blob: 06c5583be8b9680048ad42636812324de60b0693 [file] [log] [blame]
From 793eba9e90635c40052aa11dc28c41a249227c72 Mon Sep 17 00:00:00 2001
From: Linus Torvalds <>
Date: Mon, 24 Feb 2020 12:47:14 -0800
Subject: [PATCH] signal: avoid double atomic counter increments for user
commit fda31c50292a5062332fa0343c084bd9f46604d9 upstream.
When queueing a signal, we increment both the users count of pending
signals (for RLIMIT_SIGPENDING tracking) and we increment the refcount
of the user struct itself (because we keep a reference to the user in
the signal structure in order to correctly account for it when freeing).
That turns out to be fairly expensive, because both of them are atomic
updates, and particularly under extreme signal handling pressure on big
machines, you can get a lot of cache contention on the user struct.
That can then cause horrid cacheline ping-pong when you do these
multiple accesses.
So change the reference counting to only pin the user for the _first_
pending signal, and to unpin it when the last pending signal is
dequeued. That means that when a user sees a lot of concurrent signal
queuing - which is the only situation when this matters - the only
atomic access needed is generally the 'sigpending' count update.
This was noticed because of a particularly odd timing artifact on a
dual-socket 96C/192T Cascade Lake platform: when you get into bad
contention, on that machine for some reason seems to be much worse when
the contention happens in the upper 32-byte half of the cacheline.
As a result, the kernel test robot will-it-scale 'signal1' benchmark had
an odd performance regression simply due to random alignment of the
'struct user_struct' (and pointed to a completely unrelated and
apparently nonsensical commit for the regression).
Avoiding the double increments (and decrements on the dequeueing side,
of course) makes for much less contention and hugely improved
performance on that will-it-scale microbenchmark.
Quoting Feng Tang:
"It makes a big difference, that the performance score is tripled! bump
from original 17000 to 54000. Also the gap between 5.0-rc6 and
5.0-rc6+Jiri's patch is reduced to around 2%"
[ The "2% gap" is the odd cacheline placement difference on that
platform: under the extreme contention case, the effect of which half
of the cacheline was hot was 5%, so with the reduced contention the
odd timing artifact is reduced too ]
It does help in the non-contended case too, but is not nearly as
Reported-and-tested-by: Feng Tang <>
Cc: Eric W. Biederman <>
Cc: Huang, Ying <>
Cc: Philip Li <>
Cc: Andi Kleen <>
Cc: Jiri Olsa <>
Cc: Peter Zijlstra <>
Signed-off-by: Linus Torvalds <>
Signed-off-by: Paul Gortmaker <>
diff --git a/kernel/signal.c b/kernel/signal.c
index 6f332b7a8f9e..8d843001b306 100644
--- a/kernel/signal.c
+++ b/kernel/signal.c
@@ -408,27 +408,32 @@ __sigqueue_alloc(int sig, struct task_struct *t, gfp_t flags, int override_rlimi
struct sigqueue *q = NULL;
struct user_struct *user;
+ int sigpending;
* Protect access to @t credentials. This can go away when all
* callers hold rcu read lock.
+ *
+ * NOTE! A pending signal will hold on to the user refcount,
+ * and we get/put the refcount only when the sigpending count
+ * changes from/to zero.
- user = get_uid(__task_cred(t)->user);
- atomic_inc(&user->sigpending);
+ user = __task_cred(t)->user;
+ sigpending = atomic_inc_return(&user->sigpending);
+ if (sigpending == 1)
+ get_uid(user);
- if (override_rlimit ||
- atomic_read(&user->sigpending) <=
- task_rlimit(t, RLIMIT_SIGPENDING)) {
+ if (override_rlimit || likely(sigpending <= task_rlimit(t, RLIMIT_SIGPENDING))) {
q = kmem_cache_alloc(sigqueue_cachep, flags);
} else {
if (unlikely(q == NULL)) {
- atomic_dec(&user->sigpending);
- free_uid(user);
+ if (atomic_dec_and_test(&user->sigpending))
+ free_uid(user);
} else {
q->flags = 0;
@@ -442,8 +447,8 @@ static void __sigqueue_free(struct sigqueue *q)
if (q->flags & SIGQUEUE_PREALLOC)
- atomic_dec(&q->user->sigpending);
- free_uid(q->user);
+ if (atomic_dec_and_test(&q->user->sigpending))
+ free_uid(q->user);
kmem_cache_free(sigqueue_cachep, q);