From: Arnd Bergmann <arnd@arndb.de>

The definition of get_dump_page() is guarded by CONFIG_ELF_CORE, but
the caller has now moved into a function that is built based on
CONFIG_COREDUMP, which leads to a possible link failure:

ld.lld-21: error: undefined symbol: get_dump_page
>>> referenced by coredump.c
>>>               fs/coredump.o:(dump_vma_snapshot) in archive vmlinux.a

Change the #ifdef block around the definition to match the caller.
In practice there is very little difference, as setting COREDUMP
but not ELF_CORE is not useful.

Fixes: ff41385709f0 ("coredump: Only sort VMAs when truncating or core_sort_vma sysctl is set")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
---
 mm/gup.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/mm/gup.c b/mm/gup.c
index d846c0ce41d6..15d6d7b5df1d 100644
--- a/mm/gup.c
+++ b/mm/gup.c
@@ -2266,7 +2266,7 @@ EXPORT_SYMBOL(fault_in_readable);
  *
  * Called without mmap_lock (takes and releases the mmap_lock by itself).
  */
-#ifdef CONFIG_ELF_CORE
+#ifdef CONFIG_COREDUMP
 struct page *get_dump_page(unsigned long addr, int *locked)
 {
 	struct page *page;
@@ -2276,7 +2276,7 @@ struct page *get_dump_page(unsigned long addr, int *locked)
 				      FOLL_FORCE | FOLL_DUMP | FOLL_GET);
 	return (ret == 1) ? page : NULL;
 }
-#endif /* CONFIG_ELF_CORE */
+#endif /* CONFIG_COREDUMP */
 
 #ifdef CONFIG_MIGRATION
 
-- 
2.39.5

On Feb 24, 2025, at 7:12 AM, Arnd Bergmann <arnd@kernel.org> wrote:

> The definition of get_dump_page() is guarded by CONFIG_ELF_CORE, but
> the caller has now moved into a function that is built based on
> CONFIG_COREDUMP, which leads to a possible link failure:
> 
> ld.lld-21: error: undefined symbol: get_dump_page
>>>> referenced by coredump.c
>>>>              fs/coredump.o:(dump_vma_snapshot) in archive vmlinux.a

Kernel test robot reported this issue yesterday here:
https://lore.kernel.org/oe-kbuild-all/202502231914.ROOVWMZN-lkp@intel.com/
https://lore.kernel.org/oe-kbuild-all/202502231818.fP8cuxmf-lkp@intel.com/

> Change the #ifdef block around the definition to match the caller.
> In practice there is very little difference, as setting COREDUMP
> but not ELF_CORE is not useful.

Good catch! I wonder if we should guard the sparse dump size calculation
logic from the get_dump_page() caller behind CONFIG_ELF_CORE instead
though. I guess the question becomes, will we ever have a non-ELF core
dump format that won't skip zero pages?

Anyway, I'm fine with this fix as-is.

On Mon, Feb 24, 2025, at 19:42, Brian Mak wrote:
> On Feb 24, 2025, at 7:12 AM, Arnd Bergmann <arnd@kernel.org> wrote:
>
>> The definition of get_dump_page() is guarded by CONFIG_ELF_CORE, but
>> the caller has now moved into a function that is built based on
>> CONFIG_COREDUMP, which leads to a possible link failure:
>> 
>> ld.lld-21: error: undefined symbol: get_dump_page
>>>>> referenced by coredump.c
>>>>>              fs/coredump.o:(dump_vma_snapshot) in archive vmlinux.a
>
> Kernel test robot reported this issue yesterday here:
> https://lore.kernel.org/oe-kbuild-all/202502231914.ROOVWMZN-lkp@intel.com/
> https://lore.kernel.org/oe-kbuild-all/202502231818.fP8cuxmf-lkp@intel.com/
>
>> Change the #ifdef block around the definition to match the caller.
>> In practice there is very little difference, as setting COREDUMP
>> but not ELF_CORE is not useful.
>
> Good catch! I wonder if we should guard the sparse dump size calculation
> logic from the get_dump_page() caller behind CONFIG_ELF_CORE instead
> though. I guess the question becomes, will we ever have a non-ELF core
> dump format that won't skip zero pages?
>
> Anyway, I'm fine with this fix as-is.

The only other code that is guarded by CONFIG_COREDUMP now
is the powerpc spufs, but that is only used with ELF_CORE as
well.

The last non-ELF coredump support was removed in 2022, earlier
versions had some form of coredump for flat, a.out and som
binaries.

We can probably just combine ELF_CORE and COREDUMP into a
single Kconfig symbol.

      Arnd

On Mon, Feb 24, 2025 at 04:12:21PM +0100, Arnd Bergmann wrote:
> From: Arnd Bergmann <arnd@arndb.de>
> 
> The definition of get_dump_page() is guarded by CONFIG_ELF_CORE, but
> the caller has now moved into a function that is built based on
> CONFIG_COREDUMP, which leads to a possible link failure:
> 
> ld.lld-21: error: undefined symbol: get_dump_page
> >>> referenced by coredump.c
> >>>               fs/coredump.o:(dump_vma_snapshot) in archive vmlinux.a
> 
> Change the #ifdef block around the definition to match the caller.
> In practice there is very little difference, as setting COREDUMP
> but not ELF_CORE is not useful.
> 
> Fixes: ff41385709f0 ("coredump: Only sort VMAs when truncating or core_sort_vma sysctl is set")
> Signed-off-by: Arnd Bergmann <arnd@arndb.de>

Thanks for this! I think I'm going to split up ff41385709f0 -- between
this #ifdef and the recent prototype changes, it's going to be not a
great backport, so I'll pick this up too. I expect I'll have a simple
sysctl fix for -rc5 and then the rest in -next.

-Kees

> ---
>  mm/gup.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/mm/gup.c b/mm/gup.c
> index d846c0ce41d6..15d6d7b5df1d 100644
> --- a/mm/gup.c
> +++ b/mm/gup.c
> @@ -2266,7 +2266,7 @@ EXPORT_SYMBOL(fault_in_readable);
>   *
>   * Called without mmap_lock (takes and releases the mmap_lock by itself).
>   */
> -#ifdef CONFIG_ELF_CORE
> +#ifdef CONFIG_COREDUMP
>  struct page *get_dump_page(unsigned long addr, int *locked)
>  {
>  	struct page *page;
> @@ -2276,7 +2276,7 @@ struct page *get_dump_page(unsigned long addr, int *locked)
>  				      FOLL_FORCE | FOLL_DUMP | FOLL_GET);
>  	return (ret == 1) ? page : NULL;
>  }
> -#endif /* CONFIG_ELF_CORE */
> +#endif /* CONFIG_COREDUMP */
>  
>  #ifdef CONFIG_MIGRATION
>  
> -- 
> 2.39.5
> 

-- 
Kees Cook