Quoting Christophe Leroy <christophe.le...@csgroup.eu>:
Le 03/11/2020 à 14:29, Michael Ellerman a écrit :
Andreas reported that commit ee0a49a6870e ("powerpc/uaccess: Switch
__put_user_size_allowed() to __put_user_asm_goto()") broke
CLONE_CHILD_SETTID.
Further inspection showed that the put_user() in schedule_tail() was
missing entirely, the store not emitted by the compiler.
Notice there are no stores other than to the stack. There should be a
stw in there for the store to current->set_child_tid.
This is only seen with GCC 4.9 era compilers (tested with 4.9.3 and
4.9.4), and only when CONFIG_PPC_KUAP is disabled.
When CONFIG_PPC_KUAP=y, the memory clobber that's part of the isync()
and mtspr() inlined via allow_user_access() seems to be enough to
avoid the bug.
For now though let's just not use asm goto with GCC 4.9, to avoid this
bug and any other issues we haven't noticed yet. Possibly in future we
can find a smaller workaround.
Is that https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58670 ?
Should we use asm_volatile_goto() defined in include/linux/compiler-gcc.h ?
It seems to be OK with asm_volatile_goto() with GCC 4.9, and it is
already what is used in our asm/jump_label.h
diff --git a/arch/powerpc/include/asm/uaccess.h
b/arch/powerpc/include/asm/uaccess.h
index ef5bbb705c08..501c9a79038c 100644
--- a/arch/powerpc/include/asm/uaccess.h
+++ b/arch/powerpc/include/asm/uaccess.h
@@ -178,7 +178,7 @@ do {
\
* are no aliasing issues.
*/
#define __put_user_asm_goto(x, addr, label, op) \
- asm volatile goto( \
+ asm_volatile_goto( \
"1: " op "%U1%X1 %0,%1 # put_user\n" \
EX_TABLE(1b, %l2) \
: \
@@ -191,7 +191,7 @@ do {
\
__put_user_asm_goto(x, ptr, label, "std")
#else /* __powerpc64__ */
#define __put_user_asm2_goto(x, addr, label) \
- asm volatile goto( \
+ asm_volatile_goto( \
"1: stw%X1 %0, %1\n" \
"2: stw%X1 %L0, %L1\n" \
EX_TABLE(1b, %l2) \
---
Christophe
