[RFC PATCH 0/5] Arch-independent livepatch

From: Jessica Yu
Date: Mon Nov 09 2015 - 23:46:06 EST


This patchset removes livepatch's need for architecture-specific relocation
code by leveraging existing code in the module loader to perform
arch-dependent work. Specifically, instead of duplicating code and
re-implementing what the apply_relocate_add() function in the module loader
already does in livepatch's klp_write_module_reloc(), we reuse
apply_relocate_add() to write relocations. The hope is that this will make
livepatch more easily portable to other architectures and greatly reduce
the amount of arch-specific code required to port livepatch to a particular
architecture.

Background: Why does livepatch need to write its own relocations?
==
A typical livepatch module contains patched versions of functions that can
reference non-exported global symbols and non-included local symbols.
Relocations referencing these types of symbols cannot be left in as-is
since the kernel module loader cannot resolve them and will therefore
reject the livepatch module. Furthermore, we cannot apply relocations that
affect modules not loaded yet at run time (e.g. a patch to a driver). The
current kpatch build system therefore solves this problem by embedding
special "dynrela" (dynamic reloc) sections in the resulting patch module
elf output. Using these dynrela sections, livepatch can correctly resolve
symbols while taking into account its scope and what module the symbol
belongs to, and then manually apply the dynamic relocations.

Motivation: Why is having arch-dependent relocation code a problem?
==
The original motivation for this patchset stems from the increasing
roadblocks encountered while attempting to port livepatch to s390.
Specifically, there were problems dealing with s390 PLT and GOT relocation
types (R_390_{PLT,GOT}), which are handled differently from x86's
relocation types (which are much simpler to deal with, and a single
livepatch function (klp_write_module_reloc()) has been sufficient enough).
These s390 reloc types cannot be handled by simply performing a calculation
(as in the x86 case). For s390 modules with PLT/GOT relocations, the kernel
module loader allocates and fills in PLT+GOT table entries for every symbol
referenced by a PLT/GOT reloc in module core memory. So the problem of
porting livepatch to s390 became much more complicated than simply writing
an s390-specific klp_write_module_reloc() function. How can livepatch
handle these relocation types if the s390 module loader needs to allocate
and fill PLT/GOT entries ahead of time? The potential solutions were: 1)
have livepatch possibly allocate and maintain its own PLT/GOT tables for
every patch module (requiring even more arch-specific code), 2) modify the
s390 module loader heavily to accommodate livepatch modules (i.e. allocate
all the needed PLT/GOT entries for livepatch in advance but refrain from
applying relocations for to-be-patched modules), or 3) eliminate this
potential mess by leveraging module loader code to do all the relocation
work, letting livepatch off the hook completely. Solution #3 is what this
patchset implements.

How does this patchset remedy these problems?
==
Reusing the module loader code to perform livepatch relocations means that
livepatch no longer needs arch-specific reloc code and the aforementioned
problems with s390 PLT/GOT reloc types disappear (because we let the module
loader do all the relocation work for us). It will enable livepatch to be
more easily ported to other architectures.

Summary of proposed changes
==
This patch series enables livepatch to use the module loader's
apply_relocate_add() function to resolve livepatch relocations (i.e. what
used to be dynrelas). apply_relocate_add() requires access to a patch
module's section headers, symbol table, reloc section indices, etc., and all
of these are accessible through the load_info struct used in the module
loader. Therefore we persist this struct for livepatch modules and it is
made available through module->info.

The ELF-related changes enable livepatch to patch modules that are not
loaded yet. In order to use apply_relocate_add(), we need real SHT_RELA
sections to pass in. A complication here is that relocations for
not-yet-loaded modules should not be applied when the patch module loads;
they should only be applied once the target module is loaded. Thus kpatch
build scripts were modified to output a livepatch module that contains
special __klp_rela sections that correspond to the modules being patched.
They are marked with a special SHF_RELA_LIVEPATCH section flag to indicate
to the module loader that it should ignore that reloc section and that
livepatch will handle them. The SHN_LIVEPATCH shndx marks symbols that will
have to be resolved once their respective target module loads. So, the
module loader ignores these symbols (and does not attempt to resolve them).
Finally, the STB_LIVEPATCH_EXT symbol bind marks the scope of certain
livepatch symbols, so that livepatch can find the symbol in the right
place. These ELF constants were selected from OS-specific ranges according
to the definitions from glibc.

Jessica Yu (5):
elf: add livepatch-specific elf constants
module: save load_info for livepatch modules
livepatch: reuse module loader code to write relocations
samples: livepatch: init reloc list and mark as klp module
livepatch: x86: remove unused relocation code

arch/x86/kernel/Makefile | 1 -
arch/x86/kernel/livepatch.c | 91 ------------------------------
include/linux/livepatch.h | 11 +++-
include/linux/module.h | 31 ++++++++++
include/uapi/linux/elf.h | 3 +
kernel/livepatch/core.c | 106 ++++++++++++++++++++++++-----------
kernel/module.c | 36 +++++++-----
samples/livepatch/livepatch-sample.c | 2 +
8 files changed, 139 insertions(+), 142 deletions(-)
delete mode 100644 arch/x86/kernel/livepatch.c

--
2.4.3

--
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/