Re: [PATCH v2 bpf-next 4/4] selftests/bpf: Add test for sleepable bpf iterator programs

From: Kenny Yu <kennyyu@fb.com>
To: <phoenix1987@gmail.com>
Cc: <alexei.starovoitov@gmail.com>, <andrii@kernel.org>,
	<ast@kernel.org>, <bpf@vger.kernel.org>, <daniel@iogearbox.net>,
	<kennyyu@fb.com>, <yhs@fb.com>
Subject: Re: [PATCH v2 bpf-next 4/4] selftests/bpf: Add test for sleepable bpf iterator programs
Date: Wed, 19 Jan 2022 08:56:10 -0800	[thread overview]
Message-ID: <20220119165610.3966158-1-kennyyu@fb.com> (raw)
In-Reply-To: <CAGnuNNv0TVQ3ZSYjJgJh1Dxasc9pX-QTwVApYyW7Q0xEy0Bgng@mail.gmail.com>

> > But I also wanted to point out that this helper is logically in the
> > same family as bpf_probe_read_kernel/user and bpf_copy_from_user, etc,
> > where we have consistent pattern that first two arguments specify
> > destination buffer (so buf + len) and the remaining ones specify
> > source (in probe_read it's just an address, here it's tsk_addr). So I
> > wonder if it would be less surprising and more consistent to reorder
> > and have:
> >
> > buf, len, tsk, addr, flags
> >
> > ?
> >
>
> I would personally find it more intuitive to have process information
> passed as either the first argument (like process_vm_readv does), or
> as "last", just before the flags (as extra information required w.r.t.
> to local versions, e.g. bpf_copy_from_user).

I think that makes sense. I'll combine both Andrii's and Gabriele's suggestions
and keep the signature as close to the existing helpers
(e.g., bpf_probe_read_user) and add the additional arguments at the end.
I'll proceed with this signature:

  bpf_access_process_vm(void *dst,
  			u32 size,
			const void *unsafe_ptr,
			struct task_struct *tsk,
			unsigned int flags);

Thanks for all the suggestions!

Kenny