6208 add support for timerfd

Reviewed by: Gordon Ross <gwr@nexenta.com>
Approved by: Dan McDonald <danmcd@omniti.com>

28 files changed, 1246 insertions, 20 deletions

diff --git a/usr/src/cmd/devfsadm/misc_link.c b/usr/src/cmd/devfsadm/misc_link.c
index abb133bc6d..bf59fb5e6b 100644
--- a/usr/src/cmd/devfsadm/misc_link.c
+++ b/usr/src/cmd/devfsadm/misc_link.c
@@ -21,7 +21,7 @@
 /*
  * Copyright (c) 1998, 2010, Oracle and/or its affiliates. All rights reserved.
  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
- * Copyright (c) 2012, Joyent, Inc. All rights reserved.
+ * Copyright (c) 2015, Joyent, Inc. All rights reserved.
  */
 
 #include <regex.h>
@@ -126,6 +126,9 @@ static devfsadm_create_t misc_cbt[] = {
 	    "(^kdmouse$)|(^rootprop$)",
 	    TYPE_EXACT | DRV_RE, ILEVEL_0, node_name
 	},
+	{ "pseudo", "ddi_pseudo", "timerfd",
+	    TYPE_EXACT | DRV_EXACT, ILEVEL_0, minor_name
+	},
 	{ "pseudo", "ddi_pseudo", "tod",
 	    TYPE_EXACT | DRV_EXACT, ILEVEL_0, node_name
 	},
diff --git a/usr/src/lib/libc/amd64/Makefile b/usr/src/lib/libc/amd64/Makefile
index b5e54b19fa..dbda6c0c31 100644
--- a/usr/src/lib/libc/amd64/Makefile
+++ b/usr/src/lib/libc/amd64/Makefile
@@ -909,6 +909,7 @@ PORTSYS=			\
 	tasksys.o		\
 	time.o			\
 	time_util.o		\
+	timerfd.o		\
 	ucontext.o		\
 	unlink.o		\
 	ustat.o			\
diff --git a/usr/src/lib/libc/i386/Makefile.com b/usr/src/lib/libc/i386/Makefile.com
index d7e77502f2..25ba0a2743 100644
--- a/usr/src/lib/libc/i386/Makefile.com
+++ b/usr/src/lib/libc/i386/Makefile.com
@@ -949,6 +949,7 @@ PORTSYS=			\
 	tasksys.o		\
 	time.o			\
 	time_util.o		\
+	timerfd.o		\
 	ucontext.o		\
 	unlink.o		\
 	ustat.o			\
diff --git a/usr/src/lib/libc/port/mapfile-vers b/usr/src/lib/libc/port/mapfile-vers
index 017c7c31bc..a0e21e250f 100644
--- a/usr/src/lib/libc/port/mapfile-vers
+++ b/usr/src/lib/libc/port/mapfile-vers
@@ -93,6 +93,13 @@ $if _x86 && _ELF64
 $add amd64
 $endif
 
+SYMBOL_VERSION ILLUMOS_0.16 {	# timerfd
+    protected:
+	timerfd_create;
+	timerfd_gettime;
+	timerfd_settime;
+} ILLUMOS_0.15;
+
 SYMBOL_VERSION ILLUMOS_0.15 {	# epoll(3C)
     protected:
 	epoll_create;
diff --git a/usr/src/lib/libc/port/sys/timerfd.c b/usr/src/lib/libc/port/sys/timerfd.c
new file mode 100644
index 0000000000..cb2e17adf7
--- /dev/null
+++ b/usr/src/lib/libc/port/sys/timerfd.c
@@ -0,0 +1,93 @@
+/*
+ * This file and its contents are supplied under the terms of the
+ * Common Development and Distribution License ("CDDL"), version 1.0.
+ * You may only use this file in accordance with the terms of version
+ * 1.0 of the CDDL.
+ *
+ * A full copy of the text of the CDDL should have accompanied this
+ * source.  A copy of the CDDL is also available via the Internet at
+ * http://www.illumos.org/license/CDDL.
+ */
+
+/*
+ * Copyright (c) 2015, Joyent, Inc.  All rights reserved.
+ */
+
+#include <sys/timerfd.h>
+#include <sys/stat.h>
+#include <unistd.h>
+#include <errno.h>
+#include <fcntl.h>
+
+int
+timerfd_create(int clockid, int flags)
+{
+	int oflags = O_RDWR;
+	int fd;
+
+	if (flags & ~(TFD_NONBLOCK | TFD_CLOEXEC)) {
+		errno = EINVAL;
+		return (-1);
+	}
+
+	if (flags & TFD_NONBLOCK)
+		oflags |= O_NONBLOCK;
+
+	if (flags & TFD_CLOEXEC)
+		oflags |= O_CLOEXEC;
+
+	if ((fd = open("/dev/timerfd", oflags)) < 0)
+		return (-1);
+
+	if (ioctl(fd, TIMERFDIOC_CREATE, clockid) != 0) {
+		(void) close(fd);
+		return (-1);
+	}
+
+	return (fd);
+}
+
+int
+timerfd_settime(int fd, int flags, const struct itimerspec *new_value,
+    struct itimerspec *old_value)
+{
+	timerfd_settime_t st;
+	int rval;
+
+	if (flags & ~(TFD_TIMER_ABSTIME | TFD_TIMER_CANCEL_ON_SET)) {
+		errno = EINVAL;
+		return (-1);
+	}
+
+	st.tfd_settime_flags = flags;
+	st.tfd_settime_value = (uint64_t)(uintptr_t)new_value;
+	st.tfd_settime_ovalue = (uint64_t)(uintptr_t)old_value;
+
+	rval = ioctl(fd, TIMERFDIOC_SETTIME, &st);
+
+	if (rval == -1 && errno == ENOTTY) {
+		/*
+		 * Linux has us return EINVAL when the file descriptor is valid
+		 * but is not a timerfd file descriptor -- and LTP explicitly
+		 * checks this case.
+		 */
+		errno = EINVAL;
+	}
+
+	return (rval);
+}
+
+int
+timerfd_gettime(int fd, struct itimerspec *curr_value)
+{
+	int rval = ioctl(fd, TIMERFDIOC_GETTIME, curr_value);
+
+	if (rval == -1 && errno == ENOTTY) {
+		/*
+		 * See comment in timerfd_settime(), above.
+		 */
+		errno = EINVAL;
+	}
+
+	return (rval);
+}
diff --git a/usr/src/lib/libc/sparc/Makefile.com b/usr/src/lib/libc/sparc/Makefile.com
index dc965fe6ac..94036d831e 100644
--- a/usr/src/lib/libc/sparc/Makefile.com
+++ b/usr/src/lib/libc/sparc/Makefile.com
@@ -983,6 +983,7 @@ PORTSYS=			\
 	tasksys.o		\
 	time.o			\
 	time_util.o		\
+	timerfd.o		\
 	ucontext.o		\
 	unlink.o		\
 	ustat.o			\
diff --git a/usr/src/lib/libc/sparcv9/Makefile.com b/usr/src/lib/libc/sparcv9/Makefile.com
index 415aaf2be2..2156bad20b 100644
--- a/usr/src/lib/libc/sparcv9/Makefile.com
+++ b/usr/src/lib/libc/sparcv9/Makefile.com
@@ -927,6 +927,7 @@ PORTSYS=			\
 	tasksys.o		\
 	time.o			\
 	time_util.o		\
+	timerfd.o		\
 	ucontext.o		\
 	unlink.o		\
 	ustat.o			\
diff --git a/usr/src/man/man3c/Makefile b/usr/src/man/man3c/Makefile
index f6cadebe95..c38d65a57a 100644
--- a/usr/src/man/man3c/Makefile
+++ b/usr/src/man/man3c/Makefile
@@ -473,6 +473,7 @@ MANFILES= 	__fbufsize.3c					\
 	 	timer_delete.3c					\
 	 	timer_settime.3c				\
 	 	timeradd.3c					\
+		timerfd_create.3c				\
 	 	tmpfile.3c					\
 	 	tmpnam.3c					\
 	 	toascii.3c					\
@@ -1211,6 +1212,8 @@ MANLINKS=	FD_CLR.3c				\
 	 	timer_gettime.3c			\
 	 	timerclear.3c				\
 	 	timercmp.3c				\
+		timerfd_gettime.3c			\
+		timerfd_settime.3c			\
 	 	timerisset.3c				\
 	 	timersub.3c				\
 	 	tmpnam_r.3c				\
@@ -2250,6 +2253,9 @@ timercmp.3c				:= LINKSRC = timeradd.3c
 timerisset.3c				:= LINKSRC = timeradd.3c
 timersub.3c				:= LINKSRC = timeradd.3c
 
+timerfd_gettime.3c				:= LINKSRC = timerfd_create.3c
+timerfd_settime.3c				:= LINKSRC = timerfd_create.3c
+
 tempnam.3c				:= LINKSRC = tmpnam.3c
 tmpnam_r.3c				:= LINKSRC = tmpnam.3c
 
diff --git a/usr/src/man/man3c/timerfd_create.3c b/usr/src/man/man3c/timerfd_create.3c
new file mode 100644
index 0000000000..84df47e245
--- /dev/null
+++ b/usr/src/man/man3c/timerfd_create.3c
@@ -0,0 +1,201 @@
+.\"
+.\" This file and its contents are supplied under the terms of the
+.\" Common Development and Distribution License ("CDDL"), version 1.0.
+.\" You may only use this file in accordance with the terms of version
+.\" 1.0 of the CDDL.
+.\"
+.\" A full copy of the text of the CDDL should have accompanied this
+.\" source.  A copy of the CDDL is also available via the Internet at
+.\" http://www.illumos.org/license/CDDL.
+.\"
+.\"
+.\" Copyright (c) 2015, Joyent, Inc. All Rights Reserved.
+.\"
+.Dd Feb 23, 2105
+.Dt TIMERFD 3C
+.Os
+.Sh NAME
+.Nm timerfd_create ,
+.Nm timerfd_settime ,
+.Nm timerfd_gettime
+.Nd create and manipulate timers via a file descriptor interface
+.Sh SYNOPSIS
+.In sys/timerfd.h
+.Ft int
+.Fo timerfd_create
+.Fa "int clockid"
+.Fa "int flags"
+.Fc
+.Ft int
+.Fo timerfd_settime
+.Fa "int fd"
+.Fa "int flags"
+.Fa "const struct itimerspec *restrict value"
+.Fa "struct itimterspec *restrict ovalue"
+.Fc
+.Ft int
+.Fo timerfd_gettime
+.Fa "int fd"
+.Fa "struct itimerspec *value"
+.Fc
+.Sh DESCRIPTION
+These routines create and manipulate timers in which events are associated
+with a file descriptor, allowing for timer-based events to be used
+in file-descriptor based facilities like
+.Xr poll 2 ,
+.Xr port_get 3C
+or
+.Xr epoll_wait 3C .
+The
+.Fn timerfd_create
+function creates a timer with the clock
+type specified by
+.Fa clockid .
+The
+.Sy CLOCK_REALTIME
+and
+.Sy CLOCK_HIGHRES
+clock types, as defined in
+.Xr timer_create 3C ,
+are supported by
+.Fn timerfd_create .
+(Note that
+.Sy CLOCK_MONOTONIC
+may be used as an alias for
+.Sy CLOCK_HIGHRES Ns .)
+.Pp
+The
+.Fa flags
+argument specifies additional parameters for the timer instance, and can have
+any of the following values:
+.Bl -hang -width Ds
+.It Sy TFD_CLOEXEC
+.Bd -filled -compact
+Instance will be closed upon an
+.Xr exec 2 ;
+see
+.Xr open 2 Ns 's
+description of
+.Sy O_CLOEXEC .
+.Ed
+.It Sy TFD_NONBLOCK
+.Bd -filled -compact
+Instance will be set to be non-blocking.  A
+.Xr read 2
+on a
+.Sy timerfd
+instance that has been initialized with
+.Sy TFD_NONBLOCK
+will return
+.Sy EAGAIN
+in lieu of blocking if the
+timer has not expired since the last
+.Fn timerfd_settime
+or successful
+.Fn read .
+.Ed
+.El
+.Pp
+The following operations can be performed upon a
+.Sy timerfd
+instance:
+.Bl -hang -width Ds
+.It Sy read(2)
+.Bd -filled -compact
+Atomically reads and clears the number of timer expirations since the
+last successful
+.Xr read 2
+or
+.Fn timerfd_settime .
+Upon success,
+the number of expirations will be copied into the eight byte buffer
+passed to the system call. If there have been no expirations of the
+timer since the last successful
+.Xr read 2
+or
+.Fn timerfd_sttime ,
+.Xr read 2
+will block until at least the next expiration,
+or return
+.Sy EAGAIN
+if the instance was created with
+.Sy TFD_NONBLOCK .
+Note that if multiple threads are blocked in
+.Xr read 2
+for the same timer, only one of them will return upon
+a single timer expiration.
+.Pp
+If the buffer specified to
+.Xr read 2
+is less than
+eight bytes in length,
+.Sy EINAVL
+will be returned.
+.Ed
+.It Sy poll(2), port_get(3C), epoll_wait(3C)
+.Bd -filled -compact
+Provide notification when the timer expires or has expired in the past without
+a more recent
+.Xr read 2 .
+Note that threads being simultaneously
+blocked in
+.Xr read 2
+and
+.Xr poll 2
+(or equivalents) for the same
+timer constitute an application-level race; on a timer expiration,
+the thread blocked in
+.Xr poll 2
+may or may not return depending on how
+it is scheduled with respect to the thread blocked in
+.Xr read 2 .
+.Ed
+.It Sy timerfd_gettime()
+.Bd -filled -compact
+Returns the amount of time until the next timer expiration, with the
+same functional signature and semantics as
+.Xr timer_gettime 3C .
+.Ed
+.It Sy timerfd_settime()
+.Bd -filled -compact
+Sets or disarms the timer, with the
+same functional signature and semantics as
+.Xr timer_settime 3C .
+.Ed
+.El
+.Sh RETURN VALUES
+Upon succesful completion, a file descriptor associated with the instance
+is returned. Otherwise,
+.Sy -1
+is returned and errno is set to indicate the error.
+.Sh ERRORS
+The
+.Fn timerfd_create()
+function will fail if:
+.Bl -tag -width Er
+.It Er EINAL
+The
+.Fa flags
+are invalid.
+.It Er EMFILE
+There are currently
+.Pf { Sy OPEN_MAX Ns }
+file descriptors open in the calling process.
+.It Er EPERM
+The
+.Fa clock_id ,
+is
+.Sy CLOCK_HIGHRES
+and the
+.Pf { Sy PRIV_PROC_CLOCK_HIGHRES Ns }
+is not asserted in the effective set of the calling process.
+.El
+.Sh SEE ALSO
+.Xr poll 2 ,
+.Xr port_get 3C ,
+.Xr epoll_wait 3C ,
+.Xr timer_create 3C ,
+.Xr timer_gettime 3C ,
+.Xr timer_settime 3C ,
+.Xr privileages 5 ,
+.Xr timerfd 5
diff --git a/usr/src/man/man5/Makefile b/usr/src/man/man5/Makefile
index 4784603013..c30af8e1d1 100644
--- a/usr/src/man/man5/Makefile
+++ b/usr/src/man/man5/Makefile
@@ -14,6 +14,7 @@
 # Copyright (c) 2012 by Delphix. All rights reserved.
 # Copyright 2014 Nexenta Systems, Inc.
 # Copyright 2014 Garrett D'Amore <garrett@damore.org>
+# Copyright (c) 2015, Joyent, Inc.  All rights reserved.
 #
 
 include		$(SRC)/Makefile.master
@@ -121,6 +122,7 @@ MANFILES=	Intro.5			\
 		tecla.5			\
 		term.5			\
 		threads.5		\
+		timerfd.5		\
 		trusted_extensions.5	\
 		vgrindefs.5		\
 		zones.5			\
diff --git a/usr/src/man/man5/timerfd.5 b/usr/src/man/man5/timerfd.5
new file mode 100644
index 0000000000..3229095b49
--- /dev/null
+++ b/usr/src/man/man5/timerfd.5
@@ -0,0 +1,47 @@
+.\"
+.\" This file and its contents are supplied under the terms of the
+.\" Common Development and Distribution License ("CDDL"), version 1.0.
+.\" You may only use this file in accordance with the terms of version
+.\" 1.0 of the CDDL.
+.\"
+.\" A full copy of the text of the CDDL should have accompanied this
+.\" source.  A copy of the CDDL is also available via the Internet at
+.\" http://www.illumos.org/license/CDDL.
+.\"
+.\"
+.\" Copyright (c) 2015, Joyent, Inc. All Rights Reserved.
+.\"
+.Dd Feb 23, 2015
+.Dt TIMERFD 5
+.Os
+.Sh NAME
+.Nm timerfd
+.Nd Linux-compatible timer notification facility
+.Sh SYNOPSIS
+.In sys/timerfd.h
+.Sh DESCRIPTION
+.Nm
+is a Linux-borne facility for creating POSIX timers and
+receiving their subsequent events via a file descriptor.
+The facility itself is arguably unnecessary:
+portable code can either use the timeout value present in
+.Xr poll 2 /
+.Xr port_get 3C
+or -- if this is deemed of unacceptably poor resolution -- create a POSIX timer
+via
+.Xr timer_create 3C
+and use the resulting signal to induce an
+.Sy EINTR
+to polling threads.  (For code that need not be
+portable, the
+.Sy SIGEV_PORT
+signal notification allows for explicit, event-oriented timer notification to be
+sent to a specified port; see
+.Xr signal.h 3HEAD
+for details.) This facility therefore exists only to accommodate Linux-borne
+applications and binaries; it is compatible with its Linux antecedent in both
+binary interface and in semantics.
+.Sh SEE ALSO
+.Xr timerfd_create 3C ,
+.Xr timerfd_gettime 3C ,
+.Xr timerfd_settime 3C
diff --git a/usr/src/pkg/manifests/SUNWcs.mf b/usr/src/pkg/manifests/SUNWcs.mf
index 85285ae7cd..47923f1f78 100644
--- a/usr/src/pkg/manifests/SUNWcs.mf
+++ b/usr/src/pkg/manifests/SUNWcs.mf
@@ -323,6 +323,7 @@ driver name=logindmux
 driver name=ptm clone_perms="ptmx 0666 root sys"
 driver name=pts perms="* 0644 root sys" perms="0 0620 root tty" \
     perms="1 0620 root tty" perms="2 0620 root tty" perms="3 0620 root tty"
+driver name=timerfd perms="* 0666 root sys"
 file path=etc/.login group=sys preserve=renamenew
 file path=etc/cron.d/.proto group=sys mode=0744
 file path=etc/cron.d/at.deny group=sys preserve=true
@@ -868,6 +869,7 @@ file path=usr/kernel/drv/$(ARCH64)/ksyms group=sys
 file path=usr/kernel/drv/$(ARCH64)/logindmux group=sys
 file path=usr/kernel/drv/$(ARCH64)/ptm group=sys
 file path=usr/kernel/drv/$(ARCH64)/pts group=sys
+file path=usr/kernel/drv/$(ARCH64)/timerfd group=sys
 $(i386_ONLY)file path=usr/kernel/drv/dump group=sys
 file path=usr/kernel/drv/dump.conf group=sys
 $(i386_ONLY)file path=usr/kernel/drv/eventfd group=sys
@@ -884,6 +886,8 @@ $(i386_ONLY)file path=usr/kernel/drv/ptm group=sys
 file path=usr/kernel/drv/ptm.conf group=sys
 $(i386_ONLY)file path=usr/kernel/drv/pts group=sys
 file path=usr/kernel/drv/pts.conf group=sys
+$(i386_ONLY)file path=usr/kernel/drv/timerfd group=sys
+file path=usr/kernel/drv/timerfd.conf group=sys
 file path=usr/kernel/exec/$(ARCH64)/javaexec group=sys mode=0755
 file path=usr/kernel/exec/$(ARCH64)/shbinexec group=sys mode=0755
 $(i386_ONLY)file path=usr/kernel/exec/javaexec group=sys mode=0755
diff --git a/usr/src/pkg/manifests/system-header.mf b/usr/src/pkg/manifests/system-header.mf
index b72d713cd8..08f0b19416 100644
--- a/usr/src/pkg/manifests/system-header.mf
+++ b/usr/src/pkg/manifests/system-header.mf
@@ -1516,6 +1516,7 @@ file path=usr/include/sys/time_impl.h
 file path=usr/include/sys/time_std_impl.h
 file path=usr/include/sys/timeb.h
 file path=usr/include/sys/timer.h
+file path=usr/include/sys/timerfd.h
 file path=usr/include/sys/times.h
 file path=usr/include/sys/timex.h
 file path=usr/include/sys/timod.h
diff --git a/usr/src/pkg/manifests/system-library.man3c.inc b/usr/src/pkg/manifests/system-library.man3c.inc
index 27268505b3..ae061edac9 100644
--- a/usr/src/pkg/manifests/system-library.man3c.inc
+++ b/usr/src/pkg/manifests/system-library.man3c.inc
@@ -468,6 +468,7 @@ file path=usr/share/man/man3c/timer_create.3c
 file path=usr/share/man/man3c/timer_delete.3c
 file path=usr/share/man/man3c/timer_settime.3c
 file path=usr/share/man/man3c/timeradd.3c
+file path=usr/share/man/man3c/timerfd_create.3c
 file path=usr/share/man/man3c/tmpfile.3c
 file path=usr/share/man/man3c/tmpnam.3c
 file path=usr/share/man/man3c/toascii.3c
@@ -1281,6 +1282,8 @@ link path=usr/share/man/man3c/timer_getoverrun.3c target=timer_settime.3c
 link path=usr/share/man/man3c/timer_gettime.3c target=timer_settime.3c
 link path=usr/share/man/man3c/timerclear.3c target=timeradd.3c
 link path=usr/share/man/man3c/timercmp.3c target=timeradd.3c
+link path=usr/share/man/man3c/timerfd_gettime.3c target=timerfd_create.3c
+link path=usr/share/man/man3c/timerfd_settime.3c target=timerfd_create.3c
 link path=usr/share/man/man3c/timerisset.3c target=timeradd.3c
 link path=usr/share/man/man3c/timersub.3c target=timeradd.3c
 link path=usr/share/man/man3c/tmpnam_r.3c target=tmpnam.3c
diff --git a/usr/src/pkg/manifests/system-library.man5.inc b/usr/src/pkg/manifests/system-library.man5.inc
index fd222bcec4..63d883e984 100644
--- a/usr/src/pkg/manifests/system-library.man5.inc
+++ b/usr/src/pkg/manifests/system-library.man5.inc
@@ -61,6 +61,7 @@ file path=usr/share/man/man5/pkcs11_softtoken.5
 file path=usr/share/man/man5/pkcs11_tpm.5
 file path=usr/share/man/man5/regex.5
 file path=usr/share/man/man5/regexp.5
+file path=usr/share/man/man5/timerfd.5
 file path=usr/share/man/man5/threads.5
 link path=usr/share/man/man5/advance.5 target=regexp.5
 link path=usr/share/man/man5/compile.5 target=regexp.5
diff --git a/usr/src/uts/common/Makefile.files b/usr/src/uts/common/Makefile.files
index 3421b83719..cd9da11ac8 100644
--- a/usr/src/uts/common/Makefile.files
+++ b/usr/src/uts/common/Makefile.files
@@ -1032,6 +1032,8 @@ GEN_DRV_OBJS += gen_drv.o
 
 TCLIENT_OBJS +=	tclient.o
 
+TIMERFD_OBJS +=	timerfd.o
+
 TPHCI_OBJS += tphci.o
 
 TVHCI_OBJS += tvhci.o
diff --git a/usr/src/uts/common/io/timerfd.c b/usr/src/uts/common/io/timerfd.c
new file mode 100644
index 0000000000..29eea9b24a
--- /dev/null
+++ b/usr/src/uts/common/io/timerfd.c
@@ -0,0 +1,586 @@
+/*
+ * This file and its contents are supplied under the terms of the
+ * Common Development and Distribution License ("CDDL"), version 1.0.
+ * You may only use this file in accordance with the terms of version
+ * 1.0 of the CDDL.
+ *
+ * A full copy of the text of the CDDL should have accompanied this
+ * source.  A copy of the CDDL is also available via the Internet at
+ * http://www.illumos.org/license/CDDL.
+ */
+
+/*
+ * Copyright (c) 2015 Joyent, Inc.  All rights reserved.
+ */
+
+/*
+ * Support for the timerfd facility, a Linux-borne facility that allows
+ * POSIX.1b timers to be created and manipulated via a file descriptor
+ * interface.
+ */
+
+#include <sys/ddi.h>
+#include <sys/sunddi.h>
+#include <sys/timerfd.h>
+#include <sys/conf.h>
+#include <sys/vmem.h>
+#include <sys/sysmacros.h>
+#include <sys/filio.h>
+#include <sys/stat.h>
+#include <sys/file.h>
+#include <sys/timer.h>
+
+struct timerfd_state;
+typedef struct timerfd_state timerfd_state_t;
+
+struct timerfd_state {
+	kmutex_t tfd_lock;			/* lock protecting state */
+	kcondvar_t tfd_cv;			/* condvar */
+	pollhead_t tfd_pollhd;			/* poll head */
+	uint64_t tfd_fired;			/* # of times fired */
+	itimer_t tfd_itimer;			/* underlying itimer */
+	timerfd_state_t *tfd_next;		/* next state on global list */
+};
+
+/*
+ * Internal global variables.
+ */
+static kmutex_t		timerfd_lock;		/* lock protecting state */
+static dev_info_t	*timerfd_devi;		/* device info */
+static vmem_t		*timerfd_minor;		/* minor number arena */
+static void		*timerfd_softstate;	/* softstate pointer */
+static timerfd_state_t	*timerfd_state;		/* global list of state */
+
+static itimer_t *
+timerfd_itimer_lock(timerfd_state_t *state)
+{
+	itimer_t *it = &state->tfd_itimer;
+
+	mutex_enter(&state->tfd_lock);
+
+	while (it->it_lock & ITLK_LOCKED) {
+		it->it_blockers++;
+		cv_wait(&it->it_cv, &state->tfd_lock);
+		it->it_blockers--;
+	}
+
+	it->it_lock |= ITLK_LOCKED;
+
+	mutex_exit(&state->tfd_lock);
+
+	return (it);
+}
+
+static void
+timerfd_itimer_unlock(timerfd_state_t *state, itimer_t *it)
+{
+	VERIFY(it == &state->tfd_itimer);
+	VERIFY(it->it_lock & ITLK_LOCKED);
+
+	mutex_enter(&state->tfd_lock);
+
+	it->it_lock &= ~ITLK_LOCKED;
+
+	if (it->it_blockers)
+		cv_signal(&it->it_cv);
+
+	mutex_exit(&state->tfd_lock);
+}
+
+static void
+timerfd_fire(itimer_t *it)
+{
+	timerfd_state_t *state = it->it_frontend;
+	uint64_t oval;
+
+	mutex_enter(&state->tfd_lock);
+	oval = state->tfd_fired++;
+	mutex_exit(&state->tfd_lock);
+
+	if (oval == 0) {
+		cv_broadcast(&state->tfd_cv);
+		pollwakeup(&state->tfd_pollhd, POLLRDNORM | POLLIN);
+	}
+}
+
+/*ARGSUSED*/
+static int
+timerfd_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
+{
+	timerfd_state_t *state;
+	major_t major = getemajor(*devp);
+	minor_t minor = getminor(*devp);
+
+	if (minor != TIMERFDMNRN_TIMERFD)
+		return (ENXIO);
+
+	mutex_enter(&timerfd_lock);
+
+	minor = (minor_t)(uintptr_t)vmem_alloc(timerfd_minor, 1,
+	    VM_BESTFIT | VM_SLEEP);
+
+	if (ddi_soft_state_zalloc(timerfd_softstate, minor) != DDI_SUCCESS) {
+		vmem_free(timerfd_minor, (void *)(uintptr_t)minor, 1);
+		mutex_exit(&timerfd_lock);
+		return (NULL);
+	}
+
+	state = ddi_get_soft_state(timerfd_softstate, minor);
+	*devp = makedevice(major, minor);
+
+	state->tfd_next = timerfd_state;
+	timerfd_state = state;
+
+	mutex_exit(&timerfd_lock);
+
+	return (0);
+}
+
+/*ARGSUSED*/
+static int
+timerfd_read(dev_t dev, uio_t *uio, cred_t *cr)
+{
+	timerfd_state_t *state;
+	minor_t minor = getminor(dev);
+	uint64_t val;
+	int err;
+
+	if (uio->uio_resid < sizeof (val))
+		return (EINVAL);
+
+	state = ddi_get_soft_state(timerfd_softstate, minor);
+
+	mutex_enter(&state->tfd_lock);
+
+	while (state->tfd_fired == 0) {
+		if (uio->uio_fmode & (FNDELAY|FNONBLOCK)) {
+			mutex_exit(&state->tfd_lock);
+			return (EAGAIN);
+		}
+
+		if (!cv_wait_sig_swap(&state->tfd_cv, &state->tfd_lock)) {
+			mutex_exit(&state->tfd_lock);
+			return (EINTR);
+		}
+	}
+
+	/*
+	 * Our tfd_fired is non-zero; slurp its value and then clear it.
+	 */
+	val = state->tfd_fired;
+	state->tfd_fired = 0;
+	mutex_exit(&state->tfd_lock);
+
+	err = uiomove(&val, sizeof (val), UIO_READ, uio);
+
+	return (err);
+}
+
+/*ARGSUSED*/
+static int
+timerfd_poll(dev_t dev, short events, int anyyet, short *reventsp,
+    struct pollhead **phpp)
+{
+	timerfd_state_t *state;
+	minor_t minor = getminor(dev);
+	short revents = 0;
+
+	state = ddi_get_soft_state(timerfd_softstate, minor);
+
+	mutex_enter(&state->tfd_lock);
+
+	if (state->tfd_fired > 0)
+		revents |= POLLRDNORM | POLLIN;
+
+	if (!(*reventsp = revents & events) && !anyyet)
+		*phpp = &state->tfd_pollhd;
+
+	mutex_exit(&state->tfd_lock);
+
+	return (0);
+}
+
+static int
+timerfd_copyin(uintptr_t addr, itimerspec_t *dest)
+{
+	if (get_udatamodel() == DATAMODEL_NATIVE) {
+		if (copyin((void *)addr, dest, sizeof (itimerspec_t)) != 0)
+			return (EFAULT);
+	} else {
+		itimerspec32_t dest32;
+
+		if (copyin((void *)addr, &dest32, sizeof (itimerspec32_t)) != 0)
+			return (EFAULT);
+
+		ITIMERSPEC32_TO_ITIMERSPEC(dest, &dest32);
+	}
+
+	if (itimerspecfix(&dest->it_value) ||
+	    (itimerspecfix(&dest->it_interval) &&
+	    timerspecisset(&dest->it_value))) {
+		return (EINVAL);
+	}
+
+	return (0);
+}
+
+static int
+timerfd_copyout(itimerspec_t *src, uintptr_t addr)
+{
+	if (get_udatamodel() == DATAMODEL_NATIVE) {
+		if (copyout(src, (void *)addr, sizeof (itimerspec_t)) != 0)
+			return (EFAULT);
+	} else {
+		itimerspec32_t src32;
+
+		if (ITIMERSPEC_OVERFLOW(src))
+			return (EOVERFLOW);
+
+		ITIMERSPEC_TO_ITIMERSPEC32(&src32, src);
+
+		if (copyout(&src32, (void *)addr, sizeof (itimerspec32_t)) != 0)
+			return (EFAULT);
+	}
+
+	return (0);
+}
+
+/*ARGSUSED*/
+static int
+timerfd_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
+{
+	itimerspec_t when, oval;
+	timerfd_state_t *state;
+	minor_t minor = getminor(dev);
+	int err;
+	itimer_t *it;
+
+	state = ddi_get_soft_state(timerfd_softstate, minor);
+
+	switch (cmd) {
+	case TIMERFDIOC_CREATE: {
+		if (arg == TIMERFD_MONOTONIC)
+			arg = CLOCK_MONOTONIC;
+
+		it = timerfd_itimer_lock(state);
+
+		if (it->it_backend != NULL) {
+			timerfd_itimer_unlock(state, it);
+			return (EEXIST);
+		}
+
+		if ((it->it_backend = clock_get_backend(arg)) == NULL) {
+			timerfd_itimer_unlock(state, it);
+			return (EINVAL);
+		}
+
+		/*
+		 * We need to provide a proc structure only for purposes
+		 * of locking CLOCK_REALTIME-based timers -- it is safe to
+		 * provide p0 here.
+		 */
+		it->it_proc = &p0;
+
+		err = it->it_backend->clk_timer_create(it, timerfd_fire);
+
+		if (err != 0) {
+			it->it_backend = NULL;
+			timerfd_itimer_unlock(state, it);
+			return (err);
+		}
+
+		it->it_frontend = state;
+		timerfd_itimer_unlock(state, it);
+
+		return (0);
+	}
+
+	case TIMERFDIOC_GETTIME: {
+		it = timerfd_itimer_lock(state);
+
+		if (it->it_backend == NULL) {
+			timerfd_itimer_unlock(state, it);
+			return (ENODEV);
+		}
+
+		err = it->it_backend->clk_timer_gettime(it, &when);
+		timerfd_itimer_unlock(state, it);
+
+		if (err != 0)
+			return (err);
+
+		if ((err = timerfd_copyout(&when, arg)) != 0)
+			return (err);
+
+		return (0);
+	}
+
+	case TIMERFDIOC_SETTIME: {
+		timerfd_settime_t st;
+
+		if (copyin((void *)arg, &st, sizeof (st)) != 0)
+			return (EFAULT);
+
+		if ((err = timerfd_copyin(st.tfd_settime_value, &when)) != 0)
+			return (err);
+
+		it = timerfd_itimer_lock(state);
+
+		if (it->it_backend == NULL) {
+			timerfd_itimer_unlock(state, it);
+			return (ENODEV);
+		}
+
+		if (st.tfd_settime_ovalue != NULL) {
+			err = it->it_backend->clk_timer_gettime(it, &oval);
+
+			if (err != 0) {
+				timerfd_itimer_unlock(state, it);
+				return (err);
+			}
+		}
+
+		/*
+		 * Before we set the time, we're going to clear tfd_fired.
+		 * This can potentially race with the (old) timer firing, but
+		 * the window is deceptively difficult to close:  if we were
+		 * to simply clear tfd_fired after the call to the backend
+		 * returned, we would run the risk of plowing a firing of the
+		 * new timer.  Ultimately, the race can only be resolved by
+		 * the backend, which would likely need to be extended with a
+		 * function to call back into when the timer is between states
+		 * (that is, after the timer can no longer fire with the old
+		 * timer value, but before it can fire with the new one).
+		 * This is straightforward enough for backends that set a
+		 * timer's value by deleting the old one and adding the new
+		 * one, but for those that modify the timer value in place
+		 * (e.g., cyclics), the required serialization is necessarily
+		 * delicate:  the function would have to be callable from
+		 * arbitrary interrupt context.  While implementing all of
+		 * this is possible, it does not (for the moment) seem worth
+		 * it: if the timer is firing at essentially the same moment
+		 * that it's being reprogrammed, there is a higher-level race
+		 * with respect to timerfd usage that the progam itself will
+		 * have to properly resolve -- and it seems reasonable to
+		 * simply allow the program to resolve it in this case.
+		 */
+		mutex_enter(&state->tfd_lock);
+		state->tfd_fired = 0;
+		mutex_exit(&state->tfd_lock);
+
+		err = it->it_backend->clk_timer_settime(it,
+		    st.tfd_settime_flags & TFD_TIMER_ABSTIME ?
+		    TIMER_ABSTIME : TIMER_RELTIME, &when);
+		timerfd_itimer_unlock(state, it);
+
+		if (err != 0 || st.tfd_settime_ovalue == NULL)
+			return (err);
+
+		if ((err = timerfd_copyout(&oval, st.tfd_settime_ovalue)) != 0)
+			return (err);
+
+		return (0);
+	}
+
+	default:
+		break;
+	}
+
+	return (ENOTTY);
+}
+
+/*ARGSUSED*/
+static int
+timerfd_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
+{
+	timerfd_state_t *state, **sp;
+	itimer_t *it;
+	minor_t minor = getminor(dev);
+
+	state = ddi_get_soft_state(timerfd_softstate, minor);
+
+	if (state->tfd_pollhd.ph_list != NULL) {
+		pollwakeup(&state->tfd_pollhd, POLLERR);
+		pollhead_clean(&state->tfd_pollhd);
+	}
+
+	/*
+	 * No one can get to this timer; we don't need to lock it -- we can
+	 * just call on the backend to delete it.
+	 */
+	it = &state->tfd_itimer;
+
+	if (it->it_backend != NULL)
+		it->it_backend->clk_timer_delete(it);
+
+	mutex_enter(&timerfd_lock);
+
+	/*
+	 * Remove our state from our global list.
+	 */
+	for (sp = &timerfd_state; *sp != state; sp = &((*sp)->tfd_next))
+		VERIFY(*sp != NULL);
+
+	*sp = (*sp)->tfd_next;
+
+	ddi_soft_state_free(timerfd_softstate, minor);
+	vmem_free(timerfd_minor, (void *)(uintptr_t)minor, 1);
+
+	mutex_exit(&timerfd_lock);
+
+	return (0);
+}
+
+static int
+timerfd_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
+{
+	switch (cmd) {
+	case DDI_ATTACH:
+		break;
+
+	case DDI_RESUME:
+		return (DDI_SUCCESS);
+
+	default:
+		return (DDI_FAILURE);
+	}
+
+	mutex_enter(&timerfd_lock);
+
+	if (ddi_soft_state_init(&timerfd_softstate,
+	    sizeof (timerfd_state_t), 0) != 0) {
+		cmn_err(CE_NOTE, "/dev/timerfd failed to create soft state");
+		mutex_exit(&timerfd_lock);
+		return (DDI_FAILURE);
+	}
+
+	if (ddi_create_minor_node(devi, "timerfd", S_IFCHR,
+	    TIMERFDMNRN_TIMERFD, DDI_PSEUDO, NULL) == DDI_FAILURE) {
+		cmn_err(CE_NOTE, "/dev/timerfd couldn't create minor node");
+		ddi_soft_state_fini(&timerfd_softstate);
+		mutex_exit(&timerfd_lock);
+		return (DDI_FAILURE);
+	}
+
+	ddi_report_dev(devi);
+	timerfd_devi = devi;
+
+	timerfd_minor = vmem_create("timerfd_minor", (void *)TIMERFDMNRN_CLONE,
+	    UINT32_MAX - TIMERFDMNRN_CLONE, 1, NULL, NULL, NULL, 0,
+	    VM_SLEEP | VMC_IDENTIFIER);
+
+	mutex_exit(&timerfd_lock);
+
+	return (DDI_SUCCESS);
+}
+
+/*ARGSUSED*/
+static int
+timerfd_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
+{
+	switch (cmd) {
+	case DDI_DETACH:
+		break;
+
+	case DDI_SUSPEND:
+		return (DDI_SUCCESS);
+
+	default:
+		return (DDI_FAILURE);
+	}
+
+	mutex_enter(&timerfd_lock);
+	vmem_destroy(timerfd_minor);
+
+	ddi_remove_minor_node(timerfd_devi, NULL);
+	timerfd_devi = NULL;
+
+	ddi_soft_state_fini(&timerfd_softstate);
+	mutex_exit(&timerfd_lock);
+
+	return (DDI_SUCCESS);
+}
+
+/*ARGSUSED*/
+static int
+timerfd_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
+{
+	int error;
+
+	switch (infocmd) {
+	case DDI_INFO_DEVT2DEVINFO:
+		*result = (void *)timerfd_devi;
+		error = DDI_SUCCESS;
+		break;
+	case DDI_INFO_DEVT2INSTANCE:
+		*result = (void *)0;
+		error = DDI_SUCCESS;
+		break;
+	default:
+		error = DDI_FAILURE;
+	}
+	return (error);
+}
+
+static struct cb_ops timerfd_cb_ops = {
+	timerfd_open,		/* open */
+	timerfd_close,		/* close */
+	nulldev,		/* strategy */
+	nulldev,		/* print */
+	nodev,			/* dump */
+	timerfd_read,		/* read */
+	nodev,			/* write */
+	timerfd_ioctl,		/* ioctl */
+	nodev,			/* devmap */
+	nodev,			/* mmap */
+	nodev,			/* segmap */
+	timerfd_poll,		/* poll */
+	ddi_prop_op,		/* cb_prop_op */
+	0,			/* streamtab  */
+	D_NEW | D_MP		/* Driver compatibility flag */
+};
+
+static struct dev_ops timerfd_ops = {
+	DEVO_REV,		/* devo_rev */
+	0,			/* refcnt */
+	timerfd_info,		/* get_dev_info */
+	nulldev,		/* identify */
+	nulldev,		/* probe */
+	timerfd_attach,		/* attach */
+	timerfd_detach,		/* detach */
+	nodev,			/* reset */
+	&timerfd_cb_ops,	/* driver operations */
+	NULL,			/* bus operations */
+	nodev,			/* dev power */
+	ddi_quiesce_not_needed,	/* quiesce */
+};
+
+static struct modldrv modldrv = {
+	&mod_driverops,		/* module type (this is a pseudo driver) */
+	"timerfd support",	/* name of module */
+	&timerfd_ops,		/* driver ops */
+};
+
+static struct modlinkage modlinkage = {
+	MODREV_1,
+	(void *)&modldrv,
+	NULL
+};
+
+int
+_init(void)
+{
+	return (mod_install(&modlinkage));
+}
+
+int
+_info(struct modinfo *modinfop)
+{
+	return (mod_info(&modlinkage, modinfop));
+}
+
+int
+_fini(void)
+{
+	return (mod_remove(&modlinkage));
+}
diff --git a/usr/src/uts/common/io/timerfd.conf b/usr/src/uts/common/io/timerfd.conf
new file mode 100644
index 0000000000..c6ad86d051
--- /dev/null
+++ b/usr/src/uts/common/io/timerfd.conf
@@ -0,0 +1,16 @@
+#
+# This file and its contents are supplied under the terms of the
+# Common Development and Distribution License ("CDDL"), version 1.0.
+# You may only use this file in accordance with the terms of version
+# 1.0 of the CDDL.
+#
+# A full copy of the text of the CDDL should have accompanied this
+# source.  A copy of the CDDL is also available via the Internet at
+# http://www.illumos.org/license/CDDL.
+#
+
+#
+# Copyright (c) 2015 Joyent, Inc.  All rights reserved.
+#
+
+name="timerfd" parent="pseudo" instance=0;
diff --git a/usr/src/uts/common/os/clock_highres.c b/usr/src/uts/common/os/clock_highres.c
index bcdf20c0bd..805813037d 100644
--- a/usr/src/uts/common/os/clock_highres.c
+++ b/usr/src/uts/common/os/clock_highres.c
@@ -25,7 +25,7 @@
  */
 
 /*
- * Copyright (c) 2012, Joyent Inc. All rights reserved.
+ * Copyright (c) 2015, Joyent Inc. All rights reserved.
  */
 
 #include <sys/timer.h>
@@ -66,7 +66,7 @@ clock_highres_getres(timespec_t *ts)
 
 /*ARGSUSED*/
 static int
-clock_highres_timer_create(itimer_t *it, struct sigevent *ev)
+clock_highres_timer_create(itimer_t *it, void (*fire)(itimer_t *))
 {
 	/*
 	 * CLOCK_HIGHRES timers of sufficiently high resolution can deny
@@ -80,6 +80,7 @@ clock_highres_timer_create(itimer_t *it, struct sigevent *ev)
 	}
 
 	it->it_arg = kmem_zalloc(sizeof (cyclic_id_t), KM_SLEEP);
+	it->it_fire = fire;
 
 	return (0);
 }
@@ -95,7 +96,7 @@ clock_highres_fire(void *arg)
 		old = *addr;
 	} while (atomic_cas_64((uint64_t *)addr, old, new) != old);
 
-	timer_fire(it);
+	it->it_fire(it);
 }
 
 static int
diff --git a/usr/src/uts/common/os/clock_realtime.c b/usr/src/uts/common/os/clock_realtime.c
index ef3383fb28..4a75984b23 100644
--- a/usr/src/uts/common/os/clock_realtime.c
+++ b/usr/src/uts/common/os/clock_realtime.c
@@ -24,7 +24,9 @@
  * Use is subject to license terms.
  */
 
-#pragma ident	"%Z%%M%	%I%	%E% SMI"
+/*
+ * Copyright (c) 2015, Joyent Inc. All rights reserved.
+ */
 
 #include <sys/timer.h>
 #include <sys/systm.h>
@@ -80,8 +82,7 @@ clock_realtime_fire(void *arg)
 	/*
 	 * First call into the timer subsystem to get the signal going.
 	 */
-	timer_fire(it);
-
+	it->it_fire(it);
 	val = &it->it_itime.it_value;
 	interval = &it->it_itime.it_interval;
 
@@ -171,9 +172,10 @@ clock_realtime_fire_first(void *arg)
 
 /*ARGSUSED*/
 static int
-clock_realtime_timer_create(itimer_t *it, struct sigevent *ev)
+clock_realtime_timer_create(itimer_t *it, void (*fire)(itimer_t *))
 {
 	it->it_arg = kmem_zalloc(sizeof (timeout_id_t), KM_SLEEP);
+	it->it_fire = fire;
 
 	return (0);
 }
@@ -184,7 +186,7 @@ clock_realtime_timer_settime(itimer_t *it, int flags,
 {
 	timeout_id_t tid, *tidp = it->it_arg;
 	timespec_t now;
-	proc_t *p = curproc;
+	proc_t *p = it->it_proc;
 	clock_t ticks;
 
 	gethrestime(&now);
@@ -246,7 +248,7 @@ static int
 clock_realtime_timer_gettime(itimer_t *it, struct itimerspec *when)
 {
 	timespec_t now;
-	proc_t *p = curproc;
+	proc_t *p = it->it_proc;
 
 	/*
 	 * We always keep it_itime up to date, so we just need to snapshot
@@ -276,7 +278,7 @@ clock_realtime_timer_gettime(itimer_t *it, struct itimerspec *when)
 static int
 clock_realtime_timer_delete(itimer_t *it)
 {
-	proc_t *p = curproc;
+	proc_t *p = it->it_proc;
 	timeout_id_t tid, *tidp = it->it_arg;
 
 	mutex_enter(&p->p_lock);
diff --git a/usr/src/uts/common/os/timer.c b/usr/src/uts/common/os/timer.c
index 8559d8736c..b25a6cbcf1 100644
--- a/usr/src/uts/common/os/timer.c
+++ b/usr/src/uts/common/os/timer.c
@@ -269,6 +269,15 @@ clock_add_backend(clockid_t clock, clock_backend_t *backend)
 	clock_backend[clock] = backend;
 }
 
+clock_backend_t *
+clock_get_backend(clockid_t clock)
+{
+	if (clock < 0 || clock >= CLOCK_MAX)
+		return (NULL);
+
+	return (clock_backend[clock]);
+}
+
 int
 clock_settime(clockid_t clock, timespec_t *tp)
 {
@@ -398,7 +407,7 @@ timer_signal(sigqueue_t *sigq)
 /*
  * This routine is called from the clock backend.
  */
-void
+static void
 timer_fire(itimer_t *it)
 {
 	proc_t *p;
@@ -672,7 +681,7 @@ timer_create(clockid_t clock, struct sigevent *evp, timer_t *tid)
 	 * Call on the backend to verify the event argument (or return
 	 * EINVAL if this clock type does not support timers).
 	 */
-	if ((error = backend->clk_timer_create(it, &ev)) != 0)
+	if ((error = backend->clk_timer_create(it, timer_fire)) != 0)
 		goto err;
 
 	it->it_lwp = ttolwp(curthread);
diff --git a/usr/src/uts/common/sys/Makefile b/usr/src/uts/common/sys/Makefile
index c2bf2f0483..39288d5cc0 100644
--- a/usr/src/uts/common/sys/Makefile
+++ b/usr/src/uts/common/sys/Makefile
@@ -587,6 +587,7 @@ CHKHDRS=			\
 	time_std_impl.h		\
 	timeb.h			\
 	timer.h			\
+	timerfd.h		\
 	times.h			\
 	timex.h			\
 	timod.h			\
diff --git a/usr/src/uts/common/sys/timer.h b/usr/src/uts/common/sys/timer.h
index 604ddf5d83..ec349c962f 100644
--- a/usr/src/uts/common/sys/timer.h
+++ b/usr/src/uts/common/sys/timer.h
@@ -24,6 +24,10 @@
  * Use is subject to license terms.
  */
 
+/*
+ * Copyright (c) 2015, Joyent, Inc. All rights reserved.
+ */
+
 #ifndef	_SYS_TIMER_H
 #define	_SYS_TIMER_H
 
@@ -55,32 +59,45 @@ extern	int	timer_max;		/* patchable via /etc/system */
 
 struct clock_backend;
 
-typedef struct itimer {
+struct itimer;
+typedef struct itimer itimer_t;
+
+struct itimer {
 	itimerspec_t	it_itime;
 	hrtime_t	it_hrtime;
 	ushort_t	it_flags;
 	ushort_t	it_lock;
-	void		*it_arg;
-	sigqueue_t	*it_sigq;
-	klwp_t		*it_lwp;
+	void		*it_arg;	/* clock backend-specific data */
 	struct proc	*it_proc;
+	union {
+		struct {
+			sigqueue_t	*__it_sigq;
+			klwp_t		*__it_lwp;
+		} __proc;
+		void *__it_frontend;
+	} __data;			/* timer frontend-specific data */
 	kcondvar_t	it_cv;
 	int		it_blockers;
 	int		it_pending;
 	int		it_overrun;
 	struct clock_backend *it_backend;
+	void		(*it_fire)(itimer_t *);
 	kmutex_t	it_mutex;
 	void		*it_portev;	/* port_kevent_t pointer */
 	void		*it_portsrc;	/* port_source_t pointer */
 	int		it_portfd;	/* port file descriptor */
-} itimer_t;
+};
+
+#define	it_sigq		__data.__proc.__it_sigq
+#define	it_lwp		__data.__proc.__it_lwp
+#define	it_frontend	__data.__it_frontend
 
 typedef struct clock_backend {
 	struct sigevent clk_default;
 	int (*clk_clock_settime)(timespec_t *);
 	int (*clk_clock_gettime)(timespec_t *);
 	int (*clk_clock_getres)(timespec_t *);
-	int (*clk_timer_create)(itimer_t *, struct sigevent *);
+	int (*clk_timer_create)(itimer_t *, void (*)(itimer_t *));
 	int (*clk_timer_settime)(itimer_t *, int, const struct itimerspec *);
 	int (*clk_timer_gettime)(itimer_t *, struct itimerspec *);
 	int (*clk_timer_delete)(itimer_t *);
@@ -88,8 +105,8 @@ typedef struct clock_backend {
 } clock_backend_t;
 
 extern void clock_add_backend(clockid_t clock, clock_backend_t *backend);
+extern clock_backend_t *clock_get_backend(clockid_t clock);
 
-extern void timer_fire(itimer_t *);
 extern void timer_lwpbind();
 
 extern	void	timer_func(sigqueue_t *);
diff --git a/usr/src/uts/common/sys/timerfd.h b/usr/src/uts/common/sys/timerfd.h
new file mode 100644
index 0000000000..66cb50ac88
--- /dev/null
+++ b/usr/src/uts/common/sys/timerfd.h
@@ -0,0 +1,81 @@
+/*
+ * This file and its contents are supplied under the terms of the
+ * Common Development and Distribution License ("CDDL"), version 1.0.
+ * You may only use this file in accordance with the terms of version
+ * 1.0 of the CDDL.
+ *
+ * A full copy of the text of the CDDL should have accompanied this
+ * source.  A copy of the CDDL is also available via the Internet at
+ * http://www.illumos.org/license/CDDL.
+ */
+
+/*
+ * Copyright (c) 2015 Joyent, Inc.  All rights reserved.
+ */
+
+/*
+ * Header file to support for the timerfd facility.
+ */
+
+#ifndef _SYS_TIMERFD_H
+#define	_SYS_TIMERFD_H
+
+#include <sys/types.h>
+#include <sys/time_impl.h>
+
+#ifdef	__cplusplus
+extern "C" {
+#endif
+
+/*
+ * To assure binary compatibility with Linux, these values are fixed at their
+ * Linux equivalents, not their native ones.
+ */
+#define	TFD_CLOEXEC		02000000		/* LX_O_CLOEXEC */
+#define	TFD_NONBLOCK		04000			/* LX_O_NONBLOCK */
+#define	TFD_TIMER_ABSTIME	(1 << 0)
+#define	TFD_TIMER_CANCEL_ON_SET	(1 << 1)
+
+/*
+ * These ioctl values are specific to the native implementation; applications
+ * shouldn't be using them directly, and they should therefore be safe to
+ * change without breaking apps.
+ */
+#define	TIMERFDIOC		(('t' << 24) | ('f' << 16) | ('d' << 8))
+#define	TIMERFDIOC_CREATE	(TIMERFDIOC | 1)	/* create timer */
+#define	TIMERFDIOC_SETTIME	(TIMERFDIOC | 2)	/* timerfd_settime() */
+#define	TIMERFDIOC_GETTIME	(TIMERFDIOC | 3)	/* timerfd_gettime() */
+
+typedef struct timerfd_settime {
+	uint64_t tfd_settime_flags;	/* flags (e.g., TFD_TIMER_ABSTIME) */
+	uint64_t tfd_settime_value;	/* pointer to value */
+	uint64_t tfd_settime_ovalue;	/* pointer to old value, if any */
+} timerfd_settime_t;
+
+#ifndef _KERNEL
+
+extern int timerfd_create(int, int);
+extern int timerfd_settime(int, int,
+    const struct itimerspec *, struct itimerspec *);
+extern int timerfd_gettime(int, struct itimerspec *);
+
+#else
+
+#define	TIMERFDMNRN_TIMERFD	0
+#define	TIMERFDMNRN_CLONE	1
+#define	TIMERFD_VALMAX		(ULLONG_MAX - 1ULL)
+
+/*
+ * Fortunately, the values for the Linux clocks that are valid for timerfd
+ * (namely, CLOCK_REALTIME and CLOCK_MONOTONIC) don't overlap with our values
+ * the same.
+ */
+#define	TIMERFD_MONOTONIC	1	/* Linux value for CLOCK_MONOTONIC */
+
+#endif /* _KERNEL */
+
+#ifdef	__cplusplus
+}
+#endif
+
+#endif	/* _SYS_TIMERFD_H */
diff --git a/usr/src/uts/intel/Makefile.intel b/usr/src/uts/intel/Makefile.intel
index 64b4962cae..18d05f23d0 100644
--- a/usr/src/uts/intel/Makefile.intel
+++ b/usr/src/uts/intel/Makefile.intel
@@ -21,6 +21,7 @@
 # Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
 # Copyright (c) 2015 Nexenta Systems, Inc. All rights reserved.
 # Copyright (c) 2013 Andrew Stormont.  All rights reserved.
+# Copyright (c) 2015, Joyent, Inc.  All rights reserved.
 
 #
 #	This makefile contains the common definitions for all intel
@@ -338,6 +339,7 @@ DRV_KMODS	+= sysevent
 DRV_KMODS	+= sysmsg
 DRV_KMODS	+= tcp
 DRV_KMODS	+= tcp6
+DRV_KMODS	+= timerfd
 DRV_KMODS	+= tl
 DRV_KMODS	+= tnf
 DRV_KMODS	+= tpm
diff --git a/usr/src/uts/intel/timerfd/Makefile b/usr/src/uts/intel/timerfd/Makefile
new file mode 100644
index 0000000000..28cf8fd41e
--- /dev/null
+++ b/usr/src/uts/intel/timerfd/Makefile
@@ -0,0 +1,68 @@
+#
+# This file and its contents are supplied under the terms of the
+# Common Development and Distribution License ("CDDL"), version 1.0.
+# You may only use this file in accordance with the terms of version
+# 1.0 of the CDDL.
+#
+# A full copy of the text of the CDDL should have accompanied this
+# source.  A copy of the CDDL is also available via the Internet at
+# http://www.illumos.org/license/CDDL.
+#
+
+#
+# Copyright (c) 2015 Joyent, Inc.  All rights reserved.
+#
+
+#
+#	Path to the base of the uts directory tree (usually /usr/src/uts).
+#
+UTSBASE	= ../..
+
+#
+#	Define the module and object file sets.
+#
+MODULE		= timerfd
+OBJECTS		= $(TIMERFD_OBJS:%=$(OBJS_DIR)/%)
+LINTS		= $(TIMERFD_OBJS:%.o=$(LINTS_DIR)/%.ln)
+ROOTMODULE	= $(USR_DRV_DIR)/$(MODULE)
+CONF_SRCDIR	= $(UTSBASE)/common/io
+
+#
+#	Include common rules.
+#
+include $(UTSBASE)/intel/Makefile.intel
+
+CERRWARN	+= -_gcc=-Wno-parentheses
+
+#
+#	Define targets
+#
+ALL_TARGET	= $(BINARY) $(SRC_CONFILE)
+LINT_TARGET	= $(MODULE).lint
+INSTALL_TARGET	= $(BINARY) $(ROOTMODULE) $(ROOT_CONFFILE)
+
+#
+#	Default build targets.
+#
+.KEEP_STATE:
+
+def:		$(DEF_DEPS)
+
+all:		$(ALL_DEPS)
+
+clean:		$(CLEAN_DEPS)
+
+clobber:	$(CLOBBER_DEPS)
+
+lint:		$(LINT_DEPS)
+
+modlintlib:	$(MODLINTLIB_DEPS)
+
+clean.lint:	$(CLEAN_LINT_DEPS)
+
+install:	$(INSTALL_DEPS)
+
+#
+#	Include common targets.
+#
+include $(UTSBASE)/intel/Makefile.targ
diff --git a/usr/src/uts/sparc/Makefile.sparc b/usr/src/uts/sparc/Makefile.sparc
index d7c45a396e..29e4cd7ccf 100644
--- a/usr/src/uts/sparc/Makefile.sparc
+++ b/usr/src/uts/sparc/Makefile.sparc
@@ -21,6 +21,7 @@
 
 # Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
 # Copyright (c) 2013 Andrew Stormont.  All rights reserved.
+# Copyright (c) 2015, Joyent, Inc. All rights reserved.
 # Copyright 2015 Nexenta Systems, Inc. All rights reserved.
 
 
diff --git a/usr/src/uts/sparc/timerfd/Makefile b/usr/src/uts/sparc/timerfd/Makefile
new file mode 100644
index 0000000000..969ca6dd1a
--- /dev/null
+++ b/usr/src/uts/sparc/timerfd/Makefile
@@ -0,0 +1,68 @@
+#
+# This file and its contents are supplied under the terms of the
+# Common Development and Distribution License ("CDDL"), version 1.0.
+# You may only use this file in accordance with the terms of version
+# 1.0 of the CDDL.
+#
+# A full copy of the text of the CDDL should have accompanied this
+# source.  A copy of the CDDL is also available via the Internet at
+# http://www.illumos.org/license/CDDL.
+#
+
+#
+# Copyright (c) 2015 Joyent, Inc.  All rights reserved.
+#
+
+#
+#	Path to the base of the uts directory tree (usually /usr/src/uts).
+#
+UTSBASE	= ../..
+
+#
+#	Define the module and object file sets.
+#
+MODULE		= timerfd
+OBJECTS		= $(TIMERFD_OBJS:%=$(OBJS_DIR)/%)
+LINTS		= $(TIMERFD_OBJS:%.o=$(LINTS_DIR)/%.ln)
+ROOTMODULE	= $(USR_DRV_DIR)/$(MODULE)
+CONF_SRCDIR	= $(UTSBASE)/common/io
+
+#
+#	Include common rules.
+#
+include $(UTSBASE)/sparc/Makefile.sparc
+
+CERRWARN	+= -_gcc=-Wno-parentheses
+
+#
+#	Define targets
+#
+ALL_TARGET	= $(BINARY) $(SRC_CONFILE)
+LINT_TARGET	= $(MODULE).lint
+INSTALL_TARGET	= $(BINARY) $(ROOTMODULE) $(ROOT_CONFFILE)
+
+#
+#	Default build targets.
+#
+.KEEP_STATE:
+
+def:		$(DEF_DEPS)
+
+all:		$(ALL_DEPS)
+
+clean:		$(CLEAN_DEPS)
+
+clobber:	$(CLOBBER_DEPS)
+
+lint:		$(LINT_DEPS)
+
+modlintlib:	$(MODLINTLIB_DEPS)
+
+clean.lint:	$(CLEAN_LINT_DEPS)
+
+install:	$(INSTALL_DEPS)
+
+#
+#	Include common targets.
+#
+include $(UTSBASE)/sparc/Makefile.targ