Re: [PATCH v3 3/4] arm64: implement live patching

[Miroslav Benes]

On Thu, 18 Oct 2018, Jessica Yu wrote:

> +++ Miroslav Benes [17/10/18 15:39 +0200]:
> >On Mon, 1 Oct 2018, Torsten Duwe wrote:
> >
> >Ad relocations. I checked that everything in struct mod_arch_specific
> >stays after the module is load. Both core and init get SHF_ALLOC set
> >(mod->arch.core.plt->sh_flags in module_frob_arch_sections(). It is
> >important because apply_relocate_add() may use those sections
> >through module_emit_plt_entry() call.
> 
> Yes, it looks like the needed .plt sections will remain in module
> memory. However, I think I found a slight issue... :/
> 
> In module_frob_arch_sections(), mod->arch.core.plt is set to an offset
> within info->sechdrs:
> 
>             if (!strcmp(secstrings + sechdrs[i].sh_name, ".plt"))
>                     mod->arch.core.plt = sechdrs + i;
> 
> sechdrs is from info->sechdrs, which is freed at the end of
> load_module() via free_copy(). So although the relevant plt section(s)
> are in module memory, mod->arch.core.plt will point to invalid memory
> after info is freed. In other words, the section (.plt) *is* in memory
> but the section header (struct elf64_shdr) containing section metadata
> like sh_addr, sh_size etc., is not.
> 
> But we have mod->klp_info->sechdrs (which is a copy of the section
> headers) for this very reason. It is initialized only at the very end
> of load_module(). I don't think copy_module_elf() is dependent on
> anything, so it can just be moved earlier.
> 
> Note that we cannot set mod->arch.core.plt to mod->klp_info->sechdrs + i
> during module_frob_arch_sections() because it is still too early, none
> of the module sections have been copied and none of their sh_addr's
> have been set to their final locations as this is all happening before
> move_module() is called. So we can use a module_finalize() function
> for arm64 to switch the mod->arch.core.plt to use the klp_info section
> headers.

Thanks for the analysis. You seem to be right.
 
> Maybe this will be more clear in the example fix below (completely
> untested and unreviewed!):
> 
> diff --git a/arch/arm64/include/asm/module.h b/arch/arm64/include/asm/module.h
> index 97d0ef12e2ff..150afc29171b 100644
> --- a/arch/arm64/include/asm/module.h
> +++ b/arch/arm64/include/asm/module.h
> @@ -25,6 +25,7 @@ struct mod_plt_sec {
> 	struct elf64_shdr	*plt;
> 	int			plt_num_entries;
> 	int			plt_max_entries;
> +	int			plt_shndx;
> };
> 
> struct mod_arch_specific {
> diff --git a/arch/arm64/kernel/module-plts.c b/arch/arm64/kernel/module-plts.c
> index f0690c2ca3e0..c23cef8f0165 100644
> --- a/arch/arm64/kernel/module-plts.c
> +++ b/arch/arm64/kernel/module-plts.c
> @@ -196,6 +196,21 @@ static unsigned int count_plts(Elf64_Sym *syms,
> Elf64_Rela *rela, int num,
> 	return ret;
> }
> 
> +int module_finalize(const Elf_Ehdr *hdr, const Elf_Shdr *sechdrs,
> +		    struct module *mod)
> +{
> +	/*
> +	 * Livepatch modules need to keep access to section headers to apply
> +	 * relocations. Note mod->klp_info should have already been
> initialized
> +	 * and all section sh_addr's should have their final addresses by the
> +	 * time module_finalize() is called.
> +	 */
> +	if (is_livepatch_module(mod))
> +		mod->arch.core.plt = mod->klp_info->sechdrs +
> mod->arch.core.plt_shndx;
> +
> +	return 0;
> +}

There is already module_finalize() in arch/arm64/kernel/module.c, so this 
should probably go there.

If I am not mistaken, we do not care for arch.init.plt in livepatch. Is 
that correct?

> int module_frob_arch_sections(Elf_Ehdr *ehdr, Elf_Shdr *sechdrs,
> 			      char *secstrings, struct module *mod)
> {
> @@ -210,11 +225,13 @@ int module_frob_arch_sections(Elf_Ehdr *ehdr, Elf_Shdr
> *sechdrs,
> 	 * entries. Record the symtab address as well.
> 	 */
> 	for (i = 0; i < ehdr->e_shnum; i++) {
> -		if (!strcmp(secstrings + sechdrs[i].sh_name, ".plt"))
> +		if (!strcmp(secstrings + sechdrs[i].sh_name, ".plt")) {
> 			mod->arch.core.plt = sechdrs + i;
> -		else if (!strcmp(secstrings + sechdrs[i].sh_name,
> ".init.plt"))
> +			mod->arch.core.plt_shndx = i;
> +		} else if (!strcmp(secstrings + sechdrs[i].sh_name,
> ".init.plt")) {
> 			mod->arch.init.plt = sechdrs + i;
> -		else if (IS_ENABLED(CONFIG_DYNAMIC_FTRACE) &&
> +			mod->arch.init.plt_shndx = i;

It is initialized here, but that's not necessarily a bad thing.

> +		} else if (IS_ENABLED(CONFIG_DYNAMIC_FTRACE) &&
> 			 !strcmp(secstrings + sechdrs[i].sh_name,
> 				 ".text.ftrace_trampoline"))
> 			tramp = sechdrs + i;
> diff --git a/kernel/module.c b/kernel/module.c
> index 6746c85511fe..2fc4d74288dd 100644
> --- a/kernel/module.c
> +++ b/kernel/module.c
> @@ -3363,6 +3363,12 @@ static int post_relocation(struct module *mod, const
> struct load_info *info)
> 	/* Setup kallsyms-specific fields. */
> 	add_kallsyms(mod, info);
> 
> +	if (is_livepatch_module(mod)) {
> +		err = copy_module_elf(mod, info);
> +		if (err < 0)
> +			return err;
> +	}
> +
> 	/* Arch-specific module finalizing. */
> 	return module_finalize(info->hdr, info->sechdrs, mod);
> }
> @@ -3775,12 +3781,6 @@ static int load_module(struct load_info *info, const
> char __user *uargs,
> 	if (err < 0)
> 		goto coming_cleanup;
> 
> -	if (is_livepatch_module(mod)) {
> -		err = copy_module_elf(mod, info);
> -		if (err < 0)
> -			goto sysfs_cleanup;
> -	}
> -
> 	/* Get rid of temporary copy. */
> 	free_copy(info);

Hm, this is hopefully all right. We'd need to change the error handling a 
bit. free_module_elf() is now called in free_module() and it should be 
moved somewhere to load_module(). Probably under free_arch_cleanup label. 
Although it would be much better to rework post_relocation() and handle it 
right there.

Something like

-	return module_finalize(info->hdr, info->sechdrs, mod);
+	err = module_finalize(info->hdr, info->sechdrs, mod);
+	if (err < 0)
+		free_module_elf();
+
+	return err;

> Thoughts? Does the fix make sense?

I think so.

> >The last thing is count_plts() function called from
> >module_frob_arch_sections(). It needed to be changed in 2016 as well. See
> >Jessica's patch
> >(20160713001113.GA30925@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx). The logic
> >in the function has changed since then. If I am not mistaken,
> >count_plts() is fine as it is right now. It does not consider SHN_UNDEF
> >anymore, it looks at the destination section (where a symbol should
> >resolved to) only.
> >
> >Jessica, could you doublecheck, please?
> 
> Yes, I think count_plts() is fine now in this case (because it is
> always true that sym->st_shndx != dstidx for SHN_LIVEPATCH symbols)
> and my old patch from 2016 is no longer needed.

Thanks a lot, Jessica.

Torsten, could you include the outcome to your patch set once we settle on 
it? Thanks.

Miroslav