[PATCH 1/2] drm/amdgpu: make duplicated EOP packet for GFX7/8 have real content

[Icenowy Zheng]

在 2024-06-17星期一的 16:42 +0200，Christian König写道：
> Am 17.06.24 um 16:30 schrieb Icenowy Zheng:
> > 在 2024-06-17星期一的 15:59 +0200，Christian König写道：
> > > Am 17.06.24 um 15:43 schrieb Icenowy Zheng:
> > > > 在 2024-06-17星期一的 15:09 +0200，Christian König写道：
> > > > > Am 17.06.24 um 15:03 schrieb Icenowy Zheng:
> > > > > > 在 2024-06-17星期一的 14:35 +0200，Christian König写道：
> > > > > > > Am 17.06.24 um 12:58 schrieb Icenowy Zheng:
> > > > > > > > The duplication of EOP packets for GFX7/8, with the
> > > > > > > > former
> > > > > > > > one
> > > > > > > > have
> > > > > > > > seq-1 written and the latter one have seq written,
> > > > > > > > seems to
> > > > > > > > confuse
> > > > > > > > some
> > > > > > > > hardware platform (e.g. Loongson 7A series PCIe
> > > > > > > > controllers).
> > > > > > > > 
> > > > > > > > Make the content of the duplicated EOP packet the same
> > > > > > > > with
> > > > > > > > the
> > > > > > > > real
> > > > > > > > one, only masking any possible interrupts.
> > > > > > > Well completely NAK to that, exactly that disables the
> > > > > > > workaround.
> > > > > > > 
> > > > > > > The CPU needs to see two different values written here.
> > > > > > Why do the CPU need to see two different values here? Only
> > > > > > the
> > > > > > second
> > > > > > packet will raise an interrupt before and after applying
> > > > > > this
> > > > > > patch,
> > > > > > and the first packet's result should just be overriden on
> > > > > > ordinary
> > > > > > platforms. The CPU won't see the first one, until it's
> > > > > > polling
> > > > > > for
> > > > > > the
> > > > > > address for a very short interval, so short that the GPU CP
> > > > > > couldn't
> > > > > > execute 2 commands.
> > > > > Yes exactly that. We need to make two writes, one with the
> > > > > old
> > > > > value
> > > > > (seq - 1) and a second with the real value (seq).
> > > > > 
> > > > > Otherwise it is possible that a polling CPU would see the
> > > > > sequence
> > > > > before the second EOP is issued with results in incoherent
> > > > > view
> > > > > of
> > > > > memory.
> > > > In this case shouldn't we write seq-1 before any work, and then
> > > > write
> > > > seq after work, like what is done in Mesa?
> > > No. This hw workaround requires that two consecutive write
> > > operations
> > > happen directly behind each other on the PCIe bus with two
> > > different
> > > values.
> > Well to be honest the workaround code in Mesa seems to not be
> > working
> > in this way ...
> 
> Mesa doesn't have any workaround for that hw issue, the code there
> uses 
> a quite different approach.

Ah? Commit bf26da927a1c ("drm/amdgpu: add cache flush workaround to
gfx8 emit_fence") says "Both PAL and Mesa use it for gfx8 too, so port
this commit to gfx_v8_0_ring_emit_fence_gfx", so maybe the workaround
should just be not necessary here?


> 
> > > To make the software logic around that work without any changes
> > > we
> > > use
> > > the values seq - 1 and seq because those are guaranteed to be
> > > different
> > > and not trigger any unwanted software behavior.
> > > 
> > > Only then we can guarantee that we have a coherent view of system
> > > memory.
> > Any more details about it?
> 
> No, sorry. All I know is that it's a bug in the cache flush logic
> which 
> can be worked around by issuing two write behind each other to the
> same 
> location.

So the issue is that the first EOP write does not properly flush the
cache? Could EVENT_WRITE be used instead of EVENT_WRITE_EOP in this
workaround to properly flush it without hurting the fence value?


> 
> > BTW in this case, could I try to write it for 3 times instead of 2,
> > with seq-1, seq and seq?
> 
> That could potentially work as well, but at some point we would need
> to 
> increase the EOP ring buffer size or could run into performance
> issues.

Well I will try this. I think the buffer is enlarged in the original
workaround commit.

> 
> > > > As what I see, Mesa uses another command buffer to emit a
> > > > EVENT_WRITE_EOP writing 0, and commit this command buffer
> > > > before
> > > > the
> > > > real command buffer.
> > > > 
> > > > > > Or do you mean the GPU needs to see two different values
> > > > > > being
> > > > > > written,
> > > > > > or they will be merged into only one write request?
> > > > > > 
> > > > > > Please give out more information about this workaround,
> > > > > > otherwise
> > > > > > the
> > > > > > GPU hang problem on Loongson platforms will persist.
> > > > > Well if Loongson can't handle two consecutive write
> > > > > operations to
> > > > > the
> > > > > same address with different values then you have a massive
> > > > > platform
> > > > > bug.
> > > > I think the issue is triggered when two consecutive write
> > > > operations
> > > > and one IRQ is present, which is exactly the case of this
> > > > function.
> > > Well then you have a massive platform bug.
> > > 
> > > Two consecutive writes to the same bus address are perfectly
> > > legal
> > > from
> > > the PCIe specification and can happen all the time, even without
> > > this
> > > specific hw workaround.
> > Yes I know it, and I am not from Loongson, just some user trying to
> > mess around it.
> 
> Well to be honest on a platform where even two consecutive writes to
> the 
> same location doesn't work I would have strong doubts that it is
> stable 
> in general.

Well I think the current situation is that the IRQ triggered by the
second EOP packet arrives before the second write is finished, not the
second write is totally dropped.

> 
> Regards,
> Christian.