1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2014 Intel Corporation
9 #define ACL_POOL_ALIGN 8
10 #define ACL_POOL_ALLOC_MIN 0x800000
12 /* number of pointers per alloc */
13 #define ACL_PTR_ALLOC 32
15 /* macros for dividing rule sets heuristics */
16 #define NODE_MAX 0x4000
17 #define NODE_MIN 0x800
19 /* TALLY are statistics per field */
21 TALLY_0 = 0, /* number of rules that are 0% or more wild. */
22 TALLY_25, /* number of rules that are 25% or more wild. */
26 TALLY_DEACTIVATED, /* deactivated fields (100% wild in all rules). */
28 /* number of rules that are 100% wild for this field and higher. */
32 static const uint32_t wild_limits[TALLY_DEACTIVATED] = {0, 25, 50, 75, 100};
35 ACL_INTERSECT_NONE = 0,
36 ACL_INTERSECT_A = 1, /* set A is a superset of A and B intersect */
37 ACL_INTERSECT_B = 2, /* set B is a superset of A and B intersect */
38 ACL_INTERSECT = 4, /* sets A and B intersect */
42 ACL_PRIORITY_EQUAL = 0,
43 ACL_PRIORITY_NODE_A = 1,
44 ACL_PRIORITY_NODE_B = 2,
45 ACL_PRIORITY_MIXED = 3
49 struct acl_mem_block {
54 #define MEM_BLOCK_NUM 16
56 /* Single ACL rule, build representation.*/
57 struct rte_acl_build_rule {
58 struct rte_acl_build_rule *next;
59 struct rte_acl_config *config;
60 /**< configuration for each field in the rule. */
61 const struct rte_acl_rule *f;
65 /* Context for build phase */
66 struct acl_build_context {
67 const struct rte_acl_ctx *acx;
68 struct rte_acl_build_rule *build_rules;
69 struct rte_acl_config cfg;
74 uint32_t category_mask;
78 uint32_t num_build_rules;
80 struct tb_mem_pool pool;
81 struct rte_acl_trie tries[RTE_ACL_MAX_TRIES];
82 struct rte_acl_bld_trie bld_tries[RTE_ACL_MAX_TRIES];
83 uint32_t data_indexes[RTE_ACL_MAX_TRIES][RTE_ACL_MAX_FIELDS];
85 /* memory free lists for nodes and blocks used for node ptrs */
86 struct acl_mem_block blocks[MEM_BLOCK_NUM];
87 struct rte_acl_node *node_free_list;
90 static int acl_merge_trie(struct acl_build_context *context,
91 struct rte_acl_node *node_a, struct rte_acl_node *node_b,
92 uint32_t level, struct rte_acl_node **node_c);
95 acl_deref_ptr(struct acl_build_context *context,
96 struct rte_acl_node *node, int index);
99 acl_build_alloc(struct acl_build_context *context, size_t n, size_t s)
103 size_t alloc_size = n * s;
106 * look for memory in free lists
108 for (m = 0; m < RTE_DIM(context->blocks); m++) {
109 if (context->blocks[m].block_size ==
110 alloc_size && context->blocks[m].mem_ptr != NULL) {
111 p = context->blocks[m].mem_ptr;
112 context->blocks[m].mem_ptr = *((void **)p);
113 memset(p, 0, alloc_size);
119 * return allocation from memory pool
121 p = tb_alloc(&context->pool, alloc_size);
126 * Free memory blocks (kept in context for reuse).
129 acl_build_free(struct acl_build_context *context, size_t s, void *p)
133 for (n = 0; n < RTE_DIM(context->blocks); n++) {
134 if (context->blocks[n].block_size == s) {
135 *((void **)p) = context->blocks[n].mem_ptr;
136 context->blocks[n].mem_ptr = p;
140 for (n = 0; n < RTE_DIM(context->blocks); n++) {
141 if (context->blocks[n].block_size == 0) {
142 context->blocks[n].block_size = s;
143 *((void **)p) = NULL;
144 context->blocks[n].mem_ptr = p;
151 * Allocate and initialize a new node.
153 static struct rte_acl_node *
154 acl_alloc_node(struct acl_build_context *context, int level)
156 struct rte_acl_node *node;
158 if (context->node_free_list != NULL) {
159 node = context->node_free_list;
160 context->node_free_list = node->next;
161 memset(node, 0, sizeof(struct rte_acl_node));
163 node = acl_build_alloc(context, sizeof(struct rte_acl_node), 1);
169 node->node_type = RTE_ACL_NODE_UNDEFINED;
170 node->node_index = RTE_ACL_NODE_UNDEFINED;
171 context->num_nodes++;
172 node->id = context->node_id++;
178 * Dereference all nodes to which this node points
181 acl_free_node(struct acl_build_context *context,
182 struct rte_acl_node *node)
186 if (node->prev != NULL)
187 node->prev->next = NULL;
188 for (n = 0; n < node->num_ptrs; n++)
189 acl_deref_ptr(context, node, n);
191 /* free mrt if this is a match node */
192 if (node->mrt != NULL) {
193 acl_build_free(context, sizeof(struct rte_acl_match_results),
198 /* free transitions to other nodes */
199 if (node->ptrs != NULL) {
200 acl_build_free(context,
201 node->max_ptrs * sizeof(struct rte_acl_ptr_set),
206 /* put it on the free list */
207 context->num_nodes--;
208 node->next = context->node_free_list;
209 context->node_free_list = node;
214 * Include src bitset in dst bitset
217 acl_include(struct rte_acl_bitset *dst, struct rte_acl_bitset *src, bits_t mask)
221 for (n = 0; n < RTE_ACL_BIT_SET_SIZE; n++)
222 dst->bits[n] = (dst->bits[n] & mask) | src->bits[n];
226 * Set dst to bits of src1 that are not in src2
229 acl_exclude(struct rte_acl_bitset *dst,
230 struct rte_acl_bitset *src1,
231 struct rte_acl_bitset *src2)
236 for (n = 0; n < RTE_ACL_BIT_SET_SIZE; n++) {
237 dst->bits[n] = src1->bits[n] & ~src2->bits[n];
238 all_bits |= dst->bits[n];
240 return all_bits != 0;
244 * Add a pointer (ptr) to a node.
247 acl_add_ptr(struct acl_build_context *context,
248 struct rte_acl_node *node,
249 struct rte_acl_node *ptr,
250 struct rte_acl_bitset *bits)
252 uint32_t n, num_ptrs;
253 struct rte_acl_ptr_set *ptrs = NULL;
256 * If there's already a pointer to the same node, just add to the bitset
258 for (n = 0; n < node->num_ptrs; n++) {
259 if (node->ptrs[n].ptr != NULL) {
260 if (node->ptrs[n].ptr == ptr) {
261 acl_include(&node->ptrs[n].values, bits, -1);
262 acl_include(&node->values, bits, -1);
268 /* if there's no room for another pointer, make room */
269 if (node->num_ptrs >= node->max_ptrs) {
270 /* add room for more pointers */
271 num_ptrs = node->max_ptrs + ACL_PTR_ALLOC;
272 ptrs = acl_build_alloc(context, num_ptrs, sizeof(*ptrs));
274 /* copy current points to new memory allocation */
275 if (node->ptrs != NULL) {
276 memcpy(ptrs, node->ptrs,
277 node->num_ptrs * sizeof(*ptrs));
278 acl_build_free(context, node->max_ptrs * sizeof(*ptrs),
282 node->max_ptrs = num_ptrs;
285 /* Find available ptr and add a new pointer to this node */
286 for (n = node->min_add; n < node->max_ptrs; n++) {
287 if (node->ptrs[n].ptr == NULL) {
288 node->ptrs[n].ptr = ptr;
289 acl_include(&node->ptrs[n].values, bits, 0);
290 acl_include(&node->values, bits, -1);
293 if (node->num_ptrs <= n)
294 node->num_ptrs = n + 1;
303 * Add a pointer for a range of values
306 acl_add_ptr_range(struct acl_build_context *context,
307 struct rte_acl_node *root,
308 struct rte_acl_node *node,
313 struct rte_acl_bitset bitset;
315 /* clear the bitset values */
316 for (n = 0; n < RTE_ACL_BIT_SET_SIZE; n++)
319 /* for each bit in range, add bit to set */
320 for (n = 0; n < UINT8_MAX + 1; n++)
321 if (n >= low && n <= high)
322 bitset.bits[n / (sizeof(bits_t) * 8)] |=
323 1U << (n % (sizeof(bits_t) * CHAR_BIT));
325 return acl_add_ptr(context, root, node, &bitset);
329 * Generate a bitset from a byte value and mask.
332 acl_gen_mask(struct rte_acl_bitset *bitset, uint32_t value, uint32_t mask)
337 /* clear the bitset values */
338 for (n = 0; n < RTE_ACL_BIT_SET_SIZE; n++)
341 /* for each bit in value/mask, add bit to set */
342 for (n = 0; n < UINT8_MAX + 1; n++) {
343 if ((n & mask) == value) {
345 bitset->bits[n / (sizeof(bits_t) * 8)] |=
346 1U << (n % (sizeof(bits_t) * CHAR_BIT));
353 * Determine how A and B intersect.
354 * Determine if A and/or B are supersets of the intersection.
357 acl_intersect_type(const struct rte_acl_bitset *a_bits,
358 const struct rte_acl_bitset *b_bits,
359 struct rte_acl_bitset *intersect)
362 bits_t intersect_bits = 0;
363 bits_t a_superset = 0;
364 bits_t b_superset = 0;
367 * calculate and store intersection and check if A and/or B have
368 * bits outside the intersection (superset)
370 for (n = 0; n < RTE_ACL_BIT_SET_SIZE; n++) {
371 intersect->bits[n] = a_bits->bits[n] & b_bits->bits[n];
372 a_superset |= a_bits->bits[n] ^ intersect->bits[n];
373 b_superset |= b_bits->bits[n] ^ intersect->bits[n];
374 intersect_bits |= intersect->bits[n];
377 n = (intersect_bits == 0 ? ACL_INTERSECT_NONE : ACL_INTERSECT) |
378 (b_superset == 0 ? 0 : ACL_INTERSECT_B) |
379 (a_superset == 0 ? 0 : ACL_INTERSECT_A);
387 static struct rte_acl_node *
388 acl_dup_node(struct acl_build_context *context, struct rte_acl_node *node)
391 struct rte_acl_node *next;
393 next = acl_alloc_node(context, node->level);
395 /* allocate the pointers */
396 if (node->num_ptrs > 0) {
397 next->ptrs = acl_build_alloc(context,
399 sizeof(struct rte_acl_ptr_set));
400 next->max_ptrs = node->max_ptrs;
403 /* copy over the pointers */
404 for (n = 0; n < node->num_ptrs; n++) {
405 if (node->ptrs[n].ptr != NULL) {
406 next->ptrs[n].ptr = node->ptrs[n].ptr;
407 next->ptrs[n].ptr->ref_count++;
408 acl_include(&next->ptrs[n].values,
409 &node->ptrs[n].values, -1);
413 next->num_ptrs = node->num_ptrs;
415 /* copy over node's match results */
416 if (node->match_flag == 0)
417 next->match_flag = 0;
419 next->match_flag = -1;
420 next->mrt = acl_build_alloc(context, 1, sizeof(*next->mrt));
421 memcpy(next->mrt, node->mrt, sizeof(*next->mrt));
424 /* copy over node's bitset */
425 acl_include(&next->values, &node->values, -1);
434 * Dereference a pointer from a node
437 acl_deref_ptr(struct acl_build_context *context,
438 struct rte_acl_node *node, int index)
440 struct rte_acl_node *ref_node;
442 /* De-reference the node at the specified pointer */
443 if (node != NULL && node->ptrs[index].ptr != NULL) {
444 ref_node = node->ptrs[index].ptr;
445 ref_node->ref_count--;
446 if (ref_node->ref_count == 0)
447 acl_free_node(context, ref_node);
452 * acl_exclude rte_acl_bitset from src and copy remaining pointer to dst
455 acl_copy_ptr(struct acl_build_context *context,
456 struct rte_acl_node *dst,
457 struct rte_acl_node *src,
459 struct rte_acl_bitset *b_bits)
462 struct rte_acl_bitset bits;
465 if (!acl_exclude(&bits, &src->ptrs[index].values, b_bits))
468 rc = acl_add_ptr(context, dst, src->ptrs[index].ptr, &bits);
475 * Fill in gaps in ptrs list with the ptr at the end of the list
478 acl_compact_node_ptrs(struct rte_acl_node *node_a)
481 int min_add = node_a->min_add;
483 while (node_a->num_ptrs > 0 &&
484 node_a->ptrs[node_a->num_ptrs - 1].ptr == NULL)
487 for (n = min_add; n + 1 < node_a->num_ptrs; n++) {
489 /* if this entry is empty */
490 if (node_a->ptrs[n].ptr == NULL) {
492 /* move the last pointer to this entry */
493 acl_include(&node_a->ptrs[n].values,
494 &node_a->ptrs[node_a->num_ptrs - 1].values,
496 node_a->ptrs[n].ptr =
497 node_a->ptrs[node_a->num_ptrs - 1].ptr;
500 * mark the end as empty and adjust the number
501 * of used pointer enum_tries
503 node_a->ptrs[node_a->num_ptrs - 1].ptr = NULL;
504 while (node_a->num_ptrs > 0 &&
505 node_a->ptrs[node_a->num_ptrs - 1].ptr == NULL)
512 acl_resolve_leaf(struct acl_build_context *context,
513 struct rte_acl_node *node_a,
514 struct rte_acl_node *node_b,
515 struct rte_acl_node **node_c)
518 int combined_priority = ACL_PRIORITY_EQUAL;
520 for (n = 0; n < context->cfg.num_categories; n++) {
521 if (node_a->mrt->priority[n] != node_b->mrt->priority[n]) {
522 combined_priority |= (node_a->mrt->priority[n] >
523 node_b->mrt->priority[n]) ?
524 ACL_PRIORITY_NODE_A : ACL_PRIORITY_NODE_B;
529 * if node a is higher or equal priority for all categories,
530 * then return node_a.
532 if (combined_priority == ACL_PRIORITY_NODE_A ||
533 combined_priority == ACL_PRIORITY_EQUAL) {
539 * if node b is higher or equal priority for all categories,
540 * then return node_b.
542 if (combined_priority == ACL_PRIORITY_NODE_B) {
548 * mixed priorities - create a new node with the highest priority
552 /* force new duplication. */
555 *node_c = acl_dup_node(context, node_a);
556 for (n = 0; n < context->cfg.num_categories; n++) {
557 if ((*node_c)->mrt->priority[n] < node_b->mrt->priority[n]) {
558 (*node_c)->mrt->priority[n] = node_b->mrt->priority[n];
559 (*node_c)->mrt->results[n] = node_b->mrt->results[n];
566 * Merge nodes A and B together,
567 * returns a node that is the path for the intersection
569 * If match node (leaf on trie)
571 * return node = highest priority result
573 * Create C as a duplicate of A to point to child intersections
574 * If any pointers in C intersect with any in B
575 * For each intersection
577 * remove intersection from C pointer
578 * add a pointer from C to child intersection node
579 * Compact the pointers in A and B
580 * Copy any B pointers that are outside of the intersection to C
581 * If C has no references to the B trie
582 * free C and return A
583 * Else If C has no references to the A trie
584 * free C and return B
589 acl_merge_trie(struct acl_build_context *context,
590 struct rte_acl_node *node_a, struct rte_acl_node *node_b,
591 uint32_t level, struct rte_acl_node **return_c)
593 uint32_t n, m, ptrs_c, ptrs_b;
594 uint32_t min_add_c, min_add_b;
595 int node_intersect_type;
596 struct rte_acl_bitset node_intersect;
597 struct rte_acl_node *node_c;
598 struct rte_acl_node *node_a_next;
603 node_a_next = node_a->next;
606 node_a_refs = node_a->num_ptrs;
608 node_intersect_type = 0;
610 /* Resolve leaf nodes (matches) */
611 if (node_a->match_flag != 0) {
612 acl_resolve_leaf(context, node_a, node_b, return_c);
617 * Create node C as a copy of node A, and do: C = merge(A,B);
618 * If node A can be used instead (A==C), then later we'll
619 * destroy C and return A.
622 node_c = acl_dup_node(context, node_a);
625 * If the two node transitions intersect then merge the transitions.
626 * Check intersection for entire node (all pointers)
628 node_intersect_type = acl_intersect_type(&node_c->values,
632 if (node_intersect_type & ACL_INTERSECT) {
634 min_add_b = node_b->min_add;
635 node_b->min_add = node_b->num_ptrs;
636 ptrs_b = node_b->num_ptrs;
638 min_add_c = node_c->min_add;
639 node_c->min_add = node_c->num_ptrs;
640 ptrs_c = node_c->num_ptrs;
642 for (n = 0; n < ptrs_c; n++) {
643 if (node_c->ptrs[n].ptr == NULL) {
647 node_c->ptrs[n].ptr->next = NULL;
648 for (m = 0; m < ptrs_b; m++) {
650 struct rte_acl_bitset child_intersect;
651 int child_intersect_type;
652 struct rte_acl_node *child_node_c = NULL;
654 if (node_b->ptrs[m].ptr == NULL ||
655 node_c->ptrs[n].ptr ==
659 child_intersect_type = acl_intersect_type(
660 &node_c->ptrs[n].values,
661 &node_b->ptrs[m].values,
664 if ((child_intersect_type & ACL_INTERSECT) !=
666 if (acl_merge_trie(context,
673 if (child_node_c != NULL &&
675 node_c->ptrs[n].ptr) {
680 * Added link from C to
681 * child_C for all transitions
682 * in the intersection.
684 acl_add_ptr(context, node_c,
689 * inc refs if pointer is not
692 node_a_refs += (child_node_c !=
693 node_b->ptrs[m].ptr);
696 * Remove intersection from C
700 &node_c->ptrs[n].values,
701 &node_c->ptrs[n].values,
703 acl_deref_ptr(context,
705 node_c->ptrs[n].ptr =
714 /* Compact pointers */
715 node_c->min_add = min_add_c;
716 acl_compact_node_ptrs(node_c);
717 node_b->min_add = min_add_b;
718 acl_compact_node_ptrs(node_b);
722 * Copy pointers outside of the intersection from B to C
724 if ((node_intersect_type & ACL_INTERSECT_B) != 0) {
726 for (m = 0; m < node_b->num_ptrs; m++)
727 if (node_b->ptrs[m].ptr != NULL)
728 acl_copy_ptr(context, node_c,
729 node_b, m, &node_intersect);
733 * Free node C if top of trie is contained in A or B
734 * if node C is a duplicate of node A &&
735 * node C was not an existing duplicate
737 if (node_c != node_a && node_c != node_a_next) {
740 * if the intersection has no references to the
741 * B side, then it is contained in A
743 if (node_b_refs == 0) {
744 acl_free_node(context, node_c);
748 * if the intersection has no references to the
749 * A side, then it is contained in B.
751 if (node_a_refs == 0) {
752 acl_free_node(context, node_c);
758 if (return_c != NULL)
762 acl_free_node(context, node_b);
768 * Reset current runtime fields before next build:
769 * - free allocated RT memory.
770 * - reset all RT related fields to zero.
773 acl_build_reset(struct rte_acl_ctx *ctx)
776 memset(&ctx->num_categories, 0,
777 sizeof(*ctx) - offsetof(struct rte_acl_ctx, num_categories));
781 acl_gen_full_range(struct acl_build_context *context, struct rte_acl_node *root,
782 struct rte_acl_node *end, int size, int level)
784 struct rte_acl_node *node, *prev;
788 for (n = size - 1; n > 0; n--) {
789 node = acl_alloc_node(context, level++);
790 acl_add_ptr_range(context, prev, node, 0, UINT8_MAX);
793 acl_add_ptr_range(context, prev, end, 0, UINT8_MAX);
797 acl_gen_range_mdl(struct acl_build_context *context, struct rte_acl_node *root,
798 struct rte_acl_node *end, uint8_t lo, uint8_t hi, int size, int level)
800 struct rte_acl_node *node;
802 node = acl_alloc_node(context, level++);
803 acl_add_ptr_range(context, root, node, lo, hi);
804 acl_gen_full_range(context, node, end, size - 1, level);
808 acl_gen_range_low(struct acl_build_context *context, struct rte_acl_node *root,
809 struct rte_acl_node *end, const uint8_t *lo, int size, int level)
811 struct rte_acl_node *node;
816 acl_add_ptr_range(context, root, end, lo[0], UINT8_MAX);
820 node = acl_alloc_node(context, level++);
821 acl_add_ptr_range(context, root, node, lo[n], lo[n]);
823 /* generate lower-bound sub-trie */
824 acl_gen_range_low(context, node, end, lo, n, level);
826 /* generate middle sub-trie */
827 if (n > 1 && lo[n - 1] != UINT8_MAX)
828 acl_gen_range_mdl(context, node, end, lo[n - 1] + 1, UINT8_MAX,
833 acl_gen_range_high(struct acl_build_context *context, struct rte_acl_node *root,
834 struct rte_acl_node *end, const uint8_t *hi, int size, int level)
836 struct rte_acl_node *node;
841 acl_add_ptr_range(context, root, end, 0, hi[0]);
845 node = acl_alloc_node(context, level++);
846 acl_add_ptr_range(context, root, node, hi[n], hi[n]);
848 /* generate upper-bound sub-trie */
849 acl_gen_range_high(context, node, end, hi, n, level);
851 /* generate middle sub-trie */
852 if (n > 1 && hi[n - 1] != 0)
853 acl_gen_range_mdl(context, node, end, 0, hi[n - 1] - 1,
857 static struct rte_acl_node *
858 acl_gen_range_trie(struct acl_build_context *context,
859 const void *min, const void *max,
860 int size, int level, struct rte_acl_node **pend)
863 uint8_t hi_ff, lo_00;
864 struct rte_acl_node *node, *prev, *root;
871 *pend = acl_alloc_node(context, level + size);
872 root = acl_alloc_node(context, level++);
875 /* build common sub-trie till possible */
876 for (n = size - 1; n > 0 && lo[n] == hi[n]; n--) {
877 node = acl_alloc_node(context, level++);
878 acl_add_ptr_range(context, prev, node, lo[n], hi[n]);
882 /* no branch needed, just one sub-trie */
884 acl_add_ptr_range(context, prev, *pend, lo[0], hi[0]);
888 /* gather information about divirgent paths */
891 for (k = n - 1; k >= 0; k--) {
896 /* generate left (lower-bound) sub-trie */
898 acl_gen_range_low(context, prev, *pend, lo, n + 1, level);
900 /* generate right (upper-bound) sub-trie */
901 if (hi_ff != UINT8_MAX)
902 acl_gen_range_high(context, prev, *pend, hi, n + 1, level);
904 /* generate sub-trie in the middle */
905 if (lo[n] + 1 != hi[n] || lo_00 == 0 || hi_ff == UINT8_MAX) {
906 lo_00 = lo[n] + (lo_00 != 0);
907 hi_ff = hi[n] - (hi_ff != UINT8_MAX);
908 acl_gen_range_mdl(context, prev, *pend, lo_00, hi_ff,
915 static struct rte_acl_node *
916 acl_gen_mask_trie(struct acl_build_context *context,
917 const void *value, const void *mask,
918 int size, int level, struct rte_acl_node **pend)
921 struct rte_acl_node *root;
922 struct rte_acl_node *node, *prev;
923 struct rte_acl_bitset bits;
924 const uint8_t *val = value;
925 const uint8_t *msk = mask;
927 root = acl_alloc_node(context, level++);
930 for (n = size - 1; n >= 0; n--) {
931 node = acl_alloc_node(context, level++);
932 acl_gen_mask(&bits, val[n] & msk[n], msk[n]);
933 acl_add_ptr(context, prev, node, &bits);
941 static struct rte_acl_node *
942 build_trie(struct acl_build_context *context, struct rte_acl_build_rule *head,
943 struct rte_acl_build_rule **last, uint32_t *count)
946 int field_index, node_count;
947 struct rte_acl_node *trie;
948 struct rte_acl_build_rule *prev, *rule;
949 struct rte_acl_node *end, *merge, *root, *end_prev;
950 const struct rte_acl_field *fld;
956 trie = acl_alloc_node(context, 0);
958 while (rule != NULL) {
960 root = acl_alloc_node(context, 0);
965 for (n = 0; n < rule->config->num_fields; n++) {
967 field_index = rule->config->defs[n].field_index;
968 fld = rule->f->field + field_index;
971 /* build a mini-trie for this field */
972 switch (rule->config->defs[n].type) {
974 case RTE_ACL_FIELD_TYPE_BITMASK:
975 merge = acl_gen_mask_trie(context,
978 rule->config->defs[n].size,
983 case RTE_ACL_FIELD_TYPE_MASK:
986 * set msb for the size of the field and
990 mask = RTE_ACL_MASKLEN_TO_BITMASK(
992 rule->config->defs[n].size);
994 /* gen a mini-trie for this field */
995 merge = acl_gen_mask_trie(context,
998 rule->config->defs[n].size,
1004 case RTE_ACL_FIELD_TYPE_RANGE:
1005 merge = acl_gen_range_trie(context,
1006 &rule->f->field[field_index].value,
1007 &rule->f->field[field_index].mask_range,
1008 rule->config->defs[n].size,
1015 "Error in rule[%u] type - %hhu\n",
1016 rule->f->data.userdata,
1017 rule->config->defs[n].type);
1021 /* merge this field on to the end of the rule */
1022 if (acl_merge_trie(context, end_prev, merge, 0,
1028 end->match_flag = ++context->num_build_rules;
1031 * Setup the results for this rule.
1032 * The result and priority of each category.
1034 if (end->mrt == NULL)
1035 end->mrt = acl_build_alloc(context, 1,
1038 for (m = context->cfg.num_categories; 0 != m--; ) {
1039 if (rule->f->data.category_mask & (1U << m)) {
1040 end->mrt->results[m] = rule->f->data.userdata;
1041 end->mrt->priority[m] = rule->f->data.priority;
1043 end->mrt->results[m] = 0;
1044 end->mrt->priority[m] = 0;
1048 node_count = context->num_nodes;
1051 /* merge this rule into the trie */
1052 if (acl_merge_trie(context, trie, root, 0, NULL))
1055 node_count = context->num_nodes - node_count;
1056 if (node_count > context->cur_node_max) {
1070 acl_calc_wildness(struct rte_acl_build_rule *head,
1071 const struct rte_acl_config *config)
1074 struct rte_acl_build_rule *rule;
1076 for (rule = head; rule != NULL; rule = rule->next) {
1078 for (n = 0; n < config->num_fields; n++) {
1081 uint32_t bit_len = CHAR_BIT * config->defs[n].size;
1082 uint64_t msk_val = RTE_LEN2MASK(bit_len,
1084 double size = bit_len;
1085 int field_index = config->defs[n].field_index;
1086 const struct rte_acl_field *fld = rule->f->field +
1089 switch (rule->config->defs[n].type) {
1090 case RTE_ACL_FIELD_TYPE_BITMASK:
1091 wild = (size - __builtin_popcountll(
1092 fld->mask_range.u64 & msk_val)) /
1096 case RTE_ACL_FIELD_TYPE_MASK:
1097 wild = (size - fld->mask_range.u32) / size;
1100 case RTE_ACL_FIELD_TYPE_RANGE:
1101 wild = (fld->mask_range.u64 & msk_val) -
1102 (fld->value.u64 & msk_val);
1103 wild = wild / msk_val;
1107 rule->wildness[field_index] = (uint32_t)(wild * 100);
1113 acl_rule_stats(struct rte_acl_build_rule *head, struct rte_acl_config *config)
1115 struct rte_acl_build_rule *rule;
1116 uint32_t n, m, fields_deactivated = 0;
1117 uint32_t start = 0, deactivate = 0;
1118 int tally[RTE_ACL_MAX_LEVELS][TALLY_NUM];
1120 memset(tally, 0, sizeof(tally));
1122 for (rule = head; rule != NULL; rule = rule->next) {
1124 for (n = 0; n < config->num_fields; n++) {
1125 uint32_t field_index = config->defs[n].field_index;
1127 tally[n][TALLY_0]++;
1128 for (m = 1; m < RTE_DIM(wild_limits); m++) {
1129 if (rule->wildness[field_index] >=
1135 for (n = config->num_fields - 1; n > 0; n--) {
1136 uint32_t field_index = config->defs[n].field_index;
1138 if (rule->wildness[field_index] == 100)
1139 tally[n][TALLY_DEPTH]++;
1146 * Look for any field that is always wild and drop it from the config
1147 * Only deactivate if all fields for a given input loop are deactivated.
1149 for (n = 1; n < config->num_fields; n++) {
1150 if (config->defs[n].input_index !=
1151 config->defs[n - 1].input_index) {
1152 for (m = start; m < n; m++)
1153 tally[m][TALLY_DEACTIVATED] = deactivate;
1154 fields_deactivated += deactivate;
1159 /* if the field is not always completely wild */
1160 if (tally[n][TALLY_100] != tally[n][TALLY_0])
1164 for (m = start; m < n; m++)
1165 tally[m][TALLY_DEACTIVATED] = deactivate;
1167 fields_deactivated += deactivate;
1169 /* remove deactivated fields */
1170 if (fields_deactivated) {
1173 for (k = 0; k < config->num_fields; k++) {
1174 if (tally[k][TALLY_DEACTIVATED] == 0) {
1175 memmove(&tally[l][0], &tally[k][0],
1176 TALLY_NUM * sizeof(tally[0][0]));
1177 memmove(&config->defs[l++],
1179 sizeof(struct rte_acl_field_def));
1182 config->num_fields = l;
1187 rule_cmp_wildness(struct rte_acl_build_rule *r1, struct rte_acl_build_rule *r2)
1191 for (n = 1; n < r1->config->num_fields; n++) {
1192 int field_index = r1->config->defs[n].field_index;
1194 if (r1->wildness[field_index] != r2->wildness[field_index])
1195 return r1->wildness[field_index] -
1196 r2->wildness[field_index];
1202 * Split the rte_acl_build_rule list into two lists.
1205 rule_list_split(struct rte_acl_build_rule *source,
1206 struct rte_acl_build_rule **list_a,
1207 struct rte_acl_build_rule **list_b)
1209 struct rte_acl_build_rule *fast;
1210 struct rte_acl_build_rule *slow;
1212 if (source == NULL || source->next == NULL) {
1213 /* length < 2 cases */
1218 fast = source->next;
1219 /* Advance 'fast' two nodes, and advance 'slow' one node */
1220 while (fast != NULL) {
1227 /* 'slow' is before the midpoint in the list, so split it in two
1230 *list_b = slow->next;
1236 * Merge two sorted lists.
1238 static struct rte_acl_build_rule *
1239 rule_list_sorted_merge(struct rte_acl_build_rule *a,
1240 struct rte_acl_build_rule *b)
1242 struct rte_acl_build_rule *result = NULL;
1243 struct rte_acl_build_rule **last_next = &result;
1249 } else if (b == NULL) {
1253 if (rule_cmp_wildness(a, b) >= 0) {
1255 last_next = &a->next;
1259 last_next = &b->next;
1267 * Sort list of rules based on the rules wildness.
1268 * Use recursive mergesort algorithm.
1270 static struct rte_acl_build_rule *
1271 sort_rules(struct rte_acl_build_rule *head)
1273 struct rte_acl_build_rule *a;
1274 struct rte_acl_build_rule *b;
1276 /* Base case -- length 0 or 1 */
1277 if (head == NULL || head->next == NULL)
1280 /* Split head into 'a' and 'b' sublists */
1281 rule_list_split(head, &a, &b);
1283 /* Recursively sort the sublists */
1287 /* answer = merge the two sorted lists together */
1288 return rule_list_sorted_merge(a, b);
1292 acl_build_index(const struct rte_acl_config *config, uint32_t *data_index)
1295 int32_t last_header;
1300 for (n = 0; n < config->num_fields; n++) {
1301 if (last_header != config->defs[n].input_index) {
1302 last_header = config->defs[n].input_index;
1303 data_index[m++] = config->defs[n].offset;
1310 static struct rte_acl_build_rule *
1311 build_one_trie(struct acl_build_context *context,
1312 struct rte_acl_build_rule *rule_sets[RTE_ACL_MAX_TRIES],
1313 uint32_t n, int32_t node_max)
1315 struct rte_acl_build_rule *last;
1316 struct rte_acl_config *config;
1318 config = rule_sets[n]->config;
1320 acl_rule_stats(rule_sets[n], config);
1321 rule_sets[n] = sort_rules(rule_sets[n]);
1323 context->tries[n].type = RTE_ACL_FULL_TRIE;
1324 context->tries[n].count = 0;
1326 context->tries[n].num_data_indexes = acl_build_index(config,
1327 context->data_indexes[n]);
1328 context->tries[n].data_index = context->data_indexes[n];
1330 context->cur_node_max = node_max;
1332 context->bld_tries[n].trie = build_trie(context, rule_sets[n],
1333 &last, &context->tries[n].count);
1339 acl_build_tries(struct acl_build_context *context,
1340 struct rte_acl_build_rule *head)
1342 uint32_t n, num_tries;
1343 struct rte_acl_config *config;
1344 struct rte_acl_build_rule *last;
1345 struct rte_acl_build_rule *rule_sets[RTE_ACL_MAX_TRIES];
1347 config = head->config;
1348 rule_sets[0] = head;
1350 /* initialize tries */
1351 for (n = 0; n < RTE_DIM(context->tries); n++) {
1352 context->tries[n].type = RTE_ACL_UNUSED_TRIE;
1353 context->bld_tries[n].trie = NULL;
1354 context->tries[n].count = 0;
1357 context->tries[0].type = RTE_ACL_FULL_TRIE;
1359 /* calc wildness of each field of each rule */
1360 acl_calc_wildness(head, config);
1362 for (n = 0;; n = num_tries) {
1366 last = build_one_trie(context, rule_sets, n, context->node_max);
1367 if (context->bld_tries[n].trie == NULL) {
1368 RTE_LOG(ERR, ACL, "Build of %u-th trie failed\n", n);
1372 /* Build of the last trie completed. */
1376 if (num_tries == RTE_DIM(context->tries)) {
1378 "Exceeded max number of tries: %u\n",
1383 /* Trie is getting too big, split remaining rule set. */
1384 rule_sets[num_tries] = last->next;
1386 acl_free_node(context, context->bld_tries[n].trie);
1388 /* Create a new copy of config for remaining rules. */
1389 config = acl_build_alloc(context, 1, sizeof(*config));
1390 memcpy(config, rule_sets[n]->config, sizeof(*config));
1392 /* Make remaining rules use new config. */
1393 for (head = rule_sets[num_tries]; head != NULL;
1395 head->config = config;
1398 * Rebuild the trie for the reduced rule-set.
1399 * Don't try to split it any further.
1401 last = build_one_trie(context, rule_sets, n, INT32_MAX);
1402 if (context->bld_tries[n].trie == NULL || last != NULL) {
1403 RTE_LOG(ERR, ACL, "Build of %u-th trie failed\n", n);
1409 context->num_tries = num_tries;
1414 acl_build_log(const struct acl_build_context *ctx)
1418 RTE_LOG(DEBUG, ACL, "Build phase for ACL \"%s\":\n"
1419 "node limit for tree split: %u\n"
1420 "nodes created: %u\n"
1421 "memory consumed: %zu\n",
1427 for (n = 0; n < RTE_DIM(ctx->tries); n++) {
1428 if (ctx->tries[n].count != 0)
1430 "trie %u: number of rules: %u, indexes: %u\n",
1431 n, ctx->tries[n].count,
1432 ctx->tries[n].num_data_indexes);
1437 acl_build_rules(struct acl_build_context *bcx)
1439 struct rte_acl_build_rule *br, *head;
1440 const struct rte_acl_rule *rule;
1442 uint32_t fn, i, n, num;
1445 fn = bcx->cfg.num_fields;
1446 n = bcx->acx->num_rules;
1447 ofs = n * sizeof(*br);
1448 sz = ofs + n * fn * sizeof(*wp);
1450 br = tb_alloc(&bcx->pool, sz);
1452 wp = (uint32_t *)((uintptr_t)br + ofs);
1456 for (i = 0; i != n; i++) {
1457 rule = (const struct rte_acl_rule *)
1458 ((uintptr_t)bcx->acx->rules + bcx->acx->rule_sz * i);
1459 if ((rule->data.category_mask & bcx->category_mask) != 0) {
1460 br[num].next = head;
1461 br[num].config = &bcx->cfg;
1463 br[num].wildness = wp;
1470 bcx->num_rules = num;
1471 bcx->build_rules = head;
1477 * Copy data_indexes for each trie into RT location.
1480 acl_set_data_indexes(struct rte_acl_ctx *ctx)
1485 for (i = 0; i != ctx->num_tries; i++) {
1486 n = ctx->trie[i].num_data_indexes;
1487 memcpy(ctx->data_indexes + ofs, ctx->trie[i].data_index,
1488 n * sizeof(ctx->data_indexes[0]));
1489 ctx->trie[i].data_index = ctx->data_indexes + ofs;
1490 ofs += RTE_ACL_MAX_FIELDS;
1495 * Internal routine, performs 'build' phase of trie generation:
1496 * - setups build context.
1497 * - analizes given set of rules.
1498 * - builds internal tree(s).
1501 acl_bld(struct acl_build_context *bcx, struct rte_acl_ctx *ctx,
1502 const struct rte_acl_config *cfg, uint32_t node_max)
1506 /* setup build context. */
1507 memset(bcx, 0, sizeof(*bcx));
1509 bcx->pool.alignment = ACL_POOL_ALIGN;
1510 bcx->pool.min_alloc = ACL_POOL_ALLOC_MIN;
1512 bcx->category_mask = RTE_LEN2MASK(bcx->cfg.num_categories,
1513 typeof(bcx->category_mask));
1514 bcx->node_max = node_max;
1516 rc = sigsetjmp(bcx->pool.fail, 0);
1518 /* build phase runs out of memory. */
1521 "ACL context: %s, %s() failed with error code: %d\n",
1522 bcx->acx->name, __func__, rc);
1526 /* Create a build rules copy. */
1527 rc = acl_build_rules(bcx);
1531 /* No rules to build for that context+config */
1532 if (bcx->build_rules == NULL) {
1535 /* build internal trie representation. */
1536 rc = acl_build_tries(bcx, bcx->build_rules);
1542 * Check that parameters for acl_build() are valid.
1545 acl_check_bld_param(struct rte_acl_ctx *ctx, const struct rte_acl_config *cfg)
1547 static const size_t field_sizes[] = {
1548 sizeof(uint8_t), sizeof(uint16_t),
1549 sizeof(uint32_t), sizeof(uint64_t),
1554 if (ctx == NULL || cfg == NULL || cfg->num_categories == 0 ||
1555 cfg->num_categories > RTE_ACL_MAX_CATEGORIES ||
1556 cfg->num_fields == 0 ||
1557 cfg->num_fields > RTE_ACL_MAX_FIELDS)
1560 for (i = 0; i != cfg->num_fields; i++) {
1561 if (cfg->defs[i].type > RTE_ACL_FIELD_TYPE_BITMASK) {
1563 "ACL context: %s, invalid type: %hhu for %u-th field\n",
1564 ctx->name, cfg->defs[i].type, i);
1568 j != RTE_DIM(field_sizes) &&
1569 cfg->defs[i].size != field_sizes[j];
1573 if (j == RTE_DIM(field_sizes)) {
1575 "ACL context: %s, invalid size: %hhu for %u-th field\n",
1576 ctx->name, cfg->defs[i].size, i);
1585 * With current ACL implementation first field in the rule definition
1586 * has always to be one byte long. Though for optimising *classify*
1587 * implementation it might be useful to be able to use 4B reads
1588 * (as we do for rest of the fields).
1589 * This function checks input config to determine is it safe to do 4B
1590 * loads for first ACL field. For that we need to make sure that
1591 * first field in our rule definition doesn't have the biggest offset,
1592 * i.e. we still do have other fields located after the first one.
1593 * Contrary if first field has the largest offset, then it means
1594 * first field can occupy the very last byte in the input data buffer,
1595 * and we have to do single byte load for it.
1598 get_first_load_size(const struct rte_acl_config *cfg)
1600 uint32_t i, max_ofs, ofs;
1605 for (i = 0; i != cfg->num_fields; i++) {
1606 if (cfg->defs[i].field_index == 0)
1607 ofs = cfg->defs[i].offset;
1608 else if (max_ofs < cfg->defs[i].offset)
1609 max_ofs = cfg->defs[i].offset;
1612 return (ofs < max_ofs) ? sizeof(uint32_t) : sizeof(uint8_t);
1616 rte_acl_build(struct rte_acl_ctx *ctx, const struct rte_acl_config *cfg)
1621 struct acl_build_context bcx;
1623 rc = acl_check_bld_param(ctx, cfg);
1627 acl_build_reset(ctx);
1629 if (cfg->max_size == 0) {
1631 max_size = SIZE_MAX;
1634 max_size = cfg->max_size;
1637 for (rc = -ERANGE; n >= NODE_MIN && rc == -ERANGE; n /= 2) {
1639 /* perform build phase. */
1640 rc = acl_bld(&bcx, ctx, cfg, n);
1643 /* allocate and fill run-time structures. */
1644 rc = rte_acl_gen(ctx, bcx.tries, bcx.bld_tries,
1645 bcx.num_tries, bcx.cfg.num_categories,
1646 RTE_ACL_MAX_FIELDS * RTE_DIM(bcx.tries) *
1647 sizeof(ctx->data_indexes[0]), max_size);
1649 /* set data indexes. */
1650 acl_set_data_indexes(ctx);
1652 /* determine can we always do 4B load */
1653 ctx->first_load_sz = get_first_load_size(cfg);
1655 /* copy in build config. */
1660 acl_build_log(&bcx);
1662 /* cleanup after build. */
1663 tb_free_pool(&bcx.pool);