Re: [PATCH v4] x86/kexec: fix potential cmem->ranges out of bounds

[Coiby Xu]

On Wed, May 07, 2025 at 10:59:59PM -0700, Andrew Morton wrote:
> On Thu, 8 May 2025 12:25:15 +0800 Coiby Xu <coxu@xxxxxxxxxx> wrote:
>
> > >
> > >Acked-by: Baoquan He <bhe@xxxxxxxxxx>
> >
> > Hi Andrew,
> >
> > It seems this patch was missed.
>
> January 2024.  Yes, it's fair to assume that it was missed ;)
>
> > Will you pick it up?
>
> Sure.

Thanks for quickly processing this patch! Sorry I didn't reply yesterday
as I was trying to reproduce the UBSAN warning and truly understand the
it.

> > Without this patch,
> > kdump kernel will fail to be loaded by the kexec_file_load,

As already pointed out by Baoquan, a manual test shows kexec_file_load
actually works despite the UBSAN warning. Sorry I misinterpreted the
UBSAN warning and the automated test result failure (somehow sysrq
wasn't be triggered and vmcore wasn't saved either).


> >   [  139.736948] UBSAN: array-index-out-of-bounds in arch/x86/kernel/crash.c:350:25
> >   [  139.742360] index 0 is out of range for type 'range [*]'
[...]
> >
>
> Do we know why this has appeared at such a late date?  The reporter
> must be doing something rare.

The UBSAN warning happens because flexible array members annotated with
__counted_by are accessed without assigning an array element count i.e.
crash_mem->ranges[0] is accessed without setting max_nr_ranges after
vzalloc,

    // include/linux/crash_core.h
    struct crash_mem {
    	unsigned int max_nr_ranges;
    	unsigned int nr_ranges;
    	struct range ranges[] __counted_by(max_nr_ranges);
    };

The bad commit was introduced in 2021 but only recent gcc-15 supports
__counted_by. That's why we don't see this UBSAN warning until this
year. And although this UBSAN warning is scary enough, fortunately it
doesn't cause a real problem. 

> Baoquan, please re-review this?
>
> A -stable backport is clearly required.  A Fixes: would be nice, but I
> assume this goes back a long time so it isn't worth spending a lot of
> time working out when this was introduced.

So I believe the correct fix should be as follows,

--- a/arch/x86/kernel/crash.c
+++ b/arch/x86/kernel/crash.c
@@ -301,6 +301,7 @@ int crash_setup_memmap_entries(struct kimage *image, struct boot_params *params)
        cmem = vzalloc(struct_size(cmem, ranges, 1));
        if (!cmem)
                return -ENOMEM;
+       cmem->max_nr_ranges = 1;
 
        memset(&cmd, 0, sizeof(struct crash_memmap_data));
        cmd.params = params;


And a Fixes tag should be dedicated to commit
5849cdf8c120 ("x86/crash: Fix crash_setup_memmap_entries() out-of-bounds access")
which forgot to set cmem->max_nr_ranges=1.

> The patch needed a bit of work to apply to current code.  I did the
> below.  It compiles.
>
> --- a/arch/x86/kernel/crash.c~x86-kexec-fix-potential-cmem-ranges-out-of-bounds
> +++ a/arch/x86/kernel/crash.c
> @@ -165,8 +165,18 @@ static struct crash_mem *fill_up_crash_e
> 	/*
> 	 * Exclusion of crash region and/or crashk_low_res may cause
> 	 * another range split. So add extra two slots here.
> +	 *
> +	 * Exclusion of low 1M may not cause another range split, because the
> +	 * range of exclude is [0, 1M] and the condition for splitting a new
> +	 * region is that the start, end parameters are both in a certain
> +	 * existing region in cmem and cannot be equal to existing region's
> +	 * start or end. Obviously, the start of [0, 1M] cannot meet this
> +	 * condition.
> +	 *
> +	 * But in order to lest the low 1M could be changed in the future,
> +	 * (e.g. [stare, 1M]), add a extra slot.
> 	 */
> -	nr_ranges += 2;
> +	nr_ranges += 3;
> 	cmem = vzalloc(struct_size(cmem, ranges, nr_ranges));
> 	if (!cmem)
> 		return NULL;
> @@ -317,9 +327,16 @@ int crash_setup_memmap_entries(struct ki
> 	 * split. So use two slots here.
> 	 */
> 	nr_ranges = 2;
> -	cmem = vzalloc(struct_size(cmem, ranges, nr_ranges));
> +	/*
> +	 * In the current x86 architecture code, the elfheader is always
> +	 * allocated at crashk_res.start. But it depends on the allocation
> +	 * position of elfheader in crashk_res. To avoid potential out of
> +	 * bounds in future, add a extra slot.
> +	 */
> +	cmem = vzalloc(struct_size(cmem, ranges, 2));
> 	if (!cmem)
> 		return -ENOMEM;
> +	cmem->max_nr_ranges = 2;

Thanks for coming up with the above patch! I think the goal of this
patch is addressing a different issue but it also fixes the UBSAN
warning because cmem->max_nr_ranges is now set.

> 	cmem->max_nr_ranges = nr_ranges;
> 	cmem->nr_ranges = 0;
> _


--
Best regards,
Coiby