| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <20230208172245.783099843@infradead.org>
Date: Wed, 08 Feb 2023 18:18:04 +0100
From: Peter Zijlstra <peterz@...radead.org>
To: x86@...nel.org, jpoimboe@...hat.com, linux@...ssschuh.net
Cc: linux-kernel@...r.kernel.org, peterz@...radead.org
Subject: [PATCH 08/10] x86: Fix FILL_RETURN_BUFFER
With overlapping alternative validation fixed, objtool promptly
complains:
vmlinux.o: warning: objtool: __switch_to_asm+0x2c: stack layout conflict in alternatives: .altinstr_replacement+0x47
.rela.altinstructions:
000000000000009c 0000000200000002 R_X86_64_PC32 0000000000000000 .text + 16dc
00000000000000a0 0000000600000002 R_X86_64_PC32 0000000000000000 .altinstr_replacement + 3a
00000000000000a8 0000000200000002 R_X86_64_PC32 0000000000000000 .text + 16dc
00000000000000ac 0000000600000002 R_X86_64_PC32 0000000000000000 .altinstr_replacement + 66
.text:
00000000000016b0 <__switch_to_asm>:
16b0: f3 0f 1e fa endbr64
16b4: 55 push %rbp
16b5: 53 push %rbx
16b6: 41 54 push %r12
16b8: 41 55 push %r13
16ba: 41 56 push %r14
16bc: 41 57 push %r15
16be: 48 89 a7 18 0b 00 00 mov %rsp,0xb18(%rdi)
16c5: 48 8b a6 18 0b 00 00 mov 0xb18(%rsi),%rsp
16cc: 48 8b 9e 28 05 00 00 mov 0x528(%rsi),%rbx
16d3: 65 48 89 1c 25 00 00 00 00 mov %rbx,%gs:0x0 16d8: R_X86_64_32S fixed_percpu_data+0x28
16dc: eb 2a jmp 1708 <__switch_to_asm+0x58>
16de: 90 nop
16df: 90 nop
16e0: 90 nop
16e1: 90 nop
16e2: 90 nop
16e3: 90 nop
16e4: 90 nop
16e5: 90 nop
16e6: 90 nop
16e7: 90 nop
16e8: 90 nop
16e9: 90 nop
16ea: 90 nop
16eb: 90 nop
16ec: 90 nop
16ed: 90 nop
16ee: 90 nop
16ef: 90 nop
16f0: 90 nop
16f1: 90 nop
16f2: 90 nop
16f3: 90 nop
16f4: 90 nop
16f5: 90 nop
16f6: 90 nop
16f7: 90 nop
16f8: 90 nop
16f9: 90 nop
16fa: 90 nop
16fb: 90 nop
16fc: 90 nop
16fd: 90 nop
16fe: 90 nop
16ff: 90 nop
1700: 90 nop
1701: 90 nop
1702: 90 nop
1703: 90 nop
1704: 90 nop
1705: 90 nop
1706: 90 nop
1707: 90 nop
1708: 41 5f pop %r15
170a: 41 5e pop %r14
170c: 41 5d pop %r13
170e: 41 5c pop %r12
1710: 5b pop %rbx
1711: 5d pop %rbp
1712: e9 00 00 00 00 jmp 1717 <__switch_to_asm+0x67> 1713: R_X86_64_PLT32 __switch_to-0x4
.altinstr_replacement:
3a: 49 c7 c4 10 00 00 00 mov $0x10,%r12
41: e8 01 00 00 00 call 47 <.altinstr_replacement+0x47>
46: cc int3
47: e8 01 00 00 00 call 4d <.altinstr_replacement+0x4d>
4c: cc int3
4d: 48 83 c4 10 add $0x10,%rsp
51: 49 ff cc dec %r12
54: 75 eb jne 41 <.altinstr_replacement+0x41>
56: 0f ae e8 lfence
59: 65 48 c7 04 25 00 00 00 00 ff ff ff ff movq $0xffffffffffffffff,%gs:0x0 5e: R_X86_64_32S pcpu_hot+0x10
66: e8 01 00 00 00 call 6c <.altinstr_replacement+0x6c>
6b: cc int3
6c: 48 83 c4 08 add $0x8,%rsp
70: 0f ae e8 lfence
As can be seen from the two alternatives, when overlaid, the NOP after
the shorter (starting at 66) coinsides with the call at 47, leading to
conflicting CFI state for that instruction.
By offsetting the shorter alternative by 2 bytes, this alignment is
undone.
TODO: arguably objtool should be taught about the max nop length used
by the kernel for tail padding, or unconditionally use JMP.d8 to
not use the intervening bytes at all.
Signed-off-by: Peter Zijlstra (Intel) <peterz@...radead.org>
---
arch/x86/include/asm/nospec-branch.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
--- a/arch/x86/include/asm/nospec-branch.h
+++ b/arch/x86/include/asm/nospec-branch.h
@@ -261,7 +261,7 @@
.macro FILL_RETURN_BUFFER reg:req nr:req ftr:req ftr2=ALT_NOT(X86_FEATURE_ALWAYS)
ALTERNATIVE_2 "jmp .Lskip_rsb_\@", \
__stringify(__FILL_RETURN_BUFFER(\reg,\nr)), \ftr, \
- __stringify(__FILL_ONE_RETURN), \ftr2
+ __stringify(nop;nop;__FILL_ONE_RETURN), \ftr2
.Lskip_rsb_\@:
.endm
Powered by blists - more mailing lists