Skip to contentSkip to conversation

signal: Let tasks cache one sigqueue struct.

PATCH Superseded Sebastian Andrzej Siewior 3 participants 5 messages Apr 6, 2023 – May 24, 2023

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH v3 1/1 signal: Let tasks cache one sigqueue struct.

The sigqueue caching originated in the PREEMPT_RT tree. A few of the applications, that were ported to Linux, were ported from OS-9. Sending notifications from one task to another via a signal was a common communication model there and so the applications are heavy signal users. Removing the allocation reduces the critical path, avoids locks and so lowers the maximal latency of the task while sending a signal.

After posting the first version a discussion came up whether it wouldn't make sense to have this caching unconditionally and not restricted to PREEMPT_RT only.

The sigqueue flagged SIGQUEUE_PREALLOC is used by the POSIX timer code. It is allocated on initialisation and reused during the lifetime of the timer. The sigqueue is freed once the timer is deleted. The POSIX timer sigqueue has its own caching and unique lifetime pattern and therefore does not fit for the generic caching.

In the common case the signal is dequeued and freed in collect_signal(). At this point, the 'current' task receives the signal and its sighand_struct::siglock is held.

__sigqueue_alloc() is used to allocate a new sigqueue. The task_struct passed as argument is the task that will receive the signal. Its sighand_struct::siglock is acquired (except for SIGQUEUE_PREALLOC allocation which is ignored). Use a cached sigqueue before allocating a new one.

As a result of this pattern, the task sending a signal will use the cache from the task that will receive the signal which in turn caches the signal.

The numbers of system boot followed by an allmod kernel build:
Out of 333216 allocations, 194876 (~58%) were served from the cache. From all free invocations, 4212 were in a path were caching is not done and 329002 sigqueue were cached.

Cache the struct sigqueue in collect_signal() and reuse it for the allocation. Rely on sighand_struct::siglock locking for cache handling which is held during allocation and free. The cache is cleaned once sighand_struct is freed.

---
v2…v3: Wrote the patch and its description from scratch without looking at the previous version.

 include/linux/sched/signal.h |    2 +
 kernel/fork.c                |   11 +++++++++
 kernel/signal.c              |   48 +++++++++++++++++++++++++++++++++++++++----
 3 files changed, 57 insertions(+), 4 deletions(-)
include/linux/sched/signal.h +2 -0
@@ -22,6 +22,7 @@ struct sighand_struct {
 	refcount_t		count;
 	wait_queue_head_t	signalfd_wqh;
 	struct k_sigaction	action[_NSIG];
+	struct sigqueue		*sigqueue_cache;
 };
 
 /*
@@ -349,6 +350,7 @@ extern int send_sig(int, struct task_str
 extern int zap_other_threads(struct task_struct *p);
 extern struct sigqueue *sigqueue_alloc(void);
 extern void sigqueue_free(struct sigqueue *);
+extern void sigqueue_free_cached_entry(struct sigqueue *q);
 extern int send_sigqueue(struct sigqueue *, struct pid *, enum pid_type);
 extern int do_sigaction(int, struct k_sigaction *, struct k_sigaction *);
 
kernel/fork.c +11 -0
@@ -1661,6 +1661,7 @@ static int copy_sighand(unsigned long cl
 	RCU_INIT_POINTER(tsk->sighand, sig);
 	if (!sig)
 		return -ENOMEM;
+	sig->sigqueue_cache = NULL;
 
 	refcount_set(&sig->count, 1);
 	spin_lock_irq(&current->sighand->siglock);
@@ -1677,7 +1678,17 @@ static int copy_sighand(unsigned long cl
 void __cleanup_sighand(struct sighand_struct *sighand)
 {
 	if (refcount_dec_and_test(&sighand->count)) {
+		struct sigqueue *sigqueue = NULL;
+
 		signalfd_cleanup(sighand);
+		spin_lock_irq(&sighand->siglock);
+		if (sighand->sigqueue_cache) {
+			sigqueue = sighand->sigqueue_cache;
+			sighand->sigqueue_cache = NULL;
+		}
+		spin_unlock_irq(&sighand->siglock);
+
+		sigqueue_free_cached_entry(sigqueue);
 		/*
 		 * sighand_cachep is SLAB_TYPESAFE_BY_RCU so we can free it
 		 * without an RCU grace period, see __lock_task_sighand().
kernel/signal.c +44 -4

72 lines not on this page · raw message

#1 lore raw

PATCH v4 1/1 signal: Let tasks cache one sigqueue struct.

The sigqueue caching originated in the PREEMPT_RT tree. A few of the applications, that were ported to Linux, were ported from OS-9. Sending notifications from one task to another via a signal was a common communication model there and so the applications are heavy signal users. Removing the allocation reduces the critical path, avoids locks and so lowers the maximal latency of the task while sending a signal.

After posting the first version a discussion came up whether it wouldn't make sense to have this caching unconditionally and not restricted to PREEMPT_RT only.

The sigqueue flagged SIGQUEUE_PREALLOC is used by the POSIX timer code. It is allocated on initialisation and reused during the lifetime of the timer. The sigqueue is freed once the timer is deleted. The POSIX timer sigqueue has its own caching and unique lifetime pattern and therefore does not fit for the generic caching.

In the common case the signal is dequeued and freed in collect_signal(). At this point, the 'current' task receives the signal and its sighand_struct::siglock is held.

__sigqueue_alloc() is used to allocate a new sigqueue. The task_struct passed as argument is the task that will receive the signal. Its sighand_struct::siglock is acquired (except for SIGQUEUE_PREALLOC allocation which is ignored). Use a cached sigqueue before allocating a new one.

As a result of this pattern, the task sending a signal will use the cache from the task that will receive the signal which in turn caches the signal.

The numbers of system boot followed by an allmod kernel build:
Out of 333216 allocations, 194876 (~58%) were served from the cache. From all free invocations, 4212 were in a path were caching is not done and 329002 sigqueue were cached.

Cache the struct sigqueue in collect_signal() and reuse it for the allocation. Rely on sighand_struct::siglock locking for cache handling which is held during allocation and free. The cache is cleaned once sighand_struct is freed.

---
v3…v4: Repost with Brauner and Torvalds on Cc.
v2…v3: Wrote the patch and its description from scratch without looking
at the previous version.

 include/linux/sched/signal.h |    2 +
 kernel/fork.c                |   11 +++++++++
 kernel/signal.c              |   48 +++++++++++++++++++++++++++++++++++++++----
 3 files changed, 57 insertions(+), 4 deletions(-)
include/linux/sched/signal.h +2 -0
@@ -22,6 +22,7 @@ struct sighand_struct {
 	refcount_t		count;
 	wait_queue_head_t	signalfd_wqh;
 	struct k_sigaction	action[_NSIG];
+	struct sigqueue		*sigqueue_cache;
 };
 
 /*
@@ -349,6 +350,7 @@ extern int send_sig(int, struct task_str
 extern int zap_other_threads(struct task_struct *p);
 extern struct sigqueue *sigqueue_alloc(void);
 extern void sigqueue_free(struct sigqueue *);
+extern void sigqueue_free_cached_entry(struct sigqueue *q);
 extern int send_sigqueue(struct sigqueue *, struct pid *, enum pid_type);
 extern int do_sigaction(int, struct k_sigaction *, struct k_sigaction *);
 
kernel/fork.c +11 -0
@@ -1661,6 +1661,7 @@ static int copy_sighand(unsigned long cl
 	RCU_INIT_POINTER(tsk->sighand, sig);
 	if (!sig)
 		return -ENOMEM;
+	sig->sigqueue_cache = NULL;
 
 	refcount_set(&sig->count, 1);
 	spin_lock_irq(&current->sighand->siglock);
@@ -1677,7 +1678,17 @@ static int copy_sighand(unsigned long cl
 void __cleanup_sighand(struct sighand_struct *sighand)
 {
 	if (refcount_dec_and_test(&sighand->count)) {
+		struct sigqueue *sigqueue = NULL;
+
 		signalfd_cleanup(sighand);
+		spin_lock_irq(&sighand->siglock);
+		if (sighand->sigqueue_cache) {
+			sigqueue = sighand->sigqueue_cache;
+			sighand->sigqueue_cache = NULL;
+		}
+		spin_unlock_irq(&sighand->siglock);
+
+		sigqueue_free_cached_entry(sigqueue);
 		/*
 		 * sighand_cachep is SLAB_TYPESAFE_BY_RCU so we can free it
 		 * without an RCU grace period, see __lock_task_sighand().
kernel/signal.c +44 -4

72 lines not on this page · raw message

#2 lore raw

On Thu, 6 Apr 2023 22:47:21 +0200 Sebastian Andrzej Siewior <[email protected]>

Quoted 6 lines

The sigqueue caching originated in the PREEMPT_RT tree. A few of the applications, that were ported to Linux, were ported from OS-9. Sending notifications from one task to another via a signal was a common communication model there and so the applications are heavy signal users. Removing the allocation reduces the critical path, avoids locks and so lowers the maximal latency of the task while sending a signal.

This is needed only if slab does not work for you.
Why not make it work better instead of peeling slab one more off just to make signal/rt richer?

Quoted 24 lines
kernel/signal.c +0 -0
@@ -432,7 +432,18 @@ static struct sigqueue *
 		return NULL;

=20
if (override_rlimit || likely(sigpending <=3D task_rlimit(t, RLIMIT_SIGPE=
NDING))) {

kernel/signal.c +12 -1
-		q =3D kmem_cache_alloc(sigqueue_cachep, gfp_flags);
+
+		if (!sigqueue_flags) {
+			struct sighand_struct *sighand =3D t->sighand;
+
+			lockdep_assert_held(&sighand->siglock);
+			if (sighand->sigqueue_cache) {
+				q =3D sighand->sigqueue_cache;
+				sighand->sigqueue_cache =3D NULL;
+			}
+		}
+		if (!q)
+			q =3D kmem_cache_alloc(sigqueue_cachep, gfp_flags);
 	} else {
 		print_dropped_signal(sig);
 	}
#3 lore raw
Quoted 7 lines · Hillf Danton

On Thu, 6 Apr 2023 22:47:21 +0200 Sebastian Andrzej Siewior <[email protected]>

Quoted 6 lines

The sigqueue caching originated in the PREEMPT_RT tree. A few of the applications, that were ported to Linux, were ported from OS-9. Sending notifications from one task to another via a signal was a common communication model there and so the applications are heavy signal users. Removing the allocation reduces the critical path, avoids locks and so lowers the maximal latency of the task while sending a signal.

It might lower the _average_ latency, but it certainly doesn't lower the _maximum_ latency, because you have to assume worst case scenario for maximum latency. Which is that there's no sigqueue cached, so you have to go into the slab allocator. And again, worst case scenario is that you have to go into the page allocator from there, and further that you have to run reclaim, and ...

What I find odd about the numbers that you quote:

Quoted 4 lines

The numbers of system boot followed by an allmod kernel build:
Out of 333216 allocations, 194876 (~58%) were served from the cache. From all free invocations, 4212 were in a path were caching is not done and 329002 sigqueue were cached.

... is that there's no absolute numbers. Does it save 1% of the cost of sending a signal? 10%? What does perf say about the cost saved by no longer going into slab? Because the fast path in slab is very fast. It might even be quicker than your fast path for multithreaded applications which have threads running on different NUMA nodes.

#4 lore raw
Quoted 9 lines · Matthew Wilcox
Quoted 7 lines · Hillf Danton

On Thu, 6 Apr 2023 22:47:21 +0200 Sebastian Andrzej Siewior <[email protected]>

Quoted 6 lines

The sigqueue caching originated in the PREEMPT_RT tree. A few of the applications, that were ported to Linux, were ported from OS-9. Sending notifications from one task to another via a signal was a common communication model there and so the applications are heavy signal users. Removing the allocation reduces the critical path, avoids locks and so lowers the maximal latency of the task while sending a signal.

Quoted 6 lines

It might lower the _average_ latency, but it certainly doesn't lower the _maximum_ latency, because you have to assume worst case scenario for maximum latency. Which is that there's no sigqueue cached, so you have to go into the slab allocator. And again, worst case scenario is that you have to go into the page allocator from there, and further that you have to run reclaim, and ...

Yes. The idea is in general not to send more signals in parallel than the available number cached slots.

Quoted 7 lines

What I find odd about the numbers that you quote:

Quoted 4 lines

The numbers of system boot followed by an allmod kernel build:
Out of 333216 allocations, 194876 (~58%) were served from the cache. From all free invocations, 4212 were in a path were caching is not done and 329002 sigqueue were cached.

Quoted 5 lines

... is that there's no absolute numbers. Does it save 1% of the cost of sending a signal? 10%? What does perf say about the cost saved by no longer going into slab? Because the fast path in slab is very fast. It might even be quicker than your fast path for multithreaded applications which have threads running on different NUMA nodes.

I asked for updated numbers and the improvement is not as big as initially reported. There might have been a change in the configuration for the testing an so the improvement is not as big as initially assumed. I'm sorry, but I didn't get any numbers to back anything up. I'm dropping the effort here, thanks for the review :)

Sebastian