Validate that all indirect calls adhere to kCFI rules. Notably doing
nocfi indirect call to a cfi function is broken.
Apparently some Rust 'core' code violates this and explodes when ran
with FineIBT.
All the ANNOTATE_NOCFI_SYM sites are prime targets for attackers.
- runtime EFI is especially henous because it also needs to disable
IBT. Basically calling unknown code without CFI protection at
runtime is a massice security issue.
- Kexec image handover; if you can exploit this, you get to keep it :-)
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Josh Poimboeuf <jpoimboe@kernel.org>
Acked-by: Sean Christopherson <seanjc@google.com>
Link: https://lkml.kernel.org/r/20250714103441.496787279@infradead.org
After elf_update_group_sh_info() was introduced, a prototype version of
"objtool klp diff" went from taking ~1s to several minutes, due to
looping almost endlessly in elf_update_group_sh_info() while creating
thousands of local symbols in a file with thousands of sections.
Dramatically improve the performance by marking all symbols' correlated
SHT_GROUP sections while reading the object. That way there's no need
to search for it every time a symbol gets reindexed.
Fixes: 2cb291596e ("objtool: Fix up st_info in COMDAT group section")
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Rong Xu <xur@google.com>
Link: https://lkml.kernel.org/r/2a33e583c87e3283706f346f9d59aac20653b7fd.1746662991.git.jpoimboe@kernel.org
After commit a0f7085f6a ("LoongArch: Add RANDOMIZE_KSTACK_OFFSET
support"), there are three new instructions "addi.d $fp, $sp, 32",
"sub.d $sp, $sp, $t0" and "addi.d $sp, $fp, -32" for the secondary
stack in do_syscall(), then there is a objtool warning "return with
modified stack frame" and no handle_syscall() which is the previous
frame of do_syscall() in the call trace when executing the command
"echo l > /proc/sysrq-trigger".
objdump shows something like this:
0000000000000000 <do_syscall>:
0: 02ff8063 addi.d $sp, $sp, -32
4: 29c04076 st.d $fp, $sp, 16
8: 29c02077 st.d $s0, $sp, 8
c: 29c06061 st.d $ra, $sp, 24
10: 02c08076 addi.d $fp, $sp, 32
...
74: 0011b063 sub.d $sp, $sp, $t0
...
a8: 4c000181 jirl $ra, $t0, 0
...
dc: 02ff82c3 addi.d $sp, $fp, -32
e0: 28c06061 ld.d $ra, $sp, 24
e4: 28c04076 ld.d $fp, $sp, 16
e8: 28c02077 ld.d $s0, $sp, 8
ec: 02c08063 addi.d $sp, $sp, 32
f0: 4c000020 jirl $zero, $ra, 0
The instruction "sub.d $sp, $sp, $t0" changes the stack bottom and the
new stack size is a random value, in order to find the return address of
do_syscall() which is stored in the original stack frame after executing
"jirl $ra, $t0, 0", it should use fp which points to the original stack
top.
At the beginning, the thought is tended to decode the secondary stack
instruction "sub.d $sp, $sp, $t0" and set it as a label, then check this
label for the two frame pointer instructions to change the cfa base and
cfa offset during the period of secondary stack in update_cfi_state().
This is valid for GCC but invalid for Clang due to there are different
secondary stack instructions for ClangBuiltLinux on LoongArch, something
like this:
0000000000000000 <do_syscall>:
...
88: 00119064 sub.d $a0, $sp, $a0
8c: 00150083 or $sp, $a0, $zero
...
Actually, it equals to a single instruction "sub.d $sp, $sp, $a0", but
there is no proper condition to check it as a label like GCC, and so the
beginning thought is not a good way.
Essentially, there are two special frame pointer instructions which are
"addi.d $fp, $sp, imm" and "addi.d $sp, $fp, imm", the first one points
fp to the original stack top and the second one restores the original
stack bottom from fp.
Based on the above analysis, in order to avoid adding an arch-specific
update_cfi_state(), we just add a member "frame_pointer" in the "struct
symbol" as a label to avoid affecting the current normal case, then set
it as true only if there is "addi.d $sp, $fp, imm". The last is to check
this label for the two frame pointer instructions to change the cfa base
and cfa offset in update_cfi_state().
Tested with the following two configs:
(1) CONFIG_RANDOMIZE_KSTACK_OFFSET=y &&
CONFIG_RANDOMIZE_KSTACK_OFFSET_DEFAULT=n
(2) CONFIG_RANDOMIZE_KSTACK_OFFSET=y &&
CONFIG_RANDOMIZE_KSTACK_OFFSET_DEFAULT=y
By the way, there is no effect for x86 with this patch, tested on the
x86 machine with Fedora 40 system.
Cc: stable@vger.kernel.org # 6.9+
Signed-off-by: Tiezhu Yang <yangtiezhu@loongson.cn>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
When update the latest upstream gcc and binutils, it generates more
objtool warnings on LoongArch, like this:
init/main.o: warning: objtool: unexpected relocation symbol type in .rela.discard.unreachable
We can see that the reloc sym name is local label instead of section in
relocation section '.rela.discard.unreachable', in this case, the reloc
sym type is STT_NOTYPE instead of STT_SECTION.
As suggested by Peter Zijlstra, we add a "local_label" member in struct
symbol, then set it as true if symbol type is STT_NOTYPE and symbol name
starts with ".L" string in classify_symbols().
Let's check reloc->sym->local_label to not return -1 in add_dead_ends(),
and also use reloc->sym->offset instead of reloc addend which is 0 to
find the corresponding instruction. At the same time, let's replace the
variable "addend" with "offset" to reflect the reality.
Here are some detailed info:
[fedora@linux 6.8.test]$ gcc --version
gcc (GCC) 14.0.1 20240129 (experimental)
[fedora@linux 6.8.test]$ as --version
GNU assembler (GNU Binutils) 2.42.50.20240129
[fedora@linux 6.8.test]$ readelf -r init/main.o | grep -A 2 "rela.discard.unreachable"
Relocation section '.rela.discard.unreachable' at offset 0x6028 contains 1 entry:
Offset Info Type Sym. Value Sym. Name + Addend
000000000000 00d900000063 R_LARCH_32_PCREL 00000000000002c4 .L500^B1 + 0
Signed-off-by: Tiezhu Yang <yangtiezhu@loongson.cn>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
Objtool --rethunk does two things:
- it collects all (tail) call's of __x86_return_thunk and places them
into .return_sites. These are typically compiler generated, but
RET also emits this same.
- it fudges the validation of the __x86_return_thunk symbol; because
this symbol is inside another instruction, it can't actually find
the instruction pointed to by the symbol offset and gets upset.
Because these two things pertained to the same symbol, there was no
pressing need to separate these two separate things.
However, alas, along comes SRSO and more crazy things to deal with
appeared.
The SRSO patch itself added the following symbol names to identify as
rethunk:
'srso_untrain_ret', 'srso_safe_ret' and '__ret'
Where '__ret' is the old retbleed return thunk, 'srso_safe_ret' is a
new similarly embedded return thunk, and 'srso_untrain_ret' is
completely unrelated to anything the above does (and was only included
because of that INT3 vs UD2 issue fixed previous).
Clear things up by adding a second category for the embedded instruction
thing.
Fixes: fb3bd914b3 ("x86/srso: Add a Speculative RAS Overflow mitigation")
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Link: https://lore.kernel.org/r/20230814121148.704502245@infradead.org
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>
The GElf_Rel[a] structs have more similarities than differences. It's
safe to hard-code the assumptions about their shared fields as they will
never change. Consolidate their handling where possible, getting rid of
duplicated code.
Also, at least for now we only ever create rela sections, so simplify
the relocation creation code to be rela-only.
Link: https://lore.kernel.org/r/dcabf6df400ca500ea929f1e4284f5e5ec0b27c8.1685464332.git.jpoimboe@kernel.org
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
- The term "reloc" is overloaded to mean both "an instance of struct
reloc" and "a reloc section". Change the latter to "rsec".
- For variable names, use "sec" for regular sections and "rsec" for rela
sections to prevent them getting mixed up.
- For struct reloc variables, use "reloc" instead of "rel" everywhere
for consistency.
Link: https://lore.kernel.org/r/8b790e403df46f445c21003e7893b8f53b99a6f3.1685464332.git.jpoimboe@kernel.org
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Unreachable instruction warnings are limited to once per object file.
That no longer makes sense for vmlinux validation, which might have
more unreachable instructions lurking in other places. Change it to
once per function.
Note this affects some other (much rarer) non-fatal warnings as well.
In general I think one-warning-per-function makes sense, as related
warnings can accumulate quickly and we want to eventually get back to
failing the build with -Werror anyway.
Reviewed-by: Miroslav Benes <mbenes@suse.cz>
Link: https://lore.kernel.org/r/9d38f881bfc34e031c74e4e90064ccb3e49f599a.1681853186.git.jpoimboe@kernel.org
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
By using calloc() instead of malloc() in a loop, libc does not have to
keep around bookkeeping information for each single structure.
This reduces maximum memory usage while processing vmlinux.o from
3153325 KB to 3035668 KB (-3.7%) on my notebooks "localmodconfig".
Note this introduces memory leaks, because some additional structs get
added to the lists later after reading the symbols and sections from the
original object. Luckily we don't really care about memory leaks in
objtool.
Signed-off-by: Thomas Weißschuh <linux@weissschuh.net>
Link: https://lore.kernel.org/r/20221216-objtool-memory-v2-3-17968f85a464@weissschuh.net
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Pull powerpc updates from Michael Ellerman:
- Add powerpc qspinlock implementation optimised for large system
scalability and paravirt. See the merge message for more details
- Enable objtool to be built on powerpc to generate mcount locations
- Use a temporary mm for code patching with the Radix MMU, so the
writable mapping is restricted to the patching CPU
- Add an option to build the 64-bit big-endian kernel with the ELFv2
ABI
- Sanitise user registers on interrupt entry on 64-bit Book3S
- Many other small features and fixes
Thanks to Aboorva Devarajan, Angel Iglesias, Benjamin Gray, Bjorn
Helgaas, Bo Liu, Chen Lifu, Christoph Hellwig, Christophe JAILLET,
Christophe Leroy, Christopher M. Riedl, Colin Ian King, Deming Wang,
Disha Goel, Dmitry Torokhov, Finn Thain, Geert Uytterhoeven, Gustavo A.
R. Silva, Haowen Bai, Joel Stanley, Jordan Niethe, Julia Lawall, Kajol
Jain, Laurent Dufour, Li zeming, Miaoqian Lin, Michael Jeanson, Nathan
Lynch, Naveen N. Rao, Nayna Jain, Nicholas Miehlbradt, Nicholas Piggin,
Pali Rohár, Randy Dunlap, Rohan McLure, Russell Currey, Sathvika
Vasireddy, Shaomin Deng, Stephen Kitt, Stephen Rothwell, Thomas
Weißschuh, Tiezhu Yang, Uwe Kleine-König, Xie Shaowen, Xiu Jianfeng,
XueBing Chen, Yang Yingliang, Zhang Jiaming, ruanjinjie, Jessica Yu,
and Wolfram Sang.
* tag 'powerpc-6.2-1' of git://git.kernel.org/pub/scm/linux/kernel/git/powerpc/linux: (181 commits)
powerpc/code-patching: Fix oops with DEBUG_VM enabled
powerpc/qspinlock: Fix 32-bit build
powerpc/prom: Fix 32-bit build
powerpc/rtas: mandate RTAS syscall filtering
powerpc/rtas: define pr_fmt and convert printk call sites
powerpc/rtas: clean up includes
powerpc/rtas: clean up rtas_error_log_max initialization
powerpc/pseries/eeh: use correct API for error log size
powerpc/rtas: avoid scheduling in rtas_os_term()
powerpc/rtas: avoid device tree lookups in rtas_os_term()
powerpc/rtasd: use correct OF API for event scan rate
powerpc/rtas: document rtas_call()
powerpc/pseries: unregister VPA when hot unplugging a CPU
powerpc/pseries: reset the RCU watchdogs after a LPM
powerpc: Take in account addition CPU node when building kexec FDT
powerpc: export the CPU node count
powerpc/cpuidle: Set CPUIDLE_FLAG_POLLING for snooze state
powerpc/dts/fsl: Fix pca954x i2c-mux node names
cxl: Remove unnecessary cxl_pci_window_alignment()
selftests/powerpc: Fix resource leaks
...
When moving a symbol in the symtab its index changes and any reloc
referring that symtol-table-index will need to be rewritten too.
In order to facilitate this, objtool simply marks the whole reloc
section 'changed' which will cause the whole section to be
re-generated.
However, finding the relocs that use any given symbol is implemented
rather crudely -- a fully iteration of all sections and their relocs.
Given that some builds have over 20k sections (kallsyms etc..)
iterating all that for *each* symbol moved takes a bit of time.
Instead have each symbol keep a list of relocs that reference it.
This *vastly* improves build times for certain configs.
Reported-by: Borislav Petkov <bp@alien8.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/Y2LlRA7x+8UsE1xf@hirez.programming.kicks-ass.net
When code is compiled with:
-fpatchable-function-entry=${PADDING_BYTES},${PADDING_BYTES}
functions will have PADDING_BYTES of NOP in front of them. Unwinders
and other things that symbolize code locations will typically
attribute these bytes to the preceding function.
Given that these bytes nominally belong to the following symbol this
mis-attribution is confusing.
Inspired by the fact that CFI_CLANG emits __cfi_##name symbols to
claim these bytes, allow objtool to emit __pfx_##name symbols to do
the same.
Therefore add the objtool --prefix=N argument, to conditionally place
a __pfx_##name symbol at N bytes ahead of symbol 'name' when: all
these preceding bytes are NOP and name-N is an instruction boundary.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Yujie Liu <yujie.liu@intel.com>
Link: https://lkml.kernel.org/r/20221028194453.526899822@infradead.org
Due to how gelf_update_sym*() requires an Elf_Data pointer, and how
libelf keeps Elf_Data in a linked list per section,
elf_update_symbol() ends up having to iterate this list on each
update to find the correct Elf_Data for the index'ed symbol.
By allocating one Elf_Data per new symbol, the list grows per new
symbol, giving an effective O(n^2) insertion time. This is obviously
bloody terrible.
Therefore over-allocate the Elf_Data when an extention is needed.
Except it turns out libelf disregards Elf_Scn::sh_size in favour of
the sum of Elf_Data::d_size. IOW it will happily write out all the
unused space and fill it with:
0000000000000000 0 NOTYPE LOCAL DEFAULT UND
entries (aka zeros). Which obviously violates the STB_LOCAL placement
rule, and is a general pain in the backside for not being the desired
behaviour.
Manually fix-up the Elf_Data size to avoid this problem before calling
elf_update().
This significantly improves performance when adding a significant
number of symbols.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Yujie Liu <yujie.liu@intel.com>
Link: https://lkml.kernel.org/r/20221028194453.461658986@infradead.org
The current find_{symbol,func}_containing() functions are broken in
the face of overlapping symbols, exactly the case that is needed for a
new ibt/endbr supression.
Import interval_tree_generic.h into the tools tree and convert the
symbol tree to an interval tree to support proper range stabs.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20220915111146.330203761@infradead.org
Find all the return-thunk sites and record them in a .return_sites
section such that the kernel can undo this.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Borislav Petkov <bp@suse.de>
The '--lto' option is a confusing way of telling objtool to do stack
validation despite it being a linked object. It's no longer needed now
that an explicit '--stackval' option exists. The '--vmlinux' option is
also redundant.
Remove both options in favor of a straightforward '--link' option which
identifies a linked object.
Also, implicitly set '--link' with a warning if the user forgets to do
so and we can tell that it's a linked object. This makes it easier for
manual vmlinux runs.
Signed-off-by: Josh Poimboeuf <jpoimboe@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Miroslav Benes <mbenes@suse.cz>
Link: https://lkml.kernel.org/r/dcd3ceffd15a54822c6183e5766d21ad06082b45.1650300597.git.jpoimboe@redhat.com
There's a fun implementation detail on linking STB_WEAK symbols. When
the linker combines two translation units, where one contains a weak
function and the other an override for it. It simply strips the
STB_WEAK symbol from the symbol table, but doesn't actually remove the
code.
The result is that when objtool is ran in a whole-archive kind of way,
it will encounter *heaps* of unused (and unreferenced) code. All
rudiments of weak functions.
Additionally, when a weak implementation is split into a .cold
subfunction that .cold symbol is left in place, even though completely
unused.
Teach objtool to ignore such rudiments by searching for symbol holes;
that is, code ranges that fall outside the given symbol bounds.
Specifically, ignore a sequence of unreachable instruction iff they
occupy a single hole, additionally ignore any .cold subfunctions
referenced.
Both ld.bfd and ld.lld behave like this. LTO builds otoh can (and do)
properly DCE weak functions.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Josh Poimboeuf <jpoimboe@redhat.com>
Link: https://lore.kernel.org/r/20220308154319.232019347@infradead.org
Teach objtool to turn instrumentation required for memory barrier
modeling into nops in noinstr text.
The __tsan_func_entry/exit calls are still emitted by compilers even
with the __no_sanitize_thread attribute. The memory barrier
instrumentation will be inserted explicitly (without compiler help), and
thus needs to also explicitly be removed.
Signed-off-by: Marco Elver <elver@google.com>
Acked-by: Josh Poimboeuf <jpoimboe@redhat.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Normally objtool will now follow indirect calls; there is no need.
However, this becomes a problem with noinstr validation; if there's an
indirect call from noinstr code, we very much need to know it is to
another noinstr function. Luckily there aren't many indirect calls in
entry code with the obvious exception of paravirt. As such, noinstr
validation didn't work with paravirt kernels.
In order to track pv_ops[] call targets, objtool reads the static
pv_ops[] tables as well as direct assignments to the pv_ops[] array,
provided the compiler makes them a single instruction like:
bf87: 48 c7 05 00 00 00 00 00 00 00 00 movq $0x0,0x0(%rip)
bf92 <xen_init_spinlocks+0x5f>
bf8a: R_X86_64_PC32 pv_ops+0x268
There are, as of yet, no warnings for when this goes wrong :/
Using the functions found with the above means, all pv_ops[] calls are
now subject to noinstr validation.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20210624095149.118815755@infradead.org
Nathan reported that LLVM ThinLTO builds have a performance regression
with commit 25cf0d8aa2 ("objtool: Rewrite hashtable sizing"). Sami
was quick to note that this is due to their use of -ffunction-sections.
As a result the .text section is small and basing the number of relocs
off of that no longer works. Instead have read_sections() compute the
sum of all SHF_EXECINSTR sections and use that.
Fixes: 25cf0d8aa2 ("objtool: Rewrite hashtable sizing")
Reported-by: Nathan Chancellor <nathan@kernel.org>
Debugged-by: Sami Tolvanen <samitolvanen@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Nathan Chancellor <nathan@kernel.org>
Link: https://lkml.kernel.org/r/YMJpGLuGNsGtA5JJ@hirez.programming.kicks-ass.net
Currently objtool has 5 hashtables and sizes them 16 or 20 bits
depending on the --vmlinux argument.
However, a single side doesn't really work well for the 5 tables,
which among them, cover 3 different uses. Also, while vmlinux is
larger, there is still a very wide difference between a defconfig and
allyesconfig build, which again isn't optimally covered by a single
size.
Another aspect is the cost of elf_hash_init(), which for large tables
dominates the runtime for small input files. It turns out that all it
does it assign NULL, something that is required when using malloc().
However, when we allocate memory using mmap(), we're guaranteed to get
zero filled pages.
Therefore, rewrite the whole thing to:
1) use more dynamic sized tables, depending on the input file,
2) avoid the need for elf_hash_init() entirely by using mmap().
This speeds up a regular kernel build (100s to 98s for
x86_64-defconfig), and potentially dramatically speeds up vmlinux
processing.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20210506194157.452881700@infradead.org
