btf_encoder: Fix functions BTF data generation
Subsystems (2)
Jiri Olsa
reviewerv2 cover btf_encoder: Fix functions BTF data generation
hi,
recent btf encoder's changes brakes BTF data for some gcc versions. The problem is that some functions can appear in dwarf data in some instances without arguments, while they are defined with some.
v2 changes:
- drop patch 3 logic and just change conditions based on Andrii's suggestion
- drop patch 2
- add ack for patch 1
Andrii,
please check if it's working with your gcc.
thanks,
jirka
---
Jiri Olsa (2):
btf_encoder: Generate also .init functions
btf_encoder: Fix function generation
btf_encoder.c | 67 ++++++++++++------------------------------------------------------- 1 file changed, 12 insertions(+), 55 deletions(-)
Jiri Olsa
reviewerPATCH v2 1/2 btf_encoder: Generate also .init functions
Currently we skip functions under .init* sections, Removing the .init* section check, BTF now contains also functions from .init* sections.
Andrii's explanation from email:
Quoted 9 lines
... I think we should just drop the
__initcheck and
include all the__initfunctions into BTF. There could be cases where we'd need to attach BPF programs to__initfunctions (e.g.,bpf_lsmsecurity cases), so having BTFs for those FUNCs are necessary as well. Ftrace currently disallows that, but it's only because no user-space application has a way to attach probes early enough. This might change in the future, so there is no need to invent special mechanisms now forbpf_iterfunction preservation. Let's just include all__initfunctions in BTF.
It's over ~2000 functions on my .config:
$ bpftool btf dump file ./vmlinux | grep 'FUNC ' | wc -l 41505 $ bpftool btf dump file /sys/kernel/btf/vmlinux | grep 'FUNC ' | wc -l 39256
--- btf_encoder.c | 43 ++----------------------------------------- 1 file changed, 2 insertions(+), 41 deletions(-)
btf_encoder.c +2 -41
@@ -29,10 +29,6 @@struct funcs_layout {unsigned long mcount_start;unsigned long mcount_stop; -unsigned long init_begin; -unsigned long init_end; -unsigned long init_bpf_begin; -unsigned long init_bpf_end;unsigned long mcount_sec_idx; };
@@ -104,16 +100,6 @@ static int addrs_cmp(const void *_a, const void *_b)return *a < *b ?-1 :1 ; } -static bool is_init (struct funcs_layout *fl,unsigned long addr) -{ -return addr >= fl->init_begin && addr < fl->init_end; -} - -static bool is_bpf_init (struct funcs_layout *fl,unsigned long addr) -{ -return addr >= fl->init_bpf_begin && addr < fl->init_bpf_end; -} -static int filter_functions (struct btf_elf *btfe,struct funcs_layout *fl) {unsigned long *addrs, count, offset, i;
@@ -155,18 +141,11 @@ static int filter_functions(struct btf_elf *btfe, struct funcs_layout *fl)/* * Let's got through all collected functions and filter -* out those that are not in ftrace and init code. +* out those that are not in ftrace. */ for (i =0 ; i < functions_cnt; i++) {struct elf_function *func = &functions[i]; -/* -* Do not enable .init section functions, -* but keep .init.bpf.preserve_type functions. -*/ -if (is_init (fl, func->addr) && !is_bpf_init (fl, func->addr)) -continue ; -/* Make sure function is within ftrace addresses. */ if (bsearch (&func->addr, addrs, count,sizeof (addrs[0 ]), addrs_cmp)) {/*
@@ -493,29 +472,11 @@ static void collect_symbol(GElf_Sym *sym, struct funcs_layout *fl)if (!fl->mcount_stop && !strcmp ("__stop_mcount_loc" ,elf_sym__name (sym, btfe->symtab))) fl->mcount_stop = sym->st_value; - -if (!fl->init_begin && - !strcmp ("__init_begin" ,elf_sym__name (sym, btfe->symtab))) - fl->init_begin = sym->st_value; - -if (!fl->init_end && - !strcmp ("__init_end" ,elf_sym__name (sym, btfe->symtab))) - fl->init_end = sym->st_value; - -if (!fl->init_bpf_begin && - !strcmp ("__init_bpf_preserve_type_begin" ,elf_sym__name (sym, btfe->symtab))) - fl->init_bpf_begin = sym->st_value; - -if (!fl->init_bpf_end && - !strcmp ("__init_bpf_preserve_type_end" ,elf_sym__name (sym, btfe->symtab))) - fl->init_bpf_end = sym->st_value; }static int has_all_symbols (struct funcs_layout *fl) { -return fl->mcount_start && fl->mcount_stop && - fl->init_begin && fl->init_end && - fl->init_bpf_begin && fl->init_bpf_end; +return fl->mcount_start && fl->mcount_stop; }static int collect_symbols (struct btf_elf *btfe,bool collect_percpu_vars)
Signature · 1 line
2.26.2
Jiri Olsa
reviewerPATCH v2 2/2 btf_encoder: Fix function generation
In reply to Jiri OlsaCurrent conditions for picking up function records break BTF data on some gcc versions.
Some function records can appear with no arguments but with declaration tag set, so moving the 'fn->declaration' in front of other checks.
Then checking if argument names are present and finally checking ftrace filter if it's present. If ftrace filter is not available, using the external tag to filter out non external functions.
--- btf_encoder.c | 24 ++++++++++-------------- 1 file changed, 10 insertions(+), 14 deletions(-)
btf_encoder.c +10 -14
@@ -612,25 +612,21 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,const char *name;/* -* The functions_cnt != 0 means we parsed all necessary -* kernel symbols and we are using ftrace location filter -* for functions. If it's not available keep the current -* dwarf declaration check. +* Skip functions that: +* - are marked as declarations +* - do not have full argument names +* - are not in ftrace list (if it's available) +* - are not external (in case ftrace filter is not available) */ +if (fn->declaration) +continue ; +if (!has_arg_names (cu, &fn->proto)) +continue ;if (functions_cnt) { -/* -* We check following conditions: -* - argument names are defined -* - there's symbol and address defined for the function -* - function address belongs to ftrace locations -* - function is generated only once -*/ -if (!has_arg_names (cu, &fn->proto)) -continue ;if (!should_generate_function (btfe,function__name (fn, cu)))continue ; }else { -if (fn->declaration || !fn->external) +if (!fn->external)continue ; }
Signature · 1 line
2.26.2
Re: [PATCH 2/2] btf_encoder: Fix function generation
Em Fri, Nov 13, 2020 at 04:12:22PM +0100, Jiri Olsa escreveu:
Quoted 10 lines
Current conditions for picking up function records break BTF data on some gcc versions.
Some function records can appear with no arguments but with declaration tag set, so moving the '
fn->declaration' in front of other checks.Then checking if argument names are present and finally checking ftrace filter if it's present. If ftrace filter is not available, using the external tag to filter out non external functions.
Humm has_arg_names() will return true for a:
void foo(void)
function, which I think is right, but can't this function appear multiple times in different CUs and we end up with the same function multiple times in the BTF info?
- Arnaldo
Quoted 48 lines
Signed-off-by Jiri Olsa <[email protected]>--- btf_encoder.c | 24 ++++++++++-------------- 1 file changed, 10 insertions(+), 14 deletions(-)btf_encoder.c +10
-14@@ -612,25 +612,21 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,const char *name;/* -* The functions_cnt != 0 means we parsed all necessary -* kernel symbols and we are using ftrace location filter -* for functions. If it's not available keep the current -* dwarf declaration check. +* Skip functions that: +* - are marked as declarations +* - do not have full argument names +* - are not in ftrace list (if it's available) +* - are not external (in case ftrace filter is not available) */ +if (fn->declaration) +continue ; +if (!has_arg_names (cu, &fn->proto)) +continue ;if (functions_cnt) { -/* -* We check following conditions: -* - argument names are defined -* - there's symbol and address defined for the function -* - function address belongs to ftrace locations -* - function is generated only once -*/ -if (!has_arg_names (cu, &fn->proto)) -continue ;if (!should_generate_function (btfe,function__name (fn, cu)))continue ; }else { -if (fn->declaration || !fn->external) +if (!fn->external)continue ; }Signature · 1 line
2.26.2
Signature · 1 line
- Arnaldo
Re: [PATCH 2/2] btf_encoder: Fix function generation
Quoted 12 lines · Arnaldo Carvalho de Melo
Em Fri, Nov 13, 2020 at 04:12:22PM +0100, Jiri Olsa escreveu:
Quoted 10 lines
Current conditions for picking up function records break BTF data on some gcc versions.
Some function records can appear with no arguments but with declaration tag set, so moving the '
fn->declaration' in front of other checks.Then checking if argument names are present and finally checking ftrace filter if it's present. If ftrace filter is not available, using the external tag to filter out non external functions.
Quoted 7 lines
Humm
has_arg_names()will return true for a:void foo(void)
function, which I think is right, but can't this function appear multiple times in different CUs and we end up with the same function multiple times in the BTF info?
so declarations should be filtered by the fn->declaration check, if gcc has it ;-)
and if it goes through, the should_generate_function sets/checks flag if the function was already generated
jirka
Quoted 56 lines
- Arnaldo
Quoted 48 lines
Signed-off-by Jiri Olsa <[email protected]>--- btf_encoder.c | 24 ++++++++++-------------- 1 file changed, 10 insertions(+), 14 deletions(-)btf_encoder.c +10
-14@@ -612,25 +612,21 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,const char *name;/* -* The functions_cnt != 0 means we parsed all necessary -* kernel symbols and we are using ftrace location filter -* for functions. If it's not available keep the current -* dwarf declaration check. +* Skip functions that: +* - are marked as declarations +* - do not have full argument names +* - are not in ftrace list (if it's available) +* - are not external (in case ftrace filter is not available) */ +if (fn->declaration) +continue ; +if (!has_arg_names (cu, &fn->proto)) +continue ;if (functions_cnt) { -/* -* We check following conditions: -* - argument names are defined -* - there's symbol and address defined for the function -* - function address belongs to ftrace locations -* - function is generated only once -*/ -if (!has_arg_names (cu, &fn->proto)) -continue ;if (!should_generate_function (btfe,function__name (fn, cu)))continue ; }else { -if (fn->declaration || !fn->external) +if (!fn->external)continue ; }Signature · 1 line
2.26.2Signature · 1 line
- Arnaldo
Re: [PATCH 2/2] btf_encoder: Fix function generation
In reply to Jiri OlsaQuoted 8 lines · Jiri Olsa
Current conditions for picking up function records break BTF data on some gcc versions.
Some function records can appear with no arguments but with declaration tag set, so moving the '
fn->declaration' in front of other checks.
Quoted 6 lines
Then checking if argument names are present and finally checking ftrace filter if it's present. If ftrace filter is not available, using the external tag to filter out non external functions.
Signed-off-by Jiri Olsa <[email protected]>---
I tested locally, all seems to work fine. Left few suggestions below, but those could be done in follow ups (or argued to not be done).
BTW, for some stats.
BEFORE allowing static funcs:
.BTF ELF section
=======================================
Data size: 4101624 Header size: 24 Types size: 2472836 Strings size: 1628764
BTF types
=======================================
Total 2472836 bytes (83310 types)
Struct: 920436 bytes (10305 types) FuncProto: 638668 bytes (18869 types) Func: 308304 bytes (25692 types) Enum: 184308 bytes (2293 types) Ptr: 173484 bytes (14457 types) Array: 89064 bytes (3711 types) Union: 81552 bytes (1961 types) Const: 34368 bytes (2864 types) Typedef: 32124 bytes (2677 types) Var: 4688 bytes (293 types) Datasec: 3528 bytes (1 types) Fwd: 1656 bytes (138 types) Volatile: 360 bytes (30 types) Int: 272 bytes (17 types) Restrict: 24 bytes (2 types)
AFTER allowing static funcs:
.BTF ELF section
=======================================
Data size: 4930558 Header size: 24 Types size: 2914016 Strings size: 2016518
BTF types
=======================================
Total 2914016 bytes (108282 types)
Struct: 920436 bytes (10305 types) FuncProto: 851528 bytes (24814 types) Func: 536664 bytes (44722 types) Enum: 184308 bytes (2293 types) Ptr: 173484 bytes (14457 types) Array: 89064 bytes (3711 types) Union: 81552 bytes (1961 types) Const: 34368 bytes (2864 types) Typedef: 32124 bytes (2677 types) Var: 4688 bytes (293 types) Datasec: 3528 bytes (1 types) Fwd: 1656 bytes (138 types) Volatile: 360 bytes (30 types) Int: 256 bytes (16 types)
So 25692 vs 44722 functions, but the increase in func_proto is smaller due to dedup. Good chunk is strings data for all those function and parameter names.
Quoted 26 lines
btf_encoder.c | 24 ++++++++++-------------- 1 file changed, 10 insertions(+), 14 deletions(-)btf_encoder.c +9
-4@@ -612,25 +612,21 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,const char *name;/* -* The functions_cnt != 0 means we parsed all necessary -* kernel symbols and we are using ftrace location filter -* for functions. If it's not available keep the current -* dwarf declaration check. +* Skip functions that: +* - are marked as declarations +* - do not have full argument names +* - are not in ftrace list (if it's available) +* - are not external (in case ftrace filter is not available) */ +if (fn->declaration) +continue ; +if (!has_arg_names (cu, &fn->proto)) +continue ;if (functions_cnt) {
btf_encoder.c int cu__encode_btf(struct cu *cu, int verbose, bool force, · line 626
Quoted 10 lines
btf_encoder.c +0
-9-/* -* We check following conditions: -* - argument names are defined -* - there's symbol and address defined for the function -* - function address belongs to ftrace locations -* - function is generated only once -*/ -if (!has_arg_names (cu, &fn->proto)) -continue ;if (!should_generate_function (btfe,function__name (fn, cu)))
Seeing Arnaldo's confusion, I remember initially I was similarly confused. I think this p->generated = true should be moved out of should_generate_function() and done here explicitly. Let's turn should_generate_function() into find_allowed_function() or something, to encapsulate bsearch. Checking !p || p->generated could be done here explicitly.
btf_encoder.c int cu__encode_btf(struct cu *cu, int verbose, bool force, · line 629
Quoted 4 lines
btf_encoder.c +1
-1continue ; }else { -if (fn->declaration || !fn->external) +if (!fn->external)
Hm.. why didn't you drop this fallback? For non-vmlinux, do you think it's a problem to generate all FUNCs? Mostly theoretical question, though.
Quoted 6 lines
btf_encoder.c +0
-1continue ; } --2.26.2
Re: [PATCH 2/2] btf_encoder: Fix function generation
Quoted 64 lines · Andrii Nakryiko
Quoted 8 lines · Jiri Olsa
Current conditions for picking up function records break BTF data on some gcc versions.
Some function records can appear with no arguments but with declaration tag set, so moving the '
fn->declaration' in front of other checks.Quoted 6 lines
Then checking if argument names are present and finally checking ftrace filter if it's present. If ftrace filter is not available, using the external tag to filter out non external functions.
Signed-off-by Jiri Olsa <[email protected]>---
I tested locally, all seems to work fine. Left few suggestions below, but those could be done in follow ups (or argued to not be done).
Acked-by Andrii Nakryiko <[email protected]>BTW, for some stats.
BEFORE allowing static funcs:
.BTF ELF section
=======================================Data size: 4101624 Header size: 24 Types size: 2472836 Strings size: 1628764BTF types
=======================================
Total 2472836 bytes (83310 types)Struct: 920436 bytes (10305 types) FuncProto: 638668 bytes (18869 types) Func: 308304 bytes (25692 types) Enum: 184308 bytes (2293 types) Ptr: 173484 bytes (14457 types) Array: 89064 bytes (3711 types) Union: 81552 bytes (1961 types) Const: 34368 bytes (2864 types) Typedef: 32124 bytes (2677 types) Var: 4688 bytes (293 types) Datasec: 3528 bytes (1 types) Fwd: 1656 bytes (138 types) Volatile: 360 bytes (30 types) Int: 272 bytes (17 types) Restrict: 24 bytes (2 types)AFTER allowing static funcs:
.BTF ELF section
=======================================Data size: 4930558 Header size: 24 Types size: 2914016 Strings size: 2016518BTF types
=======================================
Total 2914016 bytes (108282 types)
Quoted 14 lines
Struct: 920436 bytes (10305 types) FuncProto: 851528 bytes (24814 types) Func: 536664 bytes (44722 types) Enum: 184308 bytes (2293 types) Ptr: 173484 bytes (14457 types) Array: 89064 bytes (3711 types) Union: 81552 bytes (1961 types) Const: 34368 bytes (2864 types) Typedef: 32124 bytes (2677 types) Var: 4688 bytes (293 types) Datasec: 3528 bytes (1 types) Fwd: 1656 bytes (138 types) Volatile: 360 bytes (30 types) Int: 256 bytes (16 types)
nice, is this tool somewhere in the tree?
Quoted 41 lines
So 25692 vs 44722 functions, but the increase in
func_protois smaller due to dedup. Good chunk is strings data for all those function and parameter names.Quoted 35 lines
btf_encoder.c | 24 ++++++++++-------------- 1 file changed, 10 insertions(+), 14 deletions(-)btf_encoder.c +9
-13@@ -612,25 +612,21 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,const char *name;/* -* The functions_cnt != 0 means we parsed all necessary -* kernel symbols and we are using ftrace location filter -* for functions. If it's not available keep the current -* dwarf declaration check. +* Skip functions that: +* - are marked as declarations +* - do not have full argument names +* - are not in ftrace list (if it's available) +* - are not external (in case ftrace filter is not available) */ +if (fn->declaration) +continue ; +if (!has_arg_names (cu, &fn->proto)) +continue ;if (functions_cnt) { -/* -* We check following conditions: -* - argument names are defined -* - there's symbol and address defined for the function -* - function address belongs to ftrace locations -* - function is generated only once -*/ -if (!has_arg_names (cu, &fn->proto)) -continue ;
Quoted 8 lines
btf_encoder.c int cu__encode_btf(struct cu *cu, int verbose, bool force, · line 626
Quoted 1 line
btf_encoder.c +0
-0if (!should_generate_function (btfe,function__name (fn, cu)))Seeing Arnaldo's confusion, I remember initially I was similarly confused. I think this
p->generated= true should be moved out ofshould_generate_function()and done here explicitly. Let's turnshould_generate_function()intofind_allowed_function()or something, to encapsulate bsearch. Checking !p ||p->generatedcould be done here explicitly.
ok, that should be more obvious, I'll send new version
Quoted 9 lines
btf_encoder.c int cu__encode_btf(struct cu *cu, int verbose, bool force, · line 629
Quoted 5 lines
btf_encoder.c +1
-1continue ; }else { -if (fn->declaration || !fn->external) +if (!fn->external)Hm.. why didn't you drop this fallback? For non-vmlinux, do you think it's a problem to generate all FUNCs? Mostly theoretical question, though.
because it would probably allowed all static functions, (ftrace data has only static functions that are traceable) and who knows what a can of worms we'd open here ;-)
jirka
Re: [PATCH 2/2] btf_encoder: Fix function generation
Quoted 81 lines · Jiri Olsa
Quoted 64 lines · Andrii Nakryiko
Quoted 8 lines · Jiri Olsa
Current conditions for picking up function records break BTF data on some gcc versions.
Some function records can appear with no arguments but with declaration tag set, so moving the '
fn->declaration' in front of other checks.Quoted 6 lines
Then checking if argument names are present and finally checking ftrace filter if it's present. If ftrace filter is not available, using the external tag to filter out non external functions.
Signed-off-by Jiri Olsa <[email protected]>---
I tested locally, all seems to work fine. Left few suggestions below, but those could be done in follow ups (or argued to not be done).
Acked-by Andrii Nakryiko <[email protected]>BTW, for some stats.
BEFORE allowing static funcs:
.BTF ELF section
=======================================Data size: 4101624 Header size: 24 Types size: 2472836 Strings size: 1628764BTF types
=======================================
Total 2472836 bytes (83310 types)Struct: 920436 bytes (10305 types) FuncProto: 638668 bytes (18869 types) Func: 308304 bytes (25692 types) Enum: 184308 bytes (2293 types) Ptr: 173484 bytes (14457 types) Array: 89064 bytes (3711 types) Union: 81552 bytes (1961 types) Const: 34368 bytes (2864 types) Typedef: 32124 bytes (2677 types) Var: 4688 bytes (293 types) Datasec: 3528 bytes (1 types) Fwd: 1656 bytes (138 types) Volatile: 360 bytes (30 types) Int: 272 bytes (17 types) Restrict: 24 bytes (2 types)AFTER allowing static funcs:
.BTF ELF section
=======================================Data size: 4930558 Header size: 24 Types size: 2914016 Strings size: 2016518BTF types
=======================================
Total 2914016 bytes (108282 types)Quoted 14 lines
Struct: 920436 bytes (10305 types) FuncProto: 851528 bytes (24814 types) Func: 536664 bytes (44722 types) Enum: 184308 bytes (2293 types) Ptr: 173484 bytes (14457 types) Array: 89064 bytes (3711 types) Union: 81552 bytes (1961 types) Const: 34368 bytes (2864 types) Typedef: 32124 bytes (2677 types) Var: 4688 bytes (293 types) Datasec: 3528 bytes (1 types) Fwd: 1656 bytes (138 types) Volatile: 360 bytes (30 types) Int: 256 bytes (16 types)
Quoted 1 line
nice, is this tool somewhere in the tree?
Nope, this is from my personal BTF inspection tool, which I never got to open-sourcing, too low on the priority list...
Quoted 59 lines
Quoted 42 lines
So 25692 vs 44722 functions, but the increase in
func_protois smaller due to dedup. Good chunk is strings data for all those function and parameter names.Quoted 35 lines
btf_encoder.c | 24 ++++++++++-------------- 1 file changed, 10 insertions(+), 14 deletions(-)btf_encoder.c +9
-13@@ -612,25 +612,21 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,const char *name;/* -* The functions_cnt != 0 means we parsed all necessary -* kernel symbols and we are using ftrace location filter -* for functions. If it's not available keep the current -* dwarf declaration check. +* Skip functions that: +* - are marked as declarations +* - do not have full argument names +* - are not in ftrace list (if it's available) +* - are not external (in case ftrace filter is not available) */ +if (fn->declaration) +continue ; +if (!has_arg_names (cu, &fn->proto)) +continue ;if (functions_cnt) { -/* -* We check following conditions: -* - argument names are defined -* - there's symbol and address defined for the function -* - function address belongs to ftrace locations -* - function is generated only once -*/ -if (!has_arg_names (cu, &fn->proto)) -continue ;Quoted 8 lines
btf_encoder.c int cu__encode_btf(struct cu *cu, int verbose, bool force, · line 626
Quoted 1 line
btf_encoder.c +0
-0if (!should_generate_function (btfe,function__name (fn, cu)))Seeing Arnaldo's confusion, I remember initially I was similarly confused. I think this
p->generated= true should be moved out ofshould_generate_function()and done here explicitly. Let's turnshould_generate_function()intofind_allowed_function()or something, to encapsulate bsearch. Checking !p ||p->generatedcould be done here explicitly.ok, that should be more obvious, I'll send new version
Quoted 6 lines
Quoted 5 lines
btf_encoder.c +1
-1continue ; }else { -if (fn->declaration || !fn->external) +if (!fn->external)
Quoted 8 lines
Quoted 3 lines
Hm.. why didn't you drop this fallback? For non-vmlinux, do you think it's a problem to generate all FUNCs? Mostly theoretical question, though.
because it would probably allowed all static functions, (ftrace data has only static functions that are traceable) and who knows what a can of worms we'd open here ;-)
Fair enough.
Quoted 2 lines
jirka
Re: [PATCH 2/2] btf_encoder: Fix function generation
Em Fri, Nov 13, 2020 at 01:43:47PM -0800, Andrii Nakryiko escreveu:
Quoted 18 lines · Jiri Olsa
Quoted 16 lines · Andrii Nakryiko
Quoted 14 lines · Jiri Olsa
Quoted 13 lines
Current conditions for picking up function records break BTF data on some gcc versions.
Some function records can appear with no arguments but with declaration tag set, so moving the '
fn->declaration' in front of other checks.Then checking if argument names are present and finally checking ftrace filter if it's present. If ftrace filter is not available, using the external tag to filter out non external functions.
Signed-off-by Jiri Olsa <[email protected]>
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
I tested locally, all seems to work fine. Left few suggestions below, but those could be done in follow ups (or argued to not be done).
Acked-by Andrii Nakryiko <[email protected]>BTW, for some stats.
BEFORE allowing static funcs:
Nowhere in the last patchkit comments is some explanation for the inclusion of static functions :-\ After the first patch in the last series I get:
$ llvm-objcopy --remove-section=.BTF vmlinux $ readelf -SW vmlinux | grep BTF $ pahole -J vmlinux $ bpftool btf dump file ./vmlinux | grep 'FUNC '| cut -d\' -f2 | sort > before.bpftool $ cp vmlinux vmlinux.before.all $ wc -l before.bpftool 28829 before.bpftool $ $ llvm-objcopy --remove-section=.BTF vmlinux $ readelf -SW vmlinux | grep BTF $ bpftool btf dump file ./vmlinux | grep 'FUNC '| cut -d\' -f2 | sort > after-1st-patch.bpftool Error: failed to load BTF from ./vmlinux: No such file or directory $ pahole -J vmlinux $ $ bpftool btf dump file ./vmlinux | grep 'FUNC '| cut -d\' -f2 | sort > after-1st-patch.bpftool $ wc -l after-1st-patch.bpftool 41030 after-1st-patch.bpftool $ diff -u before.bpftool after-1st-patch.bpftool | less $ diff -u before.bpftool after-1st-patch.bpftool | less $
BTF: I built this kernel without CONFIG_DEBUG_INFO_BTF=y, so that I could use: llvm-objcopy --remove-section=.BTF
That matches what you see here, i.e.
BEFORE:
Quoted 1 line
Quoted 1 line
Quoted 1 line
Func: 308304 bytes (25692 types)
AFTER:
Quoted 1 line
Quoted 1 line
Quoted 1 line
Func: 536664 bytes (44722 types)
The number of functions, that is. I'm scratching my head to figure out why:
https://lore.kernel.org/bpf/[email protected]/
[PATCH 1/2] btf_encoder: Generate also .init functions
Causes this, can you guys please explain it so that we have this in the change log?
As the diff shows that the increase in the number of functions is due to static functions being added:
[acme@five pahole]$ diff -u before.bpftool after-1st-patch.bpftool | tail +__zswap_pool_release
+zswap_writeback_entry
+zswap_zpool_param_set
+zs_zpool_create
+zs_zpool_destroy
+zs_zpool_free
+zs_zpool_malloc
+zs_zpool_map
+zs_zpool_total_size
+zs_zpool_unmap
[acme@five pahole]$
static void zs_zpool_unmap(void *pool, unsigned long handle)
{
zs_unmap_object(pool, handle);
}Reading below Jiri says:
Quoted 9 lines
Quoted 9 lines
Quoted 8 lines
btf_encoder.c int cu__encode_btf(struct cu *cu, int verbose, bool force, · line 629
Quoted 4 lines
continue; } else {btf_encoder.c +1
-1-if (fn->declaration || !fn->external) +if (!fn->external)Hm.. why didn't you drop this fallback? For non-vmlinux, do you think it's a problem to generate all FUNCs? Mostly theoretical question, though.
Quoted 6 lines
Quoted 4 lines
because it would probably allowed all static functions, (ftrace data has only static functions that are traceable) and who knows what a can of worms we'd open here ;-)
Fair enough.
But I'm getting these results (the addition of static variables) after applying:
btf_encoder: Generate also .init functions
:-\
- Arnaldo
Quoted 131 lines
Quoted 54 lines
Quoted 39 lines
.BTF ELF section
=======================================Data size: 4101624 Header size: 24 Types size: 2472836 Strings size: 1628764BTF types
=======================================
Total 2472836 bytes (83310 types)Struct: 920436 bytes (10305 types) FuncProto: 638668 bytes (18869 types) Func: 308304 bytes (25692 types) Enum: 184308 bytes (2293 types) Ptr: 173484 bytes (14457 types) Array: 89064 bytes (3711 types) Union: 81552 bytes (1961 types) Const: 34368 bytes (2864 types) Typedef: 32124 bytes (2677 types) Var: 4688 bytes (293 types) Datasec: 3528 bytes (1 types) Fwd: 1656 bytes (138 types) Volatile: 360 bytes (30 types) Int: 272 bytes (17 types) Restrict: 24 bytes (2 types)AFTER allowing static funcs:
.BTF ELF section
=======================================Data size: 4930558 Header size: 24 Types size: 2914016 Strings size: 2016518BTF types
=======================================
Total 2914016 bytes (108282 types)Quoted 14 lines
Struct: 920436 bytes (10305 types) FuncProto: 851528 bytes (24814 types) Func: 536664 bytes (44722 types) Enum: 184308 bytes (2293 types) Ptr: 173484 bytes (14457 types) Array: 89064 bytes (3711 types) Union: 81552 bytes (1961 types) Const: 34368 bytes (2864 types) Typedef: 32124 bytes (2677 types) Var: 4688 bytes (293 types) Datasec: 3528 bytes (1 types) Fwd: 1656 bytes (138 types) Volatile: 360 bytes (30 types) Int: 256 bytes (16 types)Quoted 1 line
nice, is this tool somewhere in the tree?
Nope, this is from my personal BTF inspection tool, which I never got to open-sourcing, too low on the priority list...
Quoted 59 lines
Quoted 42 lines
So 25692 vs 44722 functions, but the increase in
func_protois smaller due to dedup. Good chunk is strings data for all those function and parameter names.Quoted 35 lines
btf_encoder.c | 24 ++++++++++-------------- 1 file changed, 10 insertions(+), 14 deletions(-)btf_encoder.c +9
-13@@ -612,25 +612,21 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,const char *name;/* -* The functions_cnt != 0 means we parsed all necessary -* kernel symbols and we are using ftrace location filter -* for functions. If it's not available keep the current -* dwarf declaration check. +* Skip functions that: +* - are marked as declarations +* - do not have full argument names +* - are not in ftrace list (if it's available) +* - are not external (in case ftrace filter is not available) */ +if (fn->declaration) +continue ; +if (!has_arg_names (cu, &fn->proto)) +continue ;if (functions_cnt) { -/* -* We check following conditions: -* - argument names are defined -* - there's symbol and address defined for the function -* - function address belongs to ftrace locations -* - function is generated only once -*/ -if (!has_arg_names (cu, &fn->proto)) -continue ;Quoted 8 lines
btf_encoder.c int cu__encode_btf(struct cu *cu, int verbose, bool force, · line 626
Quoted 1 line
btf_encoder.c +0
-0if (!should_generate_function (btfe,function__name (fn, cu)))Seeing Arnaldo's confusion, I remember initially I was similarly confused. I think this
p->generated= true should be moved out ofshould_generate_function()and done here explicitly. Let's turnshould_generate_function()intofind_allowed_function()or something, to encapsulate bsearch. Checking !p ||p->generatedcould be done here explicitly.ok, that should be more obvious, I'll send new version
Quoted 6 lines
Quoted 5 lines
btf_encoder.c +1
-1continue ; }else { -if (fn->declaration || !fn->external) +if (!fn->external)Quoted 8 lines
Quoted 3 lines
Hm.. why didn't you drop this fallback? For non-vmlinux, do you think it's a problem to generate all FUNCs? Mostly theoretical question, though.
because it would probably allowed all static functions, (ftrace data has only static functions that are traceable) and who knows what a can of worms we'd open here ;-)
Fair enough.
Quoted 2 lines
jirka
Signature · 1 line
- Arnaldo
Re: [PATCH 2/2] btf_encoder: Fix function generation
Quoted 33 lines · Arnaldo Carvalho de Melo
Em Fri, Nov 13, 2020 at 01:43:47PM -0800, Andrii Nakryiko escreveu:
Quoted 18 lines · Jiri Olsa
Quoted 16 lines · Andrii Nakryiko
Quoted 14 lines · Jiri Olsa
Quoted 13 lines
Current conditions for picking up function records break BTF data on some gcc versions.
Some function records can appear with no arguments but with declaration tag set, so moving the '
fn->declaration' in front of other checks.Then checking if argument names are present and finally checking ftrace filter if it's present. If ftrace filter is not available, using the external tag to filter out non external functions.
Signed-off-by Jiri Olsa <[email protected]>Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
I tested locally, all seems to work fine. Left few suggestions below, but those could be done in follow ups (or argued to not be done).
Acked-by Andrii Nakryiko <[email protected]>BTW, for some stats.
BEFORE allowing static funcs:
Nowhere in the last patchkit comments is some explanation for the inclusion of static functions :-\ After the first patch in the last series I get:
Quoted 7 lines
$ llvm-objcopy --remove-section=.BTF vmlinux $ readelf -SW vmlinux | grep BTF $ pahole -J vmlinux $ bpftool btf dump file ./vmlinux | grep 'FUNC '| cut -d\' -f2 | sort > before.bpftool $ cp vmlinux vmlinux.before.all $ wc -l before.bpftool 28829 before.bpftool
I think you see the original number of functions, because without the 'not merged' kernel patch, that added the special init section, pahole will fail to detect vmlinux and fall back to checking dwarf declarations
there's a verbose message for the fall back, but it is not displayed at the moment ;-) with the fix below you should see it:
$ LLVM_OBJCOPY=objcopy ./pahole -V -J vmlinux >out $ cat out | grep 'vmlinux not detected' vmlinux not detected, falling back to dwarf data
I'll check on the verbose setup and send full patch, I did not expect it would not get printed, sry
so the new numebr ~41k functions is together static functions and init functions
jirka
---
btf_encoder.c +2 -1
@@ -584,6 +584,8 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,struct tag *pos;int err =0 ; + btf_elf__verbose = verbose; +if (btfe &&strcmp (btfe->filename, cu->filename)) { err =btf_encoder__encode ();if (err)
@@ -623,7 +625,6 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force, } } - btf_elf__verbose = verbose; btf_elf__force = force; type_id_off =btf__get_nr_types (btfe->btf);
lib/bpf +1 -1
@@ -1 +1 @@ -Subproject commit ff797cc905d9c5fe9acab92d2da127342b20f80f +Subproject commit ff797cc905d9c5fe9acab92d2da127342b20f80f-dirty
Re: [PATCH 2/2] btf_encoder: Fix function generation
Em Mon, Nov 16, 2020 at 07:21:45PM +0100, Jiri Olsa escreveu:
Quoted 42 lines · Arnaldo Carvalho de Melo
Quoted 33 lines
Em Fri, Nov 13, 2020 at 01:43:47PM -0800, Andrii Nakryiko escreveu:
Quoted 18 lines · Jiri Olsa
Quoted 16 lines · Andrii Nakryiko
Quoted 14 lines · Jiri Olsa
Quoted 13 lines
Current conditions for picking up function records break BTF data on some gcc versions.
Some function records can appear with no arguments but with declaration tag set, so moving the '
fn->declaration' in front of other checks.Then checking if argument names are present and finally checking ftrace filter if it's present. If ftrace filter is not available, using the external tag to filter out non external functions.
Signed-off-by Jiri Olsa <[email protected]>Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
I tested locally, all seems to work fine. Left few suggestions below, but those could be done in follow ups (or argued to not be done).
Acked-by Andrii Nakryiko <[email protected]>BTW, for some stats.
BEFORE allowing static funcs:
Nowhere in the last patchkit comments is some explanation for the inclusion of static functions :-\ After the first patch in the last series I get:
Quoted 7 lines
$ llvm-objcopy --remove-section=.BTF vmlinux $ readelf -SW vmlinux | grep BTF $ pahole -J vmlinux $ bpftool btf dump file ./vmlinux | grep 'FUNC '| cut -d\' -f2 | sort > before.bpftool $ cp vmlinux vmlinux.before.all $ wc -l before.bpftool 28829 before.bpftool
Quoted 4 lines
I think you see the original number of functions, because without the 'not merged' kernel patch, that added the special init section, pahole will fail to detect vmlinux and fall back to checking dwarf declarations
Indeed, I moved the verbose/force setting to the beggining of the encoder and:
------------
Found 352 per-CPU variables!
vmlinux not detected, falling back to dwarf data
File vmlinux:
search cu '/home/acme/git/linux/arch/x86/kernel/head_64.S' for percpu global variables.
-----------------
Now I have to read that code to figure out what that "vmlinux not detected, falling back to dwarf data" message means, as vmlinux is where DWARF data is, so what is that isn't being "detected", /me checks...
- Arnaldo
Quoted 45 lines
there's a verbose message for the fall back, but it is not displayed at the moment ;-) with the fix below you should see it:
$ LLVM_OBJCOPY=objcopy ./pahole -V -J vmlinux >out $ cat out | grep 'vmlinux not detected' vmlinux not detected, falling back to dwarf dataI'll check on the verbose setup and send full patch, I did not expect it would not get printed, sry
so the new numebr ~41k functions is together static functions and init functions
jirka
---
btf_encoder.c +2
-1@@ -584,6 +584,8 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force,struct tag *pos;int err =0 ; + btf_elf__verbose = verbose; +if (btfe &&strcmp (btfe->filename, cu->filename)) { err =btf_encoder__encode ();if (err)@@ -623,7 +625,6 @@ int cu__encode_btf(struct cu *cu, int verbose, bool force, } } - btf_elf__verbose = verbose; btf_elf__force = force; type_id_off =btf__get_nr_types (btfe->btf);lib/bpf +1
-1@@ -1 +1 @@ -Subproject commit ff797cc905d9c5fe9acab92d2da127342b20f80f +Subproject commit ff797cc905d9c5fe9acab92d2da127342b20f80f-dirty
Signature · 1 line
- Arnaldo
Re: [PATCH 2/2] btf_encoder: Fix function generation
Em Mon, Nov 16, 2020 at 04:15:44PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 58 lines
Em Mon, Nov 16, 2020 at 07:21:45PM +0100, Jiri Olsa escreveu:
Quoted 42 lines · Arnaldo Carvalho de Melo
Quoted 33 lines
Em Fri, Nov 13, 2020 at 01:43:47PM -0800, Andrii Nakryiko escreveu:
Quoted 18 lines · Jiri Olsa
Quoted 16 lines · Andrii Nakryiko
Quoted 14 lines · Jiri Olsa
Quoted 13 lines
Current conditions for picking up function records break BTF data on some gcc versions.
Some function records can appear with no arguments but with declaration tag set, so moving the '
fn->declaration' in front of other checks.Then checking if argument names are present and finally checking ftrace filter if it's present. If ftrace filter is not available, using the external tag to filter out non external functions.
Signed-off-by Jiri Olsa <[email protected]>Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
I tested locally, all seems to work fine. Left few suggestions below, but those could be done in follow ups (or argued to not be done).
Acked-by Andrii Nakryiko <[email protected]>BTW, for some stats.
BEFORE allowing static funcs:
Nowhere in the last patchkit comments is some explanation for the inclusion of static functions :-\ After the first patch in the last series I get:
Quoted 7 lines
$ llvm-objcopy --remove-section=.BTF vmlinux $ readelf -SW vmlinux | grep BTF $ pahole -J vmlinux $ bpftool btf dump file ./vmlinux | grep 'FUNC '| cut -d\' -f2 | sort > before.bpftool $ cp vmlinux vmlinux.before.all $ wc -l before.bpftool 28829 before.bpftoolQuoted 4 lines
I think you see the original number of functions, because without the 'not merged' kernel patch, that added the special init section, pahole will fail to detect vmlinux and fall back to checking dwarf declarations
Indeed, I moved the verbose/force setting to the beggining of the encoder and:
------------
Found 352 per-CPU variables!
vmlinux not detected, falling back to dwarf data
File vmlinux:
search cu '/home/acme/git/linux/arch/x86/kernel/head_64.S' for percpu global variables.
-----------------
Quoted 3 lines
Now I have to read that code to figure out what that "vmlinux not detected, falling back to dwarf data" message means, as vmlinux is where DWARF data is, so what is that isn't being "detected", /me checks...
So with some debugging I see, the message is just confusing:
"vmlinux not detected, falling back to dwarf data (functions_cnt=53238, has_all_symbols(&fl)=0"
It finds the ELF symtab, finds the percpu variables there, tons of functions, matching the number after this approach of marking BPF init functions was dropped its just that vague "has_all_symbols()" routine that fails to find all the symbols it needs in the vmlinux file.
- Arnaldo
Re: [PATCH 2/2] btf_encoder: Fix function generation
Quoted 59 lines · Arnaldo Carvalho de Melo
Em Mon, Nov 16, 2020 at 04:15:44PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 58 lines
Em Mon, Nov 16, 2020 at 07:21:45PM +0100, Jiri Olsa escreveu:
Quoted 42 lines · Arnaldo Carvalho de Melo
Quoted 33 lines
Em Fri, Nov 13, 2020 at 01:43:47PM -0800, Andrii Nakryiko escreveu:
Quoted 18 lines · Jiri Olsa
Quoted 16 lines · Andrii Nakryiko
Quoted 14 lines · Jiri Olsa
Quoted 13 lines
Current conditions for picking up function records break BTF data on some gcc versions.
Some function records can appear with no arguments but with declaration tag set, so moving the '
fn->declaration' in front of other checks.Then checking if argument names are present and finally checking ftrace filter if it's present. If ftrace filter is not available, using the external tag to filter out non external functions.
Signed-off-by Jiri Olsa <[email protected]>Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
I tested locally, all seems to work fine. Left few suggestions below, but those could be done in follow ups (or argued to not be done).
Acked-by Andrii Nakryiko <[email protected]>BTW, for some stats.
BEFORE allowing static funcs:
Nowhere in the last patchkit comments is some explanation for the inclusion of static functions :-\ After the first patch in the last series I get:
Quoted 7 lines
$ llvm-objcopy --remove-section=.BTF vmlinux $ readelf -SW vmlinux | grep BTF $ pahole -J vmlinux $ bpftool btf dump file ./vmlinux | grep 'FUNC '| cut -d\' -f2 | sort > before.bpftool $ cp vmlinux vmlinux.before.all $ wc -l before.bpftool 28829 before.bpftoolQuoted 4 lines
I think you see the original number of functions, because without the 'not merged' kernel patch, that added the special init section, pahole will fail to detect vmlinux and fall back to checking dwarf declarations
Indeed, I moved the verbose/force setting to the beggining of the encoder and:
------------
Found 352 per-CPU variables!
vmlinux not detected, falling back to dwarf data
File vmlinux:
search cu '/home/acme/git/linux/arch/x86/kernel/head_64.S' for percpu global variables.
-----------------
Quoted 7 lines
Quoted 3 lines
Now I have to read that code to figure out what that "vmlinux not detected, falling back to dwarf data" message means, as vmlinux is where DWARF data is, so what is that isn't being "detected", /me checks...
So with some debugging I see, the message is just confusing:
"vmlinux not detected, falling back to dwarf data (functions_cnt=53238, has_all_symbols(&fl)=0"
how about:
"ftrace data not detected, falling back to dwarf data"
Quoted 5 lines
It finds the ELF symtab, finds the percpu variables there, tons of functions, matching the number after this approach of marking BPF init functions was dropped its just that vague "
has_all_symbols()" routine that fails to find all the symbols it needs in the vmlinux file.
we collect functions and other symbols in one loop over the symtab, so thats why we have all those collected and still can decide to fall back
before we needed also the init section symbols, now with this patch we need to know only mcount section begin/end
jirka
Quoted 3 lines
- Arnaldo
Re: [PATCH 2/2] btf_encoder: Fix function generation
Em Mon, Nov 16, 2020 at 08:40:08PM +0100, Jiri Olsa escreveu:
Quoted 64 lines · Arnaldo Carvalho de Melo
Quoted 63 lines
Em Mon, Nov 16, 2020 at 04:15:44PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 61 lines
Em Mon, Nov 16, 2020 at 07:21:45PM +0100, Jiri Olsa escreveu:
Quoted 42 lines · Arnaldo Carvalho de Melo
Quoted 33 lines
Em Fri, Nov 13, 2020 at 01:43:47PM -0800, Andrii Nakryiko escreveu:
Quoted 18 lines · Jiri Olsa
Quoted 16 lines · Andrii Nakryiko
Quoted 14 lines · Jiri Olsa
Quoted 13 lines
Current conditions for picking up function records break BTF data on some gcc versions.
Some function records can appear with no arguments but with declaration tag set, so moving the '
fn->declaration' in front of other checks.Then checking if argument names are present and finally checking ftrace filter if it's present. If ftrace filter is not available, using the external tag to filter out non external functions.
Signed-off-by Jiri Olsa <[email protected]>Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
I tested locally, all seems to work fine. Left few suggestions below, but those could be done in follow ups (or argued to not be done).
Acked-by Andrii Nakryiko <[email protected]>BTW, for some stats.
BEFORE allowing static funcs:
Nowhere in the last patchkit comments is some explanation for the inclusion of static functions :-\ After the first patch in the last series I get:
Quoted 7 lines
$ llvm-objcopy --remove-section=.BTF vmlinux $ readelf -SW vmlinux | grep BTF $ pahole -J vmlinux $ bpftool btf dump file ./vmlinux | grep 'FUNC '| cut -d\' -f2 | sort > before.bpftool $ cp vmlinux vmlinux.before.all $ wc -l before.bpftool 28829 before.bpftoolQuoted 4 lines
I think you see the original number of functions, because without the 'not merged' kernel patch, that added the special init section, pahole will fail to detect vmlinux and fall back to checking dwarf declarations
Indeed, I moved the verbose/force setting to the beggining of the encoder and:
------------
Found 352 per-CPU variables!
vmlinux not detected, falling back to dwarf data
File vmlinux:
search cu '/home/acme/git/linux/arch/x86/kernel/head_64.S' for percpu global variables.
-----------------Now I have to read that code to figure out what that "vmlinux not detected, falling back to dwarf data" message means, as vmlinux is where DWARF data is, so what is that isn't being "detected", /me checks...
Quoted 7 lines
Quoted 3 lines
So with some debugging I see, the message is just confusing:
"vmlinux not detected, falling back to dwarf data (functions_cnt=53238, has_all_symbols(&fl)=0"
how about:
"ftrace data not detected, falling back to dwarf data"
Much better!
Quoted 11 lines
Quoted 5 lines
It finds the ELF symtab, finds the percpu variables there, tons of functions, matching the number after this approach of marking BPF init functions was dropped its just that vague "
has_all_symbols()" routine that fails to find all the symbols it needs in the vmlinux file.we collect functions and other symbols in one loop over the symtab, so thats why we have all those collected and still can decide to fall back
before we needed also the init section symbols, now with this patch we need to know only mcount section begin/end
Thanks for the explanations, match my observations, its just that the functions could have some more descriptive names :)
Please send the patch with the s/vmlinux/ftrace symbols/g and please also s/dwarf/DWARF/g as its an acronym.
- Arnaldo