The __startup_64 function assumes all symbols have relocated addresses
instead of the current boot virtual address. PIE generated code favor
relative addresses making all virtual and physical address math incorrect.
If PIE is enabled, build head64.c as mcmodel large instead to ensure absolute
references on all memory access. Add a global __force_order variable required
when using a large model with read_cr* functions.
To build head64.c as mcmodel=large, disable the retpoline gcc flags.
This code is used at early boot and removed later, it doesn't need
retpoline mitigation.
Position Independent Executable (PIE) support will allow to extend the
KASLR randomization range below 0xffffffff80000000.
Signed-off-by: Thomas Garnier <[email protected]>
---
arch/x86/kernel/Makefile | 6 ++++++
arch/x86/kernel/head64.c | 3 +++
2 files changed, 9 insertions(+)
diff --git a/arch/x86/kernel/Makefile b/arch/x86/kernel/Makefile
index 00b7e27bc2b7..1f98f52eab9f 100644
--- a/arch/x86/kernel/Makefile
+++ b/arch/x86/kernel/Makefile
@@ -22,6 +22,12 @@ CFLAGS_REMOVE_early_printk.o = -pg
CFLAGS_REMOVE_head64.o = -pg
endif
+ifdef CONFIG_X86_PIE
+# Remove PIE and retpoline flags that are incompatible with mcmodel=large
+CFLAGS_REMOVE_head64.o += -fPIE -mindirect-branch=thunk-extern -mindirect-branch-register
+CFLAGS_head64.o = -mcmodel=large
+endif
+
KASAN_SANITIZE_head$(BITS).o := n
KASAN_SANITIZE_dumpstack.o := n
KASAN_SANITIZE_dumpstack_$(BITS).o := n
diff --git a/arch/x86/kernel/head64.c b/arch/x86/kernel/head64.c
index 16b1cbd3a61e..22e81275495b 100644
--- a/arch/x86/kernel/head64.c
+++ b/arch/x86/kernel/head64.c
@@ -63,6 +63,9 @@ EXPORT_SYMBOL(vmemmap_base);
#define __head __section(.head.text)
+/* Required for read_cr3 when building as PIE */
+unsigned long __force_order;
+
static void __head *fixup_pointer(void *ptr, unsigned long physaddr)
{
return ptr - (void *)_text + (void *)physaddr;
--
2.20.1.495.gaa96b0ce6b-goog
On Thu, Jan 31, 2019 at 11:24:20AM -0800, Thomas Garnier wrote:
> The __startup_64 function assumes all symbols have relocated addresses
> instead of the current boot virtual address. PIE generated code favor
> relative addresses making all virtual and physical address math incorrect.
> If PIE is enabled, build head64.c as mcmodel large instead to ensure absolute
> references on all memory access. Add a global __force_order variable required
> when using a large model with read_cr* functions.
>
> To build head64.c as mcmodel=large, disable the retpoline gcc flags.
> This code is used at early boot and removed later, it doesn't need
> retpoline mitigation.
>
> Position Independent Executable (PIE) support will allow to extend the
> KASLR randomization range below 0xffffffff80000000.
>
> Signed-off-by: Thomas Garnier <[email protected]>
> ---
> arch/x86/kernel/Makefile | 6 ++++++
> arch/x86/kernel/head64.c | 3 +++
> 2 files changed, 9 insertions(+)
>
> diff --git a/arch/x86/kernel/Makefile b/arch/x86/kernel/Makefile
> index 00b7e27bc2b7..1f98f52eab9f 100644
> --- a/arch/x86/kernel/Makefile
> +++ b/arch/x86/kernel/Makefile
> @@ -22,6 +22,12 @@ CFLAGS_REMOVE_early_printk.o = -pg
> CFLAGS_REMOVE_head64.o = -pg
> endif
>
> +ifdef CONFIG_X86_PIE
> +# Remove PIE and retpoline flags that are incompatible with mcmodel=large
> +CFLAGS_REMOVE_head64.o += -fPIE -mindirect-branch=thunk-extern -mindirect-branch-register
> +CFLAGS_head64.o = -mcmodel=large
> +endif
> +
> KASAN_SANITIZE_head$(BITS).o := n
> KASAN_SANITIZE_dumpstack.o := n
> KASAN_SANITIZE_dumpstack_$(BITS).o := n
> diff --git a/arch/x86/kernel/head64.c b/arch/x86/kernel/head64.c
> index 16b1cbd3a61e..22e81275495b 100644
> --- a/arch/x86/kernel/head64.c
> +++ b/arch/x86/kernel/head64.c
> @@ -63,6 +63,9 @@ EXPORT_SYMBOL(vmemmap_base);
>
> #define __head __section(.head.text)
>
> +/* Required for read_cr3 when building as PIE */
> +unsigned long __force_order;
> +
I believe it only needed for GCC < 5. Newer GCC can eliminate the
reference. See my comment in arch/x86/boot/compressed/pgtable_64.c.
Maybe we should expand the comment here too?
--
Kirill A. Shutemov
On Fri, Feb 1, 2019 at 3:15 AM Kirill A. Shutemov <[email protected]> wrote:
>
> On Thu, Jan 31, 2019 at 11:24:20AM -0800, Thomas Garnier wrote:
> > The __startup_64 function assumes all symbols have relocated addresses
> > instead of the current boot virtual address. PIE generated code favor
> > relative addresses making all virtual and physical address math incorrect.
> > If PIE is enabled, build head64.c as mcmodel large instead to ensure absolute
> > references on all memory access. Add a global __force_order variable required
> > when using a large model with read_cr* functions.
> >
> > To build head64.c as mcmodel=large, disable the retpoline gcc flags.
> > This code is used at early boot and removed later, it doesn't need
> > retpoline mitigation.
> >
> > Position Independent Executable (PIE) support will allow to extend the
> > KASLR randomization range below 0xffffffff80000000.
> >
> > Signed-off-by: Thomas Garnier <[email protected]>
> > ---
> > arch/x86/kernel/Makefile | 6 ++++++
> > arch/x86/kernel/head64.c | 3 +++
> > 2 files changed, 9 insertions(+)
> >
> > diff --git a/arch/x86/kernel/Makefile b/arch/x86/kernel/Makefile
> > index 00b7e27bc2b7..1f98f52eab9f 100644
> > --- a/arch/x86/kernel/Makefile
> > +++ b/arch/x86/kernel/Makefile
> > @@ -22,6 +22,12 @@ CFLAGS_REMOVE_early_printk.o = -pg
> > CFLAGS_REMOVE_head64.o = -pg
> > endif
> >
> > +ifdef CONFIG_X86_PIE
> > +# Remove PIE and retpoline flags that are incompatible with mcmodel=large
> > +CFLAGS_REMOVE_head64.o += -fPIE -mindirect-branch=thunk-extern -mindirect-branch-register
> > +CFLAGS_head64.o = -mcmodel=large
> > +endif
> > +
> > KASAN_SANITIZE_head$(BITS).o := n
> > KASAN_SANITIZE_dumpstack.o := n
> > KASAN_SANITIZE_dumpstack_$(BITS).o := n
> > diff --git a/arch/x86/kernel/head64.c b/arch/x86/kernel/head64.c
> > index 16b1cbd3a61e..22e81275495b 100644
> > --- a/arch/x86/kernel/head64.c
> > +++ b/arch/x86/kernel/head64.c
> > @@ -63,6 +63,9 @@ EXPORT_SYMBOL(vmemmap_base);
> >
> > #define __head __section(.head.text)
> >
> > +/* Required for read_cr3 when building as PIE */
> > +unsigned long __force_order;
> > +
>
> I believe it only needed for GCC < 5. Newer GCC can eliminate the
> reference. See my comment in arch/x86/boot/compressed/pgtable_64.c.
>
> Maybe we should expand the comment here too?
Make sense, I will add a similar comment in the next iteration. Thanks
for pointing that out.
>
> --
> Kirill A. Shutemov