Re: [PATCH RFC net-next 2/4] net: page_pool: avoid calling no-op externals when possible

[Alexander H Duyck <alexander.duyck@gmail.com>]

On Thu, 2023-06-29 at 17:23 +0200, Alexander Lobakin wrote:
> Turned out page_pool_put{,_full}_page() can burn quite a bunch of cycles
> even when on DMA-coherent platforms (like x86) with no active IOMMU or
> swiotlb, just for the call ladder.
> Indeed, it's
> 
> page_pool_put_page()
>   page_pool_put_defragged_page()                  <- external
>     __page_pool_put_page()
>       page_pool_dma_sync_for_device()             <- non-inline
>         dma_sync_single_range_for_device()
>           dma_sync_single_for_device()            <- external
>             dma_direct_sync_single_for_device()
>               dev_is_dma_coherent()               <- exit
> 
> For the inline functions, no guarantees the compiler won't uninline them
> (they're clearly not one-liners and sometimes compilers uninline even
> 2 + 2). The first external call is necessary, but the rest 2+ are done
> for nothing each time, plus a bunch of checks here and there.
> Since Page Pool mappings are long-term and for one "device + addr" pair
> dma_need_sync() will always return the same value (basically, whether it
> belongs to an swiotlb pool), addresses can be tested once right after
> they're obtained and the result can be reused until the page is unmapped.
> Define new PP flag, which will mean "do DMA syncs for device, but only
> when needed" and turn it on by default when the driver asks to sync
> pages. When a page is mapped, check whether it needs syncs and if so,
> replace that "sync when needed" back to "always do syncs" globally for
> the whole pool (better safe than sorry). As long as a pool has no pages
> requiring DMA syncs, this cuts off a good piece of calls and checks.
> On my x86_64, this gives from 2% to 5% performance benefit with no
> negative impact for cases when IOMMU is on and the shortcut can't be
> used.
> 
> Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
> ---
>  include/net/page_pool.h |  3 +++
>  net/core/page_pool.c    | 10 ++++++++++
>  2 files changed, 13 insertions(+)
> 
> diff --git a/include/net/page_pool.h b/include/net/page_pool.h
> index 829dc1f8ba6b..ff3772fab707 100644
> --- a/include/net/page_pool.h
> +++ b/include/net/page_pool.h
> @@ -23,6 +23,9 @@
>  					* Please note DMA-sync-for-CPU is still
>  					* device driver responsibility
>  					*/
> +#define PP_FLAG_DMA_MAYBE_SYNC	BIT(2) /* Internal, should not be used in
> +					* drivers
> +					*/
>  #define PP_FLAG_ALL		(PP_FLAG_DMA_MAP |\
>  				 PP_FLAG_DMA_SYNC_DEV)
>  
> diff --git a/net/core/page_pool.c b/net/core/page_pool.c
> index dff0b4fa2316..498e058140b3 100644
> --- a/net/core/page_pool.c
> +++ b/net/core/page_pool.c
> @@ -197,6 +197,10 @@ static int page_pool_init(struct page_pool *pool,
>  		/* pool->p.offset has to be set according to the address
>  		 * offset used by the DMA engine to start copying rx data
>  		 */
> +
> +		/* Try to avoid calling no-op syncs */
> +		pool->p.flags |= PP_FLAG_DMA_MAYBE_SYNC;
> +		pool->p.flags &= ~PP_FLAG_DMA_SYNC_DEV;
>  	}
>  
>  #ifdef CONFIG_PAGE_POOL_STATS
> @@ -341,6 +345,12 @@ static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
>  
>  	page_pool_set_dma_addr(page, dma);
>  
> +	if ((pool->p.flags & PP_FLAG_DMA_MAYBE_SYNC) &&
> +	    dma_need_sync(pool->p.dev, dma)) {
> +		pool->p.flags |= PP_FLAG_DMA_SYNC_DEV;
> +		pool->p.flags &= ~PP_FLAG_DMA_MAYBE_SYNC;
> +	}
> +
>  	if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
>  		page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
>  

I am pretty sure the logic is flawed here. The problem is
dma_needs_sync depends on the DMA address being used. In the worst case
scenario we could have a device that has something like a 32b DMA
address space on a system with over 4GB of memory. In such a case the
higher addresses would need to be synced because they will go off to a
swiotlb bounce buffer while the lower addresses wouldn't.

If you were to store a flag like this it would have to be generated per
page.