Re: livepatch: reuse module loader code to write relocations

From: Jessica Yu
Date: Thu Nov 12 2015 - 14:15:15 EST

+++ Miroslav Benes [12/11/15 16:27 +0100]:
On Wed, 11 Nov 2015, Jessica Yu wrote:

+++ Miroslav Benes [11/11/15 15:30 +0100]:
> On Mon, 9 Nov 2015, Jessica Yu wrote:
>
> So I guess we don't need klp_reloc anymore.

Yes, that's correct. I am noticing just now that I forgot to remove
the klp_reloc struct definition from livepatch.h. That change will be
reflected in v2...

> If true, we should really
> start thinking about proper documentation because there are going to be
> plenty of assumptions about a patch module and we need to have it written
> somewhere. Especially how the relocation sections look like.

Agreed. As a first step the patch module format can perhaps be
documented somewhere. Perhaps it's time we create
Documentation/livepatch/? :-)

Yes, I think so.

> > diff --git a/kernel/livepatch/core.c b/kernel/livepatch/core.c
> > index 087a8c7..26c419f 100644
> > --- a/kernel/livepatch/core.c
> > +++ b/kernel/livepatch/core.c
> > @@ -28,6 +28,8 @@
> > #include <linux/list.h>
> > #include <linux/kallsyms.h>
> > #include <linux/livepatch.h>
> > +#include <linux/elf.h>
> > +#include <asm/cacheflush.h>
> >
> > /**
> > * struct klp_ops - structure for tracking registered ftrace ops structs
> > @@ -281,46 +283,54 @@ static int klp_find_external_symbol(struct module
> > *pmod, const char *name,
> > }
> >
> > static int klp_write_object_relocations(struct module *pmod,
> > - struct klp_object *obj)
> > + struct klp_object *obj,
> > + struct klp_patch *patch)
> > {
> > - int ret;
> > - struct klp_reloc *reloc;
> > + int relindex, num_relas;
> > + int i, ret = 0;
> > + unsigned long addr;
> > + unsigned int bind;
> > + char *symname;
> > + struct klp_reloc_sec *reloc_sec;
> > + struct load_info *info;
> > + Elf_Rela *rela;
> > + Elf_Sym *sym, *symtab;
> > + Elf_Shdr *symsect;
> >
> > if (WARN_ON(!klp_is_object_loaded(obj)))
> > return -EINVAL;
> >
> > - if (WARN_ON(!obj->relocs))
> > - return -EINVAL;
> > -
> > - for (reloc = obj->relocs; reloc->name; reloc++) {
> > - if (!klp_is_module(obj)) {
> > - ret = klp_verify_vmlinux_symbol(reloc->name,
> > - reloc->val);
> > - if (ret)
> > - return ret;
> > - } else {
> > - /* module, reloc->val needs to be discovered */
> > - if (reloc->external)
> > - ret = klp_find_external_symbol(pmod,
> > - reloc->name,
> > - &reloc->val);
> > - else
> > - ret = klp_find_object_symbol(obj->mod->name,
> > - reloc->name,
> > - &reloc->val);
> > - if (ret)
> > - return ret;
> > - }
> > - ret = klp_write_module_reloc(pmod, reloc->type, reloc->loc,
> > - reloc->val + reloc->addend);
> > - if (ret) {
> > - pr_err("relocation failed for symbol '%s' at 0x%016lx
> > (%d)\n",
> > - reloc->name, reloc->val, ret);
> > - return ret;
> > + info = pmod->info;
> > + symsect = info->sechdrs + info->index.sym;
> > + symtab = (void *)info->hdr + symsect->sh_offset;
> > +
> > + /* For each __klp_rela section for this object */
> > + list_for_each_entry(reloc_sec, &obj->reloc_secs, list) {
> > + relindex = reloc_sec->index;
> > + num_relas = info->sechdrs[relindex].sh_size /
> > sizeof(Elf_Rela);
> > + rela = (Elf_Rela *) info->sechdrs[relindex].sh_addr;
> > +
> > + /* For each rela in this __klp_rela section */
> > + for (i = 0; i < num_relas; i++, rela++) {
> > + sym = symtab + ELF_R_SYM(rela->r_info);
> > + symname = info->strtab + sym->st_name;
> > + bind = ELF_ST_BIND(sym->st_info);
> > +
> > + if (sym->st_shndx == SHN_LIVEPATCH) {
> > + if (bind == STB_LIVEPATCH_EXT)
> > + ret = klp_find_external_symbol(pmod,
> > symname, &addr);
> > + else
> > + ret =
> > klp_find_object_symbol(obj->name, symname, &addr);
> > + if (ret)
> > + return ret;
> > + sym->st_value = addr;
> > + }
> > }
> > + ret = apply_relocate_add(info->sechdrs, info->strtab,
> > + info->index.sym, relindex, pmod);
> > }
> >
> > - return 0;
> > + return ret;
> > }
>
> Looking at this... do we even need reloc_secs in klp_object? Question is
> whether we need more than one dynrela section for an object. If not then
> the binding between klp_reloc_sec and an object is the only relevant thing
> in the structure, be it index or objname. So we can replace the
> list of structures with just the index in klp_object, or get rid of it
> completely and rely on the name of dynrela section be something like
> __klp_rela_{objname}.

Hm, you bring up a good point. I think theoretically yes, it is
possible to just have one klp_reloc_sec for each object and therefore
a list is not required (I have not checked yet how difficult it would
be to implement this on the kpatch-build side of things). However,
considering the final format of the patch module, I think it is
semantically clearer to leave it as a list, and for each object to
possibly have more than one __klp_rela section.

For example, say we are patching two functions in ext4. In my
resulting kpatch module I will have two __klp_rela_ext4 sections, and
they might look like this when we run readelf --sections:

[34] __klp_rela_ext4.text.ext4_attr_store RELA ...
[35] __klp_rela_ext4.text.ext4_attr_show RELA ...

Then these two klp rela sections end up as two elements in the
reloc_secs list for the ext4 patch object. I think this way, we can
better tell which rela is being applied to what function. Might be
easier to understand what's happening from the developer's point of
view.

> You see, we go through elf sections here which were preserved by module
> loader. We even have relevant sections marked with SHF_RELA_LIVEPATCH. So
> maybe all the stuff around klp_reloc_sec is not necessary.
>
> Thoughts?

Ah, so this is where descriptive comments and documentation might have
been useful :-) So I think we will still need to keep the
klp_reloc_sec struct to help the patch module initialize. Though the
name and objname fields aren't used in this patchset, they are used in
the kpatch patch module code [1], where we iterate through each elf
section, find the ones marked with SHF_RELA_LIVEPATCH, set the
klp_reloc_sec's objname (which we find from the "name" field,
formatted as __klp_rela_{objname}.text..). Once we have the objname
set, we can then find the object to attach the reloc_sec to (i.e. add
it to its list of reloc_secs).

Hope that clears some things up.

Ok, I'll try to explain myself and it is gonna be long. I'll try to
describe how we deal with dynrelas in klp today, how you use it in kpatch
(and this is the place where my knowledge can be wrong or obsolete), what
you propose and what I'd like to propose.

1. First let's look on what we have now.

We have a patch module in which there is a section with all dynrelas
needed to be resolved (it was like this in kpatch some time ago and maybe
it is different now so just have a patience, I'll get to it). In the init
function of the module kpatch builds all the relevant info from dynrela
section. It goes through it, creates an array of klp_reloc for each object
and includes each dynrela record with an objname to the array of
klp_object with that objname. Later when we need to apply relocations for
patched object we just go through the list (array) of its dynrelas in
klp_object and call our arch-specific klp_write_module_reloc().

Sounds correct to me.

Now this was probably changed in kpatch and you do not have one dynrela
section but one dynrela section for each patched function. Is that
correct? (and can you tell us what the reason for the change was? It might
be crucial because I might be missing something.). Which leads us to your
proposal...

Your original assumption was correct; current kpatch has one big
.kpatch.dynrelas section, and each dynrela entry within that single
section gets sorted to the correct object as you described above. The
one dynrela section per patched function only came with this patchset,
but for no particular reason other than to make the kpatch-build code
for generating the patch module slightly less complicated. But I
haven't checked how big of a change it would be to do
one-dynrela-section per object, perhaps (and I hope) it will be an
easy change.

2. So we have several dynrela section for one patched object. During init
function in a patch module kpatch once again builds needed structures from
these sections. Now they are called klp_reloc_sec and contain different
kind of info. There is no val, loc and such, only name of the symbol,
objname and index to dynrela section in ELF. So when you need to write
relocations for the patched object you go through all relevant dynrela
sections (because they are stored in the klp_object), all dynrela records
in each section and you resolve the undefined symbols. All needed info is
stored in ELF. Then you just call apply_relocate_add().

3. I propose to go one step further. I think we don't need klp_reloc_sec
if there is only one dynrela section for patched object (and I currently
cannot see why this is not possible. It is possible even with one dynrela
section for whole patch module, that is for all patched objects.).

I think the furthest we can go in terms of simplifying klp rela secs
is to have one __klp_rela section per object. We can't smush all the
__klp_rela_objname sections into one big __klp_rela section since we
could be patching some objects that won't be loaded yet, and
apply_relocate_add() needs to work with real SHT_RELA sections + their
corresponding section indices (i.e., we cannot call
apply_relocate_add() with that single, combined klp rela section).

So, I think I would be OK with one __klp_rela section per object.

In my proposal there would be nothing done in init function of the patched
module (albeit some optimization mentioned later). When you call
klp_write_object_relocations you would go through all ELF section and find
the relevant one for the object (it is marked with SHF_RELA_LIVEPATCH and
objname is in the name of the section. It is the same thing you do in 2.
in the init function.). Then you go through all dynrela records in the
section, you do the same things which you do in the proposed patch above,
and call apply_relocate_add.

I'm not sure I like having klp_write_object_relocations() repeatedly
perform a loop through all the elf sections when we can pre-process
this information in the patch module's init function, and "cache" the
relevant klp section indices before passing things off to
klp_write_object_relocations(). So how about this: we do the
__klp_rela sec sorting in the init function of the patched module. The
patch module would iterate through its own elf sections, and when it
encounters a __klp_rela section, it looks at its objname, find the
corresponding klp_object, and save the section index of the __klp_rela
section in that klp_object. Then in klp_write_object_relocations, we
just have to look at the saved section index for the corresponding
object and access the klp rela section through that index, do the same
processing in this patch and call apply_relocate_add().

Now it would be crazy to go through all sections each time
klp_write_object_relocations is called (maybe it does not even matter but
nevertheless). So klp_object could have an index to its ELF dynrela
section. This index can be retrieved in the init function the same way you
do all the stuff with klp_reloc_sec.

Ah, exactly what I said above :-D

If you insisted on more than one dynrela section for a patched object we
could have an array of indices there. Or whatever.

It is almost the same as your proposal but in my opinion somewhat nicer.
We just use the info stored in ELF directly without unnecessary middle
layer (== klp_reloc_sec).

Does it make sense? I hope it does. Would it work?

It does make sense, and I think we can make it work without
klp_reloc_sec. Thanks for the explanations.

Jessica
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/