Skip to contentSkip to conversation

bpf: Move iterator functions into special init section

PATCH In review Jiri Olsa 5 participants 6 messages Nov 10, 2020 – Nov 11, 2020

Subsystems (5)
lore
Share & tools
lore

Jiri Olsa

author reviewer
#0 lore raw

PATCH v6 1/1 bpf: Move iterator functions into special init section

With upcoming changes to pahole, that change the way how and which kernel functions are stored in BTF data, we need a way to recognize iterator functions.

Iterator functions need to be in BTF data, but have no real body and are currently placed in .init.text section, so they are freed after kernel init and are filtered out of BTF data because of that.

The solution is to place these functions under new section:
.init.bpf.preserve_type

And add 2 new symbols to mark that area:
__init_bpf_preserve_type_begin
__init_bpf_preserve_type_end

The code in pahole responsible for picking up the functions will be able to recognize functions from this section and add them to the BTF data and filter out all other .init.text functions.

Cc Arnaldo Carvalho de Melo <[email protected]>

---
v6 changes:
- simplify the change and add comment to bpf.h header based on David's suggestion and remove the init.h change
- removing acks, because of the new changes

 include/asm-generic/vmlinux.lds.h | 16 +++++++++++++++-
 include/linux/bpf.h               | 12 +++++++++++-
 2 files changed, 26 insertions(+), 2 deletions(-)
include/asm-generic/vmlinux.lds.h +15 -1
index b2b3d81b1535..f91029b3443b 100644
@@ -685,8 +685,21 @@
 	.BTF_ids : AT(ADDR(.BTF_ids) - LOAD_OFFSET) {			\
 		*(.BTF_ids)						\
 	}
+
+/*
+ * .init.bpf.preserve_type
+ *
+ * This section store special BPF function and marks them
+ * with begin/end symbols pair for the sake of pahole tool.
+ */
+#define INIT_BPF_PRESERVE_TYPE						\
+	__init_bpf_preserve_type_begin = .;                             \
+	*(.init.bpf.preserve_type)                                      \
+	__init_bpf_preserve_type_end = .;				\
+	MEM_DISCARD(init.bpf.preserve_type)
 #else
 #define BTF
+#define INIT_BPF_PRESERVE_TYPE
 #endif
 
 /*
@@ -741,7 +754,8 @@
 #define INIT_TEXT							\
 	*(.init.text .init.text.*)					\
 	*(.text.startup)						\
-	MEM_DISCARD(init.text*)
+	MEM_DISCARD(init.text*)						\
+	INIT_BPF_PRESERVE_TYPE
 
 #define EXIT_DATA							\
 	*(.exit.data .exit.data.*)					\
include/linux/bpf.h +11 -1
index 2b16bf48aab6..1739a92516ed 100644
@@ -1276,10 +1276,20 @@ struct bpf_link *bpf_link_get_from_fd(u32 ufd);
 int bpf_obj_pin_user(u32 ufd, const char __user *pathname);
 int bpf_obj_get_user(const char __user *pathname, int flags);
 
+/* In case we generate BTF data, we need to group all iterator
+ * functions into special init section, so pahole can track them.
+ * Otherwise pure __init section is enough.
+ */
+#ifdef CONFIG_DEBUG_INFO_BTF
+#define __init_bpf_preserve_type __section(".init.bpf.preserve_type")
+#else
+#define __init_bpf_preserve_type __init
+#endif
+
 #define BPF_ITER_FUNC_PREFIX "bpf_iter_"
 #define DEFINE_BPF_ITER_FUNC(target, args...)			\
 	extern int bpf_iter_ ## target(args);			\
-	int __init bpf_iter_ ## target(args) { return 0; }
+	int __init_bpf_preserve_type bpf_iter_ ## target(args) { return 0; }
 
 struct bpf_iter_aux_info {
 	struct bpf_map *map;
Signature · 1 line
2.26.2
maintainer
#1 lore raw
Quoted 16 lines · Jiri Olsa

With upcoming changes to pahole, that change the way how and which kernel functions are stored in BTF data, we need a way to recognize iterator functions.

Iterator functions need to be in BTF data, but have no real body and are currently placed in .init.text section, so they are freed after kernel init and are filtered out of BTF data because of that.

The solution is to place these functions under new section:
.init.bpf.preserve_type

And add 2 new symbols to mark that area:
__init_bpf_preserve_type_begin
__init_bpf_preserve_type_end

Quoted 7 lines

The code in pahole responsible for picking up the functions will be able to recognize functions from this section and add them to the BTF data and filter out all other .init.text functions.

Cc Arnaldo Carvalho de Melo <[email protected]>

LGTM, applied, thanks! Also added a reference to the pahole commit to the commit log so that this info doesn't get lost in the void plus carried over prior Acks given nothing changed logically in the patch.

P.s.: I've been wondering whether we also need to align the begin/end symbols via ALIGN_FUNCTION() in case ld might realign to a different boundary on later passes but this seems neither the case for .init.text right now, likely since it doesn't matter for kallsyms data in our particular case.

#2 lore raw
Quoted 30 lines · Daniel Borkmann
Quoted 16 lines · Jiri Olsa

With upcoming changes to pahole, that change the way how and which kernel functions are stored in BTF data, we need a way to recognize iterator functions.

Iterator functions need to be in BTF data, but have no real body and are currently placed in .init.text section, so they are freed after kernel init and are filtered out of BTF data because of that.

The solution is to place these functions under new section:
.init.bpf.preserve_type

And add 2 new symbols to mark that area:
__init_bpf_preserve_type_begin
__init_bpf_preserve_type_end

Quoted 7 lines

The code in pahole responsible for picking up the functions will be able to recognize functions from this section and add them to the BTF data and filter out all other .init.text functions.

Cc Arnaldo Carvalho de Melo <[email protected]>

LGTM, applied, thanks! Also added a reference to the pahole commit to the commit log so that this info doesn't get lost in the void plus carried over prior Acks given nothing changed logically in the patch.

Quoted 6 lines

P.s.: I've been wondering whether we also need to align the begin/end symbols via ALIGN_FUNCTION() in case ld might realign to a different boundary on later passes but this seems neither the case for .init.text right now, likely since it doesn't matter for kallsyms data in our particular case.

I'll check but I think it's not a problem as you said

thanks,
jirka

#3 lore raw
In reply to Daniel Borkmann

Em Wed, Nov 11, 2020 at 12:26:29PM +0100, Daniel Borkmann escreveu:

Quoted 21 lines · Jiri Olsa
Quoted 20 lines

With upcoming changes to pahole, that change the way how and which kernel functions are stored in BTF data, we need a way to recognize iterator functions.

Iterator functions need to be in BTF data, but have no real body and are currently placed in .init.text section, so they are freed after kernel init and are filtered out of BTF data because of that.

The solution is to place these functions under new section:
.init.bpf.preserve_type

And add 2 new symbols to mark that area:
__init_bpf_preserve_type_begin
__init_bpf_preserve_type_end

The code in pahole responsible for picking up the functions will be able to recognize functions from this section and add them to the BTF data and filter out all other .init.text functions.

Quoted 5 lines
Quoted 3 lines
Cc Arnaldo Carvalho de Melo <[email protected]>

LGTM, applied, thanks! Also added a reference to the pahole commit

Applied to what branch? I'm trying to test it now :-)

- Arnaldo

Quoted 9 lines

to the commit log so that this info doesn't get lost in the void plus carried over prior Acks given nothing changed logically in the patch.

P.s.: I've been wondering whether we also need to align the begin/end symbols via ALIGN_FUNCTION() in case ld might realign to a different boundary on later passes but this seems neither the case for .init.text right now, likely since it doesn't matter for kallsyms data in our particular case.

Signature · 1 line
- Arnaldo
#4 lore raw

Em Wed, Nov 11, 2020 at 09:37:38AM -0300, Arnaldo Carvalho de Melo escreveu:

Quoted 22 lines

Em Wed, Nov 11, 2020 at 12:26:29PM +0100, Daniel Borkmann escreveu:

Quoted 21 lines · Jiri Olsa
Quoted 20 lines

With upcoming changes to pahole, that change the way how and which kernel functions are stored in BTF data, we need a way to recognize iterator functions.

Iterator functions need to be in BTF data, but have no real body and are currently placed in .init.text section, so they are freed after kernel init and are filtered out of BTF data because of that.

The solution is to place these functions under new section:
.init.bpf.preserve_type

And add 2 new symbols to mark that area:
__init_bpf_preserve_type_begin
__init_bpf_preserve_type_end

The code in pahole responsible for picking up the functions will be able to recognize functions from this section and add them to the BTF data and filter out all other .init.text functions.

Quoted 7 lines
Quoted 5 lines
Quoted 3 lines
Cc Arnaldo Carvalho de Melo <[email protected]>

LGTM, applied, thanks! Also added a reference to the pahole commit

Applied to what branch? I'm trying to test it now :-)

Nevermind, bpf/master, I was looking at bpf-next/master.

- Arnaldo

#5 lore raw
Quoted 28 lines · Arnaldo Carvalho de Melo

Em Wed, Nov 11, 2020 at 09:37:38AM -0300, Arnaldo Carvalho de Melo escreveu:

Quoted 26 lines

Em Wed, Nov 11, 2020 at 12:26:29PM +0100, Daniel Borkmann escreveu:

Quoted 25 lines · Jiri Olsa
Quoted 23 lines

With upcoming changes to pahole, that change the way how and which kernel functions are stored in BTF data, we need a way to recognize iterator functions.

Iterator functions need to be in BTF data, but have no real body and are currently placed in .init.text section, so they are freed after kernel init and are filtered out of BTF data because of that.

The solution is to place these functions under new section:
.init.bpf.preserve_type

And add 2 new symbols to mark that area:
__init_bpf_preserve_type_begin
__init_bpf_preserve_type_end

The code in pahole responsible for picking up the functions will be able to recognize functions from this section and add them to the BTF data and filter out all other .init.text functions.

Cc Arnaldo Carvalho de Melo <[email protected]>
Quoted 5 lines
Quoted 3 lines
Quoted 1 line

LGTM, applied, thanks! Also added a reference to the pahole commit

Applied to what branch? I'm trying to test it now :-)

Nevermind, bpf/master, I was looking at bpf-next/master.

I've dropped this patch from bpf tree.
I think we need to agree on the whole approach first.
This filtering based on section name with special handling in pahole doesn't feel like solid long term direction.
I think we have to brainstorm more on it.
I'm not saying we will not go back to a special section approach. This revert is only buying us time to discuss what's the right path here. Mainly I reverted to unbreak bpf tree CI which currently fails due to two tests in test_progs
failing with the latest pahole and this patch.