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] Exclude time spent in hypervisor from profiling
When profiling, exclude any time spent inside the hypervisor
responding to our MMIO accesses. This substantially reduces the
variance accumulated on many other profilers.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Inside a virtual machine, writing the RX ring tail pointer may incur a
substantial overhead of processing inside the hypervisor. Minimise
this overhead by writing the tail pointer once per batch of
descriptors, rather than once per descriptor.
Profiling under qemu-kvm (version 1.6.2) shows that this reduces the
amount of time taken to refill the RX descriptor ring by around 90%.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Operations which are negligible on physical hardware (such as issuing
a posted write to the transmit ring tail register) may involve
substantial amounts of processing within the hypervisor if running in
a virtual machine.
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>
On an Asus Z87-K motherboard with an onboard 8168 NIC, booting into
Windows 7 and then warm rebooting into iPXE results in a broken RX
datapath: packets can be transmitted successfully but garbage is
received. A cold reboot clears the problem.
A dump of the PHY registers reveals only one difference: in the
failure case the bits ADVERTISE_PAUSE_CAP and ADVERTISE_PAUSE_ASYM are
cleared. Explicitly setting these bits does not fix the problem.
A dump of the MAC registers reveals a few differences, of which the
most obvious culprit is the undocumented bit 24 of the Receive
Configuration Register (RCR), which is set in the failure case.
Explicitly clearing this bit does fix the problem.
Reported-by: Sebastian Nielsen <ipxe@sebbe.eu>
Reported-by: Oliver Rath <rath@mglug.de>
Debugged-by: Sebastian Nielsen <ipxe@sebbe.eu>
Tested-by: Sebastian Nielsen <ipxe@sebbe.eu>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
[settings] Explicitly separate the concept of a completed fetched setting
The fetch_setting() family of functions may currently modify the
definition of the specified setting (e.g. to add missing type
information). Clean up this interface by requiring callers to provide
an explicit buffer to contain the completed definition of the fetched
setting, if required.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Give tap devices a meaningful name, and avoid segmentation faults when
attempting to retrieve ${net0/bustype} by assigning a new bus type for
tap devices.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Prevent the card from flagging packets of 1518 bytes length as
overlength.
This fixes the High-MTU loopback test.
Signed-off-by: Thomas Miletich <thomas.miletich@gmail.com>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
The 3c90x B and C revisions support rounding up the packet length to a
specific boundary. Disable this feature to avoid overlength packets.
This fixes the loopback test.
Signed-off-by: Thomas Miletich <thomas.miletich@gmail.com>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
[3c90x] Stall upload engine before setting RX ring address
According to the 3c90x datasheet we have to stall the upload (receive)
engine before setting the receive ring address.
Signed-off-by: Thomas Miletich <thomas.miletich@gmail.com>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Some hardware (observed with an onboard RTL8168) will erroneously
report a buffer overflow error if the received packet exactly fills
the receive buffer.
Fix by adding an extra four bytes of padding to each receive buffer.
Debugged-by: Thomas Miletich <thomas.miletich@gmail.com>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Replace the old via-rhine driver with a new version using the iPXE
API.
Includes fixes by Thomas Miletich for:
- MMIO access
- Link detection
- RX completion in RX overflow case
- Reset and EEPROM reloading
- CRC stripping
- Missing cpu_to_le32() calls
- Missing memory barriers
Signed-off-by: Adrian Jamróz <adrian.jamroz@gmail.com>
Modified-by: Thomas Miletich <thomas.miletich@gmail.com>
Tested-by: Thomas Miletich <thomas.miletich@gmail.com>
Tested-by: Robin Smidsrød <robin@smidsrod.no>
Modified-by: Michael Brown <mcb30@ipxe.org>
Tested-by: Michael Brown <mcb30@ipxe.org>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
[settings] Expose PCI configuration space via settings mechanism
Allow values to be read from PCI configuration space using the syntax
${pci/<busdevfn>.<offset>.<length>}
where <busdevfn> is the bus:dev.fn address of the PCI device
(expressed as a single integer, as returned by ${net0/busloc}),
<offset> is the offset within PCI configuration space, and <length> is
the length within PCI configuration space.
Values are returned in reverse byte order, since PCI configuration
space is little-endian by definition.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
realtek_destroy_ring() currently does nothing if the card is operating
in legacy (pre-RTL8139C+) mode. In particular, the producer and
consumer counters are incorrectly left holding their current values.
Virtual hardware (e.g. the emulated RTL8139 in qemu and similar VMs)
is tolerant of this behaviour, but real hardware will fail to transmit
if the descriptors are not used in the correct order.
Fix by resetting the producer and consumer counters in
realtek_destroy_ring() even if the card is operating in legacy mode.
Reported-by: Gelip <mrgelip@gmail.com>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Create an explicit concept of "settings scope" and eliminate the magic
values used for numerical setting tags.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
[realtek] Ensure EEPROM writes reach chip before starting udelay()
On some systems, it appears to be possible for writes to the EEPROM
registers to be delayed for long enough that the EEPROM's setup and
hold times are violated, resulting in invalid data being read from the
EEPROM.
Fix by inserting a PCI read cycle immediately after writes to
RTL_9346CR, to ensure that the write has completed before starting the
udelay() used to time the SPI bus transitions.
Reported-by: Gelip <mrgelip@gmail.com>
Tested-by: Gelip <mrgelip@gmail.com>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
[realtek] Allow reaction time between writing RTL_CAPR and reading RTL_CR
Some older RTL8139 chips seem to not immediately update the
RTL_CR.BUFE bit in response to a write to RTL_CAPR. This results in
iPXE seeing a spurious zero-length received packet, and thereafter
being out of sync with the hardware's RX ring offset.
Fix by inserting an extra PCI read cycle after writing to RTL_CAPR, to
give the chip time to react before we next read RTL_CR.
Reported-by: Gelip <mrgelip@gmail.com>
Tested-by: Gelip <mrgelip@gmail.com>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Some onboard RTL8169 NICs seem to leave the EEPROM pins disconnected.
The existing is_valid_ether_addr() test will not necessarily catch
this, since it expects a missing EEPROM to show up as a MAC address of
00:00:00:00:00:00 or ff:ff:ff:ff:ff:ff. When the EEPROM pins are
floating the MAC address may read as e.g. 00:00:00:00:0f:00, which
will not be detected as invalid.
Check the ID word in the first two bytes of the EEPROM (which should
have the value 0x8129 for all RTL8139 and RTL8169 chips), and use this
to determine whether or not an EEPROM is present.
Reported-by: Carl Karsten <carl@nextdayvideo.com>
Tested-by: Carl Karsten <carl@nextdayvideo.com>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Exploit the redefinition of iPXE error codes to include a "platform
error code" to allow for meaningful conversion of EFI_STATUS values to
iPXE errors and vice versa.
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>
Avoid using UINT16 and similar typedefs, which are non-standard in the
iPXE codebase and generate conflicts when trying to include any of the
EFI headers.
Also fix trailing whitespace in the affected files, to prevent
complaints from git.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Remove macros which aren't used anywhere in the driver, and which
conflict with macros of the same name used in the EFI headers.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Remove macros which aren't used anywhere in the driver, and which
conflict with macros of the same name used in the EFI headers.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
The iBFT NIC section has a VLAN field which must be filled in so that
iSCSI booting works over VLANs.
Unfortunately it is unclear from the IBM specification linked in
ibft.c whether the VLAN field is just the 802.1Q VLAN Identifier or
the full 802.1Q TCI. For now just fill in the VID, the Priority Code
Point and Drop Eligible Indicator could be set in the future if it
turns out they should be present too.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Modified-by: Michael Brown <mcb30@ipxe.org>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
[forcedeth] Increase TX ring size to prevent dropped packets
Commit 947976d ("[netdevice] Do not force a poll on net_tx()")
requires network devices to have TX rings that are sufficiently large
to allow a transmitted response to all packets received during a
single poll.
Reported-by: Robin Smidsrød <robin@smidsrod.no>
Tested-by: Robin Smidsrød <robin@smidsrod.no>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
[intel] Poll RX queue if hardware reports RX overflow
The Intel NIC emulation in some versions of VMware seems to suffer
from a flaw whereby the Interrupt Cause Register (ICR) fails to assert
the usual "packet received" bit (ICR.RXT0) if a receive overflow
(ICR.RXO) has also occurred.
Work around this flaw by polling for completed descriptors whenever
either ICR.RXT0 or ICR.RXO is asserted.
Reported-by: Miroslav Halas <miroslav.halas@bankofamerica.com>
Debugged-by: Miroslav Halas <miroslav.halas@bankofamerica.com>
Tested-by: Miroslav Halas <miroslav.halas@bankofamerica.com>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Michael Brown