Re: [PATCH v2 bpf-next 1/3] bpf: add mmap() support for BPF_MAP_TYPE_ARRAY

From: Andrii Nakryiko <andrii.nakryiko@gmail.com>
To: Jakub Kicinski <jakub.kicinski@netronome.com>
Cc: Andrii Nakryiko <andriin@fb.com>, bpf <bpf@vger.kernel.org>,
	Networking <netdev@vger.kernel.org>,
	Alexei Starovoitov <ast@fb.com>,
	Daniel Borkmann <daniel@iogearbox.net>,
	Kernel Team <kernel-team@fb.com>, Rik van Riel <riel@surriel.com>,
	Johannes Weiner <hannes@cmpxchg.org>
Subject: Re: [PATCH v2 bpf-next 1/3] bpf: add mmap() support for BPF_MAP_TYPE_ARRAY
Date: Mon, 11 Nov 2019 18:06:42 -0800	[thread overview]
Message-ID: <CAEf4Bzay-sCd5+5Y1+toJuEd6vNh+R7pkosYA7V7wDqTdoDxdw@mail.gmail.com> (raw)
In-Reply-To: <20191111103743.1c3a38a3@cakuba>

On Mon, Nov 11, 2019 at 10:37 AM Jakub Kicinski
<jakub.kicinski@netronome.com> wrote:
>
> On Sat, 9 Nov 2019 00:06:30 -0800, Andrii Nakryiko wrote:
> > With BPF_F_MMAPABLE array allocating data in separate chunk of memory,
> > array_map_gen_lookup has to accomodate these changes. For non-memory-mapped
> > there are no changes and no extra instructions. For BPF_F_MMAPABLE case,
> > pointer to where array data is stored has to be dereferenced first.
> >
> > Generated code for non-memory-mapped array:
> >
> > ; p = bpf_map_lookup_elem(&data_map, &zero);
> >   22: (18) r1 = map[id:19]
> >   24: (07) r1 += 408                  /* array->inline_data offset */
> >   25: (61) r0 = *(u32 *)(r2 +0)
> >   26: (35) if r0 >= 0x3 goto pc+3
> >   27: (67) r0 <<= 3
> >   28: (0f) r0 += r1
> >   29: (05) goto pc+1
> >   30: (b7) r0 = 0
> >
> > Generated code for memory-mapped array:
> >
> > ; p = bpf_map_lookup_elem(&data_map, &zero);
> >   22: (18) r1 = map[id:27]
> >   24: (07) r1 += 400                  /* array->data offset */
> >   25: (79) r1 = *(u64 *)(r1 +0)               /* extra dereference */
> >   26: (61) r0 = *(u32 *)(r2 +0)
> >   27: (35) if r0 >= 0x3 goto pc+3
> >   28: (67) r0 <<= 3
> >   29: (0f) r0 += r1
> >   30: (05) goto pc+1
> >   31: (b7) r0 = 0
>
> Would it not be possible to overallocate the memory and align the start
> of the bpf_map in case of BPF_F_MMAPABLE so that no extra dereference
> is needed?

So let's say if sizeof(struct bpf_array) is 300, then I'd have to either:

- somehow make sure that I allocate 4k (for data) + 300 (for struct
bpf_array) in such a way that those 4k of data are 4k-aligned. Is
there any way to do that?
- assuming there isn't, then another way would be to allocate entire
4k page for struct bpf_array itself, but put it at the end of that
page, so that 4k of data is 4k-aligned. While wasteful, the bigger
problem is that pointer to bpf_array is not a pointer to allocated
memory anymore, so we'd need to remember that and adjust address
before calling vfree().

Were you suggesting #2 as a solution? Or am I missing some other way to do this?