perf: Disallow mis-matched inherited group reads

jira VULN-4127
cve CVE-2023-5717
commit-author Peter Zijlstra <[email protected]>
commit 32671e3
upstream-diff The mainline fix 32671e3
  adds a new `group_generation' field to the `perf_event' struct. This
  breaks LTS 8.6 kABI. The new field was preserved, but moved to the end
  of the struct and wrapped in the `RH_KABI_EXTEND' macro. The kABI in
  this particular case is preserved, as the `perf_event' struct is always
  dynamically allocated through `perf_event_alloc()' and it's never used
  as an array.

Because group consistency is non-atomic between parent (filedesc) and children
(inherited) events, it is possible for PERF_FORMAT_GROUP read() to try and sum
non-matching counter groups -- with non-sensical results.

Add group_generation to distinguish the case where a parent group removes and
adds an event and thus has the same number, but a different configuration of
events as inherited groups.

This became a problem when commit fa8c269 ("perf/core: Invert
perf_read_group() loops") flipped the order of child_list and sibling_list.
Previously it would iterate the group (sibling_list) first, and for each
sibling traverse the child_list. In this order, only the group composition of
the parent is relevant. By flipping the order the group composition of the
child (inherited) events becomes an issue and the mis-match in group
composition becomes evident.

That said; even prior to this commit, while reading of a group that is not
equally inherited was not broken, it still made no sense.

(Ab)use ECHILD as error return to indicate issues with child process group
composition.

Fixes: fa8c269 ("perf/core: Invert perf_read_group() loops")
	Reported-by: Budimir Markovic <[email protected]>
	Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]
(cherry picked from commit 32671e3)
	Signed-off-by: Marcin Wcisło <[email protected]>

Author: pvts-mat

include/linux/perf_event.h

@@ -57,6 +57,7 @@ struct perf_guest_info_callbacks {
 #include <linux/cgroup.h>
 #include <linux/refcount.h>
 #include <linux/security.h>
+#include <linux/rh_kabi.h>
 #include <asm/local.h>
 
 struct perf_callchain_entry {
@@ -800,6 +801,8 @@ struct perf_event {
 	RH_KABI_BROKEN_INSERT(void *security)
 #endif
 	struct list_head		sb_list;
+
+	RH_KABI_EXTEND(unsigned int	group_generation)
 #endif /* CONFIG_PERF_EVENTS */
 };
 

kernel/events/core.c

@@ -2020,6 +2020,7 @@ static void perf_group_attach(struct perf_event *event)
 
 	list_add_tail(&event->sibling_list, &group_leader->sibling_list);
 	group_leader->nr_siblings++;
+	group_leader->group_generation++;
 
 	perf_event__header_size(group_leader);
 
@@ -2212,6 +2213,7 @@ static void perf_group_detach(struct perf_event *event)
 	if (leader != event) {
 		list_del_init(&event->sibling_list);
 		event->group_leader->nr_siblings--;
+		event->group_leader->group_generation++;
 		goto out;
 	}
 
@@ -5278,7 +5280,7 @@ static int __perf_read_group_add(struct perf_event *leader,
 					u64 read_format, u64 *values)
 {
 	struct perf_event_context *ctx = leader->ctx;
-	struct perf_event *sub;
+	struct perf_event *sub, *parent;
 	unsigned long flags;
 	int n = 1; /* skip @nr */
 	int ret;
@@ -5288,6 +5290,33 @@ static int __perf_read_group_add(struct perf_event *leader,
 		return ret;
 
 	raw_spin_lock_irqsave(&ctx->lock, flags);
+	/*
+	 * Verify the grouping between the parent and child (inherited)
+	 * events is still in tact.
+	 *
+	 * Specifically:
+	 *  - leader->ctx->lock pins leader->sibling_list
+	 *  - parent->child_mutex pins parent->child_list
+	 *  - parent->ctx->mutex pins parent->sibling_list
+	 *
+	 * Because parent->ctx != leader->ctx (and child_list nests inside
+	 * ctx->mutex), group destruction is not atomic between children, also
+	 * see perf_event_release_kernel(). Additionally, parent can grow the
+	 * group.
+	 *
+	 * Therefore it is possible to have parent and child groups in a
+	 * different configuration and summing over such a beast makes no sense
+	 * what so ever.
+	 *
+	 * Reject this.
+	 */
+	parent = leader->parent;
+	if (parent &&
+	    (parent->group_generation != leader->group_generation ||
+	     parent->nr_siblings != leader->nr_siblings)) {
+		ret = -ECHILD;
+		goto unlock;
+	}
 
 	/*
 	 * Since we co-schedule groups, {enabled,running} times of siblings
@@ -5317,8 +5346,9 @@ static int __perf_read_group_add(struct perf_event *leader,
 			values[n++] = primary_event_id(sub);
 	}
 
+unlock:
 	raw_spin_unlock_irqrestore(&ctx->lock, flags);
-	return 0;
+	return ret;
 }
 
 static int perf_read_group(struct perf_event *event,
@@ -5337,10 +5367,6 @@ static int perf_read_group(struct perf_event *event,
 
 	values[0] = 1 + leader->nr_siblings;
 
-	/*
-	 * By locking the child_mutex of the leader we effectively
-	 * lock the child list of all siblings.. XXX explain how.
-	 */
 	mutex_lock(&leader->child_mutex);
 
 	ret = __perf_read_group_add(leader, read_format, values);
@@ -12994,6 +13020,7 @@ static int inherit_group(struct perf_event *parent_event,
 		    !perf_get_aux_event(child_ctr, leader))
 			return -EINVAL;
 	}
+	leader->group_generation = parent_event->group_generation;
 	return 0;
 }