[intel] Apply PBS/PBA errata workaround only to ICH8 PCI device IDs
ICH8 devices have an errata which requires us to reconfigure the
packet buffer size (PBS) register, and correspondingly adjust the
packet buffer allocation (PBA) register. The "Intel I/O Controller
Hub ICH8/9/10 and 82566/82567/82562V Software Developer's Manual"
notes for the PBS register that:
10.4.20 Packet Buffer Size - PBS (01008h; R/W)
Note: The default setting of this register is 20 KB and is
incorrect. This register must be programmed to 16 KB.
Initial value: 0014h
0018h (ICH9/ICH10)
It is unclear from this comment precisely which devices require the
workaround to be applied. We currently attempt to err on the side of
caution: if we detect an initial value of either 0x14 or 0x18 then the
workaround will be applied. If the workaround is applied
unnecessarily, then the effect should be just that we use less than
the full amount of the available packet buffer memory.
Unfortunately this approach does not play nicely with other device
drivers. For example, the Linux e1000e driver will rewrite PBA while
assuming that PBS still contains the default value, which can result
in inconsistent values between the two registers, and a corresponding
inability to transmit or receive packets. Even more unfortunately,
the contents of PBS and PBA are not reset by anything less than a
power cycle, meaning that this error condition will survive a hardware
reset.
The Linux driver (written and maintained by Intel) applies the PBS/PBA
errata workaround only for devices in the ICH8 family, identified via
the PCI device ID. Adopt a similar approach, using the PCI_ROM()
driver data field to indicate when the workaround is required.
Reported-by: Donald Bindner <dbindner@truman.edu>
Debugged-by: Donald Bindner <dbindner@truman.edu>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
As of commit d28bb51 ("[tcp] Defer sending ACKs until all received
packets have been processed"), increasing the RX ring size will
increase the number of received packets per transmitted ACK (since
each poll will process up to one complete receive ring). Under KVM,
this can make a substantial (up to ~200%) difference to the overall
download speed, since transmissions are very expensive.
Increase the ring fill level from four to eight packets: this
increases the download speed by around 50% at a cost of around 8kB of
heap space. Further speedups are possible by increasing the ring size
further, but it would be preferable to find alternative methods which
do not use noticeable amounts of heap space.
Tested-by: Robin Smidsrød <robin@smidsrod.no>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
[intel] Avoid completely filling the TX descriptor ring
It is unclear from the datasheets whether or not the TX ring can be
completely filled (i.e. whether writing the tail value as equal to the
current head value will cause the ring to be treated as completely
full or completely empty). It is very plausible that this edge case
could differ in behaviour between real hardware and the many
implementations of an emulated Intel NIC found in various virtual
machines. Err on the side of caution and always leave at least one
ring entry empty.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
[intel] Expose functionality to be shared with intelx driver
The Intel 10 Gigabit NICs have a datapath that is almost
register-compatible with the Intel 1 Gigabit NICs. Expose common
functionality to avoid duplication of code in the new "intelx" driver.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
[intel] Remove hardcoded offsets for descriptor ring registers
The Intel 10 Gigabit NICs use the same simplified (aka "legacy")
descriptor format and the same layout for descriptor register blocks
as the Intel 1 Gigabit NICs. The offsets of the descriptor register
blocks are not the same.
Simplify reuse of the existing code by removing all hardcoded offsets
for registers within descriptor register blocks, and ensuring that all
offsets are calculated using the descriptor register block base
address provided via intel_init_ring().
Signed-off-by: Michael Brown <mcb30@ipxe.org>
On i350 the datasheet contradicts itself in stating that the default
value of RXDCTL.ENABLE for queue zero is both set (according to the
"Receive Initialization" section) and unset (according to the "Receive
Descriptor Control - RXDCTL" section). Empirical evidence suggests
that the default value is unset.
Explicitly enable both transmit and receive queues to avoid any
ambiguity.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Michael Brown