Message-ID: <2AAC8B8C-96F1-400F-AFA6-D4AF41EC82F4@fb.com>
Date:   Mon, 24 Jan 2022 18:27:11 +0000
From:   Song Liu <songliubraving@...com>
To:     Ilya Leoshkevich <iii@...ux.ibm.com>
CC:     Alexei Starovoitov <alexei.starovoitov@...il.com>,
        Song Liu <song@...nel.org>, bpf <bpf@...r.kernel.org>,
        Network Development <netdev@...r.kernel.org>,
        LKML <linux-kernel@...r.kernel.org>,
        "Alexei Starovoitov" <ast@...nel.org>,
        Daniel Borkmann <daniel@...earbox.net>,
        "Andrii Nakryiko" <andrii@...nel.org>,
        Kernel Team <Kernel-team@...com>,
        "Peter Zijlstra" <peterz@...radead.org>, X86 ML <x86@...nel.org>
Subject: Re: [PATCH v6 bpf-next 6/7] bpf: introduce bpf_prog_pack allocator



> On Jan 24, 2022, at 4:29 AM, Ilya Leoshkevich <iii@...ux.ibm.com> wrote:
> 
> 
> 
> On 1/23/22 02:03, Song Liu wrote:
>>> On Jan 21, 2022, at 6:12 PM, Alexei Starovoitov <alexei.starovoitov@...il.com> wrote:
>>> 
>>> On Fri, Jan 21, 2022 at 5:30 PM Song Liu <songliubraving@...com> wrote:
>>>> 
>>>> 
>>>> 
>>>>> On Jan 21, 2022, at 5:12 PM, Alexei Starovoitov <alexei.starovoitov@...il.com> wrote:
>>>>> 
>>>>> On Fri, Jan 21, 2022 at 5:01 PM Song Liu <songliubraving@...com> wrote:
>>>>>> 
>>>>>> In this way, we need to allocate rw_image here, and free it in
>>>>>> bpf_jit_comp.c. This feels a little weird to me, but I guess that
>>>>>> is still the cleanest solution for now.
>>>>> 
>>>>> You mean inside bpf_jit_binary_alloc?
>>>>> That won't be arch independent.
>>>>> It needs to be split into generic piece that stays in core.c
>>>>> and callbacks like bpf_jit_fill_hole_t
>>>>> or into multiple helpers with prep in-between.
>>>>> Don't worry if all archs need to be touched.
>>>> 
>>>> How about we introduce callback bpf_jit_set_header_size_t? Then we
>>>> can split x86's jit_fill_hole() into two functions, one to fill the
>>>> hole, the other to set size. The rest of the logic gonna stay the same.
>>>> 
>>>> Archs that do not use bpf_prog_pack won't need bpf_jit_set_header_size_t.
>>> 
>>> That's not any better.
>>> 
>>> Currently the choice of bpf_jit_binary_alloc_pack vs bpf_jit_binary_alloc
>>> leaks into arch bits and bpf_prog_pack_max_size() doesn't
>>> really make it generic.
>>> 
>>> Ideally all archs continue to use bpf_jit_binary_alloc()
>>> and magic happens in a generic code.
>>> If not then please remove bpf_prog_pack_max_size(),
>>> since it doesn't provide much value and pick
>>> bpf_jit_binary_alloc_pack() signature to fit x86 jit better.
>>> It wouldn't need bpf_jit_fill_hole_t callback at all.
>>> Please think it through so we don't need to redesign it
>>> when another arch will decide to use huge pages for bpf progs.
>>> 
>>> cc-ing Ilya for ideas on how that would fit s390.
>> I guess we have a few different questions here:
>> 1. Can we use bpf_jit_binary_alloc() for both regular page and shared
>> huge page?
>> I think the answer is no, as bpf_jit_binary_alloc() allocates a rw
>> buffer, and arch calls bpf_jit_binary_lock_ro after JITing. The new
>> allocator will return a slice of a shared huge page, which is locked
>> RO before JITing.
>> 2. The problem with bpf_prog_pack_max_size() limitation.
>> I think this is the worst part of current version of bpf_prog_pack,
>> but it shouldn't be too hard to fix. I will remove this limitation
>> in the next version.
>> 3. How to set proper header->size?
>> I guess we can introduce something similar to bpf_arch_text_poke()
>> for this?
>> My proposal for the next version is:
>> 1. No changes to archs that do not use huge page, just keep using
>>    bpf_jit_binary_alloc.
>> 2. For x86_64 (and other arch that would support bpf program on huge
>>    pages):
>>    2.1 arch/bpf_jit_comp calls bpf_jit_binary_alloc_pack() to allocate
>>        an RO bpf_binary_header;
>>    2.2 arch allocates a temporary buffer for JIT. Once JIT is done,
>>        use text_poke_copy to copy the code to the RO bpf_binary_header.
> 
> 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.

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().

The syntax of bpf_arch_text_copy is quite different to existing 
bpf_arch_text_poke, so I guess a new API is better. 

> 
> If this won't work, I also don't see any big problems in the scheme
> that you propose (especially if bpf_prog_pack_max_size() limitation is
> gone).
> 
> [...]
> 
> Btw, are there any existing benchmarks that I can use to check whether
> this is worth enabling on s390?

Unfortunately, we don't have a benchmark to share. Most of our benchmarks
are shadow tests that cannot run out of production environment. We have 
issues with iTLB misses for most of our big services. A typical system 
may see hundreds of iTLB misses per million instruction. Some sched_cls
programs are often the top triggers of these iTLB misses. 

Thanks,
Song

Powered by blists - more mailing lists