Date:   Fri, 23 Jun 2023 21:08:49 +0800
From:   Menglong Dong <menglong8.dong@...il.com>
To:     David Laight <David.Laight@...lab.com>
Cc:     Yonghong Song <yhs@...a.com>,
        "alexei.starovoitov@...il.com" <alexei.starovoitov@...il.com>,
        "ast@...nel.org" <ast@...nel.org>,
        "daniel@...earbox.net" <daniel@...earbox.net>,
        "andrii@...nel.org" <andrii@...nel.org>,
        "martin.lau@...ux.dev" <martin.lau@...ux.dev>,
        "song@...nel.org" <song@...nel.org>, "yhs@...com" <yhs@...com>,
        "john.fastabend@...il.com" <john.fastabend@...il.com>,
        "kpsingh@...nel.org" <kpsingh@...nel.org>,
        "sdf@...gle.com" <sdf@...gle.com>,
        "haoluo@...gle.com" <haoluo@...gle.com>,
        "jolsa@...nel.org" <jolsa@...nel.org>,
        "benbjiang@...cent.com" <benbjiang@...cent.com>,
        "bpf@...r.kernel.org" <bpf@...r.kernel.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        Menglong Dong <imagedong@...cent.com>
Subject: Re: [PATCH bpf-next v5 2/3] bpf, x86: allow function arguments up to
 12 for TRACING

On Thu, Jun 22, 2023 at 10:19 PM David Laight <David.Laight@...lab.com> wrote:
>
> ...
> > > Is that right for 86-64?
> > >
> > > IIRC arguments always take (at least) 64bits.
> > > For any 32bit argument (register or stack) the high bits are undefined.
> > > (Maybe in kernel they are always zero?
> > > From 32bit userspace they are definitely random.)
> > >
> >
> > Hello,
> >
> > According to my testing, the compiler will always
> > pass the arguments on 8-byte size with "push" insn
> > if the count of the arguments that need to be passed
> > on stack more than 1 and the size of the argument
> > doesn't exceed 8-byte. In this case, there won't be
> > garbage. For example, the high 4-byte will be made 0
> > if the size of the argument is 4-byte, as the "push" insn
> > will copy the argument from regs or imm into stack
> > in 8-byte.
>
> You have to know whether a value is expected to be 4 or 8
> bytes - a negative 32bit value is zero extended so can't
> be treated as a 64bit value.
>
> That is even true for values passed in registers.
>
> There is also a common problem with values passed in registers
> to system calls by 32bit code (maybe bpf is tracing these).
> In this case the high 32 bits of the register are random.
> They don't get zerod in 32bit mode.
>
> > If the count of the arguments on-stack is 1 and its size
> > doesn't exceed 4-byte, some compiler, like clang, may
> > not use the "push" insn. Instead, it allocates 4 bytes in the
> > stack, and copies the arguments from regs or imm into
> > stack in 4-byte. This is the case we deal with here.
>
> If the compiler sometimes writes a 4 byte (or smaller) value
> to pre-allocated stack then it is always allowed to do that.
> So the high bytes of the stack slot that contains a 32bit
> argument might always be junk.
> The count of on-stack arguments isn't relevant.
>

Yes, the way we clean garbage values is not
relevant, which comes from assumption. However,
It should be ok with the BPF program? like what Yonghong
said.

> > I'm not sure if I understand you correctly. Do you mean
> > that there will be garbage values for 32bit args?
>
> I'm pretty sure that the function call ABI doesn't require the
> caller set the high bits of sub-64bit arguments.
> The fact that they are often written with a push instruction
> that zeros the high bytes isn't really relevant.
>
> > > I think the called code is also responsible form masking 8 and 16bit
> > > values (in reality char/short args and return values just add code
> > > bloat).
> > >
> > > A 128bit value is either passed in two registers or two stack
> > > slots. If the last register is skipped it will be used for the
> > > next argument.
> > >
> >
> > Yeah, this point is considered in save_args(). Once
> > this happen, the count of stack slots should more
> > then 1, and the arguments on-stack will be stored with
> > "push" insn in 8-byte. Therefore, there shouldn't be garbage
> > values in this case?
> >
> > Do I miss something?
>
> The register/stack for these two calls is the same:
>         foo(1, 2, 3, 4, 5, 6, (int128_t)7);
>         bar(1, 2, 3, 4, 5, (int128_t)7, 6);
>

It is ok, as we already consider such cases. For
the foo(), the order we copy args is:

reg1, reg2, reg3, reg4, reg5, reg6, stack1, stack2

and for the bar (), it is:

reg1, reg2, reg3, reg4, reg5, stack1,stack2, reg6

The order of the arguments in the array we passed
to the BPF program is ok.

Thanks!
Menglong Dong

>         David
>
> -
> Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
> Registration No: 1397386 (Wales)

Powered by blists - more mailing lists