Integrate kernel module BTF support
Subsystems (8)
Andrii Nakryiko
maintainerv4 cover Integrate kernel module BTF support
This patch set adds BTF generation for kernel modules using a compact split BTF approach. Respective patches have all the details.
Kernel module BTFs rely on pahole's split BTF support, which is added in [0] and will be available starting from v1.19. Support for it is detected automatically during kernel build time.
This patch set implements in-kernel support for split BTF loading and validation. It also extends GET_OBJ_INFO API for BTFs to return BTF's module name and a flag whether BTF itself is in-kernel or user-provided. vmlinux BTF is also exposed to user-space through the same BTF object iteration APIs.
Follow up patch set will utilize the fact that vmlinux and module BTFs now have associated ID to provide ability to attach BPF fentry/fexit/etc programs to functions defined in kernel modules.
bpftool is also extended to show module/vmlinux BTF's name.
[0] https://patchwork.kernel.org/project/netdevbpf/list/?series=378699&state=*
v3->v4:
- copy_to_user() on ENOSPC in btf_get_info_by_fd() (Martin);v2->v3:
- get rid of unnecessary gotos (Song);v2->v1:
- drop WARNs, add fewer pr_warn()'s instead (Greg);
- properly initialize sysfs binary attribute structure (Greg);
- add __maybe_unused to any_section_objs, used conditionally by module BTF;rfc->v1:
- CONFIG_DEBUG_INFO_BTF_MODULES is derived automatically (Alexei);
- vmlinux BTF now has explicit "vmlinux" name (Alexei);
- added sysfs ABI documentation for /sys/kernel/btf/<module> (Greg).
Andrii Nakryiko (5):
bpf: add in-kernel split BTF support
bpf: assign ID to vmlinux BTF and return extra info for BTF in
GET_OBJ_INFO
kbuild: build kernel module BTFs if BTF is enabled and pahole supports
it
bpf: load and verify kernel module BTFs
tools/bpftool: add support for in-kernel and named BTF in btf show
Documentation/ABI/testing/sysfs-kernel-btf | 8 + include/linux/bpf.h | 2 + include/linux/module.h | 4 + include/uapi/linux/bpf.h | 3 + kernel/bpf/btf.c | 406 ++++++++++++++++++--- kernel/bpf/sysfs_btf.c | 2 +- kernel/module.c | 32 ++ lib/Kconfig.debug | 9 + scripts/Makefile.modfinal | 20 +- tools/bpf/bpftool/btf.c | 28 +- tools/include/uapi/linux/bpf.h | 3 + 11 files changed, 459 insertions(+), 58 deletions(-)
Signature · 1 line
2.24.1
Andrii Nakryiko
maintainerPATCH v4 1/5 bpf: add in-kernel split BTF support
Adjust in-kernel BTF implementation to support a split BTF mode of operation. Changes are mostly mirroring libbpf split BTF changes, with the exception of start_id being 0 for in-kernel implementation due to simpler read-only mode.
Otherwise, for split BTF logic, most of the logic of jumping to base BTF, where necessary, is encapsulated in few helper functions. Type numbering and string offset in a split BTF are logically continuing where base BTF ends, so most of the high-level logic is kept without changes.
Type verification and size resolution is only doing an added resolution of new split BTF types and relies on already cached size and type resolution results in the base BTF.
--- kernel/bpf/btf.c | 171 +++++++++++++++++++++++++++++++++-------------- 1 file changed, 119 insertions(+), 52 deletions(-)
kernel/bpf/btf.c +119 -52
@@ -203,12 +203,17 @@ struct btf {const char *strings;void *nohdr_data;struct btf_header hdr; -u32 nr_types; +u32 nr_types;/* includes VOID for base BTF */ u32 types_size;u32 data_size;refcount_t refcnt;u32 id;struct rcu_head rcu; + +/* split BTF support */ +struct btf *base_btf; +u32 start_id;/* first type ID in this BTF (0 for base BTF) */ +u32 start_str_off;/* first string offset (0 for base BTF) */ };enum verifier_phase {
@@ -449,14 +454,27 @@ static bool btf_type_is_datasec(const struct btf_type *t)return BTF_INFO_KIND (t->info) ==BTF_KIND_DATASEC ; } +static u32 btf_nr_types_total (const struct btf *btf) +{ +u32 total =0 ; + +while (btf) { + total += btf->nr_types; + btf = btf->base_btf; + } + +return total; +} +s32 btf_find_by_name_kind (const struct btf *btf,const char *name,u8 kind) {const struct btf_type *t;const char *tname; -u32 i; +u32 i, total; -for (i =1 ; i <= btf->nr_types; i++) { - t = btf->types[i]; + total =btf_nr_types_total (btf); +for (i =1 ; i < total; i++) { + t =btf_type_by_id (btf, i);if (BTF_INFO_KIND (t->info) != kind)continue ;
@@ -599,8 +617,14 @@ static const struct btf_kind_operations *btf_type_ops(const struct btf_type *t)static bool btf_name_offset_valid (const struct btf *btf,u32 offset) { -return BTF_STR_OFFSET_VALID (offset) && - offset < btf->hdr.str_len; +if (!BTF_STR_OFFSET_VALID (offset)) +return false; + +while (offset < btf->start_str_off) + btf = btf->base_btf; + + offset -= btf->start_str_off; +return offset < btf->hdr.str_len; }static bool __btf_name_char_ok (char c,bool first,bool dot_ok)
@@ -614,10 +638,22 @@ static bool __btf_name_char_ok(char c, bool first, bool dot_ok)return true; } +static const char *btf_str_by_offset (const struct btf *btf,u32 offset) +{ +while (offset < btf->start_str_off) + btf = btf->base_btf; + + offset -= btf->start_str_off; +if (offset < btf->hdr.str_len) +return &btf->strings[offset]; + +return NULL ; +} +static bool __btf_name_valid (const struct btf *btf,u32 offset,bool dot_ok) {/* offset must be valid */ -const char *src = &btf->strings[offset]; +const char *src =btf_str_by_offset (btf, offset);const char *src_limit;if (!__btf_name_char_ok (*src, true, dot_ok))
@@ -650,27 +686,28 @@ static bool btf_name_valid_section(const struct btf *btf, u32 offset)static const char *__btf_name_by_offset (const struct btf *btf,u32 offset) { +const char *name; +if (!offset)return "(anon)" ; -else if (offset < btf->hdr.str_len) -return &btf->strings[offset]; -else -return "(invalid-name-offset)" ; + + name =btf_str_by_offset (btf, offset); +return name ?:"(invalid-name-offset)" ; }const char *btf_name_by_offset (const struct btf *btf,u32 offset) { -if (offset < btf->hdr.str_len) -return &btf->strings[offset]; - -return NULL ; +return btf_str_by_offset (btf, offset); }const struct btf_type *btf_type_by_id (const struct btf *btf,u32 type_id) { -if (type_id > btf->nr_types) -return NULL ; +while (type_id < btf->start_id) + btf = btf->base_btf; + type_id -= btf->start_id; +if (type_id >= btf->nr_types) +return NULL ;return btf->types[type_id]; }
@@ -1390,17 +1427,13 @@ static int btf_add_type(struct btf_verifier_env *env, struct btf_type *t) {struct btf *btf = env->btf; -/* < 2 because +1 for btf_void which is always in btf->types[0]. -* btf_void is not accounted in btf->nr_types because btf_void -* does not come from the BTF file. -*/ -if (btf->types_size - btf->nr_types <2 ) { +if (btf->types_size == btf->nr_types) {/* Expand 'types' array */ struct btf_type **new_types;u32 expand_by, new_size; -if (btf->types_size ==BTF_MAX_TYPE ) { +if (btf->start_id + btf->types_size ==BTF_MAX_TYPE ) {btf_verifier_log (env,"Exceeded max num of types" );return -E2BIG ; }
@@ -1414,18 +1447,23 @@ static int btf_add_type(struct btf_verifier_env *env, struct btf_type *t)if (!new_types)return -ENOMEM ; -if (btf->nr_types ==0 ) - new_types[0 ] = &btf_void; -else +if (btf->nr_types ==0 ) { +if (!btf->base_btf) { +/* lazily init VOID type */ + new_types[0 ] = &btf_void; + btf->nr_types++; + } + }else {memcpy (new_types, btf->types, -sizeof (*btf->types) * (btf->nr_types +1 )); +sizeof (*btf->types) * btf->nr_types); + }kvfree (btf->types); btf->types = new_types; btf->types_size = new_size; } - btf->types[++(btf->nr_types)] = t; + btf->types[btf->nr_types++] = t;return 0 ; }
@@ -1498,18 +1536,17 @@ static int env_resolve_init(struct btf_verifier_env *env)u32 *resolved_ids =NULL ;u8 *visit_states =NULL ; -/* +1 for btf_void */ - resolved_sizes =kvcalloc (nr_types +1 ,sizeof (*resolved_sizes), + resolved_sizes =kvcalloc (nr_types,sizeof (*resolved_sizes),GFP_KERNEL | __GFP_NOWARN);if (!resolved_sizes) goto nomem; - resolved_ids =kvcalloc (nr_types +1 ,sizeof (*resolved_ids), + resolved_ids =kvcalloc (nr_types,sizeof (*resolved_ids),GFP_KERNEL | __GFP_NOWARN);if (!resolved_ids) goto nomem; - visit_states =kvcalloc (nr_types +1 ,sizeof (*visit_states), + visit_states =kvcalloc (nr_types,sizeof (*visit_states),GFP_KERNEL | __GFP_NOWARN);if (!visit_states) goto nomem;
@@ -1561,21 +1598,27 @@ static bool env_type_is_resolve_sink(const struct btf_verifier_env *env,static bool env_type_is_resolved (const struct btf_verifier_env *env,u32 type_id) { -return env->visit_states[type_id] ==RESOLVED ; +/* base BTF types should be resolved by now */ +if (type_id < env->btf->start_id) +return true; + +return env->visit_states[type_id - env->btf->start_id] ==RESOLVED ; }static int env_stack_push (struct btf_verifier_env *env,const struct btf_type *t,u32 type_id) { +const struct btf *btf = env->btf;struct resolve_vertex *v;if (env->top_stack ==MAX_RESOLVE_DEPTH )return -E2BIG ; -if (env->visit_states[type_id] !=NOT_VISITED ) +if (type_id < btf->start_id + || env->visit_states[type_id - btf->start_id] !=NOT_VISITED )return -EEXIST ; - env->visit_states[type_id] =VISITED ; + env->visit_states[type_id - btf->start_id] =VISITED ; v = &env->stack[env->top_stack++]; v->t = t;
@@ -1605,6 +1648,7 @@ static void env_stack_pop_resolved(struct btf_verifier_env *env,u32 type_id = env->stack[--(env->top_stack)].type_id;struct btf *btf = env->btf; + type_id -= btf->start_id;/* adjust to local type id */ btf->resolved_sizes[type_id] = resolved_size; btf->resolved_ids[type_id] = resolved_type_id; env->visit_states[type_id] =RESOLVED ;
@@ -1709,14 +1753,30 @@ btf_resolve_size(const struct btf *btf, const struct btf_type *type,return __btf_resolve_size (btf, type, type_size,NULL ,NULL ,NULL ,NULL ); } +static u32 btf_resolved_type_id (const struct btf *btf,u32 type_id) +{ +while (type_id < btf->start_id) + btf = btf->base_btf; + +return btf->resolved_ids[type_id - btf->start_id]; +} +/* The input param "type_id" must point to a needs_resolve type */ static const struct btf_type *btf_type_id_resolve (const struct btf *btf,u32 *type_id) { - *type_id = btf->resolved_ids[*type_id]; + *type_id =btf_resolved_type_id (btf, *type_id);return btf_type_by_id (btf, *type_id); } +static u32 btf_resolved_type_size (const struct btf *btf,u32 type_id) +{ +while (type_id < btf->start_id) + btf = btf->base_btf; + +return btf->resolved_sizes[type_id - btf->start_id]; +} +const struct btf_type *btf_type_id_size (const struct btf *btf,u32 *type_id,u32 *ret_size) {
@@ -1731,7 +1791,7 @@ const struct btf_type *btf_type_id_size(const struct btf *btf,if (btf_type_has_size (size_type)) { size = size_type->size; }else if (btf_type_is_array (size_type)) { - size = btf->resolved_sizes[size_type_id]; + size =btf_resolved_type_size (btf, size_type_id); }else if (btf_type_is_ptr (size_type)) { size =sizeof (void *); }else {
@@ -1739,14 +1799,14 @@ const struct btf_type *btf_type_id_size(const struct btf *btf, !btf_type_is_var(size_type)))return NULL ; - size_type_id = btf->resolved_ids[size_type_id]; + size_type_id =btf_resolved_type_id (btf, size_type_id); size_type =btf_type_by_id (btf, size_type_id);if (btf_type_nosize_or_null (size_type))return NULL ;else if (btf_type_has_size (size_type)) size = size_type->size;else if (btf_type_is_array (size_type)) - size = btf->resolved_sizes[size_type_id]; + size =btf_resolved_type_size (btf, size_type_id);else if (btf_type_is_ptr (size_type)) size =sizeof (void *);else
@@ -3798,7 +3858,7 @@ static int btf_check_all_metas(struct btf_verifier_env *env) cur = btf->nohdr_data + hdr->type_off; end = cur + hdr->type_len; - env->log_type_id =1 ; + env->log_type_id = btf->base_btf ? btf->start_id :1 ;while (cur < end) {struct btf_type *t = cur;s32 meta_size;
@@ -3825,8 +3885,8 @@ static bool btf_resolve_valid(struct btf_verifier_env *env,return false;if (btf_type_is_struct (t) ||btf_type_is_datasec (t)) -return !btf->resolved_ids[type_id] && - !btf->resolved_sizes[type_id]; +return !btf_resolved_type_id (btf, type_id) && + !btf_resolved_type_size (btf, type_id);if (btf_type_is_modifier (t) ||btf_type_is_ptr (t) ||btf_type_is_var (t)) {
@@ -3846,7 +3906,7 @@ static bool btf_resolve_valid(struct btf_verifier_env *env, elem_type =btf_type_id_size (btf, &elem_type_id, &elem_size);return elem_type && !btf_type_is_modifier (elem_type) && (array->nelems * elem_size == - btf->resolved_sizes[type_id]); +btf_resolved_type_size (btf, type_id)); }return false;
@@ -3888,7 +3948,8 @@ static int btf_resolve(struct btf_verifier_env *env,static int btf_check_all_types (struct btf_verifier_env *env) {struct btf *btf = env->btf; -u32 type_id; +const struct btf_type *t; +u32 type_id, i;int err; err =env_resolve_init (env);
@@ -3896,8 +3957,9 @@ static int btf_check_all_types(struct btf_verifier_env *env)return err; env->phase++; -for (type_id =1 ; type_id <= btf->nr_types; type_id++) { -const struct btf_type *t =btf_type_by_id (btf, type_id); +for (i = btf->base_btf ?0 :1 ; i < btf->nr_types; i++) { + type_id = btf->start_id + i; + t =btf_type_by_id (btf, type_id); env->log_type_id = type_id; if (btf_type_needs_resolve(t) &&
@@ -3934,7 +3996,7 @@ static int btf_parse_type_sec(struct btf_verifier_env *env)return -EINVAL ; } -if (!hdr->type_len) { +if (!env->btf->base_btf && !hdr->type_len) {btf_verifier_log (env,"No type found" );return -EINVAL ; }
@@ -3961,13 +4023,18 @@ static int btf_parse_str_sec(struct btf_verifier_env *env)return -EINVAL ; } -if (!hdr->str_len || hdr->str_len -1 >BTF_MAX_NAME_OFFSET || - start[0 ] || end[-1 ]) { + btf->strings = start; + +if (btf->base_btf && !hdr->str_len) +return 0 ; +if (!hdr->str_len || hdr->str_len -1 >BTF_MAX_NAME_OFFSET || end[-1 ]) { +btf_verifier_log (env,"Invalid string section" ); +return -EINVAL ; + } +if (!btf->base_btf && start[0 ]) {btf_verifier_log (env,"Invalid string section" );return -EINVAL ; } - - btf->strings = start;return 0 ; }
@@ -4908,7 +4975,7 @@ static int __get_type_size(struct btf *btf, u32 btf_id,while (t &&btf_type_is_modifier (t)) t =btf_type_by_id (btf, t->type);if (!t) { - *bad_type = btf->types[0 ]; + *bad_type =btf_type_by_id (btf,0 );return -EINVAL ; }if (btf_type_is_ptr (t))
Signature · 1 line
2.24.1
Re: [PATCH v4 bpf-next 1/5] bpf: add in-kernel split BTF support
Quoted 77 lines · Andrii Nakryiko
Adjust in-kernel BTF implementation to support a split BTF mode of operation. Changes are mostly mirroring libbpf split BTF changes, with the exception of
start_idbeing 0 for in-kernel implementation due to simpler read-only mode.Otherwise, for split BTF logic, most of the logic of jumping to base BTF, where necessary, is encapsulated in few helper functions. Type numbering and string offset in a split BTF are logically continuing where base BTF ends, so most of the high-level logic is kept without changes.
Type verification and size resolution is only doing an added resolution of new split BTF types and relies on already cached size and type resolution results in the base BTF.
Signed-off-by Andrii Nakryiko <[email protected]>---
kernel/bpf/btf.c| 171 +++++++++++++++++++++++++++++++++--------------1 file changed, 119 insertions(+), 52 deletions(-)kernel/bpf/btf.c +22
-4@@ -203,12 +203,17 @@ struct btf {const char *strings;void *nohdr_data;struct btf_header hdr; -u32 nr_types; +u32 nr_types;/* includes VOID for base BTF */ u32 types_size;u32 data_size;refcount_t refcnt;u32 id;struct rcu_head rcu; + +/* split BTF support */ +struct btf *base_btf; +u32 start_id;/* first type ID in this BTF (0 for base BTF) */ +u32 start_str_off;/* first string offset (0 for base BTF) */ };enum verifier_phase {@@ -449,14 +454,27 @@ static bool btf_type_is_datasec(const struct btf_type *t)return BTF_INFO_KIND (t->info) ==BTF_KIND_DATASEC ; } +static u32 btf_nr_types_total (const struct btf *btf) +{ +u32 total =0 ; + +while (btf) { + total += btf->nr_types; + btf = btf->base_btf; + } + +return total; +} +s32 btf_find_by_name_kind (const struct btf *btf,const char *name,u8 kind) {const struct btf_type *t;const char *tname; -u32 i; +u32 i, total; -for (i =1 ; i <= btf->nr_types; i++) { - t = btf->types[i]; + total =btf_nr_types_total (btf); +for (i =1 ; i < total; i++) { + t =btf_type_by_id (btf, i);if (BTF_INFO_KIND (t->info) != kind)continue ;
kernel/bpf/btf.c static const struct btf_kind_operations *btf_type_ops(const struct btf_type *t) · line 624
Quoted 10 lines
kernel/bpf/btf.c +5
-2@@ -599,8 +617,14 @@ static const struct btf_kind_operations *btf_type_ops(const struct btf_type *t)static bool btf_name_offset_valid (const struct btf *btf,u32 offset) { -return BTF_STR_OFFSET_VALID (offset) && - offset < btf->hdr.str_len; +if (!BTF_STR_OFFSET_VALID (offset)) +return false; + +while (offset < btf->start_str_off) + btf = btf->base_btf;
Do we need "if (!btf) return false;" in the while loop? (and some other loops below)
[...]
Re: [PATCH v4 bpf-next 1/5] bpf: add in-kernel split BTF support
Quoted 79 lines · Song Liu
Quoted 79 lines · Andrii Nakryiko
Adjust in-kernel BTF implementation to support a split BTF mode of operation. Changes are mostly mirroring libbpf split BTF changes, with the exception of
start_idbeing 0 for in-kernel implementation due to simpler read-only mode.Otherwise, for split BTF logic, most of the logic of jumping to base BTF, where necessary, is encapsulated in few helper functions. Type numbering and string offset in a split BTF are logically continuing where base BTF ends, so most of the high-level logic is kept without changes.
Type verification and size resolution is only doing an added resolution of new split BTF types and relies on already cached size and type resolution results in the base BTF.
Signed-off-by Andrii Nakryiko <[email protected]>---
kernel/bpf/btf.c| 171 +++++++++++++++++++++++++++++++++--------------1 file changed, 119 insertions(+), 52 deletions(-)kernel/bpf/btf.c +22
-4@@ -203,12 +203,17 @@ struct btf {const char *strings;void *nohdr_data;struct btf_header hdr; -u32 nr_types; +u32 nr_types;/* includes VOID for base BTF */ u32 types_size;u32 data_size;refcount_t refcnt;u32 id;struct rcu_head rcu; + +/* split BTF support */ +struct btf *base_btf; +u32 start_id;/* first type ID in this BTF (0 for base BTF) */ +u32 start_str_off;/* first string offset (0 for base BTF) */ };enum verifier_phase {@@ -449,14 +454,27 @@ static bool btf_type_is_datasec(const struct btf_type *t)return BTF_INFO_KIND (t->info) ==BTF_KIND_DATASEC ; } +static u32 btf_nr_types_total (const struct btf *btf) +{ +u32 total =0 ; + +while (btf) { + total += btf->nr_types; + btf = btf->base_btf; + } + +return total; +} +s32 btf_find_by_name_kind (const struct btf *btf,const char *name,u8 kind) {const struct btf_type *t;const char *tname; -u32 i; +u32 i, total; -for (i =1 ; i <= btf->nr_types; i++) { - t = btf->types[i]; + total =btf_nr_types_total (btf); +for (i =1 ; i < total; i++) { + t =btf_type_by_id (btf, i);if (BTF_INFO_KIND (t->info) != kind)continue ;@@ -599,8 +617,14 @@ static const struct btf_kind_operations *btf_type_ops(const struct btf_type *t)
Quoted 11 lines
kernel/bpf/btf.c static const struct btf_kind_operations *btf_type_ops(const struct btf_type *t) · line 624
Quoted 9 lines
kernel/bpf/btf.c +5
-2static bool btf_name_offset_valid (const struct btf *btf,u32 offset) { -return BTF_STR_OFFSET_VALID (offset) && - offset < btf->hdr.str_len; +if (!BTF_STR_OFFSET_VALID (offset)) +return false; + +while (offset < btf->start_str_off) + btf = btf->base_btf;Do we need "if (!btf) return false;" in the while loop? (and some other loops below)
No, because for base btf start_str_off and start_type_id are always zero, so loop condition is always false.
Quoted 3 lines
[...]
Song Liu
Acked-byRe: [PATCH v4 bpf-next 1/5] bpf: add in-kernel split BTF support
Quoted 96 lines · Andrii Nakryiko
Quoted 79 lines · Song Liu
Quoted 79 lines · Andrii Nakryiko
Adjust in-kernel BTF implementation to support a split BTF mode of operation. Changes are mostly mirroring libbpf split BTF changes, with the exception of
start_idbeing 0 for in-kernel implementation due to simpler read-only mode.Otherwise, for split BTF logic, most of the logic of jumping to base BTF, where necessary, is encapsulated in few helper functions. Type numbering and string offset in a split BTF are logically continuing where base BTF ends, so most of the high-level logic is kept without changes.
Type verification and size resolution is only doing an added resolution of new split BTF types and relies on already cached size and type resolution results in the base BTF.
Signed-off-by Andrii Nakryiko <[email protected]>---
kernel/bpf/btf.c| 171 +++++++++++++++++++++++++++++++++--------------1 file changed, 119 insertions(+), 52 deletions(-)kernel/bpf/btf.c +22
-4@@ -203,12 +203,17 @@ struct btf {const char *strings;void *nohdr_data;struct btf_header hdr; -u32 nr_types; +u32 nr_types;/* includes VOID for base BTF */ u32 types_size;u32 data_size;refcount_t refcnt;u32 id;struct rcu_head rcu; + +/* split BTF support */ +struct btf *base_btf; +u32 start_id;/* first type ID in this BTF (0 for base BTF) */ +u32 start_str_off;/* first string offset (0 for base BTF) */ };enum verifier_phase {@@ -449,14 +454,27 @@ static bool btf_type_is_datasec(const struct btf_type *t)return BTF_INFO_KIND (t->info) ==BTF_KIND_DATASEC ; } +static u32 btf_nr_types_total (const struct btf *btf) +{ +u32 total =0 ; + +while (btf) { + total += btf->nr_types; + btf = btf->base_btf; + } + +return total; +} +s32 btf_find_by_name_kind (const struct btf *btf,const char *name,u8 kind) {const struct btf_type *t;const char *tname; -u32 i; +u32 i, total; -for (i =1 ; i <= btf->nr_types; i++) { - t = btf->types[i]; + total =btf_nr_types_total (btf); +for (i =1 ; i < total; i++) { + t =btf_type_by_id (btf, i);if (BTF_INFO_KIND (t->info) != kind)continue ;@@ -599,8 +617,14 @@ static const struct btf_kind_operations *btf_type_ops(const struct btf_type *t)Quoted 11 lines
kernel/bpf/btf.c static const struct btf_kind_operations *btf_type_ops(const struct btf_type *t) · line 624
Quoted 9 lines
kernel/bpf/btf.c +5
-2static bool btf_name_offset_valid (const struct btf *btf,u32 offset) { -return BTF_STR_OFFSET_VALID (offset) && - offset < btf->hdr.str_len; +if (!BTF_STR_OFFSET_VALID (offset)) +return false; + +while (offset < btf->start_str_off) + btf = btf->base_btf;Do we need "if (!btf) return false;" in the while loop? (and some other loops below)
Quoted 2 lines
No, because for base btf
start_str_offandstart_type_idare always zero, so loop condition is always false.
Ah, I misread the code. Thanks for the explanation.
Andrii Nakryiko
maintainerPATCH v4 2/5 bpf: assign ID to vmlinux BTF and return extra info for BTF in GET_OBJ_INFO
In reply to Andrii NakryikoAllocate ID for vmlinux BTF. This makes it visible when iterating over all BTF objects in the system. To allow distinguishing vmlinux BTF (and later kernel module BTF) from user-provided BTFs, expose extra kernel_btf flag, as well as BTF name ("vmlinux" for vmlinux BTF, will equal to module's name for module BTF). We might want to later allow specifying BTF name for user-provided BTFs as well, if that makes sense. But currently this is reserved only for in-kernel BTFs.
Having in-kernel BTFs exposed IDs will allow to extend BPF APIs that require in-kernel BTF type with ability to specify BTF types from kernel modules, not just vmlinux BTF. This will be implemented in a follow up patch set for fentry/fexit/fmod_ret/lsm/etc.
--- include/uapi/linux/bpf.h | 3 +++ kernel/bpf/btf.c | 43 +++++++++++++++++++++++++++++++--- tools/include/uapi/linux/bpf.h | 3 +++ 3 files changed, 46 insertions(+), 3 deletions(-)
include/uapi/linux/bpf.h +3 -0
@@ -4466,6 +4466,9 @@ struct bpf_btf_info {__aligned_u64 btf;__u32 btf_size;__u32 id; +__aligned_u64 name; +__u32 name_len; +__u32 kernel_btf; } __attribute__((aligned(8)));struct bpf_link_info {
kernel/bpf/btf.c +40 -3
78 lines not on this page · raw message
tools/include/uapi/linux/bpf.h +3 -0
@@ -4466,6 +4466,9 @@ struct bpf_btf_info {__aligned_u64 btf;__u32 btf_size;__u32 id; +__aligned_u64 name; +__u32 name_len; +__u32 kernel_btf; } __attribute__((aligned(8)));struct bpf_link_info {
Signature · 1 line
2.24.1
Re: [PATCH v4 bpf-next 2/5] bpf: assign ID to vmlinux BTF and return extra info for BTF in GET_OBJ_INFO
Quoted 12 lines · Andrii Nakryiko
Allocate ID for vmlinux BTF. This makes it visible when iterating over all BTF objects in the system. To allow distinguishing vmlinux BTF (and later kernel module BTF) from user-provided BTFs, expose extra
kernel_btfflag, as well as BTF name ("vmlinux" for vmlinux BTF, will equal to module's name for module BTF). We might want to later allow specifying BTF name for user-provided BTFs as well, if that makes sense. But currently this is reserved only for in-kernel BTFs.
Quoted 6 lines
Having in-kernel BTFs exposed IDs will allow to extend BPF APIs that require in-kernel BTF type with ability to specify BTF types from kernel modules, not just vmlinux BTF. This will be implemented in a follow up patch set for fentry/fexit/fmod_ret/lsm/etc.
Signed-off-by Andrii Nakryiko <[email protected]>
Andrii Nakryiko
maintainerPATCH v4 3/5 kbuild: build kernel module BTFs if BTF is enabled and pahole supports it
In reply to Andrii NakryikoDetect if pahole supports split BTF generation, and generate BTF for each selected kernel module, if it does. This is exposed to Makefiles and C code as CONFIG_DEBUG_INFO_BTF_MODULES flag.
Kernel module BTF has to be re-generated if either vmlinux's BTF changes or module's .ko changes. To achieve that, I needed a helper similar to if_changed, but that would allow to filter out vmlinux from the list of updated dependencies for .ko building. I've put it next to the only place that uses and needs it, but it might be a better idea to just add it along the other if_changed variants into scripts/Kbuild.include.
Each kernel module's BTF deduplication is pretty fast, as it does only incremental BTF deduplication on top of already deduplicated vmlinux BTF. To show the added build time, I've first ran make only just built kernel (to establish the baseline) and then forced only BTF re-generation, without regenerating .ko files. The build was performed with -j60 parallelization on 56-core machine. The final time also includes bzImage building, so it's not a pure BTF overhead.
$ time make -j60 ... make -j60 27.65s user 10.96s system 782% cpu 4.933 total $ touch ~/linux-build/default/vmlinux && time make -j60 ... make -j60 123.69s user 27.85s system 1566% cpu 9.675 total
So 4.6 seconds real time, with noticeable part spent in compressed vmlinux and bzImage building.
To show size savings, I've built my kernel configuration with about 700 kernel modules with full BTF per each kernel module (without deduplicating against vmlinux) and with split BTF against deduplicated vmlinux (approach in this patch). Below are top 10 modules with biggest BTF sizes. And total size of BTF data across all kernel modules.
It shows that split BTF "compresses" 115MB down to 5MB total. And the biggest kernel modules get a downsize from 500-570KB down to 200-300KB.
FULL BTF
========
$ for f in $(find . -name '*.ko'); do size -A -d $f | grep BTF | awk '{print $2}'; done | awk '{ s += $1 } END { print s }'
115710691$ for f in $(find . -name '*.ko'); do printf "%s %d\n" $f $(size -A -d $f | grep BTF | awk '{print $2}'); done | sort -nr -k2 | head -n10
./drivers/gpu/drm/i915/i915.ko 570570
./drivers/net/ethernet/mellanox/mlx5/core/mlx5_core.ko 520240
./drivers/gpu/drm/radeon/radeon.ko 503849
./drivers/infiniband/hw/mlx5/mlx5_ib.ko 491777
./fs/xfs/xfs.ko 411544
./drivers/net/ethernet/intel/i40e/i40e.ko 403904
./drivers/net/ethernet/broadcom/bnx2x/bnx2x.ko 398754
./drivers/infiniband/core/ib_core.ko 397224
./fs/cifs/cifs.ko 386249
./fs/nfsd/nfsd.ko 379738SPLIT BTF
=========
$ for f in $(find . -name '*.ko'); do size -A -d $f | grep BTF | awk '{print $2}'; done | awk '{ s += $1 } END { print s }'
5194047$ for f in $(find . -name '*.ko'); do printf "%s %d\n" $f $(size -A -d $f | grep BTF | awk '{print $2}'); done | sort -nr -k2 | head -n10
./drivers/gpu/drm/i915/i915.ko 293206
./drivers/gpu/drm/radeon/radeon.ko 282103
./fs/xfs/xfs.ko 222150
./drivers/net/ethernet/mellanox/mlx5/core/mlx5_core.ko 198503
./drivers/infiniband/hw/mlx5/mlx5_ib.ko 198356
./drivers/net/ethernet/broadcom/bnx2x/bnx2x.ko 113444
./fs/cifs/cifs.ko 109379
./arch/x86/kvm/kvm.ko 100225
./drivers/gpu/drm/drm.ko 94827
./drivers/infiniband/core/ib_core.ko 91188--- lib/Kconfig.debug | 9 +++++++++ scripts/Makefile.modfinal | 20 ++++++++++++++++++-- 2 files changed, 27 insertions(+), 2 deletions(-)
lib/Kconfig.debug +9 -0
@@ -274,6 +274,15 @@ config DEBUG_INFO_BTF Turning this on expects presence of pahole tool, which will convert DWARF type info into equivalent deduplicated BTF type info. +config PAHOLE_HAS_SPLIT_BTF + def_bool $(success, test `$(PAHOLE) --version | sed -E 's/v([0-9]+)\.([0-9]+)/\1\2/'` -ge "119") + +config DEBUG_INFO_BTF_MODULES + def_bool y + depends on DEBUG_INFO_BTF && MODULES && PAHOLE_HAS_SPLIT_BTF + help + Generate compact split BTF type information for kernel modules. + config GDB_SCRIPTS bool "Provide GDB scripts for kernel debugging" help
scripts/Makefile.modfinal +18 -2
@@ -6,6 +6,7 @@ PHONY := __modfinal __modfinal: +include include/config/auto.conf include $(srctree)/scripts/Kbuild.include # for c_flags
@@ -36,8 +37,23 @@ quiet_cmd_ld_ko_o = LD [M] $@ -T scripts/module.lds -o $@ $(filter %.o, $^); \ $(if $(ARCH_POSTLINK), $(MAKE) -f $(ARCH_POSTLINK) $@, true) -$(modules): %.ko: %.o %.mod.o scripts/module.lds FORCE - +$(call if_changed,ld_ko_o) +quiet_cmd_btf_ko = BTF [M] $@ + cmd_btf_ko = LLVM_OBJCOPY=$(OBJCOPY) $(PAHOLE) -J --btf_base vmlinux $@ + +# Same as newer-prereqs, but allows to exclude specified extra dependencies +newer_prereqs_except = $(filter-out $(PHONY) $(1),$?) + +# Same as if_changed, but allows to exclude specified extra dependencies +if_changed_except = $(if $(call newer_prereqs_except,$(2))$(cmd-check), \ + $(cmd); \ + printf '%s\n' 'cmd_$@ := $(make-cmd)' > $(dot-target).cmd, @:) + +# Re-generate module BTFs if either module's .ko or vmlinux changed +$(modules): %.ko: %.o %.mod.o scripts/module.lds vmlinux FORCE + +$(call if_changed_except,ld_ko_o,vmlinux) +ifdef CONFIG_DEBUG_INFO_BTF_MODULES + +$(if $(newer-prereqs),$(call cmd,btf_ko)) +endif targets += $(modules) $(modules:.ko=.mod.o)
Signature · 1 line
2.24.1
Re: [PATCH v4 bpf-next 3/5] kbuild: build kernel module BTFs if BTF is enabled and pahole supports it
Quoted 3 lines · Andrii Nakryiko
[...]
Quoted 18 lines
SPLIT BTF
=========$ for f in $(find . -name '*.ko'); do size -A -d $f | grep BTF | awk '{print $2}'; done | awk '{ s += $1 } END { print s }' 5194047$ for f in $(find . -name '*.ko'); do printf "%s %d\n" $f $(size -A -d $f | grep BTF | awk '{print $2}'); done | sort -nr -k2 | head -n10 ./drivers/gpu/drm/i915/i915.ko 293206 ./drivers/gpu/drm/radeon/radeon.ko 282103 ./fs/xfs/xfs.ko 222150 ./drivers/net/ethernet/mellanox/mlx5/core/mlx5_core.ko 198503 ./drivers/infiniband/hw/mlx5/mlx5_ib.ko 198356 ./drivers/net/ethernet/broadcom/bnx2x/bnx2x.ko 113444 ./fs/cifs/cifs.ko 109379 ./arch/x86/kvm/kvm.ko 100225 ./drivers/gpu/drm/drm.ko 94827 ./drivers/infiniband/core/ib_core.ko 91188
Quoted 2 lines
Cc Masahiro Yamada <[email protected]>Signed-off-by Andrii Nakryiko <[email protected]>
RE: [PATCH v4 bpf-next 3/5] kbuild: build kernel module BTFs if BTF is enabled and pahole supports it
Quoted 40 lines
-----Original Message-----
From: Song Liu <[email protected]>
Sent: Tuesday, November 10, 2020 5:05 PM
To: Andrii Nakryiko <[email protected]>
Cc: bpf <[email protected]>; Networking <[email protected]>; Starovoitov, Alexei <[email protected]>; Daniel Borkmann <[email protected]>; Kernel Team <[email protected]>; open list <linux- [email protected]>; [email protected]; [email protected]; Arnaldo Carvalho de Melo <[email protected]>; Greg Kroah-Hartman <[email protected]>; Masahiro Yamada <[email protected]>
Subject: Re: [PATCH v4 bpf-next 3/5] kbuild: build kernel module BTFs if BTF is enabled and pahole supports itQuoted 1 line · Andrii Nakryiko
[...]
Quoted 4 lines
SPLIT BTF
=========$ for f in $(find . -name '*.ko'); do size -A -d $f | grep BTF | awk '{print $2}';done | awk '{ s += $1 } END { print s }'
Quoted 3 lines
5194047
$ for f in $(find . -name '*.ko'); do printf "%s %d\n" $f $(size -A -d $f | grepBTF | awk '{print $2}'); done | sort -nr -k2 | head -n10
Quoted 11 lines
./drivers/gpu/drm/i915/i915.ko293206./drivers/gpu/drm/radeon/radeon.ko282103./fs/xfs/xfs.ko222150./drivers/net/ethernet/mellanox/mlx5/core/mlx5_core.ko198503./drivers/infiniband/hw/mlx5/mlx5_ib.ko198356./drivers/net/ethernet/broadcom/bnx2x/bnx2x.ko113444./fs/cifs/cifs.ko109379./arch/x86/kvm/kvm.ko100225./drivers/gpu/drm/drm.ko94827./drivers/infiniband/core/ib_core.ko91188
Quoted 4 lines
Quoted 2 lines
Cc Masahiro Yamada <[email protected]>Signed-off-by Andrii Nakryiko <[email protected]>Acked-by Song Liu <[email protected]>
This change, commit 5f9ae91f7c0d ("kbuild: Build kernel module BTFs if BTF is enabled and pahole supports it") currently in net-next, linux-next, etc. breaks the use-case of compiling only a specific kernel module (both in-tree and out-of-tree, e.g. 'make M=drivers/net/ethernet/intel/ice') after first doing a 'make modules_prepare'. Previously, that use-case would result in a warning noting "Symbol info of vmlinux is missing. Unresolved symbol check will be entirely skipped" but now it errors out after noting "No rule to make target 'vmlinux', needed by '<...>.ko'. Stop."
Is that intentional?
Thanks,
Bruce.
Re: [PATCH v4 bpf-next 3/5] kbuild: build kernel module BTFs if BTF is enabled and pahole supports it
Quoted 46 lines · Bruce W
Quoted 41 lines
-----Original Message-----
From: Song Liu <[email protected]>
Sent: Tuesday, November 10, 2020 5:05 PM
To: Andrii Nakryiko <[email protected]>
Cc: bpf <[email protected]>; Networking <[email protected]>; Starovoitov, Alexei <[email protected]>; Daniel Borkmann <[email protected]>; Kernel Team <[email protected]>; open list <linux- [email protected]>; [email protected]; [email protected]; Arnaldo Carvalho de Melo <[email protected]>; Greg Kroah-Hartman <[email protected]>; Masahiro Yamada <[email protected]>
Subject: Re: [PATCH v4 bpf-next 3/5] kbuild: build kernel module BTFs if BTF is enabled and pahole supports itQuoted 1 line · Andrii Nakryiko
[...]
Quoted 4 lines
SPLIT BTF
=========$ for f in $(find . -name '*.ko'); do size -A -d $f | grep BTF | awk '{print $2}';done | awk '{ s += $1 } END { print s }'
Quoted 3 lines
5194047
$ for f in $(find . -name '*.ko'); do printf "%s %d\n" $f $(size -A -d $f | grepBTF | awk '{print $2}'); done | sort -nr -k2 | head -n10
Quoted 11 lines
./drivers/gpu/drm/i915/i915.ko293206./drivers/gpu/drm/radeon/radeon.ko282103./fs/xfs/xfs.ko222150./drivers/net/ethernet/mellanox/mlx5/core/mlx5_core.ko198503./drivers/infiniband/hw/mlx5/mlx5_ib.ko198356./drivers/net/ethernet/broadcom/bnx2x/bnx2x.ko113444./fs/cifs/cifs.ko109379./arch/x86/kvm/kvm.ko100225./drivers/gpu/drm/drm.ko94827./drivers/infiniband/core/ib_core.ko91188Quoted 4 lines
Quoted 2 lines
Cc Masahiro Yamada <[email protected]>Signed-off-by Andrii Nakryiko <[email protected]>Acked-by Song Liu <[email protected]>
Quoted 8 lines
This change, commit
5f9ae91f7c0d("kbuild: Build kernel module BTFs if BTF is enabled and pahole supports it") currently in net-next, linux-next, etc. breaks the use-case of compiling only a specific kernel module (both in-tree and out-of-tree, e.g. 'make M=drivers/net/ethernet/intel/ice') after first doing a 'makemodules_prepare'. Previously, that use-case would result in a warning noting "Symbol info of vmlinux is missing. Unresolved symbol check will be entirely skipped" but now it errors out after noting "No rule to make target 'vmlinux', needed by '<...>.ko'. Stop."Is that intentional?
I wasn't aware of such a use pattern, so definitely not intentional. But vmlinux is absolutely necessary to generate the module BTF. So I'm wondering what's the proper fix here? Leave it as is (that error message is actually surprisingly descriptive, btw)? Force vmlinux build? Or skip BTF generation for that module?
Quoted 3 lines
Thanks,
Bruce.
Jakub Kicinski
maintainerRe: [PATCH v4 bpf-next 3/5] kbuild: build kernel module BTFs if BTF is enabled and pahole supports it
Quoted 7 lines · On Mon, 16 Nov 2020 12:34:17 -0800 Andrii Nakryiko wrote:
Quoted 7 lines
This change, commit
5f9ae91f7c0d("kbuild: Build kernel module BTFs if BTF is enabled and pahole supports it") currently in net-next, linux-next, etc. breaks the use-case of compiling only a specific kernel module (both in-tree and out-of-tree, e.g. 'make M=drivers/net/ethernet/intel/ice') after first doing a 'makemodules_prepare'. Previously, that use-case would result in a warning noting "Symbol info of vmlinux is missing. Unresolved symbol check will be entirely skipped" but now it errors out after noting "No rule to make target 'vmlinux', needed by '<...>.ko'. Stop."
Quoted 7 lines
Quoted 1 line
Is that intentional?
I wasn't aware of such a use pattern, so definitely not intentional. But vmlinux is absolutely necessary to generate the module BTF. So I'm wondering what's the proper fix here? Leave it as is (that error message is actually surprisingly descriptive, btw)? Force vmlinux build? Or skip BTF generation for that module?
For an external out-of-tree module there is no guarantee that vmlinux will even be on the system, no? So only the last option can work in that case.
Re: [PATCH v4 bpf-next 3/5] kbuild: build kernel module BTFs if BTF is enabled and pahole supports it
Quoted 17 lines · Jakub Kicinski
Quoted 7 lines · On Mon, 16 Nov 2020 12:34:17 -0800 Andrii Nakryiko wrote:
Quoted 7 lines
This change, commit
5f9ae91f7c0d("kbuild: Build kernel module BTFs if BTF is enabled and pahole supports it") currently in net-next, linux-next, etc. breaks the use-case of compiling only a specific kernel module (both in-tree and out-of-tree, e.g. 'make M=drivers/net/ethernet/intel/ice') after first doing a 'makemodules_prepare'. Previously, that use-case would result in a warning noting "Symbol info of vmlinux is missing. Unresolved symbol check will be entirely skipped" but now it errors out after noting "No rule to make target 'vmlinux', needed by '<...>.ko'. Stop."Quoted 7 lines
Quoted 1 line
Is that intentional?
I wasn't aware of such a use pattern, so definitely not intentional. But vmlinux is absolutely necessary to generate the module BTF. So I'm wondering what's the proper fix here? Leave it as is (that error message is actually surprisingly descriptive, btw)? Force vmlinux build? Or skip BTF generation for that module?
Quoted 3 lines
For an external out-of-tree module there is no guarantee that vmlinux will even be on the system, no? So only the last option can work in that case.
Ok, how about something like the patch below. With that I seem to be getting the desired behavior:
$ make clean $ touch drivers/acpi/button.c $ make M=drivers/acpi make[1]: Entering directory `/data/users/andriin/linux-build/default-x86_64' CC [M] drivers/acpi/button.o MODPOST drivers/acpi/Module.symvers LD [M] drivers/acpi/button.ko BTF [M] drivers/acpi/button.ko Skipping BTF generation for drivers/acpi/button.ko due to unavailability of vmlinux make[1]: Leaving directory `/data/users/andriin/linux-build/default-x86_64' $ readelf -S ~/linux-build/default/drivers/acpi/button.ko | grep BTF -A1 ... empty ...
Now with normal build:
$ make all
...
LD [M] drivers/acpi/button.ko
BTF [M] drivers/acpi/button.ko
...
$ readelf -S ~/linux-build/default/drivers/acpi/button.ko | grep BTF -A1
[60] .BTF PROGBITS 0000000000000000 00029310
000000000000ab3f 0000000000000000 0 0 1scripts/Makefile.modfinal +5 -1
@@ -38,7 +38,12 @@ quiet_cmd_ld_ko_o = LD [M] $@ $(if $(ARCH_POSTLINK), $(MAKE) -f $(ARCH_POSTLINK) $@, true) quiet_cmd_btf_ko = BTF [M] $@ - cmd_btf_ko = LLVM_OBJCOPY=$(OBJCOPY) $(PAHOLE) -J --btf_base vmlinux $@ + cmd_btf_ko = \ + if [ -f vmlinux ]; then \ + LLVM_OBJCOPY=$(OBJCOPY) $(PAHOLE) -J --btf_base vmlinux $@; \ + else \ + printf "Skipping BTF generation for %s due to unavailability
of vmlinux\n" $@ 1>&2; \
+ fi;
# Same as newer-prereqs, but allows to exclude specified extra dependencies newer_prereqs_except = $(filter-out $(PHONY) $(1),$?)
@@ -49,7 +54,7 @@ if_changed_except = $(if $(call
newer_prereqs_except,$(2))$(cmd-check), \
printf '%s\n' 'cmd_$@ := $(make-cmd)' > $(dot-target).cmd, @:) # Re-generate module BTFs if either module's .ko or vmlinux changed
-$(modules): %.ko: %.o %.mod.o scripts/module.lds vmlinux FORCE +$(modules): %.ko: %.o %.mod.o scripts/module.lds $(if $(KBUILD_BUILTIN),vmlinux) FORCE
+$(call if_changed_except,ld_ko_o,vmlinux)
ifdef CONFIG_DEBUG_INFO_BTF_MODULES
+$(if $(newer-prereqs),$(call cmd,btf_ko))Re: [PATCH v4 bpf-next 3/5] kbuild: build kernel module BTFs if BTF is enabled and pahole supports it
In reply to Jakub KicinskiQuoted 16 lines · Jakub Kicinski
Quoted 7 lines · On Mon, 16 Nov 2020 12:34:17 -0800 Andrii Nakryiko wrote:
Quoted 7 lines
This change, commit
5f9ae91f7c0d("kbuild: Build kernel module BTFs if BTF is enabled and pahole supports it") currently in net-next, linux-next, etc. breaks the use-case of compiling only a specific kernel module (both in-tree and out-of-tree, e.g. 'make M=drivers/net/ethernet/intel/ice') after first doing a 'makemodules_prepare'. Previously, that use-case would result in a warning noting "Symbol info of vmlinux is missing. Unresolved symbol check will be entirely skipped" but now it errors out after noting "No rule to make target 'vmlinux', needed by '<...>.ko'. Stop."Quoted 7 lines
Quoted 1 line
Is that intentional?
I wasn't aware of such a use pattern, so definitely not intentional. But vmlinux is absolutely necessary to generate the module BTF. So I'm wondering what's the proper fix here? Leave it as is (that error message is actually surprisingly descriptive, btw)? Force vmlinux build? Or skip BTF generation for that module?
Quoted 3 lines
For an external out-of-tree module there is no guarantee that vmlinux will even be on the system, no? So only the last option can work in that case.
a year late to the party, but it seems to me that this patch series/features also missed another, not yet fixed scenario. I have to admit I am not very well-versed in BTF/BPF matters though, so please take the analysis below with a grain of salt or two ;)
(am subscribed to LKML/netdev, but not the bpf list, so please keep me CCed if discussion moves there! apologies if too many people are CCed here, feel free to trim down to relevant people/lists)
many distros do their own tracking of kernel <-> module ABI (usually these checks use Module.symvers and some combination of lists/symbols/.. to skip/ignore[0]).
depending on detected changes, a kernel update can either
- bump ABI, resulting in a new kernel/modules package that is installed next to the current one
- keep ABI, resulting in an updated kernel/modules package that is installed over/instead of the current one
the former case is obviously not an issue, since the modules and vmlinux image shipped in that (set of) package(s) match. but in the later case of updated, compatible kernel image + modules with unchanged ABI (which is important, as it allows shipping fixed modules that are loadable for a compatible, older, booted kernel image), the following is possible:
- install kernel+modules with ABI 1
- boot kernel with ABI 1
- install ABI-compatible upgrade (e.g. a security fix)
- load module
- BTF validation fails, because the base_btf (loaded at boot time) and the offsets in the module's .BTF section (loaded at module load time) aren't matching
of course the validation might also not fail but the parsed BTF info might be bogus, or the base_btf might be similar enough that validation passes and the parsed BTF data is correct.
in our case the symptoms look like this (exact details vary with kernel builds/modules, but likely not relevant):
Nov 24 11:39:11 host kernel: BPF: type_id=7 bits_offset=0 Nov 24 11:39:11 host kernel: BPF: Nov 24 11:39:11 host kernel: BPF:Invalid name Nov 24 11:39:11 host kernel: BPF: Nov 24 11:39:11 host kernel: failed to validate module [overlay] BTF: -22
where the booted kernel and the (attempted to get) loaded module are not from the same build, but the Module.symvers is matching and loading should thus work. adding some more debug logging reveals that the root cause is the module's BTF start_str_off being, well, off, since it's derived from vmlinux' BTF/base_btf. if it is too big, the name/type lookups will wrongly look in the base_btf, if it's too small, the name/type lookups will be offset within the module or wrongly look inside the module when they should look inside base_btf/vmlinux. in any case, random garbage is the result, usually tripping up some validation check (e.g. the first byte not being 0 when checking a name). but even if it's correct (old and new vmlinux image have the same nr_types/hdr.str_len), there is no guarantuee that the offsets into base_btf are pointing at the right stuff.
example with debug logging patched in, note the garbled names, and offset slightly below the (wrong) start_str_off:
----8<----
BPF:magic: 0xeb9f
BPF:version: 1
BPF:flags: 0x0
BPF:hdr_len: 24
BPF:type_off: 0
BPF:type_len: 9264
BPF:str_off: 9264
BPF:str_len: 5511
BPF:btf_total_size: 14799
BPF:[106314] STRUCT rimary_device
BPF:size=56 vlen=14
BPF:
BPF:offset at call: 1915394
BPF:offset too small, choosing base_btf: 1915397
BPF:offset after base_btf: 1915394
BPF: ce type_id=49 bits_offset=0
BPF:
BPF:offset at call: 1915403
BPF:offset after base_btf: 6
BPF: nfig type_id=49 bits_offset=64
BPF:
BPF:offset at call: 1915412
BPF:offset after base_btf: 15
BPF: rdir type_id=49 bits_offset=128
BPF:
BPF:offset at call: 768428
BPF:offset too small, choosing base_btf: 1915397
BPF:offset after base_btf: 768428
BPF: _dio type_id=56 bits_offset=192
BPF:
BPF:offset at call: 1915420
BPF:offset after base_btf: 23
BPF: erdir type_id=56 bits_offset=200
BPF:
BPF:offset at call: 1915433
BPF:offset after base_btf: 36
BPF:first char wrong - 0
BPF: type_id=56 bits_offset=208
BPF:
BPF:Invalid name STRUCT MEMBER (name offset 1915433) BPF:
failed to validate module [overlay] BTF: -22
---->8----
also note how it's only after a few botched entries that a check actually trips up - not sure what the impliciations for crafted BTF info are, but might be worthy a closer look by someone more knowledgable as well..
it seems to me this can be solved on the distro/user side by tracking vmlinux BTF infos as part of the ABI tracking (how stable are those compared to the existing interfaces making up the kernel <-> module ABI/Module.symvers? does this effectively mean bumping ABI for any change anyway?) or by disabling CONFIG_DEBUG_INFO_BTF_MODULES.
on the kernel/libbpf side it could maybe be solved by storing a hash of the base_btf data used to generate the split BTF-sections inside the modules, and skip BTF loading/validating if another base_btf is currently loaded (so BTF is best-effort, if the booted kernel and the module are matching it works, if not module loading works but no BTF support). this might be a good safe-guard for split-BTF in general?
I'd appreciate input on how to proceed (we were recently hit by this in a downstream Debian derivative, and will disable BTF info for modules as an interim measure).
thanks!
0: e.g., Debian's: https://salsa.debian.org/kernel-team/linux/-/blob/master/debian/bin/abiupdate.py
Re: [PATCH v4 bpf-next 3/5] kbuild: build kernel module BTFs if BTF is enabled and pahole supports it
Quoted 18 lines · Fabian Grünbichler
Quoted 16 lines · Jakub Kicinski
Quoted 14 lines · On Mon, 16 Nov 2020 12:34:17 -0800 Andrii Nakryiko wrote:
Quoted 8 lines
This change, commit
5f9ae91f7c0d("kbuild: Build kernel module BTFs if BTF is enabled and pahole supports it") currently in net-next, linux-next, etc. breaks the use-case of compiling only a specific kernel module (both in-tree and out-of-tree, e.g. 'make M=drivers/net/ethernet/intel/ice') after first doing a 'makemodules_prepare'. Previously, that use-case would result in a warning noting "Symbol info of vmlinux is missing. Unresolved symbol check will be entirely skipped" but now it errors out after noting "No rule to make target 'vmlinux', needed by '<...>.ko'. Stop."Is that intentional?
I wasn't aware of such a use pattern, so definitely not intentional. But vmlinux is absolutely necessary to generate the module BTF. So I'm wondering what's the proper fix here? Leave it as is (that error message is actually surprisingly descriptive, btw)? Force vmlinux build? Or skip BTF generation for that module?
Quoted 5 lines
Quoted 3 lines
For an external out-of-tree module there is no guarantee that vmlinux will even be on the system, no? So only the last option can work in that case.
a year late to the party, but it seems to me that this patch
hi, sorry, you email slipped through the cracks initially, just getting back to it...
Quoted 139 lines
series/features also missed another, not yet fixed scenario. I have to admit I am not very well-versed in BTF/BPF matters though, so please take the analysis below with a grain of salt or two ;)
(am subscribed to LKML/netdev, but not the bpf list, so please keep me CCed if discussion moves there! apologies if too many people are CCed here, feel free to trim down to relevant people/lists)
many distros do their own tracking of kernel <-> module ABI (usually these checks use Module.symvers and some combination of lists/symbols/.. to skip/ignore[0]).
depending on detected changes, a kernel update can either
- bump ABI, resulting in a new kernel/modules package that is installed next to the current one
- keep ABI, resulting in an updated kernel/modules package that is installed over/instead of the current onethe former case is obviously not an issue, since the modules and vmlinux image shipped in that (set of) package(s) match. but in the later case of updated, compatible kernel image + modules with unchanged ABI (which is important, as it allows shipping fixed modules that are loadable for a compatible, older, booted kernel image), the following is possible:
- install kernel+modules with ABI 1
- boot kernel with ABI 1
- install ABI-compatible upgrade (e.g. a security fix)
- load module
- BTF validation fails, because thebase_btf(loaded at boot time) and the offsets in the module's .BTF section (loaded at module load time) aren't matchingof course the validation might also not fail but the parsed BTF info might be bogus, or the
base_btfmight be similar enough that validation passes and the parsed BTF data is correct.in our case the symptoms look like this (exact details vary with kernel builds/modules, but likely not relevant):
Nov 24 11:39:11 host kernel: BPF: type_id=7 bits_offset=0 Nov 24 11:39:11 host kernel: BPF: Nov 24 11:39:11 host kernel: BPF:Invalid name Nov 24 11:39:11 host kernel: BPF: Nov 24 11:39:11 host kernel: failed to validate module [overlay] BTF: -22where the booted kernel and the (attempted to get) loaded module are not from the same build, but the Module.symvers is matching and loading should thus work. adding some more debug logging reveals that the root cause is the module's BTF
start_str_offbeing, well, off, since it's derived from vmlinux' BTF/base_btf. if it is too big, the name/type lookups will wrongly look in thebase_btf, if it's too small, the name/type lookups will be offset within the module or wrongly look inside the module when they should look inside base_btf/vmlinux. in any case, random garbage is the result, usually tripping up some validation check (e.g. the first byte not being 0 when checking a name). but even if it's correct (old and new vmlinux image have the same nr_types/hdr.str_len), there is no guarantuee that the offsets intobase_btfare pointing at the right stuff.example with debug logging patched in, note the garbled names, and offset slightly below the (wrong)
start_str_off:----8<----
BPF:magic:
0xeb9fBPF:version: 1
BPF:flags:
0x0BPF:hdr_len: 24
BPF:type_off: 0
BPF:type_len: 9264
BPF:str_off: 9264
BPF:str_len: 5511
BPF:btf_total_size: 14799
BPF:[106314] STRUCT
rimary_device
BPF:size=56 vlen=14
BPF:BPF:offset at call: 1915394
BPF:offset too small, choosingbase_btf: 1915397BPF:offset after
base_btf: 1915394BPF: ce type_id=49 bits_offset=0
BPF:BPF:offset at call: 1915403
BPF:offset afterbase_btf: 6BPF: nfig type_id=49 bits_offset=64
BPF:BPF:offset at call: 1915412
BPF:offset afterbase_btf: 15BPF: rdir type_id=49 bits_offset=128
BPF:BPF:offset at call: 768428
BPF:offset too small, choosingbase_btf: 1915397BPF:offset after
base_btf: 768428BPF:
_diotype_id=56 bits_offset=192
BPF:BPF:offset at call: 1915420
BPF:offset afterbase_btf: 23BPF: erdir type_id=56 bits_offset=200
BPF:BPF:offset at call: 1915433
BPF:offset afterbase_btf: 36BPF:first char wrong - 0
BPF: type_id=56 bits_offset=208
BPF:
BPF:Invalid name STRUCT MEMBER (name offset 1915433) BPF:failed to validate module [overlay] BTF: -22
---->8----
also note how it's only after a few botched entries that a check actually trips up - not sure what the impliciations for crafted BTF info are, but might be worthy a closer look by someone more knowledgable as well..
Quoted 4 lines
it seems to me this can be solved on the distro/user side by tracking vmlinux BTF infos as part of the ABI tracking (how stable are those compared to the existing interfaces making up the kernel <-> module ABI/Module.symvers? does this effectively mean bumping ABI for any
Yes, I think so, unfortunately. Any change in order of types emitted by DWARF or any extra string might cause a big change in string offsets or type IDs.
Quoted 9 lines
change anyway?) or by disabling
CONFIG_DEBUG_INFO_BTF_MODULES.on the kernel/libbpf side it could maybe be solved by storing a hash of the
base_btfdata used to generate the split BTF-sections inside the modules, and skip BTF loading/validating if anotherbase_btfis currently loaded (so BTF is best-effort, if the booted kernel and the module are matching it works, if not module loading works but no BTF support). this might be a good safe-guard for split-BTF in general?
Quoted 4 lines
I'd appreciate input on how to proceed (we were recently hit by this in a downstream Debian derivative, and will disable BTF info for modules as an interim measure).
Yeah, I think "BTF is best-effort" is a necessity for the complicated setups you described above. It probably is good to keep strict behavior (BTF fails to load -> fail loading the module), but allow you to tune it with the Kconfig option. Most modules probably would be just fine without its BTF loaded successfully.
As for the check sums, this would add another layer of protection, so I think storing base vmlinux BTF checksum in some special ELF section in module' ELF (e.g. ".BTF.base_checksum" or something along those lines) is also a good idea and would be easy to implement. Insmod and related tooling might even provide a validation check that vmlinux BTF matches module's expectation, if necessary.
Quoted 4 lines
thanks!
0: e.g., Debian's: https://salsa.debian.org/kernel-team/linux/-/blob/master/debian/bin/abiupdate.py
Re: [PATCH v4 bpf-next 3/5] kbuild: build kernel module BTFs if BTF is enabled and pahole supports it
In reply to Andrii NakryikoQuoted 10 lines · Andrii Nakryiko
lib/Kconfig.debug +2
-0@@ -274,6 +274,15 @@ config DEBUG_INFO_BTF Turning this on expects presence of pahole tool, which will convert DWARF type info into equivalent deduplicated BTF type info. +config PAHOLE_HAS_SPLIT_BTF + def_bool $(success, test `$(PAHOLE) --version | sed -E 's/v([0-9]+)\.([0-9]+)/\1\2/'` -ge "119")
lib/Kconfig.debug config DEBUG_INFO_BTF · line 288
Quoted 10 lines
lib/Kconfig.debug +7
-0+ +config DEBUG_INFO_BTF_MODULES + def_bool y + depends on DEBUG_INFO_BTF && MODULES && PAHOLE_HAS_SPLIT_BTF + help + Generate compact split BTF type information for kernel modules. + config GDB_SCRIPTS bool "Provide GDB scripts for kernel debugging" help
Thank you for adding a config option for this feature vs bumping the required pahole version in scripts/link-vmlinux.sh. This is a much more friendly way of handling kernel features that require support from build tools.
Andrii Nakryiko
maintainerPATCH v4 4/5 bpf: load and verify kernel module BTFs
In reply to Andrii NakryikoAdd kernel module listener that will load/validate and unload module BTF. Module BTFs gets ID generated for them, which makes it possible to iterate them with existing BTF iteration API. They are given their respective module's names, which will get reported through GET_OBJ_INFO API. They are also marked as in-kernel BTFs for tooling to distinguish them from user-provided BTFs.
Also, similarly to vmlinux BTF, kernel module BTFs are exposed through sysfs as /sys/kernel/btf/<module-name>. This is convenient for user-space tools to inspect module BTF contents and dump their types with existing tools:
[vmuser@archvm bpf]$ ls -la /sys/kernel/btf
total 0
drwxr-xr-x 2 root root 0 Nov 4 19:46 . drwxr-xr-x 13 root root 0 Nov 4 19:46 ..
...
-r--r--r-- 1 root root 888 Nov 4 19:46 irqbypass -r--r--r-- 1 root root 100225 Nov 4 19:46 kvm -r--r--r-- 1 root root 35401 Nov 4 19:46 kvm_intel -r--r--r-- 1 root root 120 Nov 4 19:46 pcspkr -r--r--r-- 1 root root 399 Nov 4 19:46 serio_raw -r--r--r-- 1 root root 4094095 Nov 4 19:46 vmlinux
--- Documentation/ABI/testing/sysfs-kernel-btf | 8 + include/linux/bpf.h | 2 + include/linux/module.h | 4 + kernel/bpf/btf.c | 194 +++++++++++++++++++++ kernel/bpf/sysfs_btf.c | 2 +- kernel/module.c | 32 ++++ 6 files changed, 241 insertions(+), 1 deletion(-)
Documentation/ABI/testing/sysfs-kernel-btf +8 -0
@@ -15,3 +15,11 @@ Description: information with description of all internal kernel types. See Documentation/bpf/btf.rst for detailed description of format itself. + +What: /sys/kernel/btf/<module-name> +Date: Nov 2020 +KernelVersion: 5.11 +Contact: [email protected] +Description: + Read-only binary attribute exposing kernel module's BTF type + information as an add-on to the kernel's BTF (/sys/kernel/btf/vmlinux).
include/linux/bpf.h +2 -0
@@ -36,9 +36,11 @@ struct seq_operations;struct bpf_iter_aux_info ;struct bpf_local_storage ;struct bpf_local_storage_map ; +struct kobject ;extern struct idr btf_idr;extern spinlock_t btf_idr_lock; +extern struct kobject *btf_kobj;typedef int (*bpf_iter_init_seq_priv_t )(void *private_data,struct bpf_iter_aux_info *aux);
include/linux/module.h +4 -0
@@ -475,6 +475,10 @@ struct module {unsigned int num_bpf_raw_events;struct bpf_raw_event_map *bpf_raw_events;#endif +#ifdef CONFIG_DEBUG_INFO_BTF_MODULES +unsigned int btf_data_size; +void *btf_data; +#endif #ifdef CONFIG_JUMP_LABEL struct jump_entry *jump_entries;unsigned int num_jump_entries;
kernel/bpf/btf.c +194 -0
209 lines not on this page · raw message
kernel/bpf/sysfs_btf.c +1 -1
@@ -26,7 +26,7 @@ static struct bin_attribute bin_attr_btf_vmlinux __ro_after_init = { .read = btf_vmlinux_read, }; -static struct kobject *btf_kobj; +struct kobject *btf_kobj;static int __init btf_vmlinux_init (void ) {
kernel/module.c +32 -0
@@ -380,6 +380,35 @@ static void *section_objs(const struct load_info *info,return (void *)info->sechdrs[sec].sh_addr; } +/* Find a module section: 0 means not found. Ignores SHF_ALLOC flag. */ +static unsigned int find_any_sec (const struct load_info *info,const char *name) +{ +unsigned int i; + +for (i =1 ; i < info->hdr->e_shnum; i++) { +Elf_Shdr *shdr = &info->sechdrs[i]; +if (strcmp (info->secstrings + shdr->sh_name, name) ==0 ) +return i; + } +return 0 ; +} + +/* +* Find a module section, or NULL. Fill in number of "objects" in section. +* Ignores SHF_ALLOC flag. +*/ +static __maybe_unused void *any_section_objs (const struct load_info *info, +const char *name, +size_t object_size, +unsigned int *num) +{ +unsigned int sec =find_any_sec (info, name); + +/* Section 0 has sh_addr 0 and sh_size 0. */ + *num = info->sechdrs[sec].sh_size / object_size; +return (void *)info->sechdrs[sec].sh_addr; +} +/* Provided by the linker */ extern const struct kernel_symbol __start___ksymtab[];extern const struct kernel_symbol __stop___ksymtab[];
@@ -3250,6 +3279,9 @@ static int find_module_sections(struct module *mod, struct load_info *info)sizeof (*mod->bpf_raw_events), &mod->num_bpf_raw_events);#endif +#ifdef CONFIG_DEBUG_INFO_BTF_MODULES + mod->btf_data =any_section_objs (info,".BTF" ,1 , &mod->btf_data_size); +#endif #ifdef CONFIG_JUMP_LABEL mod->jump_entries = section_objs(info, "__jump_table", sizeof(*mod->jump_entries),
Signature · 1 line
2.24.1
kernel test robot
build botRe: [PATCH v4 bpf-next 4/5] bpf: load and verify kernel module BTFs
I love your patch! Perhaps something to improve:
Hi Andrii,
I love your patch! Perhaps something to improve:
[auto build test WARNING on bpf-next/master]
url: https://github.com/0day-ci/linux/commits/Andrii-Nakryiko/Integrate-kernel-module-BTF-support/20201110-095309 base: https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git master config: mips-randconfig-r002-20201110 (attached as .config) compiler: clang version 12.0.0 (https://github.com/llvm/llvm-project 4d81c8adb6ed9840257f6cb6b93f60856d422a15)
reproduce (this is a W=1 build):
wget https://raw.githubusercontent.com/intel/lkp-tests/master/sbin/make.cross -O ~/bin/make.cross chmod +x ~/bin/make.cross
# install mips cross compiling tool for clang build
# apt-get install binutils-mips-linux-gnu
# https://github.com/0day-ci/linux/commit/dcd763b7808fdc01ebf70bbe07ba92388df4d20d
git remote add linux-review https://github.com/0day-ci/linux
git fetch --no-tags linux-review Andrii-Nakryiko/Integrate-kernel-module-BTF-support/20201110-095309 git checkout dcd763b7808fdc01ebf70bbe07ba92388df4d20d
# save the attached .config to linux build tree
COMPILER_INSTALL_PATH=$HOME/0day COMPILER=clang make.cross ARCH=mips
If you fix the issue, kindly add following tag as appropriate
All warnings (new ones prefixed by >>):
>> kernel/bpf/btf.c:4481:20: warning: unused function 'btf_parse_module'
static struct btf char const void unsigned int data_size)
^
fatal error: error in backend: Nested variants found in inline asm string: ' .set push
.set mips64r2
.if ( 0x00 ) != -1)) 0x00 ) != -1)) : ($( static struct ftrace_branch_data __attribute__((__aligned__(4))) __attribute__((__section__("_ftrace_branch"))) __if_trace = $( .func = __func__, .file = "arch/mips/include/asm/atomic.h", .line = 154, $); 0x00 ) != -1)) : $))) ) && ( 0 ); .set push; .set mips64r2; .rept 1; sync 0x00; .endr; .set pop; .else; ; .endif 1: ll $1, $2 # atomic_fetch_sub
subu $0, $1, $3
sc $0, $2
beqz $0, 1b
.set pop
move $0, $1
'
clang-12: error: clang frontend command failed with exit code 70 (use -v to see invocation)
clang version 12.0.0 (git://gitmirror/llvm_project 874b0a0b9db93f5d3350ffe6b5efda2d908415d0)
Target: mipsel-unknown-linux-gnu
Thread model: posix
InstalledDir: /opt/cross/clang-874b0a0b9d/bin
clang-12: note: diagnostic msg:
Makefile arch drivers include kernel mm scripts source usr
vim +/btf_parse_module +4481 kernel/bpf/btf.c
4480
> 4481 static struct btf *btf_parse_module(const char *module_name, const void *data, unsigned int data_size)
4482 {
4483 struct btf_verifier_env *env = NULL;
4484 struct bpf_verifier_log *log;
4485 struct btf *btf = NULL, *base_btf;
4486 int err;
4487
4488 base_btf = bpf_get_btf_vmlinux();
4489 if (IS_ERR(base_btf))
4490 return base_btf;
4491 if (!base_btf)
4492 return ERR_PTR(-EINVAL);
4493
4494 env = kzalloc(sizeof(*env), GFP_KERNEL | __GFP_NOWARN);
4495 if (!env)
4496 return ERR_PTR(-ENOMEM);
4497
4498 log = &env->log;
4499 log->level = BPF_LOG_KERNEL;
4500
4501 btf = kzalloc(sizeof(*btf), GFP_KERNEL | __GFP_NOWARN);
4502 if (!btf) {
4503 err = -ENOMEM;
4504 goto errout;
4505 }
4506 env->btf = btf;
4507
4508 btf->base_btf = base_btf;
4509 btf->start_id = base_btf->nr_types;
4510 btf->start_str_off = base_btf->hdr.str_len;
4511 btf->kernel_btf = true;
4512 snprintf(btf->name, sizeof(btf->name), "%s", module_name);
4513
4514 btf->data = kvmalloc(data_size, GFP_KERNEL | __GFP_NOWARN);
4515 if (!btf->data) {
4516 err = -ENOMEM;
4517 goto errout;
4518 }
4519 memcpy(btf->data, data, data_size);
4520 btf->data_size = data_size;
4521
4522 err = btf_parse_hdr(env);
4523 if (err)
4524 goto errout;
4525
4526 btf->nohdr_data = btf->data + btf->hdr.hdr_len;
4527
4528 err = btf_parse_str_sec(env);
4529 if (err)
4530 goto errout;
4531
4532 err = btf_check_all_metas(env);
4533 if (err)
4534 goto errout;
4535
4536 btf_verifier_env_free(env);
4537 refcount_set(&btf->refcnt, 1);
4538 return btf;
4539
4540 errout:
4541 btf_verifier_env_free(env);
4542 if (btf) {
4543 kvfree(btf->data);
4544 kvfree(btf->types);
4545 kfree(btf);
4546 }
4547 return ERR_PTR(err);
4548 }
4549
---
0-DAY CI Kernel Test Service, Intel Corporation
https://lists.01.org/hyperkitty/list/kbuild-all(a)lists.01.org
Re: [PATCH v4 bpf-next 4/5] bpf: load and verify kernel module BTFs
In reply to Andrii Nakryiko+++ Andrii Nakryiko [09/11/20 17:19 -0800]:
[snipped]
Quoted 27 lines
kernel/module.c +19
-0@@ -380,6 +380,35 @@ static void *section_objs(const struct load_info *info,return (void *)info->sechdrs[sec].sh_addr; } +/* Find a module section: 0 means not found. Ignores SHF_ALLOC flag. */ +static unsigned int find_any_sec (const struct load_info *info,const char *name) +{ +unsigned int i; + +for (i =1 ; i < info->hdr->e_shnum; i++) { +Elf_Shdr *shdr = &info->sechdrs[i]; +if (strcmp (info->secstrings + shdr->sh_name, name) ==0 ) +return i; + } +return 0 ; +} + +/* +* Find a module section, or NULL. Fill in number of "objects" in section. +* Ignores SHF_ALLOC flag. +*/ +static __maybe_unused void *any_section_objs(const struct load_info *info, + const char *name,
kernel/module.c static void *section_objs(const struct load_info *info, · line 411
Quoted 10 lines
kernel/module.c +10
-0+size_t object_size, + unsigned int *num) +{ + unsigned int sec = find_any_sec(info, name); + +/* Section 0 has sh_addr 0 and sh_size 0. */ + *num = info->sechdrs[sec].sh_size / object_size; +return (void *)info->sechdrs[sec].sh_addr; +} +
Hm, I see this patchset has already been applied to bpf-next, but I guess that doesn't preclude any follow-up patches :-)
I am not a huge fan of the code duplication here, and also the fact that they're only called in one place. any_section_objs() and find_any_sec() are pretty much identical to section_objs() and find_sec(), other than the fact the former drops the SHF_ALLOC check.
Moreover, since it appears that the ".BTF" section is not marked SHF_ALLOC, I think this will leave mod->btf_data as a dangling pointer after the module is done loading and the module's load_info has been deallocated, since SHF_ALLOC sections are not allocated nor copied to the module's final location in memory.
Why not simply mark the ".BTF" section in the module SHF_ALLOC? We already do some sh_flags rewriting in rewrite_section_headers(). Then the module loader knows to keep the section in memory and you can use section_objs(). And since the .BTF section stays in module memory, that might save you the memcpy() to btf->data in btf_parse_module() (unless that is still needed for some reason).
Thanks,
Jessica
Quoted 16 lines
kernel/module.c +3
-0/* Provided by the linker */ extern const struct kernel_symbol __start___ksymtab[];extern const struct kernel_symbol __stop___ksymtab[];@@ -3250,6 +3279,9 @@ static int find_module_sections(struct module *mod, struct load_info *info)sizeof (*mod->bpf_raw_events), &mod->num_bpf_raw_events);#endif +#ifdef CONFIG_DEBUG_INFO_BTF_MODULES + mod->btf_data =any_section_objs (info,".BTF" ,1 , &mod->btf_data_size); +#endif #ifdef CONFIG_JUMP_LABEL mod->jump_entries = section_objs(info, "__jump_table", sizeof(*mod->jump_entries),Signature · 1 line
2.24.1
Re: [PATCH v4 bpf-next 4/5] bpf: load and verify kernel module BTFs
Quoted 35 lines · Jessica Yu
+++ Andrii Nakryiko [09/11/20 17:19 -0800]:
[snipped]Quoted 32 lines
kernel/module.c +24
-0@@ -380,6 +380,35 @@ static void *section_objs(const struct load_info *info,return (void *)info->sechdrs[sec].sh_addr; } +/* Find a module section: 0 means not found. Ignores SHF_ALLOC flag. */ +static unsigned int find_any_sec (const struct load_info *info,const char *name) +{ +unsigned int i; + +for (i =1 ; i < info->hdr->e_shnum; i++) { +Elf_Shdr *shdr = &info->sechdrs[i]; +if (strcmp (info->secstrings + shdr->sh_name, name) ==0 ) +return i; + } +return 0 ; +} + +/* +* Find a module section, or NULL. Fill in number of "objects" in section. +* Ignores SHF_ALLOC flag. +*/ +static __maybe_unused void *any_section_objs (const struct load_info *info, +const char *name, +size_t object_size, +unsigned int *num) +{ +unsigned int sec =find_any_sec (info, name); +
Quoted 8 lines
kernel/module.c static void *section_objs(const struct load_info *info, · line 411
Quoted 5 lines
kernel/module.c +5
-0+/* Section 0 has sh_addr 0 and sh_size 0. */ + *num = info->sechdrs[sec].sh_size / object_size; +return (void *)info->sechdrs[sec].sh_addr; +} +Hm, I see this patchset has already been applied to bpf-next, but I guess that doesn't preclude any follow-up patches :-)
Of course!
Quoted 5 lines
I am not a huge fan of the code duplication here, and also the fact that they're only called in one place.
any_section_objs()andfind_any_sec()are pretty much identical tosection_objs()andfind_sec(), other than the fact the former drops theSHF_ALLOCcheck.
Right, but the alternative was to add a new flag to existing section_objs() and find_sec() functions, which would cause much more code churn for no good reason (besides saving some trivial code duplication). And those true/false flags are harder to read in code anyways.
Quoted 6 lines
Moreover, since it appears that the ".BTF" section is not marked
SHF_ALLOC, I think this will leavemod->btf_dataas a dangling pointer after the module is done loading and the module'sload_infohas been deallocated, sinceSHF_ALLOCsections are not allocated nor copied to the module's final location in memory.
I can make sure that we also reset the btf_data pointer back to NULL, if that's a big concern.
Quoted 7 lines
Why not simply mark the ".BTF" section in the module
SHF_ALLOC? We already do somesh_flagsrewriting inrewrite_section_headers(). Then the module loader knows to keep the section in memory and you can usesection_objs(). And since the .BTF section stays in module memory, that might save you thememcpy()tobtf->datainbtf_parse_module()(unless that is still needed for some reason).
Wasn't aware about rewrite_section_headers() manipulations. Are you suggesting to just add SHF_ALLOC there for the .BTF section from the kernel side? I guess that would work, but won't avoid memory copy (so actually would waste kernel memory, if I understand correctly). The reason being that the module's BTF is registered as an independently ref-counted BTF object, which could be held past the kernel module being unloaded. So I can't directly reference module's .BTF data anyways.
Also, marking .BTF with SHF_ALLOC with pahole or objcopy tool actually might generate warnings because SHF_ALLOC sections need to be allocated to data segments, which neither of those tools know how to do, it requires a linker support. We do that for vmlinux with extra linker script logic, but for kernel modules we don't have and probably don't want to do that.
So in the end, the cleanest approach still seems like not doing SHF_ALLOC but allowing "capturing" .BTF data with an extra helper.
Quoted 21 lines
Thanks,
Jessica
Quoted 16 lines
kernel/module.c +3
-1/* Provided by the linker */ extern const struct kernel_symbol __start___ksymtab[];extern const struct kernel_symbol __stop___ksymtab[];@@ -3250,6 +3279,9 @@ static int find_module_sections(struct module *mod, struct load_info *info)sizeof (*mod->bpf_raw_events), &mod->num_bpf_raw_events);#endif +#ifdef CONFIG_DEBUG_INFO_BTF_MODULES + mod->btf_data =any_section_objs (info,".BTF" ,1 , &mod->btf_data_size); +#endif #ifdef CONFIG_JUMP_LABEL mod->jump_entries = section_objs(info, "__jump_table", sizeof(*mod->jump_entries), --2.24.1
Re: [PATCH v4 bpf-next 4/5] bpf: load and verify kernel module BTFs
+++ Andrii Nakryiko [11/11/20 12:11 -0800]:
Quoted 53 lines · Jessica Yu
Quoted 35 lines
+++ Andrii Nakryiko [09/11/20 17:19 -0800]:
[snipped]Quoted 32 lines
kernel/module.c +24
-0@@ -380,6 +380,35 @@ static void *section_objs(const struct load_info *info,return (void *)info->sechdrs[sec].sh_addr; } +/* Find a module section: 0 means not found. Ignores SHF_ALLOC flag. */ +static unsigned int find_any_sec (const struct load_info *info,const char *name) +{ +unsigned int i; + +for (i =1 ; i < info->hdr->e_shnum; i++) { +Elf_Shdr *shdr = &info->sechdrs[i]; +if (strcmp (info->secstrings + shdr->sh_name, name) ==0 ) +return i; + } +return 0 ; +} + +/* +* Find a module section, or NULL. Fill in number of "objects" in section. +* Ignores SHF_ALLOC flag. +*/ +static __maybe_unused void *any_section_objs (const struct load_info *info, +const char *name, +size_t object_size, +unsigned int *num) +{ +unsigned int sec =find_any_sec (info, name); +Quoted 8 lines
kernel/module.c static void *section_objs(const struct load_info *info, · line 411
Quoted 5 lines
kernel/module.c +5
-0+/* Section 0 has sh_addr 0 and sh_size 0. */ + *num = info->sechdrs[sec].sh_size / object_size; +return (void *)info->sechdrs[sec].sh_addr; +} +Hm, I see this patchset has already been applied to bpf-next, but I guess that doesn't preclude any follow-up patches :-)
Of course!
Quoted 5 lines
I am not a huge fan of the code duplication here, and also the fact that they're only called in one place.
any_section_objs()andfind_any_sec()are pretty much identical tosection_objs()andfind_sec(), other than the fact the former drops theSHF_ALLOCcheck.
Quoted 5 lines
Right, but the alternative was to add a new flag to existing
section_objs()andfind_sec()functions, which would cause much more code churn for no good reason (besides saving some trivial code duplication). And those true/false flags are harder to read in code anyways.
That's true, all fair points. I thought there was the possibility to avoid the code duplication if .BTF were also set to SHF_ALLOC, but I see for reasons you explained below it is more trouble than it's worth.
Quoted 9 lines
Quoted 6 lines
Moreover, since it appears that the ".BTF" section is not marked
SHF_ALLOC, I think this will leavemod->btf_dataas a dangling pointer after the module is done loading and the module'sload_infohas been deallocated, sinceSHF_ALLOCsections are not allocated nor copied to the module's final location in memory.I can make sure that we also reset the
btf_datapointer back to NULL, if that's a big concern.
It's not a terribly huge concern, since mod->btf_data is only accessed in the btf coming notifier at the moment, but it's probably best to at least not advertise it as a valid pointer anymore after the module is done loading. We do some pointer and section size cleanup at the end of do_init_module() for sections that are deallocated at the end of module load (starting where init_layout.base is reset to NULL), we could just tack on mod->btf_data = NULL there as well.
Quoted 8 lines
Quoted 7 lines
Why not simply mark the ".BTF" section in the module
SHF_ALLOC? We already do somesh_flagsrewriting inrewrite_section_headers(). Then the module loader knows to keep the section in memory and you can usesection_objs(). And since the .BTF section stays in module memory, that might save you thememcpy()tobtf->datainbtf_parse_module()(unless that is still needed for some reason).
Quoted 8 lines
Wasn't aware about
rewrite_section_headers()manipulations. Are you suggesting to just addSHF_ALLOCthere for the .BTF section from the kernel side? I guess that would work, but won't avoid memory copy (so actually would waste kernel memory, if I understand correctly). The reason being that the module's BTF is registered as an independently ref-counted BTF object, which could be held past the kernel module being unloaded. So I can't directly reference module's .BTF data anyways.
Ah OK, I was not aware that the section could be held past the module being unloaded. Then yeah, it would be a memory waste to keep them in memory if they are being memcpy'd anyway. Thanks for clarifying!
Jessica
Re: [PATCH v4 bpf-next 4/5] bpf: load and verify kernel module BTFs
Quoted 75 lines · Jessica Yu
+++ Andrii Nakryiko [11/11/20 12:11 -0800]:
Quoted 53 lines · Jessica Yu
Quoted 35 lines
+++ Andrii Nakryiko [09/11/20 17:19 -0800]:
[snipped]Quoted 32 lines
kernel/module.c +24
-0@@ -380,6 +380,35 @@ static void *section_objs(const struct load_info *info,return (void *)info->sechdrs[sec].sh_addr; } +/* Find a module section: 0 means not found. Ignores SHF_ALLOC flag. */ +static unsigned int find_any_sec (const struct load_info *info,const char *name) +{ +unsigned int i; + +for (i =1 ; i < info->hdr->e_shnum; i++) { +Elf_Shdr *shdr = &info->sechdrs[i]; +if (strcmp (info->secstrings + shdr->sh_name, name) ==0 ) +return i; + } +return 0 ; +} + +/* +* Find a module section, or NULL. Fill in number of "objects" in section. +* Ignores SHF_ALLOC flag. +*/ +static __maybe_unused void *any_section_objs (const struct load_info *info, +const char *name, +size_t object_size, +unsigned int *num) +{ +unsigned int sec =find_any_sec (info, name); +Quoted 8 lines
kernel/module.c static void *section_objs(const struct load_info *info, · line 411
Quoted 5 lines
kernel/module.c +5
-0+/* Section 0 has sh_addr 0 and sh_size 0. */ + *num = info->sechdrs[sec].sh_size / object_size; +return (void *)info->sechdrs[sec].sh_addr; +} +Hm, I see this patchset has already been applied to bpf-next, but I guess that doesn't preclude any follow-up patches :-)
Of course!
Quoted 5 lines
I am not a huge fan of the code duplication here, and also the fact that they're only called in one place.
any_section_objs()andfind_any_sec()are pretty much identical tosection_objs()andfind_sec(), other than the fact the former drops theSHF_ALLOCcheck.Quoted 5 lines
Right, but the alternative was to add a new flag to existing
section_objs()andfind_sec()functions, which would cause much more code churn for no good reason (besides saving some trivial code duplication). And those true/false flags are harder to read in code anyways.That's true, all fair points. I thought there was the possibility to avoid the code duplication if .BTF were also set to
SHF_ALLOC, but I see for reasons you explained below it is more trouble than it's worth.Quoted 9 lines
Quoted 6 lines
Moreover, since it appears that the ".BTF" section is not marked
SHF_ALLOC, I think this will leavemod->btf_dataas a dangling pointer after the module is done loading and the module'sload_infohas been deallocated, sinceSHF_ALLOCsections are not allocated nor copied to the module's final location in memory.I can make sure that we also reset the
btf_datapointer back to NULL, if that's a big concern.
Quoted 7 lines
It's not a terribly huge concern, since
mod->btf_datais only accessed in the btf coming notifier at the moment, but it's probably best to at least not advertise it as a valid pointer anymore after the module is done loading. We do some pointer and section size cleanup at the end ofdo_init_module()for sections that are deallocated at the end of module load (starting where init_layout.base is reset to NULL), we could just tack onmod->btf_data= NULL there as well.
Sounds good, I'll send a follow up patch. Thanks!
Quoted 23 lines
Quoted 9 lines
Quoted 8 lines
Why not simply mark the ".BTF" section in the module
SHF_ALLOC? We already do somesh_flagsrewriting inrewrite_section_headers(). Then the module loader knows to keep the section in memory and you can usesection_objs(). And since the .BTF section stays in module memory, that might save you thememcpy()tobtf->datainbtf_parse_module()(unless that is still needed for some reason).Quoted 8 lines
Wasn't aware about
rewrite_section_headers()manipulations. Are you suggesting to just addSHF_ALLOCthere for the .BTF section from the kernel side? I guess that would work, but won't avoid memory copy (so actually would waste kernel memory, if I understand correctly). The reason being that the module's BTF is registered as an independently ref-counted BTF object, which could be held past the kernel module being unloaded. So I can't directly reference module's .BTF data anyways.Ah OK, I was not aware that the section could be held past the module being unloaded. Then yeah, it would be a memory waste to keep them in memory if they are being memcpy'd anyway. Thanks for clarifying!
Jessica
Andrii Nakryiko
maintainerPATCH v4 5/5 tools/bpftool: add support for in-kernel and named BTF in `btf show`
In reply to Andrii NakryikoDisplay vmlinux BTF name and kernel module names when listing available BTFs on the system.
In human-readable output mode, module BTFs are reported with "name [module-name]", while vmlinux BTF will be reported as "name [vmlinux]". Square brackets are added by bpftool and follow kernel convention when displaying modules in human-readable text outputs.
[vmuser@archvm bpf]$ sudo ../../../bpf/bpftool/bpftool btf s 1: name [vmlinux] size 4082281B
6: size 2365B prog_ids 8,6 map_ids 3
7: name [button] size 46895B
8: name [pcspkr] size 42328B
9: name [serio_raw] size 39375B
10: name [floppy] size 57185B
11: name [i2c_core] size 76186B
12: name [crc32c_intel] size 16036B
13: name [i2c_piix4] size 50497B
14: name [irqbypass] size 14124B
15: name [kvm] size 197985B
16: name [kvm_intel] size 123564B
17: name [cryptd] size 42466B
18: name [crypto_simd] size 17187B
19: name [glue_helper] size 39205B
20: name [aesni_intel] size 41034B
25: size 36150B
pids bpftool(2519)In JSON mode, two fields (boolean "kernel" and string "name") are reported for each BTF object. vmlinux BTF is reported with name "vmlinux" (kernel itself returns and empty name for vmlinux BTF).
[vmuser@archvm bpf]$ sudo ../../../bpf/bpftool/bpftool btf s -jp
[{
"id": 1,
"size": 4082281,
"prog_ids": [],
"map_ids": [],
"kernel": true,
"name": "vmlinux"
},{
"id": 6,
"size": 2365,
"prog_ids": [8,6
],
"map_ids": [3
],
"kernel": false
},{
"id": 7,
"size": 46895,
"prog_ids": [],
"map_ids": [],
"kernel": true,
"name": "button"
},{...
--- tools/bpf/bpftool/btf.c | 28 +++++++++++++++++++++++++++- 1 file changed, 27 insertions(+), 1 deletion(-)
tools/bpf/bpftool/btf.c +27 -1
@@ -742,9 +742,14 @@ show_btf_plain(struct bpf_btf_info *info, int fd, struct btf_attach_table *btf_map_table) {struct btf_attach_point *obj; +const char *name =u64_to_ptr (info->name);int n;printf ("%u: " , info->id); +if (info->kernel_btf) +printf ("name [%s] " , name); +else if (name && name[0 ]) +printf ("name %s " , name);printf ("size %uB" , info->btf_size); n =0 ;
@@ -771,6 +776,7 @@ show_btf_json(struct bpf_btf_info *info, int fd, struct btf_attach_table *btf_map_table) {struct btf_attach_point *obj; +const char *name =u64_to_ptr (info->name);jsonw_start_object (json_wtr);/* btf object */ jsonw_uint_field (json_wtr,"id" , info->id);
@@ -796,6 +802,11 @@ show_btf_json(struct bpf_btf_info *info, int fd,emit_obj_refs_json (&refs_table, info->id, json_wtr);/* pids */ +jsonw_bool_field (json_wtr,"kernel" , info->kernel_btf); + +if (name && name[0 ]) +jsonw_string_field (json_wtr,"name" , name); +jsonw_end_object (json_wtr);/* btf object */ }
@@ -803,15 +814,30 @@ static int show_btf(int fd, struct btf_attach_table *btf_prog_table, struct btf_attach_table *btf_map_table) { - struct bpf_btf_info info = {}; + struct bpf_btf_info info;__u32 len =sizeof (info); +char name[64 ];int err; +memset (&info,0 ,sizeof (info)); err =bpf_obj_get_info_by_fd (fd, &info, &len);if (err) {p_err ("can't get BTF object info: %s" ,strerror (errno));return -1 ; } +/* if kernel support emitting BTF object name, pass name pointer */ +if (info.name_len) { +memset (&info,0 ,sizeof (info)); + info.name_len =sizeof (name); + info.name =ptr_to_u64 (name); + len =sizeof (info); + + err =bpf_obj_get_info_by_fd (fd, &info, &len); +if (err) { +p_err ("can't get BTF object info: %s" ,strerror (errno)); +return -1 ; + } + }if (json_output)show_btf_json (&info, fd, btf_prog_table, btf_map_table);
Signature · 1 line
2.24.1
Re: [PATCH v4 bpf-next 5/5] tools/bpftool: add support for in-kernel and named BTF in `btf show`
Quoted 3 lines · Andrii Nakryiko
[...]
Quoted 4 lines
...
Tested-by Alan Maguire <[email protected]>Signed-off-by Andrii Nakryiko <[email protected]>
With one nit:
Quoted 10 lines
---
tools/bpf/bpftool/btf.c| 28 +++++++++++++++++++++++++++-1 file changed, 27 insertions(+), 1 deletion(-)tools/bpf/bpftool/btf.c +0
-0@@ -742,9 +742,14 @@ show_btf_plain(struct bpf_btf_info *info, int fd, struct btf_attach_table *btf_map_table)
tools/bpf/bpftool/btf.c show_btf_plain(struct bpf_btf_info *info, int fd, · line 752
Quoted 10 lines
tools/bpf/bpftool/btf.c +5
-0{struct btf_attach_point *obj; +const char *name =u64_to_ptr (info->name);int n;printf ("%u: " , info->id); +if (info->kernel_btf) +printf ("name [%s] " , name); +else if (name && name[0 ]) +printf ("name %s " , name);
Maybe explicitly say "name <anonymous>" for btf without a name? I think it will benefit plain output.
Quoted 3 lines
tools/bpf/bpftool/btf.c +0
-0printf ("size %uB" , info->btf_size); n =0 ;
[...]
Re: [PATCH v4 bpf-next 5/5] tools/bpftool: add support for in-kernel and named BTF in `btf show`
Quoted 31 lines · Song Liu
Quoted 4 lines · Andrii Nakryiko
[...]
Quoted 4 lines
...
Tested-by Alan Maguire <[email protected]>Signed-off-by Andrii Nakryiko <[email protected]>Acked-by Song Liu <[email protected]>With one nit:
Quoted 15 lines
---
tools/bpf/bpftool/btf.c| 28 +++++++++++++++++++++++++++-1 file changed, 27 insertions(+), 1 deletion(-)tools/bpf/bpftool/btf.c +1
-0@@ -742,9 +742,14 @@ show_btf_plain(struct bpf_btf_info *info, int fd, struct btf_attach_table *btf_map_table) {struct btf_attach_point *obj; +const char *name =u64_to_ptr (info->name);int n;
Quoted 8 lines
tools/bpf/bpftool/btf.c show_btf_plain(struct bpf_btf_info *info, int fd, · line 752
Quoted 5 lines
tools/bpf/bpftool/btf.c +4
-0printf ("%u: " , info->id); +if (info->kernel_btf) +printf ("name [%s] " , name); +else if (name && name[0 ]) +printf ("name %s " , name);Maybe explicitly say "name <anonymous>" for btf without a name? I think it will benefit plain output.
This patch set is already landed. But I can do a follow-up patch to add this.
Quoted 6 lines
Quoted 4 lines
tools/bpf/bpftool/btf.c +0
-0printf ("size %uB" , info->btf_size); n =0 ;[...]
Re: [PATCH v4 bpf-next 5/5] tools/bpftool: add support for in-kernel and named BTF in `btf show`
Quoted 45 lines · Andrii Nakryiko
Quoted 31 lines · Song Liu
Quoted 4 lines · Andrii Nakryiko
[...]
Quoted 4 lines
...
Tested-by Alan Maguire <[email protected]>Signed-off-by Andrii Nakryiko <[email protected]>Acked-by Song Liu <[email protected]>With one nit:
Quoted 15 lines
---
tools/bpf/bpftool/btf.c| 28 +++++++++++++++++++++++++++-1 file changed, 27 insertions(+), 1 deletion(-)tools/bpf/bpftool/btf.c +1
-0@@ -742,9 +742,14 @@ show_btf_plain(struct bpf_btf_info *info, int fd, struct btf_attach_table *btf_map_table) {struct btf_attach_point *obj; +const char *name =u64_to_ptr (info->name);int n;Quoted 8 lines
tools/bpf/bpftool/btf.c show_btf_plain(struct bpf_btf_info *info, int fd, · line 752
Quoted 5 lines
tools/bpf/bpftool/btf.c +4
-0printf ("%u: " , info->id); +if (info->kernel_btf) +printf ("name [%s] " , name); +else if (name && name[0 ]) +printf ("name %s " , name);Maybe explicitly say "name <anonymous>" for btf without a name? I think it will benefit plain output.
Quoted 1 line
This patch set is already landed. But I can do a follow-up patch to add this.
I realized this was applied soon after sending this. Yeah, a follow-up patch would be great.
Thanks,
Song