lkml.org 
[lkml]   [2011]   [Oct]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v5 3.1.0-rc4-tip 26/26] uprobes: queue signals while thread is singlestepping.
> 
> Yes, yes, I see. But, once again, this can only protect from kill().
>
> The task can stop even if you block all signals, another thread
> can initiate the group stop and set JOBCTL_STOP_PENDING + TIF_SIGPENDING.
> And note that it can stop _before_ it returns to user mode to step
> over the xol insn.
>
> In theory the tasks like this can consume all slots, and if we have
> yet another thread waiting in xol_wait_event(), we deadlock. Although
> in this case SIGCONT helps, but this group stop can never finish.
>

Okay.

> > > Another problem is that it is not possible to block the "implicit"
> > > SIGKILL sent by exec/exit_group/etc. This mean the task can exit
> > > without sstep_complete/xol_free_insn_slot/etc. Mostly this is fine,
> > > we have free_uprobe_utask()->xol_free_insn_slot(). But in theory
> > > this can deadlock afaics. Suppose that the coredumping is in progress,
> > > the killed UTASK_SSTEP task hangs in exit_mm() waiting for other
> > > threads. If we have enough threads like this, we can deadlock with
> > > another thread sleeping in xol_wait_event().
> >
> > Shouldnt the behaviour be the same as threads that did a
> > select,sigsuspend?
>
> Hmm. I don't understand... Could you explain?
>
> Firstly, select/sigsuspend can't block SIGKILL, but this doesn't matter.
> My point was, the task can exit in UTASK_SSTEP state, and without
> xol_free_insn_slot(). And this (in theory) can lead to the "real"
> deadlock.

I think we should be okay if the test exits in UTASK_SSTEP state.
All I thought we needed to do was block it from doing anything except
exit or singlestep. Our exit hook should cleanup any references that we
hold.

>
> > > However the first problem looks nasty, even if it is not very serious.
> > > And, otoh, it doesn't look right to block SIGKILL, the task can loop
> > > forever executing the xol insn (see below).
> > >
> > >
> > >
> > > What do you think about the patch below? On top of 25/26, uncompiled,
> > > untested. With this patch the task simply refuses to react to
> > > TIF_SIGPENDING until sstep_complete().
> > >
> >
> > Your patch looks very simple and clean.
> > Will test this patch and revert.
>
> Great. I'll think a bit more and send you the "final" version tomorrow.
> Assuming we can change sstep_complete() as we discussed, it doesn't need
> fatal_signal_pending().

Okay.

>
> HOWEVER. There is yet another problem. Another thread can, say, unmap()
> xol_vma. In this case we should ensure that the task can't fault in an
> endless loop.
>

Hmm should we add a check in unmap() to see if the vma that we are
trying to unmap is the xol_vma and if so return?
Our assumption has been that once an xol_vma has been created, it should
be around till the process gets killed.

> > > And, unless I missed something, this makes
> > > free_uprobe_utask()->xol_free_insn_slot() unnecessary.
> >
> > What if a fatal (SIGKILL) signal was delivered only to that thread
>
> this is not possible, in this case all threads are killed. But,
>
> > or a fatal signal for a thread-group but more
> > than one thread-group share the mm?
>
> Yes, this is possible.
>
> Sorry for confusion. Yes, if we have the fatal_signal_pending() check
> in sstep_complete(), then we do need
> free_uprobe_utask()->xol_free_insn_slot(). But this check was added
> only to illustrate another problem with the self-repeating insns.
>
> And. With "HOWEVER" above, we probably need this xol_free anyway.
>
> > you have already commented why blocking signals is a problem, but I
> > still thought I will post the patch that I had to let you know what I
> > was thinking before I saw your patch.
> >
> > While task is processing a singlestep due to uprobes breakpoint hit,
> > block signals from the time it enables singlestep to the time it disables
> > singlestep.
>
> OK, it is too late for me today, I'll take a look tomorrow.
>
> This approach has some advantages too, perhaps we should make something
> "in between".
>

Okay.

--
Thanks and Regards
Srikar


\
 
 \ /
  Last update: 2011-10-12 14:23    [W:0.120 / U:1.868 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site