[Intel-gfx] [PATCH] semaphore: might_lock(->pi_lock) in up()

[Daniel Vetter]

In one of our bugs we very rarely see a lockdep splat on boot-up. The
locking loop only happens when the console_sem semaphore is contended
in a specific path, which doesn't happen all too often. Help out
lockdep a bit with a might_lock annotation.

For reference, the full splat:

WARNING: CPU: 0 PID: 216 at lib/debugobjects.c:291 debug_print_object+0x67/0x80
---[ end trace 1dcad5e101ee43df ]---

======================================================
WARNING: possible circular locking dependency detected
4.16.0-rc5-g178cfb9373cc-drmtip_2+ #1 Not tainted
------------------------------------------------------
systemd-udevd/216 is trying to acquire lock:
 ((console_sem).lock){..-.}, at: [<0000000073293a25>] down_trylock+0xa/0x30

but task is already holding lock:
 (&obj_hash[i].lock){-.-.}, at: [<00000000023c4e37>] debug_check_no_obj_freed+0xa4/0x230

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #3 (&obj_hash[i].lock){-.-.}:
       __debug_object_init+0x5f/0x430
       hrtimer_init+0x1b/0x150
       init_dl_task_timer+0x17/0x30
       __sched_fork.isra.16+0xa7/0x100
       init_idle+0x53/0x270
       sched_init+0x42f/0x4a5
       start_kernel+0x20e/0x3f0
       secondary_startup_64+0xa5/0xb0

-> #2 (&rq->lock){-.-.}:
       task_fork_fair+0x36/0x120
       sched_fork+0x12f/0x2a0
       copy_process.part.7+0x5e4/0x1d90
       _do_fork+0xc0/0x6b0
       kernel_thread+0x20/0x30
       rest_init+0x1d/0x220
       start_kernel+0x3e8/0x3f0
       secondary_startup_64+0xa5/0xb0

-> #1 (&p->pi_lock){-.-.}:
       try_to_wake_up+0x27/0x610
       up+0x3b/0x50
       __up_console_sem+0x2e/0x50
       console_unlock+0x31f/0x650
       do_con_write.part.12+0x6fd/0x9a0
       con_write+0x4d/0x60
       do_output_char+0x168/0x1d0
       n_tty_write+0x1f8/0x450
       tty_write+0x1b6/0x2e0
       do_iter_write+0x140/0x180
       vfs_writev+0x84/0xf0
       do_writev+0x4c/0xd0
       do_syscall_64+0x65/0x1a0
       entry_SYSCALL_64_after_hwframe+0x42/0xb7

-> #0 ((console_sem).lock){..-.}:
       _raw_spin_lock_irqsave+0x33/0x50
       down_trylock+0xa/0x30
       __down_trylock_console_sem+0x20/0x80
       console_trylock+0xe/0x60
       vprintk_emit+0x24f/0x4f0
       printk+0x3e/0x46
       __warn_printk+0x37/0x70
       debug_print_object+0x67/0x80
       debug_check_no_obj_freed+0x1ef/0x230
       __vunmap+0x5b/0x110
       load_module+0x2347/0x2dc0
       SyS_finit_module+0xa5/0xe0
       do_syscall_64+0x65/0x1a0
       entry_SYSCALL_64_after_hwframe+0x42/0xb7

other info that might help us debug this:

Chain exists of:
  (console_sem).lock --> &rq->lock --> &obj_hash[i].lock

 Possible unsafe locking scenario:

       CPU0                    CPU1
       ----                    ----
  lock(&obj_hash[i].lock);
                               lock(&rq->lock);
                               lock(&obj_hash[i].lock);
  lock((console_sem).lock);

 *** DEADLOCK ***

1 lock held by systemd-udevd/216:
 #0:  (&obj_hash[i].lock){-.-.}, at: [<00000000023c4e37>] debug_check_no_obj_freed+0xa4/0x230

stack backtrace:
CPU: 0 PID: 216 Comm: systemd-udevd Not tainted 4.16.0-rc5-g178cfb9373cc-drmtip_2+ #1
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.10.1-0-g8891697-prebuilt.qemu-project.org 04/01/2014
Call Trace:
 dump_stack+0x5f/0x86
 print_circular_bug.isra.18+0x1d0/0x2c0
 __lock_acquire+0x14ae/0x1b60
 ? lock_acquire+0xaf/0x200
 lock_acquire+0xaf/0x200
 ? down_trylock+0xa/0x30
 ? vprintk_emit+0x24f/0x4f0
 _raw_spin_lock_irqsave+0x33/0x50
 ? down_trylock+0xa/0x30
 down_trylock+0xa/0x30
 __down_trylock_console_sem+0x20/0x80
 console_trylock+0xe/0x60
 vprintk_emit+0x24f/0x4f0
 printk+0x3e/0x46
 ? clock_was_set_work+0x20/0x20
 __warn_printk+0x37/0x70
 ? debug_check_no_obj_freed+0xa4/0x230
 ? clock_was_set_work+0x20/0x20
 debug_print_object+0x67/0x80
 debug_check_no_obj_freed+0x1ef/0x230
 __vunmap+0x5b/0x110
 load_module+0x2347/0x2dc0
 ? show_coresize+0x20/0x20
 ? vfs_read+0x126/0x150
 ? SyS_finit_module+0xa5/0xe0
 SyS_finit_module+0xa5/0xe0
 do_syscall_64+0x65/0x1a0
 entry_SYSCALL_64_after_hwframe+0x42/0xb7
RIP: 0033:0x7f5037e42839
RSP: 002b:00007fff88b69df8 EFLAGS: 00000246 ORIG_RAX: 0000000000000139
RAX: ffffffffffffffda RBX: 00005599ad713550 RCX: 00007f5037e42839
RDX: 0000000000000000 RSI: 00007f5037b210e5 RDI: 000000000000000f
RBP: 00007f5037b210e5 R08: 0000000000000000 R09: 00007fff88b69f10
R10: 000000000000000f R11: 0000000000000246 R12: 0000000000000000
R13: 00005599ad707370 R14: 0000000000020000 R15: 00005599ad713550

References: https://bugs.freedesktop.org/show_bug.cgi?id=105600
Signed-off-by: Daniel Vetter <daniel.vetter at intel.com>

---

For now this is just a quick test patch for CI. I'll resubmit once we
gathered some data.
-Daniel
---
 kernel/locking/semaphore.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/kernel/locking/semaphore.c b/kernel/locking/semaphore.c
index 561acdd39960..6a9a671cef33 100644
--- a/kernel/locking/semaphore.c
+++ b/kernel/locking/semaphore.c
@@ -181,6 +181,9 @@ void up(struct semaphore *sem)
 	unsigned long flags;
 
 	raw_spin_lock_irqsave(&sem->lock, flags);
+	/* if we wake up a task we need its ->pi_lock, just pretend to take ours
+	 * for the lockdep accounting. */
+	might_lock(&current->pi_lock);
 	if (likely(list_empty(&sem->wait_list)))
 		sem->count++;
 	else
-- 
2.16.2