[PATCH v2 4/9] drm/panthor: Implement optional reset

[Alexander Stein]

Hi,

Am Donnerstag, 4. September 2025, 08:22:24 CEST schrieb Boris Brezillon:
> On Wed, 3 Sep 2025 23:44:59 +0200
> Marek Vasut <marek.vasut at mailbox.org> wrote:
> 
> > On 3/25/25 3:52 PM, Boris Brezillon wrote:
> > 
> > Hello Boris,
> > 
> > sorry for the late reply.
> > 
> > >>>>>>> Hm, that might be the cause of the fast reset issue (which is a fast
> > >>>>>>> resume more than a fast reset BTW): if you re-assert the reset line on
> > >>>>>>> runtime suspend, I guess this causes a full GPU reset, and the MCU ends
> > >>>>>>> up in a state where it needs a slow reset (all data sections reset to
> > >>>>>>> their initial state). Can you try to move the reset_control_[de]assert
> > >>>>>>> to the unplug/init functions?  
> > >>>>>> Is it correct to assume , that if I remove all reset_control_assert()
> > >>>>>> calls (and keep only the _deassert() calls), the slow resume problem
> > >>>>>> should go away too ?  
> > >>>>>
> > >>>>> Yeah, dropping the _assert()s should do the trick.  
> > >>>> Hmmm, no, that does not help. I was hoping maybe NXP can chime in and
> > >>>> suggest something too ?  
> > >>>
> > >>> Can you try keep all the clks/regulators/power-domains/... on after
> > >>> init, and see if the fast resume works with that. If it does,
> > >>> re-introduce one resource at a time to find out which one causes the
> > >>> MCU to lose its state.  
> > >>
> > >> I already tried that too . I spent quite a while until I reached that L2
> > >> workaround in fact.  
> > > 
> > > So, with your RPM suspend/resume being NOPs, it still doesn't work?
> > > Unless the FW is doing something behind our back, I don't really see
> > > why this would fail on your platform, but not on the rk3588. Are you
> > > sure the power domains are kept on at all times. I'm asking, because if
> > > you linked all the PDs, the on/off sequence is automatically handled by
> > > the RPM core at suspend/resume time.  
> > 
> > I revisited this now.
> > 
> > Can you please test the following patch (also attached) on one of your 
> > devices, and tell me what the status is at the end . The diff sets the 
> > GLB_HALT bit and then clears it again, which I suspect should first halt 
> > the GPU and (this is what I am unsure about) then again un-halt/resume 
> > the GPU ?
> 
> It doesn't work like that. What you're describing is like executing
> "shutdown" on your terminal and then typing "boot" on the keyboard
> after your computer has been shut down.
> 
> > 
> > "
> > diff --git a/drivers/gpu/drm/panthor/panthor_fw.c 
> > b/drivers/gpu/drm/panthor/panthor_fw.c
> > index 9bf06e55eaeea..57c0d4fd29aa2 100644
> > --- a/drivers/gpu/drm/panthor/panthor_fw.c
> > +++ b/drivers/gpu/drm/panthor/panthor_fw.c
> > @@ -1087,8 +1087,16 @@ void panthor_fw_pre_reset(struct panthor_device 
> > *ptdev, bool on_hang)
> >   		struct panthor_fw_global_iface *glb_iface = 
> > panthor_fw_get_glb_iface(ptdev);
> >   		u32 status;
> > 
> > +pr_err("%s[%d] pre-halt status=%x\n", __func__, __LINE__, 
> > gpu_read(ptdev, MCU_STATUS));
> > +
> >   		panthor_fw_update_reqs(glb_iface, req, GLB_HALT, GLB_HALT);
> >   		gpu_write(ptdev, CSF_DOORBELL(CSF_GLB_DOORBELL_ID), 1);
> > +mdelay(100);
> > +pr_err("%s[%d] likely-halted status=%x\n", __func__, __LINE__, 
> > gpu_read(ptdev, MCU_STATUS));
> > +		panthor_fw_update_reqs(glb_iface, req, 0, GLB_HALT);
> > +mdelay(100);
> > +pr_err("%s[%d] likely-running ? status=%x\n", __func__, __LINE__, 
> > gpu_read(ptdev, MCU_STATUS));
> > +
> >   		if (!gpu_read_poll_timeout(ptdev, MCU_STATUS, status,
> >   					   status == MCU_STATUS_HALT, 10,
> >   					   100000)) {
> > "
> > 
> > In my case, the relevant output looks like this:
> > 
> > "
> > [    3.326805] panthor_fw_pre_reset[1090] pre-halt status=1
> > [    3.432151] panthor_fw_pre_reset[1095] likely-halted status=2
> > [    3.542179] panthor_fw_pre_reset[1098] likely-running ? status=2
> > "
> > 
> > That means, the GPU remains halted at the end, even if the "GLB_HALT" 
> > bit is cleared before the last print. The clearing of GLB_HALT is also 
> > what panthor_fw_post_reset() does.
> 
> After the halt has been processed by the FW, the memory region where
> you check the halt status again is inert, since the micro-controller
> (MCU) supposed to update those bits is off at this point. The FW
> interface is really just a shared memory region between the CPU and
> MCU, nothing more.
> 
> > 
> > I suspect the extra soft reset I did before "un-halted" the GPU and 
> > allowed it to proceed.
> 
> Hm, not quite. I mean, you still need to explicitly boot the MCU after
> a reset, which is what the write to MCU_CONTROL [1] does. What the
> soft-reset does though, is reset all GPU blocks, including the MCU.
> This means the MCU starts from a fresh state when you reach [1].
> 
> If I had to guess, I'd say something is messed up when the GPU is
> halted, and you need a soft-reset to recover from that. Unfortunately,
> I don't know enough about what your FW is doing to help. Maybe
> Arm/Freescale could...
> 
> > 
> > I wonder if there is some way to un-halt the GPU using some gpu_write() 
> > direct register access, is there ?
> 
> That's MCU_CONTROL, yes. And it's done here [1] already.
> 
> > Maybe the GPU remains halted because 
> > setting the GLB_HALT stops command stream processing, and the GPU never 
> > samples the clearing of GLB_HALT and therefore remains halted forever ?
> 
> Exactly that, and that's expected.

FYI: in a new release of system manager software (starting from lf-6.12.3-1.0.0),
the GPU reset is reasserted in SM software already [1] and access to GPU
block control has been removed from Cortex-A [2]. Starting from B0 step this
version is required AFAIK.

Best regards
Alexander

[1] https://github.com/nxp-imx/imx-sm/commit/2dcc0409ede82eef54857be50daa588b23b3ba7b
[2] https://github.com/nxp-imx/imx-sm/commit/a3e5da9ea51144f513ac3909fa151fa7df394100

-- 
TQ-Systems GmbH | Mühlstraße 2, Gut Delling | 82229 Seefeld, Germany
Amtsgericht München, HRB 105018
Geschäftsführer: Detlef Schneider, Rüdiger Stahl, Stefan Schneider
http://www.tq-group.com/