[PATCH v5 0/2] uprobes: Improve scalability by reducing the contention on siglock

[PATCH v5 0/2] uprobes: Improve scalability by reducing the contention on siglock

[Liao Chang]

The profiling result of BPF selftest on ARM64 platform reveals the
significant contention on the current->sighand->siglock is the
scalability bottleneck. The reason is also very straightforward that all
producer threads of benchmark have to contend the spinlock mentioned to
resume the TIF_SIGPENDING bit in thread_info that might be removed in
uprobe_deny_signal().

The contention on current->sighand->siglock is unnecessary, this series
remove them thoroughly. I've use the script developed by Andrii in [1]
to run benchmark. The CPU used was Kunpeng916 (Hi1616), 4 NUMA nodes,
64 cores@2.4GHz running the kernel on next tree + the optimization in
[2] for get_xol_insn_slot().

before-opt
----------
uprobe-nop      ( 1 cpus):    0.907 ± 0.003M/s  (  0.907M/s/cpu)
uprobe-nop      ( 2 cpus):    1.676 ± 0.008M/s  (  0.838M/s/cpu)
uprobe-nop      ( 4 cpus):    3.210 ± 0.003M/s  (  0.802M/s/cpu)
uprobe-nop      ( 8 cpus):    4.457 ± 0.003M/s  (  0.557M/s/cpu)
uprobe-nop      (16 cpus):    3.724 ± 0.011M/s  (  0.233M/s/cpu)
uprobe-nop      (32 cpus):    2.761 ± 0.003M/s  (  0.086M/s/cpu)
uprobe-nop      (64 cpus):    1.293 ± 0.015M/s  (  0.020M/s/cpu)

uprobe-push     ( 1 cpus):    0.883 ± 0.001M/s  (  0.883M/s/cpu)
uprobe-push     ( 2 cpus):    1.642 ± 0.005M/s  (  0.821M/s/cpu)
uprobe-push     ( 4 cpus):    3.086 ± 0.002M/s  (  0.771M/s/cpu)
uprobe-push     ( 8 cpus):    3.390 ± 0.003M/s  (  0.424M/s/cpu)
uprobe-push     (16 cpus):    2.652 ± 0.005M/s  (  0.166M/s/cpu)
uprobe-push     (32 cpus):    2.713 ± 0.005M/s  (  0.085M/s/cpu)
uprobe-push     (64 cpus):    1.313 ± 0.009M/s  (  0.021M/s/cpu)

uprobe-ret      ( 1 cpus):    1.774 ± 0.000M/s  (  1.774M/s/cpu)
uprobe-ret      ( 2 cpus):    3.350 ± 0.001M/s  (  1.675M/s/cpu)
uprobe-ret      ( 4 cpus):    6.604 ± 0.000M/s  (  1.651M/s/cpu)
uprobe-ret      ( 8 cpus):    6.706 ± 0.005M/s  (  0.838M/s/cpu)
uprobe-ret      (16 cpus):    5.231 ± 0.001M/s  (  0.327M/s/cpu)
uprobe-ret      (32 cpus):    5.743 ± 0.003M/s  (  0.179M/s/cpu)
uprobe-ret      (64 cpus):    4.726 ± 0.016M/s  (  0.074M/s/cpu)

after-opt
---------
uprobe-nop      ( 1 cpus):    0.985 ± 0.002M/s  (  0.985M/s/cpu)
uprobe-nop      ( 2 cpus):    1.773 ± 0.005M/s  (  0.887M/s/cpu)
uprobe-nop      ( 4 cpus):    3.304 ± 0.001M/s  (  0.826M/s/cpu)
uprobe-nop      ( 8 cpus):    5.328 ± 0.002M/s  (  0.666M/s/cpu)
uprobe-nop      (16 cpus):    6.475 ± 0.002M/s  (  0.405M/s/cpu)
uprobe-nop      (32 cpus):    4.831 ± 0.082M/s  (  0.151M/s/cpu)
uprobe-nop      (64 cpus):    2.564 ± 0.053M/s  (  0.040M/s/cpu)

uprobe-push     ( 1 cpus):    0.964 ± 0.001M/s  (  0.964M/s/cpu)
uprobe-push     ( 2 cpus):    1.766 ± 0.002M/s  (  0.883M/s/cpu)
uprobe-push     ( 4 cpus):    3.290 ± 0.009M/s  (  0.823M/s/cpu)
uprobe-push     ( 8 cpus):    4.670 ± 0.002M/s  (  0.584M/s/cpu)
uprobe-push     (16 cpus):    5.197 ± 0.004M/s  (  0.325M/s/cpu)
uprobe-push     (32 cpus):    5.068 ± 0.161M/s  (  0.158M/s/cpu)
uprobe-push     (64 cpus):    2.605 ± 0.026M/s  (  0.041M/s/cpu)

uprobe-ret      ( 1 cpus):    1.833 ± 0.001M/s  (  1.833M/s/cpu)
uprobe-ret      ( 2 cpus):    3.384 ± 0.003M/s  (  1.692M/s/cpu)
uprobe-ret      ( 4 cpus):    6.677 ± 0.004M/s  (  1.669M/s/cpu)
uprobe-ret      ( 8 cpus):    6.854 ± 0.005M/s  (  0.857M/s/cpu)
uprobe-ret      (16 cpus):    6.508 ± 0.006M/s  (  0.407M/s/cpu)
uprobe-ret      (32 cpus):    5.793 ± 0.009M/s  (  0.181M/s/cpu)
uprobe-ret      (64 cpus):    4.743 ± 0.016M/s  (  0.074M/s/cpu)

Above benchmark results demonstrates a obivious improvement in the
scalability of trig-uprobe-nop and trig-uprobe-push, the peak throughput
of which are from 4.5M/s to 6.4M/s and 3.3M/s to 5.1M/s individually.

v5->v4:
Nothing new, just rebase to next-20250124.

v4->v3:
1. Rebase v3 [3] to the lateset tip/perf/core.
2. Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
3. Acked-by: Oleg Nesterov <oleg@redhat.com>

v3->v2:
Renaming the flag in [2/2], s/deny_signal/signal_denied/g.

v2->v1:
Oleg pointed out the _DENY_SIGNAL will be replaced by _ACK upon the
completion of singlestep which leads to handle_singlestep() has no
chance to restore the removed TIF_SIGPENDING [3] and some case in
question. So this revision proposes to use a flag in uprobe_task to
track the denied TIF_SIGPENDING instead of new UPROBE_SSTEP state.

[1] https://lore.kernel.org/all/20240731214256.3588718-1-andrii@kernel.org
[2] https://lore.kernel.org/all/20240727094405.1362496-1-liaochang1@huawei.com
[3] https://lore.kernel.org/all/20240815014629.2685155-1-liaochang1@huawei.com/ 


Liao Chang (2):
  uprobes: Remove redundant spinlock in uprobe_deny_signal()
  uprobes: Remove the spinlock within handle_singlestep()

 include/linux/uprobes.h |  1 +
 kernel/events/uprobes.c | 10 +++++-----
 2 files changed, 6 insertions(+), 5 deletions(-)

-- 
2.34.1

[PATCH v5 1/2] uprobes: Remove redundant spinlock in uprobe_deny_signal()

[Liao Chang]

Since clearing a bit in thread_info is an atomic operation, the spinlock
is redundant and can be removed, reducing lock contention is good for
performance.

Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
Acked-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Liao Chang <liaochang1@huawei.com>
---
 kernel/events/uprobes.c | 2 --
 1 file changed, 2 deletions(-)

diff --git a/kernel/events/uprobes.c b/kernel/events/uprobes.c
index e421a5f2ec7d..7a3348dfedeb 100644
--- a/kernel/events/uprobes.c
+++ b/kernel/events/uprobes.c
@@ -2298,9 +2298,7 @@ bool uprobe_deny_signal(void)
 	WARN_ON_ONCE(utask->state != UTASK_SSTEP);
 
 	if (task_sigpending(t)) {
-		spin_lock_irq(&t->sighand->siglock);
 		clear_tsk_thread_flag(t, TIF_SIGPENDING);
-		spin_unlock_irq(&t->sighand->siglock);
 
 		if (__fatal_signal_pending(t) || arch_uprobe_xol_was_trapped(t)) {
 			utask->state = UTASK_SSTEP_TRAPPED;
-- 
2.34.1

[PATCH v5 2/2] uprobes: Remove the spinlock within handle_singlestep()

[Liao Chang]

This patch introduces a flag to track TIF_SIGPENDING is suppress
temporarily during the uprobe single-step. Upon uprobe singlestep is
handled and the flag is confirmed, it could resume the TIF_SIGPENDING
directly without acquiring the siglock in most case, then reducing
contention and improving overall performance.

I've use the script developed by Andrii in [1] to run benchmark. The CPU
used was Kunpeng916 (Hi1616), 4 NUMA nodes, 64 cores@2.4GHz running the
kernel on next tree + the optimization for get_xol_insn_slot() [2].

before-opt
----------
uprobe-nop      ( 1 cpus):    0.907 ± 0.003M/s  (  0.907M/s/cpu)
uprobe-nop      ( 2 cpus):    1.676 ± 0.008M/s  (  0.838M/s/cpu)
uprobe-nop      ( 4 cpus):    3.210 ± 0.003M/s  (  0.802M/s/cpu)
uprobe-nop      ( 8 cpus):    4.457 ± 0.003M/s  (  0.557M/s/cpu)
uprobe-nop      (16 cpus):    3.724 ± 0.011M/s  (  0.233M/s/cpu)
uprobe-nop      (32 cpus):    2.761 ± 0.003M/s  (  0.086M/s/cpu)
uprobe-nop      (64 cpus):    1.293 ± 0.015M/s  (  0.020M/s/cpu)

uprobe-push     ( 1 cpus):    0.883 ± 0.001M/s  (  0.883M/s/cpu)
uprobe-push     ( 2 cpus):    1.642 ± 0.005M/s  (  0.821M/s/cpu)
uprobe-push     ( 4 cpus):    3.086 ± 0.002M/s  (  0.771M/s/cpu)
uprobe-push     ( 8 cpus):    3.390 ± 0.003M/s  (  0.424M/s/cpu)
uprobe-push     (16 cpus):    2.652 ± 0.005M/s  (  0.166M/s/cpu)
uprobe-push     (32 cpus):    2.713 ± 0.005M/s  (  0.085M/s/cpu)
uprobe-push     (64 cpus):    1.313 ± 0.009M/s  (  0.021M/s/cpu)

uprobe-ret      ( 1 cpus):    1.774 ± 0.000M/s  (  1.774M/s/cpu)
uprobe-ret      ( 2 cpus):    3.350 ± 0.001M/s  (  1.675M/s/cpu)
uprobe-ret      ( 4 cpus):    6.604 ± 0.000M/s  (  1.651M/s/cpu)
uprobe-ret      ( 8 cpus):    6.706 ± 0.005M/s  (  0.838M/s/cpu)
uprobe-ret      (16 cpus):    5.231 ± 0.001M/s  (  0.327M/s/cpu)
uprobe-ret      (32 cpus):    5.743 ± 0.003M/s  (  0.179M/s/cpu)
uprobe-ret      (64 cpus):    4.726 ± 0.016M/s  (  0.074M/s/cpu)

after-opt
---------
uprobe-nop      ( 1 cpus):    0.985 ± 0.002M/s  (  0.985M/s/cpu)
uprobe-nop      ( 2 cpus):    1.773 ± 0.005M/s  (  0.887M/s/cpu)
uprobe-nop      ( 4 cpus):    3.304 ± 0.001M/s  (  0.826M/s/cpu)
uprobe-nop      ( 8 cpus):    5.328 ± 0.002M/s  (  0.666M/s/cpu)
uprobe-nop      (16 cpus):    6.475 ± 0.002M/s  (  0.405M/s/cpu)
uprobe-nop      (32 cpus):    4.831 ± 0.082M/s  (  0.151M/s/cpu)
uprobe-nop      (64 cpus):    2.564 ± 0.053M/s  (  0.040M/s/cpu)

uprobe-push     ( 1 cpus):    0.964 ± 0.001M/s  (  0.964M/s/cpu)
uprobe-push     ( 2 cpus):    1.766 ± 0.002M/s  (  0.883M/s/cpu)
uprobe-push     ( 4 cpus):    3.290 ± 0.009M/s  (  0.823M/s/cpu)
uprobe-push     ( 8 cpus):    4.670 ± 0.002M/s  (  0.584M/s/cpu)
uprobe-push     (16 cpus):    5.197 ± 0.004M/s  (  0.325M/s/cpu)
uprobe-push     (32 cpus):    5.068 ± 0.161M/s  (  0.158M/s/cpu)
uprobe-push     (64 cpus):    2.605 ± 0.026M/s  (  0.041M/s/cpu)

uprobe-ret      ( 1 cpus):    1.833 ± 0.001M/s  (  1.833M/s/cpu)
uprobe-ret      ( 2 cpus):    3.384 ± 0.003M/s  (  1.692M/s/cpu)
uprobe-ret      ( 4 cpus):    6.677 ± 0.004M/s  (  1.669M/s/cpu)
uprobe-ret      ( 8 cpus):    6.854 ± 0.005M/s  (  0.857M/s/cpu)
uprobe-ret      (16 cpus):    6.508 ± 0.006M/s  (  0.407M/s/cpu)
uprobe-ret      (32 cpus):    5.793 ± 0.009M/s  (  0.181M/s/cpu)
uprobe-ret      (64 cpus):    4.743 ± 0.016M/s  (  0.074M/s/cpu)

Above benchmark results demonstrates a obivious improvement in the
scalability of trig-uprobe-nop and trig-uprobe-push, the peak throughput
of which are from 4.5M/s to 6.4M/s and 3.3M/s to 5.1M/s individually.

[1] https://lore.kernel.org/all/20240731214256.3588718-1-andrii@kernel.org
[2] https://lore.kernel.org/all/20240727094405.1362496-1-liaochang1@huawei.com

Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
Acked-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Liao Chang <liaochang1@huawei.com>
---
 include/linux/uprobes.h | 1 +
 kernel/events/uprobes.c | 8 +++++---
 2 files changed, 6 insertions(+), 3 deletions(-)

diff --git a/include/linux/uprobes.h b/include/linux/uprobes.h
index b1df7d792fa1..a40efdda9052 100644
--- a/include/linux/uprobes.h
+++ b/include/linux/uprobes.h
@@ -143,6 +143,7 @@ struct uprobe_task {
 
 	struct uprobe			*active_uprobe;
 	unsigned long			xol_vaddr;
+	bool				signal_denied;
 
 	struct arch_uprobe              *auprobe;
 };
diff --git a/kernel/events/uprobes.c b/kernel/events/uprobes.c
index 7a3348dfedeb..597b9e036e5f 100644
--- a/kernel/events/uprobes.c
+++ b/kernel/events/uprobes.c
@@ -2298,6 +2298,7 @@ bool uprobe_deny_signal(void)
 	WARN_ON_ONCE(utask->state != UTASK_SSTEP);
 
 	if (task_sigpending(t)) {
+		utask->signal_denied = true;
 		clear_tsk_thread_flag(t, TIF_SIGPENDING);
 
 		if (__fatal_signal_pending(t) || arch_uprobe_xol_was_trapped(t)) {
@@ -2731,9 +2732,10 @@ static void handle_singlestep(struct uprobe_task *utask, struct pt_regs *regs)
 	utask->state = UTASK_RUNNING;
 	xol_free_insn_slot(utask);
 
-	spin_lock_irq(&current->sighand->siglock);
-	recalc_sigpending(); /* see uprobe_deny_signal() */
-	spin_unlock_irq(&current->sighand->siglock);
+	if (utask->signal_denied) {
+		set_thread_flag(TIF_SIGPENDING);
+		utask->signal_denied = false;
+	}
 
 	if (unlikely(err)) {
 		uprobe_warn(current, "execute the probed insn, sending SIGILL.");
-- 
2.34.1

Re: [PATCH v5 1/2] uprobes: Remove redundant spinlock in uprobe_deny_signal()

[Steven Rostedt]

On Fri, 24 Jan 2025 09:38:25 +0000
Liao Chang <liaochang1@huawei.com> wrote:

> Since clearing a bit in thread_info is an atomic operation, the spinlock
> is redundant and can be removed, reducing lock contention is good for
> performance.

Although this patch is probably fine, the change log suggests a dangerous
precedence. Just because clearing a flag is atomic, that alone does not
guarantee that it doesn't need spin locks around it.

There may be another path that tests the flag within a spin lock, and then
does a bunch of work assuming that the flag does not change while it is
doing that work. That other path would require a spin lock around the
clearing of the flag elsewhere.

I don't know this code well enough to know if this has that scenario, and
seeing the Acked-by from Oleg, I'm assuming it does not. But in any case,
the change log needs to give a better rationale for removing a spin lock than
just "clearing a flag atomically doesn't need a spin lock"!

-- Steve


> 
> Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
> Acked-by: Oleg Nesterov <oleg@redhat.com>
> Signed-off-by: Liao Chang <liaochang1@huawei.com>

Re: [PATCH v5 1/2] uprobes: Remove redundant spinlock in uprobe_deny_signal()

[Oleg Nesterov]

On 01/24, Steven Rostedt wrote:
>
> On Fri, 24 Jan 2025 09:38:25 +0000
> Liao Chang <liaochang1@huawei.com> wrote:
>
> > Since clearing a bit in thread_info is an atomic operation, the spinlock
> > is redundant and can be removed, reducing lock contention is good for
> > performance.
>
> Although this patch is probably fine, the change log suggests a dangerous
> precedence. Just because clearing a flag is atomic, that alone does not
> guarantee that it doesn't need spin locks around it.

Yes. And iirc we already have the lockless users of clear(TIF_SIGPENDING)
(some if not most of them look buggy). But afaics in this (very special)
case it should be fine.

See also https://lore.kernel.org/all/20240812120738.GC11656@redhat.com/

> There may be another path that tests the flag within a spin lock,

Yes, retarget_shared_pending() or the complete_signal/wants_signal loop.
That is why it was decided to take siglock in uprobe_deny_signal(), just
to be "safe".

But I still think this patch is fine. The current task is going to execute
a single insn which can't enter the kernel and/or return to the userspace
before it calls handle_singlestep() and restores TIF_SIGPENDING. We do not
care if it races with another source of TIF_SIGPENDING.

The only problem is that task_sigpending() from another task can "wrongly"
return false in this window, but I don't see any problem.

Oleg.

Re: [PATCH v5 1/2] uprobes: Remove redundant spinlock in uprobe_deny_signal()

[Oleg Nesterov]

On 01/24, Oleg Nesterov wrote:
>
> But I still think this patch is fine. The current task is going to execute
> a single insn which can't enter the kernel and/or return to the userspace
                      ^^^^^^^^^^^^^^^^^^^^^^
I mean't, it can't do syscall, sorry for the possible confusion.

Oleg.

Re: [PATCH v5 0/2] uprobes: Improve scalability by reducing the contention on siglock

[Peter Zijlstra]

On Fri, Jan 24, 2025 at 09:38:24AM +0000, Liao Chang wrote:
> Liao Chang (2):
>   uprobes: Remove redundant spinlock in uprobe_deny_signal()
>   uprobes: Remove the spinlock within handle_singlestep()
> 
>  include/linux/uprobes.h |  1 +
>  kernel/events/uprobes.c | 10 +++++-----
>  2 files changed, 6 insertions(+), 5 deletions(-)

Thanks, I've picked up the patches but will not merge them until post
-rc1.