Re: [PATCH 1/2] alpha/ptrace: Record and handle the absence of switch_stack

From: Al Viro
Date: Sun Jun 20 2021 - 23:44:45 EST


On Mon, Jun 21, 2021 at 03:18:35PM +1200, Michael Schmitz wrote:

> This is what I get from WARN_ONCE:
>
> ------------[ cut here ]------------
> WARNING: CPU: 0 PID: 1177 at arch/m68k/kernel/ptrace.c:91 get_reg+0x90/0xb8
> Modules linked in:
> CPU: 0 PID: 1177 Comm: strace Not tainted 5.13.0-rc1-atari-fpuemu-exitfix+
> #1146
> Stack from 014b7f04:
> 014b7f04 00336401 00336401 000278f0 0032c015 0000005b 00000005
> 0002795a
> 0032c015 0000005b 0000338c 00000009 00000000 00000000 ffffffe4
> 00000005
> 00000003 00000014 00000003 00000014 efc2b90c 0000338c 0032c015
> 0000005b
> 00000009 00000000 efc2b908 00912540 efc2b908 000034cc 00912540
> 00000005
> 00000000 efc2b908 00000003 00912540 8000110c c010b0a4 efc2b90c
> 0002d1d8
> 00912540 00000003 00000014 efc2b908 0000049a 00000014 efc2b908
> 800acaa8
> Call Trace: [<000278f0>] __warn+0x9e/0xb4
> [<0002795a>] warn_slowpath_fmt+0x54/0x62
> [<0000338c>] get_reg+0x90/0xb8
> [<0000338c>] get_reg+0x90/0xb8
> [<000034cc>] arch_ptrace+0x7e/0x250
> [<0002d1d8>] sys_ptrace+0x232/0x2f8
> [<00002ab6>] syscall+0x8/0xc
> [<0000c00b>] lower+0x7/0x20
>
> ---[ end trace ee4be53b94695793 ]---
>
> Syscall numbers are actually 90 and 192 - sys_old_mmap and sys_mmap2 on
> m68k. Used the calculator on my Ubuntu desktop, that appears to be a little
> confused about hex to decimal conversions.
>
> I hope that makes more sense?

Not really; what is the condition you are checking? The interesting trace
is not that with get_reg() - it's that of the process being traced. You
are not accessing the stack of caller of ptrace(2) here, so you want to
know that SAVE_SWITCH_STACK had been done by the tracee, not tracer.

And if that had been strace ls, you have TIF_SYSCALL_TRACE set for ls, so
* ls hits system_call
* notices TIF_SYSCALL_TRACE and goes to do_trace_entry
* does SAVE_SWITCH_STACK there
* calls syscall_trace(), which calls ptrace_notify()
* ptrace_notify() calls ptrace_do_notify(), which calls ptrace_stop()
* ptrace_stop() arranges for tracer to be woken up and gives CPU up,
with TASK_TRACED as process state.

That's the callchain in ls, and switch_stack accessed by get_reg() from
strace is the one on ls(1) stack created by SAVE_SWITCH_STACK.