Skip to contentSkip to conversation

Fixing i915 PMU use after free after driver unbind

RFC New Tvrtko Ursulin 1 participant 4 messages Jan 15, 2024

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

cover Fixing i915 PMU use after free after driver unbind

From: Tvrtko Ursulin <[email protected]>

Hi Peter, all,

This is an early RFC to outline a newly discovered problem in the current handling of driver unbind with active perf fds.

The sequence is basically this:

1. Open a perf fd
2. Unbind a driver
3. Close the dangling fd

Or a slightly more evil variant:

1. Open a perf fd
2. Unbind a driver
3. Bind the driver again
4. Close the dangling fd

I thought we had this covered by recording the unbound status (pmu->closed in i915_pmu.c) and making sure the struct i915_pmu (and struct perf_pmu) remain active until the last event is closed (via internal reference counting). But what I missed until now are two things:

1)
core.c: _free_event() will dereference event->pmu _after_ event->destroy().

KASAN catches this easily and patches 1 & 2 are the attempt to fix that.

2)
A more evil case where pmu->cpu_pmu_context per-cpu allocation gets re-used _before_ the old perf fd is closed.

There things can nicely explode on the list_del_init inside event_sched_out on list_del_init(&event->active_list); (with list debugging turned on of course).

Most easily reproducible by simply re-binding i915, which happens to grab the same per-cpu block and then the new perf_pmu_register zaps the list_head which the old event will try to unlink itself from.

This is what the third patch attempts to deal with. It is a bit incomplete though, as I was unsure what is the best approach to fix and so thought to send it out early for some guidance.

Cc Peter Zijlstra <[email protected]>
Cc Umesh Nerlige Ramappa <[email protected]>
Cc Aravind Iddamsetty <[email protected]>

Tvrtko Ursulin (3):
perf: Add new late event free callback
drm/i915/pmu: Move i915 reference drop to new event->free()
perf: Reference count struct perf_cpu_pmu_context to fix driver unbind

 drivers/gpu/drm/i915/i915_pmu.c |  4 ++--
 include/linux/perf_event.h      |  2 ++
 kernel/events/core.c            | 34 ++++++++++++++++++++++++---------
 3 files changed, 29 insertions(+), 11 deletions(-)
Signature · 1 line
2.40.1
#1 lore raw

RFC 1/3 perf: Add new late event free callback

From: Tvrtko Ursulin <[email protected]>

This allows drivers to implement a PMU with support for unbinding the device, for example by making event->pmu reference counted on the driver side and its lifetime matching the struct perf_event init/free.

Otherwise, if an open perf fd is kept past driver unbind, the perf code can dereference the potentially freed struct pmu from the _free_event steps which follow the existing destroy callback.

TODO/FIXME/QQQ:

A simpler version could be to simply move the ->destroy() callback to later in _free_event(). However a comment there claims there are steps which need to run after the existing destroy callbacks, hence I opted for an initially cautious approach.

Cc Peter Zijlstra <[email protected]>
Cc Umesh Nerlige Ramappa <[email protected]>
Cc Aravind Iddamsetty <[email protected]>
---
 include/linux/perf_event.h |  1 +
 kernel/events/core.c       | 13 +++++++++++--
 2 files changed, 12 insertions(+), 2 deletions(-)
include/linux/perf_event.h +1 -0
index 5547ba68e6e4..a567d2d98be1 100644
@@ -799,6 +799,7 @@ struct perf_event {
 	struct perf_event		*aux_event;
 
 	void (*destroy)(struct perf_event *);
+	void (*free)(struct perf_event *);
 	struct rcu_head			rcu_head;
 
 	struct pid_namespace		*ns;
kernel/events/core.c +11 -2
index a64165af45c1..4b62d2201ca7 100644
@@ -5242,6 +5242,9 @@ static void _free_event(struct perf_event *event)
 	exclusive_event_destroy(event);
 	module_put(event->pmu->module);
 
+	if (event->free)
+		event->free(event);
+
 	call_rcu(&event->rcu_head, free_event_rcu);
 }
 
@@ -11662,8 +11665,12 @@ static int perf_try_init_event(struct pmu *pmu, struct perf_event *event)
 		    event_has_any_exclude_flag(event))
 			ret = -EINVAL;
 
-		if (ret && event->destroy)
-			event->destroy(event);
+		if (ret) {
+			if (event->destroy)
+				event->destroy(event);
+			if (event->free)
+				event->free(event);
+		}
 	}
 
 	if (ret)
@@ -12090,6 +12097,8 @@ perf_event_alloc(struct perf_event_attr *attr, int cpu,
 		perf_detach_cgroup(event);
 	if (event->destroy)
 		event->destroy(event);
+	if (event->free)
+		event->free(event);
 	module_put(pmu->module);
 err_ns:
 	if (event->hw.target)
Signature · 1 line
2.40.1
#2 lore raw

RFC 2/3 drm/i915/pmu: Move i915 reference drop to new event->free()

In reply to Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

Avoids use after free in the perf core code on the event destruction path, after the PCI driver has been unbound with the active perf file descriptors.

Cc Peter Zijlstra <[email protected]>
Cc Umesh Nerlige Ramappa <[email protected]>
Cc Aravind Iddamsetty <[email protected]>
---
 drivers/gpu/drm/i915/i915_pmu.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_pmu.c +2 -2
index 21eb0c5b320d..010763a5bc39 100644
@@ -514,7 +514,7 @@ static enum hrtimer_restart i915_sample(struct hrtimer *hrtimer)
 	return HRTIMER_RESTART;
 }
 
-static void i915_pmu_event_destroy(struct perf_event *event)
+static void i915_pmu_event_free(struct perf_event *event)
 {
 	struct i915_pmu *pmu = event_to_pmu(event);
 	struct drm_i915_private *i915 = pmu_to_i915(pmu);
@@ -630,7 +630,7 @@ static int i915_pmu_event_init(struct perf_event *event)
 
 	if (!event->parent) {
 		drm_dev_get(&i915->drm);
-		event->destroy = i915_pmu_event_destroy;
+		event->free = i915_pmu_event_free;
 	}
 
 	return 0;
Signature · 1 line
2.40.1
#3 lore raw

RFC 3/3 perf: Reference count struct perf_cpu_pmu_context to fix driver unbind

In reply to Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

If a PMU driver is a PCI driver which can be unbound at runtime there is currently an use after free issue with CPU contexts when an active perf fd is kept around.

Specifically, when perf_pmu_unregister() calls free_pmu_context() on the PMU provided per cpu struct perf_cpu_pmu_context storage, any call path which ends up in event_sched_out() (such as __perf_remove_from_context() via perf_event_release_kernel()) will dereference a freed event->pmu_ctx. Furthermore if the same percpu area has in the meantime been re-allocated, the use after free will corrupt someone elses per cpu storage area.

To fix it we attempt to add reference counting to struct perf_cpu_pmu_context such that the object remains until the last user is done with it.

TODO/FIXME/QQQ:

1)

I am really not sure about the locking here. I *think* I needed a per struct pmu counter and by looking at what find_get_pmu_context does when it takes a slot from driver provided pmu->cpu_pmu_context under the ctx->lock, it looked like that should be sufficient. Maybe even if not atomic_t. Or maybe ctx->lock is not enough.

2)

I believe pmu->pmu_disable_count will need a similar treatment, but as I wasn't sure of the locking model, or even if this all makes sense on the high level I left it out for now. Like does the idea to reference count even flies or a completely different solution will be needed.

Cc Peter Zijlstra <[email protected]>
Cc Umesh Nerlige Ramappa <[email protected]>
Cc Aravind Iddamsetty <[email protected]>
---
 include/linux/perf_event.h |  1 +
 kernel/events/core.c       | 21 ++++++++++++++-------
 2 files changed, 15 insertions(+), 7 deletions(-)
include/linux/perf_event.h +1 -0
index a567d2d98be1..bd1c8f3c1736 100644
@@ -317,6 +317,7 @@ struct pmu {
 
 	int __percpu			*pmu_disable_count;
 	struct perf_cpu_pmu_context __percpu *cpu_pmu_context;
+	atomic_t 			cpu_pmu_context_refcount;
 	atomic_t			exclusive_cnt; /* < 0: cpu; > 0: tsk */
 	int				task_ctx_nr;
 	int				hrtimer_interval_ms;
kernel/events/core.c +14 -7
index 4b62d2201ca7..0c95aecf560a 100644
@@ -4873,6 +4873,9 @@ find_get_pmu_context(struct pmu *pmu, struct perf_event_context *ctx,
 			atomic_inc(&epc->refcount);
 		}
 		raw_spin_unlock_irq(&ctx->lock);
+
+		atomic_inc(&pmu->cpu_pmu_context_refcount);
+
 		return epc;
 	}
 
@@ -4928,6 +4931,12 @@ find_get_pmu_context(struct pmu *pmu, struct perf_event_context *ctx,
 	return epc;
 }
 
+static void put_pmu_context(struct pmu *pmu)
+{
+	if (atomic_dec_and_test(&pmu->cpu_pmu_context_refcount))
+		free_percpu(pmu->cpu_pmu_context);
+}
+
 static void get_pmu_ctx(struct perf_event_pmu_context *epc)
 {
 	WARN_ON_ONCE(!atomic_inc_not_zero(&epc->refcount));
@@ -4967,8 +4976,10 @@ static void put_pmu_ctx(struct perf_event_pmu_context *epc)
 
 	raw_spin_unlock_irqrestore(&ctx->lock, flags);
 
-	if (epc->embedded)
+	if (epc->embedded) {
+		put_pmu_context(epc->pmu);
 		return;
+	}
 
 	call_rcu(&epc->rcu_head, free_epc_rcu);
 }
@@ -11347,11 +11358,6 @@ static int perf_event_idx_default(struct perf_event *event)
 	return 0;
 }
 
-static void free_pmu_context(struct pmu *pmu)
-{
-	free_percpu(pmu->cpu_pmu_context);
-}
-
 /*
  * Let userspace know that this PMU supports address range filtering:
  */
@@ -11573,6 +11579,7 @@ int perf_pmu_register(struct pmu *pmu, const char *name, int type)
 		pmu->event_idx = perf_event_idx_default;
 
 	list_add_rcu(&pmu->entry, &pmus);
+	atomic_set(&pmu->cpu_pmu_context_refcount, 1);
 	atomic_set(&pmu->exclusive_cnt, 0);
 	ret = 0;
 unlock:
@@ -11615,7 +11622,7 @@ void perf_pmu_unregister(struct pmu *pmu)
 		device_del(pmu->dev);
 		put_device(pmu->dev);
 	}
-	free_pmu_context(pmu);
+	put_pmu_context(pmu);
 	mutex_unlock(&pmus_lock);
 }
 EXPORT_SYMBOL_GPL(perf_pmu_unregister);
Signature · 1 line
2.40.1