Re: [PATCH v6 bpf-next 6/7] bpf: introduce bpf_prog_pack allocator

From: Song Liu <song@kernel.org>
To: Alexei Starovoitov <alexei.starovoitov@gmail.com>
Cc: Song Liu <songliubraving@fb.com>,
	Ilya Leoshkevich <iii@linux.ibm.com>, bpf <bpf@vger.kernel.org>,
	Network Development <netdev@vger.kernel.org>,
	LKML <linux-kernel@vger.kernel.org>,
	Alexei Starovoitov <ast@kernel.org>,
	Daniel Borkmann <daniel@iogearbox.net>,
	Andrii Nakryiko <andrii@kernel.org>,
	Kernel Team <Kernel-team@fb.com>,
	Peter Zijlstra <peterz@infradead.org>, X86 ML <x86@kernel.org>
Subject: Re: [PATCH v6 bpf-next 6/7] bpf: introduce bpf_prog_pack allocator
Date: Tue, 25 Jan 2022 16:50:33 -0800	[thread overview]
Message-ID: <CAPhsuW4YUT4r+9HSXxUMXjP8KjPq__npmxo6O4K8p0FSaZ6s0A@mail.gmail.com> (raw)
In-Reply-To: <CAADnVQLn0UFjMx_5rQhWbSPXK1PUbJR04cxSgrTH-KuUVy8C9g@mail.gmail.com>

On Tue, Jan 25, 2022 at 4:38 PM Alexei Starovoitov
<alexei.starovoitov@gmail.com> wrote:
>
> On Tue, Jan 25, 2022 at 3:09 PM Song Liu <song@kernel.org> wrote:
> >
> > On Tue, Jan 25, 2022 at 2:48 PM Alexei Starovoitov
> > <alexei.starovoitov@gmail.com> wrote:
> > >
> > > On Tue, Jan 25, 2022 at 2:25 PM Song Liu <song@kernel.org> wrote:
> > > >
> > > > On Tue, Jan 25, 2022 at 12:00 PM Alexei Starovoitov
> > > > <alexei.starovoitov@gmail.com> wrote:
> > > > >
> > > > > On Mon, Jan 24, 2022 at 11:21 PM Song Liu <song@kernel.org> wrote:
> > > > > >
> > > > > > On Mon, Jan 24, 2022 at 9:21 PM Alexei Starovoitov
> > > > > > <alexei.starovoitov@gmail.com> wrote:
> > > > > > >
> > > > > > > On Mon, Jan 24, 2022 at 10:27 AM Song Liu <songliubraving@fb.com> wrote:
> > > > > > > > >
> > > > > > > > > Are arches expected to allocate rw buffers in different ways? If not,
> > > > > > > > > I would consider putting this into the common code as well. Then
> > > > > > > > > arch-specific code would do something like
> > > > > > > > >
> > > > > > > > >  header = bpf_jit_binary_alloc_pack(size, &prg_buf, &prg_addr, ...);
> > > > > > > > >  ...
> > > > > > > > >  /*
> > > > > > > > >   * Generate code into prg_buf, the code should assume that its first
> > > > > > > > >   * byte is located at prg_addr.
> > > > > > > > >   */
> > > > > > > > >  ...
> > > > > > > > >  bpf_jit_binary_finalize_pack(header, prg_buf);
> > > > > > > > >
> > > > > > > > > where bpf_jit_binary_finalize_pack() would copy prg_buf to header and
> > > > > > > > > free it.
> > > > > > >
> > > > > > > It feels right, but bpf_jit_binary_finalize_pack() sounds 100% arch
> > > > > > > dependent. The only thing it will do is perform a copy via text_poke.
> > > > > > > What else?
> > > > > > >
> > > > > > > > I think this should work.
> > > > > > > >
> > > > > > > > We will need an API like: bpf_arch_text_copy, which uses text_poke_copy()
> > > > > > > > for x86_64 and s390_kernel_write() for x390. We will use bpf_arch_text_copy
> > > > > > > > to
> > > > > > > >   1) write header->size;
> > > > > > > >   2) do finally copy in bpf_jit_binary_finalize_pack().
> > > > > > >
> > > > > > > we can combine all text_poke operations into one.
> > > > > > >
> > > > > > > Can we add an 'image' pointer into struct bpf_binary_header ?
> > > > > >
> > > > > > There is a 4-byte hole in bpf_binary_header. How about we put
> > > > > > image_offset there? Actually we only need 2 bytes for offset.
> > > > > >
> > > > > > > Then do:
> > > > > > > int bpf_jit_binary_alloc_pack(size, &ro_hdr, &rw_hdr);
> > > > > > >
> > > > > > > ro_hdr->image would be the address used to compute offsets by JIT.
> > > > > >
> > > > > > If we only do one text_poke(), we cannot write ro_hdr->image yet. We
> > > > > > can use ro_hdr + rw_hdr->image_offset instead.
> > > > >
> > > > > Good points.
> > > > > Maybe let's go back to Ilya's suggestion and return 4 pointers
> > > > > from bpf_jit_binary_alloc_pack ?
> > > >
> > > > How about we use image_offset, like:
> > > >
> > > > struct bpf_binary_header {
> > > >         u32 size;
> > > >         u32 image_offset;
> > > >         u8 image[] __aligned(BPF_IMAGE_ALIGNMENT);
> > > > };
> > > >
> > > > Then we can use
> > > >
> > > > image = (void *)header + header->image_offset;
> > >
> > > I'm not excited about it, since it leaks header details into JITs.
> > > Looks like we don't need JIT to be aware of it.
> > > How about we do random() % roundup(sizeof(struct bpf_binary_header), 64)
> > > to pick the image start and populate
> > > image-sizeof(struct bpf_binary_header) range
> > > with 'int 3'.
> > > This way we can completely hide binary_header inside generic code.
> > > The bpf_jit_binary_alloc_pack() would return ro_image and rw_image only.
> > > And JIT would pass them back into bpf_jit_binary_finalize_pack().
> > > From the image pointer it would be trivial to get to binary_header with &63.
> > > The 128 byte offset that we use today was chosen arbitrarily.
> > > We were burning the whole page for a single program, so 128 bytes zone
> > > at the front was ok.
> > > Now we will be packing progs rounded up to 64 bytes, so it's better
> > > to avoid wasting those 128 bytes regardless.
> >
> > In bpf_jit_binary_hdr(), we calculate header as image & PAGE_MASK.
> > If we want s/PAGE_MASK/63 for x86_64, we will have different versions
> > of bpf_jit_binary_hdr(). It is not on any hot path, so we can use __weak for
> > it. Other than this, I think the solution works fine.
>
> I think it can stay generic.
>
> The existing bpf_jit_binary_hdr() will do & PAGE_MASK
> while bpf_jit_binary_hdr_pack() will do & 63.

The problem with this approach is that we need bpf_prog_ksym_set_addr
to be smart to pick bpf_jit_binary_hdr() or bpf_jit_binary_hdr_pack().

Song