From: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Date: Tue, 12 Mar 2024 19:01:50 +0100
Subject: [PATCH 2/4] perf: Enqueue SIGTRAP always via task_work.

A signal is delivered by raising irq_work() which works from any context
including NMI. irq_work() can be delayed if the architecture does not
provide an interrupt vector. In order not to lose a signal, the signal
is injected via task_work during event_sched_out().

Instead going via irq_work, the signal could be added directly via
task_work. The signal is sent to current and can be enqueued on its
return path to userland instead of triggering irq_work. A dummy IRQ is
required in the NMI case to ensure the task_work is handled before
returning to user land. For this irq_work is used. An alternative would
be just raising an interrupt like arch_send_call_function_single_ipi().

During testing with `remove_on_exec' it become visible that the event
can be enqueued via NMI during execve(). The task_work must not be kept
because free_event() will complain later. Also the new task will not
have a sighandler installed.

Queue signal via task_work. Remove perf_event::pending_sigtrap and
and use perf_event::pending_work instead. Raise irq_work in the NMI case
for a dummy interrupt. Remove the task_work if the event is freed.

Tested-by: Marco Elver <elver@google.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Reported-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Link: https://lore.kernel.org/r/20240312180814.3373778-3-bigeasy@linutronix.de
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
---
 include/linux/perf_event.h |    3 --
 kernel/events/core.c       |   57 +++++++++++++++++++++++++--------------------
 2 files changed, 33 insertions(+), 27 deletions(-)

@ include/linux/perf_event.h:784 @ struct perf_event {
 	unsigned int			pending_wakeup;
 	unsigned int			pending_kill;
 	unsigned int			pending_disable;
-	unsigned int			pending_sigtrap;
 	unsigned long			pending_addr;	/* SIGTRAP */
 	struct irq_work			pending_irq;
 	struct callback_head		pending_task;
@ include/linux/perf_event.h:961 @ struct perf_event_context {
 	struct rcu_head			rcu_head;
 
 	/*
-	 * Sum (event->pending_sigtrap + event->pending_work)
+	 * Sum (event->pending_work + event->pending_work)
 	 *
 	 * The SIGTRAP is targeted at ctx->task, as such it won't do changing
 	 * that until the signal is delivered.
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@ include/linux/perf_event.h:2286 @ event_sched_out(struct perf_event *event
 		state = PERF_EVENT_STATE_OFF;
 	}
 
-	if (event->pending_sigtrap) {
-		bool dec = true;
-
-		event->pending_sigtrap = 0;
-		if (state != PERF_EVENT_STATE_OFF &&
-		    !event->pending_work) {
-			event->pending_work = 1;
-			dec = false;
-			WARN_ON_ONCE(!atomic_long_inc_not_zero(&event->refcount));
-			task_work_add(current, &event->pending_task, TWA_RESUME);
-		}
-		if (dec)
-			local_dec(&event->ctx->nr_pending);
-	}
-
 	perf_event_set_state(event, state);
 
 	if (!is_software_event(event))
@ include/linux/perf_event.h:6729 @ static void __perf_pending_irq(struct pe
 	 * Yay, we hit home and are in the context of the event.
 	 */
 	if (cpu == smp_processor_id()) {
-		if (event->pending_sigtrap) {
-			event->pending_sigtrap = 0;
-			perf_sigtrap(event);
-			local_dec(&event->ctx->nr_pending);
-		}
 		if (event->pending_disable) {
 			event->pending_disable = 0;
 			perf_event_disable_local(event);
@ include/linux/perf_event.h:9571 @ static int __perf_event_overflow(struct
 
 		if (regs)
 			pending_id = hash32_ptr((void *)instruction_pointer(regs)) ?: 1;
-		if (!event->pending_sigtrap) {
-			event->pending_sigtrap = pending_id;
+		if (!event->pending_work) {
+			event->pending_work = pending_id;
 			local_inc(&event->ctx->nr_pending);
-			irq_work_queue(&event->pending_irq);
+			WARN_ON_ONCE(!atomic_long_inc_not_zero(&event->refcount));
+			task_work_add(current, &event->pending_task, TWA_RESUME);
+			/*
+			 * The NMI path returns directly to userland. The
+			 * irq_work is raised as a dummy interrupt to ensure
+			 * regular return path to user is taken and task_work
+			 * is processed.
+			 */
+			if (in_nmi())
+				irq_work_queue(&event->pending_irq);
 		} else if (event->attr.exclude_kernel && valid_sample) {
 			/*
 			 * Should not be able to return to user space without
-			 * consuming pending_sigtrap; with exceptions:
+			 * consuming pending_work; with exceptions:
 			 *
 			 *  1. Where !exclude_kernel, events can overflow again
 			 *     in the kernel without returning to user space.
@ include/linux/perf_event.h:9597 @ static int __perf_event_overflow(struct
 			 *     To approximate progress (with false negatives),
 			 *     check 32-bit hash of the current IP.
 			 */
-			WARN_ON_ONCE(event->pending_sigtrap != pending_id);
+			WARN_ON_ONCE(event->pending_work != pending_id);
 		}
 
 		event->pending_addr = 0;
@ include/linux/perf_event.h:13037 @ static void sync_child_event(struct perf
 		     &parent_event->child_total_time_running);
 }
 
+static bool task_work_cb_match(struct callback_head *cb, void *data)
+{
+	struct perf_event *event = container_of(cb, struct perf_event, pending_task);
+
+	return event == data;
+}
+
 static void
 perf_event_exit_event(struct perf_event *event, struct perf_event_context *ctx)
 {
@ include/linux/perf_event.h:13083 @ perf_event_exit_event(struct perf_event
 		 * Kick perf_poll() for is_event_hup();
 		 */
 		perf_event_wakeup(parent_event);
+		/*
+		 * Cancel pending task_work and update counters if it has not
+		 * yet been delivered to userland. free_event() expects the
+		 * reference counter at 1 and keeping the event around until the
+		 * task return to userland will be a unexpected.
+		 */
+		if (event->pending_work &&
+		    task_work_cancel_match(current, task_work_cb_match, event)) {
+			put_event(event);
+			local_dec(&event->ctx->nr_pending);
+		}
 		free_event(event);
 		put_event(parent_event);
 		return;