Re: [Intel-gfx] [PATCH 2/9] drm/i915: Use kmap_local_page() in gem/i915_gem_pyhs.c

["Fabio M. De Francesco" <fmdefrancesco@gmail.com>]

On lunedì 17 ottobre 2022 11:37:18 CEST Zhao Liu wrote:
> From: Zhao Liu <zhao1.liu@intel.com>
> 
> The use of kmap_atomic() is being deprecated in favor of
> kmap_local_page()[1].
> 
> The main difference between atomic and local mappings is that local
> mappings doesn't disable page faults or preemption.
> 
> In drm/i915/gem/i915_gem_phys.c, the functions
> i915_gem_object_get_pages_phys() and i915_gem_object_put_pages_phys()
> don't need to disable pagefaults and preemption for mapping because of
> these 2 reasons:
> 
> 1. The flush operation is safe for CPU hotplug when preemption is not
> disabled. In drm/i915/gem/i915_gem_object.c, the functions
> i915_gem_object_get_pages_phys() and i915_gem_object_put_pages_phys()
> calls drm_clflush_virt_range() to use CLFLUSHOPT or WBINVD to flush.
> Since CLFLUSHOPT is global on x86 and WBINVD is called on each cpu in
> drm_clflush_virt_range(), the flush operation is global and any issue
> with cpu's being added or removed can be handled safely.
> 
> 2. Any context switch caused by preemption or sleep (pagefault may
> cause sleep) doesn't affect the validity of local mapping.
> 
> Therefore, i915_gem_object_get_pages_phys() and
> i915_gem_object_put_pages_phys() are two functions where the use of
> kmap_local_page() in place of kmap_atomic() is correctly suited.
> 
> Convert the calls of kmap_atomic() / kunmap_atomic() to
> kmap_local_page() / kunmap_local().
> 

I have here the same questions as in 1/9.

> [1]: https://lore.kernel.org/all/20220813220034.806698-1-ira.weiny@intel.com
> 
> Suggested-by: Dave Hansen <dave.hansen@intel.com>
> Suggested-by: Ira Weiny <ira.weiny@intel.com>
> Suggested-by: Fabio M. De Francesco <fmdefrancesco@gmail.com>
> Signed-off-by: Zhao Liu <zhao1.liu@intel.com>
> ---
> Suggested by credits:
>   Dave: Referred to his explanation about cache flush.
>   Ira: Referred to his task document, review comments and explanation about
>        cache flush.
>   Fabio: Referred to his boiler plate commit message.
> ---
>  drivers/gpu/drm/i915/gem/i915_gem_phys.c | 8 ++++----
>  1 file changed, 4 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_phys.c
> b/drivers/gpu/drm/i915/gem/i915_gem_phys.c index 0d0e46dae559..d602ba19ecb2 
100644
> --- a/drivers/gpu/drm/i915/gem/i915_gem_phys.c
> +++ b/drivers/gpu/drm/i915/gem/i915_gem_phys.c
> @@ -66,10 +66,10 @@ static int i915_gem_object_get_pages_phys(struct 
drm_i915_gem_object
> *obj) if (IS_ERR(page))
>  			goto err_st;
> 
> -		src = kmap_atomic(page);
> +		src = kmap_local_page(page);
>  		memcpy(dst, src, PAGE_SIZE);
>  		drm_clflush_virt_range(dst, PAGE_SIZE);
> -		kunmap_atomic(src);
> +		kunmap_local(src);

Please use memcpy_from_page() instead of open coding mapping + memcpy() + 
unmapping.

> 
>  		put_page(page);
>  		dst += PAGE_SIZE;
> @@ -114,10 +114,10 @@ i915_gem_object_put_pages_phys(struct 
drm_i915_gem_object *obj,
>  			if (IS_ERR(page))
>  				continue;
> 
> -			dst = kmap_atomic(page);
> +			dst = kmap_local_page(page);
>  			drm_clflush_virt_range(src, PAGE_SIZE);
>  			memcpy(dst, src, PAGE_SIZE);
> -			kunmap_atomic(dst);
> +			kunmap_local(dst);

For the same reasons said above, memcpy_to_page() should be used here and 
avoid open coding of three functions.

Using those helpers forces you to move drm_clflush_virt_range() out of the 
mapping / un-mapping region. I may be wrong, however I'm pretty sure that the 
relative positions of each of those call sites is something that cannot be 
randomly chosen.

Thanks,

Fabio

> 
>  			set_page_dirty(page);
>  			if (obj->mm.madv == I915_MADV_WILLNEED)