Skip to contentSkip to conversation

uprobes: clear extra_consumers before pooling return instances

PATCH In review Keenan Dong 3 participants 3 messages Apr 8 – Nov 20

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

PATCH 1/1 uprobes: clear extra_consumers before pooling return instances

ri_pool_push() returns a return_instance to the per-task pool for later reuse. The pool reset clears cons_cnt, but it leaves extra_consumers behind.

A reused return_instance can later grow a fresh extra_consumers array and then reach the cleanup path with a stale pointer from its previous lifetime, leading to a double free of the recycled object.

Free and clear extra_consumers before putting the instance back into the pool so every reused entry starts from a clean state.

Fixes 8622e45b5da1 ("uprobes: Reuse return_instances between multiple uretprobes within task")
---
 kernel/events/uprobes.c | 2 ++
 1 file changed, 2 insertions(+)
kernel/events/uprobes.c +2 -0
index 923b24b321cc..24b9884a2667 100644
@@ -1945,6 +1945,8 @@ unsigned long uprobe_get_trap_addr(struct pt_regs *regs)
 
 static void ri_pool_push(struct uprobe_task *utask, struct return_instance *ri)
 {
+	kfree(ri->extra_consumers);
+	ri->extra_consumers = NULL;
 	ri->cons_cnt = 0;
 	ri->next = utask->ri_pool;
 	utask->ri_pool = ri;
Signature · 1 line
2.43.0
#1 lore raw
Sashiko AI review found 1 potential issue(s): - [Critical] Missing full memory barrier in `ri_timer` allows Store-Load reordering, causing a race condition with `free_ret_instance` that can lead to…

Sashiko AI review found 1 potential issue(s):
- [Critical] Missing full memory barrier in ri_timer allows Store-Load reordering, causing a race condition with free_ret_instance that can lead to Use-After-Free or uprobe corruption. --

commit 892d89a5acfb2bb4da3888666af1d58594ed9e7d
Author: Keenan Dong <[email protected]>

uprobes: clear extra_consumers before pooling return instances

This commit clears the extra_consumers array and frees its memory before pushing a return_instance back to the per-task pool. This prevents a potential double free if a recycled object retains a stale pointer from its previous lifetime.

Quoted 4 lines
kernel/events/uprobes.c +0 -0
index 4084e926e2844..4695cb812fe26 100644

[ ... ]

kernel/events/uprobes.c unsigned long uprobe_get_trap_addr(struct pt_regs *regs) · line 1952

Quoted 9 lines
kernel/events/uprobes.c +2 -0
@@ -1945,6 +1945,8 @@ unsigned long uprobe_get_trap_addr(struct pt_regs *regs)
 
 static void ri_pool_push(struct uprobe_task *utask, struct return_instance *ri)
 {
+	kfree(ri->extra_consumers);
+	ri->extra_consumers = NULL;
 	ri->cons_cnt = 0;
 	ri->next = utask->ri_pool;
 	utask->ri_pool = ri;

While looking at how return instances are recycled into the pool here, I noticed a potential race condition in the pre-existing code involving ri_timer() and free_ret_instance().

In uprobes.c, ri_seqcount is used in a two-way synchronization pattern between free_ret_instance() and ri_timer().

In free_ret_instance(), the task unlinks ri, executes hprobe_consume() (which contains a full memory barrier via xchg()), and then reads ri_seqcount via raw_seqcount_try_begin().

In ri_timer(), the timer asserts its presence by incrementing the seqcount via raw_write_seqcount_begin(), and then iterates over utask->return_instances.

For this mutual exclusion to work, both sides require a full memory barrier. However, raw_write_seqcount_begin() only provides a write barrier (smp_wmb()), which does not prevent Store-Load reordering.

Could this allow the timer CPU to execute the load of utask->return_instances before the store to ri_seqcount.sequence becomes globally visible?

If this reordering occurs, the following sequence seems possible:

1. ri_timer() reads a pointer to ri and stalls.
2. The task unlinks ri, executes its full barrier, reads the even seqcount, and pushes ri to ri_pool.
3. The task then quickly reuses ri for a new uretprobe, setting hprobe->state = HPROBE_LEASED.
4. ri_timer() resumes and executes try_cmpxchg() on hprobe->state.

Because the state is HPROBE_LEASED again, the cmpxchg succeeds. It transitions the new uprobe's state to STABLE, dropping the SRCU lock. When the task later consumes the new uprobe, it will call put_uprobe() without a matching get_uprobe().

Can this sequence lead to a premature refcount drop and use-after-free of the new uprobe?

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
Tested-by on v1 1/1
#2 lore raw
In reply to Keenan Dong

Hi Keenan, all,

I ran into this crash on my own and can confirm your patch fixes it, so here is a Tested-by.

I was seeing it on an unpatched 6.18.63 (kernel.org source, built with a Debian 13 config). With libfakechroot in /etc/ld.so.preload and a couple of "setarch -R ls" loops running, I attach entry and return uprobes with bpftrace to libfakechroot and to the same functions in libc, detaching and reattaching every few seconds. The kernel oopses in handler_chain() under krealloc() within a minute or so, which is the extra_consumers path your patch is about. I haven't pinned down exactly which probes end up sharing an address in this setup.

With only your patch on top of 6.18.63, the same probes ran 20 attach/detach cycles with no oops, and a fork-heavy variant ran 40 more, also clean. 6.12.117 doesn't crash, which fits the Fixes: tag.

One more thing: Keke Ming's v2 for dup_return_instance() has a hunk that clears extra_consumers when cons_cnt <= 1. With only your patch applied the fork-heavy runs were clean, so I haven't seen a need for that hunk for this crash.

I can send the reproducer and a full splat if that would help.

Thanks,
Reza