This is a bug that can cause early crashes in configurations with a
.exit.text section smaller than a page and a .init.text section that
ends in the beginning of a physical page (this is kinda random, which
might explain why this wasn't really encountered before).
The init sections are ordered like this:
.init.text
.exit.text
.init.data
Currently, these sections aren't page aligned.
Because the init code is mapped read-only at runtime and because the
.init.text section can potentially reside on the same physical page as
.init.data, the beginning of .init.data might be mapped read-only along
with .init.text.
Then when the kernel tries to modify a variable in .init.data (like
kthreadd_done, used in kernel_init()) the kernel panics.
To avoid this, I made these sections page aligned.
Fixes: 060ef9d89d18 ("powerpc32: PAGE_EXEC required for inittext")
Signed-off-by: Ariel Marcovitch <[email protected]>
---
arch/powerpc/kernel/vmlinux.lds.S | 7 +++++++
1 file changed, 7 insertions(+)
diff --git a/arch/powerpc/kernel/vmlinux.lds.S b/arch/powerpc/kernel/vmlinux.lds.S
index 326e113d2e45..e3a7c90c03f4 100644
--- a/arch/powerpc/kernel/vmlinux.lds.S
+++ b/arch/powerpc/kernel/vmlinux.lds.S
@@ -179,6 +179,11 @@ SECTIONS
#endif
} :text
+ /* .init.text is made RO and .exit.text is not, so we must
+ * ensure these sections reside in separate physical pages.
+ */
+ . = ALIGN(PAGE_SIZE);
+
/* .exit.text is discarded at runtime, not link time,
* to deal with references from __bug_table
*/
@@ -186,6 +191,8 @@ SECTIONS
EXIT_TEXT
}
+ . = ALIGN(PAGE_SIZE);
+
.init.data : AT(ADDR(.init.data) - LOAD_OFFSET) {
INIT_DATA
}
base-commit: 1398820fee515873379809a6415930ad0764b2f6
--
2.17.1
Le 13/12/2020 à 19:35, Ariel Marcovitch a écrit :
> This is a bug that can cause early crashes in configurations with a
> .exit.text section smaller than a page and a .init.text section that
> ends in the beginning of a physical page (this is kinda random, which
> might explain why this wasn't really encountered before).
It can cause, or it causes ? Did you encounter the issue ?
>
> The init sections are ordered like this:
> .init.text
> .exit.text
> .init.data
>
> Currently, these sections aren't page aligned.
>
> Because the init code is mapped read-only at runtime and because the
> .init.text section can potentially reside on the same physical page as
> .init.data, the beginning of .init.data might be mapped read-only along
> with .init.text.
init code is mapped PAGE_KERNEL_TEXT.
Whether PAGE_KERNEL_TEXT is read-only or not depends on the selected options.
>
> Then when the kernel tries to modify a variable in .init.data (like
> kthreadd_done, used in kernel_init()) the kernel panics.
>
> To avoid this, I made these sections page aligned.
Should write this unpersonal, something like "To avoid this, make these sections page aligned"
>
> Fixes: 060ef9d89d18 ("powerpc32: PAGE_EXEC required for inittext")
> Signed-off-by: Ariel Marcovitch <[email protected]>
> ---
> arch/powerpc/kernel/vmlinux.lds.S | 7 +++++++
> 1 file changed, 7 insertions(+)
>
> diff --git a/arch/powerpc/kernel/vmlinux.lds.S b/arch/powerpc/kernel/vmlinux.lds.S
> index 326e113d2e45..e3a7c90c03f4 100644
> --- a/arch/powerpc/kernel/vmlinux.lds.S
> +++ b/arch/powerpc/kernel/vmlinux.lds.S
> @@ -179,6 +179,11 @@ SECTIONS
> #endif
> } :text
>
> + /* .init.text is made RO and .exit.text is not, so we must
> + * ensure these sections reside in separate physical pages.
> + */
> + . = ALIGN(PAGE_SIZE);
> +
In principle, as it is text, it should be made RO as well. But what happens at the begining doesn't
really matter, anyway .exit.text should never be executed and is discarded together with init text.
So, I think it is OK the live with it as is for the time being.
Making it page aligned makes sense anyway.
Should we make _einittext page aligned instead, just like _etext ?
> /* .exit.text is discarded at runtime, not link time,
> * to deal with references from __bug_table
> */
> @@ -186,6 +191,8 @@ SECTIONS
> EXIT_TEXT
> }
>
> + . = ALIGN(PAGE_SIZE);
> +
Here for sure, as you explain in the coming log, this needs to be separated from init text. So
making it aligned is a must.
> .init.data : AT(ADDR(.init.data) - LOAD_OFFSET) {
> INIT_DATA
> }
>
> base-commit: 1398820fee515873379809a6415930ad0764b2f6
>
Christophe
On Fri, Dec 18, 2020 at 5:39 PM Christophe Leroy <
[email protected]> wrote:
> It can cause, or it causes ? Did you encounter the issue ?
>
Yes, in configs that result in the section layout I described, the crush is
consistent.
>
> > The init sections are ordered like this:
> > .init.text
> > .exit.text
> > .init.data
> >
> > Currently, these sections aren't page aligned.
> >
> > Because the init code is mapped read-only at runtime and because the
> > .init.text section can potentially reside on the same physical page as
> > .init.data, the beginning of .init.data might be mapped read-only along
> > with .init.text.
>
> init code is mapped PAGE_KERNEL_TEXT.
>
> Whether PAGE_KERNEL_TEXT is read-only or not depends on the selected
> options.
>
You are right, of course. Should I change the commit message to 'might be
mapped' or something?
>
> > Then when the kernel tries to modify a variable in .init.data (like
> > kthreadd_done, used in kernel_init()) the kernel panics.
> >
> > To avoid this, I made these sections page aligned.
>
> Should write this unpersonal, something like "To avoid this, make these
> sections page aligned"
>
Noted, thanks.
>
> >
> > Fixes: 060ef9d89d18 ("powerpc32: PAGE_EXEC required for inittext")
> > Signed-off-by: Ariel Marcovitch <[email protected]>
> > ---
> > arch/powerpc/kernel/vmlinux.lds.S | 7 +++++++
> > 1 file changed, 7 insertions(+)
> >
> > diff --git a/arch/powerpc/kernel/vmlinux.lds.S
> b/arch/powerpc/kernel/vmlinux.lds.S
> > index 326e113d2e45..e3a7c90c03f4 100644
> > --- a/arch/powerpc/kernel/vmlinux.lds.S
> > +++ b/arch/powerpc/kernel/vmlinux.lds.S
> > @@ -179,6 +179,11 @@ SECTIONS
> > #endif
> > } :text
> >
> > + /* .init.text is made RO and .exit.text is not, so we must
> > + * ensure these sections reside in separate physical pages.
> > + */
> > + . = ALIGN(PAGE_SIZE);
> > +
>
> In principle, as it is text, it should be made RO as well. But what
> happens at the begining doesn't
> really matter, anyway .exit.text should never be executed and is discarded
> together with init text.
> So, I think it is OK the live with it as is for the time being.
> Making it page aligned makes sense anyway.
>
> Should we make _einittext page aligned instead, just like _etext ?
Yes, this will probably be better (because when _einittext is not aligned,
the part of the page after _einittext is mapped RO implicitly, and it's
hard to notice from the code). I suppose you mean something like this:
_sinittext = .;
INIT_TEXT
+
+ . = ALIGN(.);
_einittext = .;
> /* .exit.text is discarded at runtime, not link time,
> > * to deal with references from __bug_table
> > */
> > @@ -186,6 +191,8 @@ SECTIONS
> > EXIT_TEXT
> > }
> >
> > + . = ALIGN(PAGE_SIZE);
> > +
>
> Here for sure, as you explain in the coming log, this needs to be
> separated from init text. So
> making it aligned is a must.
> > .init.data : AT(ADDR(.init.data) - LOAD_OFFSET) {
> > INIT_DATA
> > }
> >
> > base-commit: 1398820fee515873379809a6415930ad0764b2f6
> >
>
> Christophe
>
Thanks for your time,
Ariel Marcovitch