GPU lockup+blackout on 2.6.35 with r600...
Daniel J Blueman
daniel.blueman at gmail.com
Tue Aug 17 15:12:31 PDT 2010
Hi chaps,
On 9 August 2010 14:41, Daniel J Blueman <daniel.blueman at gmail.com> wrote:
> On 8 August 2010 17:36, Daniel J Blueman <daniel.blueman at gmail.com> wrote:
>> On 5 August 2010 16:23, Alex Deucher <alexdeucher at gmail.com> wrote:
>>> On Thu, Aug 5, 2010 at 9:50 AM, Jerome Glisse <glisse at freedesktop.org> wrote:
>>>> On 08/05/2010 06:52 AM, Daniel J Blueman wrote:
>>>>>
>>>>> After around 10-30 mins of typical usage (ie cairo rendering to
>>>>> surfaces + compiz) on my Dell Studio 15 (model 1557), my R600 radeon
>>>>> locks up, giving a blank screen.
>>>>>
>>>>> The only information I have is a backtrace [1]. 'radeontool regs'
>>>>> doesn't give anything - what information, /sys files etc would be
>>>>> useful here, and would it be useful to log a freedesktop.org bug
>>>>> report at this stage, assuming I haven't found a reliable reproducer?
>>>>>
>>>>> OS is Ubuntu 10.10 development release, mesa 7.8.2, libdrm2 2.4.21.
>>>>>
>>>>> Thanks,
>>>>> Daniel
>>>>>
>>>>> --- [1]
>>>>>
>>>>> radeon 0000:02:00.0: GPU lockup CP stall for more than 1000msec
>>>>> ------------[ cut here ]------------
>>>>> WARNING: at
>>>>> /home/kernel-ppa/COD/linux/drivers/gpu/drm/radeon/radeon_fence.c:235
>>>>> radeon_fence_wait+0x2b7/0x320 [radeon]()
>>>>> Hardware name: Studio 1557
>>>>> GPU lockup (waiting for 0x00008C35 last fence id 0x00008C33)
>>>>> Modules linked in: binfmt_misc parport_pc ppdev kvm_intel kvm
>>>>> microcode snd_hda_codec_atihdmi joydev ipt_REJECT xt_limit xt_tcpudp
>>>>> ipt_addrtype xt_state snd_hda_codec_idt ip6table_filter ip6_tables
>>>>> snd_hda_intel nf_nat_irc nf_conntrack_irc snd_hda_codec nf_nat_ftp
>>>>> nf_nat snd_hwdep nf_conntrack_ipv4 snd_seq_midi arc4 snd_pcm
>>>>> snd_rawmidi nf_defrag_ipv4 radeon nf_conntrack_ftp nf_conntrack
>>>>> iptable_filter snd_seq_midi_event snd_seq ip_tables ttm iwlagn video
>>>>> snd_timer uvcvideo x_tables snd_seq_device output drm_kms_helper lp
>>>>> dell_laptop iwlcore drm videodev dcdbas dell_wmi v4l1_compat mac80211
>>>>> i7core_edac parport v4l2_compat_ioctl32 snd psmouse edac_core
>>>>> i2c_algo_bit soundcore cfg80211 snd_page_alloc serio_raw sdhci_pci
>>>>> sdhci ahci libahci led_class r8169 mii btrfs zlib_deflate crc32c
>>>>> libcrc32c
>>>>> Pid: 1624, comm: Xorg Tainted: G D 2.6.35-020635-generic #020635
>>>>> Call Trace:
>>>>> ? radeon_fence_wait+0x2b7/0x320 [radeon]
>>>>> warn_slowpath_common+0x90/0xc0
>>>>> warn_slowpath_fmt+0x6e/0x70
>>>>> ? schedule_timeout+0x15a/0x2e0
>>>>> ? r600_irq_set+0x27d/0xc00 [radeon]
>>>>> ? radeon_ring_commit+0xa3/0xb0 [radeon]
>>>>> ? r100_gpu_cp_is_lockup+0xc2/0xd0 [radeon]
>>>>> ? r600_gpu_is_lockup+0x1cb/0x220 [radeon]
>>>>> radeon_fence_wait+0x2b7/0x320 [radeon]
>>>>> ? autoremove_wake_function+0x0/0x40
>>>>> radeon_sync_obj_wait+0x11/0x20 [radeon]
>>>>> ttm_bo_wait+0x102/0x1b0 [ttm]
>>>>> ttm_bo_move_accel_cleanup+0x19e/0x230 [ttm]
>>>>> radeon_move_blit+0x124/0x170 [radeon]
>>>>> radeon_bo_move+0xda/0x1a0 [radeon]
>>>>> ttm_bo_handle_move_mem+0xec/0x370 [ttm]
>>>>> ttm_bo_evict+0x1cc/0x270 [ttm]
>>>>> ? drm_mm_split_at_start+0x1d/0x80 [drm]
>>>>> ttm_mem_evict_first+0xed/0x180 [ttm]
>>>>> ? ttm_bo_man_get_node+0xd5/0xe0 [ttm]
>>>>> ttm_bo_mem_force_space+0xab/0x110 [ttm]
>>>>> ttm_bo_mem_space+0x305/0x370 [ttm]
>>>>> ttm_bo_move_buffer+0xcf/0x140 [ttm]
>>>>> ? drm_mm_split_at_start+0x1d/0x80 [drm]
>>>>> ttm_bo_validate+0xd5/0x100 [ttm]
>>>>> ttm_bo_init+0x1f2/0x240 [ttm]
>>>>> radeon_bo_create+0x121/0x240 [radeon]
>>>>> ? radeon_ttm_bo_destroy+0x0/0x80 [radeon]
>>>>> radeon_gem_object_create+0x89/0x100 [radeon]
>>>>> ? drm_gem_handle_create+0x8b/0xa0 [drm]
>>>>> radeon_gem_create_ioctl+0x58/0xe0 [radeon]
>>>>> drm_ioctl+0x283/0x460 [drm]
>>>>> ? radeon_gem_create_ioctl+0x0/0xe0 [radeon]
>>>>> ? do_readv_writev+0x16b/0x1e0
>>>>> vfs_ioctl+0x3a/0xc0
>>>>> do_vfs_ioctl+0x6d/0x1f0
>>>>> sys_ioctl+0x87/0xa0
>>>>> ? sys_writev+0x55/0xb0
>>>>> system_call_fastpath+0x16/0x1b
>>>>> ---[ end trace c0bc12025fa8386c ]---
>>>>> radeon 0000:02:00.0: GPU softreset
>>>>> radeon 0000:02:00.0: R_008010_GRBM_STATUS=0xA0003028
>>>>> radeon 0000:02:00.0: R_008014_GRBM_STATUS2=0x00000002
>>>>> radeon 0000:02:00.0: R_000E50_SRBM_STATUS=0x200000C0
>>>>> radeon 0000:02:00.0: R_008020_GRBM_SOFT_RESET=0x00007FEE
>>>>> radeon 0000:02:00.0: R_008020_GRBM_SOFT_RESET=0x00000001
>>>>> radeon 0000:02:00.0: R_008010_GRBM_STATUS=0x00003028
>>>>> radeon 0000:02:00.0: R_008014_GRBM_STATUS2=0x00000002
>>>>> radeon 0000:02:00.0: R_000E50_SRBM_STATUS=0x200000C0
>>>>> radeon 0000:02:00.0: GPU reset succeed
>>>>> Clocks initialized !
>>>>> ring test succeeded in 1 usecs
>>>>> ib test succeeded in 1 usecs
>>>>
>>>> SO it keep reseting ? According to log there is a GPU lockup but then
>>>> a successfull GPU reset so GPU should resume fine (that's what log says)
>>>> Best is to open a bug and attach full dmesg + lspci -vv and context
>>>> in which the lockup happen
>>>
>>> Perhaps the lockup timeout is too short or interrupts are being
>>> delivered late or something? There's a fdo or kernel bugzilla entry
>>> where removing the lockup timeout fixed the issue. Does something
>>> like this:
>>>
>>> --- a/drivers/gpu/drm/radeon/radeon_fence.c
>>> +++ b/drivers/gpu/drm/radeon/radeon_fence.c
>>> @@ -237,9 +237,11 @@ retry:
>>> * as signaled for now
>>> */
>>> rdev->gpu_lockup = true;
>>> +#if 0
>>> r = radeon_gpu_reset(rdev);
>>> if (r)
>>> return r;
>>> +#endif
>>> WREG32(rdev->fence_drv.scratch_reg, fence->seq);
>>> rdev->gpu_lockup = false;
>>> }
>>>
>>> fix the issue?
>>
>> Correction, RV710.
>>
>> Well, RADEON_FENCE_JIFFIES_TIMEOUT plus 1000ms is already
>> excruciatingly long for the command processor to not have responded,
>> but the issue I'm really hitting here is that X if left hanging after
>> GPU reset. The GPU lockup was being reported at:
>>
>> [<ffffffffa03b66d7>] ? radeon_fence_wait+0x2b7/0x320 [radeon]
>> [<ffffffff81061370>] warn_slowpath_common+0x90/0xc0
>> [<ffffffff8106149e>] warn_slowpath_fmt+0x6e/0x70
>> [<ffffffff81575f8a>] ? schedule_timeout+0x15a/0x2e0
>> [<ffffffffa03f5a9d>] ? r600_irq_set+0x27d/0xc00 [radeon]
>> [<ffffffffa03ce1f3>] ? radeon_ring_commit+0xa3/0xb0 [radeon]
>> [<ffffffffa03d5412>] ? r100_gpu_cp_is_lockup+0xc2/0xd0 [radeon]
>> [<ffffffffa040063b>] ? r600_gpu_is_lockup+0x1cb/0x220 [radeon]
>> [<ffffffffa03b66d7>] radeon_fence_wait+0x2b7/0x320 [radeon]
>> [<ffffffff8107fcd0>] ? autoremove_wake_function+0x0/0x40
>> [<ffffffffa03b74a1>] radeon_sync_obj_wait+0x11/0x20 [radeon]
>> [<ffffffffa02ff332>] ttm_bo_wait+0x102/0x1b0 [ttm]
>> [<ffffffffa0302ede>] ttm_bo_move_accel_cleanup+0x19e/0x230 [ttm]
>> [<ffffffffa03b75f4>] radeon_move_blit+0x124/0x170 [radeon]
>>
>> X consistently hangs after GPU reset at (with sysrq-T):
>>
>> [<ffffffff81576838>] __mutex_lock_slowpath+0xe8/0x170
>> [<ffffffff8157647b>] mutex_lock+0x2b/0x50
>> [<ffffffffa0415aba>] radeon_pm_set_clocks+0x3a/0x2d0 [radeon]
>> [<ffffffff812abc51>] ? idr_get_new_above+0x11/0x40
>> [<ffffffffa01679ab>] ? drm_mode_object_get+0x6b/0xc0 [drm]
>> [<ffffffff811410ae>] ? kmem_cache_alloc_notrace+0x8e/0xd0
>> [<ffffffffa041662d>] radeon_pm_compute_clocks+0xcd/0x240 [radeon]
>> [<ffffffffa03a9d21>] atombios_crtc_mode_fixup+0x31/0x60 [radeon]
>> [<ffffffffa02ba13a>] drm_crtc_helper_set_mode+0x16a/0x370 [drm_kms_helper]
>> [<ffffffffa02ba3bd>] drm_helper_resume_force_mode+0x7d/0x150 [drm_kms_helper]
>> [<ffffffffa03a2f35>] radeon_gpu_reset+0x85/0xd0 [radeon]
>> [<ffffffffa03b66e7>] radeon_fence_wait+0x2c7/0x320 [radeon]
>> [<ffffffff8107fcd0>] ? autoremove_wake_function+0x0/0x40
>> [<ffffffffa03b74a1>] radeon_sync_obj_wait+0x11/0x20 [radeon]
>> [<ffffffffa02ff332>] ttm_bo_wait+0x102/0x1b0 [ttm]
>> [<ffffffffa0302ede>] ttm_bo_move_accel_cleanup+0x19e/0x230 [ttm]
>> [<ffffffffa03b75f4>] radeon_move_blit+0x124/0x170 [radeon]
>>
>> I couldn't reproduce it (but did hit it a number of times over the
>> last two weeks), so have been analysing with what we have. The kernel
>> the backtrace is against isn't built with debug, so:
>>
>> radeon_pm_set_clocks():
>> 8ca80: 55 push %rbp
>> 8ca81: 48 89 e5 mov %rsp,%rbp
>> 8ca84: 41 57 push %r15
>> 8ca86: 41 56 push %r14
>> 8ca88: 41 55 push %r13
>> 8ca8a: 41 54 push %r12
>> 8ca8c: 53 push %rbx
>> 8ca8d: 48 83 ec 48 sub $0x48,%rsp
>> 8ca91: e8 00 00 00 00 callq 8ca96 <radeon_pm_set_clocks+0x16>
>> 8ca96: 49 89 fc mov %rdi,%r12
>> 8ca99: 48 8b 7f 08 mov 0x8(%rdi),%rdi
>> 8ca9d: 48 83 c7 20 add $0x20,%rdi
>> 8caa1: e8 00 00 00 00 callq 8caa6 <radeon_pm_set_clocks+0x26>
>> 8caa6: 49 8d 84 24 a0 1c 00 lea 0x1ca0(%r12),%rax
>> 8caad: 00
>> 8caae: 48 89 c7 mov %rax,%rdi
>> 8cab1: 48 89 45 90 mov %rax,-0x70(%rbp)
>> 8cab5: e8 00 00 00 00 callq 8caba <radeon_pm_set_clocks+0x3a>
>> -> 8caba: 49 8d 8c 24 e8 09 00 lea 0x9e8(%r12),%rcx
>> 8cac1: 00
>> 8cac2: 48 89 cf mov %rcx,%rdi
>> 8cac5: 48 89 4d 98 mov %rcx,-0x68(%rbp)
>> 8cac9: e8 00 00 00 00 callq 8cace <radeon_pm_set_clocks+0x4e>
>> 8cace: 41 83 7c 24 78 1a cmpl $0x1a,0x78(%r12)
>> 8cad4: 0f 86 46 01 00 00 jbe 8cc20
>> <radeon_pm_set_clocks+0x1a0>
>> 8cada: 41 80 bc 24 d0 0d 00 cmpb $0x0,0xdd0(%r12)
>> 8cae1: 00 00
>>
>> This corresponds to:
>>
>> static void radeon_pm_set_clocks(struct radeon_device *rdev)
>> {
>> int i;
>>
>> mutex_lock(&rdev->ddev->struct_mutex);
>> mutex_lock(&rdev->vram_mutex);
>> -> mutex_lock(&rdev->cp.mutex);
>>
>> We know r12 is the rdev struct pointer (%rdi is the first function
>> argument gets loaded into r12), and 0x9e8 is plausable as the offset
>> of cp.mutex, as vram_mutex is further out and plausable as 0x1ca0, so
>> the code is trying to acquire cp.mutex it seems.
>>
>> From the kernel log, the command proc ring test completed after the
>> start of the GPU reset, which is the last callsite to unlock cp.mutex,
>> ie <radeon_fence_wait->radeon_gpu_reset->radeon_resume->rv770_resume->rv770_startup->r600_cp_resume->radeon_ring_test->r600_ring_test->radeon_ring_unlock_commit>
>> before the hang at
>> <radeon_fence_wait->radeon_gpu_reset->->drm_helper_resume_force_mode->drm_crtc_helper_set_mode->mode_fixup->atombios_crtc_mode_fixup->radeon_pm_compute_clocks->radeon_pm_set_clocks>.
>>
>> Looking through all the calls between these two spots, I just can't
>> see anything that _could_ have locked cp.mutex, unless I'm misreading
>> the disassembly of mutex_lock.
>>
>> Any thoughts?
>
> Well, looking further back down the call chain, radeon_bo_create takes
> the vram_mutex across the ttm_bo_init call, so I was duped by the
> compiler's instruction scheduling.
>
> I looked at various changes to the locking scheme which did address
> the X hang (thus blackouts), but all attempts failed with lockdep
> violations somewhere, so there is no minimal fix without revisiting
> recent locking scheme changes [1,2,3].
>
> I'll pass on some other minor fixes I've cooked up in the meantime.
> Note that I am running with the kernel boot options 'dynpm=1
> radeon.dynclks=1' to stop everything getting toasty; let me know if
> you need more info for this issue.
>
> Thanks,
> Daniel
>
> [1] http://kerneltrap.org/mailarchive/git-commits-head/2010/5/21/37917
> [2] http://kerneltrap.org/mailarchive/git-commits-head/2010/5/21/37923
> [3] http://kerneltrap.org/mailarchive/git-commits-head/2010/5/21/37937
I'm hitting this issue again on 2.6.36-rc1; it reproduces the 1000ms
timeout and subsequent hang every time on my RV710 when starting a
game on doom3 (eg
http://ftp.jeuxlinux.fr/files/doom3-linux-1.1.1286-demo.x86.run), even
without the dympm/dynclks options.
It looks like the locking strategy of the Matthew's patches needs
reworking. I was able to reproduce the hang by forcing a GPU reset
also.
What else would help with this?
Thanks,
Daniel
--
Daniel J Blueman
More information about the dri-devel
mailing list