Skip to contentSkip to conversation

perf/ftrace: Sanitize perf function-trace interaction

PATCH In review Peter Zijlstra 2 participants 9 messages Oct 10, 2017

Subsystems (3)
lore
Share & tools
lore

Peter Zijlstra

author maintainer
#0 lore raw

cover perf/ftrace: Sanitize perf function-trace interaction

less is more.. :-)

---
 include/linux/ftrace.h          | 57 ++-------------------------------------------------------
 include/linux/perf_event.h      |  2 +-
 include/linux/trace_events.h    |  4 ++--
 kernel/events/core.c            | 13 ++++---------
 kernel/trace/ftrace.c           | 55 ++++++-------------------------------------------------
 kernel/trace/trace_event_perf.c | 70 ++++++++++++++++++++++++++++++++++++----------------------------------
 kernel/trace/trace_kprobe.c     |  4 ++--
 kernel/trace/trace_syscalls.c   |  4 ++--
 kernel/trace/trace_uprobe.c     |  2 +-
 9 files changed, 56 insertions(+), 155 deletions(-)
author maintainer
#1 lore raw

PATCH 1/3 perf/ftrace: Revert ("perf/ftrace: Fix double traces of perf on ftrace:function")

Revert commit:

75e8387685f6 ("perf/ftrace: Fix double traces of perf on ftrace:function")

The reason I insta stumbled on that patch is that it only addresses the ftrace situation and doesn't mention the other _5_ places that use this interface. It doesn't explain why those don't have the problem and if not, why their solution doesn't work for ftrace.

It doesn't, but this is just putting more duct tape on.

Cc Zhou Chengming <[email protected]>
Cc Jiri Olsa <[email protected]>
Cc Steven Rostedt (VMware) <[email protected]>
Cc Ingo Molnar <[email protected]>
---
 include/linux/perf_event.h      |    2 +-
 include/linux/trace_events.h    |    4 ++--
 kernel/events/core.c            |   13 ++++---------
 kernel/trace/trace_event_perf.c |    4 +---
 kernel/trace/trace_kprobe.c     |    4 ++--
 kernel/trace/trace_syscalls.c   |    4 ++--
 kernel/trace/trace_uprobe.c     |    2 +-
 7 files changed, 13 insertions(+), 20 deletions(-)
include/linux/perf_event.h +1 -1
@@ -1169,7 +1169,7 @@ extern void perf_event_init(void);
 extern void perf_tp_event(u16 event_type, u64 count, void *record,
 			  int entry_size, struct pt_regs *regs,
 			  struct hlist_head *head, int rctx,
-			  struct task_struct *task, struct perf_event *event);
+			  struct task_struct *task);
 extern void perf_bp_event(struct perf_event *event, void *data);
 
 #ifndef perf_misc_flags
include/linux/trace_events.h +2 -2
@@ -507,9 +507,9 @@ void perf_trace_run_bpf_submit(void *raw
 static inline void
 perf_trace_buf_submit(void *raw_data, int size, int rctx, u16 type,
 		       u64 count, struct pt_regs *regs, void *head,
-		       struct task_struct *task, struct perf_event *event)
+		       struct task_struct *task)
 {
-	perf_tp_event(type, count, raw_data, size, regs, head, rctx, task, event);
+	perf_tp_event(type, count, raw_data, size, regs, head, rctx, task);
 }
 #endif
 
kernel/events/core.c +4 -9
@@ -7873,15 +7873,16 @@ void perf_trace_run_bpf_submit(void *raw
 		}
 	}
 	perf_tp_event(call->event.type, count, raw_data, size, regs, head,
-		      rctx, task, NULL);
+		      rctx, task);
 }
 EXPORT_SYMBOL_GPL(perf_trace_run_bpf_submit);
 
 void perf_tp_event(u16 event_type, u64 count, void *record, int entry_size,
 		   struct pt_regs *regs, struct hlist_head *head, int rctx,
-		   struct task_struct *task, struct perf_event *event)
+		   struct task_struct *task)
 {
 	struct perf_sample_data data;
+	struct perf_event *event;
 
 	struct perf_raw_record raw = {
 		.frag = {
@@ -7895,15 +7896,9 @@ void perf_tp_event(u16 event_type, u64 c
 
 	perf_trace_buf_update(record, event_type);
 
-	/* Use the given event instead of the hlist */
-	if (event) {
+	hlist_for_each_entry_rcu(event, head, hlist_entry) {
 		if (perf_tp_event_match(event, &data, regs))
 			perf_swevent_event(event, count, &data, regs);
-	} else {
-		hlist_for_each_entry_rcu(event, head, hlist_entry) {
-			if (perf_tp_event_match(event, &data, regs))
-				perf_swevent_event(event, count, &data, regs);
-		}
 	}
 
 	/*
kernel/trace/trace_event_perf.c +1 -3
@@ -306,7 +306,6 @@ static void
 perf_ftrace_function_call(unsigned long ip, unsigned long parent_ip,
 			  struct ftrace_ops *ops, struct pt_regs *pt_regs)
 {
-	struct perf_event *event;
 	struct ftrace_entry *entry;
 	struct hlist_head *head;
 	struct pt_regs regs;
@@ -330,9 +329,8 @@ perf_ftrace_function_call(unsigned long
 
 	entry->ip = ip;
 	entry->parent_ip = parent_ip;
-	event = container_of(ops, struct perf_event, ftrace_ops);
 	perf_trace_buf_submit(entry, ENTRY_SIZE, rctx, TRACE_FN,
-			      1, &regs, head, NULL, event);
+			      1, &regs, head, NULL);
 
 #undef ENTRY_SIZE
 }
kernel/trace/trace_kprobe.c +2 -2
@@ -1200,7 +1200,7 @@ kprobe_perf_func(struct trace_kprobe *tk
 	memset(&entry[1], 0, dsize);
 	store_trace_args(sizeof(*entry), &tk->tp, regs, (u8 *)&entry[1], dsize);
 	perf_trace_buf_submit(entry, size, rctx, call->event.type, 1, regs,
-			      head, NULL, NULL);
+			      head, NULL);
 }
 NOKPROBE_SYMBOL(kprobe_perf_func);
 
@@ -1236,7 +1236,7 @@ kretprobe_perf_func(struct trace_kprobe
 	entry->ret_ip = (unsigned long)ri->ret_addr;
 	store_trace_args(sizeof(*entry), &tk->tp, regs, (u8 *)&entry[1], dsize);
 	perf_trace_buf_submit(entry, size, rctx, call->event.type, 1, regs,
-			      head, NULL, NULL);
+			      head, NULL);
 }
 NOKPROBE_SYMBOL(kretprobe_perf_func);
 #endif	/* CONFIG_PERF_EVENTS */
kernel/trace/trace_syscalls.c +2 -2
@@ -622,7 +622,7 @@ static void perf_syscall_enter(void *ign
 
 	perf_trace_buf_submit(rec, size, rctx,
 			      sys_data->enter_event->event.type, 1, regs,
-			      head, NULL, NULL);
+			      head, NULL);
 }
 
 static int perf_sysenter_enable(struct trace_event_call *call)
@@ -716,7 +716,7 @@ static void perf_syscall_exit(void *igno
 	}
 
 	perf_trace_buf_submit(rec, size, rctx, sys_data->exit_event->event.type,
-			      1, regs, head, NULL, NULL);
+			      1, regs, head, NULL);
 }
 
 static int perf_sysexit_enable(struct trace_event_call *call)
kernel/trace/trace_uprobe.c +1 -1
@@ -1156,7 +1156,7 @@ static void __uprobe_perf_func(struct tr
 	}
 
 	perf_trace_buf_submit(entry, size, rctx, call->event.type, 1, regs,
-			      head, NULL, NULL);
+			      head, NULL);
  out:
 	preempt_enable();
 }
maintainer
#2 lore raw

Re: [PATCH 1/3] perf/ftrace: Revert ("perf/ftrace: Fix double traces of perf on ftrace:function")

Quoted 5 lines · On Tue, 10 Oct 2017 18:31:25 +0200 Peter Zijlstra <[email protected]> wrote:

Revert commit:

75e8387685f6 ("perf/ftrace: Fix double traces of perf on ftrace:function")

The reason I insta stumbled on that patch is that it only addresses the ftrace

I'm trying to figure out what you meant by "insta".

-- Steve

Quoted 12 lines

situation and doesn't mention the other _5_ places that use this interface. It doesn't explain why those don't have the problem and if not, why their solution doesn't work for ftrace.

It doesn't, but this is just putting more duct tape on.

Cc Zhou Chengming <[email protected]>
Cc Jiri Olsa <[email protected]>
Cc Steven Rostedt (VMware) <[email protected]>
Cc Ingo Molnar <[email protected]>
author maintainer
#3 lore raw

Re: [PATCH 1/3] perf/ftrace: Revert ("perf/ftrace: Fix double traces of perf on ftrace:function")

Quoted 10 lines · Steven Rostedt
Quoted 5 lines · On Tue, 10 Oct 2017 18:31:25 +0200 Peter Zijlstra <[email protected]> wrote:

Revert commit:

75e8387685f6 ("perf/ftrace: Fix double traces of perf on ftrace:function")

The reason I insta stumbled on that patch is that it only addresses the ftrace

I'm trying to figure out what you meant by "insta".

instantly, immediately. It triggered a significant spike on the official measure of review: WTFs/min.

author maintainer
#4 lore raw

PATCH 2/3 perf/ftrace: Fix function trace events

In reply to Peter Zijlstra

The function-trace <-> perf interface is a tad messed up. Where all the other trace <-> perf interfaces use a single trace hook registration and use per-cpu RCU based hlist to iterate the events, function-trace actually needs multiple hook registrations in order to minimize function entry patching when filters are present.

The end result is that we iterate events both on the trace hook and on the hlist, which results in reporting events multiple times.

Since function-trace cannot use the regular scheme, fix it the other way around, use singleton hlists.

---
 kernel/trace/trace_event_perf.c |   68 +++++++++++++++++++++-------------------
 1 file changed, 36 insertions(+), 32 deletions(-)
kernel/trace/trace_event_perf.c +36 -32
@@ -240,27 +240,31 @@ void perf_trace_destroy(struct perf_even
 int perf_trace_add(struct perf_event *p_event, int flags)
 {
 	struct trace_event_call *tp_event = p_event->tp_event;
-	struct hlist_head __percpu *pcpu_list;
-	struct hlist_head *list;
 
-	pcpu_list = tp_event->perf_events;
-	if (WARN_ON_ONCE(!pcpu_list))
-		return -EINVAL;
+	if (!tp_event->class->reg(tp_event, TRACE_REG_PERF_ADD, p_event)) {
+		struct hlist_head __percpu *pcpu_list;
+		struct hlist_head *list;
+
+		pcpu_list = tp_event->perf_events;
+		if (WARN_ON_ONCE(!pcpu_list))
+			return -EINVAL;
 
-	if (!(flags & PERF_EF_START))
-		p_event->hw.state = PERF_HES_STOPPED;
+		if (!(flags & PERF_EF_START))
+			p_event->hw.state = PERF_HES_STOPPED;
 
-	list = this_cpu_ptr(pcpu_list);
-	hlist_add_head_rcu(&p_event->hlist_entry, list);
+		list = this_cpu_ptr(pcpu_list);
+		hlist_add_head_rcu(&p_event->hlist_entry, list);
+	}
 
-	return tp_event->class->reg(tp_event, TRACE_REG_PERF_ADD, p_event);
+	return 0;
 }
 
 void perf_trace_del(struct perf_event *p_event, int flags)
 {
 	struct trace_event_call *tp_event = p_event->tp_event;
-	hlist_del_rcu(&p_event->hlist_entry);
-	tp_event->class->reg(tp_event, TRACE_REG_PERF_DEL, p_event);
+
+	if (!tp_event->class->reg(tp_event, TRACE_REG_PERF_DEL, p_event))
+		hlist_del_rcu(&p_event->hlist_entry);
 }
 
 void *perf_trace_buf_alloc(int size, struct pt_regs **regs, int *rctxp)
@@ -306,15 +310,19 @@ static void
 perf_ftrace_function_call(unsigned long ip, unsigned long parent_ip,
 			  struct ftrace_ops *ops, struct pt_regs *pt_regs)
 {
+	struct hlist_head head = HLIST_HEAD_INIT;
 	struct ftrace_entry *entry;
-	struct hlist_head *head;
+	struct perf_event *event;
 	struct pt_regs regs;
 	int rctx;
 
-	head = this_cpu_ptr(event_function.perf_events);
-	if (hlist_empty(head))
+	event = container_of(ops, struct perf_event, ftrace_ops);
+
+	if ((unsigned long)event->ftrace_ops.private != smp_processor_id())
 		return;
 
+	hlist_add_head(&event->hlist_entry, &head);
+
 #define ENTRY_SIZE (ALIGN(sizeof(struct ftrace_entry) + sizeof(u32), \
 		    sizeof(u64)) - sizeof(u32))
 
@@ -330,17 +338,21 @@ perf_ftrace_function_call(unsigned long
 	entry->ip = ip;
 	entry->parent_ip = parent_ip;
 	perf_trace_buf_submit(entry, ENTRY_SIZE, rctx, TRACE_FN,
-			      1, &regs, head, NULL);
+			      1, &regs, &head, NULL);
 
 #undef ENTRY_SIZE
+
+	hlist_del_init(&event->hlist_entry);
 }
 
 static int perf_ftrace_function_register(struct perf_event *event)
 {
 	struct ftrace_ops *ops = &event->ftrace_ops;
 
-	ops->flags |= FTRACE_OPS_FL_PER_CPU | FTRACE_OPS_FL_RCU;
-	ops->func = perf_ftrace_function_call;
+	ops->flags   |= FTRACE_OPS_FL_RCU;
+	ops->func    = perf_ftrace_function_call;
+	ops->private = (void *)(unsigned long)nr_cpu_ids;
+
 	return register_ftrace_function(ops);
 }
 
@@ -352,19 +364,11 @@ static int perf_ftrace_function_unregist
 	return ret;
 }
 
-static void perf_ftrace_function_enable(struct perf_event *event)
-{
-	ftrace_function_local_enable(&event->ftrace_ops);
-}
-
-static void perf_ftrace_function_disable(struct perf_event *event)
-{
-	ftrace_function_local_disable(&event->ftrace_ops);
-}
-
 int perf_ftrace_event_register(struct trace_event_call *call,
 			       enum trace_reg type, void *data)
 {
+	struct perf_event *event = data;
+
 	switch (type) {
 	case TRACE_REG_REGISTER:
 	case TRACE_REG_UNREGISTER:
@@ -377,11 +381,11 @@ int perf_ftrace_event_register(struct tr
 	case TRACE_REG_PERF_CLOSE:
 		return perf_ftrace_function_unregister(data);
 	case TRACE_REG_PERF_ADD:
-		perf_ftrace_function_enable(data);
-		return 0;
+		event->ftrace_ops.private = (void *)(unsigned long)smp_processor_id();
+		return 1;
 	case TRACE_REG_PERF_DEL:
-		perf_ftrace_function_disable(data);
-		return 0;
+		event->ftrace_ops.private = (void *)(unsigned long)nr_cpu_ids;
+		return 1;
 	}
 
 	return -EINVAL;
maintainer
#5 lore raw

Re: [PATCH 2/3] perf/ftrace: Fix function trace events

Quoted 8 lines · On Tue, 10 Oct 2017 18:31:26 +0200 Peter Zijlstra <[email protected]> wrote:
---
 kernel/trace/trace_event_perf.c |   68 +++++++++++++++++++++-------------------
 1 file changed, 36 insertions(+), 32 deletions(-)

kernel/trace/trace_event_perf.c void perf_trace_destroy(struct perf_even · line 248

Quoted 9 lines
kernel/trace/trace_event_perf.c +0 -5
@@ -240,27 +240,31 @@ void perf_trace_destroy(struct perf_even
 int perf_trace_add(struct perf_event *p_event, int flags)
 {
 	struct trace_event_call *tp_event = p_event->tp_event;
-	struct hlist_head __percpu *pcpu_list;
-	struct hlist_head *list;
 
-	pcpu_list = tp_event->perf_events;
-	if (WARN_ON_ONCE(!pcpu_list))
-		return -EINVAL;

You're going to add a comment here on v2?

Also, this is highly subtle, that ftrace reg returns non-zero, and all others return zero. It may be good to add a comment in include/linux/trace_events.h by the TRACE_REG_PERF_ADD and DEL enums, stating what is expected when they are passed in.

Quoted 40 lines
kernel/trace/trace_event_perf.c +17 -7
+	if (!tp_event->class->reg(tp_event, TRACE_REG_PERF_ADD, p_event)) {
+		struct hlist_head __percpu *pcpu_list;
+		struct hlist_head *list;
+
+		pcpu_list = tp_event->perf_events;
+		if (WARN_ON_ONCE(!pcpu_list))
+			return -EINVAL;
 
-	if (!(flags & PERF_EF_START))
-		p_event->hw.state = PERF_HES_STOPPED;
+		if (!(flags & PERF_EF_START))
+			p_event->hw.state = PERF_HES_STOPPED;
 
-	list = this_cpu_ptr(pcpu_list);
-	hlist_add_head_rcu(&p_event->hlist_entry, list);
+		list = this_cpu_ptr(pcpu_list);
+		hlist_add_head_rcu(&p_event->hlist_entry, list);
+	}
 
-	return tp_event->class->reg(tp_event, TRACE_REG_PERF_ADD, p_event);
+	return 0;
 }
 
 void perf_trace_del(struct perf_event *p_event, int flags)
 {
 	struct trace_event_call *tp_event = p_event->tp_event;
-	hlist_del_rcu(&p_event->hlist_entry);
-	tp_event->class->reg(tp_event, TRACE_REG_PERF_DEL, p_event);
+
+	if (!tp_event->class->reg(tp_event, TRACE_REG_PERF_DEL, p_event))
+		hlist_del_rcu(&p_event->hlist_entry);
 }
 
 void *perf_trace_buf_alloc(int size, struct pt_regs **regs, int *rctxp)
@@ -306,15 +310,19 @@ static void
 perf_ftrace_function_call(unsigned long ip, unsigned long parent_ip,
 			  struct ftrace_ops *ops, struct pt_regs *pt_regs)
 {
+	struct hlist_head head = HLIST_HEAD_INIT;
 	struct ftrace_entry *entry;

kernel/trace/trace_event_perf.c static void · line 321

Quoted 10 lines
kernel/trace/trace_event_perf.c +4 -3
-	struct hlist_head *head;
+	struct perf_event *event;
 	struct pt_regs regs;
 	int rctx;
 
-	head = this_cpu_ptr(event_function.perf_events);
-	if (hlist_empty(head))
+	event = container_of(ops, struct perf_event, ftrace_ops);
+
+	if ((unsigned long)event->ftrace_ops.private != smp_processor_id())

Could you also just do:

	if ((unsigned long)ops->private != smp_processor_id())

?

And then you could get the event via container_of() after the test.

Quoted 8 lines
kernel/trace/trace_event_perf.c +2 -0
 		return;
 
+	hlist_add_head(&event->hlist_entry, &head);
+
 #define ENTRY_SIZE (ALIGN(sizeof(struct ftrace_entry) + sizeof(u32), \
 		    sizeof(u64)) - sizeof(u32))
 

kernel/trace/trace_event_perf.c perf_ftrace_function_call(unsigned long · line 345

Quoted 9 lines
kernel/trace/trace_event_perf.c +3 -1
@@ -330,17 +338,21 @@ perf_ftrace_function_call(unsigned long
 	entry->ip = ip;
 	entry->parent_ip = parent_ip;
 	perf_trace_buf_submit(entry, ENTRY_SIZE, rctx, TRACE_FN,
-			      1, &regs, head, NULL);
+			      1, &regs, &head, NULL);
 
 #undef ENTRY_SIZE
+
+	hlist_del_init(&event->hlist_entry);

Hmm, is there a better way to do this? It adds and removes the entry at each function trace.

kernel/trace/trace_event_perf.c perf_ftrace_function_call(unsigned long · line 352

Quoted 9 lines
kernel/trace/trace_event_perf.c +1 -2
 }
 
 static int perf_ftrace_function_register(struct perf_event *event)
 {
 	struct ftrace_ops *ops = &event->ftrace_ops;
 
-	ops->flags |= FTRACE_OPS_FL_PER_CPU | FTRACE_OPS_FL_RCU;
-	ops->func = perf_ftrace_function_call;
+	ops->flags   |= FTRACE_OPS_FL_RCU;

I know this doesn't change the patch, but I don't believe that "|=" is necessary. "=" should work too. But you can make that a separate patch, in case it does break something, and we have a nice bisect to see this was the cause.

Quoted 29 lines
kernel/trace/trace_event_perf.c +5 -10
+	ops->func    = perf_ftrace_function_call;
+	ops->private = (void *)(unsigned long)nr_cpu_ids;
+
 	return register_ftrace_function(ops);
 }
 
@@ -352,19 +364,11 @@ static int perf_ftrace_function_unregist
 	return ret;
 }
 
-static void perf_ftrace_function_enable(struct perf_event *event)
-{
-	ftrace_function_local_enable(&event->ftrace_ops);
-}
-
-static void perf_ftrace_function_disable(struct perf_event *event)
-{
-	ftrace_function_local_disable(&event->ftrace_ops);
-}
-
 int perf_ftrace_event_register(struct trace_event_call *call,
 			       enum trace_reg type, void *data)
 {
+	struct perf_event *event = data;
+
 	switch (type) {
 	case TRACE_REG_REGISTER:
 	case TRACE_REG_UNREGISTER:

kernel/trace/trace_event_perf.c int perf_ftrace_event_register(struct tr · line 384

Quoted 6 lines
kernel/trace/trace_event_perf.c +1 -2
@@ -377,11 +381,11 @@ int perf_ftrace_event_register(struct tr
 	case TRACE_REG_PERF_CLOSE:
 		return perf_ftrace_function_unregister(data);
 	case TRACE_REG_PERF_ADD:
-		perf_ftrace_function_enable(data);
-		return 0;
+		event->ftrace_ops.private = (void *)(unsigned long)smp_processor_id();

I curious, is there a per CPU event, or is this called during sched_switch or something?

-- Steve

Quoted 10 lines
kernel/trace/trace_event_perf.c +3 -2
+		return 1;
 	case TRACE_REG_PERF_DEL:
-		perf_ftrace_function_disable(data);
-		return 0;
+		event->ftrace_ops.private = (void *)(unsigned long)nr_cpu_ids;
+		return 1;
 	}
 
 	return -EINVAL;
author maintainer
#6 lore raw

Re: [PATCH 2/3] perf/ftrace: Fix function trace events

Quoted 18 lines · Steven Rostedt

kernel/trace/trace_event_perf.c void perf_trace_destroy(struct perf_even · line 248

Quoted 12 lines · On Tue, 10 Oct 2017 18:31:26 +0200 Peter Zijlstra <[email protected]> wrote:
kernel/trace/trace_event_perf.c +0 -5
@@ -240,27 +240,31 @@ void perf_trace_destroy(struct perf_even
 int perf_trace_add(struct perf_event *p_event, int flags)
 {
 	struct trace_event_call *tp_event = p_event->tp_event;
-	struct hlist_head __percpu *pcpu_list;
-	struct hlist_head *list;
 
-	pcpu_list = tp_event->perf_events;
-	if (WARN_ON_ONCE(!pcpu_list))
-		return -EINVAL;

You're going to add a comment here on v2?

Quoted 4 lines

Also, this is highly subtle, that ftrace reg returns non-zero, and all others return zero. It may be good to add a comment in include/linux/trace_events.h by the TRACE_REG_PERF_ADD and DEL enums, stating what is expected when they are passed in.

Yeah, I was going to add a comment; your earlier email came in just after I send out these ;-)

Quoted 18 lines
Quoted 5 lines
kernel/trace/trace_event_perf.c +1 -0
 perf_ftrace_function_call(unsigned long ip, unsigned long parent_ip,
 			  struct ftrace_ops *ops, struct pt_regs *pt_regs)
 {
+	struct hlist_head head = HLIST_HEAD_INIT;
 	struct ftrace_entry *entry;

kernel/trace/trace_event_perf.c static void · line 321

Quoted 10 lines
kernel/trace/trace_event_perf.c +4 -3
-	struct hlist_head *head;
+	struct perf_event *event;
 	struct pt_regs regs;
 	int rctx;
 
-	head = this_cpu_ptr(event_function.perf_events);
-	if (hlist_empty(head))
+	event = container_of(ops, struct perf_event, ftrace_ops);
+
+	if ((unsigned long)event->ftrace_ops.private != smp_processor_id())

Could you also just do:

Quoted 3 lines
	if ((unsigned long)ops->private != smp_processor_id())

?

Ah yes of course.

Quoted 14 lines
Quoted 14 lines
kernel/trace/trace_event_perf.c +3 -1
 		return;
 
+	hlist_add_head(&event->hlist_entry, &head);
+
 #define ENTRY_SIZE (ALIGN(sizeof(struct ftrace_entry) + sizeof(u32), \
 		    sizeof(u64)) - sizeof(u32))
 
@@ -330,17 +338,21 @@ perf_ftrace_function_call(unsigned long
 	entry->ip = ip;
 	entry->parent_ip = parent_ip;
 	perf_trace_buf_submit(entry, ENTRY_SIZE, rctx, TRACE_FN,
-			      1, &regs, head, NULL);
+			      1, &regs, &head, NULL);
 
Quoted 6 lines

kernel/trace/trace_event_perf.c perf_ftrace_function_call(unsigned long · line 345

Quoted 3 lines
kernel/trace/trace_event_perf.c +2 -0
 #undef ENTRY_SIZE
+
+	hlist_del_init(&event->hlist_entry);

Hmm, is there a better way to do this? It adds and removes the entry at each function trace.

Hmm, hlist_add_head() doesn't seem to use anything from @n if !@h->first, in which case I could simply remove this. I should probably add a comment to clarify this.

Quoted 6 lines
Quoted 6 lines
kernel/trace/trace_event_perf.c +0 -0
 }
 
 static int perf_ftrace_function_register(struct perf_event *event)
 {
 	struct ftrace_ops *ops = &event->ftrace_ops;
 
Quoted 8 lines

kernel/trace/trace_event_perf.c perf_ftrace_function_call(unsigned long · line 352

Quoted 3 lines
kernel/trace/trace_event_perf.c +1 -2
-	ops->flags |= FTRACE_OPS_FL_PER_CPU | FTRACE_OPS_FL_RCU;
-	ops->func = perf_ftrace_function_call;
+	ops->flags   |= FTRACE_OPS_FL_RCU;

I know this doesn't change the patch, but I don't believe that "|=" is necessary. "=" should work too. But you can make that a separate patch, in case it does break something, and we have a nice bisect to see this was the cause.

Will add a 4th patch to that effect.

Quoted 10 lines

kernel/trace/trace_event_perf.c int perf_ftrace_event_register(struct tr · line 384

Quoted 7 lines
kernel/trace/trace_event_perf.c +1 -2
@@ -377,11 +381,11 @@ int perf_ftrace_event_register(struct tr
 	case TRACE_REG_PERF_CLOSE:
 		return perf_ftrace_function_unregister(data);
 	case TRACE_REG_PERF_ADD:
-		perf_ftrace_function_enable(data);
-		return 0;
+		event->ftrace_ops.private = (void *)(unsigned long)smp_processor_id();

I curious, is there a per CPU event, or is this called during sched_switch or something?

Yes (the latter).

author maintainer
#7 lore raw

PATCH 3/3 ftrace: Kill FTRACE_OPS_FL_PER_CPU

In reply to Peter Zijlstra

The one and only user of FTRACE_OPS_FL_PER_CPU is gone, remove the lot.

---
 include/linux/ftrace.h |   57 +------------------------------------------------
 kernel/trace/ftrace.c  |   55 +++++------------------------------------------
 2 files changed, 8 insertions(+), 104 deletions(-)
include/linux/ftrace.h +2 -55
@@ -78,10 +78,7 @@ ftrace_func_t ftrace_ops_get_func(struct
  * ENABLED - set/unset when ftrace_ops is registered/unregistered
  * DYNAMIC - set when ftrace_ops is registered to denote dynamically
  *           allocated ftrace_ops which need special care
- * PER_CPU - set manualy by ftrace_ops user to denote the ftrace_ops
- *           could be controlled by following calls:
- *             ftrace_function_local_enable
- *             ftrace_function_local_disable
+ *
  * SAVE_REGS - The ftrace_ops wants regs saved at each function called
  *            and passed to the callback. If this flag is set, but the
  *            architecture does not support passing regs
@@ -125,7 +122,7 @@ ftrace_func_t ftrace_ops_get_func(struct
 enum {
 	FTRACE_OPS_FL_ENABLED			= 1 << 0,
 	FTRACE_OPS_FL_DYNAMIC			= 1 << 1,
-	FTRACE_OPS_FL_PER_CPU			= 1 << 2,
+
 	FTRACE_OPS_FL_SAVE_REGS			= 1 << 3,
 	FTRACE_OPS_FL_SAVE_REGS_IF_SUPPORTED	= 1 << 4,
 	FTRACE_OPS_FL_RECURSION_SAFE		= 1 << 5,
@@ -172,7 +169,6 @@ struct ftrace_ops {
 	unsigned long			flags;
 	void				*private;
 	ftrace_func_t			saved_func;
-	int __percpu			*disabled;
 #ifdef CONFIG_DYNAMIC_FTRACE
 	struct ftrace_ops_hash		local_hash;
 	struct ftrace_ops_hash		*func_hash;
@@ -204,55 +200,6 @@ int register_ftrace_function(struct ftra
 int unregister_ftrace_function(struct ftrace_ops *ops);
 void clear_ftrace_function(void);
 
-/**
- * ftrace_function_local_enable - enable ftrace_ops on current cpu
- *
- * This function enables tracing on current cpu by decreasing
- * the per cpu control variable.
- * It must be called with preemption disabled and only on ftrace_ops
- * registered with FTRACE_OPS_FL_PER_CPU. If called without preemption
- * disabled, this_cpu_ptr will complain when CONFIG_DEBUG_PREEMPT is enabled.
- */
-static inline void ftrace_function_local_enable(struct ftrace_ops *ops)
-{
-	if (WARN_ON_ONCE(!(ops->flags & FTRACE_OPS_FL_PER_CPU)))
-		return;
-
-	(*this_cpu_ptr(ops->disabled))--;
-}
-
-/**
- * ftrace_function_local_disable - disable ftrace_ops on current cpu
- *
- * This function disables tracing on current cpu by increasing
- * the per cpu control variable.
- * It must be called with preemption disabled and only on ftrace_ops
- * registered with FTRACE_OPS_FL_PER_CPU. If called without preemption
- * disabled, this_cpu_ptr will complain when CONFIG_DEBUG_PREEMPT is enabled.
- */
-static inline void ftrace_function_local_disable(struct ftrace_ops *ops)
-{
-	if (WARN_ON_ONCE(!(ops->flags & FTRACE_OPS_FL_PER_CPU)))
-		return;
-
-	(*this_cpu_ptr(ops->disabled))++;
-}
-
-/**
- * ftrace_function_local_disabled - returns ftrace_ops disabled value
- *                                  on current cpu
- *
- * This function returns value of ftrace_ops::disabled on current cpu.
- * It must be called with preemption disabled and only on ftrace_ops
- * registered with FTRACE_OPS_FL_PER_CPU. If called without preemption
- * disabled, this_cpu_ptr will complain when CONFIG_DEBUG_PREEMPT is enabled.
- */
-static inline int ftrace_function_local_disabled(struct ftrace_ops *ops)
-{
-	WARN_ON_ONCE(!(ops->flags & FTRACE_OPS_FL_PER_CPU));
-	return *this_cpu_ptr(ops->disabled);
-}
-
 extern void ftrace_stub(unsigned long a0, unsigned long a1,
 			struct ftrace_ops *op, struct pt_regs *regs);
 
kernel/trace/ftrace.c +6 -49

116 lines not on this page · raw message

maintainer
#8 lore raw
In reply to Peter Zijlstra
Quoted 1 line · On Tue, 10 Oct 2017 18:31:24 +0200 Peter Zijlstra <[email protected]> wrote:

less is more.. :-)

I received underwear at a conference that says that exact quote.

-- Steve

Quoted 13 lines
---
 include/linux/ftrace.h          | 57 ++-------------------------------------------------------
 include/linux/perf_event.h      |  2 +-
 include/linux/trace_events.h    |  4 ++--
 kernel/events/core.c            | 13 ++++---------
 kernel/trace/ftrace.c           | 55 ++++++-------------------------------------------------
 kernel/trace/trace_event_perf.c | 70 ++++++++++++++++++++++++++++++++++++----------------------------------
 kernel/trace/trace_kprobe.c     |  4 ++--
 kernel/trace/trace_syscalls.c   |  4 ++--
 kernel/trace/trace_uprobe.c     |  2 +-
 9 files changed, 56 insertions(+), 155 deletions(-)