On Mon, May 06, 2024 at 02:53:00PM -0700, Alexey Makhalov wrote:
> +#define VMWARE_HYPERCALL \
> + ALTERNATIVE_3("cmpb $" \
> + __stringify(CPUID_VMWARE_FEATURES_ECX_VMMCALL) \
> + ", %[mode]\n\t" \
> + "jg 2f\n\t" \
> + "je 1f\n\t" \
> + "movw %[port], %%dx\n\t" \
> + "inl (%%dx), %%eax\n\t" \
> + "jmp 3f\n\t" \
> + "1: vmmcall\n\t" \
> + "jmp 3f\n\t" \
> + "2: vmcall\n\t" \
> + "3:\n\t", \
> + "movw %[port], %%dx\n\t" \
> + "inl (%%dx), %%eax", X86_FEATURE_HYPERVISOR, \
That's a bunch of insns and their size would inadvertently go into the final
image.
What you should try to do is something like this:
ALTERNATIVE_3("jmp .Lend_legacy_call", "", X86_FEATURE_HYPERVISOR,
"vmcall; jmp .Lend_legacy_call", X86_FEATURE_VMCALL,
"vmmcall; jmp .Lend_legacy_call", X86_FEATURE_VMW_VMMCALL)
/* bunch of conditional branches and INs and V*MCALLs, etc go here */
.Lend_legacy_call:
so that you don't have these 26 bytes, as you say, of alternatives to patch but
only the JMPs and the VM*CALLs.
See for an example the macros in arch/x86/entry/calling.h which simply jump
over the code when not needed.
Also, you could restructure the alternative differently so that that bunch of
insns call is completely out-of-line because all current machines support
VM*CALL so you won't even need to patch. You only get to patch when running on
some old rust and there you can just as well go completely out-of-line.
Something along those lines, anyway.
> - * The high bandwidth in call. The low word of edx is presumed to have the
> - * HB bit set.
> + * High bandwidth calls are not supported on encrypted memory guests.
> + * The caller should check cc_platform_has(CC_ATTR_MEM_ENCRYPT) and use
> + * low bandwidth hypercall it memory encryption is set.
s/it/if/
> -#define VMWARE_PORT(cmd, eax, ebx, ecx, edx) \
> - __asm__("inl (%%dx), %%eax" : \
> - "=a"(eax), "=c"(ecx), "=d"(edx), "=b"(ebx) : \
> - "a"(VMWARE_HYPERVISOR_MAGIC), \
> - "c"(VMWARE_CMD_##cmd), \
> - "d"(VMWARE_HYPERVISOR_PORT), "b"(UINT_MAX) : \
> - "memory")
> -
> -#define VMWARE_VMCALL(cmd, eax, ebx, ecx, edx) \
> - __asm__("vmcall" : \
> - "=a"(eax), "=c"(ecx), "=d"(edx), "=b"(ebx) : \
> - "a"(VMWARE_HYPERVISOR_MAGIC), \
> - "c"(VMWARE_CMD_##cmd), \
> - "d"(0), "b"(UINT_MAX) : \
> - "memory")
> -
> -#define VMWARE_VMMCALL(cmd, eax, ebx, ecx, edx) \
> - __asm__("vmmcall" : \
> - "=a"(eax), "=c"(ecx), "=d"(edx), "=b"(ebx) : \
> - "a"(VMWARE_HYPERVISOR_MAGIC), \
> - "c"(VMWARE_CMD_##cmd), \
> - "d"(0), "b"(UINT_MAX) : \
> - "memory")
> -
> -#define VMWARE_CMD(cmd, eax, ebx, ecx, edx) do { \
> - switch (vmware_hypercall_mode) { \
> - case CPUID_VMWARE_FEATURES_ECX_VMCALL: \
> - VMWARE_VMCALL(cmd, eax, ebx, ecx, edx); \
> - break; \
> - case CPUID_VMWARE_FEATURES_ECX_VMMCALL: \
> - VMWARE_VMMCALL(cmd, eax, ebx, ecx, edx); \
> - break; \
> - default: \
> - VMWARE_PORT(cmd, eax, ebx, ecx, edx); \
> - break; \
> - } \
> - } while (0)
You're kidding, right?
You went to all that trouble in patch 1 to move those to the header only to
*remove* them here?
You do realize that that is a unnecessary churn for no good reason, right?
So that set needs to be restructured differently.
* first patch introduces those new API calls.
* follow-on patches convert the callers to the new API
* last patch removes the old API.
Ok?
And when you redo them, make sure you drop all Reviewed-by tags because the new
versions are not reviewed anymore.
Thx.
--
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquette
