Re: [PATCH] drm/msm/adreno: De-spaghettify the use of memory barriers

[Andrew Halaney]

On Thu, May 16, 2024 at 08:20:05PM GMT, Akhil P Oommen wrote:
> On Thu, May 16, 2024 at 08:15:34AM -0500, Andrew Halaney wrote:
> > On Wed, May 15, 2024 at 12:08:49AM GMT, Akhil P Oommen wrote:
> > > On Wed, May 08, 2024 at 07:46:31PM +0200, Konrad Dybcio wrote:
> > > > Memory barriers help ensure instruction ordering, NOT time and order
> > > > of actual write arrival at other observers (e.g. memory-mapped IP).
> > > > On architectures employing weak memory ordering, the latter can be a
> > > > giant pain point, and it has been as part of this driver.
> > > > 
> > > > Moreover, the gpu_/gmu_ accessors already use non-relaxed versions of
> > > > readl/writel, which include r/w (respectively) barriers.
> > > > 
> > > > Replace the barriers with a readback that ensures the previous writes
> > > > have exited the write buffer (as the CPU must flush the write to the
> > > > register it's trying to read back) and subsequently remove the hack
> > > > introduced in commit b77532803d11 ("drm/msm/a6xx: Poll for GBIF unhalt
> > > > status in hw_init").
> > 
> > For what its worth, I've been eyeing (but haven't tested) sending some
> > patches to clean up dsi_phy_write_udelay/ndelay(). There's no ordering
> > guarantee between a writel() and a delay(), so the expected "write then
> > delay" sequence might not be happening.. you need to write, read, delay.
> > 
> > memory-barriers.txt:
> > 
> > 	5. A readX() by a CPU thread from the peripheral will complete before
> > 	   any subsequent delay() loop can begin execution on the same thread.
> > 	   This ensures that two MMIO register writes by the CPU to a peripheral
> > 	   will arrive at least 1us apart if the first write is immediately read
> > 	   back with readX() and udelay(1) is called prior to the second
> > 	   writeX():
> > 
> > 		writel(42, DEVICE_REGISTER_0); // Arrives at the device...
> > 		readl(DEVICE_REGISTER_0);
> > 		udelay(1);
> > 		writel(42, DEVICE_REGISTER_1); // ...at least 1us before this.
> 
> Yes, udelay orders only with readl(). I saw a patch from Will Deacon
> which fixes this for arm64 few years back:
> https://lore.kernel.org/all/1543251228-30001-1-git-send-email-will.deacon@xxxxxxx/T/
> 
> But this is needed only when you write io and do cpuside wait , not when
> you poll io to check status.

Sure, I'm just highlighting the bug in dsi_phy_write_*delay() functions
here, which match the udelay case you mention.

> 
> > 
> > > > 
> > > > Fixes: b77532803d11 ("drm/msm/a6xx: Poll for GBIF unhalt status in hw_init")
> > > > Signed-off-by: Konrad Dybcio <konrad.dybcio@xxxxxxxxxx>
> > > > ---
> > > >  drivers/gpu/drm/msm/adreno/a6xx_gmu.c |  5 ++---
> > > >  drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 14 ++++----------
> > > >  2 files changed, 6 insertions(+), 13 deletions(-)
> > > 
> > > I prefer this version compared to the v2. A helper routine is
> > > unnecessary here because:
> > > 1. there are very few scenarios where we have to read back the same
> > > register.
> > > 2. we may accidently readback a write only register.
> > > 
> > > > 
> > > > diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
> > > > index 0e3dfd4c2bc8..4135a53b55a7 100644
> > > > --- a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
> > > > +++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
> > > > @@ -466,9 +466,8 @@ static int a6xx_rpmh_start(struct a6xx_gmu *gmu)
> > > >  	int ret;
> > > >  	u32 val;
> > > >  
> > > > -	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1 << 1);
> > > > -	/* Wait for the register to finish posting */
> > > > -	wmb();
> > > > +	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, BIT(1));
> > > > +	gmu_read(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ);
> > > 
> > > This is unnecessary because we are polling on a register on the same port below. But I think we
> > > can replace "wmb()" above with "mb()" to avoid reordering between read
> > > and write IO instructions.
> > 
> > If I understand correctly, you don't need any memory barrier.
> > writel()/readl()'s are ordered to the same endpoint. That goes for all
> > the reordering/barrier comments mentioned below too.
> > 
> > device-io.rst:
> > 
> >     The read and write functions are defined to be ordered. That is the
> >     compiler is not permitted to reorder the I/O sequence. When the ordering
> >     can be compiler optimised, you can use __readb() and friends to
> >     indicate the relaxed ordering. Use this with care.
> > 
> > memory-barriers.txt:
> > 
> >      (*) readX(), writeX():
> > 
> > 	    The readX() and writeX() MMIO accessors take a pointer to the
> > 	    peripheral being accessed as an __iomem * parameter. For pointers
> > 	    mapped with the default I/O attributes (e.g. those returned by
> > 	    ioremap()), the ordering guarantees are as follows:
> > 
> > 	    1. All readX() and writeX() accesses to the same peripheral are ordered
> > 	       with respect to each other. This ensures that MMIO register accesses
> > 	       by the same CPU thread to a particular device will arrive in program
> > 	       order.
> > 
> 
> In arm64, a writel followed by readl translates to roughly the following
> sequence: dmb_wmb(), __raw_writel(), __raw_readl(), dmb_rmb(). I am not
> sure what is stopping compiler from reordering  __raw_writel() and __raw_readl()
> above? I am assuming iomem cookie is ignored during compilation.

It seems to me that is due to some usage of volatile there in
__raw_writel() etc, but to be honest after reading about volatile and
some threads from gcc mailing lists, I don't have a confident answer :)

> 
> Added Will to this thread if he can throw some light on this.

Hopefully Will can school us.

> 
> -Akhil
> 
> > 
> > > 
> > > >  
> > > >  	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_RSCC_CONTROL_ACK, val,
> > > >  		val & (1 << 1), 100, 10000);
> > > > diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> > > > index 973872ad0474..0acbc38b8e70 100644
> > > > --- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> > > > +++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> > > > @@ -1713,22 +1713,16 @@ static int hw_init(struct msm_gpu *gpu)
> > > >  	}
> > > >  
> > > >  	/* Clear GBIF halt in case GX domain was not collapsed */
> > > > +	gpu_write(gpu, REG_A6XX_GBIF_HALT, 0);
> > > 
> > > We need a full barrier here to avoid reordering. Also, lets add a
> > > comment about why we are doing this odd looking sequence.
> > > 
> > > > +	gpu_read(gpu, REG_A6XX_GBIF_HALT);
> > > >  	if (adreno_is_a619_holi(adreno_gpu)) {
> > > > -		gpu_write(gpu, REG_A6XX_GBIF_HALT, 0);
> > > >  		gpu_write(gpu, REG_A6XX_RBBM_GPR0_CNTL, 0);
> > > > -		/* Let's make extra sure that the GPU can access the memory.. */
> > > > -		mb();
> > > 
> > > We need a full barrier here.
> > > 
> > > > +		gpu_read(gpu, REG_A6XX_RBBM_GPR0_CNTL);
> > > >  	} else if (a6xx_has_gbif(adreno_gpu)) {
> > > > -		gpu_write(gpu, REG_A6XX_GBIF_HALT, 0);
> > > >  		gpu_write(gpu, REG_A6XX_RBBM_GBIF_HALT, 0);
> > > > -		/* Let's make extra sure that the GPU can access the memory.. */
> > > > -		mb();
> > > 
> > > We need a full barrier here.
> > > 
> > > > +		gpu_read(gpu, REG_A6XX_RBBM_GBIF_HALT);
> > > >  	}
> > > >  
> > > > -	/* Some GPUs are stubborn and take their sweet time to unhalt GBIF! */
> > > > -	if (adreno_is_a7xx(adreno_gpu) && a6xx_has_gbif(adreno_gpu))
> > > > -		spin_until(!gpu_read(gpu, REG_A6XX_GBIF_HALT_ACK));
> > > > -
> > > 
> > > Why is this removed?
> > > 
> > > -Akhil
> > > 
> > > >  	gpu_write(gpu, REG_A6XX_RBBM_SECVID_TSB_CNTL, 0);
> > > >  
> > > >  	if (adreno_is_a619_holi(adreno_gpu))
> > > > 
> > > > ---
> > > > base-commit: 93a39e4766083050ca0ecd6a3548093a3b9eb60c
> > > > change-id: 20240508-topic-adreno-a2d199cd4152
> > > > 
> > > > Best regards,
> > > > -- 
> > > > Konrad Dybcio <konrad.dybcio@xxxxxxxxxx>
> > > > 
> > > 
> > 
>