From: Alexei Starovoitov <alexei.starovoitov@gmail.com>
To: Josh Poimboeuf <jpoimboe@redhat.com>
Cc: x86@kernel.org, tglx@linutronix.de, linux-kernel@vger.kernel.org,
mingo@kernel.org, hpa@zytor.com, ast@kernel.org,
peterz@infradead.org, rdunlap@infradead.org,
Arnd Bergmann <arnd@arndb.de>,
bpf@vger.kernel.org, daniel@iogearbox.net
Subject: Re: BPF vs objtool again
Date: Wed, 29 Apr 2020 21:24:00 -0700 [thread overview]
Message-ID: <20200430042400.45vvqx4ocwwogp3j@ast-mbp.dhcp.thefacebook.com> (raw)
In-Reply-To: <20200430035315.tc74v5twfdxv2goh@treble>
On Wed, Apr 29, 2020 at 10:53:15PM -0500, Josh Poimboeuf wrote:
> On Wed, Apr 29, 2020 at 07:10:52PM -0700, Alexei Starovoitov wrote:
> > > For example:
> > >
> > > #define GOTO ({ goto *jumptable[insn->code]; })
> > >
> > > and then replace all 'goto select_insn' with 'GOTO;'
> > >
> > > The problem is that with RETPOLINE=y, the function text size grows from
> > > 5k to 7k, because for each of the 160+ retpoline JMPs, GCC (stupidly)
> > > reloads the jump table register into a scratch register.
> >
> > that would be a tiny change, right?
> > I'd rather go with that and gate it with 'ifdef CONFIG_FRAME_POINTER'
> > Like:
> > #ifndef CONFIG_FRAME_POINTER
> > #define CONT ({ insn++; goto select_insn; })
> > #define CONT_JMP ({ insn++; goto select_insn; })
> > #else
> > #define CONT ({ insn++; goto *jumptable[insn->code]; })
> > #define CONT_JMP ({ insn++; goto *jumptable[insn->code]; })
> > #endif
> >
> > The reason this CONT and CONT_JMP macros are there because a combination
> > of different gcc versions together with different cpus make branch predictor
> > behave 'unpredictably'.
> >
> > I've played with CONT and CONT_JMP either both doing direct goto or
> > indirect goto and observed quite different performance characteristics
> > from the interpreter.
> > What you see right now was the best tune I could get from a set of cpus
> > I had to play with and compilers. If I did the same tuning today the outcome
> > could have been different.
> > So I think it's totally fine to use above code. I think some cpus may actually
> > see performance gains with certain versions of gcc.
> > The retpoline text increase is unfortunate but when retpoline is on
> > other security knobs should be on too. In particular CONFIG_BPF_JIT_ALWAYS_ON
> > should be on as well. Which will remove interpreter from .text completely.
>
> This would actually be contingent on RETPOLINE, not FRAME_POINTER.
>
> (FRAME_POINTER was the other issue with the "optimize" attribute, which
> we're reverting so it'll no longer be a problem.)
>
> So if you're not concerned about the retpoline text growth, it could be
> as simple as:
>
> #define CONT ({ insn++; goto *jumptable[insn->code]; })
> #define CONT_JMP ({ insn++; goto *jumptable[insn->code]; })
>
>
> Or, if you wanted to avoid the text growth, it could be:
>
> #ifdef CONFIG_RETPOLINE
I'm a bit lost. So objtool is fine with the asm when retpoline is on?
Then pls do:
#if defined(CONFIG_RETPOLINE) || !defined(CONFIG_X86)
since there is no need to mess with other archs.
> /*
> * Avoid a 40% increase in function text size by getting GCC to generate a
> * single retpoline jump instead of 160+.
> */
> #define CONT ({ insn++; goto select_insn; })
> #define CONT_JMP ({ insn++; goto select_insn; })
>
> select_insn:
> goto *jumptable[insn->code];
>
> #else /* !CONFIG_RETPOLINE */
> #define CONT ({ insn++; goto *jumptable[insn->code]; })
> #define CONT_JMP ({ insn++; goto *jumptable[insn->code]; })
> #endif /* CONFIG_RETPOLINE */
>
>
> But since this is legacy path, I think the first one is much nicer.
>
>
> Also, JMP_TAIL_CALL has a "goto select_insn", is it ok to convert that
> to CONT?
yep
next prev parent reply other threads:[~2020-04-30 4:24 UTC|newest]
Thread overview: 60+ messages / expand[flat|nested] mbox.gz Atom feed top
2019-07-18 1:36 [PATCH v2 00/22] x86, objtool: several fixes/improvements Josh Poimboeuf
2019-07-18 1:36 ` [PATCH v2 01/22] x86/paravirt: Fix callee-saved function ELF sizes Josh Poimboeuf
2019-07-18 19:07 ` [tip:core/urgent] " tip-bot for Josh Poimboeuf
2019-07-18 1:36 ` [PATCH v2 02/22] x86/kvm: Fix fastop function ELF metadata Josh Poimboeuf
2019-07-18 19:08 ` [tip:core/urgent] " tip-bot for Josh Poimboeuf
2019-07-18 1:36 ` [PATCH v2 03/22] x86/kvm: Replace vmx_vmenter()'s call to kvm_spurious_fault() with UD2 Josh Poimboeuf
2019-07-18 8:17 ` Paolo Bonzini
2019-07-18 19:08 ` [tip:core/urgent] " tip-bot for Josh Poimboeuf
2019-07-18 1:36 ` [PATCH v2 04/22] x86/kvm: Don't call kvm_spurious_fault() from .fixup Josh Poimboeuf
2019-07-18 8:22 ` Paolo Bonzini
2019-07-18 13:16 ` Sean Christopherson
2019-07-18 13:18 ` Paolo Bonzini
2019-07-18 14:12 ` Josh Poimboeuf
2019-07-18 14:13 ` Paolo Bonzini
2019-07-18 14:03 ` Josh Poimboeuf
2019-07-18 19:09 ` [tip:core/urgent] " tip-bot for Josh Poimboeuf
2019-07-18 1:36 ` [PATCH v2 05/22] x86/entry: Fix thunk function ELF sizes Josh Poimboeuf
2019-07-18 19:10 ` [tip:core/urgent] " tip-bot for Josh Poimboeuf
2019-07-18 1:36 ` [PATCH v2 06/22] x86/head/64: Annotate start_cpu0() as non-callable Josh Poimboeuf
2019-07-18 19:11 ` [tip:core/urgent] " tip-bot for Josh Poimboeuf
2019-07-18 1:36 ` [PATCH v2 07/22] x86/uaccess: Remove ELF function annotation from copy_user_handle_tail() Josh Poimboeuf
2019-07-18 19:11 ` [tip:core/urgent] " tip-bot for Josh Poimboeuf
2019-07-18 1:36 ` [PATCH v2 08/22] x86/uaccess: Don't leak AC flag into fentry from mcsafe_handle_tail() Josh Poimboeuf
2019-07-18 19:12 ` [tip:core/urgent] " tip-bot for Josh Poimboeuf
2019-07-18 1:36 ` [PATCH v2 09/22] x86/uaccess: Remove redundant CLACs in getuser/putuser error paths Josh Poimboeuf
2019-07-18 19:13 ` [tip:core/urgent] " tip-bot for Josh Poimboeuf
2019-07-18 1:36 ` [PATCH v2 10/22] bpf: Disable GCC -fgcse optimization for ___bpf_prog_run() Josh Poimboeuf
2019-07-18 19:14 ` [tip:core/urgent] " tip-bot for Josh Poimboeuf
2020-04-29 21:51 ` BPF vs objtool again Josh Poimboeuf
2020-04-29 22:01 ` Arvind Sankar
2020-04-29 23:41 ` Alexei Starovoitov
2020-04-30 0:13 ` Josh Poimboeuf
2020-04-30 2:10 ` Alexei Starovoitov
2020-04-30 3:53 ` Josh Poimboeuf
2020-04-30 4:24 ` Alexei Starovoitov [this message]
2020-04-30 4:43 ` Josh Poimboeuf
2019-07-18 1:36 ` [PATCH v2 11/22] objtool: Add mcsafe_handle_tail() to the uaccess safe list Josh Poimboeuf
2019-07-18 19:14 ` [tip:core/urgent] " tip-bot for Josh Poimboeuf
2019-07-18 1:36 ` [PATCH v2 12/22] objtool: Track original function across branches Josh Poimboeuf
2019-07-18 19:15 ` [tip:core/urgent] " tip-bot for Josh Poimboeuf
2019-07-18 1:36 ` [PATCH v2 13/22] objtool: Refactor function alias logic Josh Poimboeuf
2019-07-18 19:16 ` [tip:core/urgent] " tip-bot for Josh Poimboeuf
2019-07-18 1:36 ` [PATCH v2 14/22] objtool: Warn on zero-length functions Josh Poimboeuf
2019-07-18 19:17 ` [tip:core/urgent] " tip-bot for Josh Poimboeuf
2019-07-18 1:36 ` [PATCH v2 15/22] objtool: Change dead_end_function() to return boolean Josh Poimboeuf
2019-07-18 19:17 ` [tip:core/urgent] " tip-bot for Josh Poimboeuf
2019-07-18 1:36 ` [PATCH v2 16/22] objtool: Do frame pointer check before dead end check Josh Poimboeuf
2019-07-18 19:18 ` [tip:core/urgent] " tip-bot for Josh Poimboeuf
2019-07-18 1:36 ` [PATCH v2 17/22] objtool: Refactor sibling call detection logic Josh Poimboeuf
2019-07-18 19:19 ` [tip:core/urgent] " tip-bot for Josh Poimboeuf
2019-07-18 1:36 ` [PATCH v2 18/22] objtool: Refactor jump table code Josh Poimboeuf
2019-07-18 19:20 ` [tip:core/urgent] " tip-bot for Josh Poimboeuf
2019-07-18 1:36 ` [PATCH v2 19/22] objtool: Support repeated uses of the same C jump table Josh Poimboeuf
2019-07-18 19:20 ` [tip:core/urgent] " tip-bot for Jann Horn
2019-07-18 1:36 ` [PATCH v2 20/22] objtool: Fix seg fault on bad switch table entry Josh Poimboeuf
2019-07-18 19:21 ` [tip:core/urgent] " tip-bot for Josh Poimboeuf
2019-07-18 1:36 ` [PATCH v2 21/22] objtool: convert insn type to enum Josh Poimboeuf
2019-07-18 19:22 ` [tip:core/urgent] objtool: Convert " tip-bot for Josh Poimboeuf
2019-07-18 1:36 ` [PATCH v2 22/22] objtool: Support conditional retpolines Josh Poimboeuf
2019-07-18 19:23 ` [tip:core/urgent] " tip-bot for Josh Poimboeuf
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=20200430042400.45vvqx4ocwwogp3j@ast-mbp.dhcp.thefacebook.com \
--to=alexei.starovoitov@gmail.com \
--cc=arnd@arndb.de \
--cc=ast@kernel.org \
--cc=bpf@vger.kernel.org \
--cc=daniel@iogearbox.net \
--cc=hpa@zytor.com \
--cc=jpoimboe@redhat.com \
--cc=linux-kernel@vger.kernel.org \
--cc=mingo@kernel.org \
--cc=peterz@infradead.org \
--cc=rdunlap@infradead.org \
--cc=tglx@linutronix.de \
--cc=x86@kernel.org \
/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
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
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).