Re: [PATCH] drm/amdgpu: capture invalid hardware access v2

[Christian König <ckoenig.leichtzumerken@xxxxxxxxx>]

> Are there chances to let driver complete scheduler fences?

Nope, see the documentation Daniel linked in this thread.

If we would want to complete the scheduler fences and release the memory 
protected by them we would need to completely disable the device.

E.g. similar to a hotplug event when you rip out the hardware and put it 
back in again.

That would mean that applications not only need to reinitialize their 
contexts, but also completely restart and reallocate their memory.

Thinking more about it when we go down that route we wouldn't even need 
the reset lock any more in the first place :)

Something else you should keep in mind is that we already have tried 
this approach two times now and both times came to the same conclusion 
after spending a lot of work trying to fix this up.

Christian.

Am 10.03.21 um 14:19 schrieb Li, Dennis:
> [AMD Official Use Only - Internal Distribution Only]
>
> Hi, Christian,
>
> > > We can only complete the low level hardware fences, but not the scheduler fences since those need to wait for the resubmission of the jobs and with it finishing the GPU reset.
> Correct. I run a deadlock test and captured this case today.  Are there chances to let driver complete scheduler fences? If decide to do GPU reset, the pending rendering job in fact should be recreated because of vram lost. Currently, user application will recreate context when if driver report VRAM lost.
>
> -----Original Message-----
> From: Christian König <ckoenig.leichtzumerken@xxxxxxxxx>
> Sent: Wednesday, March 10, 2021 9:03 PM
> To: Li, Dennis <Dennis.Li@xxxxxxx>; Grodzovsky, Andrey <Andrey.Grodzovsky@xxxxxxx>; amd-gfx@xxxxxxxxxxxxxxxxxxxxx
> Subject: Re: [PATCH] drm/amdgpu: capture invalid hardware access v2
>
> > I assumed that pre-reset functions could complete all fences
> And exactly that's incorrect.
>
> We can only complete the low level hardware fences, but not the scheduler fences since those need to wait for the resubmission of the jobs and with it finishing the GPU reset.
>
> Regards,
> Christian.
>
> Am 10.03.21 um 07:40 schrieb Li, Dennis:
> > [AMD Official Use Only - Internal Distribution Only]
> >
> > > > > But how this will help if TDR thread will start after you both took read lock and checked that adev->in_gpu_reset is false ? Since TDR  now takes write lock only after suspending HW and waiting for all fences there is nothing that prevents both threads (e.g IOCTL and TDR) to access registers concurrently.
> > When read thread both took read lock and checked that adev->in_gpu_reset is false, it means that recovery thread still not start, and read thread will continue, it will exit normally, or exit because of fence completed by pre-reset functions. I assumed that pre-reset functions could complete all fences, which could make related threads exit and release read locks. About this, Christian has different thinking, I am trying to understand his concern.  TDR will be blocked until all read locks are released.
> >
> > Best Regards
> > Dennis Li
> > -----Original Message-----
> > From: Grodzovsky, Andrey <Andrey.Grodzovsky@xxxxxxx>
> > Sent: Wednesday, March 10, 2021 1:42 PM
> > To: Li, Dennis <Dennis.Li@xxxxxxx>; Christian König
> > <ckoenig.leichtzumerken@xxxxxxxxx>; amd-gfx@xxxxxxxxxxxxxxxxxxxxx
> > Subject: Re: [PATCH] drm/amdgpu: capture invalid hardware access v2
> >
> > But how this will help if TDR thread will start after you both took read lock and checked that adev->in_gpu_reset is false ? Since TDR  now takes write lock only after suspending HW and waiting for all fences there is nothing that prevents both threads (e.g IOCTL and TDR) to access registers concurrently.
> >
> > Andrey
> >
> > On 2021-03-09 9:59 p.m., Li, Dennis wrote:
> > > [AMD Official Use Only - Internal Distribution Only]
> > >
> > > Hi, Andrey,
> > > > > > Is the problem here that HW is suspended while some other threads that rely on the read side lock still access HW ? Mostly what I am thinking about are IOCTls - we can't 'wait for them to complete' but they might be accessing HW when we start suspend.
> > > In read side, when the reader held the read lock, it will also check whether adev->in_gpu_reset is 1, if so, it will release read clock and is waiting for recovery finish event.
> > > Best Regards
> > > Dennis Li
> > >
> > > -----Original Message-----
> > > From: Grodzovsky, Andrey <Andrey.Grodzovsky@xxxxxxx>
> > > Sent: Wednesday, March 10, 2021 2:26 AM
> > > To: Christian König <ckoenig.leichtzumerken@xxxxxxxxx>; Li, Dennis
> > > <Dennis.Li@xxxxxxx>; amd-gfx@xxxxxxxxxxxxxxxxxxxxx
> > > Subject: Re: [PATCH] drm/amdgpu: capture invalid hardware access v2
> > >
> > > On 2021-03-09 12:47 p.m., Christian König wrote:
> > > > No it won't. Accessing the hardware without the lock is ok as long
> > > > as the write side isn't taken.
> > > Oh, forgot about the trylock part, sorry...
> > >
> > > > But that approach is illegal anyway because we suspend the hardware
> > > > without proper protection from concurrent access.
> > > For my understanding and from looking again at his steps related to
> > > this
> > >
> > > Step 0: atomic_cmpxchg(&adev->in_gpu_reset, 0, 1) - [AG] protects
> > > from other TDR threads
> > >
> > > Step 1: cancel all delay works, stop drm schedule, complete all
> > > unreceived fences and so on. Call amdgpu_device_pre_asic_reset...
> > > e.t.c
> > > - [AG] this is the HW suspend part
> > >
> > > Step 2: call down_write(&adev->reset_sem) to hold write lock, which will block recovery thread until other threads release read locks.
> > >
> > > Is the problem here that HW is suspended while some other threads that rely on the read side lock still access HW ? Mostly what I am thinking about are IOCTls - we can't 'wait for them to complete' but they might be accessing HW when we start suspend.
> > >
> > > Andrey
> > >
> > >
> > > > Christian.
> > > >
> > > > Am 09.03.21 um 17:40 schrieb Andrey Grodzovsky:
> > > > > Because he takes the write side lock post amdgpu_pre_asic_reset -
> > > > > where HW suspend sequence happens (touching registers) - so i think
> > > > > it will assert.
> > > > >
> > > > > Andrey
> > > > >
> > > > > On 2021-03-09 7:56 a.m., Christian König wrote:
> > > > > > Hi Dennis,
> > > > > >
> > > > > > why do you think that this will always assert in reset thread?
> > > > > >
> > > > > > In the reset thread while we are holding the reset lock write side
> > > > > > lockdep_assert_held() should be satisfied and not cause any splat
> > > > > > in the system log.
> > > > > >
> > > > > > Regards,
> > > > > > Christian.
> > > > > >
> > > > > > Am 09.03.21 um 03:03 schrieb Li, Dennis:
> > > > > > > [AMD Official Use Only - Internal Distribution Only]
> > > > > > >
> > > > > > > Hi, Christian,
> > > > > > >            amdgpu_device_skip_hw_access will always assert in
> > > > > > > reset thread, which seems not a good idea.
> > > > > > >
> > > > > > > Best Regards
> > > > > > > Dennis Li
> > > > > > > -----Original Message-----
> > > > > > > From: Christian König <ckoenig.leichtzumerken@xxxxxxxxx>
> > > > > > > Sent: Tuesday, March 9, 2021 2:07 AM
> > > > > > > To: amd-gfx@xxxxxxxxxxxxxxxxxxxxx
> > > > > > > Cc: Grodzovsky, Andrey <Andrey.Grodzovsky@xxxxxxx>; Li, Dennis
> > > > > > > <Dennis.Li@xxxxxxx>
> > > > > > > Subject: [PATCH] drm/amdgpu: capture invalid hardware access v2
> > > > > > >
> > > > > > > From: Dennis Li <Dennis.Li@xxxxxxx>
> > > > > > >
> > > > > > > When recovery thread has begun GPU reset, there should be not
> > > > > > > other threads to access hardware, otherwise system randomly hang.
> > > > > > >
> > > > > > > v2 (chk): rewritten from scratch, use trylock and lockdep instead
> > > > > > > of
> > > > > > >       hand wiring the logic.
> > > > > > >
> > > > > > > Signed-off-by: Dennis Li <Dennis.Li@xxxxxxx>
> > > > > > > Signed-off-by: Christian König <christian.koenig@xxxxxxx>
> > > > > > > ---
> > > > > > >      drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 74
> > > > > > > +++++++++++++++++-----
> > > > > > >      1 file changed, 57 insertions(+), 17 deletions(-)
> > > > > > >
> > > > > > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> > > > > > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> > > > > > > index e247c3a2ec08..c990af6a43ca 100644
> > > > > > > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> > > > > > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> > > > > > > @@ -326,6 +326,34 @@ void amdgpu_device_vram_access(struct
> > > > > > > amdgpu_device *adev, loff_t pos,
> > > > > > >      /*
> > > > > > >       * register access helper functions.
> > > > > > >       */
> > > > > > > +
> > > > > > > +/* Check if hw access should be skipped because of hotplug or
> > > > > > > +device error */ static bool amdgpu_device_skip_hw_access(struct
> > > > > > > +amdgpu_device
> > > > > > > +*adev) {
> > > > > > > +if (adev->in_pci_err_recovery)
> > > > > > > +return true;
> > > > > > > +
> > > > > > > +#ifdef CONFIG_LOCKDEP
> > > > > > > +/*
> > > > > > > + * This is a bit complicated to understand, so worth a comment.
> > > > > > > What we assert
> > > > > > > + * here is that the GPU reset is not running on another thread
> > > > > > > + in
> > > > > > > parallel.
> > > > > > > + *
> > > > > > > + * For this we trylock the read side of the reset semaphore, if
> > > > > > > that succeeds
> > > > > > > + * we know that the reset is not running in paralell.
> > > > > > > + *
> > > > > > > + * If the trylock fails we assert that we are either already
> > > > > > > holding the read
> > > > > > > + * side of the lock or are the reset thread itself and hold the
> > > > > > > write side of
> > > > > > > + * the lock.
> > > > > > > + */
> > > > > > > +if (down_read_trylock(&adev->reset_sem))
> > > > > > > +up_read(&adev->reset_sem);
> > > > > > > +else
> > > > > > > +lockdep_assert_held(&adev->reset_sem);
> > > > > > > +#endif
> > > > > > > +
> > > > > > > +return false;
> > > > > > > +}
> > > > > > > +
> > > > > > >      /**
> > > > > > >       * amdgpu_device_rreg - read a memory mapped IO or indirect
> > > > > > > register
> > > > > > >       *
> > > > > > > @@ -340,7 +368,7 @@ uint32_t amdgpu_device_rreg(struct
> > > > > > > amdgpu_device *adev,  {
> > > > > > >      uint32_t ret;
> > > > > > >
> > > > > > > -if (adev->in_pci_err_recovery)
> > > > > > > +if (amdgpu_device_skip_hw_access(adev))
> > > > > > >      return 0;
> > > > > > >
> > > > > > >      if ((reg * 4) < adev->rmmio_size) { @@ -377,7 +405,7 @@
> > > > > > > uint32_t amdgpu_device_rreg(struct amdgpu_device *adev,
> > > > > > >       */
> > > > > > >      uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t
> > > > > > > offset)  {
> > > > > > > -if (adev->in_pci_err_recovery)
> > > > > > > +if (amdgpu_device_skip_hw_access(adev))
> > > > > > >      return 0;
> > > > > > >
> > > > > > >      if (offset < adev->rmmio_size) @@ -402,7 +430,7 @@ uint8_t
> > > > > > > amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset)
> > > > > > >       */
> > > > > > >      void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t
> > > > > > > offset, uint8_t value)  { -if (adev->in_pci_err_recovery)
> > > > > > > +if (amdgpu_device_skip_hw_access(adev))
> > > > > > >      return;
> > > > > > >
> > > > > > >      if (offset < adev->rmmio_size) @@ -425,7 +453,7 @@ void
> > > > > > > amdgpu_device_wreg(struct amdgpu_device *adev,
> > > > > > >      uint32_t reg, uint32_t v,
> > > > > > >      uint32_t acc_flags)
> > > > > > >      {
> > > > > > > -if (adev->in_pci_err_recovery)
> > > > > > > +if (amdgpu_device_skip_hw_access(adev))
> > > > > > >      return;
> > > > > > >
> > > > > > >      if ((reg * 4) < adev->rmmio_size) { @@ -452,7 +480,7 @@ void
> > > > > > > amdgpu_device_wreg(struct amdgpu_device *adev,  void
> > > > > > > amdgpu_mm_wreg_mmio_rlc(struct amdgpu_device *adev,
> > > > > > >           uint32_t reg, uint32_t v)
> > > > > > >      {
> > > > > > > -if (adev->in_pci_err_recovery)
> > > > > > > +if (amdgpu_device_skip_hw_access(adev))
> > > > > > >      return;
> > > > > > >
> > > > > > >      if (amdgpu_sriov_fullaccess(adev) && @@ -475,7 +503,7 @@ void
> > > > > > > amdgpu_mm_wreg_mmio_rlc(struct amdgpu_device *adev,
> > > > > > >       */
> > > > > > >      u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)  {
> > > > > > > -if
> > > > > > > (adev->in_pci_err_recovery)
> > > > > > > +if (amdgpu_device_skip_hw_access(adev))
> > > > > > >      return 0;
> > > > > > >
> > > > > > >      if ((reg * 4) < adev->rio_mem_size) @@ -497,7 +525,7 @@ u32
> > > > > > > amdgpu_io_rreg(struct amdgpu_device *adev,
> > > > > > > u32 reg)
> > > > > > >       */
> > > > > > >      void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32
> > > > > > > v) { -if (adev->in_pci_err_recovery)
> > > > > > > +if (amdgpu_device_skip_hw_access(adev))
> > > > > > >      return;
> > > > > > >
> > > > > > >      if ((reg * 4) < adev->rio_mem_size) @@ -519,7 +547,7 @@ void
> > > > > > > amdgpu_io_wreg(struct amdgpu_device *adev,
> > > > > > > u32 reg, u32 v)
> > > > > > >       */
> > > > > > >      u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32
> > > > > > > index) { -if (adev->in_pci_err_recovery)
> > > > > > > +if (amdgpu_device_skip_hw_access(adev))
> > > > > > >      return 0;
> > > > > > >
> > > > > > >      if (index < adev->doorbell.num_doorbells) { @@ -542,7 +570,7
> > > > > > > @@
> > > > > > > u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
> > > > > > >       */
> > > > > > >      void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32
> > > > > > > index,
> > > > > > > u32 v)  {
> > > > > > > -if (adev->in_pci_err_recovery)
> > > > > > > +if (amdgpu_device_skip_hw_access(adev))
> > > > > > >      return;
> > > > > > >
> > > > > > >      if (index < adev->doorbell.num_doorbells) { @@ -563,7 +591,7
> > > > > > > @@ void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32
> > > > > > > index,
> > > > > > > u32
> > > > > > > v)
> > > > > > >       */
> > > > > > >      u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32
> > > > > > > index) { -if (adev->in_pci_err_recovery)
> > > > > > > +if (amdgpu_device_skip_hw_access(adev))
> > > > > > >      return 0;
> > > > > > >
> > > > > > >      if (index < adev->doorbell.num_doorbells) { @@ -586,7 +614,7
> > > > > > > @@
> > > > > > > u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
> > > > > > >       */
> > > > > > >      void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32
> > > > > > > index,
> > > > > > > u64 v)  {
> > > > > > > -if (adev->in_pci_err_recovery)
> > > > > > > +if (amdgpu_device_skip_hw_access(adev))
> > > > > > >      return;
> > > > > > >
> > > > > > >      if (index < adev->doorbell.num_doorbells) { @@ -610,10
> > > > > > > +638,13 @@
> > > > > > > u32 amdgpu_device_indirect_rreg(struct amdgpu_device *adev,
> > > > > > >      u32 pcie_index, u32 pcie_data,
> > > > > > >      u32 reg_addr)
> > > > > > >      {
> > > > > > > -unsigned long flags;
> > > > > > > -u32 r;
> > > > > > >      void __iomem *pcie_index_offset;
> > > > > > >      void __iomem *pcie_data_offset;
> > > > > > > +unsigned long flags;
> > > > > > > +u32 r;
> > > > > > > +
> > > > > > > +if (amdgpu_device_skip_hw_access(adev))
> > > > > > > +return 0;
> > > > > > >
> > > > > > >      spin_lock_irqsave(&adev->pcie_idx_lock, flags);
> > > > > > >      pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index
> > > > > > > * 4; @@ -641,10 +672,13 @@ u64
> > > > > > > amdgpu_device_indirect_rreg64(struct
> > > > > > > amdgpu_device *adev,
> > > > > > >        u32 pcie_index, u32 pcie_data,
> > > > > > >        u32 reg_addr)
> > > > > > >      {
> > > > > > > -unsigned long flags;
> > > > > > > -u64 r;
> > > > > > >      void __iomem *pcie_index_offset;
> > > > > > >      void __iomem *pcie_data_offset;
> > > > > > > +unsigned long flags;
> > > > > > > +u64 r;
> > > > > > > +
> > > > > > > +if (amdgpu_device_skip_hw_access(adev))
> > > > > > > +return 0;
> > > > > > >
> > > > > > >      spin_lock_irqsave(&adev->pcie_idx_lock, flags);
> > > > > > >      pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index
> > > > > > > * 4; @@ -677,9 +711,12 @@ void amdgpu_device_indirect_wreg(struct
> > > > > > > amdgpu_device *adev,
> > > > > > >       u32 pcie_index, u32 pcie_data,
> > > > > > >       u32 reg_addr, u32 reg_data)
> > > > > > >      {
> > > > > > > -unsigned long flags;
> > > > > > >      void __iomem *pcie_index_offset;
> > > > > > >      void __iomem *pcie_data_offset;
> > > > > > > +unsigned long flags;
> > > > > > > +
> > > > > > > +if (amdgpu_device_skip_hw_access(adev))
> > > > > > > +return;
> > > > > > >
> > > > > > >      spin_lock_irqsave(&adev->pcie_idx_lock, flags);
> > > > > > >      pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index
> > > > > > > * 4; @@ -706,9 +743,12 @@ void
> > > > > > > amdgpu_device_indirect_wreg64(struct
> > > > > > > amdgpu_device *adev,
> > > > > > >         u32 pcie_index, u32 pcie_data,
> > > > > > >         u32 reg_addr, u64 reg_data)
> > > > > > >      {
> > > > > > > -unsigned long flags;
> > > > > > >      void __iomem *pcie_index_offset;
> > > > > > >      void __iomem *pcie_data_offset;
> > > > > > > +unsigned long flags;
> > > > > > > +
> > > > > > > +if (amdgpu_device_skip_hw_access(adev))
> > > > > > > +return;
> > > > > > >
> > > > > > >      spin_lock_irqsave(&adev->pcie_idx_lock, flags);
> > > > > > >      pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index
> > > > > > > * 4;
> > > > > > > --
> > > > > > > 2.25.1

_______________________________________________
amd-gfx mailing list
amd-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/amd-gfx