Re: [PATCH -v5 00/17] Rewrite x86/ftrace to use text_poke (and more)

From: Alexei Starovoitov <alexei.starovoitov@gmail.com>
To: Peter Zijlstra <peterz@infradead.org>
Cc: x86@kernel.org, linux-kernel@vger.kernel.org,
	rostedt@goodmis.org, mhiramat@kernel.org, bristot@redhat.com,
	jbaron@akamai.com, torvalds@linux-foundation.org,
	tglx@linutronix.de, mingo@kernel.org, namit@vmware.com,
	hpa@zytor.com, luto@kernel.org, ard.biesheuvel@linaro.org,
	jpoimboe@redhat.com, jeyu@kernel.org, bpf@vger.kernel.org,
	daniel@iogearbox.net, davem@davemloft.net, kernel-team@fb.com
Subject: Re: [PATCH -v5 00/17] Rewrite x86/ftrace to use text_poke (and more)
Date: Mon, 11 Nov 2019 11:47:28 -0800	[thread overview]
Message-ID: <20191111194726.udayafzpqxeqjyrj@ast-mbp.dhcp.thefacebook.com> (raw)
In-Reply-To: <20191111131252.921588318@infradead.org>

On Mon, Nov 11, 2019 at 02:12:52PM +0100, Peter Zijlstra wrote:
> Ftrace is one of the last W^X violators (after this only KLP is left). These
> patches move it over to the generic text_poke() interface and thereby get rid
> of this oddity.
> 
> The first 14 patches are the same as in the -v4 posting. The last 3 patches are
> new.
> 
> Will, patch 13, arm/ftrace, is unchanged. This is because this way it preserves
> behaviour, but if you can provide me a tested-by for the simpler variant I can
> drop that in.
> 
> Patch 15 reworks ftrace's event_create_dir(), which ran module code before the
> module was finished loading (before we even applied jump_labels and all that).
> 
> Patch 16 and 17 address minor review feedback.
> 
> Ingo, Alexei wants patch #1 for some BPF stuff, can he get that in a topic branch?

Thanks Peter!
Much appreciate it.

I've re-tested the patch 1 alone (it seems to be exactly the same as you posted
it originally back on Aug 27 and then on Oct 7). And now I tested my stuff with
this whole set. No conflicts. Feel free to add to patch 1 alone or the whole set:
Acked-by: Alexei Starovoitov <ast@kernel.org>
Tested-by: Alexei Starovoitov <ast@kernel.org>
Some of the patches I think are split too fine. I would have combined them, but
we try hard to limit our sets to less than fifteen in bpf/netdev land fwiw.

It was a poor judgment on my side to use text_poke() in my patch (to avoid
explicit dependency on your patch) and not mention the obvious race in the
commit log and intended fix when trees converge:
        case BPF_MOD_CALL_TO_CALL:
                if (memcmp(ip, old_insn, X86_CALL_SIZE))
                        goto out;
-               text_poke(ip, new_insn, X86_CALL_SIZE);
+               text_poke_bp(ip, new_insn, X86_CALL_SIZE, NULL);
                break;

To avoid the issue in the first place the best is to have your 1st patch in tip
and bpf-next/net-next trees. We had "the same patch in multiple trees"
situation in the past and git did the right thing during the merge window. So I
don't anticipate any issues this time around.

One more question.
What is the reason you stick to int3 style poking when 8 byte write is atomic?
Can text_poke() patch nop5 by combining the call/jmp5 insn with extra 3 bytes
after the nop and write 8 ?