Re: [PATCH] mm: memcontrol: fix NULL pointer deref in high_work_func() on dying memcg

[Muchun Song <muchun.song@linux.dev>]

> On Oct 24, 2026, at 06:23, T.J. Mercier <tjmercier@google.com> wrote:
> 
> When memory.high is exceeded in non-task context, try_charge_memcg()
> schedules memcg->high_work without taking a CSS reference on the memcg,
> and high_work_func() invokes reclaim_high() without pinning the memcg.
> 
> Before commit f1cf8d2f36dc ("mm: memcontrol: eliminate the problem of
> dying memory cgroup for LRU folios"), memory usage above memory.high
> kept the memcg's refcount non-zero. With reparenting, folios and objcgs
> charged to a child memcg are reparented to its parent in
> mem_cgroup_css_offline(), leaving the offlined child's memory usage
> permanently above its memory.high even after its last reference is
> dropped.
> 
> cancel_work_sync(&memcg->high_work) only runs at the end of teardown in
> mem_cgroup_css_free(), so high_work_func() can start or still be running
> reclaim_high() after mem_cgroup_css_released() or shrink_many() has
> called lru_gen_release_memcg() and unhashed lruvec->lrugen.list. This
> causes lru_gen_rotate_memcg() to dereference a NULL pprev pointer.
> 
> Fix this by acquiring a reference with css_tryget_online() in
> high_work_func() before calling reclaim_high(), and by resetting
> memory.high and swap.high to PAGE_COUNTER_MAX in
> mem_cgroup_css_offline() so that charges via references on an offlined
> memcg do not trigger high_work on the dead child or skip checking live
> ancestors in try_charge_memcg().
> 
> Oops: general protection fault, probably for non-canonical address 0xdffffc0000000000: 0000 [#1] SMP KASAN NOPTI
> KASAN: null-ptr-deref in range [0x0000000000000000-0x0000000000000007]
> CPU: 1 UID: 0 PID: 27 Comm: kworker/1:0 Tainted: G    B   W           7.3.0-rc5-g2bba28e98578 #284 PREEMPT(full) <- Taint from lru_gen_rotate_memcg() VM_WARN_ON_ONCE
> Tainted: [B]=BAD_PAGE, [W]=WARN
> Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.17.0-debian-1.17.0-1 04/01/2014
> Workqueue: events high_work_func
> RIP: 0010:lru_gen_rotate_memcg+0x1a3/0xe90
> Code: fd 7f 09 45 84 ed 0f 85 64 0c 00 00 48 89 cf c6 83 6a 07 00 00 00 48 c1 ef 03 88 83 69 07 00 00 48 b8 00 00 00 00 00 fc ff df <80> 3c 07 00 0f 85 ec 0b 00 00 48 89 31 40 f6 c6 01 0f 85 4e 00
> RSP: 0018:ffa00000001ef4c8 EFLAGS: 00010046
> RAX: dffffc0000000000 RBX: ff1100000779e440 RCX: 0000000000000000
> RDX: 0000000000000001 RSI: ff1100000779b9b0 RDI: 0000000000000000
> RBP: ff110000bffd8e40 R08: 0000000000000000 R09: ff1100000779ebb8
> R10: ff1100000779ebb0 R11: 0000000000000246 R12: 0000000000000000
> R13: 0000000000000000 R14: ffa00000001efb30 R15: 0000000000000002
> FS:  0000000000000000(0000) GS:ff1100011c0e7000(0000) knlGS:0000000000000000
> CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> CR2: 000000000251ac88 CR3: 0000000009d6c000 CR4: 0000000000751ef0
> PKRU: 55555554
> Call Trace:
> <TASK>
> ? read_ctrl_pos+0x22f/0x3f0
> shrink_lruvec+0x16cb/0x1b80
> ? ret_from_fork_asm+0x1a/0x30
> ? lapic_next_event+0x15/0x20
> ? clockevents_program_event+0x2bb/0x6f0
> ? lapic_next_event+0x15/0x20
> ? __pfx_shrink_lruvec+0x10/0x10
> ? __pfx_clockevents_program_event+0x10/0x10
> ? _raw_spin_lock+0x84/0xe0
> ? __pfx__raw_spin_lock+0x10/0x10
> ? kasan_save_stack+0x30/0x50
> ? psi_group_change+0x381/0x740
> ? __pfx_sched_clock_cpu+0x10/0x10
> ? save_fpregs_to_fpstate+0xc4/0x2c0
> ? finish_task_switch.isra.0+0x1de/0xac0
> ? __schedule+0xf62/0x42c0
> ? __pfx___schedule+0x10/0x10
> ? __pfx___irq_work_queue_local+0x10/0x10
> ? shrink_node+0x626/0x31e0
> shrink_node+0x626/0x31e0
> ? ring_buffer_unlock_commit+0x11d/0x5c0
> ? kswapd_clear_hopeless+0xa0/0x150
> ? __pfx_shrink_node+0x10/0x10
> ? 0xffffffffc0000095
> do_try_to_free_pages+0x255/0x13b0
> ? ring_buffer_unlock_commit+0x11d/0x5c0
> ? reclaim_high+0x10a/0x1b0
> try_to_free_mem_cgroup_pages+0x1f9/0x5a0
> ? __pfx_try_to_free_mem_cgroup_pages+0x10/0x10
> ? irqentry_exit+0xc6/0x670
> reclaim_high+0x10a/0x1b0
> ? __pfx_reclaim_high+0x10/0x10
> ? process_one_work+0x5ff/0x1030
> ? __pfx_high_work_func+0x10/0x10
> process_one_work+0x6e2/0x1030
> ? __pfx___schedule+0x10/0x10
> ? __pfx_process_one_work+0x10/0x10
> worker_thread+0x422/0xbb0
> ? __pfx_worker_thread+0x10/0x10
> kthread+0x2cd/0x3a0
> ? recalc_sigpending+0x15c/0x200
> ? __pfx_kthread+0x10/0x10
> ret_from_fork+0x368/0x590
> ? __pfx_ret_from_fork+0x10/0x10
> ? __switch_to+0x588/0xf50
> ? __pfx_kthread+0x10/0x10
> ret_from_fork_asm+0x1a/0x30
> </TASK>
> 
> Fixes: f1cf8d2f36dc ("mm: memcontrol: eliminate the problem of dying memory cgroup for LRU folios")

Hi T.J.,

Thanks for fixing this.

I have one question about the Fixes tag.

The changelog says that, before f1cf8d2f36dc, memory usage above
memory.high kept the memcg reference count non-zero. This is true for
LRU folios: before that commit, each charged LRU folio held a CSS
reference until it was uncharged.

However, I wonder whether objcg-based kmem already provided a reachable
path before f1cf8d2f36dc.

In particular, BPF alloc_bulk() mostly runs from irq_work. It obtains a
temporary memcg reference from a persistent objcg, uses set_active_memcg()
while performing __GFP_ACCOUNT allocations, and drops the temporary
reference after the allocation. The charge can therefore queue
memcg->high_work in non-task context without transferring a CSS
reference to the work item.

If this races with memcg offlining, objcg reparenting can redirect the
objcg to the parent while the old child's page-counter usage remains
above memory.high. Once the temporary CSS reference is dropped,
mem_cgroup_css_released() can call lru_gen_release_memcg() before the
queued high_work runs. The worker can then enter reclaim_high() and
rotate an already removed lrugen.list.

This path appears to have all the required pieces after:

  e4dde56cd208 ("mm: multi-gen LRU: per-node lru_gen_folio lists")

Commit 6867c7a33206 later changed the release operation to
hlist_nulls_del_init_rcu(), which explains the exact NULL pprev seen in
the reported crash. Before that, the same invalid second deletion would
instead encounter LIST_POISON2.

So f1cf8d2f36dc clearly broadens the problem to ordinary LRU folios, but
I am not very sure it is the first commit after which the underlying race is
reachable.

Is there an additional lifetime guarantee that prevents the pre-f1cf
objcg/kmem path above? If no, would the following be more appropriate?

Fixes: e4dde56cd208 ("mm: multi-gen LRU: per-node lru_gen_folio lists")

Thanks,
Muchun