[PATCH] drm/etnaviv: try harder to dump core

[Lucas Stach]

Hi Russell,

Am Dienstag, den 17.07.2018, 17:02 +0100 schrieb Russell King:
> Try a bit harder to produce a devcoredump when things go wrong.  If we
> fail to allocate the memory for a dump including the actual bo contents,
> omit the bos and try again.  Capturing some information from the GPU is
> better than having no information.

Agreed. One nit below.

> Signed-off-by: Russell King <rmk+kernel at armlinux.org.uk>
> ---
>  drivers/gpu/drm/etnaviv/etnaviv_dump.c | 30 ++++++++++++++++++++++--------
>  1 file changed, 22 insertions(+), 8 deletions(-)
> 
> diff --git a/drivers/gpu/drm/etnaviv/etnaviv_dump.c b/drivers/gpu/drm/etnaviv/etnaviv_dump.c
> index 48aef6cf6a42..c9987bc38acc 100644
> --- a/drivers/gpu/drm/etnaviv/etnaviv_dump.c
> +++ b/drivers/gpu/drm/etnaviv/etnaviv_dump.c
> @@ -129,6 +129,7 @@ void etnaviv_core_dump(struct etnaviv_gpu *gpu)
> >  	unsigned int n_obj, n_bomap_pages;
> >  	size_t file_size, mmu_size;
> >  	__le64 *bomap, *bomap_start;
> > +	bool bos = true;
>  
> >  	/* Only catch the first event, or when manually re-armed */
> >  	if (!etnaviv_dump_core)
> @@ -137,6 +138,7 @@ void etnaviv_core_dump(struct etnaviv_gpu *gpu)
>  
> >  	mmu_size = etnaviv_iommu_dump_size(gpu->mmu);
>  
> +again:
> >  	/* We always dump registers, mmu, ring and end marker */
> >  	n_obj = 4;
> >  	n_bomap_pages = 0;
> @@ -159,9 +161,12 @@ void etnaviv_core_dump(struct etnaviv_gpu *gpu)
> >  			continue;
>  
> >  		obj = vram->object;
> > -		file_size += obj->base.size;
> >  		n_bomap_pages += obj->base.size >> PAGE_SHIFT;
> > -		n_obj++;
> +
> > +		if (bos) {
> > +			file_size += obj->base.size;
> > +			n_obj++;
> > +		}
> >  	}
>  
> >  	/* If we have any buffer objects, add a bomap object */
> @@ -177,6 +182,11 @@ void etnaviv_core_dump(struct etnaviv_gpu *gpu)
> >  	iter.start = __vmalloc(file_size, GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
> >  			       PAGE_KERNEL);
> >  	if (!iter.start) {
> > +		if (bos) {
> > +			dev_warn(gpu->dev, "devcoredump too big, trying without bos\n");
> > +			bos = false;
> > +			goto again;
> > +		}
> >  		dev_warn(gpu->dev, "failed to allocate devcoredump file\n");
> >  		return;
> >  	}
> @@ -234,14 +244,18 @@ void etnaviv_core_dump(struct etnaviv_gpu *gpu)
> >  				*bomap++ = cpu_to_le64(page_to_phys(*pages++));
> >  		}
>  
> > -		iter.hdr->iova = cpu_to_le64(vram->iova);
> > +		if (bos) {
> > +			iter.hdr->iova = cpu_to_le64(vram->iova);
>  
> > -		vaddr = etnaviv_gem_vmap(&obj->base);
> > -		if (vaddr)
> > -			memcpy(iter.data, vaddr, obj->base.size);
> > +			vaddr = etnaviv_gem_vmap(&obj->base);
> > +			dev_crit(gpu->dev, "Copying object %p ops %pS vaddr %p to %p size %zu\n",
> +				obj, obj->ops, vaddr, iter.data, obj->base.size);

Did you mean to include this output in the patch? I guess it makes the
etnaviv coredump much more noisy in the system logs and doesn't carry
much useful info for the regular user.

Regards,
Lucas

> +			if (vaddr)
> > +				memcpy(iter.data, vaddr, obj->base.size);
>  
> > -		etnaviv_core_dump_header(&iter, ETDUMP_BUF_BO, iter.data +
> > -					 obj->base.size);
> > +			etnaviv_core_dump_header(&iter, ETDUMP_BUF_BO, iter.data +
> > +						 obj->base.size);
> > +		}
> >  	}
>  
> >  	etnaviv_core_dump_header(&iter, ETDUMP_BUF_END, iter.data);