Commit f7354ccac844 ("powerpc/32: Remove CURRENT_THREAD_INFO and
rename TI_CPU") broke the CPU wake-up from sleep mode (i.e. when
_TLF_SLEEPING is set) by delaying the tovirt(r2, r2).
This is because r2 is not restored by fast_exception_return. It used
to work (by chance ?) because CPU wake-up interrupt never comes from
user, so r2 is expected to point to 'current' on return.
Commit e2fb9f544431 ("powerpc/32: Prepare for Kernel Userspace Access
Protection") broke it even more by clobbering r0 which is not
restored by fast_exception_return either.
Use r6 instead of r0. This is possible because r3-r6 are restored by
fast_exception_return and only r3-r5 are used for exception arguments.
For r2 it could be converted back to virtual address, but stay on the
safe side and restore it from the stack instead. It should be live
in the cache at that moment, so loading from the stack should make
no difference compared to converting it from phys to virt.
Fixes: f7354ccac844 ("powerpc/32: Remove CURRENT_THREAD_INFO and rename TI_CPU")
Fixes: e2fb9f544431 ("powerpc/32: Prepare for Kernel Userspace Access Protection")
Cc: [email protected]
Signed-off-by: Christophe Leroy <[email protected]>
---
arch/powerpc/kernel/entry_32.S | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/arch/powerpc/kernel/entry_32.S b/arch/powerpc/kernel/entry_32.S
index 73b80143ffac..27e2afce8b78 100644
--- a/arch/powerpc/kernel/entry_32.S
+++ b/arch/powerpc/kernel/entry_32.S
@@ -180,7 +180,7 @@ transfer_to_handler:
2: /* if from kernel, check interrupted DOZE/NAP mode and
* check for stack overflow
*/
- kuap_save_and_lock r11, r12, r9, r2, r0
+ kuap_save_and_lock r11, r12, r9, r2, r6
addi r2, r12, -THREAD
#ifndef CONFIG_VMAP_STACK
lwz r9,KSP_LIMIT(r12)
@@ -288,6 +288,7 @@ reenable_mmu:
rlwinm r9,r9,0,~MSR_EE
lwz r12,_LINK(r11) /* and return to address in LR */
kuap_restore r11, r2, r3, r4, r5
+ lwz r2, GPR2(r11)
b fast_exception_return
#endif
--
2.25.0
On Mon, 2020-01-27 at 10:42:04 UTC, Christophe Leroy wrote:
> Commit f7354ccac844 ("powerpc/32: Remove CURRENT_THREAD_INFO and
> rename TI_CPU") broke the CPU wake-up from sleep mode (i.e. when
> _TLF_SLEEPING is set) by delaying the tovirt(r2, r2).
>
> This is because r2 is not restored by fast_exception_return. It used
> to work (by chance ?) because CPU wake-up interrupt never comes from
> user, so r2 is expected to point to 'current' on return.
>
> Commit e2fb9f544431 ("powerpc/32: Prepare for Kernel Userspace Access
> Protection") broke it even more by clobbering r0 which is not
> restored by fast_exception_return either.
>
> Use r6 instead of r0. This is possible because r3-r6 are restored by
> fast_exception_return and only r3-r5 are used for exception arguments.
>
> For r2 it could be converted back to virtual address, but stay on the
> safe side and restore it from the stack instead. It should be live
> in the cache at that moment, so loading from the stack should make
> no difference compared to converting it from phys to virt.
>
> Fixes: f7354ccac844 ("powerpc/32: Remove CURRENT_THREAD_INFO and rename TI_CPU")
> Fixes: e2fb9f544431 ("powerpc/32: Prepare for Kernel Userspace Access Protection")
> Cc: [email protected]
> Signed-off-by: Christophe Leroy <[email protected]>
Applied to powerpc next, thanks.
https://git.kernel.org/powerpc/c/9933819099c4600b41a042f27a074470a43cf6b9
cheers