* [PATCH 2/3] powerpc/vdso: Don't pass 64-bit ABI cflags to 32-bit VDSO
2020-12-18 11:16 [PATCH 1/3] powerpc/vdso: Block R_PPC_REL24 relocations Michael Ellerman
@ 2020-12-18 11:16 ` Michael Ellerman
2020-12-18 11:16 ` [PATCH 3/3] powerpc/vdso: Fix DOTSYM for 32-bit LE VDSO Michael Ellerman
2020-12-22 13:11 ` [PATCH 1/3] powerpc/vdso: Block R_PPC_REL24 relocations Michael Ellerman
2 siblings, 0 replies; 4+ messages in thread
From: Michael Ellerman @ 2020-12-18 11:16 UTC (permalink / raw)
To: linuxppc-dev; +Cc: skirmisher
When building the 32-bit VDSO, we are building 32-bit code as part of
a 64-bit kernel build. That requires us to tweak the cflags to trick
the compiler into building 32-bit code for us. The main way we do that
is by passing -m32, but there are other options that affect code
generation and ABI selection.
In particular when building vgettimeofday.c, we end up passing
-mcall-aixdesc because it's in KBUILD_CFLAGS, which causes the
compiler to generate function descriptors, and dot symbols, eg:
$ nm arch/powerpc/kernel/vdso32/vgettimeofday.o
000005d0 T .__c_kernel_clock_getres
00000024 D __c_kernel_clock_getres
...
We get away with that at the moment because we also use the DOTSYM
macro, and that is also incorrectly prepending a '.' in 32-bit VDSO
code due to a separate bug.
But we shouldn't be generating function descriptors for this file,
there's no 32-bit ABI that includes function descriptors, so the
resulting object file is some frankenstein and it's surprising that it
even links.
So filter out all the ABI-related options we add to CFLAGS for 64-bit
builds, so that they're not used when building 32-bit code. With that
we only see regular text symbols:
$ nm arch/powerpc/kernel/vdso32/vgettimeofday.o michael@alpine1-p1
000005d0 T __c_kernel_clock_getres
00000000 T __c_kernel_clock_gettime
00000200 T __c_kernel_clock_gettime64
00000410 T __c_kernel_gettimeofday
00000650 T __c_kernel_time
Fixes: ab037dd87a2f ("powerpc/vdso: Switch VDSO to generic C implementation.")
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
---
arch/powerpc/kernel/vdso32/Makefile | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/arch/powerpc/kernel/vdso32/Makefile b/arch/powerpc/kernel/vdso32/Makefile
index 6616f4e794d0..9cb6f524854b 100644
--- a/arch/powerpc/kernel/vdso32/Makefile
+++ b/arch/powerpc/kernel/vdso32/Makefile
@@ -27,7 +27,7 @@ endif
CC32FLAGS :=
ifdef CONFIG_PPC64
CC32FLAGS += -m32
-KBUILD_CFLAGS := $(filter-out -mcmodel=medium,$(KBUILD_CFLAGS))
+KBUILD_CFLAGS := $(filter-out -mcmodel=medium -mabi=elfv1 -mabi=elfv2 -mcall-aixdesc,$(KBUILD_CFLAGS))
endif
targets := $(obj-vdso32) vdso32.so.dbg
--
2.25.1
^ permalink raw reply related [flat|nested] 4+ messages in thread* [PATCH 3/3] powerpc/vdso: Fix DOTSYM for 32-bit LE VDSO
2020-12-18 11:16 [PATCH 1/3] powerpc/vdso: Block R_PPC_REL24 relocations Michael Ellerman
2020-12-18 11:16 ` [PATCH 2/3] powerpc/vdso: Don't pass 64-bit ABI cflags to 32-bit VDSO Michael Ellerman
@ 2020-12-18 11:16 ` Michael Ellerman
2020-12-22 13:11 ` [PATCH 1/3] powerpc/vdso: Block R_PPC_REL24 relocations Michael Ellerman
2 siblings, 0 replies; 4+ messages in thread
From: Michael Ellerman @ 2020-12-18 11:16 UTC (permalink / raw)
To: linuxppc-dev; +Cc: skirmisher
Skirmisher reported on IRC that the 32-bit LE VDSO was hanging. This
turned out to be due to a branch to self in eg. __kernel_gettimeofday.
Looking at the disassembly with objdump -dR shows why:
00000528 <__kernel_gettimeofday>:
528: f0 ff 21 94 stwu r1,-16(r1)
52c: a6 02 08 7c mflr r0
530: f0 ff 21 94 stwu r1,-16(r1)
534: 14 00 01 90 stw r0,20(r1)
538: 05 00 9f 42 bcl 20,4*cr7+so,53c <__kernel_gettimeofday+0x14>
53c: a6 02 a8 7c mflr r5
540: ff ff a5 3c addis r5,r5,-1
544: c4 fa a5 38 addi r5,r5,-1340
548: f0 00 a5 38 addi r5,r5,240
54c: 01 00 00 48 bl 54c <__kernel_gettimeofday+0x24>
54c: R_PPC_REL24 .__c_kernel_gettimeofday
Because we don't process relocations for the VDSO, this branch remains
a branch from 0x54c to 0x54c.
With the preceding patch to prohibit R_PPC_REL24 relocations, we
instead get a build failure:
0000054c R_PPC_REL24 .__c_kernel_gettimeofday
00000598 R_PPC_REL24 .__c_kernel_clock_gettime
000005e4 R_PPC_REL24 .__c_kernel_clock_gettime64
00000630 R_PPC_REL24 .__c_kernel_clock_getres
0000067c R_PPC_REL24 .__c_kernel_time
arch/powerpc/kernel/vdso32/vdso32.so.dbg: dynamic relocations are not supported
The root cause is that we're branching to `.__c_kernel_gettimeofday`.
But this is 32-bit LE code, which doesn't use function descriptors, so
there are no dot symbols.
The reason we're trying to branch to a dot symbol is because we're
using the DOTSYM macro, but the ifdefs we use to define the DOTSYM
macro do not currently work for 32-bit LE.
So like previous commits we need to differentiate if the current
compilation unit is 64-bit, rather than the kernel as a whole. ie.
switch from CONFIG_PPC64 to __powerpc64__.
With that fixed 32-bit LE code gets the empty version of DOTSYM, which
just resolves to the original symbol name, leading to a direct branch
and no relocations:
000003f8 <__kernel_gettimeofday>:
3f8: f0 ff 21 94 stwu r1,-16(r1)
3fc: a6 02 08 7c mflr r0
400: f0 ff 21 94 stwu r1,-16(r1)
404: 14 00 01 90 stw r0,20(r1)
408: 05 00 9f 42 bcl 20,4*cr7+so,40c <__kernel_gettimeofday+0x14>
40c: a6 02 a8 7c mflr r5
410: ff ff a5 3c addis r5,r5,-1
414: f4 fb a5 38 addi r5,r5,-1036
418: f0 00 a5 38 addi r5,r5,240
41c: 85 06 00 48 bl aa0 <__c_kernel_gettimeofday>
Fixes: ab037dd87a2f ("powerpc/vdso: Switch VDSO to generic C implementation.")
Reported-by: "Will Springer <skirmisher@protonmail.com>"
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
---
arch/powerpc/include/asm/ppc_asm.h | 7 ++++++-
1 file changed, 6 insertions(+), 1 deletion(-)
diff --git a/arch/powerpc/include/asm/ppc_asm.h b/arch/powerpc/include/asm/ppc_asm.h
index cfa814824285..cc1bca571332 100644
--- a/arch/powerpc/include/asm/ppc_asm.h
+++ b/arch/powerpc/include/asm/ppc_asm.h
@@ -180,7 +180,12 @@ END_FW_FTR_SECTION_IFSET(FW_FEATURE_SPLPAR)
#define VCPU_GPR(n) __VCPU_GPR(__REG_##n)
#ifdef __KERNEL__
-#ifdef CONFIG_PPC64
+
+/*
+ * We use __powerpc64__ here because we want the compat VDSO to use the 32-bit
+ * version below in the else case of the ifdef.
+ */
+#ifdef __powerpc64__
#define STACKFRAMESIZE 256
#define __STK_REG(i) (112 + ((i)-14)*8)
--
2.25.1
^ permalink raw reply related [flat|nested] 4+ messages in thread