[PATCH] init: fix allocated page overlapping with PTR_ERR

[Joel Granados]

On Tue, Apr 30, 2024 at 10:37:59AM +0200, Alexandre Ghiti wrote:
> Hi Joel, Nam,
> 
> On 30/04/2024 09:31, Nam Cao wrote:
> > On Mon, Apr 29, 2024 at 02:52:30PM +0200, Joel Granados wrote:
> >> On Thu, Apr 18, 2024 at 12:29:43PM +0200, Nam Cao wrote:
> >>> There is nothing preventing kernel memory allocators from allocating a
> >>> page that overlaps with PTR_ERR(), except for architecture-specific
> >>> code that setup memblock.
> >>>
> >>> It was discovered that RISCV architecture doesn't setup memblock
> >>> corectly, leading to a page overlapping with PTR_ERR() being allocated,
> >>> and subsequently crashing the kernel (link in Close: )
> >>>
> >>> The reported crash has nothing to do with PTR_ERR(): the last page
> >>> (at address 0xfffff000) being allocated leads to an unexpected
> >>> arithmetic overflow in ext4; but still, this page shouldn't be
> >>> allocated in the first place.
> >>>
> >>> Because PTR_ERR() is an architecture-independent thing, we shouldn't
> >>> ask every single architecture to set this up. There may be other
> >>> architectures beside RISCV that have the same problem.
> >>>
> >>> Fix this one and for all by reserving the physical memory page that
> >>> may be mapped to the last virtual memory page as part of low memory.
> >>>
> >>> Unfortunately, this means if there is actual memory at this reserved
> >>> location, that memory will become inaccessible. However, if this page
> >>> is not reserved, it can only be accessed as high memory, so this
> >>> doesn't matter if high memory is not supported. Even if high memory is
> >>> supported, it is still only one page.
> >>>
> >>> Closes: https://lore.kernel.org/linux-riscv/878r1ibpdn.fsf@all.your.base.are.belong.to.us
> >>> Signed-off-by: Nam Cao <namcao at linutronix.de>
> >>> Cc: <stable at vger.kernel.org> # all versions
> >>> ---
> >>>   init/main.c | 1 +
> >>>   1 file changed, 1 insertion(+)
> >>>
> >>> diff --git a/init/main.c b/init/main.c
> >>> index 881f6230ee59..f8d2793c4641 100644
> >>> --- a/init/main.c
> >>> +++ b/init/main.c
> >>> @@ -900,6 +900,7 @@ void start_kernel(void)
> >>>   	page_address_init();
> >>>   	pr_notice("%s", linux_banner);
> >>>   	early_security_init();
> >>> +	memblock_reserve(__pa(-PAGE_SIZE), PAGE_SIZE); /* reserve last page for ERR_PTR */
> >>>   	setup_arch(&command_line);
> >>>   	setup_boot_config();
> >>>   	setup_command_line(command_line);
> >>> -- 
> >>> 2.39.2
> >>>
> >> I received a similar(ish) report recently
> >> https://lore.kernel.org/oe-kbuild-all/202404211031.J6l2AfJk-lkp@intel.com/
> >> regarding RISC-V in init/mail.c. Here is the meat of the report in case
> >> you want to avoid going to the actual link:
> > This issue doesn't look like it has anything to do with this patch: this
> > patch is about overlapping of dynamically allocated memory, while I think
> > the issue is about overlapping sections during linking (maybe something
> > wrong with riscv linker script?)
> >
> > Also, FWIW, this patch is not going to be in mainline because of a
> > regression.
> >
> > Nonetheless, I will have a look at this later.
> 
> 
> The config shows that it is a XIP kernel that comes with its own 
> limitations (text is limited to 32MB for example), so I'm not surprised 
> to see those overlaps.
> 
> We already discussed the removal of randconfig builds on XIP configs, 
> but IIRC it is not possible.

I just tested this going back until "2023-09-20 602bf1830798 (HEAD)
Merge branch 'for-6.7' into for-next  [Petr Mladek]" and I still saw the
overlapping errors.

Is this just something that happens always?

Best
-- 

Joel Granados
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 659 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-riscv/attachments/20240430/df469dca/attachment.sig>