Re: [PATCH 1/3] arm64: clean up trampoline vector loads

From: Mark Rutland
Date: Tue Mar 24 2020 - 06:52:28 EST


On Mon, Mar 23, 2020 at 10:42:30PM +0200, RÃmi Denis-Courmont wrote:
> Le maanantaina 23. maaliskuuta 2020, 21.04.09 EET Catalin Marinas a Ãcrit :
> > On Mon, Mar 23, 2020 at 12:14:37PM +0000, Mark Rutland wrote:
> > > On Mon, Mar 23, 2020 at 02:08:53PM +0200, RÃmi Denis-Courmont wrote:
> > > > Le maanantaina 23. maaliskuuta 2020, 14.07.00 EET Mark Rutland a Ãcrit :
> > > > > On Thu, Mar 19, 2020 at 11:14:05AM +0200, RÃmi Denis-Courmont wrote:
> > > > > > From: RÃmi Denis-Courmont <remi.denis.courmont@xxxxxxxxxx>
> > > > > >
> > > > > > This switches from custom instruction patterns to the regular large
> > > > > > memory model sequence with ADRP and LDR. In doing so, the ADD
> > > > > > instruction can be eliminated in the SDEI handler, and the code no
> > > > > > longer assumes that the trampoline vectors and the vectors address
> > > > > > both
> > > > > > start on a page boundary.
> > > > > >
> > > > > > Signed-off-by: RÃmi Denis-Courmont <remi.denis.courmont@xxxxxxxxxx>
> > > > > > ---
> > > > > >
> > > > > > arch/arm64/kernel/entry.S | 9 ++++-----
> > > > > > 1 file changed, 4 insertions(+), 5 deletions(-)
> > > > > >
> > > > > > diff --git a/arch/arm64/kernel/entry.S b/arch/arm64/kernel/entry.S
> > > > > > index e5d4e30ee242..24f828739696 100644
> > > > > > --- a/arch/arm64/kernel/entry.S
> > > > > > +++ b/arch/arm64/kernel/entry.S
> > > > > > @@ -805,9 +805,9 @@ alternative_else_nop_endif
> > > > > >
> > > > > > 2:
> > > > > > tramp_map_kernel x30
> > > > > >
> > > > > > #ifdef CONFIG_RANDOMIZE_BASE
> > > > > >
> > > > > > - adr x30, tramp_vectors + PAGE_SIZE
> > > > > > + adrp x30, tramp_vectors + PAGE_SIZE
> > > > > >
> > > > > > alternative_insn isb, nop, ARM64_WORKAROUND_QCOM_FALKOR_E1003
> > > > > >
> > > > > > - ldr x30, [x30]
> > > > > > + ldr x30, [x30, #:lo12:__entry_tramp_data_start]
> > > > >
> > > > > I think this is busted for !4K kernels once we reduce the alignment of
> > > > > __entry_tramp_data_start.
> > > > >
> > > > > The ADRP gives us a 64K aligned address (with bits 15:0 clear). The
> > > > > lo12
> > > > > relocation gives us bits 11:0, so we haven't accounted for bits 15:12.
> > > >
> > > > IMU, ADRP gives a 4K aligned value, regardless of MMU (TCR) settings.
> > >
> > > Sorry, I had erroneously assumed tramp_vectors was page aligned. The
> > > issue still stands -- we haven't accounted for bits 15:12, as those can
> > > differ between tramp_vectors and __entry_tramp_data_start.
>
> Does that mean that the SDEI code never worked with page size > 4 KiB?

I think this happens to work, but is fragile. Because nothing happens to
get placed in .rodata between the _entry_tramp_data_start data and the
__sdei_asm_trampoline_next_handler data, the
__sdei_asm_trampoline_next_handler data doesn't spill into a separate
page from the _entry_tramp_data_start data.

If we did start adding stuff into .rodata between those two, there'd be
a bigger risk of things going wrong. That was why I suggested a
.entry.tramp.data section previously.

> > Should we just use adrp on __entry_tramp_data_start? Anyway, the diff
> > below doesn't solve the issue I'm seeing (only reverting patch 3).
>
> AFAIU, the preexisting code uses the manual PAGE_SIZE offset because the offset
> in the main vmlinux does not match the architected offset inside the fixmap. If
> so, then using the symbol directly will not work at all.

Indeed. I can't see a neat way of avoiding this right now, so should we
drop these patches and leave the code as-is (but with comments as to the
special requirements that it has)?

Thanks,
Mark.