drm/i915: Annotate timeline lock nesting
authorChris Wilson <chris@chris-wilson.co.uk>
Tue, 8 May 2018 15:35:14 +0000 (16:35 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Tue, 8 May 2018 20:47:33 +0000 (21:47 +0100)
CI noticed

<4>[   23.430701] ============================================
<4>[   23.430706] WARNING: possible recursive locking detected
<4>[   23.430713] 4.17.0-rc4-CI-CI_DRM_4156+ #1 Not tainted
<4>[   23.430720] --------------------------------------------
<4>[   23.430725] systemd-udevd/169 is trying to acquire lock:
<4>[   23.430732]         (ptrval) (&(&timeline->lock)->rlock){....}, at: move_to_timeline+0x48/0x12c [i915]
<4>[   23.430888]
                  but task is already holding lock:
<4>[   23.430894]         (ptrval) (&(&timeline->lock)->rlock){....}, at: i915_request_submit+0x1a/0x40 [i915]
<4>[   23.430995]
                  other info that might help us debug this:
<4>[   23.431002]  Possible unsafe locking scenario:

<4>[   23.431007]        CPU0
<4>[   23.431010]        ----
<4>[   23.431013]   lock(&(&timeline->lock)->rlock);
<4>[   23.431021]   lock(&(&timeline->lock)->rlock);
<4>[   23.431028]
                   *** DEADLOCK ***

<4>[   23.431036]  May be due to missing lock nesting notation

<4>[   23.431044] 5 locks held by systemd-udevd/169:
<4>[   23.431049]  #0:         (ptrval) (&dev->mutex){....}, at: __driver_attach+0x42/0xe0
<4>[   23.431065]  #1:         (ptrval) (&dev->mutex){....}, at: __driver_attach+0x50/0xe0
<4>[   23.431078]  #2:         (ptrval) (&dev->struct_mutex){+.+.}, at: i915_gem_init+0xca/0x630 [i915]
<4>[   23.431174]  #3:         (ptrval) (rcu_read_lock){....}, at: submit_notify+0x35/0x124 [i915]
<4>[   23.431271]  #4:         (ptrval) (&(&timeline->lock)->rlock){....}, at: i915_request_submit+0x1a/0x40 [i915]
<4>[   23.431369]
                  stack backtrace:
<4>[   23.431377] CPU: 0 PID: 169 Comm: systemd-udevd Not tainted 4.17.0-rc4-CI-CI_DRM_4156+ #1
<4>[   23.431385] Hardware name: Dell Inc.                 OptiPlex GX280               /0G8310, BIOS A04 02/09/2005
<4>[   23.431394] Call Trace:
<4>[   23.431403]  dump_stack+0x67/0x9b
<4>[   23.431411]  __lock_acquire+0xc67/0x1b50
<4>[   23.431421]  ? ring_buffer_lock_reserve+0x154/0x3f0
<4>[   23.431429]  ? lock_acquire+0xa6/0x210
<4>[   23.431435]  lock_acquire+0xa6/0x210
<4>[   23.431530]  ? move_to_timeline+0x48/0x12c [i915]
<4>[   23.431540]  _raw_spin_lock+0x2a/0x40
<4>[   23.431634]  ? move_to_timeline+0x48/0x12c [i915]
<4>[   23.431730]  move_to_timeline+0x48/0x12c [i915]
<4>[   23.431826]  __i915_request_submit+0xfa/0x280 [i915]
<4>[   23.431923]  i915_request_submit+0x25/0x40 [i915]
<4>[   23.432024]  i9xx_submit_request+0x11/0x140 [i915]
<4>[   23.432120]  submit_notify+0x8d/0x124 [i915]
<4>[   23.432202]  __i915_sw_fence_complete+0x81/0x250 [i915]
<4>[   23.432300]  __i915_request_add+0x31c/0x7c0 [i915]
<4>[   23.432395]  i915_gem_init+0x621/0x630 [i915]
<4>[   23.432476]  i915_driver_load+0xbee/0x10b0 [i915]
<4>[   23.432485]  ? trace_hardirqs_on_caller+0xe0/0x1b0
<4>[   23.432566]  i915_pci_probe+0x29/0x90 [i915]
<4>[   23.432574]  pci_device_probe+0xa1/0x130
<4>[   23.432582]  driver_probe_device+0x306/0x480
<4>[   23.432589]  __driver_attach+0xb7/0xe0
<4>[   23.432596]  ? driver_probe_device+0x480/0x480
<4>[   23.432602]  ? driver_probe_device+0x480/0x480
<4>[   23.432609]  bus_for_each_dev+0x74/0xc0
<4>[   23.432616]  bus_add_driver+0x15f/0x250
<4>[   23.432623]  ? 0xffffffffa02d7000
<4>[   23.432629]  driver_register+0x52/0xc0
<4>[   23.432635]  ? 0xffffffffa02d7000
<4>[   23.432642]  do_one_initcall+0x58/0x370
<4>[   23.432653]  ? do_init_module+0x1d/0x1ea
<4>[   23.432660]  ? rcu_read_lock_sched_held+0x6f/0x80
<4>[   23.432667]  ? kmem_cache_alloc_trace+0x282/0x2e0
<4>[   23.432675]  do_init_module+0x56/0x1ea
<4>[   23.432682]  load_module+0x2435/0x2b20
<4>[   23.432694]  ? __se_sys_finit_module+0xd3/0xf0
<4>[   23.432701]  __se_sys_finit_module+0xd3/0xf0
<4>[   23.432710]  do_syscall_64+0x55/0x190
<4>[   23.432717]  entry_SYSCALL_64_after_hwframe+0x49/0xbe
<4>[   23.432724] RIP: 0033:0x7fa780782839
<4>[   23.432729] RSP: 002b:00007ffcea73e668 EFLAGS: 00000246 ORIG_RAX: 0000000000000139
<4>[   23.432738] RAX: ffffffffffffffda RBX: 0000561a472a4b30 RCX: 00007fa780782839
<4>[   23.432745] RDX: 0000000000000000 RSI: 00007fa7804610e5 RDI: 000000000000000e
<4>[   23.432752] RBP: 00007fa7804610e5 R08: 0000000000000000 R09: 00007ffcea73e780
<4>[   23.432758] R10: 000000000000000e R11: 0000000000000246 R12: 0000000000000000
<4>[   23.432765] R13: 0000561a47296450 R14: 0000000000020000 R15: 0000561a472a4b30

but did not report it as an issue as it only occurred during the first
module on boot. This is due to the removal of the distinct global
timeline, and its separate lock class. So instead mark up the expected
nesting. An alternative would be to define a separate lock class for the
engine, but since we only expect to have a single point of nesting, we
can avoid having multiple lock classes for the struct.

Fixes: a89d1f921c15 ("drm/i915: Split i915_gem_timeline into individual timelines")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Tested-by: Michel Thierry <michel.thierry@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180508153514.20251-1-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_request.c

index f336942229cf3e194512939917c62aaf5ff8c53d..8928894dd9c77d1b1aa4d28c5a0ab06ec4b83a25 100644 (file)
@@ -502,7 +502,7 @@ static void move_to_timeline(struct i915_request *request,
        GEM_BUG_ON(request->timeline == &request->engine->timeline);
        lockdep_assert_held(&request->engine->timeline.lock);
 
-       spin_lock(&request->timeline->lock);
+       spin_lock_nested(&request->timeline->lock, SINGLE_DEPTH_NESTING);
        list_move_tail(&request->link, &timeline->requests);
        spin_unlock(&request->timeline->lock);
 }