This overwritten previous if_setsendqlen,
if_setsendqlen(ifp, ifqmaxlen);

Small nits, align member sc_ifp

I think we need another API for dealing with the ifq_drv_maxlen.

Missing ! operator

Missing ! operator

I'd do this in D38581, and it should be MFCed as well.

Missing ! operator.

Hmm... looking at this again, I *think* the first if_setsendqlen is unnecessary; ifqmaxlen defaults to 50, so what's being said here is that the driver can handle 10x more than the input queue, which seems quite bizarre. @hselasky, or @kevlo (driver author) can you comment on that?

USB host controllers usually cannot interrupt more than 8000 times per second (High Speed and above), and queues should be large enough to buffer up to 5 Gbps worth of traffic for 125us (1/8000 second). Does that make any sense to you?

Kind of, but I don't understand the disparity between the stack-side queue and the driver queue in this case. Would it make sense for them both to be 512, or is 50 stack and 512 driver really the right way?

I checked what re(4) uses, under the assumption that it's likely comparable to ure(4) (I could be very wrong on this assumption, though), and it uses 512 for both queues.

ure(4) has special capabilities to queue multiple ethernet packets per USB host controller IRQ, and should use 512, while the others are more simple and only do one ethernet packet per IRQ. Basically the current interface queue lengths should be preserved.

Yes, the previous setting of ifqmaxlen should just be removed.

It's an artifact of mechanical replacement, as my diff that set to 512 was:

IFQ_SET_MAXLEN(&ifp->if_snd, ifqmaxlen);

• ifp->if_snd.ifq_drv_maxlen = ifqmaxlen;

+ /*
+ * Try to keep two transfers full at a time.
+ * ~(TRANSFER_SIZE / 80 bytes/pkt * 2 buffers in flight)
+ */
+ ifp->if_snd.ifq_drv_maxlen = 512;

as FreeBSD doesn't have documentation on how to set queue length, I just modified the previous code that was there when the driver was originally imported.

As hselasky says, we need to keep a queue length of 512 in order to keep small packets streaming to the device. The max transfer size is 16k, so 16k/80 = 204.8, so * 2 and rounded up slightly is 512.

This should max out 1Gbps devices. I haven't trieds 2.5Gbps or 5Gbps devices to see if this or transfer size needs to be increased.