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;
MIME-Version: 1.0
References: <20210423011517.4069221-1-revest@chromium.org> <8f89faf1-d7e6-ebe0-fb7d-c5b8243d140a@rasmusvillemoes.dk>
In-Reply-To: <8f89faf1-d7e6-ebe0-fb7d-c5b8243d140a@rasmusvillemoes.dk>
From:   Florent Revest <revest@chromium.org>
Date:   Fri, 23 Apr 2021 15:26:18 +0200
Message-ID: <CABRcYmLDBfoM8rOwPf+SdqkmJgtFRLYF94S4Fv2eU=Uwg4asTQ@mail.gmail.com>
Subject: Re: [PATCH bpf-next 0/2] Implement BPF formatted output helpers with bstr_printf
To:     Rasmus Villemoes <linux@rasmusvillemoes.dk>
Cc:     bpf <bpf@vger.kernel.org>, Alexei Starovoitov <ast@kernel.org>,
        Daniel Borkmann <daniel@iogearbox.net>,
        Andrii Nakryiko <andrii@kernel.org>,
        KP Singh <kpsingh@kernel.org>,
        Brendan Jackman <jackmanb@google.com>,
        open list <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Fri, Apr 23, 2021 at 10:50 AM Rasmus Villemoes
<linux@rasmusvillemoes.dk> wrote:
>
> On 23/04/2021 03.15, Florent Revest wrote:
> > Our formatted output helpers are currently implemented with
> > snprintf-like functions which take arguments as va_list but the types
> > stored in a va_list need to be known at compilation time which causes
> > problems when dealing with arguments from the BPF world that are always
> > u64 but considered differently depending on the format specifiers they
> > are associated with at runtime.
> >
> > This series replaces snprintf usages with bstr_printf calls. This lets
> > us construct a binary representation of arguments in bpf_printf_prepare
> > at runtime that matches an ABI that is neither arch nor compiler
> > specific.
> >
> > This solves a bug reported by Rasmus Villemoes that would mangle
> > arguments on 32 bit machines.
>
> That's not entirely accurate. The arguments are also mangled on x86-64,
> it's just that in a few cases that goes unnoticed. That's why I
> suggested you try and take your test case (which I assume had been
> passing with flying colours on x86-64) and rearrange the specifiers,
> arguments and expected output string so that the (morally) 32 bit
> arguments end up beyond those-that-end-up-in-the-reg_save_area.
>
> IOWs, it is the 32 bit arguments that are mangled (because they get
> passed as-if they were actually 64 bits), and that applies on all
> architectures; nothing to do with sizeof(long).

Mh, yes, I get your point and I agree that my description does not
really fit what you reported.

I tried what you suggested though, with the current bpf-next/master on x86_64:
BPF_SNPRINTF(out, sizeof(out),
"%u %d %u %d %u %d %u %d %u %d %u %d",
1, -2, 3, -4, 5, -6, 7, -8, 9, -10, 11, -12);

And out is "1 -2 3 -4 5 -6 7 -8 9 -10 11 -12" so i can't seem to be
able to produce the bug you described.
Do you think I'm missing something? Would you try it differently ?