When jumbo frames are enabled, the Linux ixgbe physical function
driver will disable the virtual function's receive datapath by
default, and will enable it only if the virtual function negotiates
API version 1.1 (or higher) and explicitly selects an MTU.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Virtual functions use a mailbox to communicate with the physical
function driver: this covers functionality such as obtaining the MAC
address.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Intel virtual function NICs almost work with the use of "legacy"
transmit and receive descriptors (which are backwards compatible right
back to the original Intel Gigabit NICs).
Unfortunately the "TX switching" feature (which allows for VM<->VM
traffic to be looped back within the NIC itself) does not work when a
legacy TX descriptor is used: the packet is instead sent onto the
wire.
Fix by allowing for the use of an "advanced" TX descriptor (containing
exactly the same information as is found in the "legacy" descriptor).
Signed-off-by: Michael Brown <mcb30@ipxe.org>
The current API for Base16 (and Base64) encoding requires the caller
to always provide sufficient buffer space. This prevents the use of
the generic encoding/decoding functionality in some situations, such
as in formatting the hex setting types.
Implement a generic hex_encode() (based on the existing
format_hex_setting()), implement base16_encode() and base16_decode()
in terms of the more generic hex_encode() and hex_decode(), and update
all callers to provide the additional buffer length parameter.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
This changed in Linux kernel the same way in commit 7067e701
("ath9k_hw: remove confusing logic inversion in an ANI variable") by
Felix Fietkau.
Additionally this fixes "error: logical not is only applied to the
left hand side of comparison" with GCC 5.1.0.
Signed-off-by: Christian Hesse <mail@eworm.de>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
I218-LM (rev 3) is found in Lenovo Thinkpad X250. The remaining
device IDs are from linux/drivers/net/ethernet/intel/e1000e/hw.h
Signed-off-by: Christian Hesse <mail@eworm.de>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
On some RTL8169 onboard NICs (observed with a Lenovo ThinkPad 11e),
the EEPROM is not merely not present: any attempt to read from the
non-existent EEPROM will crash and reboot the system.
The equivalent code to read from the EEPROM was removed from the Linux
r8169 driver in 2009 with a comment suggesting that it was similarly
found to be unreliable on some systems.
Fix by accessing the EEPROM only on RTL8139 NICs, and assuming that
the MAC address will always be correctly preset on RTL8169 NICs.
Reported-by: Evan Prohaska <eprohaska@edkey.org>
Tested-by: Evan Prohaska <eprohaska@edkey.org>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
The emulated Intel 82545em in some versions of VMware (observed with
ESXi v5.1) seems to sometimes fail to set the RXT0 bit in the
interrupt cause register (ICR), causing iPXE to stop receiving
packets. Work around this problem (for the 82545em only) by always
polling the receive queue regardless of the state of the ICR.
Reported-by: Slava Bendersky <volga629@networklab.ca>
Tested-by: Slava Bendersky <volga629@networklab.ca>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
At least one NII implementation (in a Microsoft Surface tablet) seems
to fail to report the absence (sic) of TX completions properly. Work
around this by checking for TX completions only when we expect to see
one.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Some NII implementations will fail the GET_STATUS operation if we
request the media status. Fix by doing so only if GET_INIT_INFO
reported that media status is supported.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
RX FIFO overflow is almost inevitable since the (usable) USB2 bus
bandwidth is approximately one quarter of the Ethernet bandwidth.
Avoid flooding the console with RX FIFO overflow messages in a
standard debug build.
With TCP SACK implemented, the RX FIFO overflow no longer causes a
catastrophic drop in throughput. Experimentation shows that HTTP
downloads now progress at a fairly smooth 250Mbps, which is around the
maximum speed attainable for a USB2 NIC.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
This driver is functional but any downloads via a TCP-based protocol
tend to perform poorly. The 1Gbps Ethernet line rate is substantially
higher than the 480Mbps (in practice around 280Mbps) provided by USB2,
and the device has only 32kB of internal buffer memory. Our 256kB TCP
receive window therefore rapidly overflows the RX FIFO, leading to
multiple dropped packets (usually within the same TCP window) and
hence a low overall throughput.
Reducing the TCP window size so that the RX FIFO does not overflow
greatly increases throughput, but is not a general-purpose solution.
Further investigation is required to determine how other OSes
(e.g. Linux) cope with this scenario. It is possible that
implementing TCP SACK would provide some benefit.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Most devices expose at least the link up/down status via a bit in a
MAC register, since the MAC generally already needs to know whether or
not the link is up. Some devices (e.g. the SMSC75xx USB NIC) expose
this information to software only via the MII registers.
Provide a generic mii_check_link() implementation to check the BMSR
and report the link status via netdev_link_{up,down}().
Signed-off-by: Michael Brown <mcb30@ipxe.org>
iPXE already sends RX notifications to the backend when needed, but
does not set the "feature-rx-notify" flag. As of XenServer 6.5, this
flag is mandatory and omitting it will cause the backend to fail.
Fix by setting the "feature-rx-notify" flag, to inform the backend
that we will send notifications.
Reported-by: Shalom Bhooshi <shalom.bhooshi@citrix.com>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Relicense files with kind permission from
Stefan Hajnoczi <stefanha@redhat.com>
alongside the contributors who have already granted such relicensing
permission.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
At some point in the past few years, binutils became more aggressive
at removing unused symbols. To function as a symbol requirement, a
relocation record must now be in a section marked with @progbits and
must not be in a section which gets discarded during the link (either
via --gc-sections or via /DISCARD/).
Update REQUIRE_SYMBOL() to generate relocation records meeting these
criteria. To minimise the impact upon the final binary size, we use
existing symbols (specified via the REQUIRING_SYMBOL() macro) as the
relocation targets where possible. We use R_386_NONE or R_X86_64_NONE
relocation types to prevent any actual unwanted relocation taking
place. Where no suitable symbol exists for REQUIRING_SYMBOL() (such
as in config.c), the macro PROVIDE_REQUIRING_SYMBOL() can be used to
generate a one-byte-long symbol to act as the relocation target.
If there are versions of binutils for which this approach fails, then
the fallback will probably involve killing off REQUEST_SYMBOL(),
redefining REQUIRE_SYMBOL() to use the current definition of
REQUEST_SYMBOL(), and postprocessing the linked ELF file with
something along the lines of "nm -u | wc -l" to check that there are
no undefined symbols remaining.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Add the standard warranty disclaimer and Free Software Foundation
address paragraphs to the licence text where these are not currently
present.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Some USB endpoints require that a short packet be used to terminate
transfers, since they have no other way to determine message
boundaries. If the message length happens to be an exact multiple of
the USB packet size, then this requires the use of an additional
zero-length packet.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Some protocols (such as ARP) may modify the received packet and re-use
the same I/O buffer for transmission of a reply. To allow this,
reserve sufficient headroom at the start of each received packet
buffer for our transmit datapath headers.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Some devices have a very small number of internal buffers, and rely on
being able to pack multiple packets into each buffer. Using 2048-byte
buffers on such devices produces throughput of around 100Mbps. Using
a small number of much larger buffers (e.g. 32kB) increases the
throughput to around 780Mbps. (The full 1Gbps is not reached because
the high RTT induced by the use of multi-packet buffers causes us to
saturate our 256kB TCP window.)
Since allocation of large buffers is very likely to fail, allocate the
buffer set only once when the device is opened and recycle buffers
immediately after use. Received data is now always copied to
per-packet buffers.
If allocation of large buffers fails, fall back to allocating a larger
number of smaller buffers. This will give reduced performance, but
the device will at least still be functional.
Share code between the interrupt and bulk IN endpoint handlers, since
the buffer handling is now very similar.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
The (undocumented) VMBus protocol seems to allow for transfer
page-based packets where the data payload is split into an arbitrary
set of ranges within the transfer page set.
The RNDIS protocol includes a length field within the header of each
message, and it is known from observation that multiple RNDIS messages
can be concatenated into a single VMBus message.
iPXE currently assumes that the transfer page range boundaries are
entirely arbitrary, and uses the RNDIS header length to determine the
RNDIS message boundaries.
Windows Server 2012 R2 generates an RNDIS_INDICATE_STATUS_MSG for an
undocumented and unknown status code (0x40020006) with a malformed
RNDIS header length: the length does not cover the StatusBuffer
portion of the message. This causes iPXE to report a malformed RNDIS
message and to discard any further RNDIS messages within the same
VMBus message.
The Linux Hyper-V driver assumes that the transfer page range
boundaries correspond to RNDIS message boundaries, and so does not
notice the malformed length field in the RNDIS header.
Match the behaviour of the Linux Hyper-V driver: assume that the
transfer page range boundaries correspond to the RNDIS message
boundaries and ignore the RNDIS header length. This avoids triggering
the "malformed packet" error and also avoids unnecessary data copying:
since we now have one I/O buffer per RNDIS message, there is no longer
any need to use iob_split().
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Empirical observation suggests that 32 is a sensible size to minimise
the number of deferred packet transmissions without overflowing the
VMBus transmit ring buffer.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Allow for elision of transmitted TCP ACKs by handling all received
VMBus messages in each network device poll operation.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
On Windows Server 2012 R2, the receive buffer teardown completion
message seems to occasionally be deferred until after the VMBus
channel has been closed. This happens even if there are no packets
currently in the receive buffer.
Work around this problem by separating the revocation and teardown of
the receive buffer, and deferring the teardown until after the VMBus
channel has been closed.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
The i350 (and possibly other Intel NICs) have a non-trivial
correspondence between the PCI function number and the external
physical port number. For example, the i350 has a "LAN Function Sel"
bit within the EEPROM which can invert the mapping so that function 0
becomes port 3, function 1 becomes port 2, etc.
Unfortunately the MAC addresses within the EEPROM are indexed by
physical port number rather than PCI function number. The end result
is that when anything other than the default mapping is used, iPXE
will use the wrong address as the base MAC address.
Fix by using the autoloaded MAC address if it is valid, and falling
back to reading the MAC address directly from the EEPROM only if no
autoloaded address is available.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
End users almost certainly don't care whether the underlying interface
is SNP or NII/UNDI. Try to minimise surprise and unnecessary
documentation by including the NII driver whenever the SNP driver is
requested.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
iPXE itself exposes a dummy NII protocol with no UNDI. Avoid
potentially dereferencing a NULL pointer by checking for a non-zero
UNDI address.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Some UEFI network drivers provide a software UNDI interface which is
exposed via the Network Interface Identifier Protocol (NII), rather
than providing a Simple Network Protocol (SNP).
The UEFI platform firmware will usually include the SnpDxe driver,
which attaches to NII and provides an SNP interface. The SNP
interface is usually provided on the same handle as the underlying NII
device. This causes problems for our EFI driver model: when
efi_driver_connect() detaches existing drivers from the handle it will
cause the SNP interface to be uninstalled, and so our SNP driver will
not be able to attach to the handle. The platform firmware will
eventually reattach the SnpDxe driver and may attach us to the SNP
handle, but we have no way to prevent other drivers from attaching
first.
Fix by providing a driver which can attach directly to the NII
protocol, using the software UNDI interface to drive the network
device.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
The snpnet driver uses netdev_tx_defer() and so must ensure that space
in the (single-entry) transmit descriptor ring is freed up before
calling netdev_tx_complete().
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Add the ID for the LM variant and differentiate it from the I217-V.
Signed-off-by: Jan Kiszka <jan.kiszka@web.de>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
We currently require information about the underlying PCI device to
populate the snpnet device's name and description. If the underlying
device is not a PCI device, this will fail and prevent the device from
being registered.
Fix by falling back to populating the device description with
information based on the EFI handle, if no PCI device information is
available.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Some systems will install a child of the SNP device and use this as
our loaded image's device handle, duplicating the installation of the
underlying SNP protocol onto the child device handle. On such
systems, we want to end up driving the parent device (and
disconnecting any other drivers, such as MNP, which may be attached to
the parent device).
Fix by recording the SNP protocol instance at initialisation time, and
using this to match against device handles (rather than simply
comparing the handles themselves).
Reported-by: Jarrod Johnson <jarrod.b.johnson@gmail.com>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
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>
Under some circumstances (e.g. if iPXE itself is booted via iSCSI, or
after an unclean reboot), the backend may not be in the expected
InitWait state when iPXE starts up.
There is no generic reset mechanism for Xenbus devices. Recent
versions of xen-netback will gracefully perform all of the required
steps if the frontend sets its state to Initialising. Older versions
(such as that found in XenServer 6.2.0) require the frontend to
transition through Closed before reaching Initialising.
Add a reset mechanism for netfront devices which does the following:
- read current backend state
- if backend state is anything other than InitWait, then set the
frontend state to Closed and wait for the backend to also reach
Closed
- set the frontend state to Initialising and wait for the backend to
reach InitWait.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Using version 1 grant tables limits guests to using 16TB of grantable
RAM, and prevents the use of subpage grants. Some versions of the Xen
hypervisor refuse to allow the grant table version to be set after the
first grant references have been created, so the loaded operating
system may be stuck with whatever choice we make here. We therefore
currently use version 2 grant tables, since they give the most
flexibility to the loaded OS.
Current versions (7.2.0) of the Windows PV drivers have no support for
version 2 grant tables, and will merrily create version 1 entries in
what the hypervisor believes to be a version 2 table. This causes
some confusion.
Avoid this problem by attempting to use version 1 tables, since
otherwise we may render Windows unable to boot.
Play nicely with other potential bootloaders by accepting either
version 1 or version 2 grant tables (if we are unable to set our
requested version).
Note that the use of version 1 tables on a 64-bit system introduces a
possible failure path in which a frame number cannot fit into the
32-bit field within the v1 structure. This in turn introduces
additional failure paths into netfront_transmit() and
netfront_refill_rx().
Signed-off-by: Michael Brown <mcb30@ipxe.org>
The behavior observed in the Apple EFI (1.10) RecieveFilters() call
is:
- failure if any of the PROMISCUOUS or MULTICAST filters are
included
- success if only UNICAST is included, however the result is
UNICAST|BROADCAST
- success if only UNICAST and BROADCAST are included
- if UNICAST, or UNICAST|BROADCAST are used, but the previous call
tried (and failed) to set UNICAST|BROADCAST|MULTICAST, then the
result is UNICAST|BROADCAST|MULTICAST
Work around this apparently broken SNP implementation by trying
RecieveFilterMask, then falling back to UNICAST|BROADCAST|MULTICAST,
then UNICAST|BROADCAST, and finally UNICAST.
Modified-by: Michael Brown <mcb30@ipxe.org>
Tested-by: Curtis Larsen <larsen@dixie.edu>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Some EFI 1.10 systems (observed on an Apple iMac) do not allow us to
open the device path protocol with an attribute of
EFI_OPEN_PROTOCOL_BY_DRIVER and so we cannot maintain a safe,
long-lived pointer to the device path. Work around this by instead
opening the device path protocol with an attribute of
EFI_OPEN_PROTOCOL_GET_PROTOCOL whenever we need to use it.
Debugged-by: Curtis Larsen <larsen@dixie.edu>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
According to the UEFI specification, the MCastFilter parameter (which
we currently pass as NULL, along with a zero MCastFilterCnt) is
optional only if ResetMCastFilter is true.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Dump the existing openers of a protocol whenever we are unable to open
a protocol using attributes of BY_DEVICE, EXCLUSIVE, or
BY_CHILD_CONTROLLER.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Using efi_devpath_text() is marginally more efficient if we already
have the device path protocol available, but the mild increase in
efficiency is not worth compromising the clarity of the pattern:
DBGC ( device, "THING %p %s ...", device, efi_handle_name ( device ) );
Signed-off-by: Michael Brown <mcb30@ipxe.org>
Rewrite the SNP NIC driver to use non-blocking and deferrable
transmissions, to provide link status detection, to provide
information about the underlying (PCI) hardware device, and to avoid
unnecessary I/O buffer allocations during receive polling.
Signed-off-by: Michael Brown <mcb30@ipxe.org>
The VF might not have assigned a MAC address upon startup, and will
end up with a random MAC address during probe(). With this patch the
MAC address can be changed later on.
Signed-off-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Michael Brown <mcb30@ipxe.org>
If the VF doesn't have a MAC address assigned we should create a
random MAC address.
Signed-off-by: Hannes Reinecke <hare@suse.de>
Modified-by: Michael Brown <mcb30@ipxe.org>
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>
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>
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>
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>
