-
Peter Zijlstra authored
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. Signed-off-by:
Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by:
Ingo Molnar <mingo@kernel.org> Acked-by:
Josh Poimboeuf <jpoimboe@kernel.org> Tested-by: Nathan Chancellor <nathan@kernel.org> # build only Tested-by: Thomas Weißschuh <linux@weissschuh.net> # compile and run Link: https://lore.kernel.org/r/20230208172245.783099843@infradead.org
6ea17e84