Before sleeping, low_sleep_handler() stores the ROM wakeup vectors and
the 'Lars' magic at physical addresses 0x0-0x84. It does so through the
kernel linear mapping at KERNELBASE with data relocation on. With
CONFIG_STRICT_KERNEL_RWX, that page is the start of kernel text and is
mapped read-only, so the first store takes a DSI and PMU sleep oopses
with interrupts disabled, leaving the machine hung:

  NIP [c0033018] low_sleep_handler+0x198/0x1b4
  LR [c0032ff4] low_sleep_handler+0x174/0x1b4
  DAR: c0000000 DSISR: 0a000000
  [f2169d20] [c086df18] powerbook_sleep+0x86c/0xbe4 (unreliable)
  [f2169db0] [c00b1178] suspend_devices_and_enter+0x680/0x700
  [f2169e00] [c00b1428] pm_suspend+0x230/0x25c
  [f2169e20] [c00af7dc] state_store+0x74/0xd0

Turn off MSR[DR] around the stores and write to physical address 0
directly, as flush_disable_75x() and flush_disable_745x() already do
for their cache flushes. The ROM reads these locations with the MMU
off, so physical address 0 is where they belong anyway.

The stores stay ahead of the low_cpu_offline_self entry point, so the
CPU hotplug path is unchanged. pmac32-cpufreq also calls
low_sleep_handler() and gets the same fix.

The Linux PTE for page 0 has been read-only since STRICT_KERNEL_RWX was
first enabled for PPC32, and the BATs since commit 63b2bc619565
("powerpc/mm/32s: Use BATs for STRICT_KERNEL_RWX").

Tested with suspend to RAM on an iMac G4 (PowerMac6,1, 7448). Not
tested on Grackle machines or with pmac32-cpufreq.

Fixes: 95902e6c8864 ("powerpc/mm: Implement STRICT_KERNEL_RWX on PPC32")
Signed-off-by: Matt Turner <[email protected]>
---
 arch/powerpc/platforms/powermac/sleep.S | 20 ++++++++++++++++----
 1 file changed, 16 insertions(+), 4 deletions(-)

diff --git a/arch/powerpc/platforms/powermac/sleep.S 
b/arch/powerpc/platforms/powermac/sleep.S
index 822ed70cdcbf..750cb0160b4c 100644
--- a/arch/powerpc/platforms/powermac/sleep.S
+++ b/arch/powerpc/platforms/powermac/sleep.S
@@ -184,19 +184,31 @@ END_MMU_FTR_SECTION_IFSET(MMU_FTR_USE_HIGH_BATS)
        addi    r5,r5,grackle_wake_up@l
        tophys(r5,r5)
        stw     r5,SL_PC(r11)
-       lis     r4,KERNELBASE@h
        tophys(r5,r11)
        addi    r5,r5,SL_PC
        lis     r6,MAGIC@ha
        addi    r6,r6,MAGIC@l
-       stw     r5,0(r4)
-       stw     r6,4(r4)
-       /* Setup stuffs at 0x80-0x84 for Core99 */
        lis     r3,core99_wake_up@ha
        addi    r3,r3,core99_wake_up@l
        tophys(r3,r3)
+
+       /* With STRICT_KERNEL_RWX, the kernel mapping of page 0 is
+        * read-only. Write the vectors with data relocation off.
+        */
+       mfmsr   r7
+       rlwinm  r8,r7,0,~MSR_DR
+       sync
+       mtmsr   r8
+       isync
+       li      r4,0
+       stw     r5,0(r4)
+       stw     r6,4(r4)
+       /* Setup stuffs at 0x80-0x84 for Core99 */
        stw     r3,0x80(r4)
        stw     r5,0x84(r4)
+       sync
+       mtmsr   r7
+       isync
 
        .globl  low_cpu_offline_self
 low_cpu_offline_self:

---
base-commit: 4405c18c95d2d0f99f6c581ef0ee8591a1aa6ccd
change-id: 20261119-pmac-sleep-strict-rwx-57e5dcacee55

Best regards,
--  
Matt Turner <[email protected]>


Reply via email to