git: bdf0f24bb16d - main - linux: implement PTRACE_GET_SYSCALL_INFO

[Konstantin Belousov]

On Tue, Sep 14, 2021 at 11:54:08PM +0100, Jessica Clarke wrote:
> On 14 Sep 2021, at 23:45, Konstantin Belousov <kostikbel at gmail.com> wrote:
> > 
> > On Tue, Sep 14, 2021 at 08:41:06PM +0000, Edward Tomasz Napierala wrote:
> >> The branch main has been updated by trasz:
> >> 
> >> URL: https://cgit.FreeBSD.org/src/commit/?id=bdf0f24bb16d556a5b1e01cdfc087d08e91ac572
> >> 
> >> commit bdf0f24bb16d556a5b1e01cdfc087d08e91ac572
> >> Author:     Edward Tomasz Napierala <trasz at FreeBSD.org>
> >> AuthorDate: 2021-09-12 11:31:10 +0000
> >> Commit:     Edward Tomasz Napierala <trasz at FreeBSD.org>
> >> CommitDate: 2021-09-14 20:19:55 +0000
> >> 
> >>    linux: implement PTRACE_GET_SYSCALL_INFO
> >> 
> >>    This is one of the pieces required to make modern (ie Focal)
> >>    strace(1) work.
> >> 
> >>    Reviewed By:    jhb (earlier version)
> >>    Sponsored by:   EPSRC
> >>    Differential Revision:  https://reviews.freebsd.org/D28212
> >> ---
> >> lib/libsysdecode/mktables             |  2 +-
> >> sys/amd64/linux/linux_ptrace.c        | 98 +++++++++++++++++++++++++++++++++--
> >> sys/compat/freebsd32/freebsd32_misc.c |  3 ++
> >> sys/kern/sys_process.c                | 17 ++++++
> >> sys/sys/ptrace.h                      |  4 ++
> >> 5 files changed, 120 insertions(+), 4 deletions(-)
> >> 
> >> +	case PT_GET_SC_ARGS_ALL:
> >> +		CTR1(KTR_PTRACE, "PT_GET_SC_ARGS_ALL: pid %d", p->p_pid);
> >> +		if ((td2->td_dbgflags & (TDB_SCE | TDB_SCX)) == 0
> >> +#ifdef COMPAT_FREEBSD32
> >> +		    || (wrap32 && !safe)
> >> +#endif
> >> +		    ) {
> >> +			error = EINVAL;
> >> +			break;
> >> +		}
> >> +		bcopy(td2->td_sa.args, addr, sizeof(td2->td_sa.args));
> >> +		break;
> > 
> > This is awful, you already got that feedback in review, as I read it.
> > I strongly suggest to remove PT_GET_SC_ARGS_ALL, and instead checks something
> > in the implementation of PT_GET_SC_ARGS to select either full copy or just
> > nargs args.
> > 
> > Easiest thing for 'something' would be SV_PROC_ABI(p) == SV_ABI_LINUX.
> 
> That is incorrect. The original review just changed PT_GET_SC_ARGS even
> for FreeBSD and that was what was described as horrible. John suggested
> two alternatives: this approach, and having the Linuxulator bypass
> kern_ptrace entirely by doing its own thing.
My view is that
- the whole thing with reading past nargs is horrible
- having hidden API interface is horrible, it is obvious layering violation
  when you need to hide an API: the consumer calls at the wrong level.

Another reason to dislike this is that we started to add new PT_ verbs
at relatively high rate recently.  I am aware of at least to more PT_
ops coming in.  In other words, we are not too far from exhausting the
MI range and then we would need to make some arrangements.  Having one
more PT_ value, esp. not useful for anything but for layering violation,
is not good.