[PATCH 3/3] net: hisilicon: add hix5hd2 mac driver

[Arnd Bergmann]

On Tuesday 27 May 2014, zhangfei wrote:
> On 05/26/2014 10:51 PM, Arnd Bergmann wrote:
> > On Monday 19 May 2014, Zhangfei Gao wrote:
> >
> > I only noticed one real issue with the driver:
> >
> >> +struct hix5hd2_desc {
> >> +    __le32 buff_addr;
> >> +    __le32 buff_len:11;
> >> +    __le32 reserve2:5;
> >> +    __le32 data_len:11;
> >> +    __le32 reserve1:2;
> >> +    __le32 fl:2;
> >> +    __le32 descvid:1;
> >> +} __aligned(32);
> >> +
> >
> > You should generall not use bitfields in hardware data structures, as that is
> > not endian safe and will prevent running a big-endian kernel on this machine.
> > Better convert this to a set of __le32 fields and explicit shifts and masks.
> 
> Got it, will update.
> 
> More knowledge about big-endian kernel is appreciated, in which case we 
> should consider such kernel.
> Can we only consider this driver is only running on arm, which is 
> little-endian.

The main user of big-endian kernels that I know of are network infrastructure
people that have a ton of legacy user-space code written in non-portable
big-endian style.

There is no real technical advantage in running one endianess or the other
on ARM, so generally everyone uses little-endian because that is what their
user space is, except for the few people that have ported over their
code from PowerPC or MIPSeb.

> > Two smaller things you should think about, I'm not entirely sure about these:
> >
> >> +static int hix5hd2_rx(struct net_device *dev, int limit)
> >> +{
> >> +    struct hix5hd2_priv *priv = netdev_priv(dev);
> >> +    struct sk_buff *skb;
> >> +    struct hix5hd2_desc *desc;
> >> +    dma_addr_t dma_addr;
> >> +    u32 start, end, num, pos, i, len;
> >> +
> >> +    /* software read pointer */
> >> +    start = dma_cnt(readl_relaxed(priv->base + RX_BQ_RD_ADDR));
> >> +    /* logic write pointer */
> >> +    end = dma_cnt(readl_relaxed(priv->base + RX_BQ_WR_ADDR));
> >
> > I think one of these needs to be readl() instead of readl_relaxed(),
> > to ensure the data is correctly ordered with regard to the pointer
> > access.
> readl_relaxed can ensure the sequence.
> 
> >
> >> +    if (pos != start)
> >> +            writel(dma_byte(pos), priv->base + TX_RQ_RD_ADDR);
> >
> > While this looks like it could be writel_relaxed().
> >
> En, I think all three cases can use xxx_relaxed.
> These accesses are just update pointer used by internal logic and get 
> pointer updated by internal logic.

You always need a barrier between the access to a buffer from the kernel
and the register access communicating with the hardware about that
buffer:

For tx:

	a) write to DMA descriptor
	b) wmb()
	c) writel_relaxed(register)

writel() is the short form of b+c. If you omit the __iowmb(), the decriptor
access may still be inside of the CPU write buffer (or the cache on coherent
systems) by the time the hardware reads the descriptor.

for rx:

	a) readl_relaxed(register)
	b) rmb()
	c) read from descriptor or skb data

If you leave out the __iormb() here, it's possible that the CPU can prefetch
the access to the descriptor while waiting for the register data, and read
stale data.

If your particular machine doesn't need rmb() or wmb() (e.g. for
non-SMP ARMv6 or older), the build process should replace them with
do{}while(0) them automatically.

	Arnd