[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <202506270854.A729825@keescook>
Date: Fri, 27 Jun 2025 09:35:45 -0700
From: Kees Cook <kees@...nel.org>
To: YinFengwei <fengwei_yin@...ux.alibaba.com>
Cc: linux-mm@...ck.org, linux-kernel@...r.kernel.org,
zhourundong.zrd@...ux.alibaba.com
Subject: Re: [Question] About the elf program header size
On Fri, Jun 27, 2025 at 09:04:11AM +0800, YinFengwei wrote:
> We had a script generated assembly code. built it with gcc and the
> output elf file had 78 program headers.
Why so many?
> On an arm64 platform, if we have 64KB base page size, the elf can
> be started correctly. But if we have 4KB base page size, the elf
> can NOT be started with:
> cannot execute binary file: Exec format error
>
> Look at the function load_elf_phdrs():
> if (size == 0 || size > 65536 || size > ELF_MIN_ALIGN)
> goto out;
>
> ELF_MIN_ALIGN is defined as PAGE_SIZE on arm64. Which can explain
> above inconsistent behaviors (from user perspetive).
>
> I didn't find the limitation definition in ELF spec(Maybe I missed
> some obvious info there). If I remove "size > ELF_MIN_ALIGN", the
> same elf can be started correctly even with 4KB page size.
>
> So my question is why we limit the who program headers total size
> to PAGE_SIZE? git history couldn't tell anything because the
> limitation was introduced when whole linux kernel tree was migrated
> to git. Is there a possible constrain on other architecture? Thanks.
Looking through
https://git.kernel.org/pub/scm/linux/kernel/git/history/history.git
(which doesn't have linked history, so you have to examine explicit "pre
git" tags), I see:
4779b38bcb96 ("[PATCH] Linux-0.99.13 (September 19, 1993)")
Which says "ELF binary support it a notable change." Here, the PAGE_SIZE
check does not exist. When ELF interp support was added in
9e11983a5a3e ("Import 0.99.15f"), we see the check appear, and I can
find no rationale.
And with 6a8d38945cf4 ("binfmt_elf: Hoist ELF program header loading to
a function"), the PAGE_SIZE check is _added_ for non-interp loads.
It seems the 64K count limit is sufficient? (If the goal was to avoid
large memory allocations happening from userspace, we're way past
PAGE_SIZE these days between IPC, BPF, etc.) Does this work for you?
diff --git a/fs/binfmt_elf.c b/fs/binfmt_elf.c
index a43363d593e5..92de44b8765f 100644
--- a/fs/binfmt_elf.c
+++ b/fs/binfmt_elf.c
@@ -519,7 +519,7 @@ static struct elf_phdr *load_elf_phdrs(const struct elfhdr *elf_ex,
/* Sanity check the number of program headers... */
/* ...and their total size. */
size = sizeof(struct elf_phdr) * elf_ex->e_phnum;
- if (size == 0 || size > 65536 || size > ELF_MIN_ALIGN)
+ if (size == 0 || size > 65536)
goto out;
elf_phdata = kmalloc(size, GFP_KERNEL);
--
Kees Cook
Powered by blists - more mailing lists