Re: [RFC v2] perf: Rewrite core context handling

[Ravi Bangoria <ravi.bangoria@amd.com>]

On 22-Aug-22 9:13 PM, Peter Zijlstra wrote:
> On Mon, Aug 22, 2022 at 05:29:11PM +0200, Peter Zijlstra wrote:
>> On Tue, Aug 02, 2022 at 11:41:42AM +0530, Ravi Bangoria wrote:
>>>
>>>> pulling up the ctx->mutex makes things simpler, but also violates the
>>>> locking order vs exec_update_lock.
>>>>
>>>> Pull that lock up as well...
>>>
>>> I'm not able to apply this patch as is but I get the idea. Few
>>> questions below...
>>
>> I was just about to rebase the 'series' to current, let me do that and
>> get back to you on the specifics.
> 
> https://git.kernel.org/pub/scm/linux/kernel/git/peterz/queue.git/log/?h=perf/wip.rewrite

Additional set of changes on top of this tree is required to build and boot,
atleast on my AMD machine:

---
diff --git a/arch/x86/events/perf_event.h b/arch/x86/events/perf_event.h
index ccd231ea6a4e..94fb65d7b291 100644
--- a/arch/x86/events/perf_event.h
+++ b/arch/x86/events/perf_event.h
@@ -1248,7 +1248,7 @@ static inline void amd_pmu_brs_add(struct perf_event *event)
 {
 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 
-	perf_sched_cb_inc(event->ctx->pmu);
+	perf_sched_cb_inc(event->pmu_ctx->pmu);
 	cpuc->lbr_users++;
 	/*
 	 * No need to reset BRS because it is reset
@@ -1263,7 +1263,7 @@ static inline void amd_pmu_brs_del(struct perf_event *event)
 	cpuc->lbr_users--;
 	WARN_ON_ONCE(cpuc->lbr_users < 0);
 
-	perf_sched_cb_dec(event->ctx->pmu);
+	perf_sched_cb_dec(event->pmu_ctx->pmu);
 }
 
 void amd_pmu_brs_sched_task(struct perf_event_pmu_context *pmu_ctx, bool sched_in);
diff --git a/kernel/events/core.c b/kernel/events/core.c
index 31ae032d6783..086e37fa32be 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -843,7 +843,7 @@ static void perf_cgroup_switch(struct task_struct *task)
 
 	WARN_ON_ONCE(cpuctx->ctx.nr_cgroups == 0);
 	if (READ_ONCE(cpuctx->cgrp) == cgrp)
-		continue;
+		return;
 
 	perf_ctx_lock(cpuctx, cpuctx->task_ctx);
 	perf_ctx_disable(&cpuctx->ctx);
@@ -881,7 +881,7 @@ static int perf_cgroup_ensure_storage(struct perf_event *event,
 		heap_size++;
 
 	for_each_possible_cpu(cpu) {
-		cpuctx = this_cpu_ptr(&cpu_context);
+		cpuctx = per_cpu_ptr(&cpu_context, cpu);
 		if (heap_size <= cpuctx->heap_size)
 			continue;
 
@@ -2315,7 +2315,7 @@ __perf_remove_from_context(struct perf_event *event,
 	if (!pmu_ctx->nr_events) {
 		pmu_ctx->rotate_necessary = 0;
 
-		if (ctx->task) {
+		if (ctx->task && ctx->is_active) {
 			struct perf_cpu_pmu_context *cpc;
 
 			cpc = this_cpu_ptr(pmu_ctx->pmu->cpu_pmu_context);
@@ -11972,6 +11972,15 @@ static int perf_copy_attr(struct perf_event_attr __user *uattr,
 	goto out;
 }
 
+static void mutex_lock_double(struct mutex *a, struct mutex *b)
+{
+	if (b < a)
+		swap(a, b);
+
+	mutex_lock(a);
+	mutex_lock_nested(b, SINGLE_DEPTH_NESTING);
+}
+
 static int
 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
 {
---

With this, I can run 'perf test' and perf_event_tests without any error in
dmesg. I'll run perf fuzzer over night and see if it reports any issue.

Thanks,
Ravi