- Build footprint & performance improvements: - Reduce memory usage with CONFIG_DEBUG_INFO=y In the worst case of an allyesconfig+CONFIG_DEBUG_INFO=y kernel, DWARF creates almost 200 million relocations, ballooning objtool's peak heap usage to 53GB. These patches reduce that to 25GB. On a distro-type kernel with kernel IBT enabled, they reduce objtool's peak heap usage from 4.2GB to 2.8GB. These changes also improve the runtime significantly. - Debuggability improvements: - Add the unwind_debug command-line option, for more extend unwinding debugging output. - Limit unreachable warnings to once per function - Add verbose option for disassembling affected functions - Include backtrace in verbose mode - Detect missing __noreturn annotations - Ignore exc_double_fault() __noreturn warnings - Remove superfluous global_noreturns entries - Move noreturn function list to separate file - Add __kunit_abort() to noreturns - Unwinder improvements: - Allow stack operations in UNWIND_HINT_UNDEFINED regions - drm/vmwgfx: Add unwind hints around RBP clobber - Cleanups: - Move the x86 entry thunk restore code into thunk functions - x86/unwind/orc: Use swap() instead of open coding it - Remove unnecessary/unused variables - Fixes for modern stack canary handling Signed-off-by: Ingo Molnar <mingo@kernel.org> -----BEGIN PGP SIGNATURE----- iQJFBAABCgAvFiEEBpT5eoXrXCwVQwEKEnMQ0APhK1gFAmSaxcoRHG1pbmdvQGtl cm5lbC5vcmcACgkQEnMQ0APhK1ht5w//f8mBoABct29pS4ib6pDwRZQDoG8fCA7M +KWjFD1AhX7RsJVEbM4uBUXdSWZD61xxIa8p8LO2jjzE5RyhM+EuNaisKujKqmfj uQTSnRhIRHMPqqVGK/gQxy1v4+3+12O32XFIJhAPYCp/dpbZJ2yKDsiHjapzZTDy BM+86hbIyHFmSl5uJcBFHEv6EGhoxwdrrrOxhpao1CqfAUi+uVgamHGwVqx+NtTY MvOmcy3/0ukHwDLON0MIMu9MSwvnXorD7+RSkYstwAM/k6ao/k78iJ31sOcynpRn ri0gmfygJsh2bxL4JUlY4ZeTs7PLWkj3i60deePc5u6EyV4JDJ2borUibs5oGoF6 pN0AwbtubLHHhUI/v74B3E6K6ZGvLiEn9dsNTuXsJffD+qU2REb+WLhr4ut+E1Wi IKWrYh811yBLyOqFEW3XudZTiXSJlgi3eYiCxspEsKw2RIFFt2g6vYcwrIb0Hatw 8R4/jCWk1nc6Wa3RQYsVnhkglAECSKQdDfS7p2e1hNUTjZuess4EEJjSLs8upIQ9 D1bmuUxEzRxVwAZtXYNh0NKe7OtyOrqgsVTQuqxvWXq2CpC7Hqj8piVJWHdBWgHO 0o2OQqjwSrzAtevpAIaYQv9zhPs1hV7CpBgzzqWGXrwJ3vM6YoSRLf0bg+5OkN8I O4U2xq2OVa8= =uNnc -----END PGP SIGNATURE----- Merge tag 'objtool-core-2023-06-27' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip Pull objtool updates from Ingo Molar: "Build footprint & performance improvements: - Reduce memory usage with CONFIG_DEBUG_INFO=y In the worst case of an allyesconfig+CONFIG_DEBUG_INFO=y kernel, DWARF creates almost 200 million relocations, ballooning objtool's peak heap usage to 53GB. These patches reduce that to 25GB. On a distro-type kernel with kernel IBT enabled, they reduce objtool's peak heap usage from 4.2GB to 2.8GB. These changes also improve the runtime significantly. Debuggability improvements: - Add the unwind_debug command-line option, for more extend unwinding debugging output - Limit unreachable warnings to once per function - Add verbose option for disassembling affected functions - Include backtrace in verbose mode - Detect missing __noreturn annotations - Ignore exc_double_fault() __noreturn warnings - Remove superfluous global_noreturns entries - Move noreturn function list to separate file - Add __kunit_abort() to noreturns Unwinder improvements: - Allow stack operations in UNWIND_HINT_UNDEFINED regions - drm/vmwgfx: Add unwind hints around RBP clobber Cleanups: - Move the x86 entry thunk restore code into thunk functions - x86/unwind/orc: Use swap() instead of open coding it - Remove unnecessary/unused variables Fixes for modern stack canary handling" * tag 'objtool-core-2023-06-27' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (42 commits) x86/orc: Make the is_callthunk() definition depend on CONFIG_BPF_JIT=y objtool: Skip reading DWARF section data objtool: Free insns when done objtool: Get rid of reloc->rel[a] objtool: Shrink elf hash nodes objtool: Shrink reloc->sym_reloc_entry objtool: Get rid of reloc->jump_table_start objtool: Get rid of reloc->addend objtool: Get rid of reloc->type objtool: Get rid of reloc->offset objtool: Get rid of reloc->idx objtool: Get rid of reloc->list objtool: Allocate relocs in advance for new rela sections objtool: Add for_each_reloc() objtool: Don't free memory in elf_close() objtool: Keep GElf_Rel[a] structs synced objtool: Add elf_create_section_pair() objtool: Add mark_sec_changed() objtool: Fix reloc_hash size objtool: Consolidate rel/rela handling ...
139 lines
4.4 KiB
C
139 lines
4.4 KiB
C
// SPDX-License-Identifier: GPL-2.0-or-later
|
|
#include <string.h>
|
|
|
|
#include <objtool/special.h>
|
|
#include <objtool/builtin.h>
|
|
|
|
#define X86_FEATURE_POPCNT (4 * 32 + 23)
|
|
#define X86_FEATURE_SMAP (9 * 32 + 20)
|
|
|
|
void arch_handle_alternative(unsigned short feature, struct special_alt *alt)
|
|
{
|
|
switch (feature) {
|
|
case X86_FEATURE_SMAP:
|
|
/*
|
|
* If UACCESS validation is enabled; force that alternative;
|
|
* otherwise force it the other way.
|
|
*
|
|
* What we want to avoid is having both the original and the
|
|
* alternative code flow at the same time, in that case we can
|
|
* find paths that see the STAC but take the NOP instead of
|
|
* CLAC and the other way around.
|
|
*/
|
|
if (opts.uaccess)
|
|
alt->skip_orig = true;
|
|
else
|
|
alt->skip_alt = true;
|
|
break;
|
|
case X86_FEATURE_POPCNT:
|
|
/*
|
|
* It has been requested that we don't validate the !POPCNT
|
|
* feature path which is a "very very small percentage of
|
|
* machines".
|
|
*/
|
|
alt->skip_orig = true;
|
|
break;
|
|
default:
|
|
break;
|
|
}
|
|
}
|
|
|
|
bool arch_support_alt_relocation(struct special_alt *special_alt,
|
|
struct instruction *insn,
|
|
struct reloc *reloc)
|
|
{
|
|
return true;
|
|
}
|
|
|
|
/*
|
|
* There are 3 basic jump table patterns:
|
|
*
|
|
* 1. jmpq *[rodata addr](,%reg,8)
|
|
*
|
|
* This is the most common case by far. It jumps to an address in a simple
|
|
* jump table which is stored in .rodata.
|
|
*
|
|
* 2. jmpq *[rodata addr](%rip)
|
|
*
|
|
* This is caused by a rare GCC quirk, currently only seen in three driver
|
|
* functions in the kernel, only with certain obscure non-distro configs.
|
|
*
|
|
* As part of an optimization, GCC makes a copy of an existing switch jump
|
|
* table, modifies it, and then hard-codes the jump (albeit with an indirect
|
|
* jump) to use a single entry in the table. The rest of the jump table and
|
|
* some of its jump targets remain as dead code.
|
|
*
|
|
* In such a case we can just crudely ignore all unreachable instruction
|
|
* warnings for the entire object file. Ideally we would just ignore them
|
|
* for the function, but that would require redesigning the code quite a
|
|
* bit. And honestly that's just not worth doing: unreachable instruction
|
|
* warnings are of questionable value anyway, and this is such a rare issue.
|
|
*
|
|
* 3. mov [rodata addr],%reg1
|
|
* ... some instructions ...
|
|
* jmpq *(%reg1,%reg2,8)
|
|
*
|
|
* This is a fairly uncommon pattern which is new for GCC 6. As of this
|
|
* writing, there are 11 occurrences of it in the allmodconfig kernel.
|
|
*
|
|
* As of GCC 7 there are quite a few more of these and the 'in between' code
|
|
* is significant. Esp. with KASAN enabled some of the code between the mov
|
|
* and jmpq uses .rodata itself, which can confuse things.
|
|
*
|
|
* TODO: Once we have DWARF CFI and smarter instruction decoding logic,
|
|
* ensure the same register is used in the mov and jump instructions.
|
|
*
|
|
* NOTE: RETPOLINE made it harder still to decode dynamic jumps.
|
|
*/
|
|
struct reloc *arch_find_switch_table(struct objtool_file *file,
|
|
struct instruction *insn)
|
|
{
|
|
struct reloc *text_reloc, *rodata_reloc;
|
|
struct section *table_sec;
|
|
unsigned long table_offset;
|
|
|
|
/* look for a relocation which references .rodata */
|
|
text_reloc = find_reloc_by_dest_range(file->elf, insn->sec,
|
|
insn->offset, insn->len);
|
|
if (!text_reloc || text_reloc->sym->type != STT_SECTION ||
|
|
!text_reloc->sym->sec->rodata)
|
|
return NULL;
|
|
|
|
table_offset = reloc_addend(text_reloc);
|
|
table_sec = text_reloc->sym->sec;
|
|
|
|
if (reloc_type(text_reloc) == R_X86_64_PC32)
|
|
table_offset += 4;
|
|
|
|
/*
|
|
* Make sure the .rodata address isn't associated with a
|
|
* symbol. GCC jump tables are anonymous data.
|
|
*
|
|
* Also support C jump tables which are in the same format as
|
|
* switch jump tables. For objtool to recognize them, they
|
|
* need to be placed in the C_JUMP_TABLE_SECTION section. They
|
|
* have symbols associated with them.
|
|
*/
|
|
if (find_symbol_containing(table_sec, table_offset) &&
|
|
strcmp(table_sec->name, C_JUMP_TABLE_SECTION))
|
|
return NULL;
|
|
|
|
/*
|
|
* Each table entry has a rela associated with it. The rela
|
|
* should reference text in the same function as the original
|
|
* instruction.
|
|
*/
|
|
rodata_reloc = find_reloc_by_dest(file->elf, table_sec, table_offset);
|
|
if (!rodata_reloc)
|
|
return NULL;
|
|
|
|
/*
|
|
* Use of RIP-relative switch jumps is quite rare, and
|
|
* indicates a rare GCC quirk/bug which can leave dead
|
|
* code behind.
|
|
*/
|
|
if (reloc_type(text_reloc) == R_X86_64_PC32)
|
|
file->ignore_unreachables = true;
|
|
|
|
return rodata_reloc;
|
|
}
|