This is the 3rd iteration of the patch, here are the changes with regard to previous one: * more precise description of the problem. * only set r30 in PIC+secureplt case as it seems indeed to be the only case which causes problem. Secureplt without PIC does not need r30. * don't rely on uclibc being already built: now we check for secure-plt by doing $(CC) --verbose and searching for the ./configure parameters of GCC for "--enable-secureplt".
Yann Sionneau (1): powerpc: fix PIE/PIC builds with newer gcc/binutils which use secureplt by default
Rules.mak | 3 ++- ldso/ldso/powerpc/dl-startup.h | 3 +++ libc/sysdeps/linux/powerpc/crt1.S | 4 ++++ 3 files changed, 9 insertions(+), 1 deletion(-)
This patch fixes segfault of all user space processes (including init, which caused a panic) on recent buildroot powerpc32 builds.
The issue has been reported by Romain Naour in this thread: https://mailman.uclibc-ng.org/pipermail/devel/2021-May/002068.html
Recent buildroot toolchain enables secure PLT in powerpc gcc. The latter will then supply -msecure-plt to gas invocations by default. Recent buildroot also enables PIE by default.
For the secure PLT to work in PIC, the r30 register needs to point to the GOT. Old "bss plt" was just a one-instruction-wide PLT slot, pointed-to by a R_PPC_JMP_SLOT relocation, which was written on-the-fly to contain a branch instruction to the correct address. It therefore had to stay writable+executable, which you generally want to avoid for security reasons. New secure PLT only contains read-only code which loads the branch address from the writable GOT.
Note: secure PLT without PIC does not need r30 to be set. Because offset between plt stub code and got is known at link-time. In this case the PLT entry looks like: 1009b3e0 __uClibc_main@plt: 1009b3e0: 3d 60 10 0e lis r11,4110 1009b3e4: 81 6b 03 74 lwz r11,884(r11) 1009b3e8: 7d 69 03 a6 mtctr r11 1009b3ec: 4e 80 04 20 bctr
Whereas secure PLT with PIC - offset between plt and got is unknown at link-time - looks like this: 000af800 <00000000.plt_pic32.__uClibc_main>: af800: 81 7e 03 80 lwz r11,896(r30) af804: 7d 69 03 a6 mtctr r11 af808: 4e 80 04 20 bctr af80c: 60 00 00 00 nop
Signed-off-by: Yann Sionneau yann@sionneau.net --- Rules.mak | 3 ++- ldso/ldso/powerpc/dl-startup.h | 3 +++ libc/sysdeps/linux/powerpc/crt1.S | 4 ++++ 3 files changed, 9 insertions(+), 1 deletion(-)
diff --git a/Rules.mak b/Rules.mak index 1fa09be23..c789aefd3 100644 --- a/Rules.mak +++ b/Rules.mak @@ -483,9 +483,10 @@ ifeq ($(TARGET_ARCH),powerpc) PICFLAG:=-fpic PIEFLAG_NAME:=-fpie PPC_HAS_REL16:=$(shell printf "\t.text\n\taddis 11,30,_GLOBAL_OFFSET_TABLE_-.@ha\n" | $(CC) -c -x assembler -o /dev/null - 2> /dev/null && echo -n y || echo -n n) + PPC_HAS_SECUREPLT:=$(shell $(CC) --verbose 2>&1 | grep -- --enable-secureplt > /dev/null && echo -n y || echo -n n) + CPU_CFLAGS-$(PPC_HAS_SECUREPLT) += -DPPC_HAS_SECUREPLT CPU_CFLAGS-$(PPC_HAS_REL16)+= -DHAVE_ASM_PPC_REL16 CPU_CFLAGS-$(CONFIG_E500) += "-D__NO_MATH_INLINES" - endif
ifeq ($(TARGET_ARCH),bfin) diff --git a/ldso/ldso/powerpc/dl-startup.h b/ldso/ldso/powerpc/dl-startup.h index 8b2a517e2..7749395eb 100644 --- a/ldso/ldso/powerpc/dl-startup.h +++ b/ldso/ldso/powerpc/dl-startup.h @@ -25,6 +25,9 @@ __asm__( #else " bl _GLOBAL_OFFSET_TABLE_-4@local\n" /* Put our GOT pointer in r31, */ " mflr 31\n" +#endif +#ifdef PPC_HAS_SECUREPLT + " mr 30,31\n" #endif " addi 1,1,16\n" /* Restore SP */ " lwz 7,_dl_skip_args@got(31)\n" /* load EA of _dl_skip_args */ diff --git a/libc/sysdeps/linux/powerpc/crt1.S b/libc/sysdeps/linux/powerpc/crt1.S index 27bfc5a5a..3f5d056c0 100644 --- a/libc/sysdeps/linux/powerpc/crt1.S +++ b/libc/sysdeps/linux/powerpc/crt1.S @@ -56,6 +56,10 @@ _start: # else bl _GLOBAL_OFFSET_TABLE_-4@local mflr r31 +# endif + /* in PIC/PIE, plt stubs need r30 to point to the GOT if using secure-plt */ +# ifdef PPC_HAS_SECUREPLT + mr 30,31 # endif #endif /* Set up the small data pointer in r13. */
Hello Yann,
Le 28/05/2021 à 08:59, Yann Sionneau a écrit :
This patch fixes segfault of all user space processes (including init, which caused a panic) on recent buildroot powerpc32 builds.
The issue has been reported by Romain Naour in this thread: https://mailman.uclibc-ng.org/pipermail/devel/2021-May/002068.html
Recent buildroot toolchain enables secure PLT in powerpc gcc. The latter will then supply -msecure-plt to gas invocations by default. Recent buildroot also enables PIE by default.
For the secure PLT to work in PIC, the r30 register needs to point to the GOT. Old "bss plt" was just a one-instruction-wide PLT slot, pointed-to by a R_PPC_JMP_SLOT relocation, which was written on-the-fly to contain a branch instruction to the correct address. It therefore had to stay writable+executable, which you generally want to avoid for security reasons. New secure PLT only contains read-only code which loads the branch address from the writable GOT.
Note: secure PLT without PIC does not need r30 to be set. Because offset between plt stub code and got is known at link-time. In this case the PLT entry looks like: 1009b3e0 __uClibc_main@plt: 1009b3e0: 3d 60 10 0e lis r11,4110 1009b3e4: 81 6b 03 74 lwz r11,884(r11) 1009b3e8: 7d 69 03 a6 mtctr r11 1009b3ec: 4e 80 04 20 bctr
Whereas secure PLT with PIC - offset between plt and got is unknown at link-time - looks like this: 000af800 <00000000.plt_pic32.__uClibc_main>: af800: 81 7e 03 80 lwz r11,896(r30) af804: 7d 69 03 a6 mtctr r11 af808: 4e 80 04 20 bctr af80c: 60 00 00 00 nop
Signed-off-by: Yann Sionneau yann@sionneau.net
Rules.mak | 3 ++- ldso/ldso/powerpc/dl-startup.h | 3 +++ libc/sysdeps/linux/powerpc/crt1.S | 4 ++++ 3 files changed, 9 insertions(+), 1 deletion(-)
diff --git a/Rules.mak b/Rules.mak index 1fa09be23..c789aefd3 100644 --- a/Rules.mak +++ b/Rules.mak @@ -483,9 +483,10 @@ ifeq ($(TARGET_ARCH),powerpc) PICFLAG:=-fpic PIEFLAG_NAME:=-fpie PPC_HAS_REL16:=$(shell printf "\t.text\n\taddis 11,30,_GLOBAL_OFFSET_TABLE_-.@ha\n" | $(CC) -c -x assembler -o /dev/null - 2> /dev/null && echo -n y || echo -n n)
- PPC_HAS_SECUREPLT:=$(shell $(CC) --verbose 2>&1 | grep -- --enable-secureplt > /dev/null && echo -n y || echo -n n)
I'm not sure if this is the recommended way to check for a compiler support but I don't have a better proposal.
- CPU_CFLAGS-$(PPC_HAS_SECUREPLT) += -DPPC_HAS_SECUREPLT CPU_CFLAGS-$(PPC_HAS_REL16)+= -DHAVE_ASM_PPC_REL16 CPU_CFLAGS-$(CONFIG_E500) += "-D__NO_MATH_INLINES"
endif
ifeq ($(TARGET_ARCH),bfin) diff --git a/ldso/ldso/powerpc/dl-startup.h b/ldso/ldso/powerpc/dl-startup.h index 8b2a517e2..7749395eb 100644 --- a/ldso/ldso/powerpc/dl-startup.h +++ b/ldso/ldso/powerpc/dl-startup.h @@ -25,6 +25,9 @@ __asm__( #else " bl _GLOBAL_OFFSET_TABLE_-4@local\n" /* Put our GOT pointer in r31, */ " mflr 31\n" +#endif +#ifdef PPC_HAS_SECUREPLT
- " mr 30,31\n"
#endif " addi 1,1,16\n" /* Restore SP */ " lwz 7,_dl_skip_args@got(31)\n" /* load EA of _dl_skip_args */ diff --git a/libc/sysdeps/linux/powerpc/crt1.S b/libc/sysdeps/linux/powerpc/crt1.S index 27bfc5a5a..3f5d056c0 100644 --- a/libc/sysdeps/linux/powerpc/crt1.S +++ b/libc/sysdeps/linux/powerpc/crt1.S @@ -56,6 +56,10 @@ _start: # else bl _GLOBAL_OFFSET_TABLE_-4@local mflr r31 +# endif
- /* in PIC/PIE, plt stubs need r30 to point to the GOT if using secure-plt */
+# ifdef PPC_HAS_SECUREPLT
- mr 30,31
We need this instruction only for PPC_HAS_SECUREPLT and PIC/PIE enabled by default. But this is enabled as soon as PPC_HAS_SECUREPLT is enabled, so the register copy is useless if PIC/PIE is not used.
As is, this patch fix the issue!
Tested-by: Romain Naour romain.naour@gmail.com
Best regards, Romain
# endif #endif /* Set up the small data pointer in r13. */
-
Romain Naour -
Yann Sionneau