Message-ID: <20190820073317.GA23965@linux-8mug>
Date:   Tue, 20 Aug 2019 07:33:48 +0000
From:   Chester Lin <clin@...e.com>
To:     Ard Biesheuvel <ard.biesheuvel@...aro.org>
CC:     Chester Lin <clin@...e.com>,
        "guillaume.gardet@....com" <guillaume.gardet@....com>,
        "linux@...linux.org.uk" <linux@...linux.org.uk>,
        "ren_guo@...ky.com" <ren_guo@...ky.com>,
        "mingo@...nel.org" <mingo@...nel.org>,
        "akpm@...ux-foundation.org" <akpm@...ux-foundation.org>,
        "geert@...ux-m68k.org" <geert@...ux-m68k.org>,
        Mike Rapoport <rppt@...ux.ibm.com>,
        "linux-arm-kernel@...ts.infradead.org" 
        <linux-arm-kernel@...ts.infradead.org>, Gary Lin <GLin@...e.com>,
        "Juergen Gross" <JGross@...e.com>, Joey Lee <JLee@...e.com>,
        "linux-efi@...r.kernel.org" <linux-efi@...r.kernel.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH] efi/arm: fix allocation failure when reserving the kernel
 base

On Mon, Aug 19, 2019 at 05:56:51PM +0300, Ard Biesheuvel wrote:
> On Mon, 19 Aug 2019 at 11:01, Chester Lin <clin@...e.com> wrote:
> >
> > Hi Mike and Ard,
> >
> > On Thu, Aug 15, 2019 at 04:37:39PM +0300, Mike Rapoport wrote:
> > > On Thu, Aug 15, 2019 at 02:32:50PM +0300, Ard Biesheuvel wrote:
> > > > (adding Mike)
> > > >
> > > > On Thu, 15 Aug 2019 at 14:28, Chester Lin <clin@...e.com> wrote:
> > > > >
> > > > > Hi Ard,
> > > > >
> > > > > On Thu, Aug 15, 2019 at 10:59:43AM +0300, Ard Biesheuvel wrote:
> > > > > > On Sun, 4 Aug 2019 at 10:57, Ard Biesheuvel <ard.biesheuvel@...aro.org> wrote:
> > > > > > >
> > > > > > > Hello Chester,
> > > > > > >
> > > > > > > On Fri, 2 Aug 2019 at 08:40, Chester Lin <clin@...e.com> wrote:
> > > > > > > >
> > > > > > > > In some cases the arm32 efistub could fail to allocate memory for
> > > > > > > > uncompressed kernel. For example, we got the following error message when
> > > > > > > > verifying EFI stub on Raspberry Pi-2 [kernel-5.2.1 + grub-2.04] :
> > > > > > > >
> > > > > > > >   EFI stub: Booting Linux Kernel...
> > > > > > > >   EFI stub: ERROR: Unable to allocate memory for uncompressed kernel.
> > > > > > > >   EFI stub: ERROR: Failed to relocate kernel
> > > > > > > >
> > > > > > > > After checking the EFI memory map we found that the first page [0 - 0xfff]
> > > > > > > > had been reserved by Raspberry Pi-2's firmware, and the efistub tried to
> > > > > > > > set the dram base at 0, which was actually in a reserved region.
> > > > > > > >
> > > > > > >
> > > > > > > This by itself is a violation of the Linux boot protocol for 32-bit
> > > > > > > ARM when using the decompressor. The decompressor rounds down its own
> > > > > > > base address to a multiple of 128 MB, and assumes the whole area is
> > > > > > > available for the decompressed kernel and related data structures.
> > > > > > > (The first TEXT_OFFSET bytes are no longer used in practice, which is
> > > > > > > why putting a reserved region of 4 KB bytes works at the moment, but
> > > > > > > this is fragile). Note that the decompressor does not look at any DT
> > > > > > > or EFI provided memory maps *at all*.
> > > > > > >
> > > > > > > So unfortunately, this is not something we can fix in the kernel, but
> > > > > > > we should fix it in the bootloader or in GRUB, so it does not put any
> > > > > > > reserved regions in the first 128 MB of memory,
> > > > > > >
> > > > > >
> > > > > > OK, perhaps we can fix this by taking TEXT_OFFSET into account. The
> > > > > > ARM boot protocol docs are unclear about whether this memory should be
> > > > > > used or not, but it is no longer used for its original purpose (page
> > > > > > tables), and the RPi loader already keeps data there.
> > > > > >
> > > > > > Can you check whether the following patch works for you?
> > > > > >
> > > > > > diff --git a/drivers/firmware/efi/libstub/Makefile
> > > > > > b/drivers/firmware/efi/libstub/Makefile
> > > > > > index 0460c7581220..ee0661ddb25b 100644
> > > > > > --- a/drivers/firmware/efi/libstub/Makefile
> > > > > > +++ b/drivers/firmware/efi/libstub/Makefile
> > > > > > @@ -52,6 +52,7 @@ lib-$(CONFIG_EFI_ARMSTUB)     += arm-stub.o fdt.o
> > > > > > string.o random.o \
> > > > > >
> > > > > >  lib-$(CONFIG_ARM)              += arm32-stub.o
> > > > > >  lib-$(CONFIG_ARM64)            += arm64-stub.o
> > > > > > +CFLAGS_arm32-stub.o            := -DTEXT_OFFSET=$(TEXT_OFFSET)
> > > > > >  CFLAGS_arm64-stub.o            := -DTEXT_OFFSET=$(TEXT_OFFSET)
> > > > > >
> > > > > >  #
> > > > > > diff --git a/drivers/firmware/efi/libstub/arm32-stub.c
> > > > > > b/drivers/firmware/efi/libstub/arm32-stub.c
> > > > > > index e8f7aefb6813..66ff0c8ec269 100644
> > > > > > --- a/drivers/firmware/efi/libstub/arm32-stub.c
> > > > > > +++ b/drivers/firmware/efi/libstub/arm32-stub.c
> > > > > > @@ -204,7 +204,7 @@ efi_status_t
> > > > > > handle_kernel_image(efi_system_table_t *sys_table,
> > > > > >          * loaded. These assumptions are made by the decompressor,
> > > > > >          * before any memory map is available.
> > > > > >          */
> > > > > > -       dram_base = round_up(dram_base, SZ_128M);
> > > > > > +       dram_base = round_up(dram_base, SZ_128M) + TEXT_OFFSET;
> > > > > >
> > > > > >         status = reserve_kernel_base(sys_table, dram_base, reserve_addr,
> > > > > >                                      reserve_size);
> > > > > >
> > > > >
> > > > > I tried your patch on rpi2 and got the following panic. Just a reminder that I
> > > > > have replaced some log messages with "......" since it might be too long to
> > > > > post all.
> > > > >
> > > >
> > > > OK. Good to know that this change helps you to get past the EFI stub boot issue.
> > > >
> > > > > In this case the kernel failed to reserve cma, which should hit the issue of
> > > > > memblock_limit=0x1000 as I had mentioned in my patch description. The first
> > > > > block [0-0xfff] was scanned in adjust_lowmem_bounds(), but it did not align
> > > > > with PMD_SIZE so the cma reservation failed because the memblock.current_limit
> > > > > was extremely low. That's why I expand the first reservation from 1 PAGESIZE to
> > > > > 1 PMD_SIZE in my patch in order to avoid this issue. Please kindly let me know
> > > > > if any suggestion, thank you.
> > >
> > >
> > > > This looks like it is a separate issue. The memblock/cma code should
> > > > not choke on a reserved page of memory at 0x0.
> > > >
> > > > Perhaps Russell or Mike (cc'ed) have an idea how to address this?
> > >
> > > Presuming that the last memblock dump comes from the end of
> > > arm_memblock_init() with the this memory map
> > >
> > > memory[0x0] [0x0000000000000000-0x0000000000000fff], 0x0000000000001000 bytes flags: 0x4
> > > memory[0x1] [0x0000000000001000-0x0000000007ef5fff], 0x0000000007ef5000 bytes flags: 0x0
> > > memory[0x2] [0x0000000007ef6000-0x0000000007f09fff], 0x0000000000014000 bytes flags: 0x4
> > > memory[0x3] [0x0000000007f0a000-0x000000003cb3efff], 0x0000000034c35000 bytes flags: 0x0
> > >
> > > adjust_lowmem_bounds() will set the memblock_limit (and respectively global
> > > memblock.current_limit) to 0x1000 and any further memblock_alloc*() will
> > > happily fail.
> > >
> > > I believe that the assumption for memblock_limit calculations was that the
> > > first bank has several megs at least.
> > >
> > > I wonder if this hack would help:
> > >
> > > diff --git a/arch/arm/mm/mmu.c b/arch/arm/mm/mmu.c
> > > index d9a0038..948e5b9 100644
> > > --- a/arch/arm/mm/mmu.c
> > > +++ b/arch/arm/mm/mmu.c
> > > @@ -1206,7 +1206,7 @@ void __init adjust_lowmem_bounds(void)
> > >                        * allocated when mapping the start of bank 0, which
> > >                        * occurs before any free memory is mapped.
> > >                        */
> > > -                     if (!memblock_limit) {
> > > +                     if (memblock_limit < PMD_SIZE) {
> > >                               if (!IS_ALIGNED(block_start, PMD_SIZE))
> > >                                       memblock_limit = block_start;
> > >                               else if (!IS_ALIGNED(block_end, PMD_SIZE))
> > >
> >
> > I applied this patch as well and it works well on rpi-2 model B.
> >
> 
> Thanks, Chester, that is good to know.
> 
> However, afaict, this only affects systems where physical memory
> starts at address 0x0, so I think we need a better fix.
> 
> I know Mike has been looking into the NOMAP stuff lately, and your
> original patch contains a hunk that makes this code (?) disregard
> nomap memblocks. That might be a better approach.
>
Hi Ard and Mike,

In my original patch, I studied map_lowmem() and found that some blocks might
not be mapped according to the current memory map. Thus I assumed maybe NOMAP
blocks could still be ignored in adjust_lowmem_bounds() since they would not
be allocated afterward. But that change in mmu.c still depends on a condition
that there should be a PMD_SIZE block or consecutive smaller NOMAP blocks which
exacly fit the PM_SIZE alignment at the beginning of memory map otherwise the
memblock_limit could still fall on a very low address. That's why I tried to
allocate pages again in arm32-stub.c in order to fill the gap between the
unaligned block_start and the PMD_SIZE aligned kernel base.

Please feel free to let me know if any idea and I am willing to help with
verification.

Powered by blists - more mailing lists