CVE-2026-98862

In the Linux kernel, the following vulnerability has been resolved: dm-pcache: reject a kset that overruns its segment. cache_replay(), the writeback worker and the GC worker read a kset of get_kset_onmedia_size() bytes and advance the position by it. A forged key_num makes that size exceed the segment's remaining space, so the advance walks past the segment and trips the cache_pos_advance() BUG_ON.

In the Linux kernel, the following vulnerability has been resolved: dm-pcache: bound the logical key offset from persistent memory. cache_key_decode() takes a key's logical off from the cache device and later indexes req_key_tree->subtrees[] by it in get_subtree(). An off past the device forms a subtree pointer outside the array, which rb_insert() writes through during replay.

In the Linux kernel, the following vulnerability has been resolved: sched/cache: Introduce task_struct->sched_cache_grp to fix UAF. Add a sched_cache_grp pointer to task_struct so that scheduler code can access the cache group directly via the task, without going through mm->sched_cache_grp. This decouples the scheduler's hot-path accesses from the mm_struct. Each task holds its own refcount on the sched_cache_group, separate from the reference held by its mm_struct. The reference is acquired in copy_mm() (fork) and exec_mmap() (exec), and released in exit_mm(). This fixes the use-after-free when account_mm_sched() reaches the group through a task whose mm is being switched.

In the Linux kernel's sched_ext scheduler, ops.dequeue() was invoked while the source user DSQ lock was still held on the consume and move paths. A BPF scheduler that locks the source DSQ from ops.dequeue() would self-deadlock. The fix moves the invocation after the DSQ unlock on all three paths.

In the Linux kernel's sched_ext, a task could be found on a DSQ while still in SCX_OPSS_DISPATCHING state. The reenqueue paths did not wait for the state to clear, so ops.enqueue() set SCX_OPSS_QUEUED and the dispatcher's final store overwrote it with SCX_OPSS_NONE, dropping later dispatches of the task. The fix waits for SCX_OPSS_DISPATCHING to clear before dequeueing for reenqueue.

In KVM on x86, GET_NESTED_STATE_PAGES was not re-pended when fetching nested pages failed in the KVM_RUN path. If userspace re-ran the vCPU and vmcs02 held valid PFNs from the previous L2 run, KVM could re-enter L2 with stale, unpinned PFNs mapped into e.g. the vAPIC page.

A vulnerability was fixed in the Linux kernel's fprobe mechanism where unregister_fprobe() and unregister_fprobe_async() relied on standard RCU grace periods that do not track preemption-disabled sections when RCU is not watching (e.g., in the idle loop or nohz_full extended quiescent states). This could lead to a use-after-free if the fprobe was freed immediately after unregistration. The fix ensures handlers exit early when !rcu_is_watching() and replaces guard(rcu)() and rcu_read_lock() with guard(rcu_sched_notrace)(), eliminating fast-path lockdep overhead.

In the Linux kernel HID BPF subsystem, the comparison of the return value from hid_bpf_hw_request against size did not cast size to ssize_t. When a transport driver returned a negative error code (e.g. -EPIPE), it was promoted to size_t and evaluated as a large positive number, making (ret > size) true. This silently converted the hardware error into a success and copied the unmodified buffer back.

In the Linux kernel, the HID multitouch driver's mt_remove() used timer_delete_sync() to stop the sticky-finger release timer. However, the timer could be rearmed by a report arriving after deletion, leading to a use-after-free when the timer fired after the driver was unbound. The fix replaces timer_delete_sync() with timer_shutdown_sync(), which also prevents any subsequent mod_timer() from rearming the timer.

In the Linux kernel io_uring with SQPOLL, the sqpoll thread could pop and complete a request as soon as mpscq_push() published it. If that was the last request holding the ring alive, io_ring_exit_work() freed the ctx, stopped the thread and tctx while io_req_normal_work_add() was still referencing them, causing a use-after-free. The fix holds an RCU read lock across the add and waits for a grace period before freeing the ctx.