On Thu, Apr 22, 2021, Paolo Bonzini wrote:
> On 22/04/21 05:04, Sean Christopherson wrote:
> > Load the per-cpu GS.base for 32-bit build by building a temporary GDT
> > and loading a "real" segment. Using MSR_GS_BASE is wrong and broken,
> > it's a 64-bit only MSR and does not exist on 32-bit CPUs. The current
> > code works only because 32-bit KVM VMX incorrectly disables interception
> > of MSR_GS_BASE, and no one runs KVM on an actual 32-bit physical CPU,
> > i.e. the MSR exists in hardware and so everything "works".
> >
> > 32-bit KVM SVM is not buggy and correctly injects #GP on the WRMSR, i.e.
> > the tests have never worked on 32-bit SVM.
>
> Hmm, this breaks task switch. But setting up separate descriptors is
> not hard:
Much better.
> diff --git a/x86/cstart.S b/x86/cstart.S
> index 489c561..7d9ed96 100644
> --- a/x86/cstart.S
> +++ b/x86/cstart.S
> @@ -58,6 +58,10 @@ tss_descr:
> .rept max_cpus
> .quad 0x000089000000ffff // 32-bit avail tss
> .endr
> +percpu_descr:
> + .rept max_cpus
> + .quad 0x00cf93000000ffff // 32-bit data segment for perCPU area
> + .endr
> gdt32_end:
>
> i = 0
> @@ -89,13 +93,23 @@ mb_flags = 0x0
> .long mb_magic, mb_flags, 0 - (mb_magic + mb_flags)
> mb_cmdline = 16
>
> -MSR_GS_BASE = 0xc0000101
> -
> .macro setup_percpu_area
> lea -4096(%esp), %eax
> - mov $0, %edx
> - mov $MSR_GS_BASE, %ecx
> - wrmsr
> +
> + /* fill GS_BASE in the GDT */
> + mov $(APIC_DEFAULT_PHYS_BASE + APIC_ID), %ebx
Using %ebx crushes the mbi_bootinfo pointer. The easiest fix is to use %edx or
%ecx.
> + mov (%ebx), %ebx
No need to load the address into a reg, just drop the "$" above and encode
"mov [imm32], <reg>".
Want to fold this into your patch?
diff --git a/x86/cstart.S b/x86/cstart.S
index 7d9ed96..fb6eda5 100644
--- a/x86/cstart.S
+++ b/x86/cstart.S
@@ -97,17 +97,16 @@ mb_cmdline = 16
lea -4096(%esp), %eax
/* fill GS_BASE in the GDT */
- mov $(APIC_DEFAULT_PHYS_BASE + APIC_ID), %ebx
- mov (%ebx), %ebx
- shr $24, %ebx
- or %ax, percpu_descr+2(,%ebx,8)
+ mov (APIC_DEFAULT_PHYS_BASE + APIC_ID), %edx
+ shr $24, %edx
+ or %ax, percpu_descr+2(,%edx,8)
shr $16, %eax
- or %al, percpu_descr+4(,%ebx,8)
- or %ah, percpu_descr+7(,%ebx,8)
+ or %al, percpu_descr+4(,%edx,8)
+ or %ah, percpu_descr+7(,%edx,8)
lgdtl gdt32_descr
- lea percpu_descr-gdt32(,%ebx,8), %eax
+ lea percpu_descr-gdt32(,%edx,8), %eax
mov %ax, %gs
.endm
> + shr $24, %ebx
> + or %ax, percpu_descr+2(,%ebx,8)
> +
> + shr $16, %eax
> + or %al, percpu_descr+4(,%ebx,8)
> + or %ah, percpu_descr+7(,%ebx,8)
> +
> + lgdtl gdt32_descr
> + lea percpu_descr-gdt32(,%ebx,8), %eax
> + mov %ax, %gs
> +
> .endm
>
> .macro setup_segments
> @@ -188,16 +202,14 @@ load_tss:
> mov (%eax), %eax
> shr $24, %eax
> mov %eax, %ebx
> - shl $3, %ebx
> mov $((tss_end - tss) / max_cpus), %edx
> imul %edx
> add $tss, %eax
> - mov %ax, tss_descr+2(%ebx)
> + mov %ax, tss_descr+2(,%ebx,8)
> shr $16, %eax
> - mov %al, tss_descr+4(%ebx)
> - shr $8, %eax
> - mov %al, tss_descr+7(%ebx)
> - lea tss_descr-gdt32(%ebx), %eax
> + mov %al, tss_descr+4(,%ebx,8)
> + mov %ah, tss_descr+7(,%ebx,8)
Is there a functional change here? If not, can you throw this into a separate
patch?
Thanks!
> + lea tss_descr-gdt32(,%ebx,8), %eax
> ltr %ax
> ret
>
>
> Paolo
>
