[PATCH bpf-next v4 0/2] bpf: htab: Reduce memory use of hash maps

[PATCH bpf-next v4 0/2] bpf: htab: Reduce memory use of hash maps

[T.J. Mercier]

Memory is expensive and scarce these days. This series reduces the
memory use of BPF hash maps by eliminating the per-element overheads
below. This saves up to 50% of per-element memory use for standard and
PCPU hash maps. The memory use of LRU hash maps is unaffected.

Map Type & Configuration            | Old size | New size | Savings
------------------------------------|----------|----------|--------
Standard (key <= 8 B, val <= 8 B)   |   64 B   |   32 B   | 50.0%
Per-CPU (prealloc) (key <= 8 B)     |   64 B   |   32 B   | 50.0%
Per-CPU (non-prealloc) (key <= 8 B) |   64 B   |   40 B   | 37.5%
LRU (Any key/value size)            |    -     |    -     | 00.0%

1) Unused LRU / PCPU fields in standard and PCPU hash maps (patch 1)
struct htab_elem is used for all hash map types, and includes fields
that are not always used (bpf_lru_node, ptr_to_pptr). For standard
(non-LRU, non-PCPU) hash maps the 24 bytes for the bpf_lru_node (union)
are entirely overhead and can be eliminated. Non-preallocated PCPU maps
only need the 8 byte ptr_to_pptr which is currently unioned with the
unneeded 24 byte bpf_lru_node, so 16 bytes of overhead can be
eliminated. Preallocated PCPU maps don't need ptr_to_pptr, so 24 bytes
of overhead can be saved.

2) Hash caching for small keys (patch 2)
For hash maps with small key sizes (<= word size), comparing keys only
requires a single instruction. Currently the 4 byte hash value (8 byte
aligned and padded) is used for this, but offers no performance
advantage in this case and can be eliminated.

The implementation splits htab_elem into dedicated structures for the
different map types (htab_elem_lru, htab_elem_pcpu, htab_elem) which
share a common initial sequence (struct htab_node), but contain
additional map-type specific fields where necessary. This means the
placement of the key for each element varies with the map type, and
key_offset is added to bpf_htab for this purpose.

While using key_offset and conditional hash checks adds new pointer
dereferences and branching during element traversal,
run_bench_htab_mem.sh shows no significant performance regression across
10 runs on my 3995WX.

Benchmark (all in kops/sec)  | Avg. Before | StDev | Avg. After | StDev
-----------------------------|-------------|-------|------------|------
prealloc overwrite           | 115.11      | 4.10  | 115.45     | 5.24
prealloc batch_add_batch_del | 127.14      | 4.32  | 127.06     | 2.32
prealloc add_del_on_diff_cpu | 23.22       | 0.93  | 22.91      | 1.60
normal overwrite             | 78.52       | 3.05  | 80.40      | 3.25
normal batch_add_batch_del   | 45.37       | 0.69  | 47.71      | 0.66
normal add_del_on_diff_cpu   | 12.02       | 0.73  | 12.48      | 0.70

---
Changes in v4:
Removed inline from new functions per BPF CI (netdev/source_inline).

From Mykyta Yatsenko:
Factor out duplicate lookup_elem code into __lookup_elem_raw.
Use offsetof instead of sizeof for key_offset assignments in
htab_map_alloc (patch 1).
Eliminate branching and htab_elem casting in htab_elem_hash /
htab_elem_set_hash.

Changes in v3:
From Sashiko on torn reads/writes:
Use a local unsigned long and READ_ONCE / WRITE_ONCE instead of memcmp /
memcpy for atomic key comparisons for hashless elements.

Changes in v2:
Make maximum key_size for !has_hash depend on word size for atomicity
on 32-bit.

From Mykyta Yatsenko:
Put the htab_elem* common initial sequence in its own struct (htab_node)
and reuse it across all element types that share it. Eliminate
associated BUILD_BUG_ON additions.
Replace both the hash and key fields with data[].
Store has_hash in struct bpf_htab, and avoid per-element reads of it.# Please edit the description for the branch

T.J. Mercier (2):
  bpf: htab: Split htab_elem_lru and htab_elem_pcpu off of htab_elem
  bpf: htab: Reduce elem_size by 8 bytes for small key sizes

 kernel/bpf/hashtab.c                          | 418 ++++++++++++------
 kernel/bpf/map_in_map.c                       |  13 +
 kernel/bpf/map_in_map.h                       |   2 +
 .../selftests/bpf/progs/map_ptr_kern.c        |   2 +-
 4 files changed, 289 insertions(+), 146 deletions(-)


base-commit: cfce77b63375dac81d53f2f85593c548415206b7
-- 
2.55.0.691.gc56d675ccc-goog

[PATCH bpf-next v4 1/2] bpf: htab: Split htab_elem_lru and htab_elem_pcpu off of htab_elem

[T.J. Mercier]

The htab_elem struct is used as the per-element type for all BPF hash
map types and includes bpf_lru_node in a union with a ptr_to_pptr
pointer. For standard (non-LRU, non-PCPU) hash maps, the 24 byte union
allocated for every element is entirely unused. For non-preallocated
PCPU maps, ptr_to_pptr only requires 8 bytes, leaving 16 bytes of unused
overhead in the union. For preallocated PCPU maps ptr_to_pptr is unused
since elements are freed to the PCPU freelist.

Eliminate this per-element memory overhead by splitting htab_elem into
dedicated structures for each map type:
   - struct htab_elem: Minimal structure for standard hash maps and
     preallocated PCPU maps (saves 24 bytes per element).
   - struct htab_elem_pcpu: Structure for non-preallocated PCPU maps
     containing ptr_to_pptr (saves 16 bytes per element).
   - struct htab_elem_lru: Retains struct bpf_lru_node for LRU maps.

Because element sizes now vary by map type, add key_offset to struct
bpf_htab to track the dynamic key offset. Update helper accessors and
lookups to compute key and value offsets using htab->key_offset.

Pointers to struct htab_elem in the existing code (e.g. htab_elem_hash)
serve as generic base element pointers. This is possible because
htab_elem, htab_elem_pcpu, and htab_elem_lru share a common initial
sequence, making pointer casts safe.

Signed-off-by: T.J. Mercier <tjmercier@google.com>
---
 kernel/bpf/hashtab.c    | 354 +++++++++++++++++++++++++---------------
 kernel/bpf/map_in_map.c |  13 ++
 kernel/bpf/map_in_map.h |   2 +
 3 files changed, 233 insertions(+), 136 deletions(-)

diff --git a/kernel/bpf/hashtab.c b/kernel/bpf/hashtab.c
index 9f394e1aa2e8..fed4b1db8b78 100644
--- a/kernel/bpf/hashtab.c
+++ b/kernel/bpf/hashtab.c
@@ -102,11 +102,13 @@ struct bpf_htab {
 	bool use_percpu_counter;
 	u32 n_buckets;	/* number of hash buckets */
 	u32 elem_size;	/* size of each element in bytes */
+	u32 key_offset;	/* offset of key in bytes */
 	u32 hashrnd;
 };
 
 /* each htab element is struct htab_elem + key + value */
-struct htab_elem {
+struct htab_elem;
+struct htab_node {
 	union {
 		struct hlist_nulls_node hash_node;
 		struct {
@@ -117,12 +119,28 @@ struct htab_elem {
 			};
 		};
 	};
-	union {
-		/* pointer to per-cpu pointer */
-		void *ptr_to_pptr;
-		struct bpf_lru_node lru_node;
-	};
-	u32 hash;
+};
+
+struct htab_elem {
+	struct htab_node node;
+	u32 hash __aligned(8);
+	char key[] __aligned(8);
+};
+
+struct htab_elem_lru {
+	struct htab_node node;
+	struct bpf_lru_node lru_node;
+	u32 hash __aligned(8);
+	char key[] __aligned(8);
+};
+
+/* Only for non-preallocated PCPU maps. Preallocated PCPU maps don't need
+ * ptr_to_pptr, and use htab_elem.
+ */
+struct htab_elem_pcpu {
+	struct htab_node node;
+	void *ptr_to_pptr;
+	u32 hash __aligned(8);
 	char key[] __aligned(8);
 };
 
@@ -136,6 +154,21 @@ static inline bool htab_is_prealloc(const struct bpf_htab *htab)
 	return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
 }
 
+static struct bpf_lru_node *htab_elem_lru_node(struct htab_elem *l)
+{
+	return &((struct htab_elem_lru *)l)->lru_node;
+}
+
+static void *htab_elem_get_ptr_to_pptr(struct htab_elem *l)
+{
+	return ((struct htab_elem_pcpu *)l)->ptr_to_pptr;
+}
+
+static void htab_elem_set_ptr_to_pptr(struct htab_elem *l, void *ptr)
+{
+	((struct htab_elem_pcpu *)l)->ptr_to_pptr = ptr;
+}
+
 static void htab_init_buckets(struct bpf_htab *htab)
 {
 	unsigned int i;
@@ -183,25 +216,30 @@ static inline bool is_fd_htab(const struct bpf_htab *htab)
 	return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS;
 }
 
-static inline void *htab_elem_value(struct htab_elem *l, u32 key_size)
+static void *htab_elem_key(struct bpf_htab *htab, struct htab_elem *l)
+{
+	return (void *)l + htab->key_offset;
+}
+
+static inline void *htab_elem_value(struct bpf_htab *htab, struct htab_elem *l)
 {
-	return l->key + round_up(key_size, 8);
+	return htab_elem_key(htab, l) + round_up(htab->map.key_size, 8);
 }
 
-static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
+static inline void htab_elem_set_ptr(struct bpf_htab *htab, struct htab_elem *l,
 				     void __percpu *pptr)
 {
-	*(void __percpu **)htab_elem_value(l, key_size) = pptr;
+	*(void __percpu **)htab_elem_value(htab, l) = pptr;
 }
 
-static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
+static inline void __percpu *htab_elem_get_ptr(struct bpf_htab *htab, struct htab_elem *l)
 {
-	return *(void __percpu **)htab_elem_value(l, key_size);
+	return *(void __percpu **)htab_elem_value(htab, l);
 }
 
-static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
+static void *fd_htab_map_get_ptr(struct bpf_htab *htab, struct htab_elem *l)
 {
-	return *(void **)htab_elem_value(l, map->key_size);
+	return *(void **)htab_elem_value(htab, l);
 }
 
 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
@@ -209,6 +247,16 @@ static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
 	return (struct htab_elem *) (htab->elems + i * (u64)htab->elem_size);
 }
 
+static u32 htab_elem_hash(struct bpf_htab *htab, struct htab_elem *l)
+{
+	return *(u32 *)(htab_elem_key(htab, l) - 8);
+}
+
+static void htab_elem_set_hash(struct bpf_htab *htab, struct htab_elem *l, u32 hash)
+{
+	*(u32 *)(htab_elem_key(htab, l) - 8) = hash;
+}
+
 /* Both percpu and fd htab support in-place update, so no need for
  * extra elem. LRU itself can remove the least used element, so
  * there is no need for an extra elem during map_update.
@@ -231,7 +279,7 @@ static void htab_free_prealloced_internal_structs(struct bpf_htab *htab)
 
 		elem = get_htab_elem(htab, i);
 		bpf_map_free_internal_structs(&htab->map,
-					      htab_elem_value(elem, htab->map.key_size));
+					      htab_elem_value(htab, elem));
 		cond_resched();
 	}
 }
@@ -254,7 +302,7 @@ static void htab_free_prealloced_fields(struct bpf_htab *htab)
 
 		elem = get_htab_elem(htab, i);
 		if (htab_is_percpu(htab)) {
-			void __percpu *pptr = htab_elem_get_ptr(elem, htab->map.key_size);
+			void __percpu *pptr = htab_elem_get_ptr(htab, elem);
 			int cpu;
 
 			for_each_possible_cpu(cpu) {
@@ -263,7 +311,7 @@ static void htab_free_prealloced_fields(struct bpf_htab *htab)
 			}
 		} else {
 			bpf_obj_free_fields(htab->map.record,
-					    htab_elem_value(elem, htab->map.key_size));
+					    htab_elem_value(htab, elem));
 			cond_resched();
 		}
 		cond_resched();
@@ -280,8 +328,7 @@ static void htab_free_elems(struct bpf_htab *htab)
 	for (i = 0; i < htab->map.max_entries; i++) {
 		void __percpu *pptr;
 
-		pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
-					 htab->map.key_size);
+		pptr = htab_elem_get_ptr(htab, get_htab_elem(htab, i));
 		free_percpu(pptr);
 		cond_resched();
 	}
@@ -308,8 +355,8 @@ static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
 
 	if (node) {
 		bpf_map_inc_elem_count(&htab->map);
-		l = container_of(node, struct htab_elem, lru_node);
-		memcpy(l->key, key, htab->map.key_size);
+		l = (struct htab_elem *)container_of(node, struct htab_elem_lru, lru_node);
+		memcpy(htab_elem_key(htab, l), key, htab->map.key_size);
 		return l;
 	}
 
@@ -340,8 +387,7 @@ static int prealloc_init(struct bpf_htab *htab)
 					    GFP_USER | __GFP_NOWARN);
 		if (!pptr)
 			goto free_elems;
-		htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
-				  pptr);
+		htab_elem_set_ptr(htab, get_htab_elem(htab, i), pptr);
 		cond_resched();
 	}
 
@@ -349,8 +395,8 @@ static int prealloc_init(struct bpf_htab *htab)
 	if (htab_is_lru(htab))
 		err = bpf_lru_init(&htab->lru,
 				   htab->map.map_flags & BPF_F_NO_COMMON_LRU,
-				   offsetof(struct htab_elem, hash) -
-				   offsetof(struct htab_elem, lru_node),
+				   offsetof(struct htab_elem_lru, hash) -
+				   offsetof(struct htab_elem_lru, lru_node),
 				   htab_lru_map_delete_node,
 				   htab);
 	else
@@ -361,11 +407,11 @@ static int prealloc_init(struct bpf_htab *htab)
 
 	if (htab_is_lru(htab))
 		bpf_lru_populate(&htab->lru, htab->elems,
-				 offsetof(struct htab_elem, lru_node),
+				 offsetof(struct htab_elem_lru, lru_node),
 				 htab->elem_size, num_entries);
 	else
 		pcpu_freelist_populate(&htab->freelist,
-				       htab->elems + offsetof(struct htab_elem, fnode),
+				       htab->elems + offsetof(struct htab_elem, node.fnode),
 				       htab->elem_size, num_entries);
 
 	return 0;
@@ -401,7 +447,7 @@ static int alloc_extra_elems(struct bpf_htab *htab)
 		/* pop will succeed, since prealloc_init()
 		 * preallocated extra num_possible_cpus elements
 		 */
-		l_new = container_of(l, struct htab_elem, fnode);
+		l_new = container_of(l, struct htab_elem, node.fnode);
 		*per_cpu_ptr(pptr, cpu) = l_new;
 	}
 	htab->extra_elems = pptr;
@@ -425,8 +471,8 @@ static int htab_map_alloc_check(union bpf_attr *attr)
 	bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
 	int numa_node = bpf_map_attr_numa_node(attr);
 
-	BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
-		     offsetof(struct htab_elem, hash_node.pprev));
+	BUILD_BUG_ON(offsetof(struct htab_node, fnode.next) !=
+		     offsetof(struct htab_node, hash_node.pprev));
 
 	if (zero_seed && !capable(CAP_SYS_ADMIN))
 		/* Guard against local DoS, and discourage production use. */
@@ -476,7 +522,7 @@ static void htab_mem_dtor(void *obj, void *ctx)
 	if (IS_ERR_OR_NULL(hrec->record))
 		return;
 
-	map_value = htab_elem_value(elem, hrec->key_size);
+	map_value = (void *)elem + sizeof(struct htab_elem) + round_up(hrec->key_size, 8);
 	bpf_obj_free_fields(hrec->record, map_value);
 }
 
@@ -583,8 +629,14 @@ static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
 
 	htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
 
-	htab->elem_size = sizeof(struct htab_elem) +
-			  round_up(htab->map.key_size, 8);
+	if (htab_is_lru(htab))
+		htab->key_offset = offsetof(struct htab_elem_lru, key);
+	else if (percpu && !prealloc)
+		htab->key_offset = offsetof(struct htab_elem_pcpu, key);
+	else
+		htab->key_offset = offsetof(struct htab_elem, key);
+
+	htab->elem_size = htab->key_offset + round_up(htab->map.key_size, 8);
 	if (percpu)
 		htab->elem_size += sizeof(void *);
 	else
@@ -692,14 +744,16 @@ static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32
 }
 
 /* this lookup function can only be called with bucket lock taken */
-static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
+static struct htab_elem *lookup_elem_raw(struct bpf_htab *htab,
+					 struct hlist_nulls_head *head, u32 hash,
 					 void *key, u32 key_size)
 {
 	struct hlist_nulls_node *n;
 	struct htab_elem *l;
 
-	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
-		if (l->hash == hash && !memcmp(&l->key, key, key_size))
+	hlist_nulls_for_each_entry_rcu(l, n, head, node.hash_node)
+		if (htab_elem_hash(htab, l) == hash &&
+		    !memcmp(htab_elem_key(htab, l), key, key_size))
 			return l;
 
 	return NULL;
@@ -709,7 +763,8 @@ static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash
  * the unlikely event when elements moved from one bucket into another
  * while link list is being walked
  */
-static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
+static struct htab_elem *lookup_nulls_elem_raw(struct bpf_htab *htab,
+					       struct hlist_nulls_head *head,
 					       u32 hash, void *key,
 					       u32 key_size, u32 n_buckets)
 {
@@ -717,8 +772,9 @@ static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
 	struct htab_elem *l;
 
 again:
-	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
-		if (l->hash == hash && !memcmp(&l->key, key, key_size))
+	hlist_nulls_for_each_entry_rcu(l, n, head, node.hash_node)
+		if (htab_elem_hash(htab, l) == hash &&
+		    !memcmp(htab_elem_key(htab, l), key, key_size))
 			return l;
 
 	if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
@@ -747,17 +803,18 @@ static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
 
 	head = select_bucket(htab, hash);
 
-	l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
+	l = lookup_nulls_elem_raw(htab, head, hash, key, key_size, htab->n_buckets);
 
 	return l;
 }
 
 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
 {
+	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
 
 	if (l)
-		return htab_elem_value(l, map->key_size);
+		return htab_elem_value(htab, l);
 
 	return NULL;
 }
@@ -775,6 +832,7 @@ static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
  */
 static int htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 {
+	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 	struct bpf_insn *insn = insn_buf;
 	const int ret = BPF_REG_0;
 
@@ -783,7 +841,7 @@ static int htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 	*insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
 	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
 	*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
-				offsetof(struct htab_elem, key) +
+				htab->key_offset +
 				round_up(map->key_size, 8));
 	return insn - insn_buf;
 }
@@ -791,12 +849,13 @@ static int htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
 							void *key, const bool mark)
 {
+	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
 
 	if (l) {
 		if (mark)
-			bpf_lru_node_set_ref(&l->lru_node);
-		return htab_elem_value(l, map->key_size);
+			bpf_lru_node_set_ref(htab_elem_lru_node(l));
+		return htab_elem_value(htab, l);
 	}
 
 	return NULL;
@@ -815,6 +874,7 @@ static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
 static int htab_lru_map_gen_lookup(struct bpf_map *map,
 				   struct bpf_insn *insn_buf)
 {
+	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 	struct bpf_insn *insn = insn_buf;
 	const int ret = BPF_REG_0;
 	const int ref_reg = BPF_REG_1;
@@ -824,15 +884,15 @@ static int htab_lru_map_gen_lookup(struct bpf_map *map,
 	*insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
 	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
 	*insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
-			      offsetof(struct htab_elem, lru_node) +
+			      offsetof(struct htab_elem_lru, lru_node) +
 			      offsetof(struct bpf_lru_node, ref));
 	*insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
 	*insn++ = BPF_ST_MEM(BPF_B, ret,
-			     offsetof(struct htab_elem, lru_node) +
+			     offsetof(struct htab_elem_lru, lru_node) +
 			     offsetof(struct bpf_lru_node, ref),
 			     1);
 	*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
-				offsetof(struct htab_elem, key) +
+				htab->key_offset +
 				round_up(map->key_size, 8));
 	return insn - insn_buf;
 }
@@ -844,13 +904,13 @@ static void check_and_cancel_fields(struct bpf_htab *htab,
 		return;
 
 	if (htab_is_percpu(htab)) {
-		void __percpu *pptr = htab_elem_get_ptr(elem, htab->map.key_size);
+		void __percpu *pptr = htab_elem_get_ptr(htab, elem);
 		int cpu;
 
 		for_each_possible_cpu(cpu)
 			bpf_obj_cancel_fields(&htab->map, per_cpu_ptr(pptr, cpu));
 	} else {
-		void *map_value = htab_elem_value(elem, htab->map.key_size);
+		void *map_value = htab_elem_value(htab, elem);
 
 		bpf_obj_cancel_fields(&htab->map, map_value);
 	}
@@ -869,17 +929,17 @@ static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
 	struct bucket *b;
 	int ret;
 
-	tgt_l = container_of(node, struct htab_elem, lru_node);
-	b = __select_bucket(htab, tgt_l->hash);
+	tgt_l = (struct htab_elem *)container_of(node, struct htab_elem_lru, lru_node);
+	b = __select_bucket(htab, htab_elem_hash(htab, tgt_l));
 	head = &b->head;
 
 	ret = htab_lock_bucket(b, &flags);
 	if (ret)
 		return false;
 
-	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
+	hlist_nulls_for_each_entry_rcu(l, n, head, node.hash_node)
 		if (l == tgt_l) {
-			hlist_nulls_del_rcu(&l->hash_node);
+			hlist_nulls_del_rcu(&l->node.hash_node);
 			bpf_map_dec_elem_count(&htab->map);
 			break;
 		}
@@ -912,18 +972,19 @@ static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 	head = select_bucket(htab, hash);
 
 	/* lookup the key */
-	l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
+	l = lookup_nulls_elem_raw(htab, head, hash, key, key_size, htab->n_buckets);
 
 	if (!l)
 		goto find_first_elem;
 
 	/* key was found, get next key in the same bucket */
-	next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
-				  struct htab_elem, hash_node);
+	next_l = hlist_nulls_entry_safe(
+			rcu_dereference_raw(hlist_nulls_next_rcu(&l->node.hash_node)),
+			struct htab_elem, node.hash_node);
 
 	if (next_l) {
 		/* if next elem in this hash list is non-zero, just return it */
-		memcpy(next_key, next_l->key, key_size);
+		memcpy(next_key, htab_elem_key(htab, next_l), key_size);
 		return 0;
 	}
 
@@ -938,10 +999,10 @@ static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 
 		/* pick first element in the bucket */
 		next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
-					  struct htab_elem, hash_node);
+					  struct htab_elem, node.hash_node);
 		if (next_l) {
 			/* if it's not empty, just return it */
-			memcpy(next_key, next_l->key, key_size);
+			memcpy(next_key, htab_elem_key(htab, next_l), key_size);
 			return 0;
 		}
 	}
@@ -955,7 +1016,7 @@ static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
 	check_and_cancel_fields(htab, l);
 
 	if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
-		bpf_mem_cache_free(&htab->pcpu_ma, l->ptr_to_pptr);
+		bpf_mem_cache_free(&htab->pcpu_ma, htab_elem_get_ptr_to_pptr(l));
 	bpf_mem_cache_free(&htab->ma, l);
 }
 
@@ -965,7 +1026,7 @@ static void htab_put_fd_value(struct bpf_htab *htab, struct htab_elem *l)
 	void *ptr;
 
 	if (map->ops->map_fd_put_ptr) {
-		ptr = fd_htab_map_get_ptr(map, l);
+		ptr = fd_htab_map_get_ptr(htab, l);
 		map->ops->map_fd_put_ptr(map, ptr, true);
 	}
 }
@@ -1006,7 +1067,7 @@ static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
 	if (htab_is_prealloc(htab)) {
 		bpf_map_dec_elem_count(&htab->map);
 		check_and_cancel_fields(htab, l);
-		pcpu_freelist_push(&htab->freelist, &l->fnode);
+		pcpu_freelist_push(&htab->freelist, &l->node.fnode);
 	} else {
 		dec_elem_count(htab);
 		htab_elem_free(htab, l);
@@ -1097,7 +1158,7 @@ static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
 			l = __pcpu_freelist_pop(&htab->freelist);
 			if (!l)
 				return ERR_PTR(-E2BIG);
-			l_new = container_of(l, struct htab_elem, fnode);
+			l_new = container_of(l, struct htab_elem, node.fnode);
 			bpf_map_inc_elem_count(&htab->map);
 		}
 	} else {
@@ -1117,10 +1178,10 @@ static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
 		}
 	}
 
-	memcpy(l_new->key, key, key_size);
+	memcpy(htab_elem_key(htab, l_new), key, key_size);
 	if (percpu) {
 		if (prealloc) {
-			pptr = htab_elem_get_ptr(l_new, key_size);
+			pptr = htab_elem_get_ptr(htab, l_new);
 		} else {
 			/* alloc_percpu zero-fills */
 			void *ptr = bpf_mem_cache_alloc(&htab->pcpu_ma);
@@ -1130,26 +1191,26 @@ static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
 				l_new = ERR_PTR(-ENOMEM);
 				goto dec_count;
 			}
-			l_new->ptr_to_pptr = ptr;
+			htab_elem_set_ptr_to_pptr(l_new, ptr);
 			pptr = *(void __percpu **)ptr;
 		}
 
 		pcpu_init_value(htab, pptr, value, onallcpus, map_flags);
 
 		if (!prealloc)
-			htab_elem_set_ptr(l_new, key_size, pptr);
+			htab_elem_set_ptr(htab, l_new, pptr);
 	} else if (fd_htab_map_needs_adjust(htab)) {
 		size = round_up(size, 8);
-		memcpy(htab_elem_value(l_new, key_size), value, size);
+		memcpy(htab_elem_value(htab, l_new), value, size);
 	} else if (map_flags & BPF_F_LOCK) {
 		copy_map_value_locked(&htab->map,
-				      htab_elem_value(l_new, key_size),
+				      htab_elem_value(htab, l_new),
 				      value, false);
 	} else {
-		copy_map_value(&htab->map, htab_elem_value(l_new, key_size), value);
+		copy_map_value(&htab->map, htab_elem_value(htab, l_new), value);
 	}
 
-	l_new->hash = hash;
+	htab_elem_set_hash(htab, l_new, hash);
 	return l_new;
 dec_count:
 	dec_elem_count(htab);
@@ -1199,7 +1260,7 @@ static long htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 		if (unlikely(!btf_record_has_field(map->record, BPF_SPIN_LOCK)))
 			return -EINVAL;
 		/* find an element without taking the bucket lock */
-		l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
+		l_old = lookup_nulls_elem_raw(htab, head, hash, key, key_size,
 					      htab->n_buckets);
 		ret = check_flags(htab, l_old, map_flags);
 		if (ret)
@@ -1207,7 +1268,7 @@ static long htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 		if (l_old) {
 			/* grab the element lock and update value in place */
 			copy_map_value_locked(map,
-					      htab_elem_value(l_old, key_size),
+					      htab_elem_value(htab, l_old),
 					      value, false);
 			return 0;
 		}
@@ -1221,7 +1282,7 @@ static long htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 	if (ret)
 		return ret;
 
-	l_old = lookup_elem_raw(head, hash, key, key_size);
+	l_old = lookup_elem_raw(htab, head, hash, key, key_size);
 
 	ret = check_flags(htab, l_old, map_flags);
 	if (ret)
@@ -1235,7 +1296,7 @@ static long htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 		 * and update element in place
 		 */
 		copy_map_value_locked(map,
-				      htab_elem_value(l_old, key_size),
+				      htab_elem_value(htab, l_old),
 				      value, false);
 		ret = 0;
 		goto err;
@@ -1252,9 +1313,9 @@ static long htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 	/* add new element to the head of the list, so that
 	 * concurrent search will find it before old elem
 	 */
-	hlist_nulls_add_head_rcu(&l_new->hash_node, head);
+	hlist_nulls_add_head_rcu(&l_new->node.hash_node, head);
 	if (l_old) {
-		hlist_nulls_del_rcu(&l_old->hash_node);
+		hlist_nulls_del_rcu(&l_old->node.hash_node);
 
 		/* l_old has already been stashed in htab->extra_elems, cancel
 		 * its reusable special fields before it is available for reuse.
@@ -1275,7 +1336,7 @@ static void htab_lru_push_free(struct bpf_htab *htab, struct htab_elem *elem)
 {
 	check_and_cancel_fields(htab, elem);
 	bpf_map_dec_elem_count(&htab->map);
-	bpf_lru_push_free(&htab->lru, &elem->lru_node);
+	bpf_lru_push_free(&htab->lru, htab_elem_lru_node(elem));
 }
 
 static long htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
@@ -1310,13 +1371,13 @@ static long htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value
 	l_new = prealloc_lru_pop(htab, key, hash);
 	if (!l_new)
 		return -ENOMEM;
-	copy_map_value(&htab->map, htab_elem_value(l_new, map->key_size), value);
+	copy_map_value(&htab->map, htab_elem_value(htab, l_new), value);
 
 	ret = htab_lock_bucket(b, &flags);
 	if (ret)
 		goto err_lock_bucket;
 
-	l_old = lookup_elem_raw(head, hash, key, key_size);
+	l_old = lookup_elem_raw(htab, head, hash, key, key_size);
 
 	ret = check_flags(htab, l_old, map_flags);
 	if (ret)
@@ -1325,10 +1386,10 @@ static long htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value
 	/* add new element to the head of the list, so that
 	 * concurrent search will find it before old elem
 	 */
-	hlist_nulls_add_head_rcu(&l_new->hash_node, head);
+	hlist_nulls_add_head_rcu(&l_new->node.hash_node, head);
 	if (l_old) {
-		bpf_lru_node_set_ref(&l_new->lru_node);
-		hlist_nulls_del_rcu(&l_old->hash_node);
+		bpf_lru_node_set_ref(htab_elem_lru_node(l_new));
+		hlist_nulls_del_rcu(&l_old->node.hash_node);
 	}
 	ret = 0;
 
@@ -1383,7 +1444,7 @@ static long htab_map_update_elem_in_place(struct bpf_map *map, void *key,
 	if (ret)
 		return ret;
 
-	l_old = lookup_elem_raw(head, hash, key, key_size);
+	l_old = lookup_elem_raw(htab, head, hash, key, key_size);
 
 	ret = check_flags(htab, l_old, map_flags);
 	if (ret)
@@ -1392,10 +1453,10 @@ static long htab_map_update_elem_in_place(struct bpf_map *map, void *key,
 	if (l_old) {
 		/* Update value in-place */
 		if (percpu) {
-			pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
+			pcpu_copy_value(htab, htab_elem_get_ptr(htab, l_old),
 					value, onallcpus, map_flags);
 		} else {
-			void **inner_map_pptr = htab_elem_value(l_old, key_size);
+			void **inner_map_pptr = htab_elem_value(htab, l_old);
 
 			old_map_ptr = *inner_map_pptr;
 			WRITE_ONCE(*inner_map_pptr, *(void **)value);
@@ -1407,7 +1468,7 @@ static long htab_map_update_elem_in_place(struct bpf_map *map, void *key,
 			ret = PTR_ERR(l_new);
 			goto err;
 		}
-		hlist_nulls_add_head_rcu(&l_new->hash_node, head);
+		hlist_nulls_add_head_rcu(&l_new->node.hash_node, head);
 	}
 err:
 	htab_unlock_bucket(b, flags);
@@ -1456,22 +1517,22 @@ static long __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
 	if (ret)
 		goto err_lock_bucket;
 
-	l_old = lookup_elem_raw(head, hash, key, key_size);
+	l_old = lookup_elem_raw(htab, head, hash, key, key_size);
 
 	ret = check_flags(htab, l_old, map_flags);
 	if (ret)
 		goto err;
 
 	if (l_old) {
-		bpf_lru_node_set_ref(&l_old->lru_node);
+		bpf_lru_node_set_ref(htab_elem_lru_node(l_old));
 
 		/* per-cpu hash map can update value in-place */
-		pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
+		pcpu_copy_value(htab, htab_elem_get_ptr(htab, l_old),
 				value, onallcpus, map_flags);
 	} else {
-		pcpu_init_value(htab, htab_elem_get_ptr(l_new, key_size),
+		pcpu_init_value(htab, htab_elem_get_ptr(htab, l_new),
 				value, onallcpus, map_flags);
-		hlist_nulls_add_head_rcu(&l_new->hash_node, head);
+		hlist_nulls_add_head_rcu(&l_new->node.hash_node, head);
 		l_new = NULL;
 	}
 	ret = 0;
@@ -1480,7 +1541,7 @@ static long __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
 err_lock_bucket:
 	if (l_new) {
 		bpf_map_dec_elem_count(&htab->map);
-		bpf_lru_push_free(&htab->lru, &l_new->lru_node);
+		bpf_lru_push_free(&htab->lru, htab_elem_lru_node(l_new));
 	}
 	return ret;
 }
@@ -1521,9 +1582,9 @@ static long htab_map_delete_elem(struct bpf_map *map, void *key)
 	if (ret)
 		return ret;
 
-	l = lookup_elem_raw(head, hash, key, key_size);
+	l = lookup_elem_raw(htab, head, hash, key, key_size);
 	if (l)
-		hlist_nulls_del_rcu(&l->hash_node);
+		hlist_nulls_del_rcu(&l->node.hash_node);
 	else
 		ret = -ENOENT;
 
@@ -1556,10 +1617,10 @@ static long htab_lru_map_delete_elem(struct bpf_map *map, void *key)
 	if (ret)
 		return ret;
 
-	l = lookup_elem_raw(head, hash, key, key_size);
+	l = lookup_elem_raw(htab, head, hash, key, key_size);
 
 	if (l)
-		hlist_nulls_del_rcu(&l->hash_node);
+		hlist_nulls_del_rcu(&l->node.hash_node);
 	else
 		ret = -ENOENT;
 
@@ -1581,8 +1642,8 @@ static void delete_all_elements(struct bpf_htab *htab)
 		struct hlist_nulls_node *n;
 		struct htab_elem *l;
 
-		hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
-			hlist_nulls_del_rcu(&l->hash_node);
+		hlist_nulls_for_each_entry_safe(l, n, head, node.hash_node) {
+			hlist_nulls_del_rcu(&l->node.hash_node);
 			htab_elem_free(htab, l);
 		}
 		cond_resched();
@@ -1599,10 +1660,10 @@ static void htab_free_malloced_internal_structs(struct bpf_htab *htab)
 		struct hlist_nulls_node *n;
 		struct htab_elem *l;
 
-		hlist_nulls_for_each_entry(l, n, head, hash_node) {
+		hlist_nulls_for_each_entry(l, n, head, node.hash_node) {
 			/* We only free internal structs on uref dropping to zero */
 			bpf_map_free_internal_structs(&htab->map,
-						      htab_elem_value(l, htab->map.key_size));
+						      htab_elem_value(htab, l));
 		}
 		cond_resched_rcu();
 	}
@@ -1697,7 +1758,7 @@ static int __htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
 	if (ret)
 		return ret;
 
-	l = lookup_elem_raw(head, hash, key, key_size);
+	l = lookup_elem_raw(htab, head, hash, key, key_size);
 	if (!l) {
 		ret = -ENOENT;
 		goto out_unlock;
@@ -1708,14 +1769,14 @@ static int __htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
 		void __percpu *pptr;
 		int off = 0, cpu;
 
-		pptr = htab_elem_get_ptr(l, key_size);
+		pptr = htab_elem_get_ptr(htab, l);
 		for_each_possible_cpu(cpu) {
 			copy_map_value_long(&htab->map, value + off, per_cpu_ptr(pptr, cpu));
 			check_and_init_map_value(&htab->map, value + off);
 			off += roundup_value_size;
 		}
 	} else {
-		void *src = htab_elem_value(l, map->key_size);
+		void *src = htab_elem_value(htab, l);
 
 		if (flags & BPF_F_LOCK)
 			copy_map_value_locked(map, value, src, true);
@@ -1724,7 +1785,7 @@ static int __htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
 		/* Zeroing special fields in the temp buffer */
 		check_and_init_map_value(map, value);
 	}
-	hlist_nulls_del_rcu(&l->hash_node);
+	hlist_nulls_del_rcu(&l->node.hash_node);
 
 out_unlock:
 	htab_unlock_bucket(b, bflags);
@@ -1860,7 +1921,7 @@ __htab_map_lookup_and_delete_batch(struct bpf_map *map,
 	}
 
 	bucket_cnt = 0;
-	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
+	hlist_nulls_for_each_entry_rcu(l, n, head, node.hash_node)
 		bucket_cnt++;
 
 	if (bucket_cnt && !locked) {
@@ -1897,14 +1958,14 @@ __htab_map_lookup_and_delete_batch(struct bpf_map *map,
 	if (!locked)
 		goto next_batch;
 
-	hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
-		memcpy(dst_key, l->key, key_size);
+	hlist_nulls_for_each_entry_safe(l, n, head, node.hash_node) {
+		memcpy(dst_key, htab_elem_key(htab, l), key_size);
 
 		if (is_percpu) {
 			int off = 0, cpu;
 			void __percpu *pptr;
 
-			pptr = htab_elem_get_ptr(l, map->key_size);
+			pptr = htab_elem_get_ptr(htab, l);
 			if (elem_map_flags & BPF_F_CPU) {
 				cpu = elem_map_flags >> 32;
 				copy_map_value(&htab->map, dst_val, per_cpu_ptr(pptr, cpu));
@@ -1918,7 +1979,7 @@ __htab_map_lookup_and_delete_batch(struct bpf_map *map,
 				}
 			}
 		} else {
-			value = htab_elem_value(l, key_size);
+			value = htab_elem_value(htab, l);
 			if (is_fd_htab(htab)) {
 				struct bpf_map **inner_map = value;
 
@@ -1936,7 +1997,7 @@ __htab_map_lookup_and_delete_batch(struct bpf_map *map,
 			check_and_init_map_value(map, dst_val);
 		}
 		if (do_delete) {
-			hlist_nulls_del_rcu(&l->hash_node);
+			hlist_nulls_del_rcu(&l->node.hash_node);
 
 			/* bpf_lru_push_free() will acquire lru_lock, which
 			 * may cause deadlock. See comments in function
@@ -1948,7 +2009,7 @@ __htab_map_lookup_and_delete_batch(struct bpf_map *map,
 			 * lock being held and it violates the lock rule, so
 			 * invoke free_htab_elem() after unlock as well.
 			 */
-			l->batch_flink = node_to_free;
+			l->node.batch_flink = node_to_free;
 			node_to_free = l;
 		}
 		dst_key += key_size;
@@ -1960,7 +2021,7 @@ __htab_map_lookup_and_delete_batch(struct bpf_map *map,
 
 	while (node_to_free) {
 		l = node_to_free;
-		node_to_free = node_to_free->batch_flink;
+		node_to_free = node_to_free->node.batch_flink;
 		if (is_lru_map)
 			htab_lru_push_free(htab, l);
 		else
@@ -2108,8 +2169,8 @@ bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
 		/* no update/deletion on this bucket, prev_elem should be still valid
 		 * and we won't skip elements.
 		 */
-		n = rcu_dereference_raw(hlist_nulls_next_rcu(&prev_elem->hash_node));
-		elem = hlist_nulls_entry_safe(n, struct htab_elem, hash_node);
+		n = rcu_dereference_raw(hlist_nulls_next_rcu(&prev_elem->node.hash_node));
+		elem = hlist_nulls_entry_safe(n, struct htab_elem, node.hash_node);
 		if (elem)
 			return elem;
 
@@ -2125,7 +2186,7 @@ bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
 
 		count = 0;
 		head = &b->head;
-		hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
+		hlist_nulls_for_each_entry_rcu(elem, n, head, node.hash_node) {
 			if (count >= skip_elems) {
 				info->bucket_id = i;
 				info->skip_elems = count;
@@ -2183,12 +2244,12 @@ static int __bpf_hash_map_seq_show(struct seq_file *seq, struct htab_elem *elem)
 		ctx.meta = &meta;
 		ctx.map = info->map;
 		if (elem) {
-			ctx.key = elem->key;
+			ctx.key = htab_elem_key(info->htab, elem);
 			if (!info->percpu_value_buf) {
-				ctx.value = htab_elem_value(elem, map->key_size);
+				ctx.value = htab_elem_value(info->htab, elem);
 			} else {
 				roundup_value_size = round_up(map->value_size, 8);
-				pptr = htab_elem_get_ptr(elem, map->key_size);
+				pptr = htab_elem_get_ptr(info->htab, elem);
 				for_each_possible_cpu(cpu) {
 					copy_map_value_long(map, info->percpu_value_buf + off,
 							    per_cpu_ptr(pptr, cpu));
@@ -2292,14 +2353,14 @@ static long bpf_for_each_hash_elem(struct bpf_map *map, bpf_callback_t callback_
 		b = &htab->buckets[i];
 		rcu_read_lock();
 		head = &b->head;
-		hlist_nulls_for_each_entry_safe(elem, n, head, hash_node) {
-			key = elem->key;
+		hlist_nulls_for_each_entry_safe(elem, n, head, node.hash_node) {
+			key = htab_elem_key(htab, elem);
 			if (is_percpu) {
 				/* current cpu value for percpu map */
-				pptr = htab_elem_get_ptr(elem, map->key_size);
+				pptr = htab_elem_get_ptr(htab, elem);
 				val = this_cpu_ptr(pptr);
 			} else {
-				val = htab_elem_value(elem, map->key_size);
+				val = htab_elem_value(htab, elem);
 			}
 			num_elems++;
 			ret = callback_fn((u64)(long)map, (u64)(long)key,
@@ -2403,10 +2464,11 @@ const struct bpf_map_ops htab_lru_map_ops = {
 /* Called from eBPF program */
 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
 {
+	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
 
 	if (l)
-		return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
+		return this_cpu_ptr(htab_elem_get_ptr(htab, l));
 	else
 		return NULL;
 }
@@ -2414,6 +2476,7 @@ static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
 /* inline bpf_map_lookup_elem() call for per-CPU hashmap */
 static int htab_percpu_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 {
+	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 	struct bpf_insn *insn = insn_buf;
 
 	if (!bpf_jit_supports_percpu_insn())
@@ -2424,7 +2487,7 @@ static int htab_percpu_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn
 	*insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
 	*insn++ = BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 3);
 	*insn++ = BPF_ALU64_IMM(BPF_ADD, BPF_REG_0,
-				offsetof(struct htab_elem, key) + roundup(map->key_size, 8));
+				htab->key_offset + roundup(map->key_size, 8));
 	*insn++ = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_0, 0);
 	*insn++ = BPF_MOV64_PERCPU_REG(BPF_REG_0, BPF_REG_0);
 
@@ -2433,6 +2496,7 @@ static int htab_percpu_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn
 
 static void *htab_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
 {
+	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 	struct htab_elem *l;
 
 	if (cpu >= nr_cpu_ids)
@@ -2440,18 +2504,19 @@ static void *htab_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *key,
 
 	l = __htab_map_lookup_elem(map, key);
 	if (l)
-		return per_cpu_ptr(htab_elem_get_ptr(l, map->key_size), cpu);
+		return per_cpu_ptr(htab_elem_get_ptr(htab, l), cpu);
 	else
 		return NULL;
 }
 
 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
 {
+	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
 
 	if (l) {
-		bpf_lru_node_set_ref(&l->lru_node);
-		return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
+		bpf_lru_node_set_ref(htab_elem_lru_node(l));
+		return this_cpu_ptr(htab_elem_get_ptr(htab, l));
 	}
 
 	return NULL;
@@ -2459,6 +2524,7 @@ static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
 
 static void *htab_lru_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
 {
+	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 	struct htab_elem *l;
 
 	if (cpu >= nr_cpu_ids)
@@ -2466,8 +2532,8 @@ static void *htab_lru_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *k
 
 	l = __htab_map_lookup_elem(map, key);
 	if (l) {
-		bpf_lru_node_set_ref(&l->lru_node);
-		return per_cpu_ptr(htab_elem_get_ptr(l, map->key_size), cpu);
+		bpf_lru_node_set_ref(htab_elem_lru_node(l));
+		return per_cpu_ptr(htab_elem_get_ptr(htab, l), cpu);
 	}
 
 	return NULL;
@@ -2475,6 +2541,7 @@ static void *htab_lru_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *k
 
 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value, u64 map_flags)
 {
+	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 	struct htab_elem *l;
 	void __percpu *pptr;
 	int ret = -ENOENT;
@@ -2494,7 +2561,7 @@ int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value, u64 map_fl
 	/* 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);
+	pptr = htab_elem_get_ptr(htab, l);
 	if (map_flags & BPF_F_CPU) {
 		cpu = map_flags >> 32;
 		copy_map_value(map, value, per_cpu_ptr(pptr, cpu));
@@ -2532,6 +2599,7 @@ int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
 					  struct seq_file *m)
 {
+	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 	struct htab_elem *l;
 	void __percpu *pptr;
 	int cpu;
@@ -2546,7 +2614,7 @@ static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
 
 	btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
 	seq_puts(m, ": {\n");
-	pptr = htab_elem_get_ptr(l, map->key_size);
+	pptr = htab_elem_get_ptr(htab, l);
 	for_each_possible_cpu(cpu) {
 		seq_printf(m, "\tcpu%d: ", cpu);
 		btf_type_seq_show(map->btf, map->btf_value_type_id,
@@ -2619,8 +2687,8 @@ static void fd_htab_map_free(struct bpf_map *map)
 	for (i = 0; i < htab->n_buckets; i++) {
 		head = select_bucket(htab, i);
 
-		hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
-			void *ptr = fd_htab_map_get_ptr(map, l);
+		hlist_nulls_for_each_entry_safe(l, n, head, node.hash_node) {
+			void *ptr = fd_htab_map_get_ptr(htab, l);
 
 			map->ops->map_fd_put_ptr(map, ptr, false);
 		}
@@ -2705,6 +2773,7 @@ static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
 static int htab_of_map_gen_lookup(struct bpf_map *map,
 				  struct bpf_insn *insn_buf)
 {
+	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 	struct bpf_insn *insn = insn_buf;
 	const int ret = BPF_REG_0;
 
@@ -2713,7 +2782,7 @@ static int htab_of_map_gen_lookup(struct bpf_map *map,
 	*insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
 	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
 	*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
-				offsetof(struct htab_elem, key) +
+				htab->key_offset +
 				round_up(map->key_size, 8));
 	*insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
 
@@ -3533,3 +3602,16 @@ const struct bpf_map_ops rhtab_map_ops = {
 	.map_btf_id = &rhtab_map_btf_ids[0],
 	.iter_seq_info = &rhash_iter_seq_info,
 };
+
+size_t bpf_htab_map_meta_size(void)
+{
+	return sizeof(struct bpf_htab);
+}
+
+void bpf_htab_map_meta_init(struct bpf_map *inner_map_meta, struct bpf_map *inner_map)
+{
+	struct bpf_htab *inner_htab_meta = container_of(inner_map_meta, struct bpf_htab, map);
+	struct bpf_htab *inner_htab = container_of(inner_map, struct bpf_htab, map);
+
+	inner_htab_meta->key_offset = inner_htab->key_offset;
+}
diff --git a/kernel/bpf/map_in_map.c b/kernel/bpf/map_in_map.c
index d2cbab4bdf64..b2db5d0c98be 100644
--- a/kernel/bpf/map_in_map.c
+++ b/kernel/bpf/map_in_map.c
@@ -7,6 +7,15 @@
 
 #include "map_in_map.h"
 
+static bool bpf_map_type_is_htab(enum bpf_map_type type)
+{
+	return type == BPF_MAP_TYPE_HASH ||
+	       type == BPF_MAP_TYPE_PERCPU_HASH ||
+	       type == BPF_MAP_TYPE_LRU_HASH ||
+	       type == BPF_MAP_TYPE_LRU_PERCPU_HASH ||
+	       type == BPF_MAP_TYPE_HASH_OF_MAPS;
+}
+
 struct bpf_map *bpf_map_meta_alloc(int inner_map_ufd)
 {
 	struct bpf_map *inner_map, *inner_map_meta;
@@ -29,6 +38,8 @@ struct bpf_map *bpf_map_meta_alloc(int inner_map_ufd)
 	/* In some cases verifier needs to access beyond just base map. */
 	if (inner_map->ops == &array_map_ops || inner_map->ops == &percpu_array_map_ops)
 		inner_map_meta_size = sizeof(struct bpf_array);
+	else if (bpf_map_type_is_htab(inner_map->map_type))
+		inner_map_meta_size = bpf_htab_map_meta_size();
 
 	inner_map_meta = kzalloc(inner_map_meta_size, GFP_USER);
 	if (!inner_map_meta)
@@ -70,6 +81,8 @@ struct bpf_map *bpf_map_meta_alloc(int inner_map_ufd)
 		inner_array_meta->index_mask = inner_array->index_mask;
 		inner_array_meta->elem_size = inner_array->elem_size;
 		inner_map_meta->bypass_spec_v1 = inner_map->bypass_spec_v1;
+	} else if (bpf_map_type_is_htab(inner_map->map_type)) {
+		bpf_htab_map_meta_init(inner_map_meta, inner_map);
 	}
 	return inner_map_meta;
 }
diff --git a/kernel/bpf/map_in_map.h b/kernel/bpf/map_in_map.h
index 7d61602354de..1edc205772d8 100644
--- a/kernel/bpf/map_in_map.h
+++ b/kernel/bpf/map_in_map.h
@@ -15,5 +15,7 @@ void *bpf_map_fd_get_ptr(struct bpf_map *map, struct file *map_file,
 			 int ufd);
 void bpf_map_fd_put_ptr(struct bpf_map *map, void *ptr, bool need_defer);
 u32 bpf_map_fd_sys_lookup_elem(void *ptr);
+size_t bpf_htab_map_meta_size(void);
+void bpf_htab_map_meta_init(struct bpf_map *inner_map_meta, struct bpf_map *inner_map);
 
 #endif
-- 
2.55.0.691.gc56d675ccc-goog

Re: [PATCH bpf-next v4 1/2] bpf: htab: Split htab_elem_lru and htab_elem_pcpu off of htab_elem

[bot+bpf-ci]

[-- Attachment #1: Type: text/plain, Size: 4784 bytes --]

> @@ -476,7 +522,7 @@ static void htab_mem_dtor(void *obj, void *ctx)
>  	if (IS_ERR_OR_NULL(hrec->record))
>  		return;
>
> -	map_value = htab_elem_value(elem, hrec->key_size);
> +	map_value = (void *)elem + sizeof(struct htab_elem) + round_up(hrec->key_size, 8);
>  	bpf_obj_free_fields(hrec->record, map_value);
>  }

This isn't a bug, but would it be worth storing the element header offset
in struct htab_btf_record (next to key_size) and keeping htab_mem_dtor on
the shared htab_elem_value() path, rather than open-coding
sizeof(struct htab_elem) here? Everywhere else in the file the header
size is now the per-map htab->key_offset, reached through
htab_elem_key()/htab_elem_value(). The hardcoded form happens to be
correct today because htab_mem_dtor is only installed for non-preallocated,
non-percpu, non-LRU BPF_MAP_TYPE_HASH maps (LRU requires prealloc, and
HASH_OF_MAPS uses map_check_no_btf), which is exactly the case where
key_offset == sizeof(struct htab_elem). The invariant is invisible at the
call site and there is no comment or assertion tying the two together, so
the open-coded copy silently diverges from the accessor if another element
layout ever registers this dtor.

> @@ -869,17 +929,17 @@ static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
>  	struct bucket *b;
>  	int ret;
>
> -	tgt_l = container_of(node, struct htab_elem, lru_node);
> -	b = __select_bucket(htab, tgt_l->hash);
> +	tgt_l = (struct htab_elem *)container_of(node, struct htab_elem_lru, lru_node);
> +	b = __select_bucket(htab, htab_elem_hash(htab, tgt_l));

This isn't a bug, but the commit adds htab_elem_lru_node() for the
htab_elem to bpf_lru_node direction, while the inverse conversion stays
open-coded and is now duplicated verbatim at two sites (line 358 in
prealloc_lru_pop() and line 932 here), each needing an explicit
(struct htab_elem *) cast to silence the incompatible-pointer warning.
Would a small counterpart helper, e.g. htab_elem_from_lru_node(), keep
both directions symmetric and confine the cast to one place?

Commit message issue:

>     bpf: htab: Split htab_elem_lru and htab_elem_pcpu off of htab_elem
>
>     The htab_elem struct is used as the per-element type for all BPF hash
>     map types and includes bpf_lru_node in a union with a ptr_to_pptr
>     pointer. For standard (non-LRU, non-PCPU) hash maps, the 24 byte union
>     allocated for every element is entirely unused. For non-preallocated
>     PCPU maps, ptr_to_pptr only requires 8 bytes, leaving 16 bytes of unused
>     overhead in the union. For preallocated PCPU maps ptr_to_pptr is unused
>     since elements are freed to the PCPU freelist.
>
>     Eliminate this per-element memory overhead by splitting htab_elem into
>     dedicated structures for each map type:
>        - struct htab_elem: Minimal structure for standard hash maps and
>          preallocated PCPU maps (saves 24 bytes per element).
>        - struct htab_elem_pcpu: Structure for non-preallocated PCPU maps
>          containing ptr_to_pptr (saves 16 bytes per element).
>        - struct htab_elem_lru: Retains struct bpf_lru_node for LRU maps.
>
>     Because element sizes now vary by map type, add key_offset to struct
>     bpf_htab to track the dynamic key offset. Update helper accessors and
>     lookups to compute key and value offsets using htab->key_offset.
>
>     Pointers to struct htab_elem in the existing code (e.g. htab_elem_hash)
>     serve as generic base element pointers. This is possible because
>     htab_elem, htab_elem_pcpu, and htab_elem_lru share a common initial
>     sequence, making pointer casts safe.

This isn't a bug, but could the changelog gain a paragraph on the
map_in_map.c side? The patch also touches kernel/bpf/map_in_map.c and
kernel/bpf/map_in_map.h: it adds bpf_map_type_is_htab(), grows
inner_map_meta to sizeof(struct bpf_htab) via the new
bpf_htab_map_meta_size(), and copies key_offset into the meta map with the
new bpf_htab_map_meta_init(). That part is the non-obvious
correctness-critical consequence of making the key offset a runtime field,
since the verifier's inlined lookups read the offset out of the
inner_map_meta. A reviewer looking at the diff would most benefit from an
explanation of why inner_map_meta must now be sizeof(struct bpf_htab) and
have key_offset copied. Also, the sentence about making pointer casts safe
mentions the common initial sequence but doesn't name which casts rely on
it (htab_elem_lru_node(), htab_elem_get_ptr_to_pptr(),
htab_elem_set_ptr_to_pptr()).


---
AI reviewed your patch. Please fix the bug or email reply why it's not a bug.
See: https://github.com/kernel-patches/vmtest/blob/master/ci/claude/README.md

CI run summary: https://github.com/kernel-patches/bpf/actions/runs/31651309438

[PATCH bpf-next v4 2/2] bpf: htab: Reduce elem_size by 8 bytes for small key sizes

[T.J. Mercier]

For standard and PCPU (non-LRU) hash maps with small key sizes (less
than or equal to the word size), comparing keys requires only a single
instruction. Storing a cached 32-bit hash value to shortcut full key
comparisons provides no performance advantage for small keys, and
consumes memory for every element.

This memory can be saved by eliminating hash along with its associated
4 byte padding before the key, reducing the elem_size (and key_offset)
by 8 bytes for standard and PCPU maps.

Introduce htab_has_hash() to check whether a map requires a cached hash
field. Update htab_elem_set_hash(), lookup_elem_raw(), and
lookup_nulls_elem_raw() to conditionally bypass hash checking and
storage when htab_has_hash() is false.

Together with the previous patch, this reduces the minimum standard and
preallocated hash map element size from 64 bytes down to 32 bytes, and
non-preallocated per-CPU element size from 64 bytes down to 40 bytes.

Signed-off-by: T.J. Mercier <tjmercier@google.com>
---
 kernel/bpf/hashtab.c                          | 104 +++++++++++++-----
 .../selftests/bpf/progs/map_ptr_kern.c        |   2 +-
 2 files changed, 76 insertions(+), 30 deletions(-)

diff --git a/kernel/bpf/hashtab.c b/kernel/bpf/hashtab.c
index fed4b1db8b78..720766121ea2 100644
--- a/kernel/bpf/hashtab.c
+++ b/kernel/bpf/hashtab.c
@@ -100,6 +100,7 @@ struct bpf_htab {
 	struct percpu_counter pcount;
 	atomic_t count;
 	bool use_percpu_counter;
+	bool has_hash;
 	u32 n_buckets;	/* number of hash buckets */
 	u32 elem_size;	/* size of each element in bytes */
 	u32 key_offset;	/* offset of key in bytes */
@@ -123,15 +124,13 @@ struct htab_node {
 
 struct htab_elem {
 	struct htab_node node;
-	u32 hash __aligned(8);
-	char key[] __aligned(8);
+	u8 data[] __aligned(8);
 };
 
 struct htab_elem_lru {
 	struct htab_node node;
 	struct bpf_lru_node lru_node;
-	u32 hash __aligned(8);
-	char key[] __aligned(8);
+	u8 data[] __aligned(8);
 };
 
 /* Only for non-preallocated PCPU maps. Preallocated PCPU maps don't need
@@ -140,13 +139,13 @@ struct htab_elem_lru {
 struct htab_elem_pcpu {
 	struct htab_node node;
 	void *ptr_to_pptr;
-	u32 hash __aligned(8);
-	char key[] __aligned(8);
+	u8 data[] __aligned(8);
 };
 
 struct htab_btf_record {
 	struct btf_record *record;
 	u32 key_size;
+	u32 key_offset;
 };
 
 static inline bool htab_is_prealloc(const struct bpf_htab *htab)
@@ -247,6 +246,11 @@ static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
 	return (struct htab_elem *) (htab->elems + i * (u64)htab->elem_size);
 }
 
+static bool htab_has_hash(const struct bpf_htab *htab)
+{
+	return htab->has_hash;
+}
+
 static u32 htab_elem_hash(struct bpf_htab *htab, struct htab_elem *l)
 {
 	return *(u32 *)(htab_elem_key(htab, l) - 8);
@@ -254,7 +258,8 @@ static u32 htab_elem_hash(struct bpf_htab *htab, struct htab_elem *l)
 
 static void htab_elem_set_hash(struct bpf_htab *htab, struct htab_elem *l, u32 hash)
 {
-	*(u32 *)(htab_elem_key(htab, l) - 8) = hash;
+	if (htab_has_hash(htab))
+		*(u32 *)(htab_elem_key(htab, l) - 8) = hash;
 }
 
 /* Both percpu and fd htab support in-place update, so no need for
@@ -395,7 +400,7 @@ static int prealloc_init(struct bpf_htab *htab)
 	if (htab_is_lru(htab))
 		err = bpf_lru_init(&htab->lru,
 				   htab->map.map_flags & BPF_F_NO_COMMON_LRU,
-				   offsetof(struct htab_elem_lru, hash) -
+				   offsetof(struct htab_elem_lru, data) -
 				   offsetof(struct htab_elem_lru, lru_node),
 				   htab_lru_map_delete_node,
 				   htab);
@@ -522,7 +527,7 @@ static void htab_mem_dtor(void *obj, void *ctx)
 	if (IS_ERR_OR_NULL(hrec->record))
 		return;
 
-	map_value = (void *)elem + sizeof(struct htab_elem) + round_up(hrec->key_size, 8);
+	map_value = (void *)elem + hrec->key_offset + round_up(hrec->key_size, 8);
 	bpf_obj_free_fields(hrec->record, map_value);
 }
 
@@ -548,7 +553,7 @@ static void htab_dtor_ctx_free(void *ctx)
 }
 
 static int bpf_ma_set_dtor(struct bpf_map *map, struct bpf_mem_alloc *ma,
-			   void (*dtor)(void *, void *))
+			   void (*dtor)(void *, void *), u32 key_offset)
 {
 	struct htab_btf_record *hrec;
 	int err;
@@ -561,6 +566,7 @@ static int bpf_ma_set_dtor(struct bpf_map *map, struct bpf_mem_alloc *ma,
 	if (!hrec)
 		return -ENOMEM;
 	hrec->key_size = map->key_size;
+	hrec->key_offset = key_offset;
 	hrec->record = btf_record_dup(map->record);
 	if (IS_ERR(hrec->record)) {
 		err = PTR_ERR(hrec->record);
@@ -583,9 +589,9 @@ static int htab_map_check_btf(struct bpf_map *map, const struct btf *btf,
 	 * populated in htab_map_alloc(), so it will always appear as NULL.
 	 */
 	if (htab_is_percpu(htab))
-		return bpf_ma_set_dtor(map, &htab->pcpu_ma, htab_pcpu_mem_dtor);
+		return bpf_ma_set_dtor(map, &htab->pcpu_ma, htab_pcpu_mem_dtor, htab->key_offset);
 	else
-		return bpf_ma_set_dtor(map, &htab->ma, htab_mem_dtor);
+		return bpf_ma_set_dtor(map, &htab->ma, htab_mem_dtor, htab->key_offset);
 }
 
 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
@@ -608,6 +614,13 @@ static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
 
 	bpf_map_init_from_attr(&htab->map, attr);
 
+	/* Avoid hash memory use and comparisons where unnecessary.
+	 * u32 hash reads are always atomic. If we elide them, key comparisons must also be atomic
+	 * to avoid false positive key matches due to torn key reads / writes. This is only possible
+	 * when the key fits within a word, so check key_size.
+	 */
+	htab->has_hash = htab_is_lru(htab) || htab->map.key_size > sizeof(unsigned long);
+
 	if (percpu_lru) {
 		/* ensure each CPU's lru list has >=1 elements.
 		 * since we are at it, make each lru list has the same
@@ -630,11 +643,13 @@ static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
 	htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
 
 	if (htab_is_lru(htab))
-		htab->key_offset = offsetof(struct htab_elem_lru, key);
+		htab->key_offset = offsetof(struct htab_elem_lru, data) + 8;
 	else if (percpu && !prealloc)
-		htab->key_offset = offsetof(struct htab_elem_pcpu, key);
+		htab->key_offset = offsetof(struct htab_elem_pcpu, data) +
+				   (htab_has_hash(htab) ? 8 : 0);
 	else
-		htab->key_offset = offsetof(struct htab_elem, key);
+		htab->key_offset = offsetof(struct htab_elem, data) +
+				   (htab_has_hash(htab) ? 8 : 0);
 
 	htab->elem_size = htab->key_offset + round_up(htab->map.key_size, 8);
 	if (percpu)
@@ -743,22 +758,44 @@ static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32
 	return &__select_bucket(htab, hash)->head;
 }
 
-/* this lookup function can only be called with bucket lock taken */
-static struct htab_elem *lookup_elem_raw(struct bpf_htab *htab,
-					 struct hlist_nulls_head *head, u32 hash,
-					 void *key, u32 key_size)
+static struct htab_elem *__lookup_elem_raw(struct bpf_htab *htab,
+					   struct hlist_nulls_head *head,
+					   u32 hash, void *key, u32 key_size,
+					   struct hlist_nulls_node **out_n)
 {
 	struct hlist_nulls_node *n;
 	struct htab_elem *l;
 
-	hlist_nulls_for_each_entry_rcu(l, n, head, node.hash_node)
-		if (htab_elem_hash(htab, l) == hash &&
-		    !memcmp(htab_elem_key(htab, l), key, key_size))
-			return l;
+	if (htab_has_hash(htab)) {
+		hlist_nulls_for_each_entry_rcu(l, n, head, node.hash_node)
+			if (htab_elem_hash(htab, l) == hash &&
+			    !memcmp(htab_elem_key(htab, l), key, key_size))
+				return l;
+	} else {
+		/* When hash is omitted, key comparisons must be atomic. Zero extend
+		 * the caller's key to the word size to support an atomic compare.
+		 */
+		unsigned long k = 0;
 
+		memcpy(&k, key, key_size);
+		hlist_nulls_for_each_entry_rcu(l, n, head, node.hash_node)
+			if (READ_ONCE(*(unsigned long *)htab_elem_key(htab, l)) == k)
+				return l;
+	}
+
+	if (out_n)
+		*out_n = n;
 	return NULL;
 }
 
+/* this lookup function can only be called with bucket lock taken */
+static struct htab_elem *lookup_elem_raw(struct bpf_htab *htab,
+					 struct hlist_nulls_head *head, u32 hash,
+					 void *key, u32 key_size)
+{
+	return __lookup_elem_raw(htab, head, hash, key, key_size, NULL);
+}
+
 /* can be called without bucket lock. it will repeat the loop in
  * the unlikely event when elements moved from one bucket into another
  * while link list is being walked
@@ -772,10 +809,9 @@ static struct htab_elem *lookup_nulls_elem_raw(struct bpf_htab *htab,
 	struct htab_elem *l;
 
 again:
-	hlist_nulls_for_each_entry_rcu(l, n, head, node.hash_node)
-		if (htab_elem_hash(htab, l) == hash &&
-		    !memcmp(htab_elem_key(htab, l), key, key_size))
-			return l;
+	l = __lookup_elem_raw(htab, head, hash, key, key_size, &n);
+	if (l)
+		return l;
 
 	if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
 		goto again;
@@ -1178,7 +1214,17 @@ static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
 		}
 	}
 
-	memcpy(htab_elem_key(htab, l_new), key, key_size);
+	if (htab_has_hash(htab)) {
+		memcpy(htab_elem_key(htab, l_new), key, key_size);
+	} else {
+		/* Zero-extend key into k for an atomic write to support
+		 * lockless RCU readers.
+		 */
+		unsigned long k = 0;
+
+		memcpy(&k, key, key_size);
+		WRITE_ONCE(*(unsigned long *)htab_elem_key(htab, l_new), k);
+	}
 	if (percpu) {
 		if (prealloc) {
 			pptr = htab_elem_get_ptr(htab, l_new);
@@ -3181,7 +3227,7 @@ static int rhtab_map_check_btf(struct bpf_map *map, const struct btf *btf,
 {
 	struct bpf_rhtab *rhtab = container_of(map, struct bpf_rhtab, map);
 
-	return bpf_ma_set_dtor(map, &rhtab->ma, rhtab_mem_dtor);
+	return bpf_ma_set_dtor(map, &rhtab->ma, rhtab_mem_dtor, offsetof(struct rhtab_elem, data));
 }
 
 static void rhtab_map_free_internal_structs(struct bpf_map *map)
diff --git a/tools/testing/selftests/bpf/progs/map_ptr_kern.c b/tools/testing/selftests/bpf/progs/map_ptr_kern.c
index 373c8d17ea55..6bd4cb68c20c 100644
--- a/tools/testing/selftests/bpf/progs/map_ptr_kern.c
+++ b/tools/testing/selftests/bpf/progs/map_ptr_kern.c
@@ -114,7 +114,7 @@ static inline int check_hash(void)
 	VERIFY(check_default_noinline(&hash->map, map));
 
 	VERIFY(hash->n_buckets == MAX_ENTRIES);
-	VERIFY(hash->elem_size == 64);
+	VERIFY(hash->elem_size == 32);
 
 	VERIFY(hash->count.counter == 0);
 	VERIFY(bpf_map_sum_elem_count(map) == 0);
-- 
2.55.0.691.gc56d675ccc-goog

Re: [PATCH bpf-next v4 2/2] bpf: htab: Reduce elem_size by 8 bytes for small key sizes

[bot+bpf-ci]

[-- Attachment #1: Type: text/plain, Size: 1364 bytes --]

> diff --git a/kernel/bpf/hashtab.c b/kernel/bpf/hashtab.c
> index fed4b1db8b78b..720766121ea2b 100644
> --- a/kernel/bpf/hashtab.c
> +++ b/kernel/bpf/hashtab.c

[ ... ]

> diff --git a/tools/testing/selftests/bpf/progs/map_ptr_kern.c b/tools/testing/selftests/bpf/progs/map_ptr_kern.c
> index 373c8d17ea557..6bd4cb68c20c1 100644
> --- a/tools/testing/selftests/bpf/progs/map_ptr_kern.c
> +++ b/tools/testing/selftests/bpf/progs/map_ptr_kern.c
> @@ -114,7 +114,7 @@ static inline int check_hash(void)
>  	VERIFY(check_default_noinline(&hash->map, map));
>
>  	VERIFY(hash->n_buckets == MAX_ENTRIES);
> -	VERIFY(hash->elem_size == 64);
> +	VERIFY(hash->elem_size == 32);
>
>  	VERIFY(hash->count.counter == 0);
>  	VERIFY(bpf_map_sum_elem_count(map) == 0);

Should this expectation change have been split across two commits for
bisectability? The preceding commit 1066507e33e67 dropped the
freelist/lru union from struct htab_elem, which moves elem_size to 40,
but that commit leaves the selftest checking for 64. This means the
map_ptr selftest fails at 1066507e33e67 and only passes again here
where elem_size becomes 32.


---
AI reviewed your patch. Please fix the bug or email reply why it's not a bug.
See: https://github.com/kernel-patches/vmtest/blob/master/ci/claude/README.md

CI run summary: https://github.com/kernel-patches/bpf/actions/runs/31651309438

Re: [PATCH bpf-next v4 2/2] bpf: htab: Reduce elem_size by 8 bytes for small key sizes

[T.J. Mercier]

On Wed, Aug 12, 2026 at 5:22 PM <bot+bpf-ci@kernel.org> wrote:
>
> > diff --git a/kernel/bpf/hashtab.c b/kernel/bpf/hashtab.c
> > index fed4b1db8b78b..720766121ea2b 100644
> > --- a/kernel/bpf/hashtab.c
> > +++ b/kernel/bpf/hashtab.c
>
> [ ... ]
>
> > diff --git a/tools/testing/selftests/bpf/progs/map_ptr_kern.c b/tools/testing/selftests/bpf/progs/map_ptr_kern.c
> > index 373c8d17ea557..6bd4cb68c20c1 100644
> > --- a/tools/testing/selftests/bpf/progs/map_ptr_kern.c
> > +++ b/tools/testing/selftests/bpf/progs/map_ptr_kern.c
> > @@ -114,7 +114,7 @@ static inline int check_hash(void)
> >       VERIFY(check_default_noinline(&hash->map, map));
> >
> >       VERIFY(hash->n_buckets == MAX_ENTRIES);
> > -     VERIFY(hash->elem_size == 64);
> > +     VERIFY(hash->elem_size == 32);
> >
> >       VERIFY(hash->count.counter == 0);
> >       VERIFY(bpf_map_sum_elem_count(map) == 0);
>
> Should this expectation change have been split across two commits for
> bisectability? The preceding commit 1066507e33e67 dropped the
> freelist/lru union from struct htab_elem, which moves elem_size to 40,
> but that commit leaves the selftest checking for 64. This means the
> map_ptr selftest fails at 1066507e33e67 and only passes again here
> where elem_size becomes 32.

Yes, it should be 40 in the last patch, and 32 here.

Re: [PATCH bpf-next v4 0/2] bpf: htab: Reduce memory use of hash maps

[Mykyta Yatsenko]

On 8/13/26 12:19 AM, T.J. Mercier wrote:
> Memory is expensive and scarce these days. This series reduces the
> memory use of BPF hash maps by eliminating the per-element overheads
> below. This saves up to 50% of per-element memory use for standard and
> PCPU hash maps. The memory use of LRU hash maps is unaffected.
> 
> Map Type & Configuration            | Old size | New size | Savings
> ------------------------------------|----------|----------|--------
> Standard (key <= 8 B, val <= 8 B)   |   64 B   |   32 B   | 50.0%
> Per-CPU (prealloc) (key <= 8 B)     |   64 B   |   32 B   | 50.0%
> Per-CPU (non-prealloc) (key <= 8 B) |   64 B   |   40 B   | 37.5%
> LRU (Any key/value size)            |    -     |    -     | 00.0%
> 

T.J. are you still interested landing this? Maybe respin the series?
Alexei was away back when you sent this.

> 1) Unused LRU / PCPU fields in standard and PCPU hash maps (patch 1)
> struct htab_elem is used for all hash map types, and includes fields
> that are not always used (bpf_lru_node, ptr_to_pptr). For standard
> (non-LRU, non-PCPU) hash maps the 24 bytes for the bpf_lru_node (union)
> are entirely overhead and can be eliminated. Non-preallocated PCPU maps
> only need the 8 byte ptr_to_pptr which is currently unioned with the
> unneeded 24 byte bpf_lru_node, so 16 bytes of overhead can be
> eliminated. Preallocated PCPU maps don't need ptr_to_pptr, so 24 bytes
> of overhead can be saved.
> 
> 2) Hash caching for small keys (patch 2)
> For hash maps with small key sizes (<= word size), comparing keys only
> requires a single instruction. Currently the 4 byte hash value (8 byte
> aligned and padded) is used for this, but offers no performance
> advantage in this case and can be eliminated.
> 
> The implementation splits htab_elem into dedicated structures for the
> different map types (htab_elem_lru, htab_elem_pcpu, htab_elem) which
> share a common initial sequence (struct htab_node), but contain
> additional map-type specific fields where necessary. This means the
> placement of the key for each element varies with the map type, and
> key_offset is added to bpf_htab for this purpose.
> 
> While using key_offset and conditional hash checks adds new pointer
> dereferences and branching during element traversal,
> run_bench_htab_mem.sh shows no significant performance regression across
> 10 runs on my 3995WX.
> 
> Benchmark (all in kops/sec)  | Avg. Before | StDev | Avg. After | StDev
> -----------------------------|-------------|-------|------------|------
> prealloc overwrite           | 115.11      | 4.10  | 115.45     | 5.24
> prealloc batch_add_batch_del | 127.14      | 4.32  | 127.06     | 2.32
> prealloc add_del_on_diff_cpu | 23.22       | 0.93  | 22.91      | 1.60
> normal overwrite             | 78.52       | 3.05  | 80.40      | 3.25
> normal batch_add_batch_del   | 45.37       | 0.69  | 47.71      | 0.66
> normal add_del_on_diff_cpu   | 12.02       | 0.73  | 12.48      | 0.70
> 
> ---
> Changes in v4:
> Removed inline from new functions per BPF CI (netdev/source_inline).
> 
> From Mykyta Yatsenko:
> Factor out duplicate lookup_elem code into __lookup_elem_raw.
> Use offsetof instead of sizeof for key_offset assignments in
> htab_map_alloc (patch 1).
> Eliminate branching and htab_elem casting in htab_elem_hash /
> htab_elem_set_hash.
> 
> Changes in v3:
> From Sashiko on torn reads/writes:
> Use a local unsigned long and READ_ONCE / WRITE_ONCE instead of memcmp /
> memcpy for atomic key comparisons for hashless elements.
> 
> Changes in v2:
> Make maximum key_size for !has_hash depend on word size for atomicity
> on 32-bit.
> 
> From Mykyta Yatsenko:
> Put the htab_elem* common initial sequence in its own struct (htab_node)
> and reuse it across all element types that share it. Eliminate
> associated BUILD_BUG_ON additions.
> Replace both the hash and key fields with data[].
> Store has_hash in struct bpf_htab, and avoid per-element reads of it.# Please edit the description for the branch
> 
> T.J. Mercier (2):
>   bpf: htab: Split htab_elem_lru and htab_elem_pcpu off of htab_elem
>   bpf: htab: Reduce elem_size by 8 bytes for small key sizes
> 
>  kernel/bpf/hashtab.c                          | 418 ++++++++++++------
>  kernel/bpf/map_in_map.c                       |  13 +
>  kernel/bpf/map_in_map.h                       |   2 +
>  .../selftests/bpf/progs/map_ptr_kern.c        |   2 +-
>  4 files changed, 289 insertions(+), 146 deletions(-)
> 
> 
> base-commit: cfce77b63375dac81d53f2f85593c548415206b7

Re: [PATCH bpf-next v4 0/2] bpf: htab: Reduce memory use of hash maps

[T.J. Mercier]

On Thu, Sep 17, 2026 at 9:10 AM Mykyta Yatsenko
<mykyta.yatsenko5@gmail.com> wrote:
>
> On 8/13/26 12:19 AM, T.J. Mercier wrote:
> > Memory is expensive and scarce these days. This series reduces the
> > memory use of BPF hash maps by eliminating the per-element overheads
> > below. This saves up to 50% of per-element memory use for standard and
> > PCPU hash maps. The memory use of LRU hash maps is unaffected.
> >
> > Map Type & Configuration            | Old size | New size | Savings
> > ------------------------------------|----------|----------|--------
> > Standard (key <= 8 B, val <= 8 B)   |   64 B   |   32 B   | 50.0%
> > Per-CPU (prealloc) (key <= 8 B)     |   64 B   |   32 B   | 50.0%
> > Per-CPU (non-prealloc) (key <= 8 B) |   64 B   |   40 B   | 37.5%
> > LRU (Any key/value size)            |    -     |    -     | 00.0%
> >
>
> T.J. are you still interested landing this? Maybe respin the series?
> Alexei was away back when you sent this.

Hi Mykyta, thanks for following up. Yes, I'd still like to land it. I
took a break during the merge window which coincided with a vacation,
where I broke my collarbone on a mountain bike jump gone wrong. Then
surgery and recovery, and I'm still catching up on everything from
while I was out. I plan to rebase this and send it out again before I
head out for pre-LPC travel next week.


> > 1) Unused LRU / PCPU fields in standard and PCPU hash maps (patch 1)
> > struct htab_elem is used for all hash map types, and includes fields
> > that are not always used (bpf_lru_node, ptr_to_pptr). For standard
> > (non-LRU, non-PCPU) hash maps the 24 bytes for the bpf_lru_node (union)
> > are entirely overhead and can be eliminated. Non-preallocated PCPU maps
> > only need the 8 byte ptr_to_pptr which is currently unioned with the
> > unneeded 24 byte bpf_lru_node, so 16 bytes of overhead can be
> > eliminated. Preallocated PCPU maps don't need ptr_to_pptr, so 24 bytes
> > of overhead can be saved.
> >
> > 2) Hash caching for small keys (patch 2)
> > For hash maps with small key sizes (<= word size), comparing keys only
> > requires a single instruction. Currently the 4 byte hash value (8 byte
> > aligned and padded) is used for this, but offers no performance
> > advantage in this case and can be eliminated.
> >
> > The implementation splits htab_elem into dedicated structures for the
> > different map types (htab_elem_lru, htab_elem_pcpu, htab_elem) which
> > share a common initial sequence (struct htab_node), but contain
> > additional map-type specific fields where necessary. This means the
> > placement of the key for each element varies with the map type, and
> > key_offset is added to bpf_htab for this purpose.
> >
> > While using key_offset and conditional hash checks adds new pointer
> > dereferences and branching during element traversal,
> > run_bench_htab_mem.sh shows no significant performance regression across
> > 10 runs on my 3995WX.
> >
> > Benchmark (all in kops/sec)  | Avg. Before | StDev | Avg. After | StDev
> > -----------------------------|-------------|-------|------------|------
> > prealloc overwrite           | 115.11      | 4.10  | 115.45     | 5.24
> > prealloc batch_add_batch_del | 127.14      | 4.32  | 127.06     | 2.32
> > prealloc add_del_on_diff_cpu | 23.22       | 0.93  | 22.91      | 1.60
> > normal overwrite             | 78.52       | 3.05  | 80.40      | 3.25
> > normal batch_add_batch_del   | 45.37       | 0.69  | 47.71      | 0.66
> > normal add_del_on_diff_cpu   | 12.02       | 0.73  | 12.48      | 0.70
> >
> > ---
> > Changes in v4:
> > Removed inline from new functions per BPF CI (netdev/source_inline).
> >
> > From Mykyta Yatsenko:
> > Factor out duplicate lookup_elem code into __lookup_elem_raw.
> > Use offsetof instead of sizeof for key_offset assignments in
> > htab_map_alloc (patch 1).
> > Eliminate branching and htab_elem casting in htab_elem_hash /
> > htab_elem_set_hash.
> >
> > Changes in v3:
> > From Sashiko on torn reads/writes:
> > Use a local unsigned long and READ_ONCE / WRITE_ONCE instead of memcmp /
> > memcpy for atomic key comparisons for hashless elements.
> >
> > Changes in v2:
> > Make maximum key_size for !has_hash depend on word size for atomicity
> > on 32-bit.
> >
> > From Mykyta Yatsenko:
> > Put the htab_elem* common initial sequence in its own struct (htab_node)
> > and reuse it across all element types that share it. Eliminate
> > associated BUILD_BUG_ON additions.
> > Replace both the hash and key fields with data[].
> > Store has_hash in struct bpf_htab, and avoid per-element reads of it.# Please edit the description for the branch
> >
> > T.J. Mercier (2):
> >   bpf: htab: Split htab_elem_lru and htab_elem_pcpu off of htab_elem
> >   bpf: htab: Reduce elem_size by 8 bytes for small key sizes
> >
> >  kernel/bpf/hashtab.c                          | 418 ++++++++++++------
> >  kernel/bpf/map_in_map.c                       |  13 +
> >  kernel/bpf/map_in_map.h                       |   2 +
> >  .../selftests/bpf/progs/map_ptr_kern.c        |   2 +-
> >  4 files changed, 289 insertions(+), 146 deletions(-)
> >
> >
> > base-commit: cfce77b63375dac81d53f2f85593c548415206b7
>

Re: [PATCH bpf-next v4 0/2] bpf: htab: Reduce memory use of hash maps

[Andrii Nakryiko]

On Thu, Sep 17, 2026 at 9:31 AM T.J. Mercier <tjmercier@google.com> wrote:
>
> On Thu, Sep 17, 2026 at 9:10 AM Mykyta Yatsenko
> <mykyta.yatsenko5@gmail.com> wrote:
> >
> > On 8/13/26 12:19 AM, T.J. Mercier wrote:
> > > Memory is expensive and scarce these days. This series reduces the
> > > memory use of BPF hash maps by eliminating the per-element overheads
> > > below. This saves up to 50% of per-element memory use for standard and
> > > PCPU hash maps. The memory use of LRU hash maps is unaffected.
> > >
> > > Map Type & Configuration            | Old size | New size | Savings
> > > ------------------------------------|----------|----------|--------
> > > Standard (key <= 8 B, val <= 8 B)   |   64 B   |   32 B   | 50.0%
> > > Per-CPU (prealloc) (key <= 8 B)     |   64 B   |   32 B   | 50.0%
> > > Per-CPU (non-prealloc) (key <= 8 B) |   64 B   |   40 B   | 37.5%
> > > LRU (Any key/value size)            |    -     |    -     | 00.0%
> > >
> >
> > T.J. are you still interested landing this? Maybe respin the series?
> > Alexei was away back when you sent this.
>
> Hi Mykyta, thanks for following up. Yes, I'd still like to land it. I
> took a break during the merge window which coincided with a vacation,
> where I broke my collarbone on a mountain bike jump gone wrong. Then
> surgery and recovery, and I'm still catching up on everything from

oh, wow, hope you'll heal fast and well!

> while I was out. I plan to rebase this and send it out again before I
> head out for pre-LPC travel next week.

have you considered splitting lru and non-lru flavors of hashtable
before doing this optimization? I'm wondering if it will allow to
clean up some parts of it, while also making map struct itself smaller
for non-lru map (there is that LRU-specific piece in the union which
artificially blows up the size of any hash map).

I am a bit concerned about that key offset, even if the microbenchmark
doesn't show much difference. What if we put hash itself before
per-element header, so that key/value are always at the same offset.
For cases where key size > 8 we'll just access hash at
addrof(htab_elem) - 8, while smaller key sizes will just directly
compare keys.

Just some high level thoughts, haven't really coded any of that, so
hard to tell upfront if it's worth doing.

>
>
> > > 1) Unused LRU / PCPU fields in standard and PCPU hash maps (patch 1)
> > > struct htab_elem is used for all hash map types, and includes fields
> > > that are not always used (bpf_lru_node, ptr_to_pptr). For standard
> > > (non-LRU, non-PCPU) hash maps the 24 bytes for the bpf_lru_node (union)
> > > are entirely overhead and can be eliminated. Non-preallocated PCPU maps
> > > only need the 8 byte ptr_to_pptr which is currently unioned with the
> > > unneeded 24 byte bpf_lru_node, so 16 bytes of overhead can be
> > > eliminated. Preallocated PCPU maps don't need ptr_to_pptr, so 24 bytes
> > > of overhead can be saved.
> > >
> > > 2) Hash caching for small keys (patch 2)
> > > For hash maps with small key sizes (<= word size), comparing keys only
> > > requires a single instruction. Currently the 4 byte hash value (8 byte
> > > aligned and padded) is used for this, but offers no performance
> > > advantage in this case and can be eliminated.
> > >
> > > The implementation splits htab_elem into dedicated structures for the
> > > different map types (htab_elem_lru, htab_elem_pcpu, htab_elem) which
> > > share a common initial sequence (struct htab_node), but contain
> > > additional map-type specific fields where necessary. This means the
> > > placement of the key for each element varies with the map type, and
> > > key_offset is added to bpf_htab for this purpose.
> > >
> > > While using key_offset and conditional hash checks adds new pointer
> > > dereferences and branching during element traversal,
> > > run_bench_htab_mem.sh shows no significant performance regression across
> > > 10 runs on my 3995WX.
> > >
> > > Benchmark (all in kops/sec)  | Avg. Before | StDev | Avg. After | StDev
> > > -----------------------------|-------------|-------|------------|------
> > > prealloc overwrite           | 115.11      | 4.10  | 115.45     | 5.24
> > > prealloc batch_add_batch_del | 127.14      | 4.32  | 127.06     | 2.32
> > > prealloc add_del_on_diff_cpu | 23.22       | 0.93  | 22.91      | 1.60
> > > normal overwrite             | 78.52       | 3.05  | 80.40      | 3.25
> > > normal batch_add_batch_del   | 45.37       | 0.69  | 47.71      | 0.66
> > > normal add_del_on_diff_cpu   | 12.02       | 0.73  | 12.48      | 0.70
> > >
> > > ---
> > > Changes in v4:
> > > Removed inline from new functions per BPF CI (netdev/source_inline).
> > >
> > > From Mykyta Yatsenko:
> > > Factor out duplicate lookup_elem code into __lookup_elem_raw.
> > > Use offsetof instead of sizeof for key_offset assignments in
> > > htab_map_alloc (patch 1).
> > > Eliminate branching and htab_elem casting in htab_elem_hash /
> > > htab_elem_set_hash.
> > >
> > > Changes in v3:
> > > From Sashiko on torn reads/writes:
> > > Use a local unsigned long and READ_ONCE / WRITE_ONCE instead of memcmp /
> > > memcpy for atomic key comparisons for hashless elements.
> > >
> > > Changes in v2:
> > > Make maximum key_size for !has_hash depend on word size for atomicity
> > > on 32-bit.
> > >
> > > From Mykyta Yatsenko:
> > > Put the htab_elem* common initial sequence in its own struct (htab_node)
> > > and reuse it across all element types that share it. Eliminate
> > > associated BUILD_BUG_ON additions.
> > > Replace both the hash and key fields with data[].
> > > Store has_hash in struct bpf_htab, and avoid per-element reads of it.# Please edit the description for the branch
> > >
> > > T.J. Mercier (2):
> > >   bpf: htab: Split htab_elem_lru and htab_elem_pcpu off of htab_elem
> > >   bpf: htab: Reduce elem_size by 8 bytes for small key sizes
> > >
> > >  kernel/bpf/hashtab.c                          | 418 ++++++++++++------
> > >  kernel/bpf/map_in_map.c                       |  13 +
> > >  kernel/bpf/map_in_map.h                       |   2 +
> > >  .../selftests/bpf/progs/map_ptr_kern.c        |   2 +-
> > >  4 files changed, 289 insertions(+), 146 deletions(-)
> > >
> > >
> > > base-commit: cfce77b63375dac81d53f2f85593c548415206b7
> >

Re: [PATCH bpf-next v4 0/2] bpf: htab: Reduce memory use of hash maps

[T.J. Mercier]

On Thu, Sep 17, 2026 at 2:21 PM Andrii Nakryiko
<andrii.nakryiko@gmail.com> wrote:
>
> On Thu, Sep 17, 2026 at 9:31 AM T.J. Mercier <tjmercier@google.com> wrote:
> >
> > On Thu, Sep 17, 2026 at 9:10 AM Mykyta Yatsenko
> > <mykyta.yatsenko5@gmail.com> wrote:
> > >
> > > On 8/13/26 12:19 AM, T.J. Mercier wrote:
> > > > Memory is expensive and scarce these days. This series reduces the
> > > > memory use of BPF hash maps by eliminating the per-element overheads
> > > > below. This saves up to 50% of per-element memory use for standard and
> > > > PCPU hash maps. The memory use of LRU hash maps is unaffected.
> > > >
> > > > Map Type & Configuration            | Old size | New size | Savings
> > > > ------------------------------------|----------|----------|--------
> > > > Standard (key <= 8 B, val <= 8 B)   |   64 B   |   32 B   | 50.0%
> > > > Per-CPU (prealloc) (key <= 8 B)     |   64 B   |   32 B   | 50.0%
> > > > Per-CPU (non-prealloc) (key <= 8 B) |   64 B   |   40 B   | 37.5%
> > > > LRU (Any key/value size)            |    -     |    -     | 00.0%
> > > >
> > >
> > > T.J. are you still interested landing this? Maybe respin the series?
> > > Alexei was away back when you sent this.
> >
> > Hi Mykyta, thanks for following up. Yes, I'd still like to land it. I
> > took a break during the merge window which coincided with a vacation,
> > where I broke my collarbone on a mountain bike jump gone wrong. Then
> > surgery and recovery, and I'm still catching up on everything from
>
> oh, wow, hope you'll heal fast and well!

Thanks! The surgery already made a huge difference, so now I'm waiting
for bone to grow and it'll be a few months before I'm back on a bike.

> > while I was out. I plan to rebase this and send it out again before I
> > head out for pre-LPC travel next week.
>
> have you considered splitting lru and non-lru flavors of hashtable
> before doing this optimization? I'm wondering if it will allow to
> clean up some parts of it, while also making map struct itself smaller
> for non-lru map (there is that LRU-specific piece in the union which
> artificially blows up the size of any hash map).
>
> I am a bit concerned about that key offset, even if the microbenchmark
> doesn't show much difference. What if we put hash itself before
> per-element header, so that key/value are always at the same offset.
> For cases where key size > 8 we'll just access hash at
> addrof(htab_elem) - 8, while smaller key sizes will just directly
> compare keys.

This is an interesting idea and I think it will work. Let me try this
first, and then I will take a look at splitting out LRU hashtables
aftewards.

> Just some high level thoughts, haven't really coded any of that, so
> hard to tell upfront if it's worth doing.
>
> >
> >
> > > > 1) Unused LRU / PCPU fields in standard and PCPU hash maps (patch 1)
> > > > struct htab_elem is used for all hash map types, and includes fields
> > > > that are not always used (bpf_lru_node, ptr_to_pptr). For standard
> > > > (non-LRU, non-PCPU) hash maps the 24 bytes for the bpf_lru_node (union)
> > > > are entirely overhead and can be eliminated. Non-preallocated PCPU maps
> > > > only need the 8 byte ptr_to_pptr which is currently unioned with the
> > > > unneeded 24 byte bpf_lru_node, so 16 bytes of overhead can be
> > > > eliminated. Preallocated PCPU maps don't need ptr_to_pptr, so 24 bytes
> > > > of overhead can be saved.
> > > >
> > > > 2) Hash caching for small keys (patch 2)
> > > > For hash maps with small key sizes (<= word size), comparing keys only
> > > > requires a single instruction. Currently the 4 byte hash value (8 byte
> > > > aligned and padded) is used for this, but offers no performance
> > > > advantage in this case and can be eliminated.
> > > >
> > > > The implementation splits htab_elem into dedicated structures for the
> > > > different map types (htab_elem_lru, htab_elem_pcpu, htab_elem) which
> > > > share a common initial sequence (struct htab_node), but contain
> > > > additional map-type specific fields where necessary. This means the
> > > > placement of the key for each element varies with the map type, and
> > > > key_offset is added to bpf_htab for this purpose.
> > > >
> > > > While using key_offset and conditional hash checks adds new pointer
> > > > dereferences and branching during element traversal,
> > > > run_bench_htab_mem.sh shows no significant performance regression across
> > > > 10 runs on my 3995WX.
> > > >
> > > > Benchmark (all in kops/sec)  | Avg. Before | StDev | Avg. After | StDev
> > > > -----------------------------|-------------|-------|------------|------
> > > > prealloc overwrite           | 115.11      | 4.10  | 115.45     | 5.24
> > > > prealloc batch_add_batch_del | 127.14      | 4.32  | 127.06     | 2.32
> > > > prealloc add_del_on_diff_cpu | 23.22       | 0.93  | 22.91      | 1.60
> > > > normal overwrite             | 78.52       | 3.05  | 80.40      | 3.25
> > > > normal batch_add_batch_del   | 45.37       | 0.69  | 47.71      | 0.66
> > > > normal add_del_on_diff_cpu   | 12.02       | 0.73  | 12.48      | 0.70
> > > >
> > > > ---
> > > > Changes in v4:
> > > > Removed inline from new functions per BPF CI (netdev/source_inline).
> > > >
> > > > From Mykyta Yatsenko:
> > > > Factor out duplicate lookup_elem code into __lookup_elem_raw.
> > > > Use offsetof instead of sizeof for key_offset assignments in
> > > > htab_map_alloc (patch 1).
> > > > Eliminate branching and htab_elem casting in htab_elem_hash /
> > > > htab_elem_set_hash.
> > > >
> > > > Changes in v3:
> > > > From Sashiko on torn reads/writes:
> > > > Use a local unsigned long and READ_ONCE / WRITE_ONCE instead of memcmp /
> > > > memcpy for atomic key comparisons for hashless elements.
> > > >
> > > > Changes in v2:
> > > > Make maximum key_size for !has_hash depend on word size for atomicity
> > > > on 32-bit.
> > > >
> > > > From Mykyta Yatsenko:
> > > > Put the htab_elem* common initial sequence in its own struct (htab_node)
> > > > and reuse it across all element types that share it. Eliminate
> > > > associated BUILD_BUG_ON additions.
> > > > Replace both the hash and key fields with data[].
> > > > Store has_hash in struct bpf_htab, and avoid per-element reads of it.# Please edit the description for the branch
> > > >
> > > > T.J. Mercier (2):
> > > >   bpf: htab: Split htab_elem_lru and htab_elem_pcpu off of htab_elem
> > > >   bpf: htab: Reduce elem_size by 8 bytes for small key sizes
> > > >
> > > >  kernel/bpf/hashtab.c                          | 418 ++++++++++++------
> > > >  kernel/bpf/map_in_map.c                       |  13 +
> > > >  kernel/bpf/map_in_map.h                       |   2 +
> > > >  .../selftests/bpf/progs/map_ptr_kern.c        |   2 +-
> > > >  4 files changed, 289 insertions(+), 146 deletions(-)
> > > >
> > > >
> > > > base-commit: cfce77b63375dac81d53f2f85593c548415206b7
> > >