Christian, Looks like we need more discuss with itâ?¦ Here is your approach: 1. Stop the scheduler from feeding more jobs to the hardware when a jobs completes. //this is where I agree with you 2. Then call hw_job_reset to remove the connection between job and hardware fence. 3. Test if job is now completed. It is entirely possible that the job completed while we started the work to reset the hardware. Removing the connection between the hardware fence and the job is the deadline, if the job completes after that it is lost. 4. Check the karma and kick out the job if we found it guilty. 5. Get the whole stuff working again, e.g. reset the hardware, restart the scheduler etc... [ML]: One thing I agree to change with your way: in gpu_reset() we should first stop the in question ringâ??s scheduler (not the all) before kick out the guilty job. > Indeed, but I still think that this is a bad approach cause we then reset the hardware without a good reason. [ML] yes, but we cannot guarantee the job is 100% really hang when entering gpu_reset(), we can only trust our amdgpu_job_timeout as a deadline for each job. You approach that check the fence first before charge it as guilty/hang is incorrect looks to me because why you not check it twice, triple, and even more loops ? You check it one time and you found it just signaled thatâ??s great and lucky(really luckyâ?¦), But what if it didnâ??t signaled (like most usual case) , why not check it again and again ? do you have a theoretic to support on how much time you need to check before finally consider it hang ? No I donâ??t think you have so please just cut this unnecessary checking, we already use amdgpu_job_timeout to give the deadline of each job. So I refuse to check if @job is just signaled in gpu_reset, because this action is vague (and no one can guarantee the job wonâ??t signal during gpu_reset, we should not argue on this event â?¦), I prefer clean and restrict rules. >Force completion is not so much of the issue, but rather in which order you do things. >See the original code first stops the scheduler and removes the connection between hardware fence and job in an atomic manner. And THEN forces the hardware fence to complete. >This way we could be sure that nothing happened in parallel, e.g. that we don't try to signal the fence twice or something like that. [ML] I donâ??t see my approach will have chance to fence twiceâ?¦ on the contrast I think my approach is more clear: no matter the in question job finally signaled or not, I just kick it out from mirror-list Without remove the callback from hw fence, that way even it really signaled during the gpu_reset() period the logic is still perfect and its sched fence will act like usual â?¦ Please point out where or how my approach will go wrong like â??e.g. that we don't try to signal the fence twice or something like that.â??, otherwise I cannot be persuaded and fix my way â?¦ At last. I run the TDR test and it ends up with Hypervisor side error, the guest side is all perfect, here is what I ran: The vk_example and vulkan CTS test under MCBP case, we have all kinds of hang on compute ring, without TDR this test wonâ??t suffer for 5 seconds, and with TDR although MCBP is buggy now but we can Finish this test (of cause test result is mismatch due to MCBP issue), and there are tongs of job_timed_out in dmesg, but guest driver didnâ??t have any error report. I was also surprised this behave really stable â?¦ The second test the using â??watchâ?? to trigger a gpu hang (bad command stream) every 2 seconds, with amdgpu_job_timeout also set to 2seconds, I can run it till hypervisor side hit VF FLR error, and guest side Still runs perfectly and nothing wrong happened, with hw fence seq number I can tell there are 3000 loops of TDR finished before Hypervisor hit error. BR Monk From: Koenig, Christian Sent: Tuesday, May 09, 2017 8:52 PM To: Liu, Monk <Monk.Liu at amd.com>; Christian König <deathsimple at vodafone.de>; amd-gfx at lists.freedesktop.org Subject: Re: [PATCH 4/4] drm/amdgpu/SRIOV:implement guilty job TDR (V2) [ML] if the job complete, the jobâ??s sched fence callback will take this spin_lock and remove itself from mirror_list, so we are still safe to call amd_sched_job_kickout(), and it will do nothing if so Indeed, but I still think that this is a bad approach cause we then reset the hardware without a good reason. Besides, original logic also force complete the hw fence, and it works well â?¦ Force completion is not so much of the issue, but rather in which order you do things. See the original code first stops the scheduler and removes the connection between hardware fence and job in an atomic manner. And THEN forces the hardware fence to complete. This way we could be sure that nothing happened in parallel, e.g. that we don't try to signal the fence twice or something like that. State like â??You are missing that it is entirely possible that the job will complete while we are trying to kick it out.â?? Sorry I should have been more clear. Is not a good reason to reject my approach, because that is okay if the job just completed â?¦ We usually try to take a defensive approach, so stopping everything, removing the hardware fence connection and then explicitly kicking out the job in question sounds much better than doing it implicitly with the hardware fence completion. Even when this works (which I'm still not sure of) that is a really awkward and hard to understand approach. Regards, Christian. Am 09.05.2017 um 13:58 schrieb Liu, Monk: You are missing that it is entirely possible that the job will complete while we are trying to kick it out. [ML] if the job complete, the jobâ??s sched fence callback will take this spin_lock and remove itself from mirror_list, so we are still safe to call amd_sched_job_kickout(), and it will do nothing if so Please go through the whole steps again, Besides, original logic also force complete the hw fence, and it works well â?¦ I donâ??t see the solid reason why you insist on your approach, please go through the steps again and give me the details about where is incorrect than I can fix it State like â??You are missing that it is entirely possible that the job will complete while we are trying to kick it out.â?? Is not a good reason to reject my approach, because that is okay if the job just completed â?¦ BR Monk From: Koenig, Christian Sent: Tuesday, May 09, 2017 3:49 PM To: Liu, Monk <Monk.Liu at amd.com><mailto:Monk.Liu at amd.com>; Christian König <deathsimple at vodafone.de><mailto:deathsimple at vodafone.de>; amd-gfx at lists.freedesktop.org<mailto:amd-gfx at lists.freedesktop.org> Subject: Re: [PATCH 4/4] drm/amdgpu/SRIOV:implement guilty job TDR (V2) [ML] Really not necessary, we have spin_lock to protect the mirror-list, nothing will be messed up ... You are missing that it is entirely possible that the job will complete while we are trying to kick it out. [ML] why don't touch hardware fence at all ? the original/bare-metal gpu reset also signal all ring's hardware fence first, I just follow the original logic ... Scheduler fence will be auto signaled after hw fence signaled, any problem with that ? what's the concern ? The hardware runs async to the CPU which tries to reset it, so we need to be careful in which order things are done. [ML] No I don't think so, the kickout must be prior to the hw_job_reset, otherwise the scheduler fence callback will be removed and you need manually install it later , which is not correct: For the guity job, we just kick it out before job reset, in job_reset we only reset other innocent jobs( and unbind the scheduler fence callback for them), after hw fence Forcely set to drv_seq, all hw fence are signaled (this is the way of original logic, I didn't change that). When go to sched_recovery(), it will recover all innocent job and hook The scheduler fence with new hw fence. That way only the guilty job is dropped forever. Again same problem here. To be absolutely sure that everything works as expected we need to do it in the following order: 1. Stop the scheduler from feeding more jobs to the hardware when a jobs completes. 2. Then call hw_job_reset to remove the connection between job and hardware fence. 3. Test if job is now completed. It is entirely possible that the job completed while we started the work to reset the hardware. Removing the connection between the hardware fence and the job is the deadline, if the job completes after that it is lost. 4. Check the karma and kick out the job if we found it guilty. 5. Get the whole stuff working again, e.g. reset the hardware, restart the scheduler etc... Regards, Christian. Am 09.05.2017 um 04:45 schrieb Liu, Monk: > > - /* block scheduler */ > - for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { > - ring = adev->rings[i]; > + /* we start from the ring trigger GPU hang */ > + j = job ? job->ring->idx : 0; > + > + if (job) > + if (amd_sched_invalidate_job(&job->base, amdgpu_job_hang_limit)) > + amd_sched_job_kickout(&job->base); Well that looks like the wrong order to me. We should probably stop the scheduler before trying to mess anything with the job. [ML] Really not necessary, we have spin_lock to protect the mirror-list, nothing will be messed up ... > > +void amdgpu_fence_driver_force_completion_ring(struct amdgpu_ring > +*ring) { > + if (ring) > + amdgpu_fence_write(ring, ring->fence_drv.sync_seq); } > + The coding style is completely off. [ML] I don't know why at email side it looks wrong coding style, but I'm sure it is correct at my side, check here: void amdgpu_fence_driver_force_completion_ring(struct amdgpu_ring *ring) { if (ring) amdgpu_fence_write(ring, ring->fence_drv.sync_seq); } Additional to that I don't think that this is a good idea. We should probably rather just signal all scheduler fences instead and don't touch the hardware fence at all. [ML] why don't touch hardware fence at all ? the original/bare-metal gpu reset also signal all ring's hardware fence first, I just follow the original logic ... Scheduler fence will be auto signaled after hw fence signaled, any problem with that ? what's the concern ? > - for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { > - struct amdgpu_ring *ring = adev->rings[i]; > + for (i = j; i < j + AMDGPU_MAX_RINGS; ++i) { > + ring = adev->rings[i % AMDGPU_MAX_RINGS]; > if (!ring || !ring->sched.thread) > continue; > > + if (job && j != i) { > + kthread_unpark(ring->sched.thread); > + continue; > + } > + Please split up that patch a bit further. E.g. first the handling to only hw_job_reset the ring in question, then the kickout handling. [ML] No I don't think so, the kickout must be prior to the hw_job_reset, otherwise the scheduler fence callback will be removed and you need manually install it later , which is not correct: For the guity job, we just kick it out before job reset, in job_reset we only reset other innocent jobs( and unbind the scheduler fence callback for them), after hw fence Forcely set to drv_seq, all hw fence are signaled (this is the way of original logic, I didn't change that). When go to sched_recovery(), it will recover all innocent job and hook The scheduler fence with new hw fence. That way only the guilty job is dropped forever. -----Original Message----- From: Christian König [mailto:deathsimple@xxxxxxxxxxx] Sent: Monday, May 08, 2017 9:12 PM To: Liu, Monk <Monk.Liu at amd.com><mailto:Monk.Liu at amd.com>; amd-gfx at lists.freedesktop.org<mailto:amd-gfx at lists.freedesktop.org> Cc: Koenig, Christian <Christian.Koenig at amd.com><mailto:Christian.Koenig at amd.com> Subject: Re: [PATCH 4/4] drm/amdgpu/SRIOV:implement guilty job TDR (V2) Am 08.05.2017 um 09:01 schrieb Liu, Monk: > @Christian > > This one is changed to guilty job scheme accordingly with your > response > > BR Monk > > -----Original Message----- > From: Monk Liu [mailto:Monk.Liu at amd.com] > Sent: Monday, May 08, 2017 3:00 PM > To: amd-gfx at lists.freedesktop.org<mailto:amd-gfx at lists.freedesktop.org> > Cc: Liu, Monk <Monk.Liu at amd.com><mailto:Monk.Liu at amd.com> > Subject: [PATCH] drm/amdgpu/SRIOV:implement guilty job TDR (V2) > > 1,TDR will kickout guilty job if it hang exceed the threshold of the given one from kernel paramter "job_hang_limit", that way a bad command stream will not infinitly cause GPU hang. > > by default this threshold is 1 so a job will be kicked out after it hang. > > 2,if a job timeout TDR routine will not reset all sched/ring, instead if will only reset on the givn one which is indicated by @job of amdgpu_sriov_gpu_reset, that way we don't need to reset and recover each sched/ring if we already know which job cause GPU hang. > > 3,unblock sriov_gpu_reset for AI family. > 4,don't init entity for KIQ > > TODO: > when a job is considered as guilty, we should mark some flag in its fence status flag, and let UMD side aware that this fence signaling is not due to job complete but job hang. > > Change-Id: I7b89c19a3de93249db570d0a80522176b1525a09 > Signed-off-by: Monk Liu <Monk.Liu at amd.com><mailto:Monk.Liu at amd.com> > --- > drivers/gpu/drm/amd/amdgpu/amdgpu.h | 1 + > drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c | 4 +++ > drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 36 ++++++++++++++++++++------- > drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 4 +++ > drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 6 +++++ > drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 1 + > drivers/gpu/drm/amd/scheduler/gpu_scheduler.c | 11 +++++++- drivers/gpu/drm/amd/scheduler/gpu_scheduler.h | 7 ++++++ > 8 files changed, 60 insertions(+), 10 deletions(-) > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h > b/drivers/gpu/drm/amd/amdgpu/amdgpu.h > index 90a69bf..93bcea2 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h > @@ -111,6 +111,7 @@ extern int amdgpu_prim_buf_per_se; extern int > amdgpu_pos_buf_per_se; extern int amdgpu_cntl_sb_buf_per_se; extern > int amdgpu_param_buf_per_se; > +extern int amdgpu_job_hang_limit; > > #define AMDGPU_DEFAULT_GTT_SIZE_MB 3072ULL /* 3GB by default */ > #define AMDGPU_WAIT_IDLE_TIMEOUT_IN_MS 3000 > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c > b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c > index b4bbbb3..23afc58 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c > @@ -52,6 +52,10 @@ static int amdgpu_ctx_init(struct amdgpu_device *adev, struct amdgpu_ctx *ctx) > struct amd_sched_rq *rq; > > rq = &ring->sched.sched_rq[AMD_SCHED_PRIORITY_NORMAL]; > + > + if (ring == &adev->gfx.kiq.ring) > + continue; > + That looks like a bug fix and should probably go into a separate patch. > r = amd_sched_entity_init(&ring->sched, &ctx->rings[i].entity, > rq, amdgpu_sched_jobs); > if (r) > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > index 0e5f314..f3990fe 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > @@ -2537,7 +2537,7 @@ static int amdgpu_recover_vram_from_shadow(struct amdgpu_device *adev, > */ > int amdgpu_sriov_gpu_reset(struct amdgpu_device *adev, struct amdgpu_job *job) { > - int i, r = 0; > + int i, j, r = 0; > int resched; > struct amdgpu_bo *bo, *tmp; > struct amdgpu_ring *ring; > @@ -2550,19 +2550,30 @@ int amdgpu_sriov_gpu_reset(struct amdgpu_device *adev, struct amdgpu_job *job) > /* block TTM */ > resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev); > > - /* block scheduler */ > - for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { > - ring = adev->rings[i]; > + /* we start from the ring trigger GPU hang */ > + j = job ? job->ring->idx : 0; > + > + if (job) > + if (amd_sched_invalidate_job(&job->base, amdgpu_job_hang_limit)) > + amd_sched_job_kickout(&job->base); Well that looks like the wrong order to me. We should probably stop the scheduler before trying to mess anything with the job. > > + /* block scheduler */ > + for (i = j; i < j + AMDGPU_MAX_RINGS; ++i) { > + ring = adev->rings[i % AMDGPU_MAX_RINGS]; > if (!ring || !ring->sched.thread) > continue; > > kthread_park(ring->sched.thread); > + > + if (job && j != i) > + continue; > + > + /* only do job_reset on the hang ring if @job not NULL */ > amd_sched_hw_job_reset(&ring->sched); > - } > > - /* after all hw jobs are reset, hw fence is meaningless, so force_completion */ > - amdgpu_fence_driver_force_completion(adev); > + /* after all hw jobs are reset, hw fence is meaningless, so force_completion */ > + amdgpu_fence_driver_force_completion_ring(ring); > + } > > /* request to take full control of GPU before re-initialization */ > if (job) > @@ -2615,11 +2626,16 @@ int amdgpu_sriov_gpu_reset(struct amdgpu_device *adev, struct amdgpu_job *job) > } > fence_put(fence); > > - for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { > - struct amdgpu_ring *ring = adev->rings[i]; > + for (i = j; i < j + AMDGPU_MAX_RINGS; ++i) { > + ring = adev->rings[i % AMDGPU_MAX_RINGS]; > if (!ring || !ring->sched.thread) > continue; > > + if (job && j != i) { > + kthread_unpark(ring->sched.thread); > + continue; > + } > + Please split up that patch a bit further. E.g. first the handling to only hw_job_reset the ring in question, then the kickout handling. > amd_sched_job_recovery(&ring->sched); > kthread_unpark(ring->sched.thread); > } > @@ -2629,6 +2645,8 @@ int amdgpu_sriov_gpu_reset(struct amdgpu_device *adev, struct amdgpu_job *job) > if (r) { > /* bad news, how to tell it to userspace ? */ > dev_info(adev->dev, "GPU reset failed\n"); > + } else { > + dev_info(adev->dev, "GPU reset successed!\n"); > } > > adev->gfx.in_reset = false; > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c > b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c > index 416908a..fd3691a8 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c > @@ -112,6 +112,7 @@ int amdgpu_prim_buf_per_se = 0; int > amdgpu_pos_buf_per_se = 0; int amdgpu_cntl_sb_buf_per_se = 0; int > amdgpu_param_buf_per_se = 0; > +int amdgpu_job_hang_limit = 0; > > MODULE_PARM_DESC(vramlimit, "Restrict VRAM for testing, in > megabytes"); module_param_named(vramlimit, amdgpu_vram_limit, int, > 0600); @@ -237,6 +238,9 @@ module_param_named(cntl_sb_buf_per_se, > amdgpu_cntl_sb_buf_per_se, int, 0444); > MODULE_PARM_DESC(param_buf_per_se, "the size of Off-Chip Pramater > Cache per Shader Engine (default depending on gfx)"); > module_param_named(param_buf_per_se, amdgpu_param_buf_per_se, int, > 0444); > > +MODULE_PARM_DESC(job_hang_limit, "how much time allow a job hang and > +not drop it (default 0)"); module_param_named(job_hang_limit, > +amdgpu_job_hang_limit, int ,0444); > + > > static const struct pci_device_id pciidlist[] = { #ifdef > CONFIG_DRM_AMDGPU_SI diff --git > a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c > b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c > index d7523d1..8de3bd3 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c > @@ -541,6 +541,12 @@ void amdgpu_fence_driver_force_completion(struct amdgpu_device *adev) > } > } > > +void amdgpu_fence_driver_force_completion_ring(struct amdgpu_ring > +*ring) { > + if (ring) > + amdgpu_fence_write(ring, ring->fence_drv.sync_seq); } > + The coding style is completely off. Additional to that I don't think that this is a good idea. We should probably rather just signal all scheduler fences instead and don't touch the hardware fence at all. > /* > * Common fence implementation > */ > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h > b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h > index 981ef08..03e88c6 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h > @@ -76,6 +76,7 @@ struct amdgpu_fence_driver { int > amdgpu_fence_driver_init(struct amdgpu_device *adev); void > amdgpu_fence_driver_fini(struct amdgpu_device *adev); void > amdgpu_fence_driver_force_completion(struct amdgpu_device *adev); > +void amdgpu_fence_driver_force_completion_ring(struct amdgpu_ring > +*ring); > > int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring, > unsigned num_hw_submission); > diff --git a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c > b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c > index 6f4e31f..4e97e6d 100644 > --- a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c > +++ b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c > @@ -390,9 +390,18 @@ void amd_sched_hw_job_reset(struct amd_gpu_scheduler *sched) > &s_job->s_fence->cb)) { > fence_put(s_job->s_fence->parent); > s_job->s_fence->parent = NULL; > + atomic_dec(&sched->hw_rq_count); > } > } > - atomic_set(&sched->hw_rq_count, 0); > + spin_unlock(&sched->job_list_lock); > +} > + > +void amd_sched_job_kickout(struct amd_sched_job *s_job) { > + struct amd_gpu_scheduler *sched = s_job->sched; > + > + spin_lock(&sched->job_list_lock); > + list_del_init(&s_job->node); > spin_unlock(&sched->job_list_lock); > } > > diff --git a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h > b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h > index 8cb41d3..59694f3 100644 > --- a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h > +++ b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h > @@ -81,6 +81,7 @@ struct amd_sched_job { > struct list_head node; > struct delayed_work work_tdr; > uint64_t id; > + atomic_t karma; > }; > > extern const struct fence_ops amd_sched_fence_ops_scheduled; @@ -96,6 +97,11 @@ static inline struct amd_sched_fence *to_amd_sched_fence(struct fence *f) > return NULL; > } > > +static inline bool amd_sched_invalidate_job(struct amd_sched_job > +*s_job, int threshold) { > + return (s_job && atomic_inc_return(&s_job->karma) > threshold); } > + Again coding style is completely off. Christian. > /** > * Define the backend operations called by the scheduler, > * these functions should be implemented in driver side @@ -158,4 +164,5 @@ int amd_sched_job_init(struct amd_sched_job *job, > void *owner); > void amd_sched_hw_job_reset(struct amd_gpu_scheduler *sched); void > amd_sched_job_recovery(struct amd_gpu_scheduler *sched); > +void amd_sched_job_kickout(struct amd_sched_job *s_job); > #endif > -- > 2.7.4 > > _______________________________________________ > amd-gfx mailing list > amd-gfx at lists.freedesktop.org<mailto:amd-gfx at lists.freedesktop.org> > https://lists.freedesktop.org/mailman/listinfo/amd-gfx -------------- next part -------------- An HTML attachment was scrubbed... URL: <https://lists.freedesktop.org/archives/amd-gfx/attachments/20170510/208fc0b4/attachment-0001.html>
