Returning fabricated CRS value (0xFFFF0001) by PCIe Root Complex to OS is
allowed only for 4-byte PCI_VENDOR_ID config read request and only when
CRSSVE bit in Root Port PCIe device is enabled. In all other error PCIe
Root Complex must return all-ones.
So implement this logic in pci-aardvark.c driver properly.
aardvark HW does not have Root Port PCIe device and U-Boot does not
implement emulation of this device. So expect that CRSSVE bit is set as
U-Boot can already handle CRS value for PCI_VENDOR_ID config read request.
More callers of pci_bus_read_config() function in U-Boot do not check for
return value, but check readback value. Therefore always fill readback
value in pcie_advk_read_config() function. On error fill all-ones of
correct size as it is required for PCIe Root Complex.
And also correctly propagates error from failed config write request to
return value of pcie_advk_write_config() function. Most U-Boot callers
ignores this return value, but it is a good idea to return correct value
from function.
These issues about return value of failed config read requests, including
special handling of CRS were reported by Lorenzo and Bjorn for Linux kernel
driver pci-aardvark together with quotes from PCIe r4.0 spec, see details:
https://lore.kernel.org/linux-pci/20210624213345.3617-1-pali@kernel.org/t/#u
Signed-off-by: Pali Rohár <pali@kernel.org>
Reviewed-by: Stefan Roese <sr@denx.de>
Remapped address of PCIe outbound window may have set only bits from the
mask. Add additional check that remapped address which is calculated from
PCIe bus address specified in DTS file is valid.
Remove also useless clearing of low 16 bits in win_mask. As win_size is
power of two and is at least 0x10000 it means that it always has zero low
16 bits.
Signed-off-by: Pali Rohár <pali@kernel.org>
Reviewed-by: Konstantin Porotchkin <kostap@marvell.com>
Reviewed-by: Stefan Roese <sr@denx.de>
The `ranges` DT property of the PCIe node is currently ignored by
Aardvark driver - all entries are used as transparent PCIe MEM, despite
some of them being defined for IO in DT.
This is because the driver does not setup PCIe outbound windows and thus
a default configuration is used.
This can cause an external abort on CPU when a device driver tries to
access non-MEM space.
Setup the PCIe windows according to the `ranges` property for all
non-MEM resources (currently only IO) and also non-transparent MEM
resources.
Because Linux expects that bootloader does not setup Aardvark PCIe
windows, disable them before booting Linux.
Signed-off-by: Pali Rohár <pali@kernel.org>
Reviewed-by: Marek Behún <marek.behun@nic.cz>
Reviewed-by: Stefan Roese <sr@denx.de>
Change DT compatible string for A3700 PCIe from 'marvell,armada-37xx-pcie'
to 'marvell,armada-3700-pcie' to make U-Boot A3700 PCIe DT node compatible
with Linux' DT node.
Signed-off-by: Pali Rohár <pali@kernel.org>
Reviewed-by: Marek Behún <marek.behun@nic.cz>
Reviewed-by: Stefan Roese <sr@denx.de>
Disable Root Bridge I/O space, memory space and bus mastering in Aardvark's
remove method, which is called before booting Linux kernel.
This ensures that PCIe device which was initialized and used by U-Boot
cannot do new DMA transfers until Linux initializes PCI subsystem and loads
appropriate drivers for the device.
During initialization of PCI subsystem Linux in fact disables this bus
mastering on Root Bridge (and later enables it when driver is loaded and
configured), but there is a possibility of a small window after U-Boot
boots Linux when bus mastering is enabled, which is not correct.
Signed-off-by: Pali Rohár <pali@kernel.org>
Reviewed-by: Marek Behún <marek.behun@nic.cz>
Reviewed-by: Stefan Roese <sr@denx.de>
During our debugging of the Aardvark driver in Linux we have discovered
that the PCIE_CORE_LINK_CTRL_STAT_REG register in fact controls standard
PCIe Link Control Register for PCIe Root Bridge. This led us to discover
that the name of the PCIE_CORE_LINK_TRAINING macro and the corresponding
comment by this macro's usage is misleading; this bit in fact controls
Retrain Link, which, according to PCIe base spec is defined as:
A write of 1b to this bit initiates Link retraining by directing the
Physical Layer LTSSM to the Recovery state. If the LTSSM is already in
Recovery or Configuration, re-entering Recovery is permitted but not
required.
Entering Recovery state is normally done from LTSSM L0, L0s and L1 states.
But since the pci-aardvark.c driver enables Link Training just a few lines
above, the controller is not in L0 ready state yet. So setting aardvark bit
PCIE_CORE_LINK_TRAINING does not actually enter Recovery state at this
place.
Moreover, trying to enter LTSSM Recovery state without other configuration
is causing issues for some cards (e.g. Atheros AR9xxx and QCA9xxx). Since
Recovery state is not entered, these issues are not triggered.
Remove code which tries to enter LTSSM Recovery state completely.
Signed-off-by: Pali Rohár <pali@kernel.org>
Reviewed-by: Marek Behún <marek.behun@nic.cz>
Reviewed-by: Stefan Roese <sr@denx.de>
Trying to clear PIO_START register when it is non-zero (which indicates
that previous PIO transfer has not finished yet) causes an External
Abort with SError 0xbf000002.
This bug is currently worked around in TF-A by handling External Aborts
in EL3 and ignoring this particular SError.
This workaround was also discussed at:
https://git.trustedfirmware.org/TF-A/trusted-firmware-a.git/commit/?id=3c7dcdac5c50https://lore.kernel.org/linux-pci/20190316161243.29517-1-repk@triplefau.lt/https://lore.kernel.org/linux-pci/971be151d24312cc533989a64bd454b4@www.loen.fr/https://review.trustedfirmware.org/c/TF-A/trusted-firmware-a/+/1541
Implement a proper fix to prevent this External Abort. As it is not
possible to cancel a pending PIO transfer, simply do not start a new one
if previous has not finished yet. In this case return an error to the
caller.
In most cases this SError happens when there is no PCIe card connected
or when PCIe link is down. The reason is that in these cases a PIO
transfer takes about 1.44 seconds. For this reason we also increase the
wait timeout in pcie_advk_wait_pio() to 1.5 seconds.
If PIO read transfer for PCI_VENDOR_ID register times out, or if it
isn't possible to read it yet because previous transfer is not finished,
return Completion Retry Status value instead of failing, to give the
caller a chance to send a new read request.
Signed-off-by: Pali Rohár <pali@kernel.org>
Reviewed-by: Marek Behún <marek.behun@nic.cz>
Reviewed-by: Stefan Roese <sr@denx.de>
Marvell Armada 3720 Functional Errata, Guidelines, and Restrictions
document describes in erratum 4.1 PCIe value of vendor ID (Ref #: 243):
The readback value of VEND_ID (RD0070000h [15:0]) is 1B4Bh, while it
should read 11ABh.
The firmware can write the correct value, 11ABh, through VEND_ID
(RD0076044h [15:0]).
Implement this workaround in U-Boot PCIe controller driver aardvark for
both PCI vendor id and PCI subsystem vendor id.
This change affects PCI vendor id of PCIe root bridge emulated by Linux
kernel. With this change Linux kernel reports correct vendor id 11AB.
Signed-off-by: Pali Rohár <pali@kernel.org>
Reviewed-by: Stefan Roese <sr@denx.de>
Fix usage of VL805 XHCI PCIe controller when it is connected via PCIe to
Armada 3720 SOC. Without this U-Boot crashes when trying to access
enumerated USB devices connected to this XHCI PCIe controller.
This should be done according to the PCIe Link Initialization sequence, as
defined in Marvell Armada 3720 Functional Specification.
Linux has this code too.
Signed-off-by: Pali Rohár <pali@kernel.org>
Reviewed-by: Marek Behún <marek.behun@nic.cz>
Reviewed-by: Stefan Roese <sr@denx.de>
Value pcie->dev in function pcie_advk_probe() is not initialized yet
so use dev as argument for dev_dbg()/dev_warn().
Function pcie_advk_wait_pio() itself prints error message on failure
so do not print duplicate error message in caller.
Signed-off-by: Pali Rohár <pali@kernel.org>
Reviewed-by: Stefan Roese <sr@denx.de>
This reverts commit 828d326216.
This change revers code which asserting PERST# signal when unloading
driver. Driver's remove callback is still there as it is used for other
functionality.
Asserting PERST# signal prior booting kernel is causing that A3720 boards
(Turris MOX and Espressobin) with stable Linux kernel versions 4.14 and
4.19 are not able to detect some PCIe cards (e.g. Compex WLE200 and WLE900)
and anymore. When PERST# signal is not asserted these cards are detected
correctly. As this is regression for existing stable Linux kernel versions
revert this problematic change in U-Boot.
To make cards working with OpenWRT 4.14 kernel it is needed to disable link
training prior booting kernel, which is already done in driver's remove
callback.
Described issue is in Linux kernel pci aardvark driver which is (hopefully)
fixed in latest upstream versions. Latest upstream versions should be able
to initialize PCIe bus and detects cards independently of the link training
and PERST# signal state.
So with this change, U-Boot on A3720 boards should be able to boot OpenWRT
4.14 kernel, stable 4.14 and 4.19 kernels and also latest mainline kernels.
Signed-off-by: Pali Rohár <pali@kernel.org>
Reviewed-by: Stefan Roese <sr@denx.de>
At present various drivers etc. access the device's 'seq' member directly.
This makes it harder to change the meaning of that member. Change access
to go through a function instead.
The drivers/i2c/lpc32xx_i2c.c file is left unchanged for now.
Signed-off-by: Simon Glass <sjg@chromium.org>
This name is far too long. Rename it to remove the 'data' bits. This makes
it consistent with the platdata->plat rename.
Signed-off-by: Simon Glass <sjg@chromium.org>
This construct is quite long-winded. In earlier days it made some sense
since auto-allocation was a strange concept. But with driver model now
used pretty universally, we can shorten this to 'auto'. This reduces
verbosity and makes it easier to read.
Coincidentally it also ensures that every declaration is on one line,
thus making dtoc's job easier.
Signed-off-by: Simon Glass <sjg@chromium.org>
As required by PCI Express spec a delay for at least 100ms after
de-asserting PERST# signal is needed before link training is enabled.
Linux kernels prior to 5.8 version do not automatically disable link
training before de-asserting PERST# signal, therefore this requirement is
not fulfilled.
Above requirement is needed for proper detection of some Compex PCIe WiFi
cards. Otherwise Linux kernel cannot detect it.
To allow using those PCIe cards with older Linux kernel versions booted by
U-Boot compiled with U-Boot a37xx pci driver, disable link training in
U-Boot when unloading this pci driver.
Thanks to DM_FLAG_OS_PREPARE flag, U-Boot automatically unload this driver
when booting Linux kernel.
Signed-off-by: Pali Rohár <pali@kernel.org>
Reviewed-by: Stefan Roese <sr@denx.de>
For proper initialization of aardvark pci driver it is required to
de-assert reset GPIO. So depeneds on DM_GPIO option.
Signed-off-by: Pali Rohár <pali@kernel.org>
Reviewed-by: Stefan Roese <sr@denx.de>
Tested-by: Andre Heider <a.heider@gmail.com>
This change ensures that PCIe card is put into reset state when U-Boot
stops using it.
DM_FLAG_OS_PREPARE ensures that U-Boot executes driver's remove callback
prior booting Linux kernel.
Linux kernel pci-aardvark driver needs to reset PCIe card via PERST# signal
prior initializing it. If it does not issue reset then some PCIe cards
(specially Compex WiFi cards) are not detected at all.
Putting PCIe card into reset state prior booting Linux kernel would ensure
that card would be properly reset at time when Linux kernel starts
initializing pci-aardvark driver.
Signed-off-by: Pali Rohár <pali@kernel.org>
Reviewed-by: Konstantin Porotchkin <kostap@marvell.com>
Tested-by: Andre Heider <a.heider@gmail.com>
Change active-high to active-low and change DT property name from
reset-gpio to reset-gpios. This format of gpio reset is used by
pci-aardvark driver in Linux kernel.
Signed-off-by: Pali Rohár <pali@kernel.org>
Reviewed-by: Stefan Roese <sr@denx.de>
Reviewed-by: Konstantin Porotchkin <kostap@marvell.com>
Tested-by: Andre Heider <a.heider@gmail.com>
At present dm/device.h includes the linux-compatible features. This
requires including linux/compat.h which in turn includes a lot of headers.
One of these is malloc.h which we thus end up including in every file in
U-Boot. Apart from the inefficiency of this, it is problematic for sandbox
which needs to use the system malloc() in some files.
Move the compatibility features into a separate header file.
Signed-off-by: Simon Glass <sjg@chromium.org>
At present this method uses a non-const udevice pointer, but the call
should not modify the device. Use a const pointer.
Signed-off-by: Simon Glass <sjg@chromium.org>
At present if CONFIG_SPL_GPIO_SUPPORT is enabled then the GPIO uclass
is included in SPL/TPL without any control for boards. Some boards may
want to disable this to reduce code size where GPIOs are not needed in
SPL or TPL.
Add a new Kconfig option to permit this. Default it to 'y' so that
existing boards work correctly.
Change existing uses of CONFIG_DM_GPIO to CONFIG_IS_ENABLED(DM_GPIO) to
preserve the current behaviour. Also update the 74x164 GPIO driver since
it cannot build with SPL.
This allows us to remove the hacks in config_uncmd_spl.h and
Makefile.uncmd_spl (eventually those files should be removed).
Signed-off-by: Simon Glass <sjg@chromium.org>
Reviewed-by: Bin Meng <bmeng.cn@gmail.com>
We have a large number of places where while we historically referenced
gd in the code we no longer do, as well as cases where the code added
that line "just in case" during development and never dropped it.
Signed-off-by: Tom Rini <trini@konsulko.com>
This patch introduced the Aardvark PCIe driver based
driver model.
The PCIe driver is supposed to work in Root Complex
mode. It only supports X1 lane width.
Signed-off-by: Wilson Ding <dingwei@marvell.com>
Reviewed-on: http://vgitil04.il.marvell.com:8080/38725
Reviewed-by: Victor Gu <xigu@marvell.com>
Reviewed-by: Hua Jing <jinghua@marvell.com>
Tested-by: Hua Jing <jinghua@marvell.com>
Cc: Simon Glass <sjg@chromium.org>
Cc: Stefan Roese <sr@denx.de>
Signed-off-by: Ken Ma <make@marvell.com>
Signed-off-by: Stefan Roese <sr@denx.de>
