Messages in this thread | | | Date | Wed, 16 Mar 2022 10:29:05 -0400 | From | Steven Rostedt <> | Subject | Re: [RFC PATCH -next v2 0/4] arm64/ftrace: support dynamic trampoline |
| |
On Wed, 16 Mar 2022 18:01:28 +0800 Wang ShaoBo <bobo.shaobowang@huawei.com> wrote:
> This implements dynamic trampoline in ARM64, as reference said, we > complete whole design of supporting long jump in dynamic trampoline: > > .text section: > funcA: | funcA(): funcB():| > `-> +-----+ | | ... mov x9 | > | ... | | | adrp <- bl <> | > | nop | | | mov > | nop | | | br x16 ---+ > funcB | nop | | | ftrace_(regs_)caller_tramp: > `-> +-----+ | `--> +---------------------+ > | nop | | | ... | > | nop | | ftrace callsite +---------------------+ > | ... | | `-----> | PLT entry: | > | nop | | | adrp | > | nop | | | add | > funcC: | nop | | ftrace graph callsite | br x16 | > `-> +-----+ | `-----> +---------------------+ > | nop | | | ... | > | nop | | +---------------------+ > > But there is still a tricky problem that is how to adjust tracing ip, > waiting to be solved: > > For ARM64, somecases there may be extra instructions inserted into the > head of tracable functions(but not all) by compiler, for instance BTI[1]. > > This dump vmlinux with CONFIG_BTI=y: > > (1) function gic_handle_irq has bti in its head, so we adjust rec->ip+=5 to last nop > ffffffc0080100e0: d53cd042 mrs x2, tpidr_el2 > ... > ffffffc0080100f0: d503201f nop //__mcount_loc tells the rec->ip > ffffffc0080100f4: d503201f nop > ffffffc0080100f8: d503201f nop > > ffffffc0080100fc <gic_handle_irq>: > ffffffc0080100fc: d503245f bti c > ffffffc008010100: d503201f nop > ffffffc008010104: d503201f nop //we adjust origin rec->ip+5 to here > ffffffc008010108: d503233f paciasp > (2) name_to_dev_t.part.0 do not has bti in its head, so we should adjust rec->ip+=4 to last nop > ffff8000080137d4: d503201f nop > ffff8000080137d8: d503201f nop > ffff8000080137dc: d503201f nop > > ffff8000080137e0 <name_to_dev_t.part.0>: > ffff8000080137e0: d503201f nop > ffff8000080137e4: d503201f nop > ffff8000080137e8: d503233f paciasp > > So at this time we have no idea to identify rec->ip for each tracable function.
This looks like the same issue that Peter Zijlstra is handling for IBT on x86, which I think can be useful for you too.
https://lore.kernel.org/all/20220308153011.021123062@infradead.org/
Specifically this patch:
https://lore.kernel.org/all/20220308154318.227581603@infradead.org/
Which modifies the ftrace_location() to return the rec->ip if you pass in the ip of the function and kallsyms returns that the ip passed in has an offset of zero.
-- Steve
> > we are looking forward to follow-up discussions. > > References: > [1] https://developer.arm.com/documentation/100076/0100/a64-instruction-set-reference/a64-general-instructions/bti > [2] https://lore.kernel.org/linux-arm-kernel/20200109142736.1122-1-cj.chengjian@huawei.com/ > > Cheng Jian (4): > arm64: introduce aarch64_insn_gen_load_literal > arm64/ftrace: introduce ftrace dynamic trampoline entrances > arm64/ftrace: support dynamically allocated trampolines > arm64/ftrace: implement long jump for dynamic trampolines > > arch/arm64/Makefile | 2 +- > arch/arm64/include/asm/ftrace.h | 10 +- > arch/arm64/include/asm/insn.h | 6 + > arch/arm64/include/asm/module.h | 9 + > arch/arm64/kernel/entry-ftrace.S | 88 ++++++-- > arch/arm64/kernel/ftrace.c | 366 ++++++++++++++++++++++++++++--- > arch/arm64/kernel/module-plts.c | 50 +++++ > arch/arm64/lib/insn.c | 49 +++++ > 8 files changed, 532 insertions(+), 48 deletions(-) >
| |