Re: [bpf-next v3 1/2] bpf: Offload kptr destructors that run from NMI

On Mon, May 11, 2026 at 08:51:53AM -0700, Alexei Starovoitov wrote:
> On Sun May 10, 2026 at 6:49 PM PDT, Justin Suess wrote:
> > On Sun, May 10, 2026 at 03:38:08PM -0700, Alexei Starovoitov wrote:
> >> On Sun, May 10, 2026 at 8:14 AM Justin Suess <[email protected]> wrote:
Here's a reproducer for the cgroup case:

https://gist.githubusercontent.com/RazeLighter777/5f77cdfe035a4e22ee2642ae7db6387d/raw/10898d27040a07098cccc5d0785d9ad6620344e7/cgroup_kptr_nmi_deadlock_repro

Hacked together with an AI prompt but functional.

Exercises a different path, but more consistently splats even without
CONFIG_RCU_NOCB_CPU / CONFIG_RCU_EXPERT since this dtor uses workqueue.

Had to use an fexit hook to get the timing condition right to release
the last cgroup reference.

But this lets you see the deadlock is indeed in the dtor in NMI.

This is on the same bpf-next/master
7e033543a2ab4c72319201298ed458e3bbddd82f:

  [   15.160694] ================================
  [   15.160695] WARNING: inconsistent lock state
  [   15.160695] 7.1.0-rc2-g7e033543a2ab-dirty #130 Not tainted
  [   15.160697] --------------------------------
  [   15.160697] inconsistent {INITIAL USE} -> {IN-NMI} usage.
  [   15.160698] test_progs/434 [HC1[1]:SC0[0]:HE0:SE1] takes:
  [   15.160700] ffff9096fd66ced8 (&pool->lock){-.-.}-{2:2}, at: __queue_work+0xde/0x720
  [   15.160707] {INITIAL USE} state was registered at:
  [   15.160708]   lock_acquire+0xbf/0x2e0
  [   15.160711]   _raw_spin_lock+0x30/0x40
  [   15.160715]   __queue_work+0xde/0x720
  [   15.160716]   queue_work_on+0x54/0xa0
  [   15.160716]   start_poll_synchronize_rcu_expedited+0xaf/0x110
  [   15.160719]   rcu_init+0x958/0x990
  [   15.160722]   start_kernel+0x746/0x980
  [   15.160725]   x86_64_start_reservations+0x24/0x30
  [   15.160727]   __pfx_reserve_bios_regions+0x0/0x10
  [   15.160729]   common_startup_64+0x12c/0x138
  [   15.160731] irq event stamp: 18704
  [   15.160732] hardirqs last  enabled at (18703): [<ffffffffa200148a>] asm_sysvec_apic_timer_interrupt+0x1a/0x20
  [   15.160734] hardirqs last disabled at (18704): [<ffffffffa30e3f8f>] exc_nmi+0x7f/0x110
  [   15.160737] softirqs last  enabled at (18698): [<ffffffffa22e5800>] __irq_exit_rcu+0xc0/0x100
  [   15.160739] softirqs last disabled at (18687): [<ffffffffa22e5800>] __irq_exit_rcu+0xc0/0x100
  [   15.160741] 
  [   15.160741] other info that might help us debug this:
  [   15.160741]  Possible unsafe locking scenario:
  [   15.160741] 
  [   15.160742]        CPU0
  [   15.160742]        ----
  [   15.160742]   lock(&pool->lock);
  [   15.160743]   <Interrupt>
  [   15.160743]     lock(&pool->lock);
  [   15.160744] 
  [   15.160744]  *** DEADLOCK ***
  [   15.160744] 
  [   15.160744] no locks held by test_progs/434.
  [   15.160745] 
  [   15.160745] stack backtrace:
  [   15.160747] CPU: 1 UID: 0 PID: 434 Comm: test_progs Not tainted 7.1.0-rc2-g7e033543a2ab-dirty #130 PREEMPT(full) 
  [   15.160749] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.17.0-0-gb52ca86e094d-prebuilt.qemu.org 04/01/2014
  [   15.160750] Call Trace:
  [   15.160751]  <TASK>
  [   15.160753]  dump_stack_lvl+0x5d/0x80
  [   15.160757]  print_usage_bug.part.0+0x22b/0x2c0
  [   15.160760]  lock_acquire+0x295/0x2e0
  [   15.160762]  ? srso_alias_return_thunk+0x5/0xfbef5
  [   15.160763]  ? __queue_work+0xde/0x720
  [   15.160767]  _raw_spin_lock+0x30/0x40
  [   15.160768]  ? __queue_work+0xde/0x720
  [   15.160769]  __queue_work+0xde/0x720
  [   15.160772]  queue_work_on+0x54/0xa0
  [   15.160774]  bpf_cgroup_release_dtor+0x12e/0x140
  [   15.160778]  bpf_obj_free_fields+0x118/0x250
  [   15.160782]  free_htab_elem+0x85/0xd0
  [   15.160785]  htab_map_delete_elem+0x168/0x230
  [   15.160790]  bpf_prog_23fcbbeb395ac6b4_clear_cgroup_kptrs_from_nmi+0x54/0x74
  [   15.160792]  bpf_trace_run3+0x126/0x430
  [   15.160795]  ? __pfx_perf_event_nmi_handler+0x10/0x10
  [   15.160799]  nmi_handle.part.0+0x15b/0x250
  [   15.160802]  ? __pfx_perf_event_nmi_handler+0x10/0x10
  [   15.160804]  default_do_nmi+0x120/0x180
  [   15.160807]  exc_nmi+0xe3/0x110
  [   15.160809]  asm_exc_nmi+0xb7/0x100
  [   15.160810] RIP: 0033:0x5607a669541b
  [   15.160813] Code: c7 45 f0 00 00 00 00 eb 1a 8b 55 f0 8b 45 f4 01 d0 48 63 d0 48 8b 45 a8 48 01 d0 48 89 45 a8 83 45 f0 01 81 7d f0 3f 42 0f 00 <7e> dd e8 7e f5 ff ff 48 89 45 f8 48 8b 45 f8 48 3b 45 e8 73 16 83
  [   15.160814] RSP: 002b:00007ffdb09c1dc0 EFLAGS: 00000293
  [   15.160816] RAX: 0000003aced4e2f4 RBX: 00007f1d8d574000 RCX: 000000000000000f
  [   15.160816] RDX: 00000000000ad857 RSI: 00007f1d8d577000 RDI: 0000000000000001
  [   15.160817] RBP: 00007ffdb09c1e30 R08: 00007ffdb09c1da0 R09: 00007f1d8d577010
  [   15.160818] R10: 0000000000001614 R11: 0009718b9187183f R12: 0000000000000003
  [   15.160818] R13: 00007f1d8d5b6000 R14: 00007ffdb09c3358 R15: 00005607a9daf890
  [   15.160824]  </TASK>
  [   15.214040] perf: interrupt took too long (2501 > 2500), lowering kernel.perf_event_max_sample_rate to 79000
  [   15.246002] perf: interrupt took too long (3135 > 3126), lowering kernel.perf_event_max_sample_rate to 63000
  [   15.308032] perf: interrupt took too long (3928 > 3918), lowering kernel.perf_event_max_sample_rate to 50000
  [   15.500072] perf: interrupt took too long (4912 > 4910), lowering kernel.perf_event_max_sample_rate to 40000

Justin