CVE-2026-46223

Published May 28, 2026

Last updated 2 months ago

CVSS medium 5.5

Overview

Description
In the Linux kernel, the following vulnerability has been resolved: cgroup: Defer css percpu_ref kill on rmdir until cgroup is depopulated A chain of commits going back to v7.0 reworked rmdir to satisfy the controller invariant that a subsystem's ->css_offline() must not run while tasks are still doing kernel-side work in the cgroup. [1] d245698d727a ("cgroup: Defer task cgroup unlink until after the task is done switching out") [2] a72f73c4dd9b ("cgroup: Don't expose dead tasks in cgroup") [3] 1b164b876c36 ("cgroup: Wait for dying tasks to leave on rmdir") [4] 4c56a8ac6869 ("cgroup: Fix cgroup_drain_dying() testing the wrong condition") [5] 13e786b64bd3 ("cgroup: Increment nr_dying_subsys_* from rmdir context") [1] moved task cset unlink from do_exit() to finish_task_switch() so a task's cset link drops only after the task has fully stopped scheduling. That made tasks past exit_signals() linger on cset->tasks until their final context switch, which led to a series of problems as what userspace expected to see after rmdir diverged from what the kernel needs to wait for. [2]-[5] tried to bridge that divergence: [2] filtered the exiting tasks from cgroup.procs; [3] had rmdir(2) sleep in TASK_UNINTERRUPTIBLE for them; [4] fixed the wait's condition; [5] made nr_dying_subsys_* visible synchronously. The cgroup_drain_dying() wait in [3] turned out to be a dead end. When the rmdir caller is also the reaper of a zombie that pins a pidns teardown (e.g. host PID 1 systemd reaping orphan pids that were re-parented to it during the same teardown), rmdir blocks in TASK_UNINTERRUPTIBLE waiting for those pids to free, the pids can't free because PID 1 is the reaper and it's stuck in rmdir, and the system A-A deadlocks. No internal lock ordering breaks this; the wait itself is the bug. The css killing side that drove the original reorder, however, can be made cleanly asynchronous: ->css_offline() is already async, run from css_killed_work_fn() driven by percpu_ref_kill_and_confirm(). The fix is to make that chain start only after all tasks have left the cgroup. rmdir's user-visible side then returns as soon as cgroup.procs and friends are empty, while ->css_offline() still runs only after the cgroup is fully drained. Verified by the original reproducer (pidns teardown + zombie reaper, runs under vng) which hangs vanilla and succeeds here, and by per-commit deterministic repros for [2], [3], [4], [5] with a boot parameter that widens the post-exit_signals() window so each state is reliably reachable. Some stress tests on top of that. cgroup_apply_control_disable() has the same shape of pre-existing race: when a controller is disabled via subtree_control, kill_css() ran synchronously while tasks past exit_signals() could still be linked to the cgroup's csets, and ->css_offline() could fire before they drained. This patch preserves the existing synchronous behavior at that call site (kill_css_sync() + kill_css_finish() back-to-back) and a follow-up patch will defer kill_css_finish() there using a per-css trigger. This seems like the right approach and I don't see problems with it. The changes are somewhat invasive but not excessively so, so backporting to -stable should be okay. If something does turn out to be wrong, the fallback is to revert the entire chain ([1]-[5]) and rework in the development branch instead. v2: Pin cgrp across the deferred destroy work with explicit cgroup_get()/cgroup_put() around queue_work() and the work_fn. v1 wasn't actually broken (ordered cgroup_offline_wq + queue_work order in cgroup_task_dead() saved it) but the explicit ref removes the dependency on those non-obvious invariants. Also note the pre-existing cgroup_apply_control_disable() race in the description; a follow-up will defer kill_css_finish() there.
Source
416baaa9-dc9f-4396-8d5f-8c081fb06d67
NVD status
Analyzed
Products
linux_kernel

Risk scores

CVSS 3.1

Type
Primary
Base score
5.5
Impact score
3.6
Exploitability score
1.8
Vector string
CVSS:3.1/AV:L/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H
Severity
MEDIUM

Weaknesses

nvd@nist.gov
CWE-667

Social media

Hype score
Not currently trending

Configurations

Related CVEs

  1. In the Linux kernel, the following vulnerability has been resolved: KVM: x86/mmu: Ensure hugepage is in by slot before checking max mapping level When recovering hugepages in the shadow MMU, verify that the base gfn of the shadow page is actually contained within the target memslot, *before* querying the max mapping level given the shadow page's gfn. Failure to pre-check the validity of the gfn can lead to an out-of-bounds access to the slot's lpage_info (which typically manifests as a host #PF because the lpage_info is vmalloc'd) if the guest creates a hugepage mapping (in its PTEs) that extends "below" the bounds of a memslot. When faulting in memory for a guest, and the size of the guest mapping is greater than KVM's (current) max mapping, then KVM will create a "direct" shadow page (direct in that there are no gPTEs to shadow, and so the target gfn is a direct calculation given the base gfn of the shadow page). The hugepage recovery flow looks for such direct shadow pages, as forcing 4KiB mappings when dirty logging generates the guest > host mapping size case. When the 4KiB restriction is lifted, then KVM can replace the shadow page with a hugepage. But if KVM originally used a smaller mapping than the guest because the range of memory covered by the guest hugepage exceeds the bounds of a memslot, then KVM will link a direct shadow page with a gfn that is outside the bounds of the memslot being used to fault in memory. The rmap entry added for the leaf mapping is correct and within bounds, but the gfn of the leaf SPTE's parent shadow page will be out of bounds. BUG: unable to handle page fault for address: ffffc90000806ffc #PF: supervisor read access in kernel mode #PF: error_code(0x0000) - not-present page PGD 100000067 P4D 100000067 PUD 1002a7067 PMD 10612f067 PTE 0 Oops: Oops: 0000 [#1] SMP CPU: 13 UID: 1000 PID: 757 Comm: mmu_stress_test Not tainted 7.1.0-rc1-48ce1e26eace-x86_pir_to_irr_comments-vm #341 PREEMPT Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015 RIP: 0010:kvm_mmu_max_mapping_level+0x79/0x2b0 [kvm] Call Trace: <TASK> kvm_mmu_recover_huge_pages+0x21b/0x320 [kvm] kvm_set_memslot+0x1ee/0x590 [kvm] kvm_set_memory_region.part.0+0x3a1/0x4d0 [kvm] kvm_vm_ioctl+0x9bf/0x15d0 [kvm] __x64_sys_ioctl+0x8a/0xd0 do_syscall_64+0xb7/0xbb0 entry_SYSCALL_64_after_hwframe+0x4b/0x53 RIP: 0033:0x7f21c0f1a9bf </TASK> Don't bother pre-checking the bounds of the potential hugepage, i.e. don't check that e.g. sp->gfn + KVM_PAGES_PER_HPAGE(sp->role.level + 1) is also within the memslot, as the checks performed by kvm_mmu_max_mapping_level() are a superset of the basic bounds checks. I.e. pre-checking the full range would be a dubious micro-optimization.•CVE-2026-63807
  2. In the Linux kernel, the following vulnerability has been resolved: KVM: Replace guest-triggerable BUG_ON() in ioeventfd datamatch with get_unaligned() Drop a BUG_ON() that has been reachable since it was first added, way back in 2009, and instead use get_unaligned() to perform potentially-unaligned accesses. For a given store, KVM x86's emulator tracks the entire value in the destination operand, x86_emulate_ctxt.dst. If the destination is memory, and the target splits multiple pages and/or is emulated MMIO, then KVM handles each fragment independently. E.g. on a page split starting at page offset 0xffc, KVM writes 4 bytes to the first page, then the remaining bytes to the second page, using ctxt->dst as the source for both (with appropriate offsets). If the destination splits a page *and* hits emulated MMIO on the second page, then KVM will complete the write to the first page, then emulate the MMIO access to the second page. If there is a datamatch-enabled ioeventfd at offset 0 of the second page, then KVM will process the remainder of the store as a potential ioeventfd signal. Putting it all together, if the guest emits a store that splits a page starting at page offset N, and the second page has a datamatch-enabled ioeventfd at offset 0, then KVM will check for datamatch using &dst.valptr[N] as the source. Due to dst (and thus dst.valptr) being 32-byte aligned, if N is not aligned to @len, the BUG_ON() fires. E.g. with a 16-byte store at page offset 0xffc, to an ioeventfd of len 8, all initial checks in ioeventfd_in_range() will succeed, and the BUG_ON() fires due to @val being 4-byte aligned, but not 8-byte aligned. ------------[ cut here ]------------ kernel BUG at arch/x86/kvm/../../../virt/kvm/eventfd.c:783! Oops: invalid opcode: 0000 [#1] SMP CPU: 0 UID: 1000 PID: 615 Comm: repro Not tainted 7.1.0-rc2-ff238429d1ea #365 PREEMPT Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015 RIP: 0010:ioeventfd_write+0x6c/0x70 [kvm] Call Trace: <TASK> __kvm_io_bus_write+0x85/0xb0 [kvm] kvm_io_bus_write+0x53/0x80 [kvm] vcpu_mmio_write+0x66/0xf0 [kvm] emulator_read_write_onepage+0x12a/0x540 [kvm] emulator_read_write+0x109/0x2b0 [kvm] x86_emulate_insn+0x4f8/0xfb0 [kvm] x86_emulate_instruction+0x181/0x790 [kvm] kvm_mmu_page_fault+0x313/0x630 [kvm] vmx_handle_exit+0x18a/0x590 [kvm_intel] kvm_arch_vcpu_ioctl_run+0xc81/0x1c90 [kvm] kvm_vcpu_ioctl+0x2d5/0x970 [kvm] __x64_sys_ioctl+0x8a/0xd0 do_syscall_64+0xb7/0x890 entry_SYSCALL_64_after_hwframe+0x4b/0x53 RIP: 0033:0x7f19c931a9bf </TASK> Modules linked in: kvm_intel kvm irqbypass ---[ end trace 0000000000000000 ]--- In a perfect world, the fix would be to simply delete the BUG_ON(), as KVM x86 doesn't perform alignment checks on "normal" memory accesses at CPL0. Sadly, C99 ruins all the fun; while the x86 architecture plays nice, dereferencing an unaligned pointer directly is undefined behavior in C, e.g. triggers splats when running with CONFIG_UBSAN_ALIGNMENT=y.•CVE-2026-63806
  3. In the Linux kernel, the following vulnerability has been resolved: crypto: nx - fix nx_crypto_ctx_exit argument nx_crypto_ctx_shash_exit calls nx_crypto_ctx_exit with crypto_shash_ctx(...) but crypto_shash_ctx gives a nx_crypto_ctx *, not a crypto_tfm *. Fix the type in nx_crypto_ctx_exit and drop the bogus crypto_tfm_ctx call. This fixes the following oops: BUG: Unable to handle kernel data access at 0xc0403effffffffc8 Faulting instruction address: 0xc000000000396cb4 Oops: Kernel access of bad area, sig: 11 [#15] Call Trace: nx_crypto_ctx_shash_exit+0x24/0x60 crypto_shash_exit_tfm+0x28/0x40 crypto_destroy_tfm+0x98/0x140 crypto_exit_ahash_using_shash+0x20/0x40 crypto_destroy_tfm+0x98/0x140 hash_release+0x1c/0x30 alg_sock_destruct+0x38/0x60 __sk_destruct+0x48/0x2b0 af_alg_release+0x58/0xb0 __sock_release+0x68/0x150 sock_close+0x20/0x40 __fput+0x110/0x3a0 sys_close+0x48/0xa0 system_call_exception+0x140/0x2d0 system_call_common+0xf4/0x258 .. which came from hardlink(1) opportunistically using AF_ALG. The same problem exists with nx_crypto_ctx_skcipher_exit getting a context it wasn't expecting, but apparently nobody hit that for years.•CVE-2026-63805
  4. In the Linux kernel, the following vulnerability has been resolved: gfs2: fix use-after-free in gfs2_qd_dealloc gfs2_qd_dealloc(), called as an RCU callback from gfs2_qd_dispose(), accesses the superblock object sdp through qd->qd_sbd after freeing qd. It does so to decrement sd_quota_count and wake up sd_kill_wait. However, by the time the RCU callback runs, gfs2_put_super() may have already freed sdp via free_sbd(). This can happen when gfs2_quota_cleanup() is called during unmount: it disposes of quota objects via call_rcu() and then waits on sd_kill_wait with a 60-second timeout. If the timeout expires, or if gfs2_gl_hash_clear() triggers additional qd_put() calls that schedule more RCU callbacks after the wait completes, gfs2_put_super() will proceed to free the superblock while RCU callbacks referencing it are still pending. Add an rcu_barrier() before free_sbd() in gfs2_put_super() to ensure all pending RCU callbacks (including gfs2_qd_dealloc) have completed before the superblock is freed.•CVE-2026-63804
  5. In the Linux kernel, the following vulnerability has been resolved: hdlc_ppp: sync per-proto timers before freeing hdlc state Each PPP control protocol (LCP/IPCP/IPV6CP) embedded in struct ppp registers a timer via timer_setup(). That struct ppp is the hdlc->state allocation, which detach_hdlc_protocol() frees with kfree() in both teardown paths: unregister_hdlc_device() and the re-attach inside attach_hdlc_protocol(). The ppp proto never registered a .detach callback, so detach_hdlc_protocol() performs no timer synchronization before the kfree(). The only cancel, timer_delete(&proto->timer) in ppp_cp_event(), is partial (it does not wait for a running callback) and only runs on the ->CLOSED transition; ppp_stop()/ppp_close() do not sync either. A ppp_timer callback already executing (blocked on ppp->lock) survives the kfree and then dereferences proto->state / ppp->lock in freed memory, leading to a use-after-free. Fix this by adding a .detach helper that calls timer_shutdown_sync() on every per-proto timer. detach_hdlc_protocol() invokes proto->detach(dev) before kfree(hdlc->state), so timer_shutdown_sync() now runs on both free paths. timer_shutdown_sync() is used instead of timer_delete_sync() because the keepalive path re-arms the timer through add_timer()/mod_timer() and shutdown blocks any re-activation during teardown. Initialize the per-protocol timers in ppp_ioctl() when the protocol is attached, and remove the now-redundant timer_setup() from ppp_start(), so that the timers are initialized exactly once at attach time and ppp_timer_release() never operates on uninitialized timer_list structures. attach_hdlc_protocol() uses kmalloc() (not kzalloc), so struct ppp's protos[i].timer is uninitialized garbage until the first timer_setup(); without this init-at-attach, attaching the PPP protocol without ever bringing the device up would leave timer_shutdown_sync() operating on uninitialized memory in .detach. Moving the init out of ppp_start() (which only runs on NETDEV_UP) into the attach path makes the initialization unconditional and avoids initializing the same timer_list twice. This bug was found by static analysis.•CVE-2026-63803

References

Sources include official advisories and independent security research.