bpf.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: jose.marchesi@oracle.com (Jose E. Marchesi)
To: Alexei Starovoitov <alexei.starovoitov@gmail.com>
Cc: bpf@vger.kernel.org, daniel@iogearbox.net, davem@davemloft.net
Subject: Re: [GCC,LLVM] bpf_helpers.h
Date: Tue, 17 Sep 2019 15:17:10 +0200	[thread overview]
Message-ID: <87sgov2hbd.fsf@oracle.com> (raw)
In-Reply-To: <20190916161742.54yabm3plqert2af@ast-mbp> (Alexei Starovoitov's message of "Mon, 16 Sep 2019 09:18:17 -0700")


    [...]
    > Would you consider implementing this attribute in llvm and adapt the
    > kernel's bpf_header accordingly?  In that respect, note that it is
    > possible to pass enum entries to the attribute (at least in GCC.)  So
    > you once you implement the attribute, you should be able to do:
    > 
    >    void *bpf_map_lookup_elem (void *map, const void *key)
    >        __attribute__ ((kernel_helper (BPF_FUNC_map_lookup_elem)));
    > 
    > instead of the current:
    > 
    >    static void *(*bpf_map_lookup_elem)(void *map, const void *key) =
    > 	(void *) BPF_FUNC_map_lookup_elem;
    
    What we've been using for long time is not exactly normal C code,
    but it's a valid C code that any compiler and any backend should
    consume and generate the code prescribed by the language.
    Here it says that it's a function pointer with fixed offset.
    x86 backends in both clang and gcc do the right thing.

    I don't understand why it's causing bpf backend for gcc to stumble.
    You mentioned "helper table in gcc bpf backend".
    That sounds like a red flag.
    The backend should not know names and numbers for helpers.
    For the following:
    static void (*foo)(void) = (void *) 123;
    int bar()
    {
            foo();
    }
    It should generate bpf instruction 'bpf_call 123', since that's
    what C language is asking compiler to do.

The C language does not prescribe compilers to generate any particular
flavor of call instruction...

... but that point is moot in this case since bpf only provides one such
kind of call instruction, and it is perfectly capable of handling the
situation above, so the compiler should use it, ideally at _any_
optimization level, and that includes -O0.

So ok.  What you say makes sense to me.  I'm convinced.

As for the table of helpers in the compiler, got the same feedback at
plumbers and I'm removing it. One less maintenance hurdle :)

    It is as you pointed out 'fragile', since it won't work with -O0,
    but that sort of the point. -O0 is too debuggy and un-optimized
    that even without this call insn quirk the verifier is not able
    to analyze even simple programs.
    Hence -O2 was a requirement for bpf development due to verifier smartness.
    One can argue that the verifier should become even smarter and analyze -O0 code,
    but I would argue otherwise. Linux kernel itself won't work with -O0.
    Same reasoning applies to bpf code. The main purpose of -O0 for user space
    development is to produce code together with -g that debugger can understand.
    The variables will stay on stack, line numbers will be intact, etc
    For bpf program development that's anti pattern.
    The 'bpf debugging' topic at plumbers showed that we still has a long way
    to go to make bpf debugging better. Single step, execution trace, nested bpf, etc
    All that will come, but -O0 support will not.

I see.  Good to know.
    
    > Please let me know what do you think.
    > 
    > SKB load built-ins
    > ------------------
    > 
    > bpf_helpers.h contains the following llvm-isms:
    > 
    >    /* llvm builtin functions that eBPF C program may use to
    >     * emit BPF_LD_ABS and BPF_LD_IND instructions
    >     */
    >    struct sk_buff;
    >    unsigned long long load_byte(void *skb,
    >                                 unsigned long long off) asm("llvm.bpf.load.byte");
    >    unsigned long long load_half(void *skb,
    > 			        unsigned long long off) asm("llvm.bpf.load.half");
    >    unsigned long long load_word(void *skb,
    > 			        unsigned long long off) asm("llvm.bpf.load.word");
    > 
    > Would you consider adopting more standard built-ins in llvm, like I
    > implemented in GCC?  These are:
    > 
    >    __builtin_bpf_load_byte (unsigned long long off)
    >    __builtin_bpf_load_half (unsigned long long off)
    >    __builtin_bpf_load_word (unsigned long long off)
    > 
    > Note that I didn't add an SKB argument to the builtins, as it is not
    > used: the pointer to the skb is implied by the instructions to be in
    > some predefined register.  I added compatibility wrappers in my
    > bpf-helpers.h:
    > 
    >   #define load_byte(SKB,OFF) __builtin_bpf_load_byte ((OFF))
    >   #define load_half(SKB,OFF) __builtin_bpf_load_half ((OFF))
    >   #define load_word(SKB,OFF) __builtin_bpf_load_word ((OFF))
    > 
    > Would you consider removing the unused SKB arguments from the built-ins
    > in llvm?  Or is there a good reason for having them, other than maybe
    > backwards compatibility?  In case backwards compatibility is a must, I
    > can add the unused argument to my builtins.
    
    llvm.bpf.load.word consumes 'skb' pointer. llvm bpf backend makes sure
    that it's in R6 before LD_ABS insn.
    __builtin_bpf_load_byte (unsigned long long off) cannot produce correct code.

So it is the 'skb' pointer that is loaded into %r6.  I misunderstood.
Will adapt the GCC built-ins accordingly.
    
    As far as doing it as __builtin_bpf_load_word(skb, off) instead of
    asm("llvm.bpf.load.word") that's fine, of course.
    Here compilers don't have to be the same.
    #ifdef clang vs gcc in bpf_helpers.h should do it.

Ok.

    Also please get rid of bpf-helpers.h from gcc tree.
    There shouldn't be such things shipped with compiler.

bpf-helpers.h is a temporal solution.  Will remove it as soon as we have
a bpf_helpers.h in the kernel that works with both compilers.

Thanks for the feedback!

  reply	other threads:[~2019-09-17 13:17 UTC|newest]

Thread overview: 6+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2019-09-12 19:26 Jose E. Marchesi
2019-09-16 16:18 ` Alexei Starovoitov
2019-09-17 13:17   ` Jose E. Marchesi [this message]
2019-10-24 16:56 ` initiated discussion to support attribute address_space in clang Yonghong Song
2019-11-04 10:21   ` Jose E. Marchesi
2019-11-07  6:49     ` Yonghong Song

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=87sgov2hbd.fsf@oracle.com \
    --to=jose.marchesi@oracle.com \
    --cc=alexei.starovoitov@gmail.com \
    --cc=bpf@vger.kernel.org \
    --cc=daniel@iogearbox.net \
    --cc=davem@davemloft.net \
    --subject='Re: [GCC,LLVM] bpf_helpers.h' \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).