bpf: Do not allocate percpu memory at init stage
Yonghong Song
reviewerPATCH 1/1 bpf: Do not allocate percpu memory at init stage
Kirill Shutemov reported significant percpu memory increase after booting in 288-cpu VM ([1]) due to commit 41a5db8d8161 ("bpf: Add support for non-fix-size percpu mem allocation"). The percpu memory is increased from 111MB to 969MB. The number is from /proc/meminfo.
I tried to reproduce the issue with my local VM which at most supports upto 255 cpus. With 252 cpus, without the above commit, the percpu memory immediately after boot is 57MB while with the above commit the percpu memory is 231MB.
This is not good since so far percpu memory from bpf memory allocator is not widely used yet. Let us change pre-allocation in init stage to on-demand allocation when verifier detects there is a need of percpu memory for bpf program. With this change, percpu memory consumption after boot can be reduced signicantly.
[1] https://lore.kernel.org/lkml/[email protected]/
--- include/linux/bpf.h | 2 +- kernel/bpf/core.c | 8 +++----- kernel/bpf/verifier.c | 17 +++++++++++++++-- 3 files changed, 19 insertions(+), 8 deletions(-)
include/linux/bpf.h +1 -1
@@ -56,7 +56,7 @@ extern struct idr btf_idr;extern spinlock_t btf_idr_lock;extern struct kobject *btf_kobj;extern struct bpf_mem_alloc bpf_global_ma, bpf_global_percpu_ma; -extern bool bpf_global_ma_set, bpf_global_percpu_ma_set; +extern bool bpf_global_ma_set;typedef u64 (*bpf_callback_t )(u64 ,u64 ,u64 ,u64 ,u64 ); typedef int (*bpf_iter_init_seq_priv_t)(void *private_data,
kernel/bpf/core.c +3 -5
@@ -64,8 +64,8 @@#define OFF insn->off#define IMM insn->imm -struct bpf_mem_alloc bpf_global_ma, bpf_global_percpu_ma; -bool bpf_global_ma_set, bpf_global_percpu_ma_set; +struct bpf_mem_alloc bpf_global_ma; +bool bpf_global_ma_set;/* No hurry in this branch *
@@ -2934,9 +2934,7 @@ static int __init bpf_global_ma_init(void) ret =bpf_mem_alloc_init (&bpf_global_ma,0 , false); bpf_global_ma_set = !ret; - ret =bpf_mem_alloc_init (&bpf_global_percpu_ma,0 , true); - bpf_global_percpu_ma_set = !ret; - return !bpf_global_ma_set || !bpf_global_percpu_ma_set; + return ret; }late_initcall (bpf_global_ma_init);#endif
kernel/bpf/verifier.c +15 -2
@@ -26,6 +26,7 @@#include <linux/poison.h> #include <linux/module.h> #include <linux/cpumask.h> +#include <linux/bpf_mem_alloc.h> #include <net/xdp.h> #include "disasm.h"
@@ -41,6 +42,9 @@ static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {#undef BPF_LINK_TYPE }; +struct bpf_mem_alloc bpf_global_percpu_ma; +static bool bpf_global_percpu_ma_set; +/* bpf_check() is a static code analyzer that walks eBPF program * instruction by instruction and updates register/stack state. * All paths of conditional branches are analyzed until 'bpf_exit' insn.
@@ -12074,8 +12078,17 @@ static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,if (meta.func_id == special_kfunc_list[KF_bpf_obj_new_impl] && !bpf_global_ma_set)return -ENOMEM ; -if (meta.func_id == special_kfunc_list[KF_bpf_percpu_obj_new_impl] && !bpf_global_percpu_ma_set) -return -ENOMEM ; +if (meta.func_id == special_kfunc_list[KF_bpf_percpu_obj_new_impl]) { +mutex_lock (&bpf_verifier_lock); +if (!bpf_global_percpu_ma_set) { + err =bpf_mem_alloc_init (&bpf_global_percpu_ma,0 , true); +if (!err) + bpf_global_percpu_ma_set = true; + } +mutex_unlock (&bpf_verifier_lock); +if (err) +return err; + }if (((u64 )(u32 )meta.arg_constant.value) != meta.arg_constant.value) {verbose (env,"local type ID argument must be in range [0, U32_MAX]\n" );
Signature · 1 line
2.34.1
Quoted 1 line · Yonghong Song
Kirill Shutemov reported significant percpu memory increase after booting
s/memory increase/memory consumption increase/ ?
Quoted 15 lines
in 288-cpu VM ([1]) due to commit
41a5db8d8161("bpf: Add support for non-fix-size percpu mem allocation"). The percpu memory is increased from 111MB to 969MB. The number is from/proc/meminfo.I tried to reproduce the issue with my local VM which at most supports upto 255 cpus. With 252 cpus, without the above commit, the percpu memory immediately after boot is 57MB while with the above commit the percpu memory is 231MB.
This is not good since so far percpu memory from bpf memory allocator is not widely used yet. Let us change pre-allocation in init stage to on-demand allocation when verifier detects there is a need of percpu memory for bpf program. With this change, percpu memory consumption after boot can be reduced signicantly.
Quoted 4 lines
[1] https://lore.kernel.org/lkml/[email protected]/
Fixes 41a5db8d8161 ("bpf: Add support for non-fix-size percpu mem allocation")Cc Kirill A. Shutemov <[email protected]>
Signature · 1 line
Kiryl Shutsemau / Kirill A. Shutemov
Quoted 89 lines · Yonghong Song
Kirill Shutemov reported significant percpu memory increase after booting in 288-cpu VM ([1]) due to commit
41a5db8d8161("bpf: Add support for non-fix-size percpu mem allocation"). The percpu memory is increased from 111MB to 969MB. The number is from/proc/meminfo.I tried to reproduce the issue with my local VM which at most supports upto 255 cpus. With 252 cpus, without the above commit, the percpu memory immediately after boot is 57MB while with the above commit the percpu memory is 231MB.
This is not good since so far percpu memory from bpf memory allocator is not widely used yet. Let us change pre-allocation in init stage to on-demand allocation when verifier detects there is a need of percpu memory for bpf program. With this change, percpu memory consumption after boot can be reduced signicantly.
[1] https://lore.kernel.org/lkml/[email protected]/
Fixes 41a5db8d8161 ("bpf: Add support for non-fix-size percpu mem allocation")Cc Kirill A. Shutemov <[email protected]>Signed-off-by Yonghong Song <[email protected]>--- include/linux/bpf.h | 2 +- kernel/bpf/core.c | 8 +++----- kernel/bpf/verifier.c | 17 +++++++++++++++-- 3 files changed, 19 insertions(+), 8 deletions(-)include/linux/bpf.h +1
-1@@ -56,7 +56,7 @@ extern struct idr btf_idr;extern spinlock_t btf_idr_lock;extern struct kobject *btf_kobj;extern struct bpf_mem_alloc bpf_global_ma, bpf_global_percpu_ma; -extern bool bpf_global_ma_set, bpf_global_percpu_ma_set; +extern bool bpf_global_ma_set;typedef u64 (*bpf_callback_t )(u64 ,u64 ,u64 ,u64 ,u64 ); typedef int (*bpf_iter_init_seq_priv_t)(void *private_data,kernel/bpf/core.c +3
-5@@ -64,8 +64,8 @@#define OFF insn->off#define IMM insn->imm -struct bpf_mem_alloc bpf_global_ma, bpf_global_percpu_ma; -bool bpf_global_ma_set, bpf_global_percpu_ma_set; +struct bpf_mem_alloc bpf_global_ma; +bool bpf_global_ma_set;/* No hurry in this branch * @@ -2934,9 +2934,7 @@ static int __init bpf_global_ma_init(void) ret =bpf_mem_alloc_init (&bpf_global_ma,0 , false); bpf_global_ma_set = !ret; - ret =bpf_mem_alloc_init (&bpf_global_percpu_ma,0 , true); - bpf_global_percpu_ma_set = !ret; - return !bpf_global_ma_set || !bpf_global_percpu_ma_set; + return ret; }late_initcall (bpf_global_ma_init);#endif kernel/bpf/verifier.c +4
-0@@ -26,6 +26,7 @@#include <linux/poison.h> #include <linux/module.h> #include <linux/cpumask.h> +#include <linux/bpf_mem_alloc.h> #include <net/xdp.h> #include "disasm.h" @@ -41,6 +42,9 @@ static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {#undef BPF_LINK_TYPE }; +struct bpf_mem_alloc bpf_global_percpu_ma; +static bool bpf_global_percpu_ma_set; +/* bpf_check() is a static code analyzer that walks eBPF program * instruction by instruction and updates register/stack state. * All paths of conditional branches are analyzed until 'bpf_exit' insn.
kernel/bpf/verifier.c static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn, · line 12082
Quoted 7 lines
kernel/bpf/verifier.c +2
-2@@ -12074,8 +12078,17 @@ static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,if (meta.func_id == special_kfunc_list[KF_bpf_obj_new_impl] && !bpf_global_ma_set)return -ENOMEM ; -if (meta.func_id == special_kfunc_list[KF_bpf_percpu_obj_new_impl] && !bpf_global_percpu_ma_set) -return -ENOMEM ; +if (meta.func_id == special_kfunc_list[KF_bpf_percpu_obj_new_impl]) { +mutex_lock (&bpf_verifier_lock);
Instead of acquiring the global lock each time, can we test whether or bpf_global_percpu_ma_set is set before acquiring the global lock ?
Quoted 12 lines
kernel/bpf/verifier.c +9
-0+if (!bpf_global_percpu_ma_set) { + err =bpf_mem_alloc_init (&bpf_global_percpu_ma,0 , true); +if (!err) + bpf_global_percpu_ma_set = true; + } +mutex_unlock (&bpf_verifier_lock); +if (err) +return err; + }if (((u64 )(u32 )meta.arg_constant.value) != meta.arg_constant.value) {verbose (env,"local type ID argument must be in range [0, U32_MAX]\n" );
Yonghong Song
reviewerQuoted 94 lines · Hou Tao
Quoted 92 lines · Yonghong Song
Kirill Shutemov reported significant percpu memory increase after booting in 288-cpu VM ([1]) due to commit
41a5db8d8161("bpf: Add support for non-fix-size percpu mem allocation"). The percpu memory is increased from 111MB to 969MB. The number is from/proc/meminfo.I tried to reproduce the issue with my local VM which at most supports upto 255 cpus. With 252 cpus, without the above commit, the percpu memory immediately after boot is 57MB while with the above commit the percpu memory is 231MB.
This is not good since so far percpu memory from bpf memory allocator is not widely used yet. Let us change pre-allocation in init stage to on-demand allocation when verifier detects there is a need of percpu memory for bpf program. With this change, percpu memory consumption after boot can be reduced signicantly.
[1] https://lore.kernel.org/lkml/[email protected]/Fixes 41a5db8d8161 ("bpf: Add support for non-fix-size percpu mem allocation")Cc Kirill A. Shutemov <[email protected]>Signed-off-by Yonghong Song <[email protected]>--- include/linux/bpf.h | 2 +- kernel/bpf/core.c | 8 +++----- kernel/bpf/verifier.c | 17 +++++++++++++++-- 3 files changed, 19 insertions(+), 8 deletions(-)include/linux/bpf.h +1
-1@@ -56,7 +56,7 @@ extern struct idr btf_idr;extern spinlock_t btf_idr_lock;extern struct kobject *btf_kobj;extern struct bpf_mem_alloc bpf_global_ma, bpf_global_percpu_ma; -extern bool bpf_global_ma_set, bpf_global_percpu_ma_set; +extern bool bpf_global_ma_set;typedef u64 (*bpf_callback_t )(u64 ,u64 ,u64 ,u64 ,u64 ); typedef int (*bpf_iter_init_seq_priv_t)(void *private_data,kernel/bpf/core.c +3
-5@@ -64,8 +64,8 @@#define OFF insn->off#define IMM insn->imm -struct bpf_mem_alloc bpf_global_ma, bpf_global_percpu_ma; -bool bpf_global_ma_set, bpf_global_percpu_ma_set; +struct bpf_mem_alloc bpf_global_ma; +bool bpf_global_ma_set;/* No hurry in this branch * @@ -2934,9 +2934,7 @@ static int __init bpf_global_ma_init(void) ret =bpf_mem_alloc_init (&bpf_global_ma,0 , false); bpf_global_ma_set = !ret; - ret =bpf_mem_alloc_init (&bpf_global_percpu_ma,0 , true); - bpf_global_percpu_ma_set = !ret; - return !bpf_global_ma_set || !bpf_global_percpu_ma_set; + return ret; }late_initcall (bpf_global_ma_init);#endif kernel/bpf/verifier.c +4
-0@@ -26,6 +26,7 @@#include <linux/poison.h> #include <linux/module.h> #include <linux/cpumask.h> +#include <linux/bpf_mem_alloc.h> #include <net/xdp.h> #include "disasm.h" @@ -41,6 +42,9 @@ static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {#undef BPF_LINK_TYPE }; +struct bpf_mem_alloc bpf_global_percpu_ma; +static bool bpf_global_percpu_ma_set; +/* bpf_check() is a static code analyzer that walks eBPF program * instruction by instruction and updates register/stack state. * All paths of conditional branches are analyzed until 'bpf_exit' insn. @@ -12074,8 +12078,17 @@ static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,if (meta.func_id == special_kfunc_list[KF_bpf_obj_new_impl] && !bpf_global_ma_set)return -ENOMEM ;
Quoted 6 lines
kernel/bpf/verifier.c static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn, · line 12082
Quoted 4 lines
kernel/bpf/verifier.c +2
-2-if (meta.func_id == special_kfunc_list[KF_bpf_percpu_obj_new_impl] && !bpf_global_percpu_ma_set) -return -ENOMEM ; +if (meta.func_id == special_kfunc_list[KF_bpf_percpu_obj_new_impl]) { +mutex_lock (&bpf_verifier_lock);Instead of acquiring the global lock each time, can we test whether or
bpf_global_percpu_ma_setis set before acquiring the global lock ?
Currently, in verifier we have two places to use bpf_verifier_lock:
(1) to get btf_vmlinux:
if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
mutex_lock(&bpf_verifier_lock);
if (!btf_vmlinux)
btf_vmlinux = btf_parse_vmlinux();
mutex_unlock(&bpf_verifier_lock);
}This will only lock once if btf_parse_vmlinux() is successful.
(2) for unprividged bpf programs in bpf_check().
A big chunk of it is under bpf_verifier_lock.
I didn't use style (1) since I assume unprividged bpf programs is rare and it should seldomly collide with percpu_obj_new_impl.
But my assumption related to (2) may be wrong and in the future verifier_lock() could be used in other places which may cause more contention.
So I now agree with you and will make appropriate change. Thanks!
Quoted 12 lines
Quoted 12 lines
kernel/bpf/verifier.c +9
-0+if (!bpf_global_percpu_ma_set) { + err =bpf_mem_alloc_init (&bpf_global_percpu_ma,0 , true); +if (!err) + bpf_global_percpu_ma_set = true; + } +mutex_unlock (&bpf_verifier_lock); +if (err) +return err; + }if (((u64 )(u32 )meta.arg_constant.value) != meta.arg_constant.value) {verbose (env,"local type ID argument must be in range [0, U32_MAX]\n" );