Re: passive references

To: Kengo NAKAHARA <k-nakahara%iij.ad.jp@localhost>
Subject: Re: passive references
From: Taylor R Campbell <riastradh%NetBSD.org@localhost>
Date: Mon, 15 Feb 2016 22:23:31 +0000

   Date: Mon, 15 Feb 2016 19:26:44 +0900
   From: Kengo NAKAHARA <k-nakahara%iij.ad.jp@localhost>

   First, I fix some bug which you point out. Here is the
   "git format-patch" patch series.
       http://www.netbsd.org/~knakahara/fix-gif-softint-using-psref1/fix-gif-softint-using-psref1.tgz
   And here is the unified patch.
       http://www.netbsd.org/~knakahara/fix-gif-softint-using-psref1/unified-fix-gif-softint-using-psref1.patch

Thanks!  I'm a little confused, though: that patch series doesn't
apply cleanly when I extract it and pass it to `git am'.  Some of the
patches conflict with each other.

I would still like to see the independent functional changes in
separate patches -- especially USE_RADIX, because our radix tree code
is not safe for pserialize(9).  But I'll just make a few notes on the
unified patch for now.

   > I added a psref_held(target, class) operation, to be used only in
   > assertions:
   > 
   > 	psref_acquire(r, t, c);
   > 	...
   > 	KASSERT(psref_held(t, c));
   > 	...
   > 	psref_release(r, t, c);

   As far as I see the implementation, it seems to be able to be used
   after psref_release(and before pserialize_read_exit()). And I want
   to use such way.
   # see the following 0010-add-psref-KASSERT-to-encap-46-_lookup.patch
   Is this wrong?

Hmm...  I think in this case it happens to be safe, but in general, I
would advise against doing that.  As soon as you call psref_release,
you generally cannot use the target any more.

Also, if you assert !psref_held, that means even the *caller* is not
allowed to hold a passive reference to the target -- which makes the
contract of your function more complex.

So I would advise not to use !psref_held -- it is even more
questionable than !mutex_owned.

In the two cases where you use it, I would just make sure that the
very last action in the loop body is an unconditional psref_release.
In fact, you seem to be missing a psref_release for ep in the case
that 0 < prio < matchprio.

   > A new patch for psref is attached.  API changes:
   > 
   > - Removed psref_target_drain; moved logic into psref_target_destroy.
   > - Made psref_acquire never fail.
   > - Added psref_copy.
   > - Added psref_held.

   Hum, it seems psref_copy is lost. I add it in 0006-add-psref_copy.patch.
   If there are mistakes, please point out.

I must have forgotten to rerun cvs diff before sending my patch,
sorry!  It looks functionally equivalent to what I had -- I had just
added a couple more comments.  New psref2.patch attached.

In your patch, you try to do a radix tree lookup in a pserialize read
section.  It would be nice if that worked -- but I am pretty sure that
our radix tree code is not going to be pserialize-safe, because it
won't do membar_producer/membar_datadep_consumer.

However, since we can drop packets here, we can use a somewhat
heavy-handed alternative: simply drop all packets while anyone is
updating the radix tree.  Something like this:

struct radix_node_head *encap_head[2];
bool encap_head_updating = false;

encap_add(ep)
{
	struct radix_node_head *rnh = encap_rnh(ep->af);

	mutex_enter(&encaptab.lock);

	/* If it takes a simple radix lookup, use the radix tree.  */
	if (!ep->func && rnh) {
		/* Disable access to the radix tree.  */
		encap_head_updating = true;

		/* Wait for all readers to drain.  */
		pserialize_perform(psz);

		/* Insert into the radix tree.  */
		... rnh->rnh_addaddr(...) ...

		/* Re-enable access to the radix tree.  */
		encap_head_updating = false;
	}

	/* Publish it in the list.  */
	LIST_INSERT_HEAD_PSZ(&encap_table, ep, chain);

	mutex_exit(&encaptab.lock);
	return 0;
}

encap4_lookup(m, ...)
{
	int s;

	s = pserialize_read_enter();

	/* Check whether there's an update in progress.  */
	if (encap_head_updating) {
		/*
		 * Update in progress.  Pretend there are no tunnels
		 * and simply drop the packet.
		 */
		pserialize_read_exit(s);
		return NULL;
	}

	/* No update in progress.  Try a radix lookup.  */
	rn = rnh->rnh_matchaddr(...);
	if (rn && ...) {
		match = (struct encaptab *)rn;
		psref_acquire(match_psref, &match->psref, encaptab.elem_class);
		matchprio = ...;
	}

	/* Now go through all the function tests.  */
	for (ep = LIST_HEAD(&encaptab.list); ep != NULL; ep = next) {
		membar_datadep_consumer();
		...
		/* Acquire a psref so we can hold outside pserialize read.  */
		psref_acquire(&elem_psref, &ep->psref, encaptab.elem_class);
		...
			pserialize_read_exit(s);
			(*ep->func)(...);
			s = pserialize_read_enter(s);
		...
		/* Get the next one while this one is still valid.  */
		next = ep->chain;
		/* Release the psref.  */
		psref_release(&elem_psref, &ep->psref, encaptab.elem_class);
	}
	pserialize_read_exit(s);

	return match;
}

Small comment: In encap4_lookup, you do the following:

	LIST_FOREACH(ep, &encap_table, chain) {
		struct psref elem_psref;
		membar_datadep_consumer();
		psref_acquire(&elem_psref, &ep->psref, encaptab.elem_class);
		if (ep->af != AF_INET) {
			psref_release(&elem_psref, &ep->psref,
			    encaptab.elem_class);
			continue;
		}
		if (ep->proto >= 0 && ep->proto != proto) {
			psref_release(&elem_psref, &ep->psref,
			    encaptab.elem_class);
			continue;
		}
		...
	}

I would move the psref_acquire until after those two conditionals.
You don't need to hold a passive reference in order to read ep->af or
ep->proto -- reading those fields is guaranteed to never sleep.  Just
do:

	LIST_FOREACH(ep, &encap_table, chain) {
		struct psref elem_psref;
		membar_datadep_consumer();
		if (ep->af != AF_INET)
			continue;
		if (ep->proto >= 0 && ep->proto != proto)
			continue;
		psref_acquire(&elem_psref, &ep->psref, encaptab.elem_class);
		...
	}

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	15 Feb 2016 20:15:32 -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	15 Feb 2016 20:15:34 -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	15 Feb 2016 20:15:35 -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	15 Feb 2016 20:15:37 -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	15 Feb 2016 20:15:37 -0000
@@ -0,0 +1,513 @@
+/*	$NetBSD$	*/
+
+/*-
+ * 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>
+
+#ifdef PSREF_DEBUG
+#define	__psref_debugused
+#else
+#define	__psref_debugused	__unused
+#endif
+
+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);
+
+	return class;
+
+fail1:	kmem_free(class, sizeof(*class));
+fail0:	return NULL;
+}
+
+#ifdef 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 after
+ *	psref_target_init and before exposing a pointer to the target
+ *	to other CPUs.
+ */
+void
+psref_target_init(struct psref_target *target,
+    struct psref_class *class __psref_debugused)
+{
+
+#ifdef PSREF_DEBUG
+	target->prt_class = class;
+#endif
+	target->prt_draining = false;
+}
+
+/*
+ * psref_acquire(psref, target, class)
+ *
+ *	Acquire a passive reference to the specified target, which must
+ *	be in the specified class.
+ *
+ *	The caller must guarantee that the target will not be destroyed
+ *	before psref_acquire returns.
+ *
+ *	The caller must additionally 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.
+ */
+void
+psref_acquire(struct psref *psref, struct psref_target *target,
+    struct psref_class *class)
+{
+	struct psref_cpu *pcpu;
+	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");
+
+#ifdef PSREF_DEBUG
+	KASSERTMSG((target->prt_class == class),
+	    "mismatched psref target class: %p (ref) != %p (expected)",
+	    target->prt_class, class);
+#endif
+	KASSERTMSG(!target->prt_draining, "psref target already destroyed: %p",
+	    target);
+
+	/* Block interrupts and acquire the current CPU's reference list.  */
+	s = splraiseipl(class->prc_iplcookie);
+	pcpu = percpu_getref(class->prc_percpu);
+
+	/* 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
+
+	/* Release the CPU list and restore interrupts.  */
+	percpu_putref(class->prc_percpu);
+	splx(s);
+}
+
+/*
+ * 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");
+
+#ifdef PSREF_DEBUG
+	KASSERTMSG((target->prt_class == class),
+	    "mismatched psref target class: %p (ref) != %p (expected)",
+	    target->prt_class, class);
+#endif
+
+	/* Make sure the psref looks sensible.  */
+	KASSERTMSG((psref->psref_target == target),
+	    "passive reference target mismatch: %p (ref) != %p (expected)",
+	    psref->psref_target, target);
+#ifdef PSREF_DEBUG
+	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);
+}
+
+/*
+ * psref_copy(pto, pfrom class)
+ *
+ *	Copy a passive reference from pfrom, which must be in the
+ *	specified class, to pto.  Both pfrom and pto must later be
+ *	released with psref_release.
+ *
+ *	The caller must not have switched CPUs or LWPs since acquiring
+ *	pfrom, and must not switch CPUs or LWPs before releasing both
+ *	pfrom and pto.
+ */
+void
+psref_copy(struct psref *pto, const struct psref *pfrom,
+    struct psref_class *class)
+{
+	struct psref_cpu *pcpu;
+	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");
+	KASSERTMSG((pto != pfrom),
+	    "can't copy passive reference to itself: %p",
+	    pto);
+
+#ifdef PSREF_DEBUG
+	/* Make sure the pfrom reference looks sensible.  */
+	KASSERTMSG((pfrom->psref_lwp == curlwp),
+	    "passive reference transferred from lwp %p to lwp %p",
+	    pfrom->psref_lwp, curlwp);
+	KASSERTMSG((pfrom->psref_cpu == curcpu()),
+	    "passive reference transferred from CPU %u to CPU %u",
+	    cpu_index(pfrom->psref_cpu), cpu_index(curcpu()));
+	KASSERTMSG((pfrom->psref_target->prt_class == class),
+	    "mismatched psref target class: %p (ref) != %p (expected)",
+	    pfrom->psref_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);
+
+	/* Record the new reference.  */
+	LIST_INSERT_HEAD(&pcpu->pcpu_head, pto, psref_entry);
+	pto->psref_target = pfrom->psref_target;
+#ifdef PSREF_DEBUG
+	pto->psref_lwp = curlwp;
+	pto->psref_cpu = curcpu();
+#endif
+
+	/* Release the CPU list and restore interrupts.  */
+	percpu_putref(class->prc_percpu);
+	splx(s);
+}
+
+/*
+ * 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;
+
+	/*
+	 * If we hold a psref to the target, then answer true.
+	 *
+	 * This is the only dynamic decision that may be made with
+	 * psref_held.
+	 *
+	 * No need to lock anything here: every write transitions from
+	 * false to true, so there can be no conflicting writes.  No
+	 * need for a memory barrier here because P->ret is read only
+	 * after xc_wait, which has already issued any necessary memory
+	 * barriers.
+	 */
+	if (psref_held(P->target, P->class))
+		P->ret = true;
+}
+
+static bool
+psreffed_p(struct psref_target *target, struct psref_class *class)
+{
+	struct psreffed P = {
+		.class = class,
+		.target = target,
+		.ret = false,
+	};
+
+	/* Ask all CPUs to say whether they hold a psref to the target.  */
+	xc_wait(xc_broadcast(0, &psreffed_p_xc, &P, NULL));
+
+	return P.ret;
+}
+
+/*
+ * psref_target_destroy(target, class)
+ *
+ *	Destroy a passive reference target.  Waits for all existing
+ *	references to drain.  Caller must guarantee no new references
+ *	will be acquired once it calls psref_target_destroy, e.g. by
+ *	removing the target from a global list first.  May sleep.
+ */
+void
+psref_target_destroy(struct psref_target *target, struct psref_class *class)
+{
+
+	ASSERT_SLEEPABLE();
+
+#ifdef PSREF_DEBUG
+	KASSERTMSG((target->prt_class == class),
+	    "mismatched psref target class: %p (ref) != %p (expected)",
+	    target->prt_class, class);
+#endif
+
+	/* Request psref_release to notify us when done.  */
+	KASSERTMSG(!target->prt_draining, "psref target already destroyed: %p",
+	    target);
+	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);
+	}
+
+#ifdef PSREF_DEBUG
+	/* No more references.  Cause subsequent psref_acquire to kassert.  */
+	target->prt_class = NULL;
+#endif
+}
+
+/*
+ * psref_held(target, class)
+ *
+ *	True if the current CPU holds a passive reference to target,
+ *	false otherwise.  May be used only inside assertions.
+ */
+bool
+psref_held(struct psref_target *target, struct psref_class *class)
+{
+	struct psref_cpu *pcpu;
+	struct psref *psref;
+	int s;
+	bool held = false;
+
+	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
+	KASSERTMSG((target->prt_class == class),
+	    "mismatched psref target class: %p (ref) != %p (expected)",
+	    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);
+
+	/* Search through all the references on this CPU.  */
+	LIST_FOREACH(psref, &pcpu->pcpu_head, psref_entry) {
+#ifdef PSREF_DEBUG
+		/* Sanity-check the reference.  */
+		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
+
+		/* If it matches, stop here and answer yes.  */
+		if (psref->psref_target == target) {
+			held = true;
+			break;
+		}
+	}
+
+	/* Release the CPU list and restore interrupts.  */
+	percpu_putref(class->prc_percpu);
+	splx(s);
+
+	return held;
+}
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	15 Feb 2016 20:15:38 -0000
@@ -0,0 +1,111 @@
+/*	$NetBSD$	*/
+
+/*-
+ * 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 has prevented new
+ *	references from being created and wants to drain the target in
+ *	order to destroy it.
+ */
+struct psref_target {
+#ifdef PSREF_DEBUG
+	struct psref_class	*prt_class;
+#endif
+	bool			prt_draining;
+};
+
+/*
+ * 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 *);
+
+void	psref_acquire(struct psref *, struct psref_target *,
+	    struct psref_class *);
+void	psref_release(struct psref *, struct psref_target *,
+	    struct psref_class *);
+void	psref_copy(struct psref *, const struct psref *,
+	    struct psref_class *);
+
+/* For use only in assertions.  */
+bool	psref_held(struct psref_target *, struct psref_class *);
+
+#endif	/* _SYS_PSREF_H */

Follow-Ups:
- Re: passive references
  - From: Kengo NAKAHARA
- Re: passive references
  - From: Taylor R Campbell

References:
- Re: passive references
  - From: Kengo NAKAHARA

Prev by Date: Re: sctp & ipv4-mapped ipv6 fix
Next by Date: Re: passive references
Previous by Thread: Re: passive references
Next by Thread: Re: passive references
Indexes:

Home | Main Index | Thread Index | Old Index