Sukadev Bhattiprolu says:
====================
ibmvnic: Reuse ltb, rx, tx pools
It can take a long time to free and reallocate rx and tx pools and long
term buffer (LTB) during each reset of the VNIC. This is specially true
when the partition (LPAR) is heavily loaded and going through a Logical
Partition Migration (LPM). The long drawn reset causes the LPAR to lose
connectivity for extended periods of time and results in "RMC connection"
errors and the LPM failing.
What is worse is that during the LPM we could get a failover because
of the lost connectivity. At that point, the vnic driver releases
even the resources it has already allocated and starts over.
As long as the resources we have already allocated are valid/applicable,
we might as well hold on to them while trying to allocate the remaining
resources. This patch set attempts to reuse the resources previously
allocated as long as they are valid. It seems to vastly improve the
time taken for the vnic reset and signficantly reduces the chances of
getting the RMC connection errors. We do get still them occasionally,
but appears to be for reasons other than memory allocation delays and
those are still being investigated.
If the backing devices for a vnic adapter are not "matched" (see "pool
parameters" in patches 8 and 9) it is possible that we will still free
all the resources and allocate them. If that becomes a common problem,
we have to address it separately.
Thanks to input and extensive testing from Brian King, Cris Forno,
Dany Madden, Rick Lindsley.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Rather than releasing the tx pools on every close and reallocating
them on open, reuse the tx pools unless the pool parameters (number
of pools, size of each pool or size of each buffer in a pool) have
changed.
If the pool parameters changed, then release the old pools (if
any) and allocate new ones.
Specifically release tx pools, if:
- adapter is removed,
- pool parameters change during reset,
- we encounter an error when opening the adapter in response
to a user request (in ibmvnic_open()).
and don't release them:
- in __ibmvnic_close() or
- on errors in __ibmvnic_open()
in the hope that we can reuse them during this or next reset.
With these changes reset_tx_pools() can be dropped because its
optimization is now included in init_tx_pools() itself.
cleanup_tx_pools() releases all the skbs associated with the pool and
is called from ibmvnic_cleanup(), which is called on every reset. Since
we want to reuse skbs across resets, move cleanup_tx_pools() out of
ibmvnic_cleanup() and call it only when user closes the adapter.
Add two new adapter fields, ->prev_mtu, ->prev_tx_pool_size to track the
previous values and use them to decide whether to reuse or realloc the
pools.
Reviewed-by: Rick Lindsley <ricklind@linux.vnet.ibm.com>
Reviewed-by: Dany Madden <drt@linux.ibm.com>
Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Rather than releasing the rx pools on and reallocating them on every
reset, reuse the rx pools unless the pool parameters (number of pools,
size of each pool or size of each buffer in a pool) have changed.
If the pool parameters changed, then release the old pools (if any)
and allocate new ones.
Specifically release rx pools, if:
- adapter is removed,
- pool parameters change during reset,
- we encounter an error when opening the adapter in response
to a user request (in ibmvnic_open()).
and don't release them:
- in __ibmvnic_close() or
- on errors in __ibmvnic_open()
in the hope that we can reuse them on the next reset.
With these, reset_rx_pools() can be dropped because its optimzation is
now included in init_rx_pools() itself.
cleanup_rx_pools() releases all the skbs associated with the pool and
is called from ibmvnic_cleanup(), which is called on every reset. Since
we want to reuse skbs across resets, move cleanup_rx_pools() out of
ibmvnic_cleanup() and call it only when user closes the adapter.
Add two new adapter fields, ->prev_rx_buf_sz, ->prev_rx_pool_size to
keep track of the previous values and use them to decide whether to
reuse or realloc the pools.
Reviewed-by: Rick Lindsley <ricklind@linux.vnet.ibm.com>
Reviewed-by: Dany Madden <drt@linux.ibm.com>
Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Reuse the long term buffer during a reset as long as its size has
not changed. If the size has changed, free it and allocate a new
one of the appropriate size.
When we do this, alloc_long_term_buff() and reset_long_term_buff()
become identical. Drop reset_long_term_buff().
Reviewed-by: Rick Lindsley <ricklind@linux.vnet.ibm.com>
Reviewed-by: Dany Madden <drt@linux.ibm.com>
Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In a follow-on patch, we will reuse long term buffers when possible.
When doing so we have to be careful to properly assign map ids. We
can no longer assign them sequentially because a lower map id may be
available and we could wrap at 255 and collide with an in-use map id.
Instead, use a bitmap to track active map ids and to find a free map id.
Don't need to take locks here since the map_id only changes during reset
and at that time only the reset worker thread should be using the adapter.
Noticed this when analyzing an error Dany Madden ran into with the
patch set.
Reported-by: Dany Madden <drt@linux.ibm.com>
Reviewed-by: Rick Lindsley <ricklind@linux.vnet.ibm.com>
Reviewed-by: Dany Madden <drt@linux.ibm.com>
Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In init_tx_pools() move some loop-invariant code out of the loop.
Reviewed-by: Rick Lindsley <ricklind@linux.vnet.ibm.com>
Reviewed-by: Dany Madden <drt@linux.ibm.com>
Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use/rename local variables in init_tx_pools() for consistency with
init_rx_pools() and for readability. Also add some comments
Reviewed-by: Rick Lindsley <ricklind@linux.vnet.ibm.com>
Reviewed-by: Dany Madden <drt@linux.ibm.com>
Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
To make the code more readable, use/rename some local variables.
Basically we have a set of pools, num_pools. Each pool has a set of
buffers, pool_size and each buffer is of size buff_size.
pool_size is a bit ambiguous (whether size in bytes or buffers). Add
a comment in the header file to make it explicit.
Reviewed-by: Rick Lindsley <ricklind@linux.vnet.ibm.com>
Reviewed-by: Dany Madden <drt@linux.ibm.com>
Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add/update some comments/function headers and fix up some messages.
Reviewed-by: Rick Lindsley <ricklind@linux.vnet.ibm.com>
Reviewed-by: Dany Madden <drt@linux.ibm.com>
Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
For better readability, consolidate related code in replenish_rx_pool()
and add some comments.
Reviewed-by: Rick Lindsley <ricklind@linux.vnet.ibm.com>
Reviewed-by: Dany Madden <drt@linux.ibm.com>
Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Jonathan Lemon says:
====================
timecard updates for v13 firmware
This update mainly deals with features for the TimeCard v13 firmware.
The signals provided from the external SMA connectors can be steered
to different locations, and the generated SMA signals can be chosen.
Future timecard revisions will allow selectable I/O on any of the
SMA connectors, so name the attributes appropriately, and set up
the ABI in preparation for the new features.
The update also adds support for IRIG-B and DCF formats, as well
as NMEA output.
A ts_window_adjust tunable is also provided to fine-tune the
PHC:SYS time mapping.
--
v1: Earlier reviewed series was for v10 firmware, this is expanded to
include the v13 features.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch describes the sysfs interface implemented by the
ptp_ocp driver, under /sys/class/timecard.
Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The following process is used to read the PHC clock and correlate
the reading with the "correct" system time.
- get starting timestamp
- issue PCI write command
- issue PCI read command
- get ending timestamp
- read latched sec/nsec registers
The write command is posted to PCI bus and returns. When the write
arrives at the FPGA, the PHC time is latched into the sec/nsec registers,
and a flag is set indicating the registers are valid. The read command
returns this flag, and the time retrieval proceeds.
Below is a non-scaled picture of the timing diagram involved. The
PHC time corresponds to some SYS time between [start, end]. Userspace
usually uses the midpoint between [start, end] to estimate the PCI
delay and match this with the PHC time.
[start] | |
write |-------+ |
| \ |
read |----+ +----->|
| \ * PHC time latched into register
| \ |
midpoint | +------->|
| |
| |
| +----|
| / |
|<--------+ |
[end] | |
As the diagram indicates, the PHC time is latched before the midpoint,
so the system clock time is slightly off the real PHC time. This shows
up as a phase error with an oscilliscope.
The workaround here is to provide a tunable which reduces (shrinks)
the end time in the above diagram. This in turn moves the calculated
midpoint so the system time and PHC time are in agreemment.
Currently, the adjustment reduces the end time by 3/16th of the entire
window. E.g.: [start, end] ==> [start, (end - (3/16 * end)], which
produces reasonably good results.
Also reduce delays by just writing to the clock control register
instead of performing a read/modify/write sequence, as the contents
of the control register are known.
Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The current implementation of adjtime uses gettime/settime to
perform nanosecond adjustments. This introduces addtional phase
errors due to delays.
Instead, use the FPGA's ability to just apply the nanosecond
adjustment to the clock directly.
Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
A 4th timestamper is added which timestamps the output of the PHC.
The clock nanosecond offset is not always zero, so when compared
to other timestampers, this provides precise measurements.
Also, the timestamper interrupt from the PHC can be used to generate
a PPS signal for /dev/pps.
Also allow PTP_CLK_REQ_PEROUT requests for a 1PPS output, but do
not actually configure any output pins, this is done via sysfs.
Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Upcoming boards may have a second GNSS receiver, getting information
from a different constellation than the first receiver, which provides
some measure of anti-spoofing.
Expose the sysfs attribute for this device, if detected.
Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The timecard can provide a NMEA-1083 ZDA (time and date) output
string on a serial port, which can be used to drive other devices.
Add the NMEA resources, and the serial port as a sysfs attribute.
Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Provide a view into the timecard internals for debugging.
Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
On startup, parts of the FPGA need to be initialized - break these
out into their own functions, separate from the purely informational
blocks.
On startup, distrbute the UTC:TAI offset from the NMEA GNSS parser,
if it is available.
Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
IRIG and DCF output time in UTC, but the timecard operates
on TAI internally. Add an attribute node which allows adding
an offset to these modes before output.
Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
IRIG-B has several different output formats, the timecard defaults
to using B007. Add a control which selects different output modes.
Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
IRIG (Inter-range Instrumentation Group) timecode format on
one of the SMA output channels is provided by the IRIG master
FPGA block. Enable the master when the IRIG output format is
selected on either one of the output channels.
By default, the output is in B007 format.
DCF output format is provided by the DCF master block.
Also enable the IRIG and DCF slaves, which parse an incoming
signal from the external SMA connectors, and may be used to
adjust the PHC.
Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The latest firmware for the TimeCard adds selectable signals for
the SMA input/outputs. Add support for SMA selectors, and the
GPIO controls needed for steering signals.
Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The firmware may provide a third signal timestamper, so make it
available for use.
Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If a resource could not be registered, report the name of
the resource and the error code.
Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The TimeCard exposes different resources, which may have their
own irqs. Space for the irqs is allocated through a MSI or MSI-X
interrupt vector. On some platforms, the interrupt allocation
fails.
Rather than making this fatal, just skip exposing those resources.
The main timecard functionality (that of a PTP clock) will work
without the additional resources.
Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If an I2C controller isn't present, don't try and read the I2C flash.
Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Only display the TOD information if there is a corresponding
TOD resource.
Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Move the xilinx i2c driver parameters to the resource block instead
of hardcoding things in the registration functions.
Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The new added properties are used for configuring burst length.
Signed-off-by: Aleksander Jan Bajkowski <olek2@wp.pl>
Signed-off-by: David S. Miller <davem@davemloft.net>
Document the Lantiq Xway SoC series External Bus Unit (ETOP) bindings.
Signed-off-by: Aleksander Jan Bajkowski <olek2@wp.pl>
Signed-off-by: David S. Miller <davem@davemloft.net>
Convert the Lantiq PMAC Device Tree binding documentation to json-schema.
Signed-off-by: Aleksander Jan Bajkowski <olek2@wp.pl>
Signed-off-by: David S. Miller <davem@davemloft.net>
Configure the burst length in Ethernet drivers. This improves
Ethernet performance by 58%. According to the vendor BSP,
8W burst length is supported by ar9 and newer SoCs.
The NAT benchmark results on xRX200 (Down/Up):
* 2W: 330 Mb/s
* 4W: 432 Mb/s 372 Mb/s
* 8W: 520 Mb/s 389 Mb/s
Tested on xRX200 and xRX330.
Signed-off-by: Aleksander Jan Bajkowski <olek2@wp.pl>
Signed-off-by: David S. Miller <davem@davemloft.net>
Make the burst length configurable by the drivers.
Signed-off-by: Aleksander Jan Bajkowski <olek2@wp.pl>
Acked-by: Hauke Mehrtens <hauke@hauke-m.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
The current definition of 2W burst length is invalid.
This patch fixes it. Current downstream DEU driver doesn't
use DMA. An incorrect burst length value doesn't cause any
errors. This patch also adds other burst length values.
Fixes: dfec1a827d ("MIPS: Lantiq: Add DMA support")
Signed-off-by: Aleksander Jan Bajkowski <olek2@wp.pl>
Signed-off-by: David S. Miller <davem@davemloft.net>
Different SoCs have a different number of channels, e.g .:
* amazon-se has 10 channels,
* danube+ar9 have 20 channels,
* vr9 has 28 channels,
* ar10 has 24 channels.
We can read the ID register and, depending on the reported
number of channels, reset the appropriate number of channels.
Signed-off-by: Aleksander Jan Bajkowski <olek2@wp.pl>
Signed-off-by: David S. Miller <davem@davemloft.net>
Reading the DMA registers immediately after the reset causes
Data Bus Error. Adding a small delay fixes this issue.
Signed-off-by: Aleksander Jan Bajkowski <olek2@wp.pl>
Signed-off-by: David S. Miller <davem@davemloft.net>
This drops the code setting bit 9 on egress frames on the
Realtek "type A" (RTL8366RB) frames.
This bit was set on ingress frames for unknown reason,
and was set on egress frames as the format of ingress
and egress frames was believed to be the same. As that
assumption turned out to be false, and since this bit
seems to have zero effect on the behaviour of the switch
let's drop this bit entirely.
Signed-off-by: Linus Walleij <linus.walleij@linaro.org>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Link: https://lore.kernel.org/r/20210913143156.1264570-1-linus.walleij@linaro.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Guangbin Huang says:
====================
PF support get MAC address space assigned by firmware
This series add support PF to get unicast/multicast MAC address space
assigned by firmware for the HNS3 ethernet driver.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
The new firmware supports to divides the whole multicast MAC address space
equally to functions of all PFs, and calculates the space size of each PF
according to its function number.
To support this feature, PF driver adds querying multicast MAC address
space size from firmware and limits used number according to space size.
Signed-off-by: Guangbin Huang <huangguangbin2@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, there are two ways for PF to set the unicast MAC address space
size: specified by config parameters in firmware or set to default value.
That's mean if the config parameters in firmware is zero, driver will
divide the whole unicast MAC address space equally to 8 PFs. However, in
this case, the unicast MAC address space will be wasted a lot when the
hardware actually has less then 8 PFs. And in the other hand, if one PF has
much more VFs than other PFs, then each function of this PF will has much
less address space than other PFs.
In order to ameliorate the above two situations, introduce the third way
of unicast MAC address space assignment: firmware divides the whole unicast
MAC address space equally to functions of all PFs, and calculates the space
size of each PF according to its function number. PF queries the space size
by the querying device specification command when in initialization
process.
The third way assignment is lower priority than specified by config
parameters, only if the config parameters is zero can be used, and if
firmware does not support the third way assignment, then driver still
divides the whole unicast MAC address space equally to 8 PFs.
Signed-off-by: Guangbin Huang <huangguangbin2@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add support for qca8327 internal phy needed for correct init of the
switch port. It does use the same qca8337 function and reg just with a
different id.
Signed-off-by: Ansuel Smith <ansuelsmth@gmail.com>
Tested-by: Rosen Penev <rosenp@gmail.com>
Tested-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
The __alloc_frag_align() is short, and only called by two functions,
so inline page_frag_alloc_align() for reduce the overhead of calls.
Reported-by: kernel test robot <oliver.sang@intel.com>
Signed-off-by: Yajun Deng <yajun.deng@linux.dev>
Signed-off-by: David S. Miller <davem@davemloft.net>
It shouldn't happen, but can happen that readable eeprom size is smaller
than announced. Then we would be stuck in an endless loop here because
after reaching the actual end reads return eeprom.len = 0. I faced this
issue when making a mistake in driver development. Detect this scenario
and return an error.
Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Removed "stdbool.h" inclusion in iosm_ipc_imem.h
Fixes: 13bb8429ca ("net: wwan: iosm: firmware flashing and coredump collection")
Reported-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: M Chetan Kumar <m.chetan.kumar@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Also correct how initialize_dco_operating_mode is called
Signed-off-by: Min Li <min.li.xe@renesas.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
So far we don't need to support new 5.2 functions but different register
addresses
Signed-off-by: Min Li <min.li.xe@renesas.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Not need since TCS firmware file will configure it properlly.
Signed-off-by: Min Li <min.li.xe@renesas.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch is a follow-up to beb401ec50 ("r8169: deprecate support for
RTL_GIGA_MAC_VER_27") that came with 5.12. Nobody complained, so let's
remove support for this chip version.
Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Guvenc Gulce says:
====================
net/smc: add EID support
please apply the following patch series for smc to netdev's net-next
tree. The series introduce the so called Enterprise ID support for smc
protocol. Including the generic netlink based interface.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
