[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20200429234159.gid6ht74qqmlpuz7@ast-mbp.dhcp.thefacebook.com>
Date: Wed, 29 Apr 2020 16:41:59 -0700
From: Alexei Starovoitov <alexei.starovoitov@...il.com>
To: Josh Poimboeuf <jpoimboe@...hat.com>
Cc: x86@...nel.org, tglx@...utronix.de, linux-kernel@...r.kernel.org,
mingo@...nel.org, hpa@...or.com, ast@...nel.org,
peterz@...radead.org, rdunlap@...radead.org,
Arnd Bergmann <arnd@...db.de>, bpf@...r.kernel.org,
daniel@...earbox.net
Subject: Re: BPF vs objtool again
On Wed, Apr 29, 2020 at 04:51:59PM -0500, Josh Poimboeuf wrote:
> On Thu, Jul 18, 2019 at 12:14:08PM -0700, tip-bot for Josh Poimboeuf wrote:
> > Commit-ID: 3193c0836f203a91bef96d88c64cccf0be090d9c
> > Gitweb: https://git.kernel.org/tip/3193c0836f203a91bef96d88c64cccf0be090d9c
> > Author: Josh Poimboeuf <jpoimboe@...hat.com>
> > AuthorDate: Wed, 17 Jul 2019 20:36:45 -0500
> > Committer: Thomas Gleixner <tglx@...utronix.de>
> > CommitDate: Thu, 18 Jul 2019 21:01:06 +0200
> >
> > bpf: Disable GCC -fgcse optimization for ___bpf_prog_run()
>
> For some reason, this
>
> __attribute__((optimize("-fno-gcse")))
>
> is disabling frame pointers in ___bpf_prog_run(). If you compile with
> CONFIG_FRAME_POINTER it'll show something like:
>
> kernel/bpf/core.o: warning: objtool: ___bpf_prog_run.cold()+0x7: call without frame pointer save/setup
you mean it started to disable frame pointers from some version of gcc?
It wasn't doing this before, since objtool wasn't complaining, right?
Sounds like gcc bug?
> Also, since GCC 9.1, the GCC docs say "The optimize attribute should be
> used for debugging purposes only. It is not suitable in production
> code." That doesn't sound too promising.
>
> So it seems like this commit should be reverted. But then we're back to
> objtool being broken again in the RETPOLINE=n case, which means no ORC
> coverage in this function. (See above commit for the details)
>
> Some ideas:
>
> - Skip objtool checking of that func/file (at least for RETPOLINE=n) --
> but then it won't have ORC coverage.
>
> - Get rid of the "double goto" in ___bpf_prog_run(), which simplifies it
> enough for objtool to understand -- but then the text explodes for
> RETPOLINE=y.
How that will look like?
That could be the best option.
> - Add -fno-gfcse to the Makefile for kernel/bpf/core.c -- but then that
> affects the optimization of other functions in the file. However I
> don't think the impact is significant.
>
> - Move ___bpf_prog_run() to its own file with the -fno-gfcse flag. I'm
> thinking this could be the least bad option. Alexei?
I think it would be easier to move some of the hot path
functions out of core.c instead.
Like *ksym*, BPF_CALL*, bpf_jit*, bpf_prog*.
I think resulting churn will be less.
imo it's more important to keep git blame history for interpreter
than for the other funcs.
Sounds like it's a fix that needs to be sent for the next RC ?
Please send a patch for bpf tree then.
Daniel, thoughts?
Powered by blists - more mailing lists