Re: drm/etnaviv: slow down FE idle polling

[Lucas Stach <l.stach@pengutronix.de>]

Am Donnerstag, dem 15.06.2023 um 12:09 +0800 schrieb Sui Jingfeng:
> Hi,
> 
> On 2023/6/7 20:59, Lucas Stach wrote:
> > Currently the FE is spinning way too fast when polling for new work in
> 'way' -> 'away'
> > the FE idleloop.
> 'idleloop' -> 'idle loop'
> >   As each poll fetches 16 bytes from memory, a GPU running
> > at 1GHz with the current setting of 200 wait cycle between fetches causes
> > 80 MB/s of memory traffic just to check for new work when the GPU is
> > otherwise idle, which is more FE traffic than in some GPU loaded cases.
> > 
> > Significantly increase the number of wait cycles to slow down the poll
> > interval to ~30µs, limiting the FE idle memory traffic to 512 KB/s, while
> > providing a max latency which should not hurt most use-cases. The FE WAIT
> > command seems to have some unknown discrete steps in the wait cycles
> add a comma here.
> >   so
> > we may over/undershoot the target a bit, but that should be harmless.
> overshoot or undershoot
> > Signed-off-by: Lucas Stach <l.stach@pengutronix.de>
> > Reviewed-by: Christian Gmeiner <cgmeiner@igalia.com>
> > ---
> >   drivers/gpu/drm/etnaviv/etnaviv_buffer.c | 11 ++++++-----
> >   drivers/gpu/drm/etnaviv/etnaviv_gpu.c    |  7 +++++++
> >   drivers/gpu/drm/etnaviv/etnaviv_gpu.h    |  1 +
> >   3 files changed, 14 insertions(+), 5 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/etnaviv/etnaviv_buffer.c b/drivers/gpu/drm/etnaviv/etnaviv_buffer.c
> > index cf741c5c82d2..384df1659be6 100644
> > --- a/drivers/gpu/drm/etnaviv/etnaviv_buffer.c
> > +++ b/drivers/gpu/drm/etnaviv/etnaviv_buffer.c
> > @@ -53,11 +53,12 @@ static inline void CMD_END(struct etnaviv_cmdbuf *buffer)
> >   	OUT(buffer, VIV_FE_END_HEADER_OP_END);
> >   }
> >   
> > -static inline void CMD_WAIT(struct etnaviv_cmdbuf *buffer)
> > +static inline void CMD_WAIT(struct etnaviv_cmdbuf *buffer,
> > +			    unsigned int waitcycles)
> >   {
> >   	buffer->user_size = ALIGN(buffer->user_size, 8);
> >   
> > -	OUT(buffer, VIV_FE_WAIT_HEADER_OP_WAIT | 200);
> > +	OUT(buffer, VIV_FE_WAIT_HEADER_OP_WAIT | waitcycles);
> >   }
> >   
> >   static inline void CMD_LINK(struct etnaviv_cmdbuf *buffer,
> > @@ -168,7 +169,7 @@ u16 etnaviv_buffer_init(struct etnaviv_gpu *gpu)
> >   	/* initialize buffer */
> >   	buffer->user_size = 0;
> >   
> > -	CMD_WAIT(buffer);
> > +	CMD_WAIT(buffer, gpu->fe_waitcycles);
> >   	CMD_LINK(buffer, 2,
> >   		 etnaviv_cmdbuf_get_va(buffer, &gpu->mmu_context->cmdbuf_mapping)
> >   		 + buffer->user_size - 4);
> > @@ -320,7 +321,7 @@ void etnaviv_sync_point_queue(struct etnaviv_gpu *gpu, unsigned int event)
> >   	CMD_END(buffer);
> >   
> >   	/* Append waitlink */
> > -	CMD_WAIT(buffer);
> > +	CMD_WAIT(buffer, gpu->fe_waitcycles);
> >   	CMD_LINK(buffer, 2,
> >   		 etnaviv_cmdbuf_get_va(buffer, &gpu->mmu_context->cmdbuf_mapping)
> >   		 + buffer->user_size - 4);
> > @@ -503,7 +504,7 @@ void etnaviv_buffer_queue(struct etnaviv_gpu *gpu, u32 exec_state,
> >   
> >   	CMD_LOAD_STATE(buffer, VIVS_GL_EVENT, VIVS_GL_EVENT_EVENT_ID(event) |
> >   		       VIVS_GL_EVENT_FROM_PE);
> > -	CMD_WAIT(buffer);
> > +	CMD_WAIT(buffer, gpu->fe_waitcycles);
> >   	CMD_LINK(buffer, 2,
> >   		 etnaviv_cmdbuf_get_va(buffer, &gpu->mmu_context->cmdbuf_mapping)
> >   		 + buffer->user_size - 4);
> > diff --git a/drivers/gpu/drm/etnaviv/etnaviv_gpu.c b/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
> > index 41aab1aa330b..8c20dff32240 100644
> > --- a/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
> > +++ b/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
> > @@ -493,6 +493,13 @@ static void etnaviv_gpu_update_clock(struct etnaviv_gpu *gpu)
> >   		clock |= VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
> >   		etnaviv_gpu_load_clock(gpu, clock);
> >   	}
> > +
> > +	/*
> > +	 * Choose number of wait cycles to target a ~30us (1/32768) max latency
> > +	 * until new work is picked up by the FE when it polls in the idle loop.
> > +	 */
> > +	gpu->fe_waitcycles = min(gpu->base_rate_core >> (15 - gpu->freq_scale),
> > +				 0xffffUL);
> 
> This patch is NOT effective on our hardware GC1000 v5037 (ls7a1000 + 
> ls3a5000).
> 
> As the gpu->base_rate_core is 0,  so, in the end gpu->fe_waitcycles is 
> also zero.
> 
Uh, that's a problem, as the patch will then have the opposite effect
on your platform by speeding up the idle loop. Thanks for catching
this! I'll improve the patch to keep a reasonable amount of wait cycles
in this case.

Regards,
Lucas

> But after apply this path, the glmark2 still run happily, no influence. So
> 
> 
> Tested-by: Sui Jingfeng <suijingfeng@loongson.cn>
> 
> >   }
> >   
> >   static int etnaviv_hw_reset(struct etnaviv_gpu *gpu)
> > diff --git a/drivers/gpu/drm/etnaviv/etnaviv_gpu.h b/drivers/gpu/drm/etnaviv/etnaviv_gpu.h
> > index 98c6f9c320fc..e1e1de59c38d 100644
> > --- a/drivers/gpu/drm/etnaviv/etnaviv_gpu.h
> > +++ b/drivers/gpu/drm/etnaviv/etnaviv_gpu.h
> > @@ -150,6 +150,7 @@ struct etnaviv_gpu {
> >   	struct clk *clk_shader;
> >   
> >   	unsigned int freq_scale;
> > +	unsigned int fe_waitcycles;
> >   	unsigned long base_rate_core;
> >   	unsigned long base_rate_shader;
> >   };
>