drm/i915: Avoid reset lock in writing fence registers
authorChris Wilson <chris@chris-wilson.co.uk>
Tue, 19 Feb 2019 12:21:54 +0000 (12:21 +0000)
committerChris Wilson <chris@chris-wilson.co.uk>
Wed, 20 Feb 2019 16:40:13 +0000 (16:40 +0000)
commitc1d1746f6d4b37518fe3dc4aba99db1f7a155bdb
tree5406946221826e46f7d5a661dbec4ded7ac3077e
parentc41166f9a145f1c4ce2961b338f9b57495ace4b5
drm/i915: Avoid reset lock in writing fence registers

The idea of taking the reset lock around writing the fence register was
to serialise the mmio write we also perform during the reset where those
registers get clobbered. However, the lock is overkill as write tearing
between reset and fence_update() is harmless; the final value of the
fence register is the same. A race between revoke_fences() and
fence_update() is also harmless at this point as on the fault path where
this is necessary, we acquire the reset lock to coordinate ourselves in
the upper layer.

The danger of acquiring the reset lock again in fence_update() is that
we may recurse from the shrinker along the i915_gem_fault() path.

<4> [125.739646] ============================================
<4> [125.739652] WARNING: possible recursive locking detected
<4> [125.739659] 5.0.0-rc6-ga6e4cbf00557-drmtip_223+ #1 Tainted: G     U
<4> [125.739666] --------------------------------------------
<4> [125.739672] gem_mmap_gtt/1017 is trying to acquire lock:
<4> [125.739679] 00000000a730190a (&dev_priv->gpu_error.reset_backoff_srcu){+.+.}, at: i915_reset_trylock+0x0/0x310 [i915]
<4> [125.739848]
but task is already holding lock:
<4> [125.739854] 00000000a730190a (&dev_priv->gpu_error.reset_backoff_srcu){+.+.}, at: i915_reset_trylock+0x192/0x310 [i915]
<4> [125.739918]
other info that might help us debug this:
<4> [125.739925]  Possible unsafe locking scenario:

<4> [125.739930]        CPU0
<4> [125.739934]        ----
<4> [125.739937]   lock(&dev_priv->gpu_error.reset_backoff_srcu);
<4> [125.739944]   lock(&dev_priv->gpu_error.reset_backoff_srcu);
<4> [125.739950]
 *** DEADLOCK ***

<4> [125.739958]  May be due to missing lock nesting notation

<4> [125.739966] 5 locks held by gem_mmap_gtt/1017:
<4> [125.739972]  #0: 00000000471f682c (&mm->mmap_sem){++++}, at: __do_page_fault+0x133/0x500
<4> [125.739987]  #1: 0000000026542685 (&dev->struct_mutex){+.+.}, at: i915_gem_fault+0x1f6/0x860 [i915]
<4> [125.740061]  #2: 00000000a730190a (&dev_priv->gpu_error.reset_backoff_srcu){+.+.}, at: i915_reset_trylock+0x192/0x310 [i915]
<4> [125.740126]  #3: 00000000c828eb4f (fs_reclaim){+.+.}, at: fs_reclaim_acquire.part.25+0x0/0x30
<4> [125.740140]  #4: 000000002d360d65 (shrinker_rwsem){++++}, at: shrink_slab+0x1cb/0x2c0
<4> [125.740151]
stack backtrace:
<4> [125.740159] CPU: 1 PID: 1017 Comm: gem_mmap_gtt Tainted: G     U            5.0.0-rc6-ga6e4cbf00557-drmtip_223+ #1
<4> [125.740170] Hardware name: Dell Inc.                 OptiPlex 745                 /0GW726, BIOS 2.3.1  05/21/2007
<4> [125.740180] Call Trace:
<4> [125.740189]  dump_stack+0x67/0x9b
<4> [125.740199]  __lock_acquire+0xc75/0x1b00
<4> [125.740209]  ? arch_tlb_finish_mmu+0x2a/0xa0
<4> [125.740216]  ? tlb_finish_mmu+0x1a/0x30
<4> [125.740222]  ? zap_page_range_single+0xe2/0x130
<4> [125.740230]  ? lock_acquire+0xa6/0x1c0
<4> [125.740237]  lock_acquire+0xa6/0x1c0
<4> [125.740296]  ? i915_clear_error_registers+0x280/0x280 [i915]
<4> [125.740357]  i915_reset_trylock+0x44/0x310 [i915]
<4> [125.740417]  ? i915_clear_error_registers+0x280/0x280 [i915]
<4> [125.740426]  ? lockdep_hardirqs_on+0xe0/0x1b0
<4> [125.740434]  ? _raw_spin_unlock_irqrestore+0x39/0x60
<4> [125.740499]  fence_update+0x218/0x470 [i915]
<4> [125.740571]  i915_vma_unbind+0xa6/0x550 [i915]
<4> [125.740640]  i915_gem_object_unbind+0xfa/0x190 [i915]
<4> [125.740711]  i915_gem_shrink+0x2dc/0x590 [i915]
<4> [125.740722]  ? ___preempt_schedule+0x16/0x18
<4> [125.740792]  ? i915_gem_shrinker_scan+0xc9/0x130 [i915]
<4> [125.740861]  i915_gem_shrinker_scan+0xc9/0x130 [i915]
<4> [125.740870]  do_shrink_slab+0x143/0x3f0
<4> [125.740878]  shrink_slab+0x228/0x2c0
<4> [125.740886]  shrink_node+0x167/0x450
<4> [125.740894]  do_try_to_free_pages+0xc4/0x340
<4> [125.740902]  try_to_free_pages+0xdc/0x2e0
<4> [125.740911]  __alloc_pages_nodemask+0x662/0x1110
<4> [125.740921]  ? reacquire_held_locks+0xb5/0x1b0
<4> [125.740928]  ? reacquire_held_locks+0xb5/0x1b0
<4> [125.740986]  ? i915_reset_trylock+0x192/0x310 [i915]
<4> [125.741045]  ? i915_memcpy_init_early+0x30/0x30 [i915]
<4> [125.741054]  pte_alloc_one+0x12/0x70
<4> [125.741060]  __pte_alloc+0x11/0xf0
<4> [125.741067]  apply_to_page_range+0x37e/0x440
<4> [125.741127]  remap_io_mapping+0x6c/0x100 [i915]
<4> [125.741196]  i915_gem_fault+0x5a9/0x860 [i915]
<4> [125.741204]  ? ptlock_alloc+0x15/0x30
<4> [125.741212]  __do_fault+0x2c/0xb0
<4> [125.741218]  __handle_mm_fault+0x8ee/0xfa0
<4> [125.741227]  handle_mm_fault+0x196/0x3a0
<4> [125.741235]  __do_page_fault+0x246/0x500
<4> [125.741243]  ? page_fault+0x8/0x30
<4> [125.741250]  page_fault+0x1e/0x30
<4> [125.741256] RIP: 0033:0x55d0cc456e12
<4> [125.741264] Code: b0 df ff ff 89 c2 8b 85 70 df ff ff 01 c2 8b 85 70 df ff ff 48 98 48 8d 0c 85 00 00 00 00 48 8b 85 e0 df ff ff 48 01 c8 f7 d2 <89> 10 83 85 70 df ff ff 01 81 bd 70 df ff ff ff 03 00 00 7e be 48
<4> [125.741280] RSP: 002b:00007ffc1bab7ab0 EFLAGS: 00010206
<4> [125.741287] RAX: 00007fc787cb6000 RBX: 0000000000000000 RCX: 0000000000000000
<4> [125.741295] RDX: 00000000ffffffff RSI: 0000000000005401 RDI: 0000000000000002
<4> [125.741303] RBP: 00007ffc1bab9b70 R08: 00007ffc1bab7920 R09: 000000000000001b
<4> [125.741310] R10: 7165722074736554 R11: 0000000000000246 R12: 000055d0cc454a80
<4> [125.741318] R13: 00007ffc1bab9f60 R14: 0000000000000000 R15: 0000000000000000

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=109665
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190219122215.8941-4-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_gem_fence_reg.c