Re: passive references

[Taylor R Campbell <riastradh%NetBSD.org@localhost>]

   Date: Mon, 1 Feb 2016 14:58:05 +0900
   From: Ryota Ozaki <ozaki-r%netbsd.org@localhost>

   On Sat, Jan 30, 2016 at 12:43 AM, Taylor R Campbell
   <riastradh%netbsd.org@localhost> wrote:
   >    Date: Fri, 29 Jan 2016 17:39:13 +0900
   >    From: Ryota Ozaki <ozaki-r%netbsd.org@localhost>
   >
   >    I'm thinking applying psref to bridge member list
   >    that is now using its own version of similar mechanism
   >    (psz + refcount).
   >
   > Nice!  That sounds like a good application to test, since most of the
   > work has been done already and it's mainly a matter of replacing a
   > refcount by a psref.

   I'm happy if there is a patch of psref.c to -current :)

Here you go, attached!  I even compile-tested this one.

   > int bound = curlwp->l_pflag & LP_BOUND;
   > curlwp->l_pflag |= LP_BOUND;
   > ...acquire and release psref...
   > curlwp->l_pflag ^= bound ^ LP_BOUND;
   >
   > [...]
   >
   > Someone more familiar with the inner workings of LWPs and the
   > scheduler should review this, though.  And maybe we should invent a
   > name lwp_bind/lwp_unbind for this pattern.

   I agree; open accesses to l_pflag looks not good idea.

Actually, I think it is fine.  There was only one exception I could
find in-tree, LP_VFORKWAIT, and it has been relegated to #if 0.  All
other flags are used in l_pflag unlocked, and I believe the scheduler
and interrupt handlers will undo any changes they make -- LP_BOUND,
LP_INTR, LP_OWEUPC, LP_RUNNING, LP_TIMEINTR.

That said, I think naming lwp_bind/lwp_unbind is still a good idea!
Maybe instead curlwp_bind/curlwp_unbind.

/* sys/lwp.h */
static inline int
curlwp_bind(void)
{
	int bound;

	bound = curlwp->l_pflag & LP_BOUND;
	curlwp->l_pflag |= LP_BOUND;

	return bound;
}

static inline void
curlwp_unbind(int bound)
{

	curlwp->l_pflag ^= bound ^ LP_BOUND;
}

Index: sys/arch/amd64/conf/ALL
===================================================================
RCS file: /cvsroot/src/sys/arch/amd64/conf/ALL,v
retrieving revision 1.33
diff -p -u -r1.33 ALL
--- sys/arch/amd64/conf/ALL	10 Nov 2015 13:01:41 -0000	1.33
+++ sys/arch/amd64/conf/ALL	2 Feb 2016 00:08:33 -0000
@@ -125,6 +125,7 @@ options 	SYSCALL_STATS	# per syscall cou
 options 	SYSCALL_TIMES	# per syscall times
 options 	SYSCALL_TIMES_HASCOUNTER	# use 'broken' rdtsc (soekris)
 options 	KDTRACE_HOOKS	# kernel DTrace hooks
+options 	PSREF_DEBUG	# debug passive references
 
 # Compatibility options
 #options 	COMPAT_NOMID	# NetBSD 0.8, 386BSD, and BSDI
Index: sys/arch/i386/conf/ALL
===================================================================
RCS file: /cvsroot/src/sys/arch/i386/conf/ALL,v
retrieving revision 1.397
diff -p -u -r1.397 ALL
--- sys/arch/i386/conf/ALL	25 Oct 2015 22:48:23 -0000	1.397
+++ sys/arch/i386/conf/ALL	2 Feb 2016 00:08:34 -0000
@@ -125,6 +125,7 @@ options 	SYSCALL_STATS	# per syscall cou
 options 	SYSCALL_TIMES	# per syscall times
 options 	SYSCALL_TIMES_HASCOUNTER	# use 'broken' rdtsc (soekris)
 options 	KDTRACE_HOOKS	# kernel DTrace hooks
+options 	PSREF_DEBUG	# debug passive references
 
 # Compatibility options
 options 	COMPAT_NOMID	# NetBSD 0.8, 386BSD, and BSDI
Index: sys/conf/files
===================================================================
RCS file: /cvsroot/src/sys/conf/files,v
retrieving revision 1.1152
diff -p -u -r1.1152 files
--- sys/conf/files	9 Dec 2015 18:25:32 -0000	1.1152
+++ sys/conf/files	2 Feb 2016 00:08:34 -0000
@@ -284,6 +284,7 @@ defparam opt_kgdb.h		KGDB_DEV KGDB_DEVNA
 defflag				LOCKDEBUG
 defflag				SYSCALL_DEBUG
 defflag	opt_kstack.h		KSTACK_CHECK_MAGIC
+defflag	opt_psref.h		PSREF_DEBUG
 
 # memory (ram) disk options
 #
Index: sys/kern/files.kern
===================================================================
RCS file: /cvsroot/src/sys/kern/files.kern,v
retrieving revision 1.10
diff -p -u -r1.10 files.kern
--- sys/kern/files.kern	3 Dec 2015 02:51:00 -0000	1.10
+++ sys/kern/files.kern	2 Feb 2016 00:08:35 -0000
@@ -125,6 +125,7 @@ file	kern/subr_pool.c		kern
 file	kern/subr_prf.c			kern
 file	kern/subr_prof.c		kern
 file	kern/subr_pserialize.c		kern
+file	kern/subr_psref.c		kern
 file	kern/subr_specificdata.c	kern
 file	kern/subr_tftproot.c		tftproot
 file	kern/subr_time.c		kern
Index: sys/kern/subr_psref.c
===================================================================
RCS file: sys/kern/subr_psref.c
diff -N sys/kern/subr_psref.c
--- /dev/null	1 Jan 1970 00:00:00 -0000
+++ sys/kern/subr_psref.c	2 Feb 2016 00:08:35 -0000
@@ -0,0 +1,409 @@
+/*	$NetBSD$	*/
+
+/*-
+ * Copyright (c) 2016 The NetBSD Foundation, Inc.
+ * All rights reserved.
+ *
+ * This code is derived from software contributed to The NetBSD Foundation
+ * by Taylor R. Campbell.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ * 1. Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
+ * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+ * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
+ * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+ * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+ * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+ * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+ * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+ * POSSIBILITY OF SUCH DAMAGE.
+ */
+
+/*
+ * Passive references
+ *
+ *	Passive references are references to objects that guarantee the
+ *	object will not be destroyed until the reference is released.
+ *
+ *	Passive references require no interprocessor synchronization to
+ *	acquire or release.  However, destroying the target of passive
+ *	references requires expensive interprocessor synchronization --
+ *	xcalls to determine on which CPUs the object is still in use.
+ *
+ *	Passive references may be held only on a single CPU and by a
+ *	single LWP.  They require the caller to allocate a little stack
+ *	space, a struct psref object.  Sleeping while a passive
+ *	reference is held is allowed, provided that the owner's LWP is
+ *	bound to a CPU -- e.g., the owner is a softint or a bound
+ *	kthread.  However, sleeping should be kept to a short duration,
+ *	e.g. sleeping on an adaptive lock.
+ *
+ *	Passive references serve as an intermediate stage between
+ *	reference counting and passive serialization (pserialize(9)):
+ *
+ *	- If you need references to transfer from CPU to CPU or LWP to
+ *	  LWP, or if you need long-term references, you must use
+ *	  reference counting, e.g. with atomic operations or locks,
+ *	  which incurs interprocessor synchronization for every use --
+ *	  cheaper than an xcall, but not scalable.
+ *
+ *	- If all users *guarantee* that they will not sleep, then it is
+ *	  not necessary to use passive references: you may as well just
+ *	  use the even cheaper pserialize(9), because you have
+ *	  satisfied the requirements of a pserialize read section.
+ */
+
+#include <sys/cdefs.h>
+__KERNEL_RCSID(0, "$NetBSD$");
+
+#include <sys/types.h>
+#include <sys/condvar.h>
+#include <sys/cpu.h>
+#include <sys/intr.h>
+#include <sys/kmem.h>
+#include <sys/lwp.h>
+#include <sys/mutex.h>
+#include <sys/percpu.h>
+#include <sys/psref.h>
+#include <sys/queue.h>
+#include <sys/xcall.h>
+
+LIST_HEAD(psref_head, psref);
+
+/*
+ * struct psref_class
+ *
+ *	Private global state for a class of passive reference targets.
+ *	Opaque to callers.
+ */
+struct psref_class {
+	kmutex_t		prc_lock;
+	kcondvar_t		prc_cv;
+	struct percpu		*prc_percpu; /* struct psref_cpu */
+	ipl_cookie_t		prc_iplcookie;
+};
+
+/*
+ * struct psref_cpu
+ *
+ *	Private per-CPU state for a class of passive reference targets.
+ *	Not exposed by the API.
+ */
+struct psref_cpu {
+	struct psref_head	pcpu_head;
+};
+
+/*
+ * psref_class_create(name, ipl)
+ *
+ *	Create a new passive reference class, with the given wchan name
+ *	and ipl.
+ */
+struct psref_class *
+psref_class_create(const char *name, int ipl)
+{
+	struct psref_class *class;
+
+	ASSERT_SLEEPABLE();
+
+	class = kmem_alloc(sizeof(*class), KM_SLEEP);
+	if (class == NULL)
+		goto fail0;
+
+	class->prc_percpu = percpu_alloc(sizeof(struct psref_cpu));
+	if (class->prc_percpu == NULL)
+		goto fail1;
+
+	mutex_init(&class->prc_lock, MUTEX_DEFAULT, ipl);
+	cv_init(&class->prc_cv, name);
+	class->prc_iplcookie = makeiplcookie(ipl);
+
+fail1:	kmem_free(class, sizeof(*class));
+fail0:	return NULL;
+}
+
+#if DIAGNOSTIC
+static void
+psref_cpu_drained_p(void *p, void *cookie, struct cpu_info *ci __unused)
+{
+	struct psref_cpu *pcpu = p;
+	bool *retp = cookie;
+
+	if (!LIST_EMPTY(&pcpu->pcpu_head))
+		*retp = false;
+}
+
+static bool
+psref_class_drained_p(struct psref_class *prc)
+{
+	bool ret = true;
+
+	percpu_foreach(prc->prc_percpu, &psref_cpu_drained_p, &ret);
+
+	return ret;
+}
+#endif	/* DIAGNOSTIC */
+
+/*
+ * psref_class_destroy(class)
+ *
+ *	Destroy a passive reference class and free memory associated
+ *	with it.  All targets in this class must have been drained and
+ *	destroyed already.
+ */
+void
+psref_class_destroy(struct psref_class *class)
+{
+
+	KASSERT(psref_class_drained_p(class));
+
+	cv_destroy(&class->prc_cv);
+	mutex_destroy(&class->prc_lock);
+	percpu_free(class->prc_percpu, sizeof(struct psref_cpu));
+	kmem_free(class, sizeof(*class));
+}
+
+/*
+ * psref_target_init(target, class)
+ *
+ *	Initialize a passive reference target in the specified class.
+ *	The caller is responsible for issuing a membar_producer before
+ *	exposing a pointer to the target to other CPUs.
+ */
+void
+psref_target_init(struct psref_target *target, struct psref_class *class)
+{
+
+	target->prt_draining = false;
+#ifdef PSREF_DEBUG
+	target->prt_class = class;
+#endif
+}
+
+/*
+ * psref_target_destroy(target, class)
+ *
+ *	Destroy a passive reference target.  It must have previously
+ *	been drained.
+ */
+void
+psref_target_destroy(struct psref_target *target, struct psref_class *class)
+{
+
+	KASSERT(target->prt_draining);
+#ifdef PSREF_DEBUG
+	KASSERT(target->prt_class == class);
+	target->prt_class = NULL;
+#endif
+}
+
+/*
+ * psref_acquire(psref, target, class)
+ *
+ *	Try to acquire a passive reference to the specified target,
+ *	which must be in the specified class.  On success, returns
+ *	zero; on failure, returns a nonzero error code.  If the target
+ *	is draining, returns ENOENT.
+ *
+ *	The caller must guarantee that it will not switch CPUs before
+ *	releasing the passive reference, either by disabling
+ *	kpreemption and avoiding sleeps, or by being in a softint or in
+ *	an LWP bound to a CPU.
+ */
+int
+psref_acquire(struct psref *psref, struct psref_target *target,
+    struct psref_class *class)
+{
+	struct psref_cpu *pcpu;
+	int s, error;
+
+	KASSERTMSG((kpreempt_disabled() || cpu_softintr_p() ||
+		ISSET(curlwp->l_pflag, LP_BOUND)),
+	    "passive references are CPU-local,"
+	    " but preemption is enabled and the caller is not"
+	    " in a softint or CPU-bound LWP");
+
+#ifdef PSREF_DEBUG
+	KASSERT(target->prt_class == class);
+#endif
+
+	/* Block interrupts and acquire the current CPU's reference list.  */
+	s = splraiseipl(class->prc_iplcookie);
+	pcpu = percpu_getref(class->prc_percpu);
+
+	/* Is this target going away?  */
+	if (__predict_false(target->prt_draining)) {
+		/* Yes: fail.  */
+		error = ENOENT;
+	} else {
+		/* No: record our reference.  */
+		LIST_INSERT_HEAD(&pcpu->pcpu_head, psref, psref_entry);
+		psref->psref_target = target;
+#ifdef PSREF_DEBUG
+		psref->psref_lwp = curlwp;
+		psref->psref_cpu = curcpu();
+#endif
+		error = 0;
+	}
+
+	/* Release the CPU list and restore interrupts.  */
+	percpu_putref(class->prc_percpu);
+	splx(s);
+
+	return error;
+}
+
+/*
+ * psref_release(psref, target, class)
+ *
+ *	Release a passive reference to the specified target, which must
+ *	be in the specified class.
+ *
+ *	The caller must not have switched CPUs or LWPs since acquiring
+ *	the passive reference.
+ */
+void
+psref_release(struct psref *psref, struct psref_target *target,
+    struct psref_class *class)
+{
+	int s;
+
+	KASSERTMSG((kpreempt_disabled() || cpu_softintr_p() ||
+		ISSET(curlwp->l_pflag, LP_BOUND)),
+	    "passive references are CPU-local,"
+	    " but preemption is enabled and the caller is not"
+	    " in a softint or CPU-bound LWP");
+
+	KASSERT(psref->psref_target == target);
+#ifdef PSREF_DEBUG
+	KASSERT(target->prt_class == class);
+	KASSERTMSG((psref->psref_lwp == curlwp),
+	    "passive reference transferred from lwp %p to lwp %p",
+	    psref->psref_lwp, curlwp);
+	KASSERTMSG((psref->psref_cpu == curcpu()),
+	    "passive reference transferred from CPU %u to CPU %u",
+	    cpu_index(psref->psref_cpu), cpu_index(curcpu()));
+#endif
+
+	/*
+	 * Block interrupts and remove the psref from the current CPU's
+	 * list.  No need to percpu_getref or get the head of the list,
+	 * and the caller guarantees that we are bound to a CPU anyway
+	 * (as does blocking interrupts).
+	 */
+	s = splraiseipl(class->prc_iplcookie);
+	LIST_REMOVE(psref, psref_entry);
+	splx(s);
+
+	/* If someone is waiting for users to drain, notify 'em.  */
+	if (__predict_false(target->prt_draining))
+		cv_broadcast(&class->prc_cv);
+}
+
+/*
+ * struct psreffed
+ *
+ *	Global state for draining a psref target.
+ */
+struct psreffed {
+	struct psref_class	*class;
+	struct psref_target	*target;
+	bool			ret;
+};
+
+static void
+psreffed_p_xc(void *cookie0, void *cookie1 __unused)
+{
+	struct psreffed *P = cookie0;
+	struct psref_class *class = P->class;
+	struct psref_target *target = P->target;
+	struct psref_cpu *pcpu;
+	struct psref *psref;
+	int s;
+
+	/* Block interrupts and acquire the current CPU's reference list.  */
+	s = splraiseipl(class->prc_iplcookie);
+	pcpu = percpu_getref(class->prc_percpu);
+
+	/*
+	 * Check the CPU's reference list for any references to this
+	 * target.  This loop shouldn't take very long because any
+	 * single CPU should hold only a small number of references at
+	 * any given time unless there is a bug.
+	 */
+	LIST_FOREACH(psref, &pcpu->pcpu_head, psref_entry) {
+		if (psref->psref_target == target) {
+			/*
+			 * No need to lock anything here: every write
+			 * transitions from false to true, so as long
+			 * as any write goes through we're good.  No
+			 * need for a memory barrier because this is
+			 * read only after xc_wait, which has already
+			 * issued any necessary memory barriers.
+			 */
+			P->ret = true;
+			break;
+		}
+	}
+
+	/* Release the CPU list and restore interrupts.  */
+	percpu_putref(class->prc_percpu);
+	splx(s);
+}
+
+static bool
+psreffed_p(struct psref_target *target, struct psref_class *class)
+{
+	struct psreffed P = {
+		.class = class,
+		.target = target,
+		.ret = false,
+	};
+
+	xc_wait(xc_broadcast(0, &psreffed_p_xc, &P, NULL));
+
+	return P.ret;
+}
+
+/*
+ * psref_target_drain(target, class)
+ *
+ *	Prevent new references to target and wait for existing ones to
+ *	drain.  May sleep.
+ */
+void
+psref_target_drain(struct psref_target *target, struct psref_class *class)
+{
+
+#ifdef PSREF_DEBUG
+	KASSERT(target->prt_class == class);
+#endif
+
+	ASSERT_SLEEPABLE();
+	KASSERT(!target->prt_draining);
+	target->prt_draining = true;
+
+	/* Wait until there are no more references on any CPU.  */
+	while (psreffed_p(target, class)) {
+		/*
+		 * This enter/wait/exit business looks wrong, but it is
+		 * both necessary, because psreffed_p performs a
+		 * low-priority xcall and hence cannot run while a
+		 * mutex is locked, and OK, because the wait is timed
+		 * -- explicit wakeups are only an optimization.
+		 */
+		mutex_enter(&class->prc_lock);
+		(void)cv_timedwait(&class->prc_cv, &class->prc_lock, 1);
+		mutex_exit(&class->prc_lock);
+	}
+}
Index: sys/sys/psref.h
===================================================================
RCS file: sys/sys/psref.h
diff -N sys/sys/psref.h
--- /dev/null	1 Jan 1970 00:00:00 -0000
+++ sys/sys/psref.h	2 Feb 2016 00:08:36 -0000
@@ -0,0 +1,107 @@
+/*	$NetBSD$	*/
+
+/*-
+ * Copyright (c) 2016 The NetBSD Foundation, Inc.
+ * All rights reserved.
+ *
+ * This code is derived from software contributed to The NetBSD Foundation
+ * by Taylor R. Campbell.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ * 1. Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
+ * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+ * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
+ * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+ * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+ * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+ * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+ * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+ * POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#ifndef	_SYS_PSREF_H
+#define	_SYS_PSREF_H
+
+#include <sys/types.h>
+#include <sys/queue.h>
+
+/*
+ * PSREF_DEBUG
+ *
+ *	If nonzero, enable debugging of psrefs.  WARNING: This changes
+ *	the ABI by adding extra fields to struct psref_target and
+ *	struct psref, which are exposed to callers and embedded in
+ *	other structures.
+ */
+#ifdef _KERNEL_OPT
+#include "opt_psref.h"
+#endif
+
+struct cpu_info;
+struct lwp;
+
+struct psref;
+struct psref_class;
+struct psref_target;
+
+/*
+ * struct psref_target
+ *
+ *	Bookkeeping for an object to which users can acquire passive
+ *	references.  This is compact so that it can easily be embedded
+ *	into many multitudes of objects, e.g. IP packet flows.
+ *
+ *	prt_draining is false on initialization, and may be written
+ *	only once, to make it true, when someone with exclusive access
+ *	to it is about to drain the target.
+ */
+struct psref_target {
+	bool			prt_draining;
+#ifdef PSREF_DEBUG
+	struct psref_class	*prt_class;
+#endif
+};
+
+/*
+ * struct psref
+ *
+ *	Bookkeeping for a single passive reference.  There should only
+ *	be a few of these per CPU in the system at once, no matter how
+ *	many targets are stored, so these are a bit larger than struct
+ *	psref_target.  The contents of struct psref may be read and
+ *	written only on the local CPU.
+ */
+struct psref {
+	LIST_ENTRY(psref)	psref_entry;
+	struct psref_target	*psref_target;
+#ifdef PSREF_DEBUG
+	struct lwp		*psref_lwp;
+	struct cpu_info		*psref_cpu;
+#endif
+};
+
+struct psref_class *
+	psref_class_create(const char *, int);
+void	psref_class_destroy(struct psref_class *);
+
+void	psref_target_init(struct psref_target *, struct psref_class *);
+void	psref_target_destroy(struct psref_target *, struct psref_class *);
+
+int	psref_acquire(struct psref *, struct psref_target *,
+	    struct psref_class *);
+void	psref_release(struct psref *, struct psref_target *,
+	    struct psref_class *);
+
+void	psref_target_drain(struct psref_target *, struct psref_class *);
+
+#endif	/* _SYS_PSREF_H */