openwrt/target/linux/layerscape/patches-5.4/701-net-0155-soc-fsl-dpio-aligned-access-of-qbman-cacheable-regio.patch
Yangbo Lu cddd459140 layerscape: add patches-5.4
Add patches for linux-5.4. The patches are from NXP LSDK-20.04 release
which was tagged LSDK-20.04-V5.4.
https://source.codeaurora.org/external/qoriq/qoriq-components/linux/

For boards LS1021A-IOT, and Traverse-LS1043 which are not involved in
LSDK, port the dts patches from 4.14.

The patches are sorted into the following categories:
  301-arch-xxxx
  302-dts-xxxx
  303-core-xxxx
  701-net-xxxx
  801-audio-xxxx
  802-can-xxxx
  803-clock-xxxx
  804-crypto-xxxx
  805-display-xxxx
  806-dma-xxxx
  807-gpio-xxxx
  808-i2c-xxxx
  809-jailhouse-xxxx
  810-keys-xxxx
  811-kvm-xxxx
  812-pcie-xxxx
  813-pm-xxxx
  814-qe-xxxx
  815-sata-xxxx
  816-sdhc-xxxx
  817-spi-xxxx
  818-thermal-xxxx
  819-uart-xxxx
  820-usb-xxxx
  821-vfio-xxxx

Signed-off-by: Yangbo Lu <yangbo.lu@nxp.com>
2020-05-07 12:53:06 +02:00

74 lines
2.9 KiB
Diff

From 353725d455cc27aa2008456c2711e7fae2b8a631 Mon Sep 17 00:00:00 2001
From: Bharat Bhushan <Bharat.Bhushan@nxp.com>
Date: Wed, 7 Mar 2018 21:35:37 +0530
Subject: [PATCH] soc: fsl: dpio: aligned access of qbman cacheable region
Alignment requirement on ARM is lenient (In Linux) for regions
mapped as "Memory Type" but have very strict policy for regions
mapped as "Device Type". Unaligned access to regions mapped
as "Device Type" will always result to unaligned fault.
DPIO driver have un-aligned access to QBman cacheable region
and the Linux driver maps the region as "Memory Type". On Host
Linux this works because MMU Stage-1 configured by driver allows
unaligned access. In Virtual Machine cases, final region mapping type
is governed by combination of Stage-1 and Stage-2 MMU mapping.
Linux driver in VM controls maps the region as "Memory Type" in
Stage-1 MMU while Stage-2 is controlled by KVM. And current KVM
implementation does not allow device region to be mapped as
"Memory Type". Till we have a working/upstream-able solution
for Virtual Machine, we need to change un-aligned access in DPIO
driver to be aligned
While we reached to this point as we observed below alignment
exception in Virtual Machine when accessing qbman cacheable region.
kvm [2347]: Unsupported FSC: EC=0x24 xFSC=0x21
ESR_EL2=0x92000061
error: kvm run failed Bad address
PC=ffff000008398e78 SP=ffff800009bcb540
X00=ffff000008041000 X01=ffff800009bcb580 X02=ffff800009bcb650
X03=0000000000000180
X04=ffff000008041001 X05=ffff800009bcb581 X06=0200000000000000
X07=0000000000000000
X08=0000000000000000 X09=ffff000008041000 X10=0000000000000001
X11=0000000000de6cb0
X12=00000000fa83b2da X13=0000000000000001 X14=000000007f605ec8
X15=00000000e26f5d5e
X16=000000008521af1e X17=000000001076277e X18=ffff800009bcb5c0
X19=ffff800079da2b00
X20=ffff800009bcb650 X21=0000000000000002 X22=0000000000000000
X23=0000000000000000
X24=0000000000000000 X25=ffff8000099e7440 X26=ffff000008da6000
X27=ffff000008e7f000
X28=00000000499e7440 X29=ffff800009bcb540 X30=ffff00000839a160
PSTATE=20000145 --C- EL1h
Signed-off-by: Bharat Bhushan <Bharat.Bhushan@nxp.com>
---
drivers/soc/fsl/dpio/qbman-portal.c | 12 +++++++++++-
1 file changed, 11 insertions(+), 1 deletion(-)
--- a/drivers/soc/fsl/dpio/qbman-portal.c
+++ b/drivers/soc/fsl/dpio/qbman-portal.c
@@ -515,7 +515,17 @@ int qbman_swp_enqueue(struct qbman_swp *
return -EBUSY;
p = qbman_get_cmd(s, QBMAN_CENA_SWP_EQCR(EQAR_IDX(eqar)));
- memcpy(&p->dca, &d->dca, 31);
+ /* This is mapped as DEVICE type memory, writes are
+ * with address alignment:
+ * desc.dca address alignment = 1
+ * desc.seqnum address alignment = 2
+ * desc.orpid address alignment = 4
+ * desc.tgtid address alignment = 8
+ */
+ p->dca = d->dca;
+ p->seqnum = d->seqnum;
+ p->orpid = d->orpid;
+ memcpy(&p->tgtid, &d->tgtid, 24);
memcpy(&p->fd, fd, sizeof(*fd));
if ((s->desc->qman_version & QMAN_REV_MASK) < QMAN_REV_5000) {