Am 05.03.2018 um 09:08 schrieb Liu, Monk: > To better approach this issue I suggest to do the following: > 1. Revert the original patch. > > 2. Stop waiting to long for writes. E.g. use a separate timeout (20ms > maybe?) to wait for the write. Then do a WARN_ON_ONCE when we timeout. > Cannot do that 20ms is not enough, sometimes you need 10 seconds since other VFs may doing bad things like occupying GPU intentionally or they are doing TDR, so > I don't think separate read and write is good idea, they should be treated equally Well the question is if separating read&writes would actually help. > > 3. To the read function add a "if (!in_intterupt()) may_sleep();" and then retest. That should at least print a nice warning when called from atomic context. > Sorry what is may_sleep() ?? > > 4. Test the whole thing and try to fix all warnings about atomic contexts from the may_sleep(); > > 5. Reapply the original patch, but this time only for the read function, not the write function. > > > From current LKG code, the only one spin lock may wrapping the kiq_rreg/wreg() is the pcie_idx_lock, and this lock is only used during init(), > Since init() is run under the case of exclusive mode for SRIOV, which means: > 1) register access is not go through KIQ (see admgpu_mm_reg) > 2) those functions are only in bif_medium_grain_xxx part (vi.c and nbio_v6.c) , and they won't hit under SRIOV ( we return in the head if SRIOV detect) > So I don' think this spin_lock may cause trouble... Ok in this case let's keep the patch for now, but please provide a new patch which adds "if (!in_intterupt()) may_sleep();" in both the read and write function. This way we should at least catch problems early on. Christian. > > /Monk > > > > -----Original Message----- > From: Christian König [mailto:ckoenig.leichtzumerken at gmail.com] > Sent: 2018å¹´3æ??5æ?¥ 15:57 > To: Liu, Monk <Monk.Liu at amd.com>; Koenig, Christian <Christian.Koenig at amd.com>; Kuehling, Felix <Felix.Kuehling at amd.com>; amd-gfx at lists.freedesktop.org > Subject: Re: [PATCH] drm/amdgpu: try again kiq access if not in IRQ(v2) > >> otherwise I don't see how it is better by reverting it > Well it's better to revert it for now because it seems to create more problems than it solves. > > To better approach this issue I suggest to do the following: > 1. Revert the original patch. > > 2. Stop waiting to long for writes. E.g. use a separate timeout (20ms > maybe?) to wait for the write. Then do a WARN_ON_ONCE when we timeout. > > 3. To the read function add a "if (!in_intterupt()) may_sleep();" and then retest. That should at least print a nice warning when called from atomic context. > > 4. Test the whole thing and try to fix all warnings about atomic contexts from the may_sleep(); > > 5. Reapply the original patch, but this time only for the read function, not the write function. > > Regards, > Christian. > > Am 05.03.2018 um 05:20 schrieb Liu, Monk: >> When there are 16 VF/VM on one GPU, we can easily hit "sys lockup to >> 22s" kernel error/warning introduced by kiq_rreg/wreg routine That's >> why I must use this patch to let thread sleep a while and try again, >> >> If you insist reverting this patch please give me a solution, >> otherwise I don't see how it is better by reverting it >> >> /Monk >> >> -----Original Message----- >> From: Christian König [mailto:ckoenig.leichtzumerken at gmail.com] >> Sent: 2018å¹´3æ??3æ?¥ 21:38 >> To: Kuehling, Felix <Felix.Kuehling at amd.com>; Liu, Monk >> <Monk.Liu at amd.com>; amd-gfx at lists.freedesktop.org >> Subject: Re: [PATCH] drm/amdgpu: try again kiq access if not in >> IRQ(v2) >> >> Am 02.03.2018 um 21:47 schrieb Felix Kuehling: >>> On 2018-03-02 04:29 AM, Liu, Monk wrote: >>>> In_atomic() isnot encouraged to be used to judge if sleep is >>>> possible, see the macros of it >>>> >>>> #define in_atomic() (preept_count() != 0) >>> OK. But my point is still that you're not testing the right thing >>> when you check in_interrupt(). The comment before the in_atomic macro >>> definition states the limitations and says "do not use in driver code". >>> Unfortunately it doesn't suggest any alternative. I think >>> in_interrupt is actually worse, because it misses even more cases than in_atomic. >> Thinking about this, Felix seems to be absolutely right. >> >> So we need to revert this patch since you can't reliable detect in a driver if sleeping is allowed or not. >> >> Regards, >> Christian. >> >>> Regards, >>>  Felix >>> >>>> /Monk >>>> >>>> -----Original Message----- >>>> From: Kuehling, Felix >>>> Sent: 2018å¹´3æ??1æ?¥ 23:50 >>>> To: amd-gfx at lists.freedesktop.org; Liu, Monk <Monk.Liu at amd.com> >>>> Subject: Re: [PATCH] drm/amdgpu: try again kiq access if not in >>>> IRQ(v2) >>>> >>>> On 2018-02-28 02:27 AM, Monk Liu wrote: >>>>> sometimes GPU is switched to other VFs and won't swich back soon, >>>>> so the kiq reg access will not signal within a short period, >>>>> instead of busy waiting a long time(MAX_KEQ_REG_WAIT) and returning >>>>> TMO we can istead sleep 5ms and try again later (non irq context) >>>>> >>>>> And since the waiting in kiq_r/weg is busy wait, so >>>>> MAX_KIQ_REG_WAIT shouldn't set to a long time, set it to 10ms is more appropriate. >>>>> >>>>> if gpu already in reset state, don't retry the KIQ reg access >>>>> otherwise it would always hang because KIQ was already die usually. >>>>> >>>>> v2: >>>>> replace schedule() with msleep() for the wait >>>>> >>>>> Change-Id: I8fc807ce85a8d30d2b50153f3f3a6eda344ef994 >>>>> Signed-off-by: Monk Liu <Monk.Liu at amd.com> >>>>> --- >>>>> drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c | 15 +++++++++++++-- >>>>> 1 file changed, 13 insertions(+), 2 deletions(-) >>>>> >>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c >>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c >>>>> index b832651..1672f5b 100644 >>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c >>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c >>>>> @@ -22,7 +22,7 @@ >>>>> */ >>>>> >>>>> #include "amdgpu.h" >>>>> -#define MAX_KIQ_REG_WAIT 100000000 /* in usecs */ >>>>> +#define MAX_KIQ_REG_WAIT 10000 /* in usecs, 10ms */ >>>>> >>>>> uint64_t amdgpu_csa_vaddr(struct amdgpu_device *adev) { @@ >>>>> -152,9 >>>>> +152,14 @@ uint32_t amdgpu_virt_kiq_rreg(struct amdgpu_device >>>>> +*adev, uint32_t reg) >>>>> amdgpu_ring_commit(ring); >>>>> spin_unlock_irqrestore(&kiq->ring_lock, flags); >>>>> >>>>> +retry_read: >>>>> r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT); >>>>> if (r < 1) { >>>>> DRM_ERROR("wait for kiq fence error: %ld\n", r); >>>>> + if (!in_interrupt() && !adev->in_gpu_reset) { >>>> You should check in_atomic here. Because it's invalid to sleep in atomic context (e.g. while holding a spin lock) even when not in an interrupt. >>>> This seems to happen a lot for indirect register access, e.g. >>>> soc15_pcie_rreg. >>>> >>>> Regards, >>>>  Felix >>>> >>>>> + msleep(5); >>>>> + goto retry_read; >>>>> + } >>>>> return ~0; >>>>> } >>>>> val = adev->wb.wb[adev->virt.reg_val_offs]; >>>>> @@ -179,9 +184,15 @@ void amdgpu_virt_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v) >>>>> amdgpu_ring_commit(ring); >>>>> spin_unlock_irqrestore(&kiq->ring_lock, flags); >>>>> >>>>> +retry_write: >>>>> r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT); >>>>> - if (r < 1) >>>>> + if (r < 1) { >>>>> DRM_ERROR("wait for kiq fence error: %ld\n", r); >>>>> + if (!in_interrupt() && !adev->in_gpu_reset) { >>>>> + msleep(5); >>>>> + goto retry_write; >>>>> + } >>>>> + } >>>>> } >>>>> >>>>> /** >>> _______________________________________________ >>> amd-gfx mailing list >>> amd-gfx at lists.freedesktop.org >>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx >> _______________________________________________ >> amd-gfx mailing list >> amd-gfx at lists.freedesktop.org >> https://lists.freedesktop.org/mailman/listinfo/amd-gfx > _______________________________________________ > amd-gfx mailing list > amd-gfx at lists.freedesktop.org > https://lists.freedesktop.org/mailman/listinfo/amd-gfx
