drm/i915: Take reference for signaling the request from hardirq
authorChris Wilson <chris@chris-wilson.co.uk>
Fri, 3 Mar 2017 14:45:57 +0000 (14:45 +0000)
committerChris Wilson <chris@chris-wilson.co.uk>
Fri, 3 Mar 2017 15:53:09 +0000 (15:53 +0000)
Being inside a spinlock signaling that the hardware just completed a
request doesn't prevent a second thread already spotting that the
request is complete, freeing it and reallocating it! The code currently
tries to prevent this using RCU -- but that only prevents the request
from being freed, it doesn't prevent us from reallocating it - that
requires us to take a reference.

[  206.922985] BUG: spinlock already unlocked on CPU#4, gem_exec_parall/7796
[  206.922994]  lock: 0xffff8801c6047120, .magic: dead4ead, .owner: <none>/-1, .owner_cpu: -1
[  206.923000] CPU: 4 PID: 7796 Comm: gem_exec_parall Not tainted 4.10.0-CI-Patchwork_4008+ #1
[  206.923006] Hardware name: System manufacturer System Product Name/Z170M-PLUS, BIOS 1805 06/20/2016
[  206.923012] Call Trace:
[  206.923014]  <IRQ>
[  206.923019]  dump_stack+0x67/0x92
[  206.923023]  spin_dump+0x73/0xc0
[  206.923027]  do_raw_spin_unlock+0x79/0xb0
[  206.923031]  _raw_spin_unlock_irqrestore+0x27/0x60
[  206.923042]  dma_fence_signal+0x160/0x230
[  206.923060]  notify_ring+0xae/0x2e0 [i915]
[  206.923073]  ? ibx_hpd_irq_handler+0xc0/0xc0 [i915]
[  206.923086]  gen8_gt_irq_handler+0x219/0x290 [i915]
[  206.923100]  gen8_irq_handler+0x8e/0x6b0 [i915]
[  206.923105]  __handle_irq_event_percpu+0x58/0x370
[  206.923109]  handle_irq_event_percpu+0x1e/0x50
[  206.923113]  handle_irq_event+0x34/0x60
[  206.923117]  handle_edge_irq+0xbe/0x150
[  206.923122]  handle_irq+0x15/0x20
[  206.923126]  do_IRQ+0x63/0x130
[  206.923142]  ? i915_mutex_lock_interruptible+0x39/0x140 [i915]
[  206.923148]  common_interrupt+0x90/0x90
[  206.923153] RIP: 0010:osq_lock+0x77/0x110
[  206.923157] RSP: 0018:ffffc90001cabaa0 EFLAGS: 00000246 ORIG_RAX: ffffffffffffff6e
[  206.923164] RAX: 0000000000000000 RBX: ffff880236d1abc0 RCX: ffff8801ef642fc0
[  206.923169] RDX: ffff8801ef6427c0 RSI: ffffffff81c6e7fd RDI: ffffffff81c7c848
[  206.923175] RBP: ffffc90001cabab8 R08: 00000000692bb19b R09: 08c1493200000000
[  206.923180] R10: 0000000000000001 R11: 0000000000000001 R12: ffff880236cdabc0
[  206.923185] R13: ffff8802207f00b0 R14: ffffffffa00b7cd9 R15: ffff8802207f0070
[  206.923191]  </IRQ>
[  206.923206]  ? i915_mutex_lock_interruptible+0x39/0x140 [i915]
[  206.923213]  __mutex_lock+0x649/0x990
[  206.923217]  ? __mutex_lock+0xb0/0x990
[  206.923221]  ? _raw_spin_unlock+0x2c/0x50
[  206.923226]  ? __pm_runtime_resume+0x56/0x80
[  206.923242]  ? i915_mutex_lock_interruptible+0x39/0x140 [i915]
[  206.923249]  mutex_lock_interruptible_nested+0x16/0x20
[  206.923264]  i915_mutex_lock_interruptible+0x39/0x140 [i915]
[  206.923270]  ? __pm_runtime_resume+0x56/0x80
[  206.923285]  i915_gem_do_execbuffer.isra.15+0x442/0x1d10 [i915]
[  206.923291]  ? __lock_acquire+0x449/0x1b50
[  206.923296]  ? __might_fault+0x3e/0x90
[  206.923301]  ? __might_fault+0x87/0x90
[  206.923305]  ? __might_fault+0x3e/0x90
[  206.923320]  i915_gem_execbuffer2+0xb5/0x220 [i915]
[  206.923327]  drm_ioctl+0x200/0x450
[  206.923341]  ? i915_gem_execbuffer+0x330/0x330 [i915]
[  206.923348]  do_vfs_ioctl+0x90/0x6e0
[  206.923352]  ? __fget+0x108/0x200
[  206.923356]  ? expand_files+0x2b0/0x2b0
[  206.923361]  SyS_ioctl+0x3c/0x70
[  206.923365]  entry_SYSCALL_64_fastpath+0x1c/0xb1
[  206.923369] RIP: 0033:0x7fdd75fc6357
[  206.923373] RSP: 002b:00007fdd20e59bf8 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
[  206.923380] RAX: ffffffffffffffda RBX: ffffffff81481ff3 RCX: 00007fdd75fc6357
[  206.923385] RDX: 00007fdd20e59c70 RSI: 0000000040406469 RDI: 0000000000000003
[  206.923390] RBP: ffffc90001cabf88 R08: 0000000000000040 R09: 00000000000003f7
[  206.923396] R10: 0000000000000000 R11: 0000000000000246 R12: 0000000000000000
[  206.923401] R13: 0000000000000003 R14: 0000000040406469 R15: 0000000001cf9cb0
[  206.923408]  ? __this_cpu_preempt_check+0x13/0x20

Fixes: 56299fb7d904 ("drm/i915: Signal first fence from irq handler if complete")
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=100051
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170303144557.4815-1-chris@chris-wilson.co.uk
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
drivers/gpu/drm/i915/i915_irq.c

index 675c28d5465867319862e543dd98be9d6d7f4098..cb1424abb7c08958681b20676d5f189bc90a5272 100644 (file)
@@ -1042,8 +1042,6 @@ static void notify_ring(struct intel_engine_cs *engine)
        atomic_inc(&engine->irq_count);
        set_bit(ENGINE_IRQ_BREADCRUMB, &engine->irq_posted);
 
-       rcu_read_lock();
-
        spin_lock(&engine->breadcrumbs.lock);
        wait = engine->breadcrumbs.first_wait;
        if (wait) {
@@ -1060,7 +1058,7 @@ static void notify_ring(struct intel_engine_cs *engine)
                 */
                if (i915_seqno_passed(intel_engine_get_seqno(engine),
                                      wait->seqno))
-                       rq = wait->request;
+                       rq = i915_gem_request_get(wait->request);
 
                wake_up_process(wait->tsk);
        } else {
@@ -1068,10 +1066,10 @@ static void notify_ring(struct intel_engine_cs *engine)
        }
        spin_unlock(&engine->breadcrumbs.lock);
 
-       if (rq)
+       if (rq) {
                dma_fence_signal(&rq->fence);
-
-       rcu_read_unlock();
+               i915_gem_request_put(rq);
+       }
 
        trace_intel_engine_notify(engine, wait);
 }