Re: [PATCH v2 7/9] x86/vdso: abstract out vdso system call internals

[H. Peter Anvin]

On November 12, 2025 11:15:08 PM PST, Uros Bizjak <ubizjak@xxxxxxxxx> wrote:
>On Wed, Nov 12, 2025 at 10:25 PM H. Peter Anvin <hpa@xxxxxxxxx> wrote:
>>
>> On 2025-11-12 02:31, Uros Bizjak wrote:
>> >
>> > Unfortunately, %ebp is still special with -fno-omit-frame-pointer, so
>> > using "ebp" as _sys_arg6 on 32-bit targets will result in:
>> >
>> > error: bp cannot be used in ‘asm’ here
>> >
>> > Please see how %ebp register is handled in
>> > arch/x86/include/asm/vmware.h, vmware_hypercall_hb_out() and
>> > vmware_hypercall_hb_in().
>> >
>>
>> #ifdef CONFIG_X86_64
>> #define VMW_BP_CONSTRAINT "r"
>> #else
>> #define VMW_BP_CONSTRAINT "m"
>> #endif
>>
>>         asm_inline volatile (
>>                 UNWIND_HINT_SAVE
>>                 "push %%" _ASM_BP "\n\t"
>>                 UNWIND_HINT_UNDEFINED
>>                 "mov %[in6], %%" _ASM_BP "\n\t"
>>                 "rep outsb\n\t"
>>                 "pop %%" _ASM_BP "\n\t"
>>                 UNWIND_HINT_RESTORE
>>                 : "=a" (out0), "=b" (*out1)
>>                 : "a" (VMWARE_HYPERVISOR_MAGIC),
>>                   "b" (cmd),
>>                   "c" (in2),
>>                   "d" (in3 | VMWARE_HYPERVISOR_PORT_HB),
>>                   "S" (in4),
>>                   "D" (in5),
>>                   [in6] VMW_BP_CONSTRAINT (in6)
>>                 : "cc", "memory");
>>         return out0;
>>
>> That code is actually incorrect, in at least two ways:
>>
>>
>> 1. It should be conditioned on frame pointers enabled, not x86-64 vs i386.
>> 2. The compiler is perfectly within its right to emit an %esp-relative
>>    reference for the "m"-constrained [in6]. This is particularly likely
>>    when *not* compiled with frame pointers, see #1.
>>
>> A better sequence might be:
>>
>>         pushl %[in6]
>>         push %ebp
>>         mov 4(%esp),%ebp
>>         <stuff>
>>         pop %ebp
>>         pop %[junk]
>>
>> Then %[in6] can even safely be a "g" constraint (hence pushl).
>
>If we want to also handle x86_64, the above code (including push)
>needs to be 64-bit, with "rme" constraint for the pushed value.
>
>I have CC'd the author of the above code, he might be interested in
>the above discussion.
>
>Uros.
>

Note: for the specific case of VDSO_SYSCALL6 the %eax value is always a constant, so:

push %%ebp
mov %%eax,%%ebp
mov %[sysnr],%%eax /* "i" constraint */
<syscall sequence>
pop %%ebp

... would be easiest.