Re: [PATCH v3 3/4] arm64: implement live patching
From: Jessica Yu
Date: Thu Oct 18 2018 - 08:58:29 EST
+++ Miroslav Benes [17/10/18 15:39 +0200]:
On Mon, 1 Oct 2018, Torsten Duwe wrote:
Based on ftrace with regs, do the usual thing. Also allocate a
task flag for whatever consistency handling will be used.
Watch out for interactions with the graph tracer.
Similar to what Mark wrote about 2/4, I'd appreciate a better commit log.
Could you explain traditional "what/why/how", please?
Signed-off-by: Torsten Duwe <duwe@xxxxxxx>
--- a/arch/arm64/Kconfig
+++ b/arch/arm64/Kconfig
@@ -119,6 +119,7 @@ config ARM64
select HAVE_GENERIC_DMA_COHERENT
select HAVE_HW_BREAKPOINT if PERF_EVENTS
select HAVE_IRQ_TIME_ACCOUNTING
+ select HAVE_LIVEPATCH
select HAVE_MEMBLOCK
select HAVE_MEMBLOCK_NODE_MAP if NUMA
select HAVE_NMI
@@ -1349,4 +1350,6 @@ if CRYPTO
source "arch/arm64/crypto/Kconfig"
endif
+source "kernel/livepatch/Kconfig"
+
source "lib/Kconfig"
--- a/arch/arm64/include/asm/thread_info.h
+++ b/arch/arm64/include/asm/thread_info.h
@@ -76,6 +76,7 @@ void arch_release_task_struct(struct tas
#define TIF_FOREIGN_FPSTATE 3 /* CPU's FP state is not current's */
#define TIF_UPROBE 4 /* uprobe breakpoint or singlestep */
#define TIF_FSCHECK 5 /* Check FS is USER_DS on return */
+#define TIF_PATCH_PENDING 6
#define TIF_NOHZ 7
#define TIF_SYSCALL_TRACE 8
#define TIF_SYSCALL_AUDIT 9
@@ -94,6 +95,7 @@ void arch_release_task_struct(struct tas
#define _TIF_NEED_RESCHED (1 << TIF_NEED_RESCHED)
#define _TIF_NOTIFY_RESUME (1 << TIF_NOTIFY_RESUME)
#define _TIF_FOREIGN_FPSTATE (1 << TIF_FOREIGN_FPSTATE)
+#define _TIF_PATCH_PENDING (1 << TIF_PATCH_PENDING)
#define _TIF_NOHZ (1 << TIF_NOHZ)
#define _TIF_SYSCALL_TRACE (1 << TIF_SYSCALL_TRACE)
#define _TIF_SYSCALL_AUDIT (1 << TIF_SYSCALL_AUDIT)
@@ -106,7 +108,8 @@ void arch_release_task_struct(struct tas
#define _TIF_WORK_MASK (_TIF_NEED_RESCHED | _TIF_SIGPENDING | \
_TIF_NOTIFY_RESUME | _TIF_FOREIGN_FPSTATE | \
- _TIF_UPROBE | _TIF_FSCHECK)
+ _TIF_UPROBE | _TIF_FSCHECK | \
+ _TIF_PATCH_PENDING)
Could you add a note to the changelog what this means? My ability to read
arm64 entry.S is very limited, but I can see that _TIF_WORK_MASK is
process in a syscall exit and irq return paths. That's good. It is also
called (via "b ret_to_user") in a couple of different places (el0_*
labels). I guess those are returns from exception handling. A comment
about it in the changelog would be appreciated.
#define _TIF_SYSCALL_WORK (_TIF_SYSCALL_TRACE | _TIF_SYSCALL_AUDIT | \
_TIF_SYSCALL_TRACEPOINT | _TIF_SECCOMP | \
--- /dev/null
+++ b/arch/arm64/include/asm/livepatch.h
@@ -0,0 +1,36 @@
+/* SPDX-License-Identifier: GPL-2.0
+ *
+ * livepatch.h - arm64-specific Kernel Live Patching Core
+ *
+ * Copyright (C) 2016,2018 SUSE
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version 2
+ * of the License, or (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, see <http://www.gnu.org/licenses/>.
+ */
+#ifndef _ASM_ARM64_LIVEPATCH_H
+#define _ASM_ARM64_LIVEPATCH_H
+
+#include <linux/module.h>
+#include <linux/ftrace.h>
I think that only
#include <asm/ptrace.h>
is in fact needed, because of "struct pt_regs".
Ad relocations. I checked that everything in struct mod_arch_specific
stays after the module is load. Both core and init get SHF_ALLOC set
(mod->arch.core.plt->sh_flags in module_frob_arch_sections(). It is
important because apply_relocate_add() may use those sections
through module_emit_plt_entry() call.
Yes, it looks like the needed .plt sections will remain in module
memory. However, I think I found a slight issue... :/
In module_frob_arch_sections(), mod->arch.core.plt is set to an offset
within info->sechdrs:
if (!strcmp(secstrings + sechdrs[i].sh_name, ".plt"))
mod->arch.core.plt = sechdrs + i;
sechdrs is from info->sechdrs, which is freed at the end of
load_module() via free_copy(). So although the relevant plt section(s)
are in module memory, mod->arch.core.plt will point to invalid memory
after info is freed. In other words, the section (.plt) *is* in memory
but the section header (struct elf64_shdr) containing section metadata
like sh_addr, sh_size etc., is not.
But we have mod->klp_info->sechdrs (which is a copy of the section
headers) for this very reason. It is initialized only at the very end
of load_module(). I don't think copy_module_elf() is dependent on
anything, so it can just be moved earlier.
Note that we cannot set mod->arch.core.plt to mod->klp_info->sechdrs + i
during module_frob_arch_sections() because it is still too early, none
of the module sections have been copied and none of their sh_addr's
have been set to their final locations as this is all happening before
move_module() is called. So we can use a module_finalize() function
for arm64 to switch the mod->arch.core.plt to use the klp_info section
headers.
Maybe this will be more clear in the example fix below (completely
untested and unreviewed!):
diff --git a/arch/arm64/include/asm/module.h b/arch/arm64/include/asm/module.h
index 97d0ef12e2ff..150afc29171b 100644
--- a/arch/arm64/include/asm/module.h
+++ b/arch/arm64/include/asm/module.h
@@ -25,6 +25,7 @@ struct mod_plt_sec {
struct elf64_shdr *plt;
int plt_num_entries;
int plt_max_entries;
+ int plt_shndx;
};
struct mod_arch_specific {
diff --git a/arch/arm64/kernel/module-plts.c b/arch/arm64/kernel/module-plts.c
index f0690c2ca3e0..c23cef8f0165 100644
--- a/arch/arm64/kernel/module-plts.c
+++ b/arch/arm64/kernel/module-plts.c
@@ -196,6 +196,21 @@ static unsigned int count_plts(Elf64_Sym *syms, Elf64_Rela *rela, int num,
return ret;
}
+int module_finalize(const Elf_Ehdr *hdr, const Elf_Shdr *sechdrs,
+ struct module *mod)
+{
+ /*
+ * Livepatch modules need to keep access to section headers to apply
+ * relocations. Note mod->klp_info should have already been initialized
+ * and all section sh_addr's should have their final addresses by the
+ * time module_finalize() is called.
+ */
+ if (is_livepatch_module(mod))
+ mod->arch.core.plt = mod->klp_info->sechdrs + mod->arch.core.plt_shndx;
+
+ return 0;
+}
+
int module_frob_arch_sections(Elf_Ehdr *ehdr, Elf_Shdr *sechdrs,
char *secstrings, struct module *mod)
{
@@ -210,11 +225,13 @@ int module_frob_arch_sections(Elf_Ehdr *ehdr, Elf_Shdr *sechdrs,
* entries. Record the symtab address as well.
*/
for (i = 0; i < ehdr->e_shnum; i++) {
- if (!strcmp(secstrings + sechdrs[i].sh_name, ".plt"))
+ if (!strcmp(secstrings + sechdrs[i].sh_name, ".plt")) {
mod->arch.core.plt = sechdrs + i;
- else if (!strcmp(secstrings + sechdrs[i].sh_name, ".init.plt"))
+ mod->arch.core.plt_shndx = i;
+ } else if (!strcmp(secstrings + sechdrs[i].sh_name, ".init.plt")) {
mod->arch.init.plt = sechdrs + i;
- else if (IS_ENABLED(CONFIG_DYNAMIC_FTRACE) &&
+ mod->arch.init.plt_shndx = i;
+ } else if (IS_ENABLED(CONFIG_DYNAMIC_FTRACE) &&
!strcmp(secstrings + sechdrs[i].sh_name,
".text.ftrace_trampoline"))
tramp = sechdrs + i;
diff --git a/kernel/module.c b/kernel/module.c
index 6746c85511fe..2fc4d74288dd 100644
--- a/kernel/module.c
+++ b/kernel/module.c
@@ -3363,6 +3363,12 @@ static int post_relocation(struct module *mod, const struct load_info *info)
/* Setup kallsyms-specific fields. */
add_kallsyms(mod, info);
+ if (is_livepatch_module(mod)) {
+ err = copy_module_elf(mod, info);
+ if (err < 0)
+ return err;
+ }
+
/* Arch-specific module finalizing. */
return module_finalize(info->hdr, info->sechdrs, mod);
}
@@ -3775,12 +3781,6 @@ static int load_module(struct load_info *info, const char __user *uargs,
if (err < 0)
goto coming_cleanup;
- if (is_livepatch_module(mod)) {
- err = copy_module_elf(mod, info);
- if (err < 0)
- goto sysfs_cleanup;
- }
-
/* Get rid of temporary copy. */
free_copy(info);
Thoughts? Does the fix make sense?
The last thing is count_plts() function called from
module_frob_arch_sections(). It needed to be changed in 2016 as well. See
Jessica's patch
(20160713001113.GA30925@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx). The logic
in the function has changed since then. If I am not mistaken,
count_plts() is fine as it is right now. It does not consider SHN_UNDEF
anymore, it looks at the destination section (where a symbol should
resolved to) only.
Jessica, could you doublecheck, please?
Yes, I think count_plts() is fine now in this case (because it is
always true that sym->st_shndx != dstidx for SHN_LIVEPATCH symbols)
and my old patch from 2016 is no longer needed.
Thanks,
Jessica