Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Fri, 29 May 2020 12:44:33 -0500
From:   Josh Poimboeuf <jpoimboe@redhat.com>
To:     Wang ShaoBo <bobo.shaobowang@huawei.com>
Cc:     huawei.libin@huawei.com, xiexiuqi@huawei.com,
        cj.chengjian@huawei.com, mingo@redhat.com, x86@kernel.org,
        linux-kernel@vger.kernel.org, live-patching@vger.kernel.org,
        mbenes@suse.cz, devel@etsukata.com, viro@zeniv.linux.org.uk,
        esyr@redhat.com
Subject: Re: Question: livepatch failed for new fork() task stack unreliable
Message-ID: <20200529174433.wpkknhypx2bmjika@treble>
References: <20200529101059.39885-1-bobo.shaobowang@huawei.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
In-Reply-To: <20200529101059.39885-1-bobo.shaobowang@huawei.com>
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Fri, May 29, 2020 at 06:10:59PM +0800, Wang ShaoBo wrote:
> Stack unreliable error is reported by stack_trace_save_tsk_reliable() when trying
> to insmod a hot patch for module modification, this results in frequent failures
> sometimes. We found this 'unreliable' stack is from task just fork.

For livepatch, this shouldn't actually be a failure.  The patch will
just stay in the transition state until after the fork has completed.
Which should happen in a reasonable amount of time, right?

> 1) The task was not actually scheduled to excute, at this time UNWIND_HINT_EMPTY in
> ret_from_fork() has not reset unwind_hint, it's sp_reg and end field remain default value
> and end up throwing an error in unwind_next_frame() when called by arch_stack_walk_reliable();

Yes, this seems to be true for forked-but-not-yet-scheduled tasks.

I can look at fixing that.  I have some ORC cleanups in progress which
are related to UNWIND_HINT_EMPTY and the end of the stack.  I can add
this issue to the list of improvements.

> 2) The task has been scheduled but UNWIND_HINT_REGS not finished, at this time
> arch_stack_walk_reliable() terminates it's backtracing loop for pt_regs unknown
> and return -EINVAL because it's a user task.

Hm, do you see this problem with upstream?  It seems like it should
work.  arch_stack_walk_reliable() has this:

			/* Success path for user tasks */
			if (user_mode(regs))
				return 0;

Where exactly is the error coming from?

-- 
Josh