Messages in this thread | | | Date | Wed, 6 Jun 2012 14:36:02 -0400 | From | Chris Metcalf <> | Subject | Re: [PATCH v9] tilegx network driver: initial support |
| |
On 6/6/2012 1:40 PM, Eric Dumazet wrote: > On Mon, 2012-06-04 at 16:12 -0400, Chris Metcalf wrote: > >> +/* Allocate and push a buffer. */ >> +static bool tile_net_provide_buffer(bool small) >> +{ >> + int stack = small ? small_buffer_stack : large_buffer_stack; >> + const unsigned long buffer_alignment = 128; >> + struct sk_buff *skb; >> + int len; >> + >> + len = sizeof(struct sk_buff **) + buffer_alignment; >> + len += (small ? 128 : 1664); > 1664 is a magic number, it should be a nice define > > #define ..... ( ETH_DATA_LEN + .... )
Fair enough. However, the magic-ness comes from the hardware header code in arch/tile/gxio/mpipe.h, which provides a limited set of allowed buffer sizes, including 1664. But I can add these #defines at the top of this driver:
/* Buffer sizes and mpipe enum codes for buffer stacks. * See arch/tile/include/gxio/mpipe.h for the set of possible values. */ #define BUFFER_SIZE_SMALL_ENUM GXIO_MPIPE_BUFFER_SIZE_128 #define BUFFER_SIZE_SMALL 128 #define BUFFER_SIZE_LARGE_ENUM GXIO_MPIPE_BUFFER_SIZE_1664 #define BUFFER_SIZE_LARGE 1664
>> + skb = dev_alloc_skb(len); >> + if (skb == NULL) >> + return false; >> + >> + /* Make room for a back-pointer to 'skb' and guarantee alignment. */ >> + skb_reserve(skb, sizeof(struct sk_buff **)); >> + skb_reserve(skb, -(long)skb->data & (buffer_alignment - 1)); >> + >> + /* Save a back-pointer to 'skb'. */ >> + *(struct sk_buff **)(skb->data - sizeof(struct sk_buff **)) = skb; >> + >> + /* Make sure "skb" and the back-pointer have been flushed. */ >> + wmb(); > Interesting, have you considered using build_skb() instead of this > convoluted thing ? > > This could save some cache misses...
I hadn't looked at build_skb() before; we built up this driver mostly on a base of 2.6.38, where it doesn't exist. That said, it doesn't seem like it matters; dev_alloc_skb() will just end up calling down to build_skb() anyway, as far as I can tell.
The code where we do the two skb_reserves and then stuff in a backpointer and do a barrier are because we track the skbuffs in hardware, and hardware ignores the low 7 bits aof the address (thus the "buffer_alignment" part) and we need to be able to pull the actual skb address out of the data when the hardware returns a pointer to the data to us.
By the way, your question about tx_queue_len is a good one; I'm roping in our other network developer folks to figure it out. Originally it was a performance optimization, I believe; I'm not sure it's still required. I'll follow up on that one when we've tracked it down.
-- Chris Metcalf, Tilera Corp. http://www.tilera.com
| |