[PATCH] kernel: Fix ath79 DSP exception at bootup

Hauke Mehrtens hauke at hauke-m.de
Thu Jul 2 18:54:38 EDT 2020

This resolves a hazard between a mtc0 and a mfc0 instruction after
activating the DSP support. Without this fix the CPU could use the old
value again and the DSP support would not be active.

Fixes: FS#2928
Signed-off-by: Hauke Mehrtens <hauke at hauke-m.de>
 ...-EHB-in-mtc0-mfc0-sequence-for-DSPen.patch | 61 +++++++++++++++++++
 1 file changed, 61 insertions(+)
 create mode 100644 target/linux/generic/pending-5.4/350-MIPS-Add-missing-EHB-in-mtc0-mfc0-sequence-for-DSPen.patch

diff --git a/target/linux/generic/pending-5.4/350-MIPS-Add-missing-EHB-in-mtc0-mfc0-sequence-for-DSPen.patch b/target/linux/generic/pending-5.4/350-MIPS-Add-missing-EHB-in-mtc0-mfc0-sequence-for-DSPen.patch
new file mode 100644
index 000000000000..063ec0e9fb5b
--- /dev/null
+++ b/target/linux/generic/pending-5.4/350-MIPS-Add-missing-EHB-in-mtc0-mfc0-sequence-for-DSPen.patch
@@ -0,0 +1,61 @@
+From db4603e30effd74d4adb6bcdf73072b2c06fafcd Mon Sep 17 00:00:00 2001
+From: Hauke Mehrtens <hauke at hauke-m.de>
+Date: Fri, 3 Jul 2020 00:07:15 +0200
+Subject: [PATCH] MIPS: Add missing EHB in mtc0 -> mfc0 sequence for DSPen
+This resolves the hazard between the mtc0 in the change_c0_status() and
+the mfc0 in configure_exception_vector(). Without resolving this hazard
+configure_exception_vector() could read an old value and would restore
+this old value again. This would revert the changes change_c0_status()
+did. I checked this by printing out the read_c0_status() at the end of
+per_cpu_trap_init() and the ST0_MX is not set without this patch.
+The hazard is documented in the MIPS Architecture Reference Manual Vol.
+III: MIPS32/microMIPS32 Privileged Resource Architecture (MD00088), rev
+6.03 table 8.1 which includes:
+   Producer | Consumer | Hazard
+  ----------|----------|----------------------------
+   mtc0     | mfc0     | any coprocessor 0 register
+I saw this hazard on an Atheros AR9344 rev 2 SoC with a MIPS 74Kc CPU.
+There the change_c0_status() function would activate the DSPen by
+setting ST0_MX in the c0_status register. This was reverted and then the
+system got a DSP exception when the DSP registers were saved in
+save_dsp() in the first process switch. The crash looks like this:
+[    0.089999] Mount-cache hash table entries: 1024 (order: 0, 4096 bytes, linear)
+[    0.097796] Mountpoint-cache hash table entries: 1024 (order: 0, 4096 bytes, linear)
+[    0.107070] Kernel panic - not syncing: Unexpected DSP exception
+[    0.113470] Rebooting in 1 seconds..
+We saw this problem in OpenWrt only on the MIPS 74Kc based Atheros SoCs,
+not on the 24Kc based SoCs. We only saw it with kernel 5.4 not with
+kernel 4.19, in addition we had to use GCC 8.4 or 9.X, with GCC 8.3 it
+did not happen.
+In the kernel I bisected this problem to commit 9012d011660e ("compiler:
+allow all arches to enable CONFIG_OPTIMIZE_INLINING"), but when this was
+reverted it also happened after commit 172dcd935c34b ("MIPS: Always
+allocate exception vector for MIPSr2+").
+Commit 0b24cae4d535 ("MIPS: Add missing EHB in mtc0 -> mfc0 sequence.")
+does similar changes to a different file. I am not sure if there are
+more places affected by this problem.
+Signed-off-by: Hauke Mehrtens <hauke at hauke-m.de>
+Cc: <stable at vger.kernel.org>
+ arch/mips/kernel/traps.c | 1 +
+ 1 file changed, 1 insertion(+)
+--- a/arch/mips/kernel/traps.c
++++ b/arch/mips/kernel/traps.c
+@@ -2126,6 +2126,7 @@ static void configure_status(void)
+ 	change_c0_status(ST0_CU|ST0_MX|ST0_RE|ST0_FR|ST0_BEV|ST0_TS|ST0_KX|ST0_SX|ST0_UX,
+ 			 status_set);
++	back_to_back_c0_hazard();
+ }
+ unsigned int hwrena;

More information about the openwrt-devel mailing list