BPF Archive on lore.kernel.org
 help / color / Atom feed
From: Masami Hiramatsu <mhiramat@kernel.org>
To: Alexei Starovoitov <alexei.starovoitov@gmail.com>
Cc: Steven Rostedt <rostedt@goodmis.org>,
	Peter Zijlstra <peterz@infradead.org>,
	Masami Hiramatsu <mhiramat@kernel.org>,
	Nikolay Borisov <nborisov@suse.com>,
	LKML <linux-kernel@vger.kernel.org>,
	Alexei Starovoitov <ast@kernel.org>, bpf <bpf@vger.kernel.org>,
	Josh Poimboeuf <jpoimboe@redhat.com>
Subject: Re: kprobes broken since 0d00449c7a28 ("x86: Replace ist_enter() with nmi_enter()")
Date: Sat, 30 Jan 2021 10:41:10 +0900
Message-ID: <20210130104110.0a25f155c00a86513e959ef0@kernel.org> (raw)
In-Reply-To: <20210129210533.7s6udd3vobkgb27u@ast-mbp.dhcp.thefacebook.com>

On Fri, 29 Jan 2021 13:05:33 -0800
Alexei Starovoitov <alexei.starovoitov@gmail.com> wrote:

> On Fri, Jan 29, 2021 at 02:01:03PM -0500, Steven Rostedt wrote:
> > On Fri, 29 Jan 2021 18:59:43 +0100
> > Peter Zijlstra <peterz@infradead.org> wrote:
> > 
> > > On Fri, Jan 29, 2021 at 09:45:48AM -0800, Alexei Starovoitov wrote:
> > > > Same things apply to bpf side. We can statically prove safety for
> > > > ftrace and kprobe attaching whereas to deal with NMI situation we
> > > > have to use run-time checks for recursion prevention, etc.  
> > > 
> > > I have no idea what you're saying. You can attach to functions that are
> > > called with random locks held, you can create kprobes in some very
> > > sensitive places.
> > > 
> > > What can you staticlly prove about that?
> > 
> > I think the main difference is, if you attach a kprobe or ftrace function,
> > you can theoretically analyze the location before you do the attachment.
> 
> Excatly.
> When we're writing bpf helpers we need to carefully think about reentrance and NMI.
> If the helper looks like:
> int nokprobe notrace bpf_something(...)
> {
>   // access variables A and B
> }
> 
> The implementation can rely on the fact that even if the helper is reentrant
> the state of A and B will be consistent. Either both got touched or none.
> Only NMI condition we have to worry about, because A could be modified 
> without touching B.
> If we write it as
> int nokprobe bpf_something(...) { ... }
> that would be another case.
> Here we need to consider the case that bpf prog can be attached to it via fentry nop.
> But no need to worry about instructions split in the middle because of kprobe via int3.
> Since we have big "if (in_nmi()) goto unsupported;" check in the beginning we
> only need to worry about combinations of kprobe at the start of the func,
> kprobe anywhere inside the func via int3, and ftrace at the start.
> Not having to think of NMI helps a ton.
> My earlier this_cpu vs __this_cpu comment is an example of that.
> If in_nmi is filtered early it's one implementation. If nmi has to be handled
> it's completely different algorithm.
> Now you've broke all this logic by making int3 to be marked as 'in_nmi' and
> bpf in kprobe in the middle of the func are now broken.
> Do people use that? Yeah they do.
> We have to fix it.
> What were your reasons to make int3 in_nmi?
> I've read the commit log, but I don't see in it the actual motivation
> for int3 other than "it looks like NMI to me. Let's make it so".
> The commit logs talk about cpu exceptions. I agree that #DB and #MC do behave like NMI.
> But #BP is not really. My understanding it's used by kprobes and text_poke_bp only.
> If the motivation was to close some issue with text_poke_bp then, sure,
> let's make handling of text_poke_bp to be treated as nmi.
> But kprobe is not that.
> I'm thinking either of the following solutions would be generic:
> - introduce another state to preempt flags like "kernel exception"

I like this solution. Or, at least there should be a way to provide the
probed context is NMI or not.
(BTW, would the NMI has a specific stack area? If so, nmi_context(regs)
 can be implemented.)

> - remove kprobe's int3 from in_nmi
> As bpf specific alternative we can do:
> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
> index 6c0018abe68a..37cc549ad52e 100644
> --- a/kernel/trace/bpf_trace.c
> +++ b/kernel/trace/bpf_trace.c
> @@ -96,7 +96,7 @@ unsigned int trace_call_bpf(struct trace_event_call *call, void *ctx)
>  {
>         unsigned int ret;
> 
> -       if (in_nmi()) /* not supported yet */
> +       if (in_nmi() && !kprobe_running()) /* not supported yet */

This doesn't make sense, because kprobe_running() always true in the kprobe handler.

The problem is that the in_nmi() checks whether the current context is NMI context,
but you want to know the context where the kprobe is invoked, is NMI context or not.

Thank you,

-- 
Masami Hiramatsu <mhiramat@kernel.org>

  reply index

Thread overview: 37+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
     [not found] <25cd2608-03c2-94b8-7760-9de9935fde64@suse.com>
     [not found] ` <20210128001353.66e7171b395473ef992d6991@kernel.org>
     [not found]   ` <20210128002452.a79714c236b69ab9acfa986c@kernel.org>
     [not found]     ` <a35a6f15-9ab1-917c-d443-23d3e78f2d73@suse.com>
     [not found]       ` <20210128103415.d90be51ec607bb6123b2843c@kernel.org>
2021-01-28  3:38         ` Masami Hiramatsu
2021-01-28  7:11           ` Nikolay Borisov
2021-01-28 16:12           ` Nikolay Borisov
2021-01-28 16:45             ` Nikolay Borisov
2021-01-28 16:50               ` Josh Poimboeuf
2021-01-28 21:52                 ` [PATCH] x86: Disable CET instrumentation in the kernel Josh Poimboeuf
2021-01-29  6:23                   ` Nikolay Borisov
2021-01-29 10:21                   ` Borislav Petkov
     [not found]                     ` <20210129151034.iba4eaa2fuxsipqa@treble>
2021-01-29 16:30                       ` Borislav Petkov
2021-01-29 16:49                         ` Josh Poimboeuf
2021-01-29 16:54                           ` Nikolay Borisov
2021-01-29 17:03                             ` Josh Poimboeuf
2021-01-29 17:07                               ` Borislav Petkov
2021-01-29 17:58                                 ` Seth Forshee
2021-01-28 18:24               ` kprobes broken since 0d00449c7a28 ("x86: Replace ist_enter() with nmi_enter()") Peter Zijlstra
2021-01-29  1:34                 ` Alexei Starovoitov
2021-01-29  6:36                   ` Nikolay Borisov
     [not found]                   ` <YBPNyRyrkzw2echi@hirez.programming.kicks-ass.net>
     [not found]                     ` <20210129224011.81bcdb3eba1227c414e69e1f@kernel.org>
     [not found]                       ` <20210129105952.74dc8464@gandalf.local.home>
2021-01-29 16:24                         ` Peter Zijlstra
2021-01-29 17:45                           ` Alexei Starovoitov
2021-01-29 17:59                             ` Peter Zijlstra
2021-01-29 19:01                               ` Steven Rostedt
2021-01-29 21:05                                 ` Alexei Starovoitov
2021-01-30  1:41                                   ` Masami Hiramatsu [this message]
2021-01-29 21:24                                 ` Steven Rostedt
2021-01-30  8:28                                   ` Peter Zijlstra
2021-01-30 12:44                                     ` Steven Rostedt
2021-02-02 10:45                                       ` Peter Zijlstra
2021-02-02 14:52                                         ` Steven Rostedt
2021-02-02 16:45                                           ` Peter Zijlstra
2021-02-02 16:56                                             ` Steven Rostedt
2021-02-02 18:30                                               ` Peter Zijlstra
2021-02-02 21:05                                                 ` Steven Rostedt
2021-02-03 13:33                                                   ` Masami Hiramatsu
2021-02-03 13:52                                                     ` Steven Rostedt
2021-01-30  2:02                               ` Masami Hiramatsu
2021-01-30  3:08                                 ` Alexei Starovoitov
2021-01-30 12:10                                   ` Masami Hiramatsu

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20210130104110.0a25f155c00a86513e959ef0@kernel.org \
    --to=mhiramat@kernel.org \
    --cc=alexei.starovoitov@gmail.com \
    --cc=ast@kernel.org \
    --cc=bpf@vger.kernel.org \
    --cc=jpoimboe@redhat.com \
    --cc=linux-kernel@vger.kernel.org \
    --cc=nborisov@suse.com \
    --cc=peterz@infradead.org \
    --cc=rostedt@goodmis.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

BPF Archive on lore.kernel.org

Archives are clonable:
	git clone --mirror https://lore.kernel.org/bpf/0 bpf/git/0.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 bpf bpf/ https://lore.kernel.org/bpf \
		bpf@vger.kernel.org
	public-inbox-index bpf

Example config snippet for mirrors

Newsgroup available over NNTP:
	nntp://nntp.lore.kernel.org/org.kernel.vger.bpf


AGPL code for this site: git clone https://public-inbox.org/public-inbox.git