Message-ID: <87eggosmvk.wl@dns1.atmark-techno.com>
Date:	Wed, 21 Oct 2015 16:26:39 +0900
From:	Yasushi SHOJI <yashi@...ark-techno.com>
To:	sergei.shtylyov@...entembedded.com
Cc:	netdev@...r.kernel.org
Subject: Re: sh_eth.c::sh_eth_rx(): mdp->rx_skbuff[entry] can be NULL

Hi Sergei,

Thank your for your reply.

On Wed, 21 Oct 2015 05:48:01 +0900,
Sergei Shtylyov wrote:
> 
> On 10/19/2015 06:01 PM, Yasushi SHOJI wrote:
> 
> > I'm not that familiar with this code base so I'm note including any
> > patch yet.  I appreciate if someone with insight in this code give a
> > quick look and tell me that it's a real one or not.  if this is a real
> > case, I can take a deep look.
> 
>    If you got the oops, it's real. Thanks for the reporting. I guess I
> should check the new ravb driver as well...
>    Do you want to try fixing the bug yourself?

Sure.  I can dive in to this.  I appreciate if someone who has worked
on sh_eth.c give me some design advises or tell me the initial design
thoughts / what was the intention when allocation if failed.

My idea right now is to simply invalidate the descriptor when
netdev_alloc_skb() failed.  When next packet arrived, in near future,
the driver can try again to allocate the buffer and update the
corresponding descriptor if succeeds.  If memory is not yet available
when the controller is trying to use the invalid descriptor, the
controller will see it and DMA will stop.

Is it acceptable path to go?


Here is how I understand this driver:

sh_eth.c uses napi to handle incoming packet.  It calls
__napi_schedule() in its interrupt handler.  sh_eth_poll() is the pool
method of the driver.  sh_eth_poll calls sh_eth_rx() to handle
incoming packet.  And this is the function I need to fix.

The driver utilizes array of sk_buffs for tx and rx.  For rx, the
driver has an array of pointers of sk_buffs, rx_skbuff[]. This
rx_skbuff[] is filled with sk_buffs in sh_eth_ring_format() which is
called when the driver is open()ed.

The controller, the driver is targeted to, is GETHER.  This controller
is capable to DMA transfer the received packets to the system memory.
The link between them is described in "Receive Descriptor".

A receive descriptor corresponds to one sk_buff.  The controller
expects array of descriptors in the system memory and treat it as a
ring, meaning that the controller process each descriptor one by one.
Once the controller finished the last descriptor, it will go back to
the first one.

To achieve zero copy, the driver push the sk_buffs filled with
received packet to the netdev core with netif_receive_skb() then
netdev_alloc_skb() sk_buffs in the sh_eth_rx(), the poll method of the
driver, and update the corresponding descriptor.

If the allocation failed, it just leave the function, leaving old
pointer in the descriptor as is.  In some future, the controller will
access the descriptor and writes to the old memory address. (I haven't
checked the state of bits in the descriptor yet)



BTW, if any one has a bit of time, I have questions regarding to the
atomic allocation:

  Q1) is it OK to _always_ use napi_schedule()?
  Q2) is it OK to netdev_alloc_skb() in napi poll method? (sh_eth_rx)
      this, combined with Q1, leads to allocate sk_buff with GFP_ATOMIC
      all the time.
  Q3) if not, any alternative way to allocate sk_buff?
      ie.
        a) allocate sk_buff in pool out of interrupt context, and
           get sk_buff from it in softirq context.
	b) use dev_alloc_pages() instead.  The slab allocator seem to
	   fail if slabs are full but still order 0 pages are available.
-- 
             yashi
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists