btf_encoder: Handle DW_TAG_variable that has DW_AT_specification
A message that is not in the archive (sent off-list or not yet received): [email protected]
A message that is not in the archive (sent off-list or not yet received): [email protected]
A message that is not in the archive (sent off-list or not yet received): CA+khW7jf7Z=sMC1u5eyn6XOZDTFJaNjV-D0ogvQSyUGSKjC3LQ@mail.gmail.com
A message that is not in the archive (sent off-list or not yet received): [email protected]
Arnaldo,
Is this patch ready to be merged into Pahole's master branch? Alexei is testing the kernel patches that need this patch. Please let me know if there is anything I can do to help merging.
Thank you,
Hao
Quoted 167 lines · Arnaldo Carvalho de Melo
Quoted 77 lines · Hao Luo
Arnaldo,
Quoted 9 lines · Arnaldo Carvalho de Melo
Em Mon, Aug 24, 2020 at 05:45:23PM -0700, Hao Luo escreveu:
Quoted 7 lines
It is found on gcc 8.2 that global percpu variables generate the following dwarf entry in the cu where the variable is defined[1].
Take the global variable "
bpf_prog_active" defined inkernel/bpf/syscall.cas an example. The debug info forsyscall.chas two dwarf entries for "bpf_prog_active".[...]
Quoted 4 lines
Interesting, here I get, with binutils' readelf:
[root@quaco perf]# readelf -wi
../build/v5.8-rc5+/kernel/bpf/syscall.o| grepbpf_prog_activeQuoted 1 line
<f6a1> DW_AT_name : (indirect string, offset: 0xb70d):
bpf_prog_activeQuoted 5 lines
[root@quaco perf]#
Just one, as:
[root@quaco perf]# readelf -wi
../build/v5.8-rc5+/kernel/bpf/syscall.o| grepbpf_prog_active-B1 -A8Quoted 2 lines
<1><f6a0>: Abbrev Number: 103 (
DW_TAG_variable)<f6a1> DW_AT_name : (indirect string, offset: 0xb70d):
bpf_prog_activeQuoted 8 lines
<f6a5> DW_AT_decl_file : 11 <f6a6> DW_AT_decl_line : 1008 <f6a8> DW_AT_decl_column : 1 <f6a9> DW_AT_type : <0xcf> <f6ad> DW_AT_external : 1 <f6ad> DW_AT_declaration : 1 <1><f6ad>: Abbrev Number: 103 (DW_TAG_variable) <f6ae> DW_AT_name : (indirect string, offset: 0x3a5d):
bpf_stats_enabled_mutexQuoted 5 lines
[root@quaco perf]#
I get what you have when I use elfutils' readelf:
[root@quaco perf]# eu-readelf -winfo
../build/v5.8-rc5+/kernel/bpf/syscall.o| grepbpf_prog_activeQuoted 5 lines
name (strp) "
bpf_prog_active"
[ 0] addr .data..percpu+0 <bpf_prog_active>
[root@quaco perf]#[root@quaco perf]# eu-readelf -winfo
../build/v5.8-rc5+/kernel/bpf/syscall.o| grep -B1 -A8 \"bpf_prog_active\"Quoted 10 lines
[ f6a0] variable abbrev: 103 name (strp) "bpf_prog_active" decl_file (data1) bpf.h (11) decl_line (data2) 1008 decl_column (data1) 1 type (ref4) [ cf] external (flag_present) yes declaration (flag_present) yes [ f6ad] variable abbrev: 103 name (strp) "bpf_stats_enabled_mutex"Quoted 5 lines
[root@quaco perf]#
And:
[root@quaco perf]# eu-readelf -winfo
../build/v5.8-rc5+/kernel/bpf/syscall.o| grep -B5 \<bpf_prog_active\>Quoted 8 lines
[ 1bdf5] variable abbrev: 212 specification (ref4) [ f6a0] decl_file (data1) syscall.c (1) decl_line (data1) 43 location (exprloc) [ 0] addr .data..percpu+0 <bpf_prog_active>[root@quaco perf]#
Quoted 1 line
In binutils readelf, there is a extra entry
Not here, tomorrow I'll triple check.
Quoted 34 lines
<1><1b24c>: Abbrev Number: 195 (
DW_TAG_variable)<1b24e> DW_AT_specification: <0xf335> <1b252> DW_AT_decl_file : 1 <1b253> DW_AT_decl_line : 43 <1b254> DW_AT_location : 9 byte block: 3 0 0 0 0 0 0 0 0(
DW_OP_addr: 0)which points to
<1><f335>: Abbrev Number: 95 (
DW_TAG_variable)
<f336>DW_AT_name: (indirect string, offset:0xb37a):bpf_prog_activeIt just doesn't have the string '
bpf_prog_active', annotating entry. So eu-readelf and binutils readelf have the same results.Quoted 5 lines
Quoted 2 lines
Note that second
DW_TAG_variableentry contains specification that points to the first entry.So you are not considering the first when encoding since it is just a
DW_AT_declaration, considers the second, as it should be, and thenneeds
Quoted 4 lines
to go see its
DW_AT_specification, right?Sounds correct, applying, will test further and then push out,
Yes, exactly. The var tags to be considered are those that either have
DW_AT_specificationor not haveDW_AT_declaration. This makes surebtf_encoderworks correctly on both old and new gcc.Quoted 2 lines
Thanks,
Quoted 8 lines
Quoted 1 line
- Arnaldo
Suggested by Yonghong, I tested this change on a larger set of compilers this time and works correctly. See below.
Could you also add 'Reported-by: Yonghong Song <[email protected]>'? I should have done that when sending out this patch. The credit goes to Yonghong.
Sure, and I'll add your results with different computers, for the record.
Thanks,
- Arnaldo
Quoted 24 lines
Thank you,
Haoclang 10:
[67] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [20168] VAR 'bpf_prog_active' type_id=67, linkage=global-allocclang 9:
[64] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [19789] VAR 'bpf_prog_active' type_id=64, linkage=global-allocgcc 10.2
[18] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [20319] VAR 'bpf_prog_active' type_id=18, linkage=global-allocgcc 9.3:
[21] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [21085] VAR 'bpf_prog_active' type_id=21, linkage=global-allocgcc 8
[21] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [21084] VAR 'bpf_prog_active' type_id=21, linkage=global-allocQuoted 7 lines
gcc 6.2
[22] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [21083] VAR 'bpf_prog_active' type_id=22, linkage=global-allocgcc 4.9
[17] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [20410] VAR 'bpf_prog_active' type_id=17, linkage=global-alloc--
Sent from my Android device with K-9 Mail. Please excuse my brevity.
Arnaldo,
ping.
Is anything blocking this fix from merging? The kernel patches are stalled waiting on the pahole.
Thanks
Quoted 179 lines · Hao Luo
Arnaldo,
Is this patch ready to be merged into Pahole's master branch? Alexei is testing the kernel patches that need this patch. Please let me know if there is anything I can do to help merging.
Thank you,
HaoQuoted 167 lines · Arnaldo Carvalho de Melo
Quoted 77 lines · Hao Luo
Arnaldo,
Quoted 9 lines · Arnaldo Carvalho de Melo
Em Mon, Aug 24, 2020 at 05:45:23PM -0700, Hao Luo escreveu:
Quoted 7 lines
It is found on gcc 8.2 that global percpu variables generate the following dwarf entry in the cu where the variable is defined[1].
Take the global variable "
bpf_prog_active" defined inkernel/bpf/syscall.cas an example. The debug info forsyscall.chas two dwarf entries for "bpf_prog_active".[...]
Quoted 4 lines
Interesting, here I get, with binutils' readelf:
[root@quaco perf]# readelf -wi
../build/v5.8-rc5+/kernel/bpf/syscall.o| grepbpf_prog_activeQuoted 1 line
<f6a1> DW_AT_name : (indirect string, offset: 0xb70d):
bpf_prog_activeQuoted 5 lines
[root@quaco perf]#
Just one, as:
[root@quaco perf]# readelf -wi
../build/v5.8-rc5+/kernel/bpf/syscall.o| grepbpf_prog_active-B1 -A8Quoted 2 lines
<1><f6a0>: Abbrev Number: 103 (
DW_TAG_variable)<f6a1> DW_AT_name : (indirect string, offset: 0xb70d):
bpf_prog_activeQuoted 8 lines
<f6a5> DW_AT_decl_file : 11 <f6a6> DW_AT_decl_line : 1008 <f6a8> DW_AT_decl_column : 1 <f6a9> DW_AT_type : <0xcf> <f6ad> DW_AT_external : 1 <f6ad> DW_AT_declaration : 1 <1><f6ad>: Abbrev Number: 103 (DW_TAG_variable) <f6ae> DW_AT_name : (indirect string, offset: 0x3a5d):
bpf_stats_enabled_mutexQuoted 5 lines
[root@quaco perf]#
I get what you have when I use elfutils' readelf:
[root@quaco perf]# eu-readelf -winfo
../build/v5.8-rc5+/kernel/bpf/syscall.o| grepbpf_prog_activeQuoted 5 lines
name (strp) "
bpf_prog_active"
[ 0] addr .data..percpu+0 <bpf_prog_active>
[root@quaco perf]#[root@quaco perf]# eu-readelf -winfo
../build/v5.8-rc5+/kernel/bpf/syscall.o| grep -B1 -A8 \"bpf_prog_active\"Quoted 10 lines
[ f6a0] variable abbrev: 103 name (strp) "bpf_prog_active" decl_file (data1) bpf.h (11) decl_line (data2) 1008 decl_column (data1) 1 type (ref4) [ cf] external (flag_present) yes declaration (flag_present) yes [ f6ad] variable abbrev: 103 name (strp) "bpf_stats_enabled_mutex"Quoted 5 lines
[root@quaco perf]#
And:
[root@quaco perf]# eu-readelf -winfo
../build/v5.8-rc5+/kernel/bpf/syscall.o| grep -B5 \<bpf_prog_active\>Quoted 8 lines
[ 1bdf5] variable abbrev: 212 specification (ref4) [ f6a0] decl_file (data1) syscall.c (1) decl_line (data1) 43 location (exprloc) [ 0] addr .data..percpu+0 <bpf_prog_active>[root@quaco perf]#
Quoted 1 line
In binutils readelf, there is a extra entry
Not here, tomorrow I'll triple check.
Quoted 34 lines
<1><1b24c>: Abbrev Number: 195 (
DW_TAG_variable)<1b24e> DW_AT_specification: <0xf335> <1b252> DW_AT_decl_file : 1 <1b253> DW_AT_decl_line : 43 <1b254> DW_AT_location : 9 byte block: 3 0 0 0 0 0 0 0 0(
DW_OP_addr: 0)which points to
<1><f335>: Abbrev Number: 95 (
DW_TAG_variable)
<f336>DW_AT_name: (indirect string, offset:0xb37a):bpf_prog_activeIt just doesn't have the string '
bpf_prog_active', annotating entry. So eu-readelf and binutils readelf have the same results.Quoted 5 lines
Quoted 2 lines
Note that second
DW_TAG_variableentry contains specification that points to the first entry.So you are not considering the first when encoding since it is just a
DW_AT_declaration, considers the second, as it should be, and thenneeds
Quoted 4 lines
to go see its
DW_AT_specification, right?Sounds correct, applying, will test further and then push out,
Yes, exactly. The var tags to be considered are those that either have
DW_AT_specificationor not haveDW_AT_declaration. This makes surebtf_encoderworks correctly on both old and new gcc.Quoted 2 lines
Thanks,
Quoted 8 lines
Quoted 1 line
- Arnaldo
Suggested by Yonghong, I tested this change on a larger set of compilers this time and works correctly. See below.
Could you also add 'Reported-by: Yonghong Song <[email protected]>'? I should have done that when sending out this patch. The credit goes to Yonghong.
Sure, and I'll add your results with different computers, for the record.
Thanks,
- Arnaldo
Quoted 24 lines
Thank you,
Haoclang 10:
[67] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [20168] VAR 'bpf_prog_active' type_id=67, linkage=global-allocclang 9:
[64] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [19789] VAR 'bpf_prog_active' type_id=64, linkage=global-allocgcc 10.2
[18] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [20319] VAR 'bpf_prog_active' type_id=18, linkage=global-allocgcc 9.3:
[21] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [21085] VAR 'bpf_prog_active' type_id=21, linkage=global-allocgcc 8
[21] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [21084] VAR 'bpf_prog_active' type_id=21, linkage=global-allocQuoted 7 lines
gcc 6.2
[22] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [21083] VAR 'bpf_prog_active' type_id=22, linkage=global-allocgcc 4.9
[17] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [20410] VAR 'bpf_prog_active' type_id=17, linkage=global-alloc--
Sent from my Android device with K-9 Mail. Please excuse my brevity.
Em Thu, Oct 01, 2020 at 08:47:51AM -0700, Alexei Starovoitov escreveu:
Quoted 5 lines
Arnaldo,
ping.
Is anything blocking this fix from merging? The kernel patches are stalled waiting on the pahole.
Applied locally, testing now, will push to the main branch ASAP.
- Arnaldo
Quoted 181 lines
Thanks
Quoted 179 lines · Hao Luo
Arnaldo,
Is this patch ready to be merged into Pahole's master branch? Alexei is testing the kernel patches that need this patch. Please let me know if there is anything I can do to help merging.
Thank you,
HaoQuoted 167 lines · Arnaldo Carvalho de Melo
Quoted 77 lines · Hao Luo
Arnaldo,
Quoted 9 lines · Arnaldo Carvalho de Melo
Em Mon, Aug 24, 2020 at 05:45:23PM -0700, Hao Luo escreveu:
Quoted 7 lines
It is found on gcc 8.2 that global percpu variables generate the following dwarf entry in the cu where the variable is defined[1].
Take the global variable "
bpf_prog_active" defined inkernel/bpf/syscall.cas an example. The debug info forsyscall.chas two dwarf entries for "bpf_prog_active".[...]
Quoted 4 lines
Interesting, here I get, with binutils' readelf:
[root@quaco perf]# readelf -wi
../build/v5.8-rc5+/kernel/bpf/syscall.o| grepbpf_prog_activeQuoted 1 line
<f6a1> DW_AT_name : (indirect string, offset: 0xb70d):
bpf_prog_activeQuoted 5 lines
[root@quaco perf]#
Just one, as:
[root@quaco perf]# readelf -wi
../build/v5.8-rc5+/kernel/bpf/syscall.o| grepbpf_prog_active-B1 -A8Quoted 2 lines
<1><f6a0>: Abbrev Number: 103 (
DW_TAG_variable)<f6a1> DW_AT_name : (indirect string, offset: 0xb70d):
bpf_prog_activeQuoted 8 lines
<f6a5> DW_AT_decl_file : 11 <f6a6> DW_AT_decl_line : 1008 <f6a8> DW_AT_decl_column : 1 <f6a9> DW_AT_type : <0xcf> <f6ad> DW_AT_external : 1 <f6ad> DW_AT_declaration : 1 <1><f6ad>: Abbrev Number: 103 (DW_TAG_variable) <f6ae> DW_AT_name : (indirect string, offset: 0x3a5d):
bpf_stats_enabled_mutexQuoted 5 lines
[root@quaco perf]#
I get what you have when I use elfutils' readelf:
[root@quaco perf]# eu-readelf -winfo
../build/v5.8-rc5+/kernel/bpf/syscall.o| grepbpf_prog_activeQuoted 5 lines
name (strp) "
bpf_prog_active"
[ 0] addr .data..percpu+0 <bpf_prog_active>
[root@quaco perf]#[root@quaco perf]# eu-readelf -winfo
../build/v5.8-rc5+/kernel/bpf/syscall.o| grep -B1 -A8 \"bpf_prog_active\"Quoted 10 lines
[ f6a0] variable abbrev: 103 name (strp) "bpf_prog_active" decl_file (data1) bpf.h (11) decl_line (data2) 1008 decl_column (data1) 1 type (ref4) [ cf] external (flag_present) yes declaration (flag_present) yes [ f6ad] variable abbrev: 103 name (strp) "bpf_stats_enabled_mutex"Quoted 5 lines
[root@quaco perf]#
And:
[root@quaco perf]# eu-readelf -winfo
../build/v5.8-rc5+/kernel/bpf/syscall.o| grep -B5 \<bpf_prog_active\>Quoted 8 lines
[ 1bdf5] variable abbrev: 212 specification (ref4) [ f6a0] decl_file (data1) syscall.c (1) decl_line (data1) 43 location (exprloc) [ 0] addr .data..percpu+0 <bpf_prog_active>[root@quaco perf]#
Quoted 1 line
In binutils readelf, there is a extra entry
Not here, tomorrow I'll triple check.
Quoted 34 lines
<1><1b24c>: Abbrev Number: 195 (
DW_TAG_variable)<1b24e> DW_AT_specification: <0xf335> <1b252> DW_AT_decl_file : 1 <1b253> DW_AT_decl_line : 43 <1b254> DW_AT_location : 9 byte block: 3 0 0 0 0 0 0 0 0(
DW_OP_addr: 0)which points to
<1><f335>: Abbrev Number: 95 (
DW_TAG_variable)
<f336>DW_AT_name: (indirect string, offset:0xb37a):bpf_prog_activeIt just doesn't have the string '
bpf_prog_active', annotating entry. So eu-readelf and binutils readelf have the same results.Quoted 5 lines
Quoted 2 lines
Note that second
DW_TAG_variableentry contains specification that points to the first entry.So you are not considering the first when encoding since it is just a
DW_AT_declaration, considers the second, as it should be, and thenneeds
Quoted 4 lines
to go see its
DW_AT_specification, right?Sounds correct, applying, will test further and then push out,
Yes, exactly. The var tags to be considered are those that either have
DW_AT_specificationor not haveDW_AT_declaration. This makes surebtf_encoderworks correctly on both old and new gcc.Quoted 2 lines
Thanks,
Quoted 8 lines
Quoted 1 line
- Arnaldo
Suggested by Yonghong, I tested this change on a larger set of compilers this time and works correctly. See below.
Could you also add 'Reported-by: Yonghong Song <[email protected]>'? I should have done that when sending out this patch. The credit goes to Yonghong.
Sure, and I'll add your results with different computers, for the record.
Thanks,
- Arnaldo
Quoted 24 lines
Thank you,
Haoclang 10:
[67] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [20168] VAR 'bpf_prog_active' type_id=67, linkage=global-allocclang 9:
[64] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [19789] VAR 'bpf_prog_active' type_id=64, linkage=global-allocgcc 10.2
[18] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [20319] VAR 'bpf_prog_active' type_id=18, linkage=global-allocgcc 9.3:
[21] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [21085] VAR 'bpf_prog_active' type_id=21, linkage=global-allocgcc 8
[21] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [21084] VAR 'bpf_prog_active' type_id=21, linkage=global-allocQuoted 7 lines
gcc 6.2
[22] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [21083] VAR 'bpf_prog_active' type_id=22, linkage=global-allocgcc 4.9
[17] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [20410] VAR 'bpf_prog_active' type_id=17, linkage=global-alloc--
Sent from my Android device with K-9 Mail. Please excuse my brevity.
Signature · 1 line
- Arnaldo
Thanks!
Quoted 197 lines · Arnaldo Carvalho de Melo
Em Thu, Oct 01, 2020 at 08:47:51AM -0700, Alexei Starovoitov escreveu:
Quoted 5 lines
Arnaldo,
ping.
Is anything blocking this fix from merging? The kernel patches are stalled waiting on the pahole.Applied locally, testing now, will push to the main branch ASAP.
- Arnaldo
Quoted 181 lines
Thanks
Quoted 179 lines · Hao Luo
Arnaldo,
Is this patch ready to be merged into Pahole's master branch? Alexei is testing the kernel patches that need this patch. Please let me know if there is anything I can do to help merging.
Thank you,
HaoQuoted 167 lines · Arnaldo Carvalho de Melo
Quoted 77 lines · Hao Luo
Arnaldo,
Quoted 9 lines · Arnaldo Carvalho de Melo
Em Mon, Aug 24, 2020 at 05:45:23PM -0700, Hao Luo escreveu:
Quoted 7 lines
It is found on gcc 8.2 that global percpu variables generate the following dwarf entry in the cu where the variable is defined[1].
Take the global variable "
bpf_prog_active" defined inkernel/bpf/syscall.cas an example. The debug info forsyscall.chas two dwarf entries for "bpf_prog_active".[...]
Quoted 4 lines
Interesting, here I get, with binutils' readelf:
[root@quaco perf]# readelf -wi
../build/v5.8-rc5+/kernel/bpf/syscall.o| grepbpf_prog_activeQuoted 1 line
<f6a1> DW_AT_name : (indirect string, offset: 0xb70d):
bpf_prog_activeQuoted 5 lines
[root@quaco perf]#
Just one, as:
[root@quaco perf]# readelf -wi
../build/v5.8-rc5+/kernel/bpf/syscall.o| grepbpf_prog_active-B1 -A8Quoted 2 lines
<1><f6a0>: Abbrev Number: 103 (
DW_TAG_variable)<f6a1> DW_AT_name : (indirect string, offset: 0xb70d):
bpf_prog_activeQuoted 8 lines
<f6a5> DW_AT_decl_file : 11 <f6a6> DW_AT_decl_line : 1008 <f6a8> DW_AT_decl_column : 1 <f6a9> DW_AT_type : <0xcf> <f6ad> DW_AT_external : 1 <f6ad> DW_AT_declaration : 1 <1><f6ad>: Abbrev Number: 103 (DW_TAG_variable) <f6ae> DW_AT_name : (indirect string, offset: 0x3a5d):
bpf_stats_enabled_mutexQuoted 5 lines
[root@quaco perf]#
I get what you have when I use elfutils' readelf:
[root@quaco perf]# eu-readelf -winfo
../build/v5.8-rc5+/kernel/bpf/syscall.o| grepbpf_prog_activeQuoted 5 lines
name (strp) "
bpf_prog_active"
[ 0] addr .data..percpu+0 <bpf_prog_active>
[root@quaco perf]#[root@quaco perf]# eu-readelf -winfo
../build/v5.8-rc5+/kernel/bpf/syscall.o| grep -B1 -A8 \"bpf_prog_active\"Quoted 10 lines
[ f6a0] variable abbrev: 103 name (strp) "bpf_prog_active" decl_file (data1) bpf.h (11) decl_line (data2) 1008 decl_column (data1) 1 type (ref4) [ cf] external (flag_present) yes declaration (flag_present) yes [ f6ad] variable abbrev: 103 name (strp) "bpf_stats_enabled_mutex"Quoted 5 lines
[root@quaco perf]#
And:
[root@quaco perf]# eu-readelf -winfo
../build/v5.8-rc5+/kernel/bpf/syscall.o| grep -B5 \<bpf_prog_active\>Quoted 8 lines
[ 1bdf5] variable abbrev: 212 specification (ref4) [ f6a0] decl_file (data1) syscall.c (1) decl_line (data1) 43 location (exprloc) [ 0] addr .data..percpu+0 <bpf_prog_active>[root@quaco perf]#
Quoted 1 line
In binutils readelf, there is a extra entry
Not here, tomorrow I'll triple check.
Quoted 34 lines
<1><1b24c>: Abbrev Number: 195 (
DW_TAG_variable)<1b24e> DW_AT_specification: <0xf335> <1b252> DW_AT_decl_file : 1 <1b253> DW_AT_decl_line : 43 <1b254> DW_AT_location : 9 byte block: 3 0 0 0 0 0 0 0 0(
DW_OP_addr: 0)which points to
<1><f335>: Abbrev Number: 95 (
DW_TAG_variable)
<f336>DW_AT_name: (indirect string, offset:0xb37a):bpf_prog_activeIt just doesn't have the string '
bpf_prog_active', annotating entry. So eu-readelf and binutils readelf have the same results.Quoted 5 lines
Quoted 2 lines
Note that second
DW_TAG_variableentry contains specification that points to the first entry.So you are not considering the first when encoding since it is just a
DW_AT_declaration, considers the second, as it should be, and thenneeds
Quoted 4 lines
to go see its
DW_AT_specification, right?Sounds correct, applying, will test further and then push out,
Yes, exactly. The var tags to be considered are those that either have
DW_AT_specificationor not haveDW_AT_declaration. This makes surebtf_encoderworks correctly on both old and new gcc.Quoted 2 lines
Thanks,
Quoted 8 lines
Quoted 1 line
- Arnaldo
Suggested by Yonghong, I tested this change on a larger set of compilers this time and works correctly. See below.
Could you also add 'Reported-by: Yonghong Song <[email protected]>'? I should have done that when sending out this patch. The credit goes to Yonghong.
Sure, and I'll add your results with different computers, for the record.
Thanks,
- Arnaldo
Quoted 24 lines
Thank you,
Haoclang 10:
[67] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [20168] VAR 'bpf_prog_active' type_id=67, linkage=global-allocclang 9:
[64] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [19789] VAR 'bpf_prog_active' type_id=64, linkage=global-allocgcc 10.2
[18] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [20319] VAR 'bpf_prog_active' type_id=18, linkage=global-allocgcc 9.3:
[21] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [21085] VAR 'bpf_prog_active' type_id=21, linkage=global-allocgcc 8
[21] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [21084] VAR 'bpf_prog_active' type_id=21, linkage=global-allocQuoted 7 lines
gcc 6.2
[22] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [21083] VAR 'bpf_prog_active' type_id=22, linkage=global-allocgcc 4.9
[17] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [20410] VAR 'bpf_prog_active' type_id=17, linkage=global-alloc--
Sent from my Android device with K-9 Mail. Please excuse my brevity.--
- Arnaldo
Em Thu, Oct 01, 2020 at 11:40:25AM -0700, Hao Luo escreveu:
Quoted 1 line
Thanks!
I must just apologise not having this in an officially released version yet, getting constantly postponed due to bug reports about corner cases and some features I got carried away working on, I'll fast pace a new version to avoid getting in the way of the larger eBPF effort.
- Arnaldo
Quoted 199 lines
Quoted 197 lines · Arnaldo Carvalho de Melo
Em Thu, Oct 01, 2020 at 08:47:51AM -0700, Alexei Starovoitov escreveu:
Quoted 5 lines
Arnaldo,
ping.
Is anything blocking this fix from merging? The kernel patches are stalled waiting on the pahole.Applied locally, testing now, will push to the main branch ASAP.
- Arnaldo
Quoted 181 lines
Thanks
Quoted 179 lines · Hao Luo
Arnaldo,
Is this patch ready to be merged into Pahole's master branch? Alexei is testing the kernel patches that need this patch. Please let me know if there is anything I can do to help merging.
Thank you,
HaoQuoted 167 lines · Arnaldo Carvalho de Melo
Quoted 77 lines · Hao Luo
Arnaldo,
Quoted 9 lines · Arnaldo Carvalho de Melo
Em Mon, Aug 24, 2020 at 05:45:23PM -0700, Hao Luo escreveu:
Quoted 7 lines
It is found on gcc 8.2 that global percpu variables generate the following dwarf entry in the cu where the variable is defined[1].
Take the global variable "
bpf_prog_active" defined inkernel/bpf/syscall.cas an example. The debug info forsyscall.chas two dwarf entries for "bpf_prog_active".[...]
Quoted 4 lines
Interesting, here I get, with binutils' readelf:
[root@quaco perf]# readelf -wi
../build/v5.8-rc5+/kernel/bpf/syscall.o| grepbpf_prog_activeQuoted 1 line
<f6a1> DW_AT_name : (indirect string, offset: 0xb70d):
bpf_prog_activeQuoted 5 lines
[root@quaco perf]#
Just one, as:
[root@quaco perf]# readelf -wi
../build/v5.8-rc5+/kernel/bpf/syscall.o| grepbpf_prog_active-B1 -A8Quoted 2 lines
<1><f6a0>: Abbrev Number: 103 (
DW_TAG_variable)<f6a1> DW_AT_name : (indirect string, offset: 0xb70d):
bpf_prog_activeQuoted 8 lines
<f6a5> DW_AT_decl_file : 11 <f6a6> DW_AT_decl_line : 1008 <f6a8> DW_AT_decl_column : 1 <f6a9> DW_AT_type : <0xcf> <f6ad> DW_AT_external : 1 <f6ad> DW_AT_declaration : 1 <1><f6ad>: Abbrev Number: 103 (DW_TAG_variable) <f6ae> DW_AT_name : (indirect string, offset: 0x3a5d):
bpf_stats_enabled_mutexQuoted 5 lines
[root@quaco perf]#
I get what you have when I use elfutils' readelf:
[root@quaco perf]# eu-readelf -winfo
../build/v5.8-rc5+/kernel/bpf/syscall.o| grepbpf_prog_activeQuoted 5 lines
name (strp) "
bpf_prog_active"
[ 0] addr .data..percpu+0 <bpf_prog_active>
[root@quaco perf]#[root@quaco perf]# eu-readelf -winfo
../build/v5.8-rc5+/kernel/bpf/syscall.o| grep -B1 -A8 \"bpf_prog_active\"Quoted 10 lines
[ f6a0] variable abbrev: 103 name (strp) "bpf_prog_active" decl_file (data1) bpf.h (11) decl_line (data2) 1008 decl_column (data1) 1 type (ref4) [ cf] external (flag_present) yes declaration (flag_present) yes [ f6ad] variable abbrev: 103 name (strp) "bpf_stats_enabled_mutex"Quoted 5 lines
[root@quaco perf]#
And:
[root@quaco perf]# eu-readelf -winfo
../build/v5.8-rc5+/kernel/bpf/syscall.o| grep -B5 \<bpf_prog_active\>Quoted 8 lines
[ 1bdf5] variable abbrev: 212 specification (ref4) [ f6a0] decl_file (data1) syscall.c (1) decl_line (data1) 43 location (exprloc) [ 0] addr .data..percpu+0 <bpf_prog_active>[root@quaco perf]#
Quoted 1 line
In binutils readelf, there is a extra entry
Not here, tomorrow I'll triple check.
Quoted 34 lines
<1><1b24c>: Abbrev Number: 195 (
DW_TAG_variable)<1b24e> DW_AT_specification: <0xf335> <1b252> DW_AT_decl_file : 1 <1b253> DW_AT_decl_line : 43 <1b254> DW_AT_location : 9 byte block: 3 0 0 0 0 0 0 0 0(
DW_OP_addr: 0)which points to
<1><f335>: Abbrev Number: 95 (
DW_TAG_variable)
<f336>DW_AT_name: (indirect string, offset:0xb37a):bpf_prog_activeIt just doesn't have the string '
bpf_prog_active', annotating entry. So eu-readelf and binutils readelf have the same results.Quoted 5 lines
Quoted 2 lines
Note that second
DW_TAG_variableentry contains specification that points to the first entry.So you are not considering the first when encoding since it is just a
DW_AT_declaration, considers the second, as it should be, and thenneeds
Quoted 4 lines
to go see its
DW_AT_specification, right?Sounds correct, applying, will test further and then push out,
Yes, exactly. The var tags to be considered are those that either have
DW_AT_specificationor not haveDW_AT_declaration. This makes surebtf_encoderworks correctly on both old and new gcc.Quoted 2 lines
Thanks,
Quoted 8 lines
Quoted 1 line
- Arnaldo
Suggested by Yonghong, I tested this change on a larger set of compilers this time and works correctly. See below.
Could you also add 'Reported-by: Yonghong Song <[email protected]>'? I should have done that when sending out this patch. The credit goes to Yonghong.
Sure, and I'll add your results with different computers, for the record.
Thanks,
- Arnaldo
Quoted 24 lines
Thank you,
Haoclang 10:
[67] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [20168] VAR 'bpf_prog_active' type_id=67, linkage=global-allocclang 9:
[64] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [19789] VAR 'bpf_prog_active' type_id=64, linkage=global-allocgcc 10.2
[18] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [20319] VAR 'bpf_prog_active' type_id=18, linkage=global-allocgcc 9.3:
[21] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [21085] VAR 'bpf_prog_active' type_id=21, linkage=global-allocgcc 8
[21] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [21084] VAR 'bpf_prog_active' type_id=21, linkage=global-allocQuoted 7 lines
gcc 6.2
[22] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [21083] VAR 'bpf_prog_active' type_id=22, linkage=global-allocgcc 4.9
[17] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [20410] VAR 'bpf_prog_active' type_id=17, linkage=global-alloc--
Sent from my Android device with K-9 Mail. Please excuse my brevity.--
- Arnaldo
Signature · 1 line
- Arnaldo
Arnaldo, thanks for the update. In that case, I think on the kernel side I need to skip encoding percpu vars for this pahole release, and re-enable for the next pahole release. (assuming the flag for opt-out is in this release). Alexei, do you have any better idea?
Hao
Quoted 214 lines · Arnaldo Carvalho de Melo
Em Thu, Oct 01, 2020 at 11:40:25AM -0700, Hao Luo escreveu:
Quoted 1 line
Thanks!
I must just apologise not having this in an officially released version yet, getting constantly postponed due to bug reports about corner cases and some features I got carried away working on, I'll fast pace a new version to avoid getting in the way of the larger eBPF effort.
- Arnaldo
Quoted 193 lines
Quoted 191 lines · Arnaldo Carvalho de Melo
Em Thu, Oct 01, 2020 at 08:47:51AM -0700, Alexei Starovoitov escreveu:
Quoted 5 lines
Arnaldo,
ping.
Is anything blocking this fix from merging? The kernel patches are stalled waiting on the pahole.Applied locally, testing now, will push to the main branch ASAP.
- Arnaldo
Quoted 179 lines
Thanks
Quoted 177 lines · Hao Luo
Arnaldo,
Is this patch ready to be merged into Pahole's master branch? Alexei is testing the kernel patches that need this patch. Please let me know if there is anything I can do to help merging.
Thank you,
HaoQuoted 165 lines · Arnaldo Carvalho de Melo
Quoted 77 lines · Hao Luo
Arnaldo,
Quoted 9 lines · Arnaldo Carvalho de Melo
Em Mon, Aug 24, 2020 at 05:45:23PM -0700, Hao Luo escreveu:
Quoted 7 lines
It is found on gcc 8.2 that global percpu variables generate the following dwarf entry in the cu where the variable is defined[1].
Take the global variable "
bpf_prog_active" defined inkernel/bpf/syscall.cas an example. The debug info forsyscall.chas two dwarf entries for "bpf_prog_active".[...]
Quoted 4 lines
Interesting, here I get, with binutils' readelf:
[root@quaco perf]# readelf -wi
../build/v5.8-rc5+/kernel/bpf/syscall.o| grepbpf_prog_activeQuoted 1 line
<f6a1> DW_AT_name : (indirect string, offset: 0xb70d):
bpf_prog_activeQuoted 5 lines
[root@quaco perf]#
Just one, as:
[root@quaco perf]# readelf -wi
../build/v5.8-rc5+/kernel/bpf/syscall.o| grepbpf_prog_active-B1 -A8Quoted 2 lines
<1><f6a0>: Abbrev Number: 103 (
DW_TAG_variable)<f6a1> DW_AT_name : (indirect string, offset: 0xb70d):
bpf_prog_activeQuoted 8 lines
<f6a5> DW_AT_decl_file : 11 <f6a6> DW_AT_decl_line : 1008 <f6a8> DW_AT_decl_column : 1 <f6a9> DW_AT_type : <0xcf> <f6ad> DW_AT_external : 1 <f6ad> DW_AT_declaration : 1 <1><f6ad>: Abbrev Number: 103 (DW_TAG_variable) <f6ae> DW_AT_name : (indirect string, offset: 0x3a5d):
bpf_stats_enabled_mutexQuoted 5 lines
[root@quaco perf]#
I get what you have when I use elfutils' readelf:
[root@quaco perf]# eu-readelf -winfo
../build/v5.8-rc5+/kernel/bpf/syscall.o| grepbpf_prog_activeQuoted 5 lines
name (strp) "
bpf_prog_active"
[ 0] addr .data..percpu+0 <bpf_prog_active>
[root@quaco perf]#[root@quaco perf]# eu-readelf -winfo
../build/v5.8-rc5+/kernel/bpf/syscall.o| grep -B1 -A8 \"bpf_prog_active\"Quoted 10 lines
[ f6a0] variable abbrev: 103 name (strp) "bpf_prog_active" decl_file (data1) bpf.h (11) decl_line (data2) 1008 decl_column (data1) 1 type (ref4) [ cf] external (flag_present) yes declaration (flag_present) yes [ f6ad] variable abbrev: 103 name (strp) "bpf_stats_enabled_mutex"Quoted 5 lines
[root@quaco perf]#
And:
[root@quaco perf]# eu-readelf -winfo
../build/v5.8-rc5+/kernel/bpf/syscall.o| grep -B5 \<bpf_prog_active\>Quoted 8 lines
[ 1bdf5] variable abbrev: 212 specification (ref4) [ f6a0] decl_file (data1) syscall.c (1) decl_line (data1) 43 location (exprloc) [ 0] addr .data..percpu+0 <bpf_prog_active>[root@quaco perf]#
Quoted 1 line
In binutils readelf, there is a extra entry
Not here, tomorrow I'll triple check.
Quoted 34 lines
<1><1b24c>: Abbrev Number: 195 (
DW_TAG_variable)<1b24e> DW_AT_specification: <0xf335> <1b252> DW_AT_decl_file : 1 <1b253> DW_AT_decl_line : 43 <1b254> DW_AT_location : 9 byte block: 3 0 0 0 0 0 0 0 0(
DW_OP_addr: 0)which points to
<1><f335>: Abbrev Number: 95 (
DW_TAG_variable)
<f336>DW_AT_name: (indirect string, offset:0xb37a):bpf_prog_activeIt just doesn't have the string '
bpf_prog_active', annotating entry. So eu-readelf and binutils readelf have the same results.Quoted 5 lines
Quoted 2 lines
Note that second
DW_TAG_variableentry contains specification that points to the first entry.So you are not considering the first when encoding since it is just a
DW_AT_declaration, considers the second, as it should be, and thenneeds
Quoted 4 lines
to go see its
DW_AT_specification, right?Sounds correct, applying, will test further and then push out,
Yes, exactly. The var tags to be considered are those that either have
DW_AT_specificationor not haveDW_AT_declaration. This makes surebtf_encoderworks correctly on both old and new gcc.Quoted 2 lines
Thanks,
Quoted 8 lines
Quoted 1 line
- Arnaldo
Suggested by Yonghong, I tested this change on a larger set of compilers this time and works correctly. See below.
Could you also add 'Reported-by: Yonghong Song <[email protected]>'? I should have done that when sending out this patch. The credit goes to Yonghong.
Sure, and I'll add your results with different computers, for the record.
Thanks,
- Arnaldo
Quoted 31 lines
Thank you,
Haoclang 10:
[67] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [20168] VAR 'bpf_prog_active' type_id=67, linkage=global-allocclang 9:
[64] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [19789] VAR 'bpf_prog_active' type_id=64, linkage=global-allocgcc 10.2
[18] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [20319] VAR 'bpf_prog_active' type_id=18, linkage=global-allocgcc 9.3:
[21] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [21085] VAR 'bpf_prog_active' type_id=21, linkage=global-allocgcc 8
[21] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [21084] VAR 'bpf_prog_active' type_id=21, linkage=global-allocgcc 6.2
[22] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [21083] VAR 'bpf_prog_active' type_id=22, linkage=global-allocgcc 4.9
[17] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED [20410] VAR 'bpf_prog_active' type_id=17, linkage=global-allocQuoted 6 lines
Quoted 6 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
--
Sent from my Android device with K-9 Mail. Please excuse my brevity.--
- Arnaldo
--
- Arnaldo
Quoted 5 lines · Hao Luo
Arnaldo, thanks for the update. In that case, I think on the kernel side I need to skip encoding percpu vars for this pahole release, and re-enable for the next pahole release. (assuming the flag for opt-out is in this release). Alexei, do you have any better idea?
I'm not following. Let's get this fix landed in pahole and release new 1.18 with it.
The opt-out flag is orthogonal. I can be done in 1.19 or whenever. With your kernel patches the kernel will reject percpu vars when pahole is too old, because they will not be found in vmlinux btf,
so I don't see any compatibility issues.
There is no need to bump the required min version of pahole in scripts/link-vmlinux.sh.
It can stay as v1.16. We only need clean verifier message that percpu BTF is not found and the kernel needs to be rebuilt with pahole 1.18.
Quoted 10 lines · Alexei Starovoitov
Quoted 5 lines · Hao Luo
Arnaldo, thanks for the update. In that case, I think on the kernel side I need to skip encoding percpu vars for this pahole release, and re-enable for the next pahole release. (assuming the flag for opt-out is in this release). Alexei, do you have any better idea?
I'm not following. Let's get this fix landed in pahole and release new 1.18 with it.
Sorry, I misunderstood Arnaldo. I thought what he was saying was, not having this fix in an officially released version (I took this as the new 1.18) and will get 1.19 quicker. It would be ideal if we can have this in 1.18. Sorry for my bad reading.
Quoted 8 lines
The opt-out flag is orthogonal. I can be done in 1.19 or whenever. With your kernel patches the kernel will reject percpu vars when pahole is too old, because they will not be found in vmlinux btf,
so I don't see any compatibility issues.
There is no need to bump the required min version of pahole inscripts/link-vmlinux.sh.
It can stay as v1.16. We only need clean verifier message that percpu BTF is not found and the kernel needs to be rebuilt with pahole 1.18.