From: Nick Desaulniers <ndesaulniers@google.com>
To: Yonghong Song <yhs@fb.com>
Cc: "Arnaldo Carvalho de Melo" <arnaldo.melo@gmail.com>,
dwarves@vger.kernel.org, "Alexei Starovoitov" <ast@kernel.org>,
"Andrii Nakryiko" <andrii@kernel.org>,
"Bill Wendling" <morbo@google.com>, bpf <bpf@vger.kernel.org>,
"David Blaikie" <dblaikie@gmail.com>,
"Fāng-ruì Sòng" <maskray@google.com>,
kernel-team@fb.com, "David Blaikie" <blaikie@google.com>
Subject: Re: [PATCH dwarves 1/2] dwarf_loader: check .debug_abbrev for cross-cu references
Date: Thu, 1 Apr 2021 11:52:05 -0700 [thread overview]
Message-ID: <CAKwvOd=mzDREDAXCxdFzSWnxC1hNc7udMXc7Lrf50qmJk9zE7Q@mail.gmail.com> (raw)
In-Reply-To: <20210401025820.2254482-1-yhs@fb.com>
On Wed, Mar 31, 2021 at 7:58 PM Yonghong Song <yhs@fb.com> wrote:
>
> Commit 39227909db3c checked compilation flags to see
> whether the binary is built with lto or not (-flto).
> Currently, for clang lto build, default setting
> won't put compilation flags in dwarf due to size
> concern.
>
> David Blaikie suggested in [1] to scan through .debug_abbrev
> for DW_FORM_ref_addr which should be most faster than
> scanning through cu's. This patch implemented this
> suggestion and replaced the previous compilation flag
> matching approach. Indeed, it seems that the overhead
> for this approach is indeed managable.
>
> I did some change to measure the overhead of cus_merging_cu():
> @@ -2650,7 +2652,15 @@ static int cus__load_module(struct cus *cus, struct conf_load *conf,
> }
> }
>
> - if (cus__merging_cu(dw)) {
> + bool do_merging;
> + struct timeval start, end;
> + gettimeofday(&start, NULL);
> + do_merging = cus__merging_cu(dw);
> + gettimeofday(&end, NULL);
> + fprintf(stderr, "%ld %ld -> %ld %ld\n", start.tv_sec, start.tv_usec,
> + end.tv_sec, end.tv_usec);
> +
> + if (do_merging) {
> res = cus__merge_and_process_cu(cus, conf, mod, dw, elf, filename,
> build_id, build_id_len,
> type_cu ? &type_dcu : NULL);
>
> For lto vmlinux, the cus__merging_cu() checking takes
> 130us over total "pahole -J vmlinux" time 65sec as the function bail out
> earlier due to detecting a merging cu condition.
> For non-lto vmlinux, the cus__merging_cu() checking takes
> ~171368us over total pahole time 36sec, roughly 0.5% overhead.
>
> [1] https://lore.kernel.org/bpf/20210328064121.2062927-1-yhs@fb.com/
>
It might be a nice little touch to add:
Suggested-by: David Blaikie <blaikie@google.com>
> Signed-off-by: Yonghong Song <yhs@fb.com>
> ---
> dwarf_loader.c | 43 ++++++++++++++++++++++++-------------------
> 1 file changed, 24 insertions(+), 19 deletions(-)
>
> diff --git a/dwarf_loader.c b/dwarf_loader.c
> index c1ca1a3..bd23751 100644
> --- a/dwarf_loader.c
> +++ b/dwarf_loader.c
> @@ -2503,35 +2503,40 @@ static int cus__load_debug_types(struct cus *cus, struct conf_load *conf,
>
> static bool cus__merging_cu(Dwarf *dw)
> {
> - uint8_t pointer_size, offset_size;
> Dwarf_Off off = 0, noff;
> size_t cuhl;
> - int cnt = 0;
>
> - /*
> - * Just checking the first cu is not enough.
> - * In linux, some C files may have LTO is disabled, e.g.,
> - * e242db40be27 x86, vdso: disable LTO only for vDSO
> - * d2dcd3e37475 x86, cpu: disable LTO for cpu.c
> - * Fortunately, disabling LTO for a particular file in a LTO build
> - * is rather an exception. Iterating 5 cu's to check whether
> - * LTO is used or not should be enough.
> - */
> - while (dwarf_nextcu(dw, off, &noff, &cuhl, NULL, &pointer_size,
> - &offset_size) == 0) {
> + while (dwarf_nextcu (dw, off, &noff, &cuhl, NULL, NULL, NULL) == 0) {
> Dwarf_Die die_mem;
> Dwarf_Die *cu_die = dwarf_offdie(dw, off + cuhl, &die_mem);
>
> if (cu_die == NULL)
> break;
>
> - if (++cnt > 5)
> - break;
> + Dwarf_Off offset = 0;
> + while (true) {
> + size_t length;
> + Dwarf_Abbrev *abbrev = dwarf_getabbrev (cu_die, offset, &length);
> + if (abbrev == NULL || abbrev == DWARF_END_ABBREV)
> + break;
>
> - const char *producer = attr_string(cu_die, DW_AT_producer);
> - if (strstr(producer, "clang version") != NULL &&
> - strstr(producer, "-flto") != NULL)
> - return true;
> + size_t attrcnt;
> + if (dwarf_getattrcnt (abbrev, &attrcnt) != 0)
> + return false;
> +
> + unsigned int attr_num, attr_form;
> + Dwarf_Off aboffset;
> + size_t j;
> + for (j = 0; j < attrcnt; ++j) {
> + if (dwarf_getabbrevattr (abbrev, j, &attr_num, &attr_form,
> + &aboffset))
> + return false;
> + if (attr_form == DW_FORM_ref_addr)
> + return true;
> + }
> +
> + offset += length;
> + }
>
> off = noff;
> }
> --
> 2.30.2
>
--
Thanks,
~Nick Desaulniers
next prev parent reply other threads:[~2021-04-01 19:26 UTC|newest]
Thread overview: 33+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-04-01 2:58 [PATCH dwarves 0/2] dwarf_loader: improve cus__merging_cu() Yonghong Song
2021-04-01 2:58 ` [PATCH dwarves 1/2] dwarf_loader: check .debug_abbrev for cross-cu references Yonghong Song
2021-04-01 18:52 ` Nick Desaulniers [this message]
2021-04-01 19:36 ` Arnaldo
2021-04-01 20:54 ` Yonghong Song
2021-04-01 2:58 ` [PATCH dwarves 2/2] dwarf_loader: check .notes section for lto build info Yonghong Song
2021-04-01 12:59 ` [PATCH dwarves 0/2] dwarf_loader: improve cus__merging_cu() Arnaldo Carvalho de Melo
2021-04-01 18:50 ` Nick Desaulniers
2021-04-01 19:35 ` Bill Wendling
2021-04-01 20:56 ` Bill Wendling
2021-04-01 21:59 ` Yonghong Song
2021-04-02 19:44 ` Bill Wendling
2021-04-06 13:22 ` Arnaldo Carvalho de Melo
2021-04-06 13:23 ` Arnaldo Carvalho de Melo
2021-04-06 16:40 ` Arnaldo Carvalho de Melo
2021-04-06 16:55 ` Arnaldo Carvalho de Melo
2021-04-06 16:59 ` Arnaldo Carvalho de Melo
2021-04-06 17:03 ` Arnaldo Carvalho de Melo
2021-04-06 17:07 ` Arnaldo Carvalho de Melo
2021-04-06 17:25 ` Arnaldo Carvalho de Melo
2021-04-06 17:23 ` Yonghong Song
2021-04-06 17:34 ` Arnaldo Carvalho de Melo
2021-04-06 17:36 ` Arnaldo Carvalho de Melo
2021-04-06 17:48 ` Yonghong Song
2021-04-06 18:03 ` Arnaldo Carvalho de Melo
2021-04-06 18:04 ` Arnaldo Carvalho de Melo
2021-04-06 18:31 ` Nick Desaulniers
2021-04-06 18:17 ` David Blaikie
2021-04-06 18:20 ` Arnaldo Carvalho de Melo
2021-04-06 18:22 ` Fāng-ruì Sòng
2021-04-06 17:09 ` Jiri Olsa
2021-04-02 13:27 ` Arnaldo Carvalho de Melo
2021-04-02 14:34 ` 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='CAKwvOd=mzDREDAXCxdFzSWnxC1hNc7udMXc7Lrf50qmJk9zE7Q@mail.gmail.com' \
--to=ndesaulniers@google.com \
--cc=andrii@kernel.org \
--cc=arnaldo.melo@gmail.com \
--cc=ast@kernel.org \
--cc=blaikie@google.com \
--cc=bpf@vger.kernel.org \
--cc=dblaikie@gmail.com \
--cc=dwarves@vger.kernel.org \
--cc=kernel-team@fb.com \
--cc=maskray@google.com \
--cc=morbo@google.com \
--cc=yhs@fb.com \
/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).