In preparation for the objtool klp diff subcommand, broaden the
elf_create_section() interface to give callers more control and reduce
duplication of some subtle setup logic.
While at it, make elf_create_rela_section() global so sections can be
created by the upcoming klp diff code.
Acked-by: Petr Mladek <pmladek@suse.com>
Tested-by: Joe Lawrence <joe.lawrence@redhat.com>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
'struct objtool_file' is specific to the check code and doesn't belong
in the elf code which is supposed to be objtool_file-agnostic. Convert
the elf iterator macros to use 'struct elf' instead.
Acked-by: Petr Mladek <pmladek@suse.com>
Tested-by: Joe Lawrence <joe.lawrence@redhat.com>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Move init_orc_entry(), write_orc_entry(), reg_name(), orc_type_name()
and print_reg() from generic orc_gen.c and orc_dump.c to arch-specific
orc.c, then introduce a new function orc_print_dump() to print info.
This is preparation for later patch, no functionality change.
Co-developed-by: Jinyang He <hejinyang@loongson.cn>
Signed-off-by: Jinyang He <hejinyang@loongson.cn>
Co-developed-by: Youling Tang <tangyouling@loongson.cn>
Signed-off-by: Youling Tang <tangyouling@loongson.cn>
Signed-off-by: Tiezhu Yang <yangtiezhu@loongson.cn>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
When creating an annotation section, allocate the reloc section data at
the beginning. This simplifies the data model a bit and also saves
memory due to the removal of malloc() in elf_rebuild_reloc_section().
With allyesconfig + CONFIG_DEBUG_INFO:
- Before: peak heap memory consumption: 53.49G
- After: peak heap memory consumption: 49.02G
Link: https://lore.kernel.org/r/048e908f3ede9b66c15e44672b6dda992b1dae3e.1685464332.git.jpoimboe@kernel.org
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Mark reported that the ORC unwinder incorrectly marks an unwind as
reliable when the unwind terminates prematurely in the dark corners of
return_to_handler() due to lack of information about the next frame.
The problem is UNWIND_HINT_EMPTY is used in two different situations:
1) The end of the kernel stack unwind before hitting user entry, boot
code, or fork entry
2) A blind spot in ORC coverage where the unwinder has to bail due to
lack of information about the next frame
The ORC unwinder has no way to tell the difference between the two.
When it encounters an undefined stack state with 'end=1', it blindly
marks the stack reliable, which can break the livepatch consistency
model.
Fix it by splitting UNWIND_HINT_EMPTY into UNWIND_HINT_UNDEFINED and
UNWIND_HINT_END_OF_STACK.
Reported-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Steven Rostedt (Google) <rostedt@goodmis.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/fd6212c8b450d3564b855e1cb48404d6277b4d9f.1677683419.git.jpoimboe@kernel.org
There have been some recently reported ORC unwinder warnings like:
WARNING: can't access registers at entry_SYSCALL_64_after_hwframe+0x63/0xcd
WARNING: stack going in the wrong direction? at __sys_setsockopt+0x2c6/0x5b0 net/socket.c:2271
And a KASAN warning:
BUG: KASAN: stack-out-of-bounds in unwind_next_frame (arch/x86/include/asm/ptrace.h:136 arch/x86/kernel/unwind_orc.c:455)
It turns out the 'signal' bit isn't getting propagated from the unwind
hints to the ORC entries, making the unwinder confused at times.
Fixes: ffb1b4a410 ("x86/unwind/orc: Add 'signal' field to ORC metadata")
Reported-by: kernel test robot <oliver.sang@intel.com>
Reported-by: Dmitry Vyukov <dvyukov@google.com>
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/97eef9db60cd86d376a9a40d49d77bb67a8f6526.1676579666.git.jpoimboe@kernel.org
Andi reported that objtool on vmlinux.o consumes more memory than his
system has, leading to horrific performance.
This is in part because we keep a struct instruction for every
instruction in the file in-memory. Shrink struct instruction by
removing the CFI state (which includes full register state) from it
and demand allocating it.
Given most instructions don't actually change CFI state, there's lots
of repetition there, so add a hash table to find previous CFI
instances.
Reduces memory consumption (and runtime) for processing an
x86_64-allyesconfig:
pre: 4:40.84 real, 143.99 user, 44.18 sys, 30624988 mem
post: 2:14.61 real, 108.58 user, 25.04 sys, 16396184 mem
Suggested-by: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20210624095147.756759107@infradead.org
Instead of manually calling elf_rebuild_reloc_section() on sections
we've called elf_add_reloc() on, have elf_write() DTRT.
This makes it easier to add random relocations in places without
carefully tracking when we're done and need to flush what section.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Borislav Petkov <bp@suse.de>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Miroslav Benes <mbenes@suse.cz>
Link: https://lkml.kernel.org/r/20210326151259.754213408@infradead.org
The ORC unwinder showed a warning [1] which revealed the stack layout
didn't match what was expected. The problem was that paravirt patching
had replaced "CALL *pv_ops.irq.save_fl" with "PUSHF;POP". That changed
the stack layout between the PUSHF and the POP, so unwinding from an
interrupt which occurred between those two instructions would fail.
Part of the agreed upon solution was to rework the custom paravirt
patching code to use alternatives instead, since objtool already knows
how to read alternatives (and converging runtime patching infrastructure
is always a good thing anyway). But the main problem still remains,
which is that runtime patching can change the stack layout.
Making stack layout changes in alternatives was disallowed with commit
7117f16bf4 ("objtool: Fix ORC vs alternatives"), but now that paravirt
is going to be doing it, it needs to be supported.
One way to do so would be to modify the ORC table when the code gets
patched. But ORC is simple -- a good thing! -- and it's best to leave
it alone.
Instead, support stack layout changes by "flattening" all possible stack
states (CFI) from parallel alternative code streams into a single set of
linear states. The only necessary limitation is that CFI conflicts are
disallowed at all possible instruction boundaries.
For example, this scenario is allowed:
Alt1 Alt2 Alt3
0x00 CALL *pv_ops.save_fl CALL xen_save_fl PUSHF
0x01 POP %RAX
0x02 NOP
...
0x05 NOP
...
0x07 <insn>
The unwind information for offset-0x00 is identical for all 3
alternatives. Similarly offset-0x05 and higher also are identical (and
the same as 0x00). However offset-0x01 has deviating CFI, but that is
only relevant for Alt3, neither of the other alternative instruction
streams will ever hit that offset.
This scenario is NOT allowed:
Alt1 Alt2
0x00 CALL *pv_ops.save_fl PUSHF
0x01 NOP6
...
0x07 NOP POP %RAX
The problem here is that offset-0x7, which is an instruction boundary in
both possible instruction patch streams, has two conflicting stack
layouts.
[ The above examples were stolen from Peter Zijlstra. ]
The new flattened CFI array is used both for the detection of conflicts
(like the second example above) and the generation of linear ORC
entries.
BTW, another benefit of these changes is that, thanks to some related
cleanups (new fake nops and alt_group struct) objtool can finally be rid
of fake jumps, which were a constant source of headaches.
[1] https://lkml.kernel.org/r/20201111170536.arx2zbn4ngvjoov7@treble
Cc: Shinichiro Kawasaki <shinichiro.kawasaki@wdc.com>
Signed-off-by: Josh Poimboeuf <jpoimboe@redhat.com>
Decouple ORC entries from instructions. This simplifies the
control/data flow, and is going to make it easier to support alternative
instructions which change the stack layout.
Signed-off-by: Josh Poimboeuf <jpoimboe@redhat.com>
Currently objtool headers are being included either by their base name
or included via ../ from a parent directory. In case of a base name usage:
#include "warn.h"
#include "arch_elf.h"
it does not make it apparent from which directory the file comes from.
To make it slightly better, and actually to avoid name clashes some arch
specific files have "arch_" suffix. And files from an arch folder have
to revert to including via ../ e.g:
#include "../../elf.h"
With additional architectures support and the code base growth there is
a need for clearer headers naming scheme for multiple reasons:
1. to make it instantly obvious where these files come from (objtool
itself / objtool arch|generic folders / some other external files),
2. to avoid name clashes of objtool arch specific headers, potential
obtool arch generic headers and the system header files (there is
/usr/include/elf.h already),
3. to avoid ../ includes and improve code readability.
4. to give a warm fuzzy feeling to developers who are mostly kernel
developers and are accustomed to linux kernel headers arranging
scheme.
Doesn't this make it instantly obvious where are these files come from?
#include <objtool/warn.h>
#include <arch/elf.h>
And doesn't it look nicer to avoid ugly ../ includes? Which also
guarantees this is elf.h from the objtool and not /usr/include/elf.h.
#include <objtool/elf.h>
This patch defines and implements new objtool headers arranging
scheme. Which is:
- all generic headers go to include/objtool (similar to include/linux)
- all arch headers go to arch/$(SRCARCH)/include/arch (to get arch
prefix). This is similar to linux arch specific "asm/*" headers but we
are not abusing "asm" name and calling it what it is. This also helps
to prevent name clashes (arch is not used in system headers or kernel
exports).
To bring objtool to this state the following things are done:
1. current top level tools/objtool/ headers are moved into
include/objtool/ subdirectory,
2. arch specific headers, currently only arch/x86/include/ are moved into
arch/x86/include/arch/ and were stripped of "arch_" suffix,
3. new -I$(srctree)/tools/objtool/include include path to make
includes like <objtool/warn.h> possible,
4. rewriting file includes,
5. make git not to ignore include/objtool/ subdirectory.
Signed-off-by: Vasily Gorbik <gor@linux.ibm.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Masami Hiramatsu <mhiramat@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@redhat.com>
Correct objtool orc generation endianness problems to enable fully
functional x86 cross-compiles on big endian hardware.
Introduce bswap_if_needed() macro, which does a byte swap if target
endianness doesn't match the host, i.e. cross-compilation for little
endian on big endian and vice versa. The macro is used for conversion
of multi-byte values which are read from / about to be written to a
target native endianness ELF file.
Signed-off-by: Vasily Gorbik <gor@linux.ibm.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Masami Hiramatsu <mhiramat@kernel.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@redhat.com>
Unwind hints are useful to provide objtool with information about stack
states in non-standard functions/code.
While the type of information being provided might be very arch
specific, the mechanism to provide the information can be useful for
other architectures.
Move the relevant unwint hint definitions for all architectures to
see.
[ jpoimboe: REGS_IRET -> REGS_PARTIAL ]
Signed-off-by: Julien Thierry <jthierry@redhat.com>
Signed-off-by: Josh Poimboeuf <jpoimboe@redhat.com>
Orc generation is only done for text sections, but some instructions
can be found in non-text sections (e.g. .discard.text sections).
Skip setting their orc sections since their whole sections will be
skipped for orc generation.
Reviewed-by: Miroslav Benes <mbenes@suse.cz>
Signed-off-by: Julien Thierry <jthierry@redhat.com>
Signed-off-by: Josh Poimboeuf <jpoimboe@redhat.com>
