bpf: Fix NULL event->prog pointer access in bpf_overflow_handler

Andrii reported that libbpf CI hit the following oops when
running selftest send_signal:
  [ 1243.160719] BUG: kernel NULL pointer dereference, address: 0000000000000030
  [ 1243.161066] #PF: supervisor read access in kernel mode
  [ 1243.161066] #PF: error_code(0x0000) - not-present page
  [ 1243.161066] PGD 0 P4D 0
  [ 1243.161066] Oops: 0000 [#1] PREEMPT SMP NOPTI
  [ 1243.161066] CPU: 1 PID: 882 Comm: new_name Tainted: G           O      5.14.0-rc5 #1
  [ 1243.161066] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.13.0-1ubuntu1.1 04/01/2014
  [ 1243.161066] RIP: 0010:bpf_overflow_handler+0x9a/0x1e0
  [ 1243.161066] Code: 5a 84 c0 0f 84 06 01 00 00 be 66 02 00 00 48 c7 c7 6d 96 07 82 48 8b ab 18 05 00 00 e8 df 55 eb ff 66 90 48 8d 75 48 48 89 e7 <ff> 55 30 41 89 c4 e8 fb c1 f0 ff 84 c0 0f 84 94 00 00 00 e8 6e 0f
  [ 1243.161066] RSP: 0018:ffffc900000c0d80 EFLAGS: 00000046
  [ 1243.161066] RAX: 0000000000000002 RBX: ffff8881002e0dd0 RCX: 00000000b4b47cf8
  [ 1243.161066] RDX: ffffffff811dcb06 RSI: 0000000000000048 RDI: ffffc900000c0d80
  [ 1243.161066] RBP: 0000000000000000 R08: 0000000000000000 R09: 1a9d56bb00000000
  [ 1243.161066] R10: 0000000000000001 R11: 0000000000080000 R12: 0000000000000000
  [ 1243.161066] R13: ffffc900000c0e00 R14: ffffc900001c3c68 R15: 0000000000000082
  [ 1243.161066] FS:  00007fc0be2d3380(0000) GS:ffff88813bd00000(0000) knlGS:0000000000000000
  [ 1243.161066] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  [ 1243.161066] CR2: 0000000000000030 CR3: 0000000104f8e000 CR4: 00000000000006e0
  [ 1243.161066] Call Trace:
  [ 1243.161066]  <IRQ>
  [ 1243.161066]  __perf_event_overflow+0x4f/0xf0
  [ 1243.161066]  perf_swevent_hrtimer+0x116/0x130
  [ 1243.161066]  ? __lock_acquire+0x378/0x2730
  [ 1243.161066]  ? __lock_acquire+0x372/0x2730
  [ 1243.161066]  ? lock_is_held_type+0xd5/0x130
  [ 1243.161066]  ? find_held_lock+0x2b/0x80
  [ 1243.161066]  ? lock_is_held_type+0xd5/0x130
  [ 1243.161066]  ? perf_event_groups_first+0x80/0x80
  [ 1243.161066]  ? perf_event_groups_first+0x80/0x80
  [ 1243.161066]  __hrtimer_run_queues+0x1a3/0x460
  [ 1243.161066]  hrtimer_interrupt+0x110/0x220
  [ 1243.161066]  __sysvec_apic_timer_interrupt+0x8a/0x260
  [ 1243.161066]  sysvec_apic_timer_interrupt+0x89/0xc0
  [ 1243.161066]  </IRQ>
  [ 1243.161066]  asm_sysvec_apic_timer_interrupt+0x12/0x20
  [ 1243.161066] RIP: 0010:finish_task_switch+0xaf/0x250
  [ 1243.161066] Code: 31 f6 68 90 2a 09 81 49 8d 7c 24 18 e8 aa d6 03 00 4c 89 e7 e8 12 ff ff ff 4c 89 e7 e8 ca 9c 80 00 e8 35 af 0d 00 fb 4d 85 f6 <58> 74 1d 65 48 8b 04 25 c0 6d 01 00 4c 3b b0 a0 04 00 00 74 37 f0
  [ 1243.161066] RSP: 0018:ffffc900001c3d18 EFLAGS: 00000282
  [ 1243.161066] RAX: 000000000000031f RBX: ffff888104cf4980 RCX: 0000000000000000
  [ 1243.161066] RDX: 0000000000000000 RSI: ffffffff82095460 RDI: ffffffff820adc4e
  [ 1243.161066] RBP: ffffc900001c3d58 R08: 0000000000000001 R09: 0000000000000001
  [ 1243.161066] R10: 0000000000000001 R11: 0000000000080000 R12: ffff88813bd2bc80
  [ 1243.161066] R13: ffff8881002e8000 R14: ffff88810022ad80 R15: 0000000000000000
  [ 1243.161066]  ? finish_task_switch+0xab/0x250
  [ 1243.161066]  ? finish_task_switch+0x70/0x250
  [ 1243.161066]  __schedule+0x36b/0xbb0
  [ 1243.161066]  ? _raw_spin_unlock_irqrestore+0x2d/0x50
  [ 1243.161066]  ? lockdep_hardirqs_on+0x79/0x100
  [ 1243.161066]  schedule+0x43/0xe0
  [ 1243.161066]  pipe_read+0x30b/0x450
  [ 1243.161066]  ? wait_woken+0x80/0x80
  [ 1243.161066]  new_sync_read+0x164/0x170
  [ 1243.161066]  vfs_read+0x122/0x1b0
  [ 1243.161066]  ksys_read+0x93/0xd0
  [ 1243.161066]  do_syscall_64+0x35/0x80
  [ 1243.161066]  entry_SYSCALL_64_after_hwframe+0x44/0xae

The oops can also be reproduced with the following steps:
  ./vmtest.sh -s
  # at qemu shell
  cd /root/bpf && while true; do ./test_progs -t send_signal

Further analysis showed that the failure is introduced with
commit b89fbfbb ("bpf: Implement minimal BPF perf link").
With the above commit, the following scenario becomes possible:
    cpu1                        cpu2
                                hrtimer_interrupt -> bpf_overflow_handler
    (due to closing link_fd)
    bpf_perf_link_release ->
    perf_event_free_bpf_prog ->
    perf_event_free_bpf_handler ->
      WRITE_ONCE(event->overflow_handler, event->orig_overflow_handler)
      event->prog = NULL
                                bpf_prog_run(event->prog, &ctx)

In the above case, the event->prog is NULL for bpf_prog_run, hence
causing oops.

To fix the issue, check whether event->prog is NULL or not. If it
is, do not call bpf_prog_run. This seems working as the above
reproducible step runs more than one hour and I didn't see any
failures.

Fixes: b89fbfbb ("bpf: Implement minimal BPF perf link")
Signed-off-by: Yonghong Song <yhs@fb.com>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/bpf/20210819155209.1927994-1-yhs@fb.com

kernel/events/core.c

@@ -9920,13 +9920,16 @@ static void bpf_overflow_handler(struct perf_event *event,
 		.data = data,
 		.event = event,
 	};
+	struct bpf_prog *prog;
 	int ret = 0;
 
 	ctx.regs = perf_arch_bpf_user_pt_regs(regs);
 	if (unlikely(__this_cpu_inc_return(bpf_prog_active) != 1))
 		goto out;
 	rcu_read_lock();
-	ret = bpf_prog_run(event->prog, &ctx);
+	prog = READ_ONCE(event->prog);
+	if (prog)
+		ret = bpf_prog_run(prog, &ctx);
 	rcu_read_unlock();
 out:
 	__this_cpu_dec(bpf_prog_active);