Commit f39703b2 authored by Alexei Starovoitov's avatar Alexei Starovoitov
Browse files

Merge branch 'bpf-introduce-bpf_f_cpu-and-bpf_f_all_cpus-flags-for-percpu-maps' 

Leon Hwang says:

====================
bpf: Introduce BPF_F_CPU and BPF_F_ALL_CPUS flags for percpu maps

This patch set introduces the BPF_F_CPU and BPF_F_ALL_CPUS flags for
percpu maps, as the requirement of BPF_F_ALL_CPUS flag for percpu_array
maps was discussed in the thread of
"[PATCH bpf-next v3 0/4] bpf: Introduce global percpu data"[1].

The goal of BPF_F_ALL_CPUS flag is to reduce data caching overhead in light
skeletons by allowing a single value to be reused to update values across all
CPUs. This avoids the M:N problem where M cached values are used to update a
map on N CPUs kernel.

The BPF_F_CPU flag is accompanied by *flags*-embedded cpu info, which
specifies the target CPU for the operation:

* For lookup operations: the flag field alongside cpu info enable querying
  a value on the specified CPU.
* For update operations: the flag field alongside cpu info enable
  updating value for specified CPU.

Links:
[1] https://lore.kernel.org/bpf/20250526162146.24429-1-leon.hwang@linux.dev/

Changes:
v12 -> v13:
* No changes, rebased on latest tree.

v11 -> v12:
* Dropped the v11 changes.
* Stabilized the lru_percpu_hash map test by keeping an extra spare entry,
  which can be used temporarily during updates to avoid unintended LRU
  evictions.

v10 -> v11:
* Support the combination of BPF_EXIST and BPF_F_CPU/BPF_F_ALL_CPUS for
  update operations.
* Fix unstable lru_percpu_hash map test using the combination of
  BPF_EXIST and BPF_F_CPU/BPF_F_ALL_CPUS to avoid LRU eviction
  (reported by Alexei).

v9 -> v10:
* Add tests to verify array and hash maps do not support BPF_F_CPU and
  BPF_F_ALL_CPUS flags.
* Address comment from Andrii:
  * Copy map value using copy_map_value_long for percpu_cgroup_storage
    maps in a separate patch.

v8 -> v9:
* Change value type from u64 to u32 in selftests.
* Address comments from Andrii:
  * Keep value_size unaligned and update everywhere for consistency when
    cpu flags are specified.
  * Update value by getting pointer for percpu hash and percpu
    cgroup_storage maps.

v7 -> v8:
* Address comments from Andrii:
  * Check BPF_F_LOCK when update percpu_array, percpu_hash and
    lru_percpu_hash maps.
  * Refactor flags check in __htab_map_lookup_and_delete_batch().
  * Keep value_size unaligned and copy value using copy_map_value() in
    __htab_map_lookup_and_delete_batch() when BPF_F_CPU is specified.
  * Update warn message in libbpf's validate_map_op().
  * Update comment of libbpf's bpf_map__lookup_elem().

v6 -> v7:
* Get correct value size for percpu_hash and lru_percpu_hash in
  update_batch API.
* Set 'count' as 'max_entries' in test cases for lookup_batch API.
* Address comment from Alexei:
  * Move cpu flags check into bpf_map_check_op_flags().

v5 -> v6:
* Move bpf_map_check_op_flags() from 'bpf.h' to 'syscall.c'.
* Address comments from Alexei:
  * Drop the refactoring code of data copying logic for percpu maps.
  * Drop bpf_map_check_op_flags() wrappers.

v4 -> v5:
* Address comments from Andrii:
  * Refactor data copying logic for all percpu maps.
  * Drop this_cpu_ptr() micro-optimization.
  * Drop cpu check in libbpf's validate_map_op().
  * Enhance bpf_map_check_op_flags() using *allowed flags* instead of
    'extra_flags_mask'.

v3 -> v4:
* Address comments from Andrii:
  * Remove unnecessary map_type check in bpf_map_value_size().
  * Reduce code churn.
  * Remove unnecessary do_delete check in
    __htab_map_lookup_and_delete_batch().
  * Introduce bpf_percpu_copy_to_user() and bpf_percpu_copy_from_user().
  * Rename check_map_flags() to bpf_map_check_op_flags() with
    extra_flags_mask.
  * Add human-readable pr_warn() explanations in validate_map_op().
  * Use flags in bpf_map__delete_elem() and
    bpf_map__lookup_and_delete_elem().
  * Drop "for alignment reasons".
v3 link: https://lore.kernel.org/bpf/20250821160817.70285-1-leon.hwang@linux.dev/

v2 -> v3:
* Address comments from Alexei:
  * Use BPF_F_ALL_CPUS instead of BPF_ALL_CPUS magic.
  * Introduce these two cpu flags for all percpu maps.
* Address comments from Jiri:
  * Reduce some unnecessary u32 cast.
  * Refactor more generic map flags check function.
  * A code style issue.
v2 link: https://lore.kernel.org/bpf/20250805163017.17015-1-leon.hwang@linux.dev/

v1 -> v2:
* Address comments from Andrii:
  * Embed cpu info as high 32 bits of *flags* totally.
  * Use ERANGE instead of E2BIG.
  * Few format issues.
====================

Link: https://patch.msgid.link/20260107022022.12843-1-leon.hwang@linux.dev


Signed-off-by: default avatarAlexei Starovoitov <ast@kernel.org>
parents a8d50675 07bf7aa5

include/linux/bpf-cgroup.h

+2 −2
Original line number Diff line number Diff line
@@ -172,7 +172,7 @@ void bpf_cgroup_storage_link(struct bpf_cgroup_storage *storage, 
void bpf_cgroup_storage_unlink(struct bpf_cgroup_storage *storage);

int bpf_cgroup_storage_assign(struct bpf_prog_aux *aux, struct bpf_map *map);



int bpf_percpu_cgroup_storage_copy(struct bpf_map *map, void *key, void *value);

int bpf_percpu_cgroup_storage_copy(struct bpf_map *map, void *key, void *value, u64 flags);

int bpf_percpu_cgroup_storage_update(struct bpf_map *map, void *key,

				     void *value, u64 flags);
@@ -470,7 +470,7 @@ static inline struct bpf_cgroup_storage *bpf_cgroup_storage_alloc( 
static inline void bpf_cgroup_storage_free(

	struct bpf_cgroup_storage *storage) {}

static inline int bpf_percpu_cgroup_storage_copy(struct bpf_map *map, void *key,

						 void *value) {

						 void *value, u64 flags) {

	return 0;

}

static inline int bpf_percpu_cgroup_storage_update(struct bpf_map *map,

include/linux/bpf.h

+32 −3
Original line number Diff line number Diff line
@@ -2847,8 +2847,8 @@ int map_set_for_each_callback_args(struct bpf_verifier_env *env, 
				   struct bpf_func_state *caller,

				   struct bpf_func_state *callee);



int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value);

int bpf_percpu_array_copy(struct bpf_map *map, void *key, void *value);

int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value, u64 flags);

int bpf_percpu_array_copy(struct bpf_map *map, void *key, void *value, u64 flags);

int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,

			   u64 flags);

int bpf_percpu_array_update(struct bpf_map *map, void *key, void *value,
@@ -3915,14 +3915,43 @@ bpf_prog_update_insn_ptrs(struct bpf_prog *prog, u32 *offsets, void *image) 
}

#endif



static inline bool bpf_map_supports_cpu_flags(enum bpf_map_type map_type)

{

	switch (map_type) {

	case BPF_MAP_TYPE_PERCPU_ARRAY:

	case BPF_MAP_TYPE_PERCPU_HASH:

	case BPF_MAP_TYPE_LRU_PERCPU_HASH:

	case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:

		return true;

	default:

		return false;

	}

}



static inline int bpf_map_check_op_flags(struct bpf_map *map, u64 flags, u64 allowed_flags)

{

	if (flags & ~allowed_flags)

	u32 cpu;



	if ((u32)flags & ~allowed_flags)

		return -EINVAL;



	if ((flags & BPF_F_LOCK) && !btf_record_has_field(map->record, BPF_SPIN_LOCK))

		return -EINVAL;



	if (!(flags & BPF_F_CPU) && flags >> 32)

		return -EINVAL;



	if (flags & (BPF_F_CPU | BPF_F_ALL_CPUS)) {

		if (!bpf_map_supports_cpu_flags(map->map_type))

			return -EINVAL;

		if ((flags & BPF_F_CPU) && (flags & BPF_F_ALL_CPUS))

			return -EINVAL;



		cpu = flags >> 32;

		if ((flags & BPF_F_CPU) && cpu >= num_possible_cpus())

			return -ERANGE;

	}



	return 0;

}

include/uapi/linux/bpf.h

+2 −0
Original line number Diff line number Diff line
@@ -1384,6 +1384,8 @@ enum { 
	BPF_NOEXIST	= 1, /* create new element if it didn't exist */

	BPF_EXIST	= 2, /* update existing element */

	BPF_F_LOCK	= 4, /* spin_lock-ed map_lookup/map_update */

	BPF_F_CPU	= 8, /* cpu flag for percpu maps, upper 32-bit of flags is a cpu number */

	BPF_F_ALL_CPUS	= 16, /* update value across all CPUs for percpu maps */

};



/* flags for BPF_MAP_CREATE command */

kernel/bpf/arraymap.c

+23 −6
Original line number Diff line number Diff line
@@ -307,7 +307,7 @@ static void *percpu_array_map_lookup_percpu_elem(struct bpf_map *map, void *key, 
	return per_cpu_ptr(array->pptrs[index & array->index_mask], cpu);

}



int bpf_percpu_array_copy(struct bpf_map *map, void *key, void *value)

int bpf_percpu_array_copy(struct bpf_map *map, void *key, void *value, u64 map_flags)

{

	struct bpf_array *array = container_of(map, struct bpf_array, map);

	u32 index = *(u32 *)key;
@@ -325,11 +325,18 @@ int bpf_percpu_array_copy(struct bpf_map *map, void *key, void *value) 
	size = array->elem_size;

	rcu_read_lock();

	pptr = array->pptrs[index & array->index_mask];

	if (map_flags & BPF_F_CPU) {

		cpu = map_flags >> 32;

		copy_map_value(map, value, per_cpu_ptr(pptr, cpu));

		check_and_init_map_value(map, value);

		goto unlock;

	}

	for_each_possible_cpu(cpu) {

		copy_map_value_long(map, value + off, per_cpu_ptr(pptr, cpu));

		check_and_init_map_value(map, value + off);

		off += size;

	}

unlock:

	rcu_read_unlock();

	return 0;

}
@@ -398,10 +405,11 @@ int bpf_percpu_array_update(struct bpf_map *map, void *key, void *value, 
	struct bpf_array *array = container_of(map, struct bpf_array, map);

	u32 index = *(u32 *)key;

	void __percpu *pptr;

	int cpu, off = 0;

	void *ptr, *val;

	u32 size;

	int cpu;



	if (unlikely(map_flags > BPF_EXIST))

	if (unlikely((map_flags & BPF_F_LOCK) || (u32)map_flags > BPF_F_ALL_CPUS))

		/* unknown flags */

		return -EINVAL;
@@ -422,11 +430,20 @@ int bpf_percpu_array_update(struct bpf_map *map, void *key, void *value, 
	size = array->elem_size;

	rcu_read_lock();

	pptr = array->pptrs[index & array->index_mask];

	if (map_flags & BPF_F_CPU) {

		cpu = map_flags >> 32;

		ptr = per_cpu_ptr(pptr, cpu);

		copy_map_value(map, ptr, value);

		bpf_obj_free_fields(array->map.record, ptr);

		goto unlock;

	}

	for_each_possible_cpu(cpu) {

		copy_map_value_long(map, per_cpu_ptr(pptr, cpu), value + off);

		bpf_obj_free_fields(array->map.record, per_cpu_ptr(pptr, cpu));

		off += size;

		ptr = per_cpu_ptr(pptr, cpu);

		val = (map_flags & BPF_F_ALL_CPUS) ? value : value + size * cpu;

		copy_map_value(map, ptr, val);

		bpf_obj_free_fields(array->map.record, ptr);

	}

unlock:

	rcu_read_unlock();

	return 0;

}

kernel/bpf/hashtab.c

+64 −30
Original line number Diff line number Diff line
@@ -932,7 +932,7 @@ static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l) 
}



static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,

			    void *value, bool onallcpus)

			    void *value, bool onallcpus, u64 map_flags)

{

	void *ptr;
@@ -943,19 +943,28 @@ static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr, 
		bpf_obj_free_fields(htab->map.record, ptr);

	} else {

		u32 size = round_up(htab->map.value_size, 8);

		int off = 0, cpu;

		void *val;

		int cpu;



		if (map_flags & BPF_F_CPU) {

			cpu = map_flags >> 32;

			ptr = per_cpu_ptr(pptr, cpu);

			copy_map_value(&htab->map, ptr, value);

			bpf_obj_free_fields(htab->map.record, ptr);

			return;

		}



		for_each_possible_cpu(cpu) {

			ptr = per_cpu_ptr(pptr, cpu);

			copy_map_value_long(&htab->map, ptr, value + off);

			val = (map_flags & BPF_F_ALL_CPUS) ? value : value + size * cpu;

			copy_map_value(&htab->map, ptr, val);

			bpf_obj_free_fields(htab->map.record, ptr);

			off += size;

		}

	}

}



static void pcpu_init_value(struct bpf_htab *htab, void __percpu *pptr,

			    void *value, bool onallcpus)

			    void *value, bool onallcpus, u64 map_flags)

{

	/* When not setting the initial value on all cpus, zero-fill element

	 * values for other cpus. Otherwise, bpf program has no way to ensure
@@ -973,7 +982,7 @@ static void pcpu_init_value(struct bpf_htab *htab, void __percpu *pptr, 
				zero_map_value(&htab->map, per_cpu_ptr(pptr, cpu));

		}

	} else {

		pcpu_copy_value(htab, pptr, value, onallcpus);

		pcpu_copy_value(htab, pptr, value, onallcpus, map_flags);

	}

}
@@ -985,7 +994,7 @@ static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab) 
static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,

					 void *value, u32 key_size, u32 hash,

					 bool percpu, bool onallcpus,

					 struct htab_elem *old_elem)

					 struct htab_elem *old_elem, u64 map_flags)

{

	u32 size = htab->map.value_size;

	bool prealloc = htab_is_prealloc(htab);
@@ -1043,7 +1052,7 @@ static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key, 
			pptr = *(void __percpu **)ptr;

		}



		pcpu_init_value(htab, pptr, value, onallcpus);

		pcpu_init_value(htab, pptr, value, onallcpus, map_flags);



		if (!prealloc)

			htab_elem_set_ptr(l_new, key_size, pptr);
@@ -1147,7 +1156,7 @@ static long htab_map_update_elem(struct bpf_map *map, void *key, void *value, 
	}



	l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,

				l_old);

				l_old, map_flags);

	if (IS_ERR(l_new)) {

		/* all pre-allocated elements are in use or memory exhausted */

		ret = PTR_ERR(l_new);
@@ -1249,6 +1258,15 @@ static long htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value 
	return ret;

}



static int htab_map_check_update_flags(bool onallcpus, u64 map_flags)

{

	if (unlikely(!onallcpus && map_flags > BPF_EXIST))

		return -EINVAL;

	if (unlikely(onallcpus && ((map_flags & BPF_F_LOCK) || (u32)map_flags > BPF_F_ALL_CPUS)))

		return -EINVAL;

	return 0;

}



static long htab_map_update_elem_in_place(struct bpf_map *map, void *key,

					  void *value, u64 map_flags,

					  bool percpu, bool onallcpus)
@@ -1262,9 +1280,9 @@ static long htab_map_update_elem_in_place(struct bpf_map *map, void *key, 
	u32 key_size, hash;

	int ret;



	if (unlikely(map_flags > BPF_EXIST))

		/* unknown flags */

		return -EINVAL;

	ret = htab_map_check_update_flags(onallcpus, map_flags);

	if (unlikely(ret))

		return ret;



	WARN_ON_ONCE(!bpf_rcu_lock_held());
@@ -1289,7 +1307,7 @@ static long htab_map_update_elem_in_place(struct bpf_map *map, void *key, 
		/* Update value in-place */

		if (percpu) {

			pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),

					value, onallcpus);

					value, onallcpus, map_flags);

		} else {

			void **inner_map_pptr = htab_elem_value(l_old, key_size);
@@ -1298,7 +1316,7 @@ static long htab_map_update_elem_in_place(struct bpf_map *map, void *key, 
		}

	} else {

		l_new = alloc_htab_elem(htab, key, value, key_size,

					hash, percpu, onallcpus, NULL);

					hash, percpu, onallcpus, NULL, map_flags);

		if (IS_ERR(l_new)) {

			ret = PTR_ERR(l_new);

			goto err;
@@ -1324,9 +1342,9 @@ static long __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key, 
	u32 key_size, hash;

	int ret;



	if (unlikely(map_flags > BPF_EXIST))

		/* unknown flags */

		return -EINVAL;

	ret = htab_map_check_update_flags(onallcpus, map_flags);

	if (unlikely(ret))

		return ret;



	WARN_ON_ONCE(!bpf_rcu_lock_held());
@@ -1363,10 +1381,10 @@ static long __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key, 


		/* per-cpu hash map can update value in-place */

		pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),

				value, onallcpus);

				value, onallcpus, map_flags);

	} else {

		pcpu_init_value(htab, htab_elem_get_ptr(l_new, key_size),

				value, onallcpus);

				value, onallcpus, map_flags);

		hlist_nulls_add_head_rcu(&l_new->hash_node, head);

		l_new = NULL;

	}
@@ -1678,9 +1696,9 @@ __htab_map_lookup_and_delete_batch(struct bpf_map *map, 
	void __user *ukeys = u64_to_user_ptr(attr->batch.keys);

	void __user *ubatch = u64_to_user_ptr(attr->batch.in_batch);

	u32 batch, max_count, size, bucket_size, map_id;

	u64 elem_map_flags, map_flags, allowed_flags;

	u32 bucket_cnt, total, key_size, value_size;

	struct htab_elem *node_to_free = NULL;

	u64 elem_map_flags, map_flags;

	struct hlist_nulls_head *head;

	struct hlist_nulls_node *n;

	unsigned long flags = 0;
@@ -1690,9 +1708,12 @@ __htab_map_lookup_and_delete_batch(struct bpf_map *map, 
	int ret = 0;



	elem_map_flags = attr->batch.elem_flags;

	if ((elem_map_flags & ~BPF_F_LOCK) ||

	    ((elem_map_flags & BPF_F_LOCK) && !btf_record_has_field(map->record, BPF_SPIN_LOCK)))

		return -EINVAL;

	allowed_flags = BPF_F_LOCK;

	if (!do_delete && is_percpu)

		allowed_flags |= BPF_F_CPU;

	ret = bpf_map_check_op_flags(map, elem_map_flags, allowed_flags);

	if (ret)

		return ret;



	map_flags = attr->batch.flags;

	if (map_flags)
@@ -1715,7 +1736,7 @@ __htab_map_lookup_and_delete_batch(struct bpf_map *map, 
	key_size = htab->map.key_size;

	value_size = htab->map.value_size;

	size = round_up(value_size, 8);

	if (is_percpu)

	if (is_percpu && !(elem_map_flags & BPF_F_CPU))

		value_size = size * num_possible_cpus();

	total = 0;

	/* while experimenting with hash tables with sizes ranging from 10 to
@@ -1798,11 +1819,18 @@ __htab_map_lookup_and_delete_batch(struct bpf_map *map, 
			void __percpu *pptr;



			pptr = htab_elem_get_ptr(l, map->key_size);

			if (elem_map_flags & BPF_F_CPU) {

				cpu = elem_map_flags >> 32;

				copy_map_value(&htab->map, dst_val, per_cpu_ptr(pptr, cpu));

				check_and_init_map_value(&htab->map, dst_val);

			} else {

				for_each_possible_cpu(cpu) {

				copy_map_value_long(&htab->map, dst_val + off, per_cpu_ptr(pptr, cpu));

					copy_map_value_long(&htab->map, dst_val + off,

							    per_cpu_ptr(pptr, cpu));

					check_and_init_map_value(&htab->map, dst_val + off);

					off += size;

				}

			}

		} else {

			value = htab_elem_value(l, key_size);

			if (is_fd_htab(htab)) {
@@ -2357,7 +2385,7 @@ static void *htab_lru_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *k 
	return NULL;

}



int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)

int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value, u64 map_flags)

{

	struct htab_elem *l;

	void __percpu *pptr;
@@ -2374,16 +2402,22 @@ int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value) 
	l = __htab_map_lookup_elem(map, key);

	if (!l)

		goto out;

	ret = 0;

	/* We do not mark LRU map element here in order to not mess up

	 * eviction heuristics when user space does a map walk.

	 */

	pptr = htab_elem_get_ptr(l, map->key_size);

	if (map_flags & BPF_F_CPU) {

		cpu = map_flags >> 32;

		copy_map_value(map, value, per_cpu_ptr(pptr, cpu));

		check_and_init_map_value(map, value);

		goto out;

	}

	for_each_possible_cpu(cpu) {

		copy_map_value_long(map, value + off, per_cpu_ptr(pptr, cpu));

		check_and_init_map_value(map, value + off);

		off += size;

	}

	ret = 0;

out:

	rcu_read_unlock();

	return ret;