| Age | Commit message (Collapse) | Author |
|---|
|
The klp test runner deletes each test's working directory when the test
finishes. On failure that removes diff.log, out.o and the fixture objects
before the developer can look at them, and the only hint was to re-run
(something) with --keep:
$ make -C tools/objtool tests
...
not ok - test-special-section: expected section '.kcfi_traps' in output
...
# pass:49 fail:1 static-skip:1 probe-skip:0 xfail:0 xpass:0
# re-run with --keep to hold on to what a failing test built
Change the default behavior to save failing test working directories. This
is friendlier to the drive-by tester. Provide the power-user with a KEEP
Makefile variable that can optionally retain 'all' or 'none' of those
directories. run-tests.sh --keep remains an alias for KEEP=all.
Now that a run can be holding objects worth looking at, say where they are
when it is cut short. A Ctrl-C left the run directory behind -- the tests
clean up after themselves as they exit, but the two levels the runner made
are its own to answer for, and nothing removed them. Take them away from
a trap if the tests left them empty, and print the path if they did not:
under the new default the test that was running keeps what it had built,
and that is the run most worth being told about.
The trap only ever rmdirs, so it cannot overrule what a test decided to
keep; the same helper now serves the end of a normal run.
Signed-off-by: Joe Lawrence <joe.lawrence@redhat.com>
Signed-off-by: Song Liu <song@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Assisted-by: Claude:claude-opus-5
Link: https://patch.msgid.link/20260916184351.2720310-11-song@kernel.org
|
|
Two documents, for two different readers.
klp-test-design.txt is for someone deciding whether to trust the suite or
change it: what makes unit tests possible for a pipeline that normally
needs two kernel builds (objtool has no configuration-dependent logic, so a
test reproduces a configuration by reproducing its input), how tests are
selected per architecture, and what the outcome classification is for.
klp-write-tests.txt is the procedure for adding one, and is written to be
handed to someone -- or something -- with the instruction "follow this and
write a test for commit <sha>", or "port the case at <location> from
another harness". It leads with the rule that matters: a test is not
finished until you have watched it fail with the code broken. Then the
traps, each of which this suite has already fallen into -- a fixture the
compiler optimised away, a named char[] which never reaches the string path
because that keys on SHF_STRINGS, per-function sections hiding the movement
a test was about.
Some of this was previously written as comments in lib.sh. That was the
wrong place: guidance on how to write a test is read once by an author,
while every comment in the harness is read by everyone who opens it.
Explanations of why the code is as it is stay where they were.
Signed-off-by: Song Liu <song@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Assisted-by: Claude:claude-opus-5
Link: https://patch.msgid.link/20260916184351.2720310-10-song@kernel.org
|
|
Most of these assertions end in a grep, and an empty pattern matches
anything. So a mistyped variable name in a test does not fail: it expands
to nothing, the grep succeeds, and the test reports a pass having checked
nothing. assert_section "$sectoin_nmae" becomes a search for two adjacent
spaces, which readelf's column padding satisfies on every line.
Refuse unset variables. lib.sh guards the few genuinely optional
expansions it has with ${x:-}.
Signed-off-by: Song Liu <song@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Assisted-by: Claude:claude-opus-5
Link: https://patch.msgid.link/20260916184351.2720310-9-song@kernel.org
|
|
Each test made its own mktemp directory, so a run scattered them through
/tmp among everything else using the same names, and there was no way to
inspect what a failing test built -- the objects are gone the moment it
exits, which is right every time but the one that matters.
Give the run one directory and each test a subdirectory of it, named for
the test and nested the way the source is:
/tmp/klp-tests.3zC6oyfd/
generic/test-basic/{orig.o,patched.o,out.o,Module.symvers,...}
x86/test-kcfi/...
--keep then reports one path instead of forty, and removing it is one
command. A test run by hand still falls back to a temp directory of its
own, and still says where that went.
Detecting a leak comes free. Cleanup is now "remove each test's directory,
then rmdir the run's", and the rmdir fails if anything is left -- so a test
which dies without running its own cleanup is reported rather than quietly
leaving something behind.
preflight names the directory the tests build under, since mktemp honours
TMPDIR and a run in a container is otherwise silent about where its work
went. The runner says what to do only when it is relevant: on a failure
without --keep, that the option exists; with it, how many directories are
waiting. Usage text is not read while something is broken.
Signed-off-by: Song Liu <song@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Assisted-by: Claude:claude-opus-5
Link: https://patch.msgid.link/20260916184351.2720310-8-song@kernel.org
|
|
Assertions for what klp diff produces, and for what went into it.
The output side covers the things a livepatch is made of: cloned sections
and symbols, klp symbols and their sympos, tombstones, relocations by
section and by count, the converted .klp.rela sections and their
SHF_RELA_LIVEPATCH flag, SHN_LIVEPATCH symbols, recorded checksums, and klp
diff's own diagnostics -- a rejection for the wrong reason is not a pass.
The input side matters just as much, and is easier to forget. A test which
asserts only on the output passes when the compiler never emitted the
construct under test, and reads as coverage it does not have. So there are
two forms, and the difference between them is a statement about why the
thing might be absent:
require_* the compiler cannot produce it here -> skip
assert_* the fixture is supposed to produce it -> fail
Names are quoted before they reach grep. Nearly every name here contains
dots -- .text.target, .klp.rela.vmlinux -- and an unescaped dot matches any
character, so an assertion for one section could be satisfied by a
different one whose name merely lines up.
Signed-off-by: Song Liu <song@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Assisted-by: Claude:claude-opus-5
Link: https://patch.msgid.link/20260916184351.2720310-7-song@kernel.org
|
|
Compiling a fixture is the one thing every test does, and it has more
invariants than it looks.
- The flags describe the kernel a fixture stands in for, and each is
written down with its reason. That substitution is the whole mechanism
by which these tests cover configurations without building a kernel --
objtool reads no .config, so a configuration reaches it only as
compiler flags and the bytes they produce -- and a flag with no stated
motive is indistinguishable from a mistake.
- -c is not one of them. Producing an object rather than a program is
the helper's contract, not something a test may reconsider, so it lives
at the compile and an override cannot drop it.
- A fixture which will not compile is a failure, not a skip. It is far
more often a mistake in the fixture than a compiler which cannot
express the construct, and the two are indistinguishable once reported
as a skip.
- A newly produced object has no checksums in it, so producing one
invalidates the record that checksumming has already been done. Left
to itself that invariant ends up in each test which rebuilds.
build_one handles a single object for the tests which need more than two or
different names; build_pair and build_module_pair are expressed in terms of
it, so none of the above can be lost by going around them.
run_checksum comes out of run_diff for the same reason: a test which wants
checksums without a diff should not have to run one.
Signed-off-by: Song Liu <song@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Assisted-by: Claude:claude-opus-5
Link: https://patch.msgid.link/20260916184351.2720310-6-song@kernel.org
|
|
A test could pass, fail, or skip, and the runner reported only the exit
status of the last one to fail. Skips are the interesting case: they are
how a suite quietly stops testing anything.
Three kinds, and the runner counts them separately:
declared the test said in advance it does not apply here, e.g. gcc_only
on a clang run. Expected indefinitely.
probe the construct did not turn up in the built object this time.
Weaker: one which becomes permanent is a fixture that stopped
testing anything.
undeclared counted as a failure. A test which gives up for a reason it
never declared is a hole, not an outcome.
xfail and xpass come with them, so a known failure is reported rather than
commented out and forgotten, and one which starts passing says so instead
of going quietly green.
The result line is what gets classified, not the whole of a test's output:
objtool warns on stderr and the runner captures it, and a stray line ahead
of the result would otherwise leave the exit status to decide -- counting
an expected failure, which exits 0, as a pass.
Every test owes the plan exactly one result line, so a substituted verdict
replaces the test's own rather than being printed beside it, and a test
which dies before reporting gets a result of its own instead of being
inferred from its exit status.
Also: a TAP version line, LC_ALL=C so the order tests run in and the
behaviour of grep's character ranges inside them do not depend on the
invoking shell's locale, and fail() on stdout with every other verdict.
Co-developed-by: Puranjay Mohan <puranjay@kernel.org>
Signed-off-by: Puranjay Mohan <puranjay@kernel.org>
Signed-off-by: Song Liu <song@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Assisted-by: Claude:claude-opus-5
Link: https://patch.msgid.link/20260916184351.2720310-5-song@kernel.org
|
|
Which architecture a test is for is expressed by where it lives: tests are
in generic/ or in a directory named for their architecture, each carrying
its own fixtures, and the runner executes generic/ plus the one that
matches. A test which cannot apply here is then not run at all, rather
than running in order to announce that it did not.
Layout does this better than a declaration would. There is no x86_only(),
and no lookup letting a fixtures/<arch>/ file shadow a generic one of the
same name -- an arch-specific test simply carries its own fixtures.
Compilers cannot work the same way: CI varies CC over the same tree, so a
compiler requirement stays a declaration in the test.
What a run leaves out is reported once:
# not run: 5 tests in x86/ (this run is arm64)
Silence would have been cheaper and wrong. A run covering less than the
tree holds must not look like a run that covered all of it.
Signed-off-by: Song Liu <song@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Assisted-by: Claude:claude-opus-5
Link: https://patch.msgid.link/20260916184351.2720310-4-song@kernel.org
|
|
Every test checked for itself that objtool exists, was built with klp
support, and that $CC runs. Three problems with that: it is the same work
done 40 times, a missing objtool reads as a per-test skip rather than as a
suite which cannot run, and a run in which everything skipped still exits
0.
Do it once, before any test, in klp_preflight() in lib.sh, and export the
answers where the tests can read them. If the suite cannot run the whole
run fails and says why; a test which gets as far as running can assume its
environment.
Nothing is written down, so nothing can go stale. Sourcing lib.sh runs the
checks, which means a test run by hand establishes its own answers rather
than inheriting those of some earlier run with a different CC, and a test
is never handed an empty value for where objtool is or which architecture
it is on: either the exports are set or the run has already failed.
preflight answers only whether the suite can run at all -- not what the
compiler is capable of. A test needing a particular compiler feature
probes for it and skips; that costs one compile and keeps the reason next
to the test that has to justify it.
Signed-off-by: Song Liu <song@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Assisted-by: Claude:claude-opus-5
Link: https://patch.msgid.link/20260916184351.2720310-3-song@kernel.org
|
|
"objtool klp checksum" and "objtool klp diff" work on object files alone,
with no kernel, vmlinux or configuration involved, so they can be tested
directly. That is worth doing: most klp generation bugs so far have been in
symbol correlation, special section extraction and relocation conversion,
and several of them failed silently, producing a livepatch which built
cleanly but was missing data.
A test compiles a fixture twice, as the original and (with -DPATCHED) the
patched object, runs both through klp checksum, diffs them and asserts on
the result. Fixtures are compiled at test time rather than committed as
binaries: codegen varies between compilers and architectures, and that
variation is where a good number of these bugs come from.
Assertions check properties rather than compare against recorded output.
Golden files would need re-recording for every compiler change and would
report churn instead of regressions.
klp diff resolves symbols against Module.symvers, so the harness writes
one. Whether a symbol is listed there decides between an ordinary
relocation and a klp relocation, which makes it the main knob tests use.
Run with:
make -C tools/objtool tests
Tests skip when objtool was built without klp support or when a fixture
does not build for the target architecture. A missing objtool binary fails
instead of skipping, since that means a broken invocation rather than an
environment which cannot run the test.
Signed-off-by: Puranjay Mohan <puranjay@kernel.org>
Co-developed-by: Joe Lawrence <joe.lawrence@redhat.com>
Signed-off-by: Joe Lawrence <joe.lawrence@redhat.com>
Signed-off-by: Song Liu <song@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Assisted-by: Claude:claude-opus-5
Link: https://patch.msgid.link/20260916184351.2720310-2-song@kernel.org
|
|
commit 3f2de814c059 ("objtool: Fix libopcodes linking with static libraries")
tested for libopcodes availability by linking a test snippet with a forward
declaration of disassemble_init_for_target().
However, testing symbol linkage with an extern declaration only verifies
the presence of the library (.so/.a) and bypasses checking for development
headers (binutils-dev). On systems where libopcodes is present without
development headers installed, the probe succeeds, enabling BUILD_DISAS.
Subsequent compilation of objtool then fails:
fatal error: 'bfd.h' file not found
113 | #include <bfd.h>
Additionally, the probe invokes $(HOSTCC) without $(HOSTCFLAGS), ignoring
any sysroot or include flags specified for the host compiler.
Fix this by including <bfd.h> and <dis-asm.h> directly in the test snippet,
passing $(HOSTCFLAGS) so host compiler options are respected, and defining
PACKAGE="objtool" to satisfy the configuration check in <bfd.h>.
Fixes: 3f2de814c059 ("objtool: Fix libopcodes linking with static libraries")
Fixes: 436326bc525d ("objtool: fix build failure due to missing libopcodes check")
Reported-by: Alice Ryhl <aliceryhl@google.com>
Assisted-by: Antigravity:Gemini-Next
Signed-off-by: Ulises Mendez Martinez <umendez@google.com>
Link: https://patch.msgid.link/20260904150710.2997558-1-umendez@google.com
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
Adding a line of code to __link_shadow_page() with a literal string
causes a false positive changed function with GCC:
arch/x86/kvm/kvm.ko.o: changed function: kvm_tdp_mmu_map_private_pfn
While the patch only touched __link_shadow_page(), the string addition
triggered a rename of .LC64 -> .LC65 in kvm_tdp_mmu_map_private_pfn()
even though the underlying referenced constant data didn't change.
So for .LC* symbols, the suffix is arbitrary but the data isn't. Add
the underlying data to the checksum calculation rather than the symbol
name.
Clang also uses .LC* symbols, but also uses anonymous data. Both
compilers put this data in .rodata.cst<num> sections.
Fixes: 0d83da43b1e1 ("objtool/klp: Add --checksum option to generate per-function checksums")
Link: https://patch.msgid.link/f3a9e74ceebc6475ce94bcfe985401140857814a.1787939301.git.jpoimboe@kernel.org
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/ojeda/linux
Pull Rust fixes from Miguel Ojeda:
"Toolchain and infrastructure:
- Fix KCFI failures, such as in Rust doctests, by disabling function
merging when CFI is enabled. Gary reported the LLVM bug to upstream
and it is now fixed in their mainline.
- Fix 'objtool' fallthrough warnings under the experimental
'CONFIG_RUST_INLINE_HELPERS' by passing (for the combined Rust and
helpers code) the LLVM options needed to preserve the unreachable
traps that 'rustc' normally emits.
In addition, fix 'objtool' errors when LTO is enabled on top, by
also filtering out the LTO flags (for the combined Rust and helpers
code) so that the traps are kept in place.
- Fix 'objtool' warnings by adding one more 'noreturn' function.
- Fix 'make rusttest' target when the 'rustc-dev' component is
installed and Rust >= 1.82.0, <= 1.87.0 is used.
'kernel' crate:
- 'num' module: fix soundness issue in the 'Bounded' conversion from
'bool' by restricting the conversions to unsigned 'Bounded'.
- 'jump_label' module: fix future 'make rusttest' target failures
when 'ARCH=' is set to an arch different than the host's.
- 'list' module: fix incorrect 'pop_back()' comment"
* tag 'rust-fixes-7.3' of git://git.kernel.org/pub/scm/linux/kernel/git/ojeda/linux:
rust: kbuild: disambiguate `zerocopy_derive` for `rusttest`
rust: num: restrict bool conversion to unsigned Bounded
kbuild: rust: keep Rust objects out of Clang LTO with inline helpers
kbuild: rust: preserve unreachable traps with inline helpers
rust: cfi: disable function merging if CFI is enabled
rust: jump_label: skip arch-specific asm in `testlib` builds
objtool/rust: add one more `noreturn` Rust function
rust: kernel: list: fix incorrect pop_back example comment
|
|
When the pointer formatting series [1] is applied and KUnit tests
are enabled, `objtool` would report an error with any of our
supported Rust versions. For instance, with Rust 1.97.1:
rust/kernel.o: error: objtool: _R..._4core3fmt7Pointer3fmtB7_()
falls through to next function _R..._4core7convert5AsRefNtB5_4BStrE6as_ref()
Or, with Rust 1.85.0:
rust/kernel.o: error: objtool: _R..._4core3fmt7Pointer3fmtB7_()
falls through to next function _R..._4core3ffi5c_str4CStrENtNtBS_3fmt7Display3fmtB7_()
This happens due to calls to the `noreturn` symbol:
core::str::slice_error_fail
Thus add the mangled one to the list so that `objtool` knows it is
actually `noreturn`.
See commit 56d680dd23c3 ("objtool/rust: list `noreturn` Rust functions")
for more details.
[ While the series [1] mentioned above was not applied in a way
that triggered this, Gary found another case [2]:
I am seeing this function causing issue on my 1.85.1 -Os build, although the
error message is different this time:
rust/kernel.o: warning: objtool: _R..._5range9RangeFromjEE5indexCsldbe0TGTmNx_6kernel+0x1d:
stack state mismatch: cfa1=4+8 cfa2=4+16
Disassembly shows that the function call before the warning location
is a call to _4core3str16slice_error_fail, and objtool thought it can
return (which will end up with a unbalanced stack if it does return).
Thus apply it. - Miguel ]
Cc: stable@vger.kernel.org # Needed in 6.12.y and later (Rust is pinned in older LTSs).
Cc: Josh Poimboeuf <jpoimboe@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Reported-by: Alice Ryhl <aliceryhl@google.com>
Link: https://lore.kernel.org/rust-for-linux/anLxXvCKBcAVf91_@google.com/
Link: https://lore.kernel.org/rust-for-linux/20260706-hashedptr-v13-0-377a07f2f78d@kylinos.cn/ [1]
Link: https://lore.kernel.org/rust-for-linux/DKS7U9UHY6XR.2JC0EBIPCI06Z@garyguo.net/ [2]
Link: https://patch.msgid.link/20260805144524.233362-1-ojeda@kernel.org
Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/efi/efi
Pull EFI updates from Ard Biesheuvel:
- Set a timeout for EFI runtime service completions, and declare the
firmware wedged if it is exceeded. Note that this requires special
handling in case the firmware does return after all
- Rate limit the efivarfs statfs() handler as the QueryVariableInfo()
runtime service can be costly
- Sanity check the size of struct properties_header on Mac/x86
- Tweak the prototype of efi_guid_to_str()
* tag 'efi-next-for-v7.3' of git://git.kernel.org/pub/scm/linux/kernel/git/efi/efi:
efivarfs: Rate limit statfs() handler
efi: apple-properties: validate setup data header length
efi: make efi_guid_to_str() take a const GUID pointer
efi/runtime-wrappers: retire the worker if a wedged call ever returns
efi/runtime-wrappers: honour EFI_RUNTIME_SERVICES in the non-blocking paths
efi/runtime-wrappers: bound the wait for EFI runtime service calls
efi/runtime-wrappers: check EFI_RUNTIME_SERVICES before using efi_rts_work
efi/runtime-wrappers: handle queue_work() failure with goto exit
efi/runtime-wrappers: factor out efi_rts_park_worker()
efi: fix stale reference to efi_recover_from_page_fault()
|
|
x86's efi_crash_gracefully_on_page_fault() ends in an infinite
schedule() loop so the kworker that faulted in firmware never runs
efi_rts_wq again. A later change needs the same "park this worker
forever" primitive on the runtime service timeout path, so factor the
loop into a shared efi_rts_park_worker() and call it from the x86
page-fault handler.
No functional change.
Signed-off-by: Breno Leitao <leitao@debian.org>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip
Pull x86 alternatives update from Borislav Petkov:
- Remove the smp_locks alternatives machinery which was used to patch
out lock prefixes when running a SMP kernel on a uniprocessor machine
* tag 'x86_alternatives_for_v7.3_rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
x86/alternative: Drop smp_locks glue
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip
Pull objtool updates from Ingo Molnar:
- Fix various klp-build bugs reported by Joe Lawrence (Josh Poimboeuf,
Joe Lawrence)
- Misc fixes and cleanups (Puranjay Mohan, Thomas Huth and Ingo Molnar)
* tag 'objtool-core-2026-08-17' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
objtool/klp: Fix vmlinux klp relocations for EXPORT_SYMBOL_FOR_MODULES()
objtool/klp: Fix .kcfi_traps special section extraction
objtool/klp: Fix vmlinux .klp.symid link error for .exitcall.exit symbols
objtool/klp: Fix line numbers in Module.symvers parse errors
objtool/klp: Fix relocations for EXPORT_SYMBOL_FOR_MODULES() symbols
objtool/klp: Allow new references to module exports
objtool/klp: Don't match local symbols against exports
objtool/klp: Fix cross-module klp relocation section naming
objtool/klp: Explicitly disallow patching or referencing init code/data
objtool/klp: Ignore replacement offset of empty x86 alternatives
objtool/klp: Fix size of empty special section entries
objtool/klp: Fix vmlinux .klp.symid link error for .no_trim_symbol symbols
objtool/headers: Sync tools/include/linux/objtool_types.h with include/linux/objtool_types.h
objtool: Replace __ASSEMBLY__ with __ASSEMBLER__ in header files
objtool/klp: Fix symbol resolution for duplicate data symbols
objtool/klp: Add .klp.symid for sympos disambiguation
objtool/klp: Skip hidden directories when finding objects
objtool/klp: Fix false module dependencies caused by dead relocs
objtool/klp: Normalize Module.symvers paths to module names
objtool/klp: Fix module name normalization for paths with dots
|
|
When a module function references a vmlinux symbol which is exported
with EXPORT_SYMBOL_FOR_MODULES(), a patch to that function needs to use
a klp reloc.
Currently, livepatch fails to load such a module:
livepatch: invalid access to vmlinux symbol 'get_task_policy' from module-specific livepatch relocation section
livepatch: failed to initialize patch 'livepatch_test' for module 'testmod' (-22)
livepatch: patch 'livepatch_test' failed for module 'testmod', refusing to load module 'testmod'
klp diff puts all klp relocs in __klp_relocs.<patched object>, so
post-link names the section .klp.rela.<patched object>.<secname>, which the
kernel rejects for vmlinux symbols.
Commit 07f14d6af9d77 ("objtool/klp: Fix cross-module klp relocation
section naming") changed the meaning of objname in the klp rela section
name to be where the referenced symbol is referenced rather than where
it lives. That premise only holds for symbols in a module: the relocs
get applied when the patched module gets patched, and the module
dependency guarantees the referenced module is loaded by then.
A vmlinux symbol needs the opposite. It's always resolvable, and it has
to be applied when the patch module loads, before the module loader
initializes the patch module's special sections, which may reference it.
That's why livepatch rejects vmlinux symbols in module-specific
sections.
Use "vmlinux" as the section objname when the referenced symbol lives in
vmlinux. This moves such klp relocs from .klp.rela.kvm..text to
.klp.rela.vmlinux..text.
Fixes: 07f14d6af9d77 ("objtool/klp: Fix cross-module klp relocation section naming")
Reported-by: Dylan Hatch <dylanbhatch@google.com>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Acked-by: Song Liu <song@kernel.org>
Link: https://patch.msgid.link/f8e3b9fae109903a6aafb2a33310e4afdcebf58e.1786761327.git.jpoimboe@kernel.org
Closes: https://lore.kernel.org/CADBMgpz7iWC0=t=_gE-tfvv0mTPq4kg0qQ2zgPH8DVPE6eQ9Kw@mail.gmail.com
|
|
create_fake_symbols() creates a symbol per entry for special sections.
It does so in two steps: first for the sections which have
ANNOTATE_DATA_SPECIAL annotations, then for the rest, using entsize or
the reloc count to infer the entry size. The second step skips the
sections already handled by the first one by looking for a symbol at
offset 0.
That heuristic is too fuzzy: with Clang and CONFIG_CFI, it misfires on
.kcfi_traps because Clang emits a .Ltmp* assembler-local label at the
start of the section, so no symbols are created and
clone_special_sections() extracts nothing. klp-build still reports
SUCCESS, but the livepatch module has no __kcfi_traps section and the
traps for the patched functions are lost.
Look for the actual fake symbols created by the first step instead.
Fixes: da4326573ae8d ("objtool/klp: Fix kCFI trap handling")
Reported-by: Joe Lawrence <joe.lawrence@redhat.com>
Closes: https://lore.kernel.org/r/akQNqlfFC0T5pcMa@redhat.com
Acked-by: Song Liu <song@kernel.org>
Link: https://patch.msgid.link/8faaead205b219607b6fc2359ae743be824056eb.1786230311.git.jpoimboe@kernel.org
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
Building a kernel via klp-build fails to link:
`__exitcall_aes_mod_exit' referenced in section `.klp.symid' of vmlinux.o: defined in discarded section `.exitcall.exit' of vmlinux.o
`__exitcall_dax_exit' referenced in section `.klp.symid' of vmlinux.o: defined in discarded section `.exitcall.exit' of vmlinux.o
`__exitcall_hid_exit' referenced in section `.klp.symid' of vmlinux.o: defined in discarded section `.exitcall.exit' of vmlinux.o
`__exitcall_usb_serial_module_exit' referenced in section `.klp.symid' of vmlinux.o: defined in discarded section `.exitcall.exit' of vmlinux.o
module_exit() on a built-in emits a static __exitcall_$fn pointer into
.exitcall.exit, which vmlinux.lds.h discards unconditionally via
EXIT_CALL. When two built-in translation units define a module_exit()
function of the same name, the resulting local symbols collide,
symid_needed() sees a duplicate and emits a .klp.symid entry for each,
referencing symbols the linker then throws away.
Same-named module_exit() functions are not rare:
dax_exit drivers/dax/device.c, drivers/dax/fsdev.c
hid_exit drivers/hid/hid-core.c, drivers/hid/usbhid/hid-core.c
aes_mod_exit arch/arm64/crypto/aes-ce-ccm-glue.c, lib/crypto/aes.c
usb_serial_module_exit module_usb_serial_driver() expands to this fixed
name in each of its ~49 users
The last one makes the collision structural rather than accidental: any
kernel with two built-in USB serial drivers has it. This is not arch
specific either; it only requires the objects to be built in rather than
modular, which is why a monolithic config trips it while a typical
distro config does not.
Add .exitcall.exit to the discarded section list so its symbols don't get
symids.
This is the same failure mode as "objtool/klp: Fix vmlinux .klp.symid
link error for .no_trim_symbol symbols", for another unconditionally
discarded allocated section.
Fixes: 029223d30162 ("objtool/klp: Add .klp.symid for sympos disambiguation")
Signed-off-by: Puranjay Mohan <puranjay@kernel.org>
Link: https://patch.msgid.link/m24igzlbxf.fsf@kernel.org
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
read_exports() reports the offending line number when it fails to parse
Module.symvers. The counter is initialized to 1 but never incremented,
so every error blames line 1 regardless of where the bad line is.
Acked-by: Joe Lawrence <joe.lawrence@redhat.com>
Acked-by: Song Liu <song@kernel.org>
Link: https://patch.msgid.link/133e16bb0c7cb916f10bbfb017eba525449ad1d6.1786138493.git.jpoimboe@kernel.org
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
EXPORT_SYMBOL_FOR_MODULES() puts a symbol in a "module:<names>"
namespace, which the module loader grants access to by matching the
importing module's name against that list.
klp_reloc_needed() only creates a klp reloc for module-owned exports; a
vmlinux export gets a normal reloc. For a vmlinux symbol exported with
EXPORT_SYMBOL_FOR_MODULES(), using a normal reloc results in a modpost
failure in klp-build:
ERROR: modpost: module livepatch-foo uses symbol mpol_shared_policy_lookup from namespace module:kvm, but does not import it.
And the modpost error is correct: even with that error removed, the
patch module would fail to load:
livepatch_foo: module uses symbol (mpol_shared_policy_lookup) from namespace module:kvm, but does not import it.
livepatch_foo: Unknown symbol mpol_shared_policy_lookup (err -22)
Treat it like an unexported symbol by using a klp reloc.
Note this only affects "module:" namespaces. Ordinary namespaced
exports continue to work with normal relocs thanks to copy_import_ns(),
which propagates the patched object's import_ns tags to the patch
module.
Fixes: dd590d4d57eb ("objtool/klp: Introduce klp diff subcommand for diffing object files")
Reported-by: Joe Lawrence <joe.lawrence@redhat.com>
Link: https://lore.kernel.org/6a6608f4-0a05-4d75-8b7f-edddfac9c5d4@redhat.com
Acked-by: Joe Lawrence <joe.lawrence@redhat.com>
Acked-by: Song Liu <song@kernel.org>
Link: https://patch.msgid.link/fe5a00818e06ec613344d41d5944de054fcd8832.1786138493.git.jpoimboe@kernel.org
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
klp_reloc_needed() returns true for module exports to support
late-module patching. However, clone_reloc_klp() unconditionally
rejects symbols without a twin (i.e., new references added by the
patch), even when the symbol is a known export from Module.symvers.
Relax the check: allow new references to exported symbols by only
erroring on !twin when there is no export. The export metadata from
Module.symvers provides sufficient context to emit the klp-relocation
without a twin.
For a module export that isn't sufficient on its own though, as the
resulting klp relocation will only be resolved at patch-enable time if
the exporting module is loaded.
If the original (unpatched) module already depends on the exporting
module, the dependency is safe: the module loader ensures the dependency
is satisfied before the patched module can be loaded, so the
klp relocation target will exist.
However, if the patch introduces a reference to a module that the
original doesn't depend on, there is no such guarantee. The exporting
module could be absent or could be unloaded at any time, leading to a
relocation failure or use-after-free.
So also add a build-time check: when a new symbol reference (no twin)
targets a module export, verify that the original module already has at
least one UNDEF symbol resolving to that same exporting module. If not,
error out with a diagnostic message.
Signed-off-by: Joe Lawrence <joe.lawrence@redhat.com>
Acked-by: Song Liu <song@kernel.org>
Acked-by: Joe Lawrence <joe.lawrence@redhat.com>
Link: https://patch.msgid.link/e0d725acb4774747f0e271308b4ca33daae2d5db.1786138493.git.jpoimboe@kernel.org
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
While cloning a reloc, klp diff calls find_export() to determine whether
the referenced symbol is exported. That decides whether the reference
needs a klp reloc, which object the klp symbol belongs to, and whether
the symbol's data needs to be copied into the patch module.
But find_export() matches purely on symbol name, so a static function or
variable which happens to share its name with an export is mistaken for
a reference to that export:
- klp_reloc_needed() creates a klp reloc pointing at the exporting
module's symbol rather than the local one. For a vmlinux export it
skips the klp reloc altogether, leaving a normal reloc which the
module loader resolves to the vmlinux symbol.
- clone_reloc() treats the symbol as external and clones it without
its data, leaving a dangling reference.
- validate_special_section_klp_reloc() attributes a static branch or
call key to the wrong module, and for a vmlinux export skips the
unsupported-key check entirely.
Exports are always global, so ignore local symbols in find_export().
Fixes: dd590d4d57eb ("objtool/klp: Introduce klp diff subcommand for diffing object files")
Acked-by: Song Liu <song@kernel.org>
Acked-by: Joe Lawrence <joe.lawrence@redhat.com>
Link: https://patch.msgid.link/dafc7c017d1f62e2190a20dc84a52fbfc4c70210.1786138493.git.jpoimboe@kernel.org
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
A klp relocation section is .klp.rela.<objname>.<secname>, where objname
is the object being patched.
klp-build wrongly derives objname from where the referenced symbol
lives, not where it's referenced. For a cross-module reference like
patched can_isotp code calling can.ko's can_rx_unregister(), that gives
.klp.rela.can..text rather than .klp.rela.can_isotp..text. Unless the
patch happens to patch can.ko as well, the relocation never gets applied
and the call goes off into the weeds.
Name the intermediate section __klp_relocs.<objname> so post-link can
read the patched object's name from there.
Fixes: dd590d4d57eb ("objtool/klp: Introduce klp diff subcommand for diffing object files")
Reported-by: Joe Lawrence <joe.lawrence@redhat.com>
Link: https://lore.kernel.org/20260720145658.1103243-2-joe.lawrence@redhat.com
Acked-by: Song Liu <song@kernel.org>
Acked-by: Joe Lawrence <joe.lawrence@redhat.com>
Link: https://patch.msgid.link/ee93a08f3e55e76ffa67d04e283917ddaa893f09.1786138493.git.jpoimboe@kernel.org
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
Explicitly disallow the patching and referencing of init code/data.
Otherwise it could potentially introduce some odd edge cases depending
on whether the target object's init section has been freed yet (note
that the init code still exists in the target module when doing late
module patching).
Such edge cases include sympos calculation and the patching and/or
referencing of non-existent (init-freed) code/data. Not to mention the
inherent differences in behavior that occur when the init code is only
patched *some* of the time depending on module loading order or kernel
config.
Acked-by: Joe Lawrence <joe.lawrence@redhat.com>
Acked-by: Song Liu <song@kernel.org>
Link: https://patch.msgid.link/516e14f84cfbffa27dc19d3dcf35097504097966.1786138493.git.jpoimboe@kernel.org
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
An x86 alternative with an empty replacement, e.g. the second entry of
ALTERNATIVE_2("orig", "repl", ft1, "", ft2)
has a replacementlen of zero. Its replacement offset still gets a
relocation, but the label it points at is the end of the previous
replacement, which is also the beginning of the *next* alternative's
replacement. The value is meaningless; get_alt_entry() already ignores
it for that reason.
klp diff doesn't ignore it. When such an alternative belongs to a
changed function, cloning its relocations drags in the unrelated
neighboring replacement, along with everything that replacement
references. On an x86 clang/lto build an empty alternative in
meminfo_proc_show() pulled in the replacement of an alternative in
proc_kcore_init(), silently emitting a klp relocation against init text
which has long since been freed by the time the patch is applied.
Add arch_alt_ignore_new_reloc() and skip such relocations when cloning.
This has to be arch specific: on arm64 a zero-length replacement instead
identifies an alternative callback, whose replacement offset points at
the callback function and must be preserved.
Fixes: dd590d4d57eb ("objtool/klp: Introduce klp diff subcommand for diffing object files")
Acked-by: Song Liu <song@kernel.org>
Acked-by: Joe Lawrence <joe.lawrence@redhat.com>
Link: https://patch.msgid.link/7a885b70974795c3417f3358869e62aafd4ef783.1786138493.git.jpoimboe@kernel.org
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
create_fake_symbols() sizes each ANNOTATE_DATA_SPECIAL entry from the
offset of the next annotation, falling back to the end of the section
for the last entry. But the last entry is detected by a zero size,
which also happens for an *empty* entry: ALTERNATIVE(oldinstr, "", ft)
still annotates its zero-length replacement, at the same offset as the
next entry's annotation.
So every empty replacement gets a fake symbol spanning the entire rest
of .altinstr_replacement. That's harmless today only because
find_symbol_containing() picks the smaller of two overlapping symbols.
Track whether a next annotation was found rather than inferring it from
the size. A zero-length fake symbol is fine: find_symbol_containing()
skips those, so the properly sized symbol at the same offset still wins.
Fixes: dd590d4d57eb ("objtool/klp: Introduce klp diff subcommand for diffing object files")
Acked-by: Song Liu <song@kernel.org>
Acked-by: Joe Lawrence <joe.lawrence@redhat.com>
Link: https://patch.msgid.link/913e691c5009397df832c7c9a18cd5cf71b42737.1786138493.git.jpoimboe@kernel.org
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
Testing klp-build with arm64 produced the following linker error during
the original kernel build:
`__notrim.1' referenced in section `.klp.symid' of vmlinux.o: defined in discarded section `.no_trim_symbol' of vmlinux.o
symbol_get() puts a static __notrim[] in .no_trim_symbol, which GCC
names __notrim.1, __notrim.2, etc. Two or more built-in translation
units calling symbol_get() thus produce duplicate names, resulting in
corresponding .klp.symid references which trigger the above error.
Add .no_trim_symbol to the discarded section list so its symbols don't
get symids.
Note this issue is not specific to arm64: it just needs two built-in
symbol_get() callers. arm64 trips over it easily because it has KVM
always compiled in vmlinux, whereas on x86 it's typically a module.
Fixes: 029223d30162 ("objtool/klp: Add .klp.symid for sympos disambiguation")
Acked-by: Song Liu <song@kernel.org>
Acked-by: Joe Lawrence <joe.lawrence@redhat.com>
Link: https://patch.msgid.link/5a3cc4cded743167dd0878220201d80f7e48e5d7.1786138493.git.jpoimboe@kernel.org
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
Starting with Rust 1.99.0 (expected 2026-10-01), `objtool` may report:
rust/kernel.o: warning: objtool: _R..._6kernel3str9parse_intaNtNtB2_7private12FromStrRadix14from_str_radix()
falls through to next function _R..._6kernel3str9parse_intaNtNtB2_7private12FromStrRadix16from_u64_negated()
due to calls to the `noreturn` symbol:
core::num::from_ascii_bytes_radix_panic
The function was renamed from `from_ascii_radix_panic` [1], which is
already in the list.
Thus add the new one to the list so that `objtool` knows it is actually
`noreturn`.
See commit 56d680dd23c3 ("objtool/rust: list `noreturn` Rust functions")
for more details.
Cc: stable@vger.kernel.org # Needed in 6.12.y and later (Rust is pinned in older LTSs).
Cc: Josh Poimboeuf <jpoimboe@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://github.com/rust-lang/rust/pull/159554 [1]
Tested-by: Alice Ryhl <aliceryhl@google.com>
Link: https://patch.msgid.link/20260728191448.349241-1-ojeda@kernel.org
Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
|
|
find_sympos() calculates a sympos used by livepatch to disambiguate
duplicately-named symbols. For function symbols, there's a hack which
counts .text.unlikely symbols before other .text symbols, matching the
linker script's section ordering.
Not only is the hack fragile, data symbols can have the same problem.
So for example, adding a reference to pwq_cache in
ep_unregister_pollwait() can trigger a corrupt sympos and a relocation
to the wrong pwq_cache symbol in the livepatch module, resulting in a
crash or undefined behavior.
Remove the existing hack in favor of a fully deterministic solution,
using the new .klp.symid table to derive the symbol-to-id mapping from
the original vmlinux.o and the id-to-address mapping from the
corresponding vmlinux, which can then be used to determine the exact
sympos associated with the original vmlinux.
Modules don't need any special treatment: the .ko has the same
section/symbol ordering as the original whole-archive symbol table.
Fixes: dd590d4d57eb ("objtool/klp: Introduce klp diff subcommand for diffing object files")
Reported-by: Ben Procknow <bprockno@redhat.com>
Reported-by: Joe Lawrence <joe.lawrence@redhat.com>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Cc: live-patching@vger.kernel.org
Link: https://lore.kernel.org/20260710153042.3156788-1-joe.lawrence@redhat.com
Link: https://lore.kernel.org/20260724221730.3126529-1-joe.lawrence@redhat.com
Link: https://patch.msgid.link/919785e3bf2245db02ff6391e735d9cb139170b1.1785727106.git.jpoimboe@kernel.org
|
|
Livepatch identifies a duplicate-named symbol by its position (sympos)
among same-named kallsyms entries, which for vmlinux are counted in
ascending address order in the final linked kernel. That order can't be
reliably derived from vmlinux.o: the final link reorders sub-sections
(.text.unlikely*, .data..*, etc).
Bridge the gap with a new .klp.symid section which can be used to
correlate symbols between vmlinux.o and vmlinux so that klp-diff can
reliably determine the sympos.
The table can't survive --gc-sections: keeping it alive would keep every
duplicate-named symbol's section alive, so the reference kernel would
stop matching the one which ships. klp-build rejects
CONFIG_LD_DEAD_CODE_DATA_ELIMINATION instead. Nothing is lost today:
x86_64 is the only HAVE_KLP_BUILD arch and doesn't select
HAVE_LD_DEAD_CODE_DATA_ELIMINATION, arm64 and s390 have never selected
it either, and on powerpc, it's still EXPERIMENTAL and disabled by every
distro kernel.
This is the build-time half of reliable vmlinux sympos computation;
"objtool klp diff" will consume the table in a subsequent commit.
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Cc: live-patching@vger.kernel.org
Link: https://patch.msgid.link/64d50f077b569f47883c015cdb7079edb068efe8.1785727106.git.jpoimboe@kernel.org
|
|
When creating a klp reloc, klp-diff keeps the original relocation but
converts the referenced symbol to an UNDEF/WEAK placeholder tombstone
symbol, which gets fully disabled later by klp post-link. The tombstone
symbol is only needed to avoid confusing objtool when it does the final
run on the patch module.
However, for references to exported symbols, modpost sees the reference
to the tombstone symbol as a real reference to an exported symbol,
resulting in a false module dependency getting created.
Further, for a reference to a tombstone symbol which is exported into a
module namespace, e.g. via EXPORT_SYMBOL_FOR_KVM_INTERNAL(), modpost
can't satisfy the dependency, resulting in a warning like the following:
module ... uses symbol kvm_flush_remote_tlbs from namespace
module:kvm-amd,kvm-intel, but does not import it.
Rename the placeholder tombstone symbols to ".klp.tombstone.<name>" so
modpost no longer recognizes them.
Fixes: dd590d4d57eb ("objtool/klp: Introduce klp diff subcommand for diffing object files")
Reported-by: Ben Procknow <bprockno@redhat.com>
Reported-by: Joe Lawrence <joe.lawrence@redhat.com>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Cc: live-patching@vger.kernel.org
Link: https://lore.kernel.org/20260720145658.1103243-5-joe.lawrence@redhat.com
Link: https://patch.msgid.link/9548393f4d89ec3b498f4f69aa6ef6b9bb7150fe.1785727106.git.jpoimboe@kernel.org
|
|
Module.symvers contains build-tree object paths as module identifiers
(e.g., "arch/x86/kvm/kvm") rather than runtime module names ("kvm").
Objtool's clone_reloc_klp() uses this field directly for exported
symbols, while unexported symbols correctly go through __find_modname().
This means that exported symbol relocations may land in a .klp.rela
section named with the build path rather than the module name. That is
a crash waiting to happen: the kernel's livepatch loader silently skips
this relocation because it doesn't match the expected klp_object name.
The unresolved relocation sits in the newly activated code, crashing
when executed.
Normalize export->mod at Module.symvers read time using the same logic
as __find_modname() (refactored into a shared normalize_modname()
helper).
Fixes: dd590d4d57eb ("objtool/klp: Introduce klp diff subcommand for diffing object files")
Reported-by: Ben Procknow <bprockno@redhat.com>
Signed-off-by: Joe Lawrence <joe.lawrence@redhat.com>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Miroslav Benes <mbenes@suse.cz>
Cc: live-patching@vger.kernel.org
Link: https://patch.msgid.link/dbe1b72931bd3c31b751fd0729613d9f2226fff6.1785727106.git.jpoimboe@kernel.org
|
|
When .modinfo has no "name=" tag, __find_modname() falls back to
converting the object's build-tree path to a runtime module name by
stripping directory components, converting '-' to '_' and truncating the
file extension.
It does all that in a single pass over the entire path, so the first dot
anywhere in the path ends the name. For an object built in a directory
whose name contains a dot, e.g. "drivers/foo-1.0/bar.o", the result is a
bogus module name.
Strip the directory components up front so only the basename is scanned
for the extension separator.
Fixes: dd590d4d57eb ("objtool/klp: Introduce klp diff subcommand for diffing object files")
Reported-by: Sashiko <sashiko-bot@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Cc: live-patching@vger.kernel.org
Link: https://patch.msgid.link/9017b4609553bed16674e8f924d34691cbc2b2c1.1785727106.git.jpoimboe@kernel.org
|
|
This was there to be able to patch out locking instructions when running
a SMP kernel build on a UP CPU. The times are long gone when single-CPU
x86 machines were relevant so drop that machinery and simplify the code
considerably.
LOCK_PREFIX needs to stay for when one wants to do a UP build for
whatever reason. That'll go away when CONFIG_SMP becomes unconditional.
Kill a bunch of leftover, unused prototypes while at it.
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Mike Rapoport (Microsoft) <rppt@kernel.org>
Link: https://lore.kernel.org/r/20260328081634.797552-1-rppt@kernel.org
|
|
Starting with Rust 1.99.0 (expected 2026-10-01), under
`CONFIG_RUST_DEBUG_ASSERTIONS=y`, `objtool` may report:
rust/kernel.o: warning: objtool: _R..._6kernel12module_param9set_paramaEB4_()
falls through to next function _R..._6kernel12module_param9set_paramhEB4_()
(and many others) due to calls to the `noreturn` symbol [1]:
core::panicking::panic_null_reference_constructed
Thus add the mangled one to the list so that `objtool` knows it is
actually `noreturn`.
See commit 56d680dd23c3 ("objtool/rust: list `noreturn` Rust functions")
for more details.
Cc: stable@vger.kernel.org # Needed in 6.12.y and later (Rust is pinned in older LTSs).
Cc: Josh Poimboeuf <jpoimboe@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Petr Pavlu <petr.pavlu@suse.com>
Link: https://github.com/rust-lang/rust/pull/158796 [1]
Reported-by: Alice Ryhl <aliceryhl@google.com>
Closes: https://lore.kernel.org/rust-for-linux/alEBInX9gD1M5NAr@google.com/
Reviewed-by: Alice Ryhl <aliceryhl@google.com>
Tested-by: Alice Ryhl <aliceryhl@google.com>
Link: https://patch.msgid.link/20260710173252.191781-1-ojeda@kernel.org
Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
|
|
Cache the dont_correlate() result once per symbol at the start of
correlate_symbols(). This reduces klp diff time on an arm64 LTO
vmlinux.o from 2m51s to 35s.
Acked-by: Song Liu <song@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
Only create prefix symbols for functions that have
__patchable_function_entries entries, since those are the only C
functions where prefix NOPs are intentional.
This both simplifies the detection and makes it more accurate.
Note that assembly functions using SYM_TYPED_FUNC_START() can also have
prefixed NOPs, but that macro already creates their __cfi_ symbols.
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
With CFI+CALL_PADDING, Clang places .Ltmp labels at the start of the NOP
padding (offset 5) between the __cfi_ prefix and the function entry
point. get_func_prefix() only checks the immediately previous symbol,
so the intervening .Ltmp label causes it to miss the __cfi_ prefix
symbol.
This results in klp-diff not cloning the kCFI type hash into the
livepatch module, causing a CFI failure at module load when calling
callback functions through indirect calls:
CFI failure at __klp_enable_patch+0xab/0x140
(target: pre_patch_callback+0x0/0x80 [livepatch_combined];
expected type: 0xde073954)
Instead of walking backward through the section's symbol list, just use
find_func_containing() for the byte before the function. This works now
that __cfi_ symbols are being grown by objtool to fill the padding.
Acked-by: Song Liu <song@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
For all CONFIG_CFI+CONFIG_CALL_PADDING configs, for C functions, the
__cfi_ symbols only cover the 5-byte kCFI type hash. After that there
also N bytes of NOP padding between the hash and the function entry
which aren't associated with any symbol.
The NOPs can be replaced with actual code at runtime. Without a symbol,
unwinders and tooling have no way of knowing where those bytes belong.
Grow the existing __cfi_* symbols to fill that gap.
Note that assembly functions with SYM_TYPED_FUNC_START() aren't affected
by this issue, their __cfi_ symbols also cover the padding.
Also, CONFIG_PREFIX_SYMBOLS has no reason to exist: CONFIG_CALL_PADDING
is what causes the compiler to emit NOP padding before function entry
(via -fpatchable-function-entry), so it's the right condition for
creating prefix symbols.
Remove CONFIG_PREFIX_SYMBOLS, as it's no longer needed. Simplify the
LONGEST_SYM_KUNIT_TEST dependency accordingly. Rework objtool's
arguments a bit to handle the variety of prefix/cfi-related cases.
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
When computing klp checksums, instructions with non-relocated jump/call
destination offsets are problematic because the offset values can change
when surrounding code has moved, causing the function to be incorrectly
marked as changed.
Specifically, that includes jumps from alternatives to the end of the
alternative, which from objtool's perspective are jumps to the end of
the alternative instruction block in the original function.
Note that 'jump_dest' jumps don't include sibling calls (those use
call_dest), nor do they include jumps to/from .cold sub functions (those
are cross-section and need a reloc).
Fix it by hashing the opcode bytes (excluding the immediate operand)
along with a position-independent representation of the destination.
For calls, use the function name, and for jumps, use the destination's
offset within its function.
[Note the "9 bit hole" comment was wrong: it has been 8 bits since
commit 70589843b36f ("objtool: Add option to trace function validation")
added the 'trace' field. Adding the 4-bit 'immediate_len' field now
leaves a 4-bit hole.]
Fixes: 0d83da43b1e1 ("objtool/klp: Add --checksum option to generate per-function checksums")
Acked-by: Song Liu <song@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
Alternative replacement instructions awkwardly have insn->sym set to the
function they get patched to rather than the symbol (or rather lack
thereof) they belong to in the file.
This makes it difficult to know where a given instruction actually
lives.
Add a new insn_sym() helper which preserves the existing semantic of
insn->sym. Rename insn->sym to insn->_sym, which contains the actual
ELF binary symbol (or NULL, for alternative replacements) an instruction
lives in.
The private insn->_sym value will be needed for a subsequent patch.
Acked-by: Song Liu <song@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
Add debugging messages to show how duplicate symbols get correlated, and
split the --debug feature into --debug-correlate and --debug-clone.
Acked-by: Song Liu <song@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
Rewrite the symbol correlation code, using a tiered list of
deterministic strategies in a loop. For duplicately named symbols, each
tier applies a filter with the goal of finding a 1:1 deterministic
correlation between the original and patched version of the symbol.
The three matching strategies are:
find_twin(): A funnel of progressively tighter filters. Candidates
with the same demangled name are counted at four levels: name, scope
(local-vs-global), file (strict file association), and checksum
(unchanged functions). The widest level that yields a 1:1 match wins,
narrower levels are only tried when the wider level is ambiguous.
find_twin_suffixed(): Uses already-correlated LLVM symbol pairs to map
.llvm.<hash> suffixes from orig to patched. Because all promoted
symbols from the same TU share the same hash, one correlated pair
seeds the mapping for the entire TU.
find_twin_positional(): Last resort, matches symbols by position among
same-named candidates, similar to livepatch sympos. Used for data
objects like __quirk variables where no deterministic filter can
distinguish the candidates.
Overall this works much better than the existing algorithm, particularly
with LTO kernels.
Acked-by: Song Liu <song@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
Start checksumming data objects in preparation for revamping the
correlation algorithm.
Acked-by: Song Liu <song@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
The checksum functionality has been moved to "objtool klp checksum"
which is now used by klp-build. Remove the now-dead --checksum and
--debug-checksum options from the default objtool command.
Acked-by: Song Liu <song@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
Move the checksum functionality out of the main objtool command into a
new "objtool klp checksum" subcommand.
This has the benefit of making the code (and the patch generation
process itself) more modular.
For bisectability, both "objtool --checksum" and "objtool klp checksum"
work for now. The former will be removed after klp-build has been
converted to use the new subcommand.
Acked-by: Song Liu <song@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
|
decode_sections() relies on CFI and cfi_hash initialization done
separately in check(), making it unusable outside of check().
Consolidate the initialization into decode_sections() and rename it to
decode_file(), and make it global along with free_insns() and
insn_reloc() for use by other objtool components -- namely, the checksum
code which will be moving to another file.
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Song Liu <song@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
|
