Re: [PATCH V3] iwlwifi: use paged Rx

[Zhu Yi <yi.zhu@intel.com>]

On Mon, 2009-10-12 at 22:20 +0800, Stanislaw Gruszka wrote:
> On Fri, Oct 09, 2009 at 05:19:45PM +0800, Zhu Yi wrote:
> > This switches the iwlwifi driver to use paged skb from linear skb for Rx
> > buffer. So that it relieves some Rx buffer allocation pressure for the
> > memory subsystem. Currently iwlwifi (4K for 3945) requests 8K bytes for
> > Rx buffer. Due to the trailing skb_shared_info in the skb->data,
> > alloc_skb() will do the next order allocation, which is 16K bytes. This
> > is suboptimal and more likely to fail when the system is under memory
> > usage pressure. Switching to paged Rx skb lets us allocate the RXB
> > directly by alloc_pages(), so that only order 1 allocation is required.
> 
> [snip]
> 
> > Finally, mac80211 doesn't support paged Rx yet. So we linearize the skb
> > for all the management frames and software decryption or defragmentation
> > required data frames before handed to mac80211. For all the other frames,
> > we __pskb_pull_tail 64 bytes in the linear area of the skb for mac80211
> > to handle them properly.
> 
> This seems to be big overhead, but since there is no way to avoid it ...

Which case? The linear one is the same as the current implementation.
For __pskb_pull_tail, it is guaranteed no new buffer will be allocated.

> If we know that we need to linearize we can alloc as big skb as needed,
> otherwise skb_linearize() need to do reallocation.

OK.

> Can we also remove IWL_LINK_HDR_MAX and do __pskb_pull_tail based on 
> real header(s) size ? 

The wireless header is a variant depending on type. And mac80211 also
needs to play with LLC and IP header (for qos). So I used the max
possible frame buffer (128 or probably 64 is enough?) mac80211 is going
to use for none software decryption and defragmentation data frames.

> Or.
> 
> If we decide to do alloc_skb(IWL_LINK_HDR_MAX, gfp_flags) can this be 
> done together with skb_add_rx_frag() in iwl_rx_allocate(), to offload
> this interrupt context ?

I'm not sure if it's a big gain. This is only a 128 bytes GFP_ATOMIC
allocation anyway. Other driver like niu also does this.

> > +	    le16_to_cpu(hdr->seq_ctrl) & IEEE80211_SCTL_FRAG)
> 
> Minor optimization:
> 	    hdr->seq_ctrl & cpu_to_le16(IEEE80211_SCTL_FRAG)

The original code returns the fragment index. The change makes it
optimized on LE arches. But is less readable.

> >  struct iwl_rx_mem_buffer {
> > -	dma_addr_t real_dma_addr;
> > -	dma_addr_t aligned_dma_addr;
> > -	struct sk_buff *skb;
> > +	dma_addr_t page_dma;
> > +	struct page *page;
> >  	struct list_head list;
> >  };
> >  
> > +#define rxb_addr(r) page_address(r->page)
> 
> Since we mostly use pointer, perhaps better would be save address of page
> in iwl_rx_mem_buffer, and use virt_to_page where struct page is needed.

Both should be fine. But I'd like to access the virtual address of
rxb->page with a micro like rxb_addr than something like "pkt =
rxb->virt_page" directly.

> >  			    net_ratelimit())
> > -				IWL_CRIT(priv, "Failed to allocate SKB buffer with %s. Only %u free buffers remaining.\n",
> > +				IWL_CRIT(priv, "Failed to alloc_pages with %s. Only %u free buffers remaining.\n",
> >  					 priority == GFP_ATOMIC ?  "GFP_ATOMIC" : "GFP_KERNEL",
> 
> priority can not be equal GFP_ATOMIC if was or'ed with __GFP_COMP or __GFP_NOWARN

Good catch. The bug also exists in the original code. Will send out a
separated patch to fix.

Thanks,
-yi