Re: [PATCH v7 2/4] module: introduce SH_ENTSIZE_STANDALONE for separately allocated sections

From
Hao Ge <hao.ge@linux.dev>
Date
2026-09-07 06:09:25
Reply to
Petr Pavlu, 2026-09-04
Hi Petr

On 2026/9/4 23:25, Petr Pavlu wrote:
On 9/3/26 8:00 AM, Hao Ge wrote:
quoted
Hi Petr and Suren

On 2026/9/2 16:33, sashiko-bot@kernel.org wrote:
quoted
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Medium] The `.data..percpu` section is incorrectly exported in sysfs with a dangling pointer to a freed temporary buffer.
--

commit 5082dc654c1b8850af405eef473b249c0d936de9
Author: Hao Ge [off-list ref]

module: introduce SH_ENTSIZE_STANDALONE for separately allocated sections

This commit introduces a new marker, SH_ENTSIZE_STANDALONE, to indicate
sections with a separate allocation. Instead of clearing the SHF_ALLOC
flag, it utilizes sh_entsize so that sections with separate allocations
like the percpu section can be correctly skipped during layout.

Link: https://lore.kernel.org/all/499bb60c-c6e3-43a3-bd92-95a0567ece5e@suse.com/ (local)
quoted
diff --git a/kernel/module/main.c b/kernel/module/main.c
index d0e1e0bd2ad06..b114b815187a9 100644
--- a/kernel/module/main.c
+++ b/kernel/module/main.c
[ ... ]
quoted
@@ -2822,7 +2810,8 @@ static int move_module(struct module *mod, struct load_info *info)
 		Elf_Shdr *shdr = &info->sechdrs[i];
 		const char *sname;
 
-		if (!(shdr->sh_flags & SHF_ALLOC))
+		if (!(shdr->sh_flags & SHF_ALLOC)
+		    || shdr->sh_entsize == SH_ENTSIZE_STANDALONE)
 			continue;
 
 		sname = info->secstrings + shdr->sh_name;
[ ... ]
quoted
@@ -2967,8 +2957,13 @@ static struct module *layout_and_allocate(struct load_info *info, int flags)
 	if (err < 0)
 		return ERR_PTR(err);
 
+	/* Repurpose sh_entsize to track where each section is allocated. */
+	for (i = 0; i < info->hdr->e_shnum; i++)
+		info->sechdrs[i].sh_entsize = ~0UL;
+
 	/* We will do a special allocation for per-cpu sections later. */
-	info->sechdrs[info->index.pcpu].sh_flags &= ~(unsigned long)SHF_ALLOC;
+	if (info->index.pcpu)
+		info->sechdrs[info->index.pcpu].sh_entsize = SH_ENTSIZE_STANDALONE;
 
 	/*
 	 * Mark relevant sections as SHF_RO_AFTER_INIT so layout_sections() can
[Severity: Medium]
Because SHF_ALLOC is no longer cleared here, does this cause the
.data..percpu section to be incorrectly exported via sysfs?

With SHF_ALLOC remaining intact, sect_empty() in kernel/module/sysfs.c now
returns false for the percpu section, causing it to be processed instead of
skipped.

Additionally, because move_module() explicitly skips sections marked with
SH_ENTSIZE_STANDALONE, the sh_addr is never updated to the final allocated
memory.

This means when add_sect_attrs() exports the section:

kernel/module/sysfs.c:add_sect_attrs() {
    ...
	if (sect_empty(sec))
		continue;
    ...
	sattr->private = (void *)sec->sh_addr;
    ...
}

Will sattr->private end up storing an unmodified sh_addr pointing to the
freed temporary load_info buffer (info->hdr)?

This appears to leak a stale vmalloc pointer address to userspace when
reading the sysfs file for the section.
Yes, Sashiko is correct. The fix is just one line after percpu_modcopy():

info->sechdrs[info->index.pcpu].sh_addr = (unsigned long)mod_percpu(mod); 
mod_percpu() (mod->percpu) isn't really the execution address of the
.data..percpu section. That section in the ELF file is only a template
to initialize the individual per-CPU variable areas.
Yes, I used mod_percpu() to stay consistent with /proc/kallsyms.
I think the two options are:

1) Do not export the section under /sys/module/*/sections/.
2) Treat the boot CPU area as the canonical location of
   .data..percpu. In other words, do:
   info->sechdrs[info->index.pcpu].sh_addr = (unsigned long)per_cpu_ptr(mod->percpu, get_boot_cpu_id());

The /sys/module/*/sections/ interface is intended for debugging. From
that perspective, exposing the .data..percpu address for the boot CPU
may be useful and seems fairly reasonable to me. On the other hand, one
could argue that this can cause confusion, since a loaded module has
multiple .data..percpu instances.

I don't see a strong argument for either option, but I would lean toward
the second.
OK. I spent the morning looking for relevant specs, but found none.
Nor is this documented in Documentation/ABI/testing/sysfs-module.

https://elixir.bootlin.com/linux/v7.2.2/source/Documentation/ABI/testing/sysfs-module
https://elixir.bootlin.com/linux/v7.2.2/source/Documentation/ABI/stable/sysfs-module

But I've noticed similar usage in __is_module_percpu_address,
https://elixir.bootlin.com/linux/v7.2.2/source/kernel/module/main.c#L516

So I will first add documentation for this in Documentation/ABI/testing/sysfs-module.
quoted
But .data..percpu will show up in /sys/module/*/sections/, it was never
exported before. Stable will get this too via backport, since later fixes
depend on it.
I think exposing .data..percpu under /sys/module/*/sections/ is a minor
change and is fine for stable. As mentioned, /sys/module/*/sections/ is
a debugging interface. Additionally, most uses of per-CPU variables are
in the core kernel, with only a handful modules defining per-CPU data.
On my system, 30 modules out of 5383 contain a .data..percpu section.
Agree.

Thanks for your help.

Best Regards
Hao
Thread (18 messages) 18 messages, 4 authors, last reply 2026-09-08
Keyboard shortcuts
hback out one level
jnext message in thread / next series-diff position
kprevious message in thread / previous series-diff position
ldrill in
Escclose help / fold thread tree
?toggle this help