On Tue, 2004-07-06 at 23:29, Jeff Garzik wrote:
On Tue, Jul 06, 2004 at 11:18:02PM -0400, jamal wrote:
You dont return a 1 anywhere.
That OK in one model.
True returning 0 this is not wrong; it
results in an extra call in the layer above the driver.
(I was trying to point that out in earlier email)
er, I'm confused now?
Every single ethernet driver returns zero, when it has queued a packet
to hardware :) That's the common case, I would hope it doesn't result
in additional work.
When you are not dealing with fragments, the most optimal model
eliminates the overflow case completely, so your ->hard_start_xmit looks
queue packet to DMA ring
if (DMA ring full)
If you can be sure -- by design -- that room is always available when
the queue is not stopped, then that's fine.
With fragments, you cannot be sure of this, if you do not wish to
reserve MY_HW_MAX_FRAGMENTS slots on the DMA. Such a case would require
moving the "if no more descriptors" check up, and returning 1 when the
ring is empty.
But ideally, you should write the driver where such a condition does not
occur at all.
Ok, I overlooked fragments. I think it would be useful to capture this
in the doc you were preping. BTW, why can you figure out the fragment
count? If you can then the check for number of descriptors availabel
could account for that.
In one design, you can say
if (free descriptors < MAX_SKB_FRAGS)
That design wastes descriptors, but ensures you always have enough room
when ->hard_start_xmit is called, and thus ensures you never have to
Another design, that attempts to use more descriptors, is
if (free descriptors < skb->frags)
if (free descriptors == 0)