An investigation of a "Trying to vfree() nonexistent vm area" bug
occurring in arch_kimage_file_post_load_cleanup() doing a
vfree(image->elf_headers) in our 5.14-based kernel yielded the following
double vfree() scenario, also present in mainline:
SYSCALL_DEFINE5(kexec_file_load)
kimage_file_alloc_init()
kimage_file_prepare_segments()
arch_kexec_kernel_image_probe()
kexec_image_load_default()
kexec_bzImage64_ops.load()
bzImage64_load()
crash_load_segments()
prepare_elf_headers(image, &kbuf.buffer, &kbuf.bufsz);
image->elf_headers = kbuf.buffer;
ret = kexec_add_buffer(&kbuf);
if (ret) vfree((void *)image->elf_headers); // first vfree()
if (ret) kimage_file_post_load_cleanup()
vfree(image->elf_headers); // second vfree()
AFAICS the scenario is possible since v5.19 commit b3e34a47f989
("x86/kexec: fix memory leak of elf header buffer") that was marked for
stable and also was backported to our kernel.
Fix the problem by setting the pointer to NULL after the first vfree().
Also set elf_headers_sz to 0, as kimage_file_post_load_cleanup() does.
Fixes: b3e34a47f989 ("x86/kexec: fix memory leak of elf header buffer")
Signed-off-by: Vlastimil Babka <[email protected]>
Cc: Baoquan He <[email protected]>
Cc: Dave Young <[email protected]>
Cc: <[email protected]>
---
arch/x86/kernel/crash.c | 2 ++
1 file changed, 2 insertions(+)
diff --git a/arch/x86/kernel/crash.c b/arch/x86/kernel/crash.c
index 9730c88530fc..0d651c05a49e 100644
--- a/arch/x86/kernel/crash.c
+++ b/arch/x86/kernel/crash.c
@@ -403,6 +403,8 @@ int crash_load_segments(struct kimage *image)
ret = kexec_add_buffer(&kbuf);
if (ret) {
vfree((void *)image->elf_headers);
+ image->elf_headers = NULL;
+ image->elf_headers_sz = 0;
return ret;
}
image->elf_load_addr = kbuf.mem;
--
2.39.0
On 1/2/23 11:39, Vlastimil Babka wrote:
> An investigation of a "Trying to vfree() nonexistent vm area" bug
> occurring in arch_kimage_file_post_load_cleanup() doing a
> vfree(image->elf_headers) in our 5.14-based kernel yielded the following
> double vfree() scenario, also present in mainline:
>
> SYSCALL_DEFINE5(kexec_file_load)
> kimage_file_alloc_init()
> kimage_file_prepare_segments()
> arch_kexec_kernel_image_probe()
> kexec_image_load_default()
> kexec_bzImage64_ops.load()
> bzImage64_load()
> crash_load_segments()
> prepare_elf_headers(image, &kbuf.buffer, &kbuf.bufsz);
> image->elf_headers = kbuf.buffer;
> ret = kexec_add_buffer(&kbuf);
> if (ret) vfree((void *)image->elf_headers); // first vfree()
> if (ret) kimage_file_post_load_cleanup()
> vfree(image->elf_headers); // second vfree()
>
> AFAICS the scenario is possible since v5.19 commit b3e34a47f989
> ("x86/kexec: fix memory leak of elf header buffer") that was marked for
> stable and also was backported to our kernel.
>
> Fix the problem by setting the pointer to NULL after the first vfree().
> Also set elf_headers_sz to 0, as kimage_file_post_load_cleanup() does.
>
> Fixes: b3e34a47f989 ("x86/kexec: fix memory leak of elf header buffer")
> Signed-off-by: Vlastimil Babka <[email protected]>
> Cc: Baoquan He <[email protected]>
> Cc: Dave Young <[email protected]>
> Cc: <[email protected]>
Takashi told me he sent a slightly different fix already in November:
https://lore.kernel.org/all/[email protected]/
Seems it wasn't picked up? You might pick his then, as Baoquan acked it, and
it's removing code, not adding it.
> ---
> arch/x86/kernel/crash.c | 2 ++
> 1 file changed, 2 insertions(+)
>
> diff --git a/arch/x86/kernel/crash.c b/arch/x86/kernel/crash.c
> index 9730c88530fc..0d651c05a49e 100644
> --- a/arch/x86/kernel/crash.c
> +++ b/arch/x86/kernel/crash.c
> @@ -403,6 +403,8 @@ int crash_load_segments(struct kimage *image)
> ret = kexec_add_buffer(&kbuf);
> if (ret) {
> vfree((void *)image->elf_headers);
> + image->elf_headers = NULL;
> + image->elf_headers_sz = 0;
> return ret;
> }
> image->elf_load_addr = kbuf.mem;
* Vlastimil Babka <[email protected]> wrote:
> On 1/2/23 11:39, Vlastimil Babka wrote:
> > An investigation of a "Trying to vfree() nonexistent vm area" bug
> > occurring in arch_kimage_file_post_load_cleanup() doing a
> > vfree(image->elf_headers) in our 5.14-based kernel yielded the following
> > double vfree() scenario, also present in mainline:
> >
> > SYSCALL_DEFINE5(kexec_file_load)
> > kimage_file_alloc_init()
> > kimage_file_prepare_segments()
> > arch_kexec_kernel_image_probe()
> > kexec_image_load_default()
> > kexec_bzImage64_ops.load()
> > bzImage64_load()
> > crash_load_segments()
> > prepare_elf_headers(image, &kbuf.buffer, &kbuf.bufsz);
> > image->elf_headers = kbuf.buffer;
> > ret = kexec_add_buffer(&kbuf);
> > if (ret) vfree((void *)image->elf_headers); // first vfree()
> > if (ret) kimage_file_post_load_cleanup()
> > vfree(image->elf_headers); // second vfree()
> >
> > AFAICS the scenario is possible since v5.19 commit b3e34a47f989
> > ("x86/kexec: fix memory leak of elf header buffer") that was marked for
> > stable and also was backported to our kernel.
> >
> > Fix the problem by setting the pointer to NULL after the first vfree().
> > Also set elf_headers_sz to 0, as kimage_file_post_load_cleanup() does.
> >
> > Fixes: b3e34a47f989 ("x86/kexec: fix memory leak of elf header buffer")
> > Signed-off-by: Vlastimil Babka <[email protected]>
> > Cc: Baoquan He <[email protected]>
> > Cc: Dave Young <[email protected]>
> > Cc: <[email protected]>
>
> Takashi told me he sent a slightly different fix already in November:
> https://lore.kernel.org/all/[email protected]/
>
> Seems it wasn't picked up? You might pick his then, as Baoquan acked it, and
> it's removing code, not adding it.
Thanks, indeed we missed that fix - Boris picked up that version in
tip:x86/urgent, via:
d00dd2f2645d ("x86/kexec: Fix double-free of elf header buffer")
Ingo