On Wed, Jan 13, 2021 at 11:25 PM Caroline Tice <cmtice@xxxxxxxxxx> wrote:
>
>
>
>
>
> On Tue, Jan 12, 2021 at 3:17 PM Sedat Dilek <sedat.dilek@xxxxxxxxx> wrote:
>>
>> On Mon, Jan 11, 2021 at 9:27 AM Sedat Dilek <sedat.dilek@xxxxxxxxx> wrote:
>> >
>> > On Fri, Dec 4, 2020 at 2:11 AM 'Nick Desaulniers' via Clang Built
>> > Linux <clang-built-linux@xxxxxxxxxxxxxxxx> wrote:
>> > >
>> > > DWARF v5 is the latest standard of the DWARF debug info format.
>> > >
>> > > DWARF5 wins significantly in terms of size when mixed with compression
>> > > (CONFIG_DEBUG_INFO_COMPRESSED).
>> > >
>> >
>> > Is this patchset bulletproof with CONFIG_DEBUG_INFO_BTF=y (and clang-cfi)?
>> >
>> > Debian has enabled this Kconfig in recent Linux v5.10 kernels which is
>> > a base for my custom kernels.
>> >
>> > It was fu**ing annoying to see I have no pahole binary installed and
>> > my build got broken after 3 hours of build.
>> > OK, I see that requirements is coded in scripts/link-vmlinux.sh.
>> >
>> > I needed to install dwarves package which provides pahole binary.
>> >
>> > I would like to see a prereq-checking for needed binaries with certain
>> > Kconfig set.
>> >
>> > After I calmed down I will - maybe - write to linux-kbuild mailing-list.
>> > Hope this will be a friendly email.
>> >
>>
>> After linux-bpf folks recommended not to use LLVM I jumped to gcc-10.
>>
>> I tried with ld.bfd first and then in a next run with LLVM=1.
>>
>> Upgraded pahole binary to latest Git plus a recommended patch from
>> linux-btf folks.
>>
>> Unfortunately, I see with CONFIG_DEBUG_INFO_DWARF5=y and
>> CONFIG_DEBUG_INFO_BTF=y:
>>
>> die__process_inline_expansion: DW_TAG_INVALID (0x48) @ <0x3f0dd5a> not handled!
>> die__process_function: DW_TAG_INVALID (0x48) @ <0x3f0dd69> not handled!
>>
>> In /usr/include/dwarf.h I found:
>>
>> 498: DW_OP_lit24 = 0x48, /* Literal 24. *
>
>
> There are multiple dwarf objects with the value 0x48, depending on which section of the dwarf.h file you search:
>
> DW_TAG_call_site = 0x48
> DW_AT_static_link = 0x48
> DW_OP_lit24 = 0x48.
>
> In this case, since the error message was about a DW_TAG, it would be complaining about DW_TAG_call_site, which is new to DWARR v5.
>
[ CC linux-bpf & Andrii and Jiri ]
Thanks for your feedback Caroline.
I ran several builds in the last 24 hours with Linux v5.11-rc3.
Setting DWARF version 2 (CONFIG_DEBUG_INFO_DWARF2=y) or version 4
(CONFIG_DEBUG_INFO_DWARF4=y) with this patchset together with GCC
v10.2.1 plus LLVM=1 does NOT show this.
BTW, it does not matter when LLVM/Clang v12 and LLVM/Clang v11 is used.
But again my compiler is here GCC plus LLVM utils like llvm-objcopy,
ld.lld, lllvm-ar, llvm-nm, etc.
( My initial problem was also seen with Clang v11.0.1 - I switched to
GCC as Debian's linux-kernel uses CONFIG_DEBUG_INFO_BTF=y
successfully. )
So, this is definitely a DWARF version 5 issue when
CONFIG_DEBUG_INFO_BTF=y (and CONFIG_DEBUG_INFO_BTF_MODULES=y).
Furthermore, my build-log says with pahole (see post-scriptum) from
dwarves package - here as an example:
WARN: multiple IDs found for 'bpf_map': 3860, 369255 - using 3860
$ grep 'WARN: multiple IDs found for'
build-log_5.11.0-rc3-6-amd64-gcc10-llvm11.txt | wc -l
1621
In the Linux code this derives from tools/bpf/resolve_btfids:
static int symbols_resolve(struct object *obj)
...
pr_info("WARN: multiple IDs found for
'%s': %d, %d - using %d\n",
( Cut-n-paste into Gmail truncates the lines and indentation, so I dropped it. )
Please see:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/bpf/resolve_btfids/main.c#n469
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/bpf/resolve_btfids/main.c#n532
I looked with llvm-dwarf tool and saw some errors concerning
".debug-ranges" (cannot remember the exact output and the command-line
I used).
Example for "DW_TAG_INVALID (0x48)" from my build-log:
die__process_inline_expansion: DW_TAG_INVALID (0x48) @ <0x1f671e7> not handled!
$ llvm-dwarfdump-11 --debug-info=0x1f671e7 vmlinux
vmlinux: file format elf64-x86-64
.debug_info contents:
0x01f671e7: DW_TAG_call_site
DW_AT_call_return_pc (0xffffffff811b16f2)
DW_AT_call_origin (0x01f67f1d)
Looking for "DW_AT_call_origin (0x01f67f1d)":
$ llvm-dwarfdump-11 --debug-info=0x01f67f1d vmlinux
vmlinux: file format elf64-x86-64
.debug_info contents:
0x01f67f1d: DW_TAG_subprogram
DW_AT_external (true)
DW_AT_declaration (true)
DW_AT_linkage_name ("fput")
DW_AT_name ("fput")
DW_AT_decl_file
("/home/dileks/src/linux-kernel/git/./include/linux/file.h")
DW_AT_decl_line (16)
DW_AT_decl_column (0x0d)
I have no experience in digging into DWARF (version 5) issues and how
to use llvm-dwarf or another appropriate tool.
If you give me a hand...
So all the above says - to be honest - nothing to me.
I hope it says something to you experts.
Regards,
- Sedat -
P.S.: I tried with a selfmade pahole from latest Git plus Jiri's v2
patch of "btf_encoder: Add extra checks for symbol names".
link="https://lore.kernel.org/bpf/20210113102509.1338601-1-jolsa@xxxxxxxxxx/T/#t
b4 -d am $link
...
Wrote v2_20210113_jolsa_btf_encoder_add_extra_checks_for_symbol_names.am
for thanks tracking
[1] https://git.kernel.org/pub/scm/devel/pahole/pahole.git/
>>
>>
>> Can someone enlighten what is going on?
>>
>> Nick, Fangrui?
>>
>> Thanks.
>>
>> - Sedat -
>>
>> P.S.: Patch from linux-bpf
>>
>> link="https://lore.kernel.org/bpf/20210112194724.GB1291051@krava/T/#t";
>> b4 -d am $link
>>
>> - EOT -
