Re: [PATCH] ravb: Use GFP_KERNEL instead of GFP_ATOMIC when possible

[Christophe JAILLET]

Le 20/02/2022 à 08:53, Biju Das a écrit :
> Hi Christophe,
>
> Thanks for the patch.
>
> Just a  question, As per [1], former can be allocated from interrupt context.
> But nothing mentioned for the allocation using the patch you mentioned[2]. I agree GFP_KERNEL
> gives more opportunities of successful allocation.

Hi,

netdev_alloc_skb() uses an implicit GFP_ATOMIC, that is why it can be 
safely called from an interrupt context.
__netdev_alloc_skb() is the same as netdev_alloc_skb(), except that you 
can choose the GFP flag you want to use. ([1])

Here, the netdev_alloc_skb() is called just after some 
"kcalloc(GFP_KERNEL);"

So this function can already NOT be called from interrupt context.

So if GFP_KERNEL is fine here for kcalloc(), it is fine also for 
netdev_alloc_skb(), hence __netdev_alloc_skb(GFP_KERNEL).

> Q1) Here it allocates 8K instead of 1K on each loop, Is there any limitation for netdev_alloc_skb for allocating 8K size?

Not sure to understand.
My patch does NOT change anything on the amount of memory allocated. it 
only changes a GFP_ATOMIC into a GFP_KERNEL.

I'm not aware of specific limitation for netdev_alloc_skb().
My understanding is that in the worst case, it will behave just like 
malloc() ([3])

So, if it was an issue before, it is still an issue after my patch.

> Q2) In terms of allocation performance which is better netdev_alloc_skb or __netdev_alloc_skb?

AFAIK, there should be no difference, but __netdev_alloc_skb(GFP_KERNEL) 
can succeed where netdev_alloc_skb() can fail. In such a case, it would 
be slower but most importantly, it would succeed.


CJ

[1]: 
https://elixir.bootlin.com/linux/v5.17-rc4/source/include/linux/skbuff.h#L2945

[2]: 
https://elixir.bootlin.com/linux/v5.17-rc4/source/drivers/net/ethernet/renesas/ravb_main.c#L470

[3]: 
https://elixir.bootlin.com/linux/v5.17-rc3/source/net/core/skbuff.c#L488


> [1] https://www.kernel.org/doc/htmldocs/networking/API-netdev-alloc-skb.html
> [2] https://www.kernel.org/doc/htmldocs/networking/API---netdev-alloc-skb.html
>
> Regards,
> Biju
>
> > Subject: [PATCH] ravb: Use GFP_KERNEL instead of GFP_ATOMIC when possible
> >
> > 'max_rx_len' can be up to GBETH_RX_BUFF_MAX (i.e. 8192) (see
> > 'gbeth_hw_info').
> > The default value of 'num_rx_ring' can be BE_RX_RING_SIZE (i.e. 1024).
> >
> > So this loop can allocate 8 Mo of memory.
> >
> > Previous memory allocations in this function already use GFP_KERNEL, so
> > use __netdev_alloc_skb() and an explicit GFP_KERNEL instead of a implicit
> > GFP_ATOMIC.
> >
> > This gives more opportunities of successful allocation.
> >
> > Signed-off-by: Christophe JAILLET <christophe.jaillet@xxxxxxxxxx>
> > ---
> >   drivers/net/ethernet/renesas/ravb_main.c | 2 +-
> >   1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/drivers/net/ethernet/renesas/ravb_main.c
> > b/drivers/net/ethernet/renesas/ravb_main.c
> > index 24e2635c4c80..525d66f71f02 100644
> > --- a/drivers/net/ethernet/renesas/ravb_main.c
> > +++ b/drivers/net/ethernet/renesas/ravb_main.c
> > @@ -475,7 +475,7 @@ static int ravb_ring_init(struct net_device *ndev, int
> > q)
> >   		goto error;
> >
> >   	for (i = 0; i < priv->num_rx_ring[q]; i++) {
> > -		skb = netdev_alloc_skb(ndev, info->max_rx_len);
> > +		skb = __netdev_alloc_skb(ndev, info->max_rx_len, GFP_KERNEL);
> >   		if (!skb)
> >   			goto error;
> >   		ravb_set_buffer_align(skb);
> > --
> > 2.32.0