[Intel-gfx] [PATCH] iosys-map: Add word-sized reads

[Thomas Zimmermann]

Hi Lucas

Am 10.06.22 um 01:20 schrieb Lucas De Marchi:
> Instead of always falling back to memcpy_fromio() for any size, prefer
> using read{b,w,l}(). When reading struct members it's common to read
> individual integer variables individually. Going through memcpy_fromio()
> for each of them poses a high penalty.
> 
> Employ a similar trick as __seqprop() by using _Generic() to generate
> only the specific call based on a type-compatible variable.
> 
> For a pariticular i915 workload producing GPU context switches,
> __get_engine_usage_record() is particularly hot since the engine usage
> is read from device local memory with dgfx, possibly multiple times
> since it's racy. Test execution time for this test shows a ~12.5%
> improvement with DG2:
> 
> Before:
> 	nrepeats = 1000; min = 7.63243e+06; max = 1.01817e+07;
> 	median = 9.52548e+06; var = 526149;
> After:
> 	nrepeats = 1000; min = 7.03402e+06; max = 8.8832e+06;
> 	median = 8.33955e+06; var = 333113;
> 
> Other things attempted that didn't prove very useful:
> 1) Change the _Generic() on x86 to just dereference the memory address
> 2) Change __get_engine_usage_record() to do just 1 read per loop,
>     comparing with the previous value read
> 3) Change __get_engine_usage_record() to access the fields directly as it
>     was before the conversion to iosys-map
> 
> (3) did gave a small improvement (~3%), but doesn't seem to scale well
> to other similar cases in the driver.
> 
> Additional test by Chris Wilson using gem_create from igt with some
> changes to track object creation time. This happens to accidentaly
> stress this code path:
> 
> 	Pre iosys_map conversion of engine busyness:
> 	lmem0: Creating    262144 4KiB objects took 59274.2ms
> 
> 	Unpatched:
> 	lmem0: Creating    262144 4KiB objects took 108830.2ms
> 
> 	With readl (this patch):
> 	lmem0: Creating    262144 4KiB objects took 61348.6ms
> 
> 	s/readl/READ_ONCE/
> 	lmem0: Creating    262144 4KiB objects took 61333.2ms
> 
> So we do take a little bit more time than before the conversion, but
> that is due to other factors: bringing the READ_ONCE back would be as
> good as just doing this conversion.
> 
> Signed-off-by: Lucas De Marchi <lucas.demarchi at intel.com>
> ---
> 
> If this is acceptable we should probably add the write counterpart, too.
> Sending here only the read for now since this fixes the issue we are
> seeing and to gather feedback.

That would not be a problem, but please only add functions that you use.

> 
>   include/linux/iosys-map.h | 26 ++++++++++++++++++++++----
>   1 file changed, 22 insertions(+), 4 deletions(-)
> 
> diff --git a/include/linux/iosys-map.h b/include/linux/iosys-map.h
> index e69a002d5aa4..4ae3e459419e 100644
> --- a/include/linux/iosys-map.h
> +++ b/include/linux/iosys-map.h
> @@ -333,6 +333,20 @@ static inline void iosys_map_memset(struct iosys_map *dst, size_t offset,
>   		memset(dst->vaddr + offset, value, len);
>   }
>   
> +#ifdef CONFIG_64BIT
> +#define __iosys_map_rd_io_u64_case(val_, vaddr_iomem_)			\
> +	u64: val_ = readq(vaddr_iomem_),
> +#else
> +#define __iosys_map_u64_case(val_, vaddr_iomem_)
> +#endif
> +
> +#define __iosys_map_rd_io(val__, vaddr_iomem__, type__) _Generic(val__,	\
> +	u8: val__ = readb(vaddr_iomem__),				\
> +	u16: val__ = readw(vaddr_iomem__),				\
> +	u32: val__ = readl(vaddr_iomem__),				\
> +	__iosys_map_rd_io_u64_case(val__, vaddr_iomem__)		\
> +	default: memcpy_fromio(&(val__), vaddr_iomem__, sizeof(val__)))
> +
>   /**
>    * iosys_map_rd - Read a C-type value from the iosys_map
>    *
> @@ -346,10 +360,14 @@ static inline void iosys_map_memset(struct iosys_map *dst, size_t offset,
>    * Returns:
>    * The value read from the mapping.
>    */
> -#define iosys_map_rd(map__, offset__, type__) ({			\
> -	type__ val;							\
> -	iosys_map_memcpy_from(&val, map__, offset__, sizeof(val));	\
> -	val;								\
> +#define iosys_map_rd(map__, offset__, type__) ({				\
> +	type__ val;								\
> +	if ((map__)->is_iomem) {						\
> +		__iosys_map_rd_io(val, (map__)->vaddr_iomem + offset__, type__);\
> +	} else {								\
> +		memcpy(&val, (map__)->vaddr + offset__, sizeof(val));		\
> +	}									\
> +	val;									\

To my knowledge, calls to readw/readl have alignment requirements on 
some platforms, while memcpy_fromio() has none. Mixing memcpy() and 
read*() sounds like a problem for subtle bugs. I'd prefer to at least 
mitigate that to some extend.

For each case in the _Generic statement, there should be an if/else 
branch on is_iomem. Here's the example

#define iosys_map_rd() \
_Generic( (val__),
   u8: {
     if (map__)->is_iomem
       val__ = readb()
     else
       val__ *(volatile u8*)(vaddr_iomem);
   },
   u16: {
     if (map__)->is_iomem
       val__ = readw()
     else
       val__ *(volatile u16*)(vaddr_iomem);
   },
   u32,
   u64,
   ...
   default: {
     if (map__)->is_iomem
       mempy_fromio()
     else
       memcpy()
   })

Using volatile with system memory enforces single instructions or even 
alignment on some platforms. While experimenting with framebuffer 
updates, I've also found this to be faster then regular code. With 
'volatile' the compiler generated a single movq instead of a number of 
shorter movs. (But I won't promise anything. :)

Within _Generic, for each type, a macro can generate the case. Like this

#define __iosys_map_rd_case(__type, __map, __read) \
   __type: if else ...

In the case of u64, you can simply do

#if CONFIG_64BIT
#define __iosys_map_rd_case_u64(__map) \
   __iosys_map_rc_case(u64, __map, readq)
#else
#define __iosys_map_rd_case_u64(__map)
#endif

and use that macro in the _Generic. On 64-bit systems, the case will be 
there. Otherwise it will be empty.

The only user of iosys_map_rd() is i915. I quickly looked through the 
usage and found no cases where the default memcpy could be used. It's 
all structs with u32. (Right?)  If so, please remove the default case 
with memcpy entirely.  This will result in clear compile-time errors if 
a certain type is not supported.  There's still iosys_mem_memcpy() for 
those who need it.

Best regards
Thomas

>   })
>   
>   /**

-- 
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5, 90409 Nürnberg, Germany
(HRB 36809, AG Nürnberg)
Geschäftsführer: Ivo Totev
-------------- next part --------------
A non-text attachment was scrubbed...
Name: OpenPGP_signature
Type: application/pgp-signature
Size: 840 bytes
Desc: OpenPGP digital signature
URL: <https://lists.freedesktop.org/archives/intel-gfx/attachments/20220613/35486d45/attachment.sig>