On Tue, 2014-03-18 at 18:28 +0000, Catalin Marinas wrote:
> On Tue, Mar 18, 2014 at 02:40:29PM +0000, Mark Salter wrote:
> > On Tue, 2014-03-18 at 12:09 +0000, Catalin Marinas wrote:
> > > On Thu, Mar 13, 2014 at 10:47:04PM +0000, Leif Lindholm wrote:
> > > > --- /dev/null
> > > > +++ b/arch/arm64/kernel/efi-entry.S
> > > > @@ -0,0 +1,93 @@
> > > > +/*
> > > > + * EFI entry point.
> > > > + *
> > > > + * Copyright (C) 2013 Red Hat, Inc.
> > > > + * Author: Mark Salter <msalter@xxxxxxxxxx>
> > > > + *
> > > > + * This program is free software; you can redistribute it and/or modify
> > > > + * it under the terms of the GNU General Public License version 2 as
> > > > + * published by the Free Software Foundation.
> > > > + *
> > > > + */
> > > > +#include <linux/linkage.h>
> > > > +#include <linux/init.h>
> > > > +
> > > > +#include <asm/assembler.h>
> > > > +
> > > > +#define EFI_LOAD_ERROR 0x8000000000000001
> > > > +
> > > > + __INIT
> > > > +
> > > > + /*
> > > > + * We arrive here from the EFI boot manager with:
> > > > + *
> > > > + * * MMU on with identity-mapped RAM.
> > > > + * * Icache and Dcache on
> > > > + *
> > > > + * We will most likely be running from some place other than where
> > > > + * we want to be. The kernel image wants to be placed at TEXT_OFFSET
> > > > + * from start of RAM.
> > > > + */
> > > > +ENTRY(efi_stub_entry)
> > > > + stp x29, x30, [sp, #-32]!
> > > > +
> > > > + /*
> > > > + * Call efi_entry to do the real work.
> > > > + * x0 and x1 are already set up by firmware. Current runtime
> > > > + * address of image is calculated and passed via *image_addr.
> > > > + *
> > > > + * unsigned long efi_entry(void *handle,
> > > > + * efi_system_table_t *sys_table,
> > > > + * unsigned long *image_addr) ;
> > > > + */
> > > > + adrp x8, _text
> > > > + add x8, x8, #:lo12:_text
> > > > + add x2, sp, 16
> > > > + str x8, [x2]
> > > > + bl efi_entry
> > > > + cmn x0, #1
> > > > + b.eq efi_load_fail
> > > > +
> > > > + /*
> > > > + * efi_entry() will have relocated the kernel image if necessary
> > > > + * and we return here with device tree address in x0 and the kernel
> > > > + * entry point stored at *image_addr. Save those values in registers
> > > > + * which are preserved by __flush_dcache_all.
> > > > + */
> > > > + ldr x1, [sp, #16]
> > > > + mov x20, x0
> > > > + mov x21, x1
> > > > +
> > > > + /* Turn off Dcache and MMU */
> > > > + mrs x0, CurrentEL
> > > > + cmp x0, #PSR_MODE_EL2t
> > > > + ccmp x0, #PSR_MODE_EL2h, #0x4, ne
> > > > + b.ne 1f
> > > > + mrs x0, sctlr_el2
> > > > + bic x0, x0, #1 << 0 // clear SCTLR.M
> > > > + bic x0, x0, #1 << 2 // clear SCTLR.C
> > > > + msr sctlr_el2, x0
> > > > + isb
> > > > + b 2f
> > > > +1:
> > > > + mrs x0, sctlr_el1
> > > > + bic x0, x0, #1 << 0 // clear SCTLR.M
> > > > + bic x0, x0, #1 << 2 // clear SCTLR.C
> > > > + msr sctlr_el1, x0
> > > > + isb
> > > > +2:
> > > > + bl __flush_dcache_all
> > >
> > > In linux-next I'm pushing a patch which no longer exports the
> > > __flush_dcache_all function. The reason is that it doesn't really work
> > > if you have a (not fully transparent) external cache like on the Applied
> > > Micro boards. There other issues when running as a guest as well.
> > >
> > > If you know exactly what needs to be flushed here, can you use a range
> > > (MVA) operation?
> >
> > This is just before the EFI stub jumps to kernel proper. The only things
> > in the dcache would be from identity mapped references to RAM used by
> > UEFI. The booting.txt doc says dcache should be off and invalidated. I
> > am just wanting to comply with that. The code here doesn't really know
> > the extent of DRAM to flush by address.
>
> Does UEFI do anything with the caches before invoking the EFI_STUB code?
> I guess it doesn't since that's just another application for it. Can
> UEFI flush the caches via exit boot? When is this called?
>
> As I said, we have a real problem here since the EFI_STUB call does not
> have information about the SoC to be able to flush all the caches. But
> UEFI should know more about the hardware.
>
> If UEFI doesn't handle the caches, the only thing left to EFI_STUB is to
> flush by MVA. We don't need to flush the whole DRAM (and I would even
> recommend it) but at least the relevant kernel code/data touched with
> the MMU disabled.
>
So, it goes like this:
1) UEFI calls stub with MMU/Caches on. Stub/kernel can be anywhere.
2) Stub runs and relocates kernel to the desired runtime location
but continues to execute from wherever UEFI loaded it until just
after ExitBootServices().
3) After ExitBootServices, efi_entry() returns relocated entry point
for kernel to efi_stub_entry() in efi-entry.S where the Dcache and
MMU are turned off, the __flush_dcache_all is called, then the
code jumps to the kernel proper entry point.
It isn't clear to me if UEFI does cache flushing at ExitBootServices
time, but even so, at least stack use will get cached between then and
the kernel entry point. The stub could conceivably get its hands on the
EFI memmap and invalidate dcache using address ranges from UEFI memory
descriptors so maybe that is the way we should do it.
--
To unsubscribe from this list: send the line "unsubscribe linux-efi" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
