From: Joerg Roedel <[email protected]>
GHCB protocol version 2 adds the MSR-based AP-reset-hold VMGEXIT which
does not need a GHCB. Use that to park APs in 16-bit protected mode on
the AP Jump Table.
Signed-off-by: Joerg Roedel <[email protected]>
---
arch/x86/include/asm/realmode.h | 3 +
arch/x86/kernel/sev.c | 48 ++++++++++++++--
arch/x86/realmode/rm/Makefile | 11 ++--
arch/x86/realmode/rm/header.S | 3 +
arch/x86/realmode/rm/sev_ap_park.S | 89 ++++++++++++++++++++++++++++++
5 files changed, 144 insertions(+), 10 deletions(-)
create mode 100644 arch/x86/realmode/rm/sev_ap_park.S
diff --git a/arch/x86/include/asm/realmode.h b/arch/x86/include/asm/realmode.h
index 29590a4ddf24..668de0a8b1ae 100644
--- a/arch/x86/include/asm/realmode.h
+++ b/arch/x86/include/asm/realmode.h
@@ -23,6 +23,9 @@ struct real_mode_header {
u32 trampoline_header;
#ifdef CONFIG_AMD_MEM_ENCRYPT
u32 sev_es_trampoline_start;
+ u32 sev_real_ap_park_asm;
+ u32 sev_real_ap_park_seg;
+ u32 sev_ap_park_gdt;
#endif
#ifdef CONFIG_X86_64
u32 trampoline_pgd;
diff --git a/arch/x86/kernel/sev.c b/arch/x86/kernel/sev.c
index a98eab926682..20b439986d86 100644
--- a/arch/x86/kernel/sev.c
+++ b/arch/x86/kernel/sev.c
@@ -27,6 +27,7 @@
#include <asm/fpu/internal.h>
#include <asm/processor.h>
#include <asm/realmode.h>
+#include <asm/tlbflush.h>
#include <asm/traps.h>
#include <asm/svm.h>
#include <asm/smp.h>
@@ -695,6 +696,35 @@ static bool __init sev_es_setup_ghcb(void)
}
#ifdef CONFIG_HOTPLUG_CPU
+void __noreturn sev_jumptable_ap_park(void)
+{
+ local_irq_disable();
+
+ write_cr3(real_mode_header->trampoline_pgd);
+
+ /* Exiting long mode will fail if CR4.PCIDE is set. */
+ if (boot_cpu_has(X86_FEATURE_PCID))
+ cr4_clear_bits(X86_CR4_PCIDE);
+
+ asm volatile("xorq %%r15, %%r15\n"
+ "xorq %%r14, %%r14\n"
+ "xorq %%r13, %%r13\n"
+ "xorq %%r12, %%r12\n"
+ "xorq %%r11, %%r11\n"
+ "xorq %%r10, %%r10\n"
+ "xorq %%r9, %%r9\n"
+ "xorq %%r8, %%r8\n"
+ "xorq %%rsi, %%rsi\n"
+ "xorq %%rdi, %%rdi\n"
+ "xorq %%rsp, %%rsp\n"
+ "xorq %%rbp, %%rbp\n"
+ "ljmpl *%0" : :
+ "m" (real_mode_header->sev_real_ap_park_asm),
+ "b" (sev_es_jump_table_pa >> 4));
+ unreachable();
+}
+STACK_FRAME_NON_STANDARD(sev_jumptable_ap_park);
+
static void sev_es_ap_hlt_loop(void)
{
struct ghcb_state state;
@@ -731,8 +761,10 @@ static void sev_es_play_dead(void)
play_dead_common();
/* IRQs now disabled */
-
- sev_es_ap_hlt_loop();
+ if (sev_ap_jumptable_blob_installed)
+ sev_jumptable_ap_park();
+ else
+ sev_es_ap_hlt_loop();
/*
* If we get here, the VCPU was woken up again. Jump to CPU
@@ -762,8 +794,9 @@ static inline void sev_es_setup_play_dead(void) { }
void __init sev_es_setup_ap_jump_table_data(void *base, u32 pa)
{
struct sev_ap_jump_table_header *header;
+ u64 *ap_jumptable_gdt, *sev_ap_park_gdt;
struct desc_ptr *gdt_descr;
- u64 *ap_jumptable_gdt;
+ int idx;
header = base;
@@ -773,8 +806,13 @@ void __init sev_es_setup_ap_jump_table_data(void *base, u32 pa)
* real-mode.
*/
ap_jumptable_gdt = (u64 *)(base + header->gdt_offset);
- ap_jumptable_gdt[SEV_APJT_CS16 / 8] = GDT_ENTRY(0x9b, pa, 0xffff);
- ap_jumptable_gdt[SEV_APJT_DS16 / 8] = GDT_ENTRY(0x93, pa, 0xffff);
+ sev_ap_park_gdt = __va(real_mode_header->sev_ap_park_gdt);
+
+ idx = SEV_APJT_CS16 / 8;
+ ap_jumptable_gdt[idx] = sev_ap_park_gdt[idx] = GDT_ENTRY(0x9b, pa, 0xffff);
+
+ idx = SEV_APJT_DS16 / 8;
+ ap_jumptable_gdt[idx] = sev_ap_park_gdt[idx] = GDT_ENTRY(0x93, pa, 0xffff);
/* Write correct GDT base address into GDT descriptor */
gdt_descr = (struct desc_ptr *)(base + header->gdt_offset);
diff --git a/arch/x86/realmode/rm/Makefile b/arch/x86/realmode/rm/Makefile
index 83f1b6a56449..a7f84d43a0a3 100644
--- a/arch/x86/realmode/rm/Makefile
+++ b/arch/x86/realmode/rm/Makefile
@@ -27,11 +27,12 @@ wakeup-objs += video-vga.o
wakeup-objs += video-vesa.o
wakeup-objs += video-bios.o
-realmode-y += header.o
-realmode-y += trampoline_$(BITS).o
-realmode-y += stack.o
-realmode-y += reboot.o
-realmode-$(CONFIG_ACPI_SLEEP) += $(wakeup-objs)
+realmode-y += header.o
+realmode-y += trampoline_$(BITS).o
+realmode-y += stack.o
+realmode-y += reboot.o
+realmode-$(CONFIG_ACPI_SLEEP) += $(wakeup-objs)
+realmode-$(CONFIG_AMD_MEM_ENCRYPT) += sev_ap_park.o
targets += $(realmode-y)
diff --git a/arch/x86/realmode/rm/header.S b/arch/x86/realmode/rm/header.S
index 8c1db5bf5d78..6c17f8fd1eb4 100644
--- a/arch/x86/realmode/rm/header.S
+++ b/arch/x86/realmode/rm/header.S
@@ -22,6 +22,9 @@ SYM_DATA_START(real_mode_header)
.long pa_trampoline_header
#ifdef CONFIG_AMD_MEM_ENCRYPT
.long pa_sev_es_trampoline_start
+ .long pa_sev_ap_park_asm
+ .long __KERNEL32_CS
+ .long pa_sev_ap_park_gdt;
#endif
#ifdef CONFIG_X86_64
.long pa_trampoline_pgd;
diff --git a/arch/x86/realmode/rm/sev_ap_park.S b/arch/x86/realmode/rm/sev_ap_park.S
new file mode 100644
index 000000000000..0b63d0569d4d
--- /dev/null
+++ b/arch/x86/realmode/rm/sev_ap_park.S
@@ -0,0 +1,89 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#include <linux/linkage.h>
+#include <asm/segment.h>
+#include <asm/page_types.h>
+#include <asm/processor-flags.h>
+#include <asm/msr-index.h>
+#include <asm/sev-ap-jumptable.h>
+#include "realmode.h"
+
+ .section ".text32", "ax"
+ .code32
+/*
+ * The following code switches to 16-bit protected mode and sets up the
+ * execution environment for the AP Jump Table blob. Then it jumps to the AP
+ * Jump Table to park the AP.
+ *
+ * The code was copied from reboot.S and modified to fit the SEV-ES requirements
+ * for AP parking. When this code is entered, all registers except %EAX-%EDX are
+ * in reset state.
+ *
+ * The AP Jump Table physical base address is in %EBX upon entry.
+ *
+ * %EAX, %ECX, %EDX and EFLAGS are undefined. Only use registers %EAX-%EDX and
+ * %ESP in this code.
+ */
+SYM_CODE_START(sev_ap_park_asm)
+
+ /* Switch to trampoline GDT as it is guaranteed < 4 GiB */
+ movl $__KERNEL_DS, %eax
+ movl %eax, %ds
+ lgdt pa_tr_gdt
+
+ /* Disable paging to drop us out of long mode */
+ movl %cr0, %eax
+ btcl $X86_CR0_PG_BIT, %eax
+ movl %eax, %cr0
+
+ ljmpl $__KERNEL32_CS, $pa_sev_ap_park_paging_off
+
+SYM_INNER_LABEL(sev_ap_park_paging_off, SYM_L_GLOBAL)
+ /* Clear EFER */
+ movl $0, %eax
+ movl $0, %edx
+ movl $MSR_EFER, %ecx
+ wrmsr
+
+ /* Clear CR3 */
+ movl $0, %ecx
+ movl %ecx, %cr3
+
+ /* Set up the IDT for real mode. */
+ lidtl pa_machine_real_restart_idt
+
+ /*
+ * Load the GDT with the 16-bit segments for the AP Jump Table
+ */
+ lgdtl pa_sev_ap_park_gdt
+
+ /* Setup Code and Data segments for AP Jump Table */
+ movw $SEV_APJT_DS16, %ax
+ movw %ax, %ds
+ movw %ax, %ss
+
+ /* Jump to the AP Jump Table into 16 bit protected mode */
+ ljmpw $SEV_APJT_CS16, $SEV_APJT_ENTRY
+SYM_CODE_END(sev_ap_park_asm)
+
+ .data
+ .balign 16
+SYM_DATA_START(sev_ap_park_gdt)
+ /* Self-pointer */
+ .word sev_ap_park_gdt_end - sev_ap_park_gdt - 1
+ .long pa_sev_ap_park_gdt
+ .word 0
+
+ /*
+ * Offset 0x8
+ * 32 bit code segment descriptor pointing to AP Jump table base
+ * Setup at runtime in sev_es_setup_ap_jump_table_data().
+ */
+ .quad 0
+
+ /*
+ * Offset 0x10
+ * 32 bit data segment descriptor pointing to AP Jump table base
+ * Setup at runtime in sev_es_setup_ap_jump_table_data().
+ */
+ .quad 0
+SYM_DATA_END_LABEL(sev_ap_park_gdt, SYM_L_GLOBAL, sev_ap_park_gdt_end)
--
2.33.0
On Mon, Sep 13, 2021 at 05:55:59PM +0200, Joerg Roedel wrote:
> From: Joerg Roedel <[email protected]>
>
> GHCB protocol version 2 adds the MSR-based AP-reset-hold VMGEXIT which
> does not need a GHCB. Use that to park APs in 16-bit protected mode on
> the AP Jump Table.
>
> Signed-off-by: Joerg Roedel <[email protected]>
> ---
> arch/x86/include/asm/realmode.h | 3 +
> arch/x86/kernel/sev.c | 48 ++++++++++++++--
> arch/x86/realmode/rm/Makefile | 11 ++--
> arch/x86/realmode/rm/header.S | 3 +
> arch/x86/realmode/rm/sev_ap_park.S | 89 ++++++++++++++++++++++++++++++
> 5 files changed, 144 insertions(+), 10 deletions(-)
> create mode 100644 arch/x86/realmode/rm/sev_ap_park.S
>
> diff --git a/arch/x86/include/asm/realmode.h b/arch/x86/include/asm/realmode.h
> index 29590a4ddf24..668de0a8b1ae 100644
> --- a/arch/x86/include/asm/realmode.h
> +++ b/arch/x86/include/asm/realmode.h
> @@ -23,6 +23,9 @@ struct real_mode_header {
> u32 trampoline_header;
> #ifdef CONFIG_AMD_MEM_ENCRYPT
> u32 sev_es_trampoline_start;
> + u32 sev_real_ap_park_asm;
sev_ap_park;
> + u32 sev_real_ap_park_seg;
sev_ap_park_seg;
> + u32 sev_ap_park_gdt;
Yap, like thist one.
> #endif
> #ifdef CONFIG_X86_64
> u32 trampoline_pgd;
> diff --git a/arch/x86/kernel/sev.c b/arch/x86/kernel/sev.c
> index a98eab926682..20b439986d86 100644
> --- a/arch/x86/kernel/sev.c
> +++ b/arch/x86/kernel/sev.c
> @@ -27,6 +27,7 @@
> #include <asm/fpu/internal.h>
> #include <asm/processor.h>
> #include <asm/realmode.h>
> +#include <asm/tlbflush.h>
> #include <asm/traps.h>
> #include <asm/svm.h>
> #include <asm/smp.h>
> @@ -695,6 +696,35 @@ static bool __init sev_es_setup_ghcb(void)
> }
>
> #ifdef CONFIG_HOTPLUG_CPU
> +void __noreturn sev_jumptable_ap_park(void)
> +{
> + local_irq_disable();
> +
> + write_cr3(real_mode_header->trampoline_pgd);
> +
> + /* Exiting long mode will fail if CR4.PCIDE is set. */
> + if (boot_cpu_has(X86_FEATURE_PCID))
cpu_feature_enabled() is what we use everywhere now.
> + cr4_clear_bits(X86_CR4_PCIDE);
> +
> + asm volatile("xorq %%r15, %%r15\n"
> + "xorq %%r14, %%r14\n"
> + "xorq %%r13, %%r13\n"
> + "xorq %%r12, %%r12\n"
> + "xorq %%r11, %%r11\n"
> + "xorq %%r10, %%r10\n"
> + "xorq %%r9, %%r9\n"
> + "xorq %%r8, %%r8\n"
> + "xorq %%rsi, %%rsi\n"
> + "xorq %%rdi, %%rdi\n"
> + "xorq %%rsp, %%rsp\n"
> + "xorq %%rbp, %%rbp\n"
Use xorl and the 32-bit regs is enough - zero extension.
> + "ljmpl *%0" : :
> + "m" (real_mode_header->sev_real_ap_park_asm),
> + "b" (sev_es_jump_table_pa >> 4));
In any case, this asm needs comments: why those regs, why
sev_es_jump_table_pa >> 4 in rbx (I found later in the patch why) and so
on.
> diff --git a/arch/x86/realmode/rm/header.S b/arch/x86/realmode/rm/header.S
> index 8c1db5bf5d78..6c17f8fd1eb4 100644
> --- a/arch/x86/realmode/rm/header.S
> +++ b/arch/x86/realmode/rm/header.S
> @@ -22,6 +22,9 @@ SYM_DATA_START(real_mode_header)
> .long pa_trampoline_header
> #ifdef CONFIG_AMD_MEM_ENCRYPT
> .long pa_sev_es_trampoline_start
> + .long pa_sev_ap_park_asm
> + .long __KERNEL32_CS
> + .long pa_sev_ap_park_gdt;
> #endif
> #ifdef CONFIG_X86_64
> .long pa_trampoline_pgd;
> diff --git a/arch/x86/realmode/rm/sev_ap_park.S b/arch/x86/realmode/rm/sev_ap_park.S
arch/x86/realmode/rm/sev.S
is perfectly fine I guess.
> new file mode 100644
> index 000000000000..0b63d0569d4d
> --- /dev/null
> +++ b/arch/x86/realmode/rm/sev_ap_park.S
> @@ -0,0 +1,89 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +#include <linux/linkage.h>
> +#include <asm/segment.h>
> +#include <asm/page_types.h>
> +#include <asm/processor-flags.h>
> +#include <asm/msr-index.h>
> +#include <asm/sev-ap-jumptable.h>
> +#include "realmode.h"
> +
> + .section ".text32", "ax"
> + .code32
> +/*
"This is executed by ... when ... "
> + * The following code switches to 16-bit protected mode and sets up the
> + * execution environment for the AP Jump Table blob. Then it jumps to the AP
> + * Jump Table to park the AP.
> + *
> + * The code was copied from reboot.S and modified to fit the SEV-ES requirements
> + * for AP parking.
That sentence belongs at most in the commit message.
> When this code is entered, all registers except %EAX-%EDX are
%eax, etc. Lowercase pls.
> + * in reset state.
> + *
> + * The AP Jump Table physical base address is in %EBX upon entry.
> + *
> + * %EAX, %ECX, %EDX and EFLAGS are undefined. Only use registers %EAX-%EDX and
> + * %ESP in this code.
> + */
> +SYM_CODE_START(sev_ap_park_asm)
sev_ap_park
> +
> + /* Switch to trampoline GDT as it is guaranteed < 4 GiB */
> + movl $__KERNEL_DS, %eax
> + movl %eax, %ds
> + lgdt pa_tr_gdt
> +
> + /* Disable paging to drop us out of long mode */
> + movl %cr0, %eax
> + btcl $X86_CR0_PG_BIT, %eax
> + movl %eax, %cr0
> +
/* Start executing from 32-bit addresses or so, I guess...
> + ljmpl $__KERNEL32_CS, $pa_sev_ap_park_paging_off
Please add a comment also about those pa_ things because they look like
magic but they're sed-generated into arch/x86/realmode/rm/pasyms.h by
the Makefile in that same dir.
> +SYM_INNER_LABEL(sev_ap_park_paging_off, SYM_L_GLOBAL)
Global symbol but used only in this file. .L-prefix then?
> + /* Clear EFER */
> + movl $0, %eax
> + movl $0, %edx
both: xorl
> + movl $MSR_EFER, %ecx
> + wrmsr
> +
> + /* Clear CR3 */
> + movl $0, %ecx
ditto
> + movl %ecx, %cr3
> +
> + /* Set up the IDT for real mode. */
> + lidtl pa_machine_real_restart_idt
> +
> + /*
> + * Load the GDT with the 16-bit segments for the AP Jump Table
> + */
/* Load the GDT with the 16-bit segments for the AP Jump Table */
works too.
> + lgdtl pa_sev_ap_park_gdt
> +
> + /* Setup Code and Data segments for AP Jump Table */
... code and data segments ...
you have been reading too much vendor text where they love to capitalize
everything.
> + movw $SEV_APJT_DS16, %ax
> + movw %ax, %ds
> + movw %ax, %ss
> +
> + /* Jump to the AP Jump Table into 16 bit protected mode */
> + ljmpw $SEV_APJT_CS16, $SEV_APJT_ENTRY
> +SYM_CODE_END(sev_ap_park_asm)
> +
> + .data
> + .balign 16
> +SYM_DATA_START(sev_ap_park_gdt)
> + /* Self-pointer */
> + .word sev_ap_park_gdt_end - sev_ap_park_gdt - 1
> + .long pa_sev_ap_park_gdt
> + .word 0
> +
> + /*
> + * Offset 0x8
> + * 32 bit code segment descriptor pointing to AP Jump table base
> + * Setup at runtime in sev_es_setup_ap_jump_table_data().
> + */
> + .quad 0
> +
> + /*
> + * Offset 0x10
> + * 32 bit data segment descriptor pointing to AP Jump table base
> + * Setup at runtime in sev_es_setup_ap_jump_table_data().
> + */
> + .quad 0
> +SYM_DATA_END_LABEL(sev_ap_park_gdt, SYM_L_GLOBAL, sev_ap_park_gdt_end)
> --
> 2.33.0
>
--
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquette
On Fri, Nov 12, 2021 at 05:33:05PM +0100, Borislav Petkov wrote:
> On Mon, Sep 13, 2021 at 05:55:59PM +0200, Joerg Roedel wrote:
> > + "ljmpl *%0" : :
> > + "m" (real_mode_header->sev_real_ap_park_asm),
> > + "b" (sev_es_jump_table_pa >> 4));
>
> In any case, this asm needs comments: why those regs, why
> sev_es_jump_table_pa >> 4 in rbx (I found later in the patch why) and so
> on.
Turned out the jump_table_pa is not used in asm code anymore. It was a
left-over from a previous version of the patch, it is removed now.
> > +SYM_INNER_LABEL(sev_ap_park_paging_off, SYM_L_GLOBAL)
>
> Global symbol but used only in this file. .L-prefix then?
It needs to be a global symbol so the pa_ variant can be generated.
Regards,
--
J?rg R?del
[email protected]
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5
90409 N?rnberg
Germany
(HRB 36809, AG N?rnberg)
Gesch?ftsf?hrer: Ivo Totev