATUS:
if (cmsg->cmsg_len < CMSG_LEN(sizeof(notify)))
die("RDS_CMSG_RDMA_DEST data too small");
memcpy(¬ify, CMSG_DATA(cmsg), sizeof(notify));
- rdma_mark_completed(tasks, notify.user_token, notify.status);
+ rdma_mark_completed(tasks, notify.user_token, notify.status, opts);
break;
}
}
@@ -1445,7 +1742,8 @@
static int recv_one(int fd, struct task *tasks,
struct options *opts,
- struct child_control *ctl)
+ struct child_control *ctl,
+ struct child_control *all_ctl)
{
char buf[max(opts->req_size, opts->ack_size)];
rds_rdma_cookie_t rdma_dest = 0;
@@ -1456,15 +1754,18 @@
uint16_t expect_index;
int task_index;
ssize_t ret;
-
- ret = recv_message(fd, buf, sizeof(buf), &rdma_dest, &sin, &tstamp, tasks);
+ int check_status;
+
+
+ ret = recv_message(fd, buf, sizeof(buf), &rdma_dest, &sin, &tstamp, tasks, opts);
if (ret < 0)
return ret;
/* If we received only RDMA completions or cong updates,
* ret will be 0 */
- if (ret == 0)
+ if (ret == 0) {
return 0;
+ }
/* check the incoming sequence number */
task_index = ntohs(sin.sin_port) - opts->starting_port - 1;
@@ -1508,15 +1809,31 @@
hdr.to_port = t->src_addr.sin_port;
hdr.index = expect_index;
- if (check_hdr(buf, ret, &hdr))
- die("header from %s:%u to id %u bogus\n",
- inet_ntoa(sin.sin_addr), htons(sin.sin_port),
- ntohs(t->src_addr.sin_port));
+ check_status = check_hdr(buf, ret, &hdr, opts);
+ if (check_status) {
+ if (check_status > 0) {
+ die("header from %s:%u to id %u bogus\n",
+ inet_ntoa(sin.sin_addr), htons(sin.sin_port),
+ ntohs(t->src_addr.sin_port));
+ } else
+ return 0;
+ }
if (hdr.op == OP_ACK) {
- stat_inc(&ctl->cur[S_RTT_USECS],
- usec_sub(&tstamp, &t->send_time[expect_index]));
- t->pending -= 1;
+ uint64_t rtt_time =
+ usec_sub(&tstamp, &t->send_time[expect_index]);
+
+ stat_inc(&ctl->cur[S_RTT_USECS], rtt_time);
+ if (rtt_time > rtt_threshold)
+ print_outlier("Found RTT = 0x%lx\n", rtt_time);
+
+ if (show_histogram)
+ {
+ ctl->latency_histogram[get_bucket(rtt_time)]++;
+ }
+
+ if (t->pending > 0)
+ t->pending -= 1;
if (in_hdr.rdma_key)
rdma_process_ack(fd, &in_hdr, ctl);
@@ -1549,6 +1866,7 @@
}
static void run_child(pid_t parent_pid, struct child_control *ctl,
+ struct child_control *all_ctl,
struct options *opts, uint16_t id, int active)
{
struct sockaddr_in sin;
@@ -1559,8 +1877,15 @@
struct task tasks[opts->nr_tasks];
struct timeval start;
int do_work = opts->simplex ? active : 1;
-
+ int j;
+
+
+#if defined(__SVR4) && defined(__sun)
+ set_my_lgrp();
+ sin.sin_family = AF_INET_OFFLOAD;
+#else
sin.sin_family = AF_INET;
+#endif
sin.sin_port = htons(opts->starting_port + 1 + id);
sin.sin_addr.s_addr = htonl(opts->receive_addr);
@@ -1572,7 +1897,11 @@
for (i = 0; i < opts->nr_tasks; i++) {
tasks[i].nr = i;
tasks[i].src_addr = sin;
+#if defined(__SVR4) && defined(__sun)
+ tasks[i].dst_addr.sin_family = AF_INET_OFFLOAD;
+#else
tasks[i].dst_addr.sin_family = AF_INET;
+#endif
tasks[i].dst_addr.sin_addr.s_addr = htonl(opts->send_addr);
tasks[i].dst_addr.sin_port = htons(opts->starting_port + 1 + i);
tasks[i].send_time = alloca(opts->req_depth * sizeof(struct timeval));
@@ -1581,6 +1910,15 @@
tasks[i].rdma_buf = alloca(opts->req_depth * sizeof(uint64_t *));
tasks[i].local_buf = alloca(opts->req_depth * sizeof(uint64_t *));
tasks[i].ack_header = alloca(opts->req_depth * sizeof(struct header));
+ tasks[i].ack2_header = alloca(opts->req_depth * sizeof(struct header));
+ for (j=0;j<opts->req_depth;j++)
+ tasks[i].ack2_header[j].pending = 0;
+
+ tasks[i].req_header = alloca(opts->req_depth * sizeof(struct header));
+ for (j=0;j<opts->req_depth;j++)
+ tasks[i].req_header[j].pending = 0;
+
+ tasks[i].retry_token = alloca(2 * opts->req_depth * sizeof(uint64_t));
tasks[i].rdma_next_op = (i & 1)? RDMA_OP_READ : RDMA_OP_WRITE;
}
@@ -1611,7 +1949,7 @@
check_parent(parent_pid);
- ret = poll(&pfd, 1, -1);
+ ret = poll(&pfd, 1, 1000);
if (ret < 0) {
if (errno == EINTR)
continue;
@@ -1621,10 +1959,14 @@
pfd.events = POLLIN;
if (pfd.revents & POLLIN) {
- while (recv_one(fd, tasks, opts, ctl) >= 0)
+ while (recv_one(fd, tasks, opts, ctl, all_ctl) >= 0)
;
}
+ /* stop sending if in shutdown phase */
+ if (ctl->stopping)
+ continue;
+
/* keep the pipeline full */
can_send = !!(pfd.revents & POLLOUT);
for (i = 0, t = tasks; i < opts->nr_tasks; i++, t++) {
@@ -1633,6 +1975,7 @@
if (t->drain_rdmas)
continue;
if (send_anything(fd, t, opts, ctl, can_send, do_work) < 0) {
+
pfd.events |= POLLOUT;
/* If the send queue is full, we will see EAGAIN.
@@ -1665,8 +2008,12 @@
uint32_t i;
len = opts->nr_tasks * sizeof(*ctl);
+#if defined(__SVR4) && defined(__sun)
+ ctl = (struct child_control *)mmap(NULL, len, PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_SHARED, -1, 0);
+#else
ctl = mmap(NULL, len, PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_SHARED,
0, 0);
+#endif
if (ctl == MAP_FAILED)
die("mmap of %u child control structs failed", opts->nr_tasks);
@@ -1688,7 +2035,7 @@
control_fd = -1;
}
rdma_key_o_meter_set_self(i);
- run_child(parent, ctl + i, opts, i, active);
+ run_child(parent, ctl + i, ctl, opts, i, active);
exit(0);
}
ctl[i].pid = pid;
@@ -1699,7 +2046,7 @@
continue;
pid = waitpid(-1, NULL, WNOHANG);
if (pid)
- die("child %u (pid %u) exited\n", i, pid);
+ die("child %u (pid %u) exited\n", i, (int)pid);
sleep(1);
i--; /* try this child again */
}
@@ -1823,6 +2170,7 @@
if (disable)
return;
+#if !(defined(__SVR4) && defined(__sun))
if ((fp = fopen("/proc/stat", "r")) == NULL) {
fprintf(stderr, "Cannot open /proc/stat (%s) - "
"not printing cpu stats\n",
@@ -1856,10 +2204,37 @@
}
}
fclose(fp);
+#else
+#define NSEC_TO_TICK(v) (v * sysconf(_SC_CLK_TCK)/1000000000)
+ sol_cpu_stats_t stats;
+
+ solaris_init();
+ if ((sol_get_cpu_stats(&stats)) < 0) {
+ disable = 1;
+ return;
+ }
+ solaris_fini();
+ current.times[0] = stats.t_user;
+ current.times[1] = 0;
+ current.times[2] = stats.t_kernel;
+ current.times[3] = stats.t_idle;
+ current.times[4] = stats.t_iowait;
+ current.times[5] = 0;
+ current.times[6] = 0;
+ current.times[7] = 0;
+ current.intr = NSEC_TO_TICK(stats.t_intr); /* NSEC_TO_TICK */
+
+#endif
if (initialize) {
+#if !(defined(__SVR4) && defined(__sun))
printf(",user:percent,system:percent,idle:percent"
",irq:percent,intr:count");
+#else
+ /* Solaris kstat doesn't provide irq/softirq info. */
+ printf(",user:percent,system:percent,idle:percent"
+ ",intr:count");
+#endif
} else {
struct sys_stats sys;
unsigned long sum = 0;
@@ -1884,12 +2259,21 @@
* 5 irq
* 6 softirq
*/
+#if !(defined(__SVR4) && defined(__sun))
printf(",%f,%f,%f,%f,%Lu",
sys.times[2] * scale,
sys.intr);
+#else
+ /* Solaris kstat doesn't provide irq/softirq info. */
+ printf(",%f,%f,%f,%Lu",
+ sys.times[2] * scale,
+ sys.intr);
+#endif
}
prev = current;
}
@@ -1903,6 +2287,10 @@
static socklen_t buflen = 0;
static int sock_fd = -1;
int i, count, item_size;
+#if defined(__SVR4) && defined(__sun)
+ socklen_t len;
+ struct rds_info_arg arg;
+#endif
if (sock_fd < 0) {
sock_fd = socket(pf, SOCK_SEQPACKET, 0);
@@ -1912,6 +2300,7 @@
/* We should only loop once on the first call; after that the
* buffer requirements for RDS counters should not change. */
+#if !(defined(__SVR4) && defined(__sun))
while ((item_size = getsockopt(sock_fd, sol, RDS_INFO_COUNTERS, curr, &buflen)) < 0) {
if (errno != ENOSPC)
die_errno("getsockopt(RDS_INFO_COUNTERS) failed");
@@ -1919,6 +2308,28 @@
if (!curr)
die_errno("Cannot allocate buffer for stats counters");
}
+#else
+ int retcode;
+
+ retcode = sol_ioctl(
+ sock_fd, RDS_INFO_COUNTERS, &arg, &buflen, &item_size);
+ if (retcode != 0) {
+ if (retcode == 1) {
+ die_errno("ioctl(RDS_INFO_COUNTERS) failed");
+ } else if (retcode == 2) {
+ fprintf(stderr, "%s: Unable to allocate memory "
+ "for %u bytes of info: %s\n",
+ "rds-stress", buflen, strerror(errno));
+ return;
+ } else {
+ fprintf(stderr, "%s: Unable to access "
+ "RDS_INFO_COUNTERS statistics: %s\n",
+ "rds-stress", strerror(errno));
+ return;
+ }
+ }
+
+#endif
if (item_size > sizeof(*ctr))
die("Bad counter item size in RDS_INFO_COUNTERS (got %d, max %zd)\n",
@@ -1932,8 +2343,11 @@
}
for (i = 0; i < count; ++i)
+#if !(defined(__SVR4) && defined(__sun))
memcpy(ctr + i, curr + i * item_size, item_size);
-
+#else
+ memcpy(ctr + i, ((void *)(uintptr_t)arg.datap) + i * item_size, item_size);
+#endif
gettimeofday(&now, NULL);
if (initialize) {
@@ -1957,6 +2371,10 @@
memcpy(prev, ctr, count * sizeof(*ctr));
last_ts = now;
+#if defined(__SVR4) && defined(__sun)
+ free((void *)(uintptr_t)arg.datap);
+#endif
+
get_stats(initialize);
}
@@ -1967,7 +2385,7 @@
pid = waitpid(-1, &status, wflags);
if (pid < 0)
- die("waitpid returned %u", pid);
+ die("waitpid returned %u", (int)pid);
if (pid == 0)
return 0;
@@ -1975,15 +2393,15 @@
if (WEXITSTATUS(status) == 0)
return 1;
die("child pid %u exited with status %d\n",
- pid, WEXITSTATUS(status));
+ (int)pid, WEXITSTATUS(status));
}
if (WIFSIGNALED(status)) {
if (WTERMSIG(status) == SIGTERM)
return 1;
die("child pid %u exited with signal %d\n",
- pid, WTERMSIG(status));
+ (int)pid, WTERMSIG(status));
}
- die("child pid %u wait status %d\n", pid, status);
+ die("child pid %u wait status %d\n", (int)pid, status);
}
static void release_children_and_wait(struct options *opts,
@@ -1995,8 +2413,12 @@
struct counter summary[NR_STATS];
struct timeval start, end, now, first_ts, last_ts;
double cpu_total = 0;
- uint16_t i, cpu_samples = 0;
+ uint16_t i, j, cpu_samples = 0;
uint16_t nr_running;
+ uint64_t latency_histogram[MAX_BUCKETS];
+
+ if (show_histogram)
+ memset(latency_histogram, 0, sizeof(latency_histogram));
gettimeofday(&start, NULL);
start.tv_sec += 2;
@@ -2139,6 +2561,11 @@
control_fd = -1;
if (nr_running) {
+ /* let everything gracefully stop before we kill the chillins */
+ for (i = 0; i < opts->nr_tasks; i++)
+ ctl[i].stopping = 1;
+ sleep(1);
+
for (i = 0; i < opts->nr_tasks; i++)
kill(ctl[i].pid, SIGTERM);
stop_soakers(soak_arr);
@@ -2167,6 +2594,19 @@
avg(&summary[S_SENDMSG_USECS]),
avg(&summary[S_RTT_USECS]),
soak_arr? scale * cpu_total : -1.0);
+
+ if (show_histogram)
+ {
+ for (i = 0; i < opts->nr_tasks; i++)
+ for (j=0;j < MAX_BUCKETS; j++)
+ latency_histogram[j] += ctl[i].latency_histogram[j];
+
+ printf("\nRTT histogram\n");
+ printf("RTT (us) \t\t Count\n");
+ for (i=0;i < MAX_BUCKETS; i++)
+ printf("[%6u - %6u] \t\t %8u\n", 1 << i, 1 << (i+1),
+ (unsigned int)latency_histogram[i]);
+ }
}
}
@@ -2220,6 +2660,21 @@
{
ssize_t ret;
+ if (size == sizeof(struct options)) {
+ memset(ptr, 0, size);
+ ret = read(fd, peer_version, VERSION_MAX_LEN);
+ if (ret != VERSION_MAX_LEN)
+ die_errno("Failed to read version");
+
+ if (strcmp(peer_version, RDS_VERSION)) {
+ ptr += ret;
+ memcpy(ptr, peer_version, VERSION_MAX_LEN);
+ size = sizeof(struct options_2_0_6) - ret;
+ } else
+ size -= ret;
+ ptr += ret;
+ }
+
while (size) {
ret = read(fd, ptr, size);
if (ret < 0)
@@ -2233,6 +2688,7 @@
static void encode_options(struct options *dst, const struct options *src)
{
+ memcpy(dst->version, src->version, VERSION_MAX_LEN);
dst->req_depth = htonl(src->req_depth);
dst->req_size = htonl(src->req_size);
dst->ack_size = htonl(src->ack_size);
@@ -2262,10 +2718,13 @@
dst->simplex = src->simplex; /* byte sized */
dst->rw_mode = src->rw_mode; /* byte sized */
dst->rdma_vector = htonl(src->rdma_vector);
+ dst->tos = src->tos;
+ dst->async = src->async;
}
static void decode_options(struct options *dst, const struct options *src)
{
+ memcpy(dst->version, src->version, VERSION_MAX_LEN);
dst->req_depth = ntohl(src->req_depth);
dst->req_size = ntohl(src->req_size);
dst->ack_size = ntohl(src->ack_size);
@@ -2295,6 +2754,8 @@
dst->simplex = src->simplex; /* byte sized */
dst->rw_mode = src->rw_mode; /* byte sized */
dst->rdma_vector = ntohl(src->rdma_vector);
+ dst->tos = src->tos;
+ dst->async = src->async;
}
static void verify_option_encdec(const struct options *opts)
@@ -2316,6 +2777,25 @@
die("encode/decode check of options struct failed");
}
+static void reset_conn(struct options *opts)
+{
+ struct rds_reset val;
+ int fd;
+ struct sockaddr_in sin;
+
+ sin.sin_family = AF_INET;
+ sin.sin_port = htons(opts->starting_port);
+ sin.sin_addr.s_addr = htonl(opts->receive_addr);
+
+ fd = bound_socket(pf, SOCK_SEQPACKET, 0, &sin);
+
+ val.tos = opts->tos;
+ val.src.s_addr = htonl(opts->receive_addr);
+ val.dst.s_addr = htonl(opts->send_addr);
+ if (setsockopt(fd, sol, RDS_CONN_RESET, &val, sizeof(val)))
+ die_errno("setsockopt RDS_CONN_RESET failed");
+}
+
static int active_parent(struct options *opts, struct soak_control *soak_arr)
{
struct options enc_options;
@@ -2324,6 +2804,11 @@
int fd;
uint8_t ok;
+ if (reset_connection) {
+ reset_conn(opts);
+ return 0;
+ }
+
if (opts->show_params) {
unsigned int k;
@@ -2387,7 +2872,11 @@
* We just tell the peer what options to use.
*/
encode_options(&enc_options, opts);
- peer_send(fd, &enc_options, sizeof(struct options));
+ if (opts->tos || opts->async)
+ peer_send(fd, &enc_options, sizeof(struct options));
+ else
+ peer_send(fd, &enc_options.req_depth,
+ sizeof(struct options_2_0_6));
printf("negotiated options, tasks will start in 2 seconds\n");
ctl = start_children(opts, 1);
@@ -2517,7 +3006,11 @@
/* an extra terminating entry which will be all 0s */
len = (nr_soak + 1) * sizeof(struct soak_control);
soak_arr = mmap(NULL, len, PROT_READ|PROT_WRITE,
+#if defined(__SVR4) && defined(__sun)
+ MAP_ANONYMOUS|MAP_SHARED, -1, 0);
+#else
MAP_ANONYMOUS|MAP_SHARED, 0, 0);
+#endif
if (soak_arr == MAP_FAILED)
die("mmap of %ld soak control structs failed", nr_soak);
@@ -2572,6 +3065,10 @@
OPT_CONNECT_RETRIES,
OPT_USE_CONG_MONITOR,
OPT_PERFDATA,
+ OPT_SHOW_OUTLIERS,
+ OPT_SHOW_HISTOGRAM,
+ OPT_RESET,
+ OPT_ASYNC,
};
static struct option long_options[] = {
@@ -2584,11 +3081,13 @@
{ "send-addr", required_argument, NULL, 's' },
{ "port", required_argument, NULL, 'p' },
{ "time", required_argument, NULL, 'T' },
+{ "tos", required_argument, NULL, 'Q' },
{ "report-cpu", no_argument, NULL, 'c' },
{ "report-summary", no_argument, NULL, 'z' },
{ "rtprio", no_argument, NULL, 'R' },
{ "verify", no_argument, NULL, 'v' },
{ "trace", no_argument, NULL, 'V' },
+{ "lgrpid", required_argument, NULL, 'g' },
{ "rdma-use-once", required_argument, NULL, OPT_RDMA_USE_ONCE },
{ "rdma-use-get-mr", required_argument, NULL, OPT_RDMA_USE_GET_MR },
@@ -2601,6 +3100,10 @@
{ "show-perfdata", no_argument, NULL, OPT_PERFDATA },
{ "connect-retries", required_argument, NULL, OPT_CONNECT_RETRIES },
{ "use-cong-monitor", required_argument, NULL, OPT_USE_CONG_MONITOR },
+{ "show-outliers", required_argument, NULL, OPT_SHOW_OUTLIERS },
+{ "show-histogram", no_argument, NULL, OPT_SHOW_HISTOGRAM },
+{ "reset", no_argument, NULL, OPT_RESET },
+{ "async", no_argument, NULL, OPT_ASYNC },
{ NULL }
};
@@ -2640,6 +3143,8 @@
opts.rdma_use_fence = 1;
opts.rdma_cache_mrs = 0;
+ opts.rdma_use_once = 0;
+ opts.rdma_use_get_mr = 0;
opts.rdma_alignment = 0;
opts.show_params = 0;
@@ -2648,11 +3153,17 @@
opts.simplex = 0;
opts.rw_mode = 0;
opts.rdma_vector = 1;
+ rtt_threshold = ~0U;
+ show_histogram = 0;
+ opts.tos = 0;
+ reset_connection = 0;
+ opts.async = 0;
+ strcpy(opts.version, RDS_VERSION);
while(1) {
int c, index;
- c = getopt_long(argc, argv, "+a:cD:d:hI:M:op:q:Rr:s:t:T:vVz",
+ c = getopt_long(argc, argv, "+a:cD:d:hI:M:op:q:Rr:s:t:T:Q:vVg:z",
long_options, &index);
if (c == -1)
break;
@@ -2702,6 +3213,9 @@
case 'T':
opts.run_time = parse_ull(optarg, (uint32_t)~0);
break;
+ case 'Q':
+ opts.tos = parse_ull(optarg, (uint8_t)~0);
+ break;
case 'z':
opts.summary_only = 1;
break;
@@ -2711,9 +3225,25 @@
case 'V':
opts.tracing = 1;
break;
+ case 'g':
+ lgrp_id = (lgrp_id_t)parse_ull(optarg,
+ (uint32_t)~0);
+ break;
+ case OPT_SHOW_OUTLIERS:
+ rtt_threshold = parse_ull(optarg, ~0U);
+ break;
+ case OPT_SHOW_HISTOGRAM:
+ show_histogram = 1;
+ break;
case OPT_USE_CONG_MONITOR:
opts.use_cong_monitor = parse_ull(optarg, 1);
break;
+ case OPT_RESET:
+ reset_connection = 1;
+ break;
+ case OPT_ASYNC:
+ opts.async = 1;
+ break;
case OPT_RDMA_USE_ONCE:
opts.rdma_use_once = parse_ull(optarg, 1);
break;
@@ -2786,6 +3316,7 @@
if (opts.rdma_size && 0)
opts.rdma_size = (opts.rdma_size + 4095) & ~4095;
+ set_my_lgrp();
opt = opts;
return active_parent(&opts, soak_arr);
}
diff -r 1afa90e87b4b rds.7
--- a/rds.7 Tue Feb 23 22:30:57 2016 -0800
+++ b/rds.7 Tue Feb 23 22:30:58 2016 -0800
@@ -6,6 +6,7 @@
.nf
.B #include <sys/socket.h>
.B #include <netinet/in.h>
+.B #include <sys/rds.h>
.fi
.SH DESCRIPTION
This is an implementation of the RDS socket API. It provides reliable,
@@ -14,18 +15,13 @@
Currently, RDS can be transported over Infiniband, and loopback.
RDS over TCP is disabled, but will be re-enabled in the near future.
.PP
-RDS uses standard
-.B AF_INET
-addresses as described in
-.BR ip (7)
+RDS uses
+.B AF_INET_OFFLOAD address family
to identify end points.
.\"------------------------------------------------------------------
.SS Socket Creation
RDS is still in development and as such does not have a reserved protocol
-family constant. Applications must read the string representation of the
-protocol family value from the
-.B pf_rds
-sysctl parameter file described below.
+family constant. Applications should use AF_INET_OFFLOAD.
.PP
.nf
.B rds_socket = socket(pf_rds, SOCK_SEQPACKET, 0);
@@ -58,9 +54,6 @@
.BR SOL_RDS ).
Just as with the RDS protocol family, an official value has not been
assigned yet, so the kernel will assign a value dynamically.
-The assigned value can be retrieved from the
-.B sol_rds
-sysctl parameter file.
.PP
RDS specific socket options will be described in a separate section
below.
@@ -77,7 +70,7 @@
.PP
For instance, when binding to the address of an Infiniband interface
such as
-.BR ib0 ,
+.BR ibd0 ,
the socket will use the Infiniband transport. If RDS is not able
to associate a transport with the given address, it will return
.BR EADDRNOTAVAIL .
@@ -394,47 +387,6 @@
be delivered in the order they're sent. Messages sent from different
sockets, or to different destinations, may be delivered in any order.
.\"------------------------------------------------------------------
-.SH SYSCTL VALUES
-These parameteres may only be accessed through their files in
-Access through
-.BR sysctl (2)
-is not supported.
-.B pf_rds
-This file contains the string representation of the protocol family
-constant passed to
-.BR socket (2)
-to create a new RDS socket.
-.B sol_rds
-This file contains the string representation of the socket level parameter
-that is passed to
-.BR getsockopt (2)
-and
-.BR setsockopt (2)
-to manipulate RDS socket options.
-.BR max_unacked_bytes " and " max_unacked_packets
-These parameters are used to tune the generation of acknowledgements. By
-default, the system receiving RDS messages does not send back explicit
-acknowledgements unless it transmits a message of its own (in which
-case the ACK is piggybacked onto the outgoing message), or when the sending
-system requests an ACK.
-However, the sender needs to see an ACK from time to time so that it
-can purge old messages from the send queue. The unacked bytes and
-packet counters are used to keep track of how much data has been
-sent without requesting an ACK. The default is to request an acknowledgement
-every 16 packets, or every 16 MB, whichever comes first.
-.BR reconnect_delay_min_ms " and " reconnect_delay_max_ms
-RDS uses host-to-host connections to transport RDS messages (both for the TCP
-and the Infiniband transport). If this connection breaks, RDS will try to
-re-establish the connection. Because this reconnect may be triggered by
-both hosts at the same time and fail, RDS uses a random backoff before attempting
-a reconnect. These two parameters specify the minimum and maximum delay in
-milliseconds. The default values are 1 and 1000, respectively.
.SH SEE ALSO
.BR rds-rdma (7),
.BR socket (2),