Hello, For a long time, we tackled to fix some issues around kretprobe. One of the latest action was the stacktrace fix on x86 in this thread. https://lore.kernel.org/bpf/162756755600.301564.4957591913842010341.stgit@devnote2/ However, there seems no progress/further discussion. So I would like to make another approach for this (and the other issues.) Here is my idea -- replace kretprobe with kprobe. In other words, put a kprobe on the "return instruction" directly instead of modifying the kernel stack. This can solve most of the kretprobe disadvantges. E.g. - Since it doesn't change the kernel stack, any special stack unwinder fixup is not needed anymore. - No "max-instance" limitations anymore, because it will use kprobes directly. - Scalability performance will be improved as same as kprobes. No list-operation in probe-runtime. Here is a PoC code which introduces "retinsn_probe" event as a part of ftrace kprobe event. I don't think we need to replace the kretprobe. This should be a higher layer feature, because some kernel functions can have multiple "return instructions". Thus, the "retinsn_probe" must manage multiple kprobes. That means the "retinsn_probe" will be a user of kprobes. I decided to make it inside the ftrace "kprobe-event". This gives us another advantage for eBPF support. Because eBPF uses "kprobe-event" instead of "kprobe" directly, if the "retinsn_probe" is implemented in the "kprobe-event", eBPF can use it without any change. Anyway, this can be co-exist with kretprobe. So as far as any user uses kretprobe, we can keep it. Example ======= For example, I ran a shell script, which was used in the stacktrace fix series. ---- mount -t debugfs debugfs /sys/kernel/debug/ cd /sys/kernel/debug/tracing echo > trace echo 1 > options/sym-offset echo r vfs_read >> kprobe_events echo r full_proxy_read >> kprobe_events echo traceoff:1 > events/kprobes/r_vfs_read_0/trigger echo stacktrace:1 > events/kprobes/r_full_proxy_read_0/trigger echo 1 > events/kprobes/enable cat /sys/kernel/debug/kprobes/list echo 0 > events/kprobes/enable cat trace ---- This is the result. ---- ffffffff813b420e k full_proxy_read+0x6e ffffffff812b7c0a k vfs_read+0xda # tracer: nop # # entries-in-buffer/entries-written: 3/3 #P:8 # # _-----=> irqs-off # / _----=> need-resched # | / _---=> hardirq/softirq # || / _--=> preempt-depth # ||| / delay # TASK-PID CPU# |||| TIMESTAMP FUNCTION # | | | |||| | | cat-136 [007] d.Z. 8.038381: r_full_proxy_read_0: (vfs_read+0x9b/0x180 <- full_proxy_read) cat-136 [007] d.Z. 8.038386: <stack trace> => kretprobe_trace_func+0x209/0x300 => retinsn_dispatcher+0x7a/0xa0 => kprobe_post_process+0x28/0x80 => kprobe_int3_handler+0x166/0x1a0 => exc_int3+0x47/0x140 => asm_exc_int3+0x31/0x40 => vfs_read+0x9b/0x180 => ksys_read+0x68/0xe0 => do_syscall_64+0x3b/0x90 => entry_SYSCALL_64_after_hwframe+0x44/0xae cat-136 [007] d.Z. 8.038387: r_vfs_read_0: (ksys_read+0x68/0xe0 <- vfs_read) ---- You can see the return probe events are translated to kprobes instead of kretprobes. And also, on the stacktrace, we can see an int3 calls the kprobe and decode stacktrace correctly. TODO ==== Of course, this is just an PoC code, there are many TODOs. - This PoC code only supports x86 at this moment. But I think this can be done on the other architectures. What it needs is to implement "find_return_instructions()". - Code cleanup is not enough. I have to remove "kretprobe" from "trace_kprobe" data structure, rewrite related functions etc. - It has to handle "tail-call" optimized code, which replaces a "call + return" into "jump". find_return_instruction() should detect it and decode the jump destination too. Thank you, --- Masami Hiramatsu (1): [PoC] tracing: kprobe: Add non-stack intrusion return probe event arch/x86/kernel/kprobes/core.c | 59 +++++++++++++++++++++ kernel/trace/trace_kprobe.c | 110 ++++++++++++++++++++++++++++++++++++++-- 2 files changed, 164 insertions(+), 5 deletions(-) -- Masami Hiramatsu (Linaro) <mhiramat@xxxxxxxxxx>
