Date:   Fri, 19 Oct 2018 14:18:48 +0200
From:   Jessica Yu <jeyu@...nel.org>
To:     Miroslav Benes <mbenes@...e.cz>
Cc:     Torsten Duwe <duwe@....de>, Will Deacon <will.deacon@....com>,
        Catalin Marinas <catalin.marinas@....com>,
        Julien Thierry <julien.thierry@....com>,
        Steven Rostedt <rostedt@...dmis.org>,
        Josh Poimboeuf <jpoimboe@...hat.com>,
        Ingo Molnar <mingo@...hat.com>,
        Ard Biesheuvel <ard.biesheuvel@...aro.org>,
        Arnd Bergmann <arnd@...db.de>,
        AKASHI Takahiro <takahiro.akashi@...aro.org>,
        linux-arm-kernel@...ts.infradead.org, linux-kernel@...r.kernel.org,
        live-patching@...r.kernel.org
Subject: Re: [PATCH v3 3/4] arm64: implement live patching

+++ Miroslav Benes [19/10/18 13:59 +0200]:
>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.

Ah yeah, I missed that :) Yep, that would be where it'd go.

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

I do not believe patching of __init functions is supported (right?) So
we do not need to keep arch.init.plt alive post-module-load.

>> 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.

I think I added this line for consistency, but I actually don't think
it is needed. We only would need to keep the section index for
arch.core.plt then.

>> +		} 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;

Ah yeah, I'll clean up the error handling. I'll re-submit this as a
real patch after cleaning it up a bit and it could be added to this
series after review.

>> Thoughts? Does the fix make sense?
>
>I think so.

Thanks Miroslav for the impromptu review :-)

Jessica

Powered by blists - more mailing lists