[OpenWrt-Devel] [PATCH] ramips: Fix pcie interrupt mapping for ZBT WE1326

[Kristian Evensen]

Hi,

On Tue, May 29, 2018 at 9:21 PM, Kristian Evensen
<kristian.evensen at gmail.com> wrote:
> +       interrupt-map = <0x10000 0 0 1 &gic GIC_SHARED 24 IRQ_TYPE_LEVEL_HIGH>,
> +                       <0x20000 0 0 1 &gic GIC_SHARED 25 IRQ_TYPE_LEVEL_HIGH>;
> +

Now that the offending commit has been reverted, this patch can be
ignored. However, I thought I should provide some context on this
change, as the recent discussions on the forum triggered me to do a
more thorough investigation of what actually went wrong. Any feedback,
or plain "YOU ARE WRONG", are very much welcome as I am not very
experience with dts, setting up pci and interrupts, etc.

Before commit 5f7396ebef09, the IRQ was set inside a large if-clause
(removed by the mentioned commit). The wifi radios on the 1326 are on
PCIe bus 1, slot 0 (mt76x2) and bus 2, slot 1 (mt7603). When PCIe is
about to be enabled, pcie_link_status is six for both radios. The
RALINK_INT_PCIE-values map directly to interrupts specified for PCIe
in the dtsi-file. I.e., RALINK_INT_PCIE0 is GIC_SHARED 4, PCIE1 is 24
and PCIE2 is 25. Before 5f7396ebef09, the IRQ for mt76x2 was always
set to PCIE1 (24) and mt7603 to PCIE2 (25). /proc/interrupts looks as
follows on a working WE1326:

24: 2792 0 0 0 MIPS GIC 31 mt76x2e
25: 13123 0 0 0 MIPS GIC 32 mt7603e

The hardware interrupt numbers are calculated by doing
GIC_SHARED_HWIRQ_BASE + x (where X is either 4, 24 or 25).
GIC_SHARED_HWIRQ_BASE is defined as GIC_NUM_LOCAL_INTRS, which is
seven. This is why we see GIC 31 and 32. See
/drivers/irqchip/irq-mips-gic.c and /arch/mips/include/asm/mips-gic.h
in the kernel sources for more details about the calculations, values,
etc..

After commit 5f7396ebef09, the static IRQ assigned for the given
bus/slot combos are removed and the IRQ values from the DTS are used
directly. A consequence of this is that the first pci device (mt76x2)
in the 1326 is incorrectly assigned IRQ 11. /proc/interrupts after
5f7396ebef09 looks as follows:

24: 0 0 0 0 MIPS GIC 11 mt76x2e

The correct interrupts should be as above. I.e, mt76x2 should have IRQ
31 (GIC_SHARED 24) and mt7603 IRQ 32 (GIC_SHARED 25). By applying my
patch and thus update the interrupt map, the correct interrupt mapping
was restored. /proc/interrupts looks as follows:

23: 23 0 0 0 MIPS GIC 32 mt7603e
24: 3010 0 0 0 MIPS GIC 31 mt76x2e

In order to confirm that the IRQ-mapping on the WE1326 are different
from (some) other mt7621-devices, I checked the interrupt mapping on
the ZBT WG3526 (which worked before and after commit). It looks as
follows:

23: 84765 0 0 0 MIPS GIC 11 mt7603e
24: 27389 0 0 0 MIPS GIC 31 mt76x2e

I.e., mt7603 (first pci device) is GIC_SHARED 4 and mt76x2 GIC_SHARED
24. The paths to the pci-devices are the same on both WE1326 and
WG3526 (0000:01:00.0 and 0000:02:00.0).

BR,
Kristian

_______________________________________________
openwrt-devel mailing list
openwrt-devel at lists.openwrt.org
https://lists.openwrt.org/listinfo/openwrt-devel