drm/i915: Add a sanity check that no request is submitted in the middle
authorChris Wilson <chris@chris-wilson.co.uk>
Wed, 11 Jan 2017 14:08:58 +0000 (14:08 +0000)
committerChris Wilson <chris@chris-wilson.co.uk>
Wed, 11 Jan 2017 14:31:48 +0000 (14:31 +0000)
It is an error to start a new request on the same timeline (ringbuffer)
as the current one before the current is submitted. If there are two
requests emitting to the ringbuffer at the same time, the operation is
undefined. We can catch this by checking for the timeline having a later
seqno than ours when we come to submit our request.

Currently we have this check at the end of __i915_add_request, but
having an early check as well isolates a failure in the caller versus a
failure in sealing the request (i.e. from inside __i915_add_request
itself). For example, CI is currently tripping over this late assertion
on ctg/ilk:

[  100.329399] [IGT] gem_cs_tlb: starting subtest basic-default
[  100.336333] ------------[ cut here ]------------
[  100.336341] kernel BUG at drivers/gpu/drm/i915/i915_gem_request.c:908!
[  100.336347] invalid opcode: 0000 [#1] PREEMPT SMP
[  100.336351] Modules linked in: snd_hda_intel i915 snd_hda_codec_generic snd_hda_codec snd_hwdep snd_hda_core snd_pcm coretemp mei_me lpc_ich mei e1000e ptp pps_core [last unloaded: i915]
[  100.336373] CPU: 0 PID: 6308 Comm: gem_cs_tlb Tainted: G     U          4.10.0-rc3-CI-CI_DRM_2045+ #1
[  100.336380] Hardware name: LENOVO 7465CTO/7465CTO, BIOS 6DET44WW (2.08 ) 04/22/2009
[  100.336386] task: ffff88012b738040 task.stack: ffffc90000560000
[  100.336441] RIP: 0010:__i915_add_request+0x4aa/0x510 [i915]
[  100.336445] RSP: 0018:ffffc90000563ac0 EFLAGS: 00010212
[  100.336451] RAX: 0000000000005d52 RBX: ffff880133bb84c0 RCX: 0000000000000001
[  100.336456] RDX: 0000000080000001 RSI: ffff88012b738860 RDI: 00000000ffffffff
[  100.336461] RBP: ffffc90000563b00 R08: ffff880133bb8780 R09: 0000000000000000
[  100.336466] R10: 0000000000000000 R11: 0000000000000000 R12: ffff88012f53d950
[  100.336472] R13: ffff88012a2b0af8 R14: ffff88012a5b0008 R15: ffff88012f53d960
[  100.336477] FS:  00007f0d19da38c0(0000) GS:ffff88013bc00000(0000) knlGS:0000000000000000
[  100.336483] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[  100.336488] CR2: 00007f0d17706000 CR3: 000000012aa3e000 CR4: 00000000000406f0
[  100.336496] Call Trace:
[  100.336527]  i915_gem_switch_to_kernel_context+0x131/0x1b0 [i915]
[  100.336559]  i915_gem_evict_vm+0x202/0x2b0 [i915]
[  100.336590]  i915_gem_execbuffer_reserve.isra.9+0x3ae/0x440 [i915]
[  100.336623]  i915_gem_do_execbuffer.isra.15+0x6d9/0x1b20 [i915]
[  100.336656]  i915_gem_execbuffer2+0xc0/0x250 [i915]
[  100.336666]  drm_ioctl+0x200/0x450
[  100.336697]  ? i915_gem_execbuffer+0x330/0x330 [i915]
[  100.336708]  do_vfs_ioctl+0x90/0x6e0
[  100.336716]  ? up_read+0x1a/0x40
[  100.336723]  ? trace_hardirqs_on_caller+0x122/0x1b0
[  100.336730]  SyS_ioctl+0x3c/0x70
[  100.336738]  entry_SYSCALL_64_fastpath+0x1c/0xb1
[  100.336745] RIP: 0033:0x7f0d187cb357
[  100.336750] RSP: 002b:00007ffe0b2f7c28 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
[  100.336761] RAX: ffffffffffffffda RBX: 00007ffe0b2f7d60 RCX: 00007f0d187cb357
[  100.336768] RDX: 00007ffe0b2f7d00 RSI: 0000000040406469 RDI: 0000000000000003
[  100.336775] RBP: 0000000000000000 R08: 0000000000000000 R09: 0000000000000022
[  100.336782] R10: 0000000000000007 R11: 0000000000000246 R12: 0000000000000002
[  100.336789] R13: 0000000000419101 R14: 00007ffe0b2f7d60 R15: 00007ffe0b2f7d50
[  100.336797] Code: 5f 74 1e e9 d4 fb ff ff e8 bc 1e 9c e0 e9 ae fb ff ff 4c 89 e7 e8 77 22 fd ff e9 88 fd ff ff 0f 0b e8 a3 1e 9c e0 e9 b1 fb ff ff <0f> 0b 0f 0b e8 fd af ab e0 85 c0 75 c2 48 c7 c2 80 2c 71 a0 be
[  100.336877] RIP: __i915_add_request+0x4aa/0x510 [i915] RSP: ffffc90000563ac0
[  100.336886] ---[ end trace 22b36545479e5eb7 ]---

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170111140858.1922-1-chris@chris-wilson.co.uk
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
drivers/gpu/drm/i915/i915_gem_request.c

index 99056b948edabe5d1945573a192727fd56af5454..e614398abe2faf0a818ffd15564909e581aad3bd 100644 (file)
@@ -851,6 +851,13 @@ void __i915_add_request(struct drm_i915_gem_request *request, bool flush_caches)
        lockdep_assert_held(&request->i915->drm.struct_mutex);
        trace_i915_gem_request_add(request);
 
+       /* Make sure that no request gazumped us - if it was allocated after
+        * our i915_gem_request_alloc() and called __i915_add_request() before
+        * us, the timeline will hold its seqno which is later than ours.
+        */
+       GEM_BUG_ON(i915_seqno_passed(timeline->last_submitted_seqno,
+                                    request->fence.seqno));
+
        /*
         * To ensure that this call will not fail, space for its emissions
         * should already have been reserved in the ring buffer. Let the ring