All functions referenced via __PV_IS_CALLEE_SAVE() need to be assembler
functions, as those functions calls are hidden from gcc. In case the
kernel is compiled with "-fzero-call-used-regs" the compiler will
clobber caller-saved registers at the end of C functions, which will
result in unexpectedly zeroed registers at the call site of the
related paravirt functions.
Replace the C functions with DEFINE_PARAVIRT_ASM() constructs using
the same instructions as the related paravirt calls in the
PVOP_ALT_[V]CALLEE*() macros.
Signed-off-by: Juergen Gross <[email protected]>
---
arch/x86/include/asm/paravirt_types.h | 8 +++++++-
arch/x86/kernel/paravirt.c | 27 ++++++---------------------
2 files changed, 13 insertions(+), 22 deletions(-)
diff --git a/arch/x86/include/asm/paravirt_types.h b/arch/x86/include/asm/paravirt_types.h
index 8c1da419260f..49f5c6955229 100644
--- a/arch/x86/include/asm/paravirt_types.h
+++ b/arch/x86/include/asm/paravirt_types.h
@@ -562,8 +562,14 @@ void paravirt_flush_lazy_mmu(void);
void _paravirt_nop(void);
void paravirt_BUG(void);
-u64 _paravirt_ident_64(u64);
unsigned long paravirt_ret0(void);
+#ifdef CONFIG_PARAVIRT_XXL
+u64 _paravirt_ident_64(u64);
+unsigned long pv_native_save_fl(void);
+void pv_native_irq_disable(void);
+void pv_native_irq_enable(void);
+unsigned long pv_native_read_cr2(void);
+#endif
#define paravirt_nop ((void *)_paravirt_nop)
diff --git a/arch/x86/kernel/paravirt.c b/arch/x86/kernel/paravirt.c
index 42e182868873..d25ac4b08c41 100644
--- a/arch/x86/kernel/paravirt.c
+++ b/arch/x86/kernel/paravirt.c
@@ -64,11 +64,11 @@ static unsigned paravirt_patch_call(void *insn_buff, const void *target,
}
#ifdef CONFIG_PARAVIRT_XXL
-/* identity function, which can be inlined */
-u64 notrace _paravirt_ident_64(u64 x)
-{
- return x;
-}
+DEFINE_PARAVIRT_ASM(_paravirt_ident_64, "mov %rdi, %rax", .text);
+DEFINE_PARAVIRT_ASM(pv_native_save_fl, "pushf; pop %rax", .text);
+DEFINE_PARAVIRT_ASM(pv_native_irq_disable, "cli", .text);
+DEFINE_PARAVIRT_ASM(pv_native_irq_enable, "sti", .text);
+DEFINE_PARAVIRT_ASM(pv_native_read_cr2, "mov %cr2, %rax", .text);
#endif
DEFINE_STATIC_KEY_TRUE(virt_spin_lock_key);
@@ -197,11 +197,6 @@ void paravirt_end_context_switch(struct task_struct *next)
arch_enter_lazy_mmu_mode();
}
-static noinstr unsigned long pv_native_read_cr2(void)
-{
- return native_read_cr2();
-}
-
static noinstr void pv_native_write_cr2(unsigned long val)
{
native_write_cr2(val);
@@ -222,16 +217,6 @@ noinstr void pv_native_wbinvd(void)
native_wbinvd();
}
-static noinstr void pv_native_irq_enable(void)
-{
- native_irq_enable();
-}
-
-static noinstr void pv_native_irq_disable(void)
-{
- native_irq_disable();
-}
-
static noinstr void pv_native_safe_halt(void)
{
native_safe_halt();
@@ -298,7 +283,7 @@ struct paravirt_patch_template pv_ops = {
.cpu.end_context_switch = paravirt_nop,
/* Irq ops. */
- .irq.save_fl = __PV_IS_CALLEE_SAVE(native_save_fl),
+ .irq.save_fl = __PV_IS_CALLEE_SAVE(pv_native_save_fl),
.irq.irq_disable = __PV_IS_CALLEE_SAVE(pv_native_irq_disable),
.irq.irq_enable = __PV_IS_CALLEE_SAVE(pv_native_irq_enable),
.irq.safe_halt = pv_native_safe_halt,
--
2.35.3
On Wed, Mar 08, 2023 at 04:42:10PM +0100, Juergen Gross wrote:
> All functions referenced via __PV_IS_CALLEE_SAVE() need to be assembler
> functions, as those functions calls are hidden from gcc. In case the
> kernel is compiled with "-fzero-call-used-regs" the compiler will
> clobber caller-saved registers at the end of C functions, which will
> result in unexpectedly zeroed registers at the call site of the
> related paravirt functions.
>
> Replace the C functions with DEFINE_PARAVIRT_ASM() constructs using
> the same instructions as the related paravirt calls in the
> PVOP_ALT_[V]CALLEE*() macros.
>
> Signed-off-by: Juergen Gross <[email protected]>
> ---
> arch/x86/include/asm/paravirt_types.h | 8 +++++++-
> arch/x86/kernel/paravirt.c | 27 ++++++---------------------
> 2 files changed, 13 insertions(+), 22 deletions(-)
Right, works with my particular reproducer.
Turning them into asm prevents the compiler from doing the
callee-clobbered zeroing and that's fine as this whole paravirt gunk is
hiding the "CALL" insn from it and you putting them in asm is in line
with this.
And a negative diffstat..
So yeah, I'll queue it soon unless someone objects.
Long term, I think we should continue switching all that pv stuff to
using the alternatives.
Thx.
--
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquette
On 09.03.23 14:39, Borislav Petkov wrote:
> On Wed, Mar 08, 2023 at 04:42:10PM +0100, Juergen Gross wrote:
>> All functions referenced via __PV_IS_CALLEE_SAVE() need to be assembler
>> functions, as those functions calls are hidden from gcc. In case the
>> kernel is compiled with "-fzero-call-used-regs" the compiler will
>> clobber caller-saved registers at the end of C functions, which will
>> result in unexpectedly zeroed registers at the call site of the
>> related paravirt functions.
>>
>> Replace the C functions with DEFINE_PARAVIRT_ASM() constructs using
>> the same instructions as the related paravirt calls in the
>> PVOP_ALT_[V]CALLEE*() macros.
>>
>> Signed-off-by: Juergen Gross <[email protected]>
>> ---
>> arch/x86/include/asm/paravirt_types.h | 8 +++++++-
>> arch/x86/kernel/paravirt.c | 27 ++++++---------------------
>> 2 files changed, 13 insertions(+), 22 deletions(-)
>
> Right, works with my particular reproducer.
>
> Turning them into asm prevents the compiler from doing the
> callee-clobbered zeroing and that's fine as this whole paravirt gunk is
> hiding the "CALL" insn from it and you putting them in asm is in line
> with this.
>
> And a negative diffstat..
>
> So yeah, I'll queue it soon unless someone objects.
Thanks.
> Long term, I think we should continue switching all that pv stuff to
> using the alternatives.
The "normal" cases not using alternatives should rather be switched to
static calls.
Whether it is possible to mix a static call with alternatives needs to
be evaluated.
Juergen
On Fri, Mar 10, 2023 at 07:24:17AM +0100, Juergen Gross wrote:
> The "normal" cases not using alternatives should rather be switched to
> static calls.
Or that.
> Whether it is possible to mix a static call with alternatives needs to
> be evaluated.
I'd prefer not to mix them. Either should be fine and if neither have
the required functionality, then it should be added depending on which
- static calls or alternatives - would make things simpler.
I'd love to get rid of the whole paravirt glue and use the facilities we
have in the tree instead.
But no hurry - it should be nice and clean work. :-)
--
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquette
On Wed, Mar 08, 2023 at 04:42:10PM +0100, Juergen Gross wrote:
> All functions referenced via __PV_IS_CALLEE_SAVE() need to be assembler
> functions, as those functions calls are hidden from gcc. In case the
> kernel is compiled with "-fzero-call-used-regs" the compiler will
> clobber caller-saved registers at the end of C functions, which will
> result in unexpectedly zeroed registers at the call site of the
> related paravirt functions.
>
> Replace the C functions with DEFINE_PARAVIRT_ASM() constructs using
> the same instructions as the related paravirt calls in the
> PVOP_ALT_[V]CALLEE*() macros.
>
> Signed-off-by: Juergen Gross <[email protected]>
> ---
> arch/x86/include/asm/paravirt_types.h | 8 +++++++-
> arch/x86/kernel/paravirt.c | 27 ++++++---------------------
> 2 files changed, 13 insertions(+), 22 deletions(-)
objtool's not happy with this for whatever reason. I'll look later as to
why. .config is allmodconfig with this patch ontop of tip:x86/paravirt:
vmlinux.o: warning: objtool: pv_ops[31]: pv_native_irq_disable
vmlinux.o: warning: objtool: default_idle+0x1e: call to {dynamic}() leaves .noinstr.text section
vmlinux.o: warning: objtool: pv_ops[31]: pv_native_irq_disable
vmlinux.o: warning: objtool: mwait_idle+0x5d: call to {dynamic}() leaves .noinstr.text section
vmlinux.o: warning: objtool: pv_ops[31]: pv_native_irq_disable
vmlinux.o: warning: objtool: cpu_idle_poll.isra.0+0x94: call to {dynamic}() leaves .noinstr.text section
vmlinux.o: warning: objtool: pv_ops[31]: pv_native_irq_disable
vmlinux.o: warning: objtool: intel_idle_irq+0xab: call to {dynamic}() leaves .noinstr.text section
vmlinux.o: warning: objtool: pv_ops[31]: pv_native_irq_disable
vmlinux.o: warning: objtool: acpi_safe_halt+0x2a: call to {dynamic}() leaves .noinstr.text section
vmlinux.o: warning: objtool: pv_ops[31]: pv_native_irq_disable
vmlinux.o: warning: objtool: poll_idle+0x86: call to {dynamic}() leaves .noinstr.text section
vmlinux.o: warning: objtool: pv_ops[42]: pv_native_read_cr2
vmlinux.o: warning: objtool: exc_double_fault+0x3b: call to {dynamic}() leaves .noinstr.text section
vmlinux.o: warning: objtool: pv_ops[42]: pv_native_read_cr2
vmlinux.o: warning: objtool: exc_nmi+0x188: call to {dynamic}() leaves .noinstr.text section
vmlinux.o: warning: objtool: pv_ops[30]: pv_native_save_fl
vmlinux.o: warning: objtool: __sev_put_ghcb+0x11: call to {dynamic}() leaves .noinstr.text section
vmlinux.o: warning: objtool: pv_ops[30]: pv_native_save_fl
vmlinux.o: warning: objtool: __sev_get_ghcb+0x13: call to {dynamic}() leaves .noinstr.text section
vmlinux.o: warning: objtool: pv_ops[42]: pv_native_read_cr2
vmlinux.o: warning: objtool: exc_page_fault+0x1e: call to {dynamic}() leaves .noinstr.text section
vmlinux.o: warning: objtool: pv_ops[30]: pv_native_save_fl
vmlinux.o: warning: objtool: lockdep_hardirqs_on+0xd0: call to {dynamic}() leaves .noinstr.text section
vmlinux.o: warning: objtool: pv_ops[30]: pv_native_save_fl
vmlinux.o: warning: objtool: lockdep_hardirqs_off+0xe7: call to {dynamic}() leaves .noinstr.text section
vmlinux.o: warning: objtool: pv_ops[30]: pv_native_save_fl
vmlinux.o: warning: objtool: look_up_lock_class+0x52: call to {dynamic}() leaves .noinstr.text section
vmlinux.o: warning: objtool: pv_ops[32]: pv_native_irq_enable
vmlinux.o: warning: objtool: lock_is_held_type+0x143: call to {dynamic}() leaves .noinstr.text section
vmlinux.o: warning: objtool: pv_ops[30]: pv_native_save_fl
vmlinux.o: warning: objtool: ct_kernel_enter.constprop.0+0x37: call to {dynamic}() leaves .noinstr.text section
vmlinux.o: warning: objtool: pv_ops[32]: pv_native_irq_enable
vmlinux.o: warning: objtool: ct_idle_exit+0x51: call to {dynamic}() leaves .noinstr.text section
vmlinux.o: warning: objtool: pv_ops[30]: pv_native_save_fl
vmlinux.o: warning: objtool: ct_idle_enter+0xe: call to {dynamic}() leaves .noinstr.text section
vmlinux.o: warning: objtool: pv_ops[30]: pv_native_save_fl
vmlinux.o: warning: objtool: check_preemption_disabled+0x4c: call to {dynamic}() leaves .noinstr.text section
--
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquette
On Wed, Mar 08, 2023 at 04:42:10PM +0100, Juergen Gross wrote:
> +DEFINE_PARAVIRT_ASM(pv_native_irq_disable, "cli", .text);
> +DEFINE_PARAVIRT_ASM(pv_native_irq_enable, "sti", .text);
> +DEFINE_PARAVIRT_ASM(pv_native_read_cr2, "mov %cr2, %rax", .text);
per these v, the above ^ should be in .noinstr.text
> -static noinstr unsigned long pv_native_read_cr2(void)
> -static noinstr void pv_native_irq_enable(void)
> -static noinstr void pv_native_irq_disable(void)
On 16.03.23 21:14, Peter Zijlstra wrote:
> On Wed, Mar 08, 2023 at 04:42:10PM +0100, Juergen Gross wrote:
>
>> +DEFINE_PARAVIRT_ASM(pv_native_irq_disable, "cli", .text);
>> +DEFINE_PARAVIRT_ASM(pv_native_irq_enable, "sti", .text);
>> +DEFINE_PARAVIRT_ASM(pv_native_read_cr2, "mov %cr2, %rax", .text);
>
> per these v, the above ^ should be in .noinstr.text
Yes, and I'm inclined to even put pv_native_save_fl into the noinstr
section. After paravirt patching it isn't called anymore anyway.
Juergen