Re: [PATCH RFC tip/core/rcu 09/16] rcu-tasks: Add an RCU-tasks rude variant

From: Steven Rostedt
Date: Mon May 11 2020 - 15:48:30 EST


On Sun, 10 May 2020 17:59:27 +0800
Lai Jiangshan <jiangshanlai@xxxxxxxxx> wrote:

> Hello
>
> I think adding a small number of instructions to preempt_schedule_irq()
> is sufficient to create the needed protected region between the start
> of a function and the trampoline body.
>
> preempt_schedule_irq() {
> + if (unlikely(is_trampoline_page(page_of(interrupted_ip)))) {
> + return; // don't do preempt schedule
> +
> + }
> preempt_schedule_irq() original body
> }
>
> // generated on trampoline pages
> trace_trampoline() {
> preempt_disable();
> trace_trampoline body
> jmp preempt_enable_traced(clobbers)
> }
>
> asm(kernel text):
> preempt_enable_traced:
> preempt_enable_notrace();
> restore cobblers
> return(the return ip on the stack is traced_function_start_code)
>
>
> If the number of instructions added in preempt_schedule_irq() and
> the complexity to make trampoline ip detectable(is_trampoline_page(),
> or is_trampoline_range()) are small, and tasks_rcu is rendered useless,
> I think it will be win-win.

To make this even more complex, with ftrace direct callers (used by bpf to
define their own non ftrace trampoline), if a direct call is on the same
location as a ftrace caller, we have something like this:

ftrace_caller:
save_regs
call ftrace_ops_list_func
cmp ORIG_RAX
jnz do_direct
restore_regs
ret

do_direct:
mov ORIG_RAX to return
restore_regs
ret

What the above is basically doing, is that the ftrace_ops_list_func() will
call the ftrace callbacks, but also a special callback to handle the direct
that is also registered to that same location. The direct callback will
place the address of the direct trampoline into ORIG_RAX. Then on return
from ftrace_ops_list_func(), it will jump directly to the direct caller.

To implement what you are proposing, you have to have a way to keep
preemption off between the setting of ORIG_RAX and the jump to the direct
caller (which would require its own preempt_disable() section). But if we
preempt between the two, the direct trampoline may disappear and then this
code will jump to it.

-- Steve