[RFC][PATCH v5 00/51] objtool: Make recordmcount a subcommand

From: Matt Helsley
Date: Thu Jun 18 2020 - 16:39:02 EST

Next message: Matt Helsley: "[RFC][PATCH v5 01/51] objtool: Factor out reasons to build objtool"
Previous message: Matthew Wilcox: "Re: [PATCH v2 09/16] rcu/tree: Maintain separate array for vmalloc ptrs"
Next in thread: Matt Helsley: "[RFC][PATCH v5 01/51] objtool: Factor out reasons to build objtool"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

recordmcount has its own ELF wrapper code and could utilize
objtool's ELF code to more-portably handle architecture variations.
This series makes recordmcount a subcommand of objtool. It very
gradually converts recordmcount to become a subcommand of objtool and
then reuses parts of objtool's ELF code.

recordmcount maps the file in and collects simple information it needs to
append a section to the object file. The only part of the original file it
modifies is the address of new section tables -- interestingly enough this
resembles RCU in that we don't really trim the old tables so
much as unlink them via a critical offset and then rely on
future tooling, in this case, to drop the unused bits.

Much of the recordmcount ELF code is only reading and walking the data
structures to collect the mcount locations it records in a separate
area of memory. This means it's safe to mix access to the mapped
file with access to the objtool-style linked data
structures as we gradually convert it to using only the linked data
structures. Once the old ELF code is no longer in use we can drop it
and use objtool to take over the task of writing the results without
using the RCU-like trick any more.

After that we greatly simplify the mcount subcommand by adding a
few flags to the ELF reading code in objtool. Overall the series
removes about 600 lines of recordmcount while adding little to
objtool's ELF code.

Testing so far:

I've been using scripts to test cross compilation and execution of
objtool, and mcount on objects built for x86, ppc64le, arm64, s390, and
sparc.

Applies on top of:
objtool/core

Peter Zijlstra's "x86/entry: noinstr fixes" [2]

Sami Tolvanen's patch enabling support for more than 64k
sections in recordmcount, already going upstream. [3]

Changes since v4[1]:
v5:
Updated the HAVE_C_RECORDMCOUNT Kconfig description.

Removed unused extern is_cmd_mcount_available() spotted by Julien Thierry.

Issue with symbol info handling reported and fixed by Kamalesh Babulal.

Removed the giant per-supported-arch series of ifdef blocks in
the objtool Makefile and replaced with new CONFIG_ variables
as suggested by Julien Thierry.

Moved the first pass through the sections into objtool's ELF
code as suggested by Peter Zijlstra.

Removed a redundant section lookup.

Removed old comments that no longer apply since conversion to objtool.

v4:
Split out recordmcount cleanups and upstreamed. [5]

Split out, iterated on objtool multi-arch support, and upstreamed. [6]

Split out expanded relocation support, renamed types, and functions
to reflect expanded relocation support, and upstreamed. [4]

This set is based on the patches sent upstream and posted above.

Adapted to renames by Ingo and Peter: s/elf_open/elf_open_read/

Added weak symbols for mcount subcommand
This nicely eliminated the need for the mcount.h header.

Added tools/objtool/Makefile per-arch SUBCMD_ blocks for each
arch recordmcount / mcount supports.

Moved ftrace/mcount/record.h from objtool_dep to recordmcount_dep
This keeps the dependencies better organized.

Fixed Makefile issue reported for PowerPC and a couple other archs
by kbuild test robot. The always-$(BUILD_C_RECORDMCOUNT)
line wasn't sufficiently replaced. Added to prepare-objtool
target in top level Makefile.

Split up dependencies to be independent of CONFIG_STACK_VALIDATION
and CONFIG_UNWINDER_ORC since these are x86-specific.
Now any arch which uses the C version of recordmcount
will build objtool if dynamic tracing is enabled.

Added a second rename at the end to be consistent with other
objtool subcommands.

v3:
Rebased on mainline. s/elf_open/elf_read/ in recordmcount.c

v2:
Fix whitespace before line continuation

Add ftrace/mcount/record.h to objtool_dep

Rename the Makefile variable BUILD_C_RECORDMCOUNT to
better reflect its purpose

Similar: rename recordmcount_source => recordmcount_dep
When using objtool we can just depend on the
binary rather than the source the binary is
built from. This should address Josh's feedback and
make the Makefile code a bit clearer

Add a comment to make reading the Makefile a little
easier

Rebased to latest mainline -rc

[1] https://lore.kernel.org/lkml/cover.1591125127.git.mhelsley@xxxxxxxxxx/
[2] https://lore.kernel.org/lkml/20200618144407.520952071@xxxxxxxxxxxxx
[3] https://lore.kernel.org/lkml/20200424193046.160744-1-samitolvanen@xxxxxxxxxx
[4] https://lore.kernel.org/lkml/cover.1590785960.git.mhelsley@xxxxxxxxxx/
[5] https://lore.kernel.org/lkml/20190802134712.2d8cc63f@xxxxxxxxxxxxxxxxxx/
[6] https://lore.kernel.org/lkml/cover.1586468801.git.mhelsley@xxxxxxxxxx/

Matt Helsley (51):
objtool: Factor out reasons to build objtool
objtool: Prepare to merge recordmcount
objtool: Make recordmcount into mcount subcmd
objtool: recordmcount: Start using objtool's elf wrapper
objtool: recordmcount: Search for __mcount_loc before walking the
sections
objtool: recordmcount: Convert do_func() relhdrs
objtool: mcount: Move nhdr into find_symtab()
objtool: mcount: Remove unused fname parameter
objtool: mcount: Use libelf for section header names
objtool: mcount: Walk objtool Elf structs in find_secsym_ndx
objtool: mcount: Use symbol structs to find mcount relocations
objtool: mcount: Walk relocation lists
objtool: mcount: Return symbol from mcountsym
objtool: mcount: Move get_mcountsym
objtool: mcount: Replace MIPS offset types
objtool: mcount: Move is_fake_mcount()
objtool: mcount: Stop using ehdr in find_section_sym_index
objtool: mcount: Move find_section_sym_index()
objtool: mcount: Restrict using ehdr in append_func()
objtool: mcount: Use objtool ELF to write
objtool: mcount: Move nop_mcount()
objtool: mcount: Move has_rel_mcount() and tot_relsize()
objtool: mcount: Move relocation entry size detection
objtool: mcount: Only keep ELF file size
objtool: mcount: Use ELF header from objtool
objtool: mcount: Remove unused file mapping
objtool: mcount: Reduce usage of _size wrapper
objtool: mcount: Move mcount_adjust out of wrapper
objtool: mcount: Pre-allocate new ELF sections
objtool: mcount: Generic location and relocation table types
objtool: mcount: Use objtool relocation section
objtool: mcount: Move sift_rel_mcount out of wrapper file
objtool: mcount: Remove wrapper for ELF relocation type
objtool: mcount: Remove wrapper double-include trick
objtool: mcount: Remove endian wrappers
objtool: mcount: Rename to mcount.c
objtool: mcount: Simplify mcount name matching
objtool: mcount: mcount symbol name simplification
objtool: mcount: Verify x86 instruction with memcmp()
objtool: mcount: const-ify ARM instruction patterns
objtool: mcount: Convert nop writes to elf_write_insn()
objtool: mcount: Move mcount symbol name testing
objtool: check: Use class to recognize kcov calls
objtool: mcount: Keep lists locations and relocations
objtool: mcount: Move mcount section test to objtool ELF
objtool: mcount: Flag mcount relocation sections
objtool: mcount: Merge section mcount flags
objtool: mcount: Eliminate first pass
objtool: mcount: Remove relocation size check
objtool: mcount: Remove useless lookup
objtool: mcount: Remove stale description

Documentation/dontdiff | 2 +-
Documentation/trace/ftrace-design.rst | 4 +-
Documentation/trace/ftrace.rst | 6 +-
Makefile | 39 +-
arch/Kconfig | 23 +
arch/arm64/include/asm/ftrace.h | 2 +-
arch/x86/Kconfig.debug | 1 +
arch/x86/include/asm/ftrace.h | 2 +-
kernel/trace/Kconfig | 9 +-
lib/Kconfig.debug | 1 +
scripts/.gitignore | 1 -
scripts/Makefile | 1 -
scripts/Makefile.build | 22 +-
scripts/recordmcount.c | 663 --------------------
scripts/recordmcount.h | 692 ---------------------
scripts/sorttable.h | 2 +-
tools/objtool/.gitignore | 1 +
tools/objtool/Build | 2 +
tools/objtool/Makefile | 14 +-
tools/objtool/builtin-mcount.c | 50 ++
tools/objtool/builtin.h | 1 +
tools/objtool/check.c | 3 +-
tools/objtool/elf.c | 60 +-
tools/objtool/elf.h | 12 +-
tools/objtool/mcount.c | 596 ++++++++++++++++++
tools/objtool/objtool.c | 1 +
tools/objtool/objtool.h | 1 +
{scripts => tools/objtool}/recordmcount.pl | 0
tools/objtool/weak.c | 5 +
29 files changed, 806 insertions(+), 1410 deletions(-)
delete mode 100644 scripts/recordmcount.c
delete mode 100644 scripts/recordmcount.h
create mode 100644 tools/objtool/builtin-mcount.c
create mode 100644 tools/objtool/mcount.c
rename {scripts => tools/objtool}/recordmcount.pl (100%)

base-commit: 14bda4e5293ed9722f1dc39b543024e37707d6c6
prerequisite-patch-id: b9f6483185e4e80fccbb6ec7276bf61ce0329472
prerequisite-patch-id: 6d3fec64974cdce64fec35141ff71f0532cb8c82
prerequisite-patch-id: b71e084f3d2adeb4d05e4327183b75b388a20e6d
prerequisite-patch-id: 261f5d8102f4ecee86df2eb93be6a29495702717
prerequisite-patch-id: 64fa3e2f92e5dfd1b64289b4118288b6d2a25a67
prerequisite-patch-id: e5671d612490912b263505907657f3197c4a0833
prerequisite-patch-id: cba407d3dc82ad54095a9b5149b0c851c4e30b19
prerequisite-patch-id: bcf669a844506eec601cba1797f954e59cc7a2ad
--
2.20.1

Next message: Matt Helsley: "[RFC][PATCH v5 01/51] objtool: Factor out reasons to build objtool"
Previous message: Matthew Wilcox: "Re: [PATCH v2 09/16] rcu/tree: Maintain separate array for vmalloc ptrs"
Next in thread: Matt Helsley: "[RFC][PATCH v5 01/51] objtool: Factor out reasons to build objtool"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]