> On Sat, 16 Nov 2019 13:36:30 +0200
> Lorenzo Bianconi <lorenzo@kernel.org> wrote:
> 
> > > On Fri, 15 Nov 2019 21:01:38 +0200
> > > Lorenzo Bianconi <lorenzo@kernel.org> wrote:
> > >   
> > > >  static bool __page_pool_recycle_into_ring(struct page_pool *pool,
> > > > -				   struct page *page)
> > > > +					  struct page *page,
> > > > +					  unsigned int dma_sync_size)
> > > >  {
> > > >  	int ret;
> > > >  	/* BH protection not needed if current is serving softirq */
> > > > @@ -264,6 +285,9 @@ static bool __page_pool_recycle_into_ring(struct page_pool *pool,
> > > >  	else
> > > >  		ret = ptr_ring_produce_bh(&pool->ring, page);
> > > >  
> > > > +	if (ret == 0 && (pool->p.flags & PP_FLAG_DMA_SYNC_DEV))
> > > > +		page_pool_dma_sync_for_device(pool, page, dma_sync_size);
> > > > +
> > > >  	return (ret == 0) ? true : false;
> > > >  }  
> > > 
> > > 
> > > I do wonder if we should DMA-sync-for-device BEFORE putting page into
> > > ptr_ring, as this is a channel between several concurrent CPUs.  
> > 
> > Hi Jesper,
> > 
> > in this way we can end up syncing the DMA page even if it is unmapped in
> > __page_pool_clean_page (e.g. if the ptr_ring is full), right?
> 
> Yes.  The call __page_pool_clean_page() will do a dma_unmap_page, so it
> should still be safe/correct.   I can see, that it is not optimal
> performance wise, in-case the ptr_ring is full, as DMA-sync-for-device
> is wasted work.
> 
> I don't know if you can find an argument, that proves that it cannot
> happen, that a remote CPU can dequeue/consume the page from ptr_ring
> and give it to the device, while you (the CPU the enqueued) are still
> doing the DMA-sync-for-device.

right, I can see it now :)

> 
>  
> > > > @@ -273,18 +297,22 @@ static bool __page_pool_recycle_into_ring(struct page_pool *pool,
> > > >   * Caller must provide appropriate safe context.
> > > >   */
> > > >  static bool __page_pool_recycle_direct(struct page *page,
> > > > -				       struct page_pool *pool)
> > > > +				       struct page_pool *pool,
> > > > +				       unsigned int dma_sync_size)
> > > >  {
> > > >  	if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE))
> > > >  		return false;
> > > >  
> > > >  	/* Caller MUST have verified/know (page_ref_count(page) == 1) */
> > > >  	pool->alloc.cache[pool->alloc.count++] = page;
> > > > +
> > > > +	if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
> > > > +		page_pool_dma_sync_for_device(pool, page, dma_sync_size);
> > > >  	return true;
> > > >  }  
> > > 
> > > We know __page_pool_recycle_direct() is concurrency safe, and only a
> > > single (NAPI processing) CPU can enter. (So, the DMA-sync order is not
> > > wrong here, but it could be swapped)  
> > 
> > do you mean move it before putting the page in the cache?
> > 
> > pool->alloc.cache[pool->alloc.count++] = page;
> 
> Yes, but here the order doesn't matter.
> 
> If you choose to do the DMA-sync-for-device earlier/before, then look
> at the code, and see of it makes sense to do it in __page_pool_put_page() ?
> (I've not checked the details)

I guess we can move page_pool_dma_sync_for_device() before
__page_pool_recycle_direct and __page_pool_recycle_into_ring since even if
__page_pool_put_page is not running in NAPI context or if alloc.cache is full
we will end up calling page_pool_dma_sync_for_device as first task in
__page_pool_recycle_into_ring. I will fix in v4.

Regards,
Lorenzo

> 
> -- 
> Best regards,
>   Jesper Dangaard Brouer
>   MSc.CS, Principal Kernel Engineer at Red Hat
>   LinkedIn: http://www.linkedin.com/in/brouer
>