1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2014 Intel Corporation
8 #define QRANGE_MIN ((uint8_t)INT8_MIN)
10 #define RTE_ACL_VERIFY(exp) do { \
12 rte_panic("line %d\tassert \"" #exp "\" failed\n", __LINE__); \
15 struct acl_node_counters {
25 struct rte_acl_indices {
34 acl_gen_log_stats(const struct rte_acl_ctx *ctx,
35 const struct acl_node_counters *counts,
36 const struct rte_acl_indices *indices,
39 RTE_LOG(DEBUG, ACL, "Gen phase for ACL \"%s\":\n"
40 "runtime memory footprint on socket %d:\n"
41 "single nodes/bytes used: %d/%zu\n"
42 "quad nodes/vectors/bytes used: %d/%d/%zu\n"
43 "DFA nodes/group64/bytes used: %d/%d/%zu\n"
44 "match nodes/bytes used: %d/%zu\n"
46 "max limit: %zu bytes\n",
47 ctx->name, ctx->socket_id,
48 counts->single, counts->single * sizeof(uint64_t),
49 counts->quad, counts->quad_vectors,
50 (indices->quad_index - indices->dfa_index) * sizeof(uint64_t),
51 counts->dfa, counts->dfa_gr64,
52 indices->dfa_index * sizeof(uint64_t),
54 counts->match * sizeof(struct rte_acl_match_results),
60 acl_dfa_gen_idx(const struct rte_acl_node *node, uint32_t index)
66 for (i = 0; i != RTE_DIM(node->dfa_gr64); i++) {
67 RTE_ACL_VERIFY(node->dfa_gr64[i] < RTE_ACL_DFA_GR64_NUM);
68 RTE_ACL_VERIFY(node->dfa_gr64[i] < node->fanout);
69 idx |= (i - node->dfa_gr64[i]) <<
70 (6 + RTE_ACL_DFA_GR64_BIT * i);
73 return idx << (CHAR_BIT * sizeof(index)) | index | node->node_type;
77 acl_dfa_fill_gr64(const struct rte_acl_node *node,
78 const uint64_t src[RTE_ACL_DFA_SIZE], uint64_t dst[RTE_ACL_DFA_SIZE])
82 for (i = 0; i != RTE_DIM(node->dfa_gr64); i++) {
83 memcpy(dst + node->dfa_gr64[i] * RTE_ACL_DFA_GR64_SIZE,
84 src + i * RTE_ACL_DFA_GR64_SIZE,
85 RTE_ACL_DFA_GR64_SIZE * sizeof(dst[0]));
90 acl_dfa_count_gr64(const uint64_t array_ptr[RTE_ACL_DFA_SIZE],
91 uint8_t gr64[RTE_ACL_DFA_GR64_NUM])
96 for (i = 0; i != RTE_ACL_DFA_GR64_NUM; i++) {
98 for (j = 0; j != i; j++) {
99 if (memcmp(array_ptr + i * RTE_ACL_DFA_GR64_SIZE,
100 array_ptr + j * RTE_ACL_DFA_GR64_SIZE,
101 RTE_ACL_DFA_GR64_SIZE *
102 sizeof(array_ptr[0])) == 0)
105 gr64[i] = (j != i) ? gr64[j] : k++;
112 acl_node_fill_dfa(const struct rte_acl_node *node,
113 uint64_t dfa[RTE_ACL_DFA_SIZE], uint64_t no_match, int32_t resolved)
116 uint32_t ranges, last_bit;
117 struct rte_acl_node *child;
118 struct rte_acl_bitset *bits;
123 for (n = 0; n < RTE_ACL_DFA_SIZE; n++)
126 for (x = 0; x < node->num_ptrs; x++) {
128 child = node->ptrs[x].ptr;
132 bits = &node->ptrs[x].values;
133 for (n = 0; n < RTE_ACL_DFA_SIZE; n++) {
135 if (bits->bits[n / (sizeof(bits_t) * CHAR_BIT)] &
136 (1U << (n % (sizeof(bits_t) * CHAR_BIT)))) {
138 dfa[n] = resolved ? child->node_index : x;
139 ranges += (last_bit == 0);
151 * Counts the number of groups of sequential bits that are
152 * either 0 or 1, as specified by the zero_one parameter. This is used to
153 * calculate the number of ranges in a node to see if it fits in a quad range
157 acl_count_sequential_groups(struct rte_acl_bitset *bits, int zero_one)
159 int n, ranges, last_bit;
162 last_bit = zero_one ^ 1;
164 for (n = QRANGE_MIN; n < UINT8_MAX + 1; n++) {
165 if (bits->bits[n / (sizeof(bits_t) * 8)] &
166 (1U << (n % (sizeof(bits_t) * 8)))) {
167 if (zero_one == 1 && last_bit != 1)
171 if (zero_one == 0 && last_bit != 0)
176 for (n = 0; n < QRANGE_MIN; n++) {
177 if (bits->bits[n / (sizeof(bits_t) * 8)] &
178 (1U << (n % (sizeof(bits_t) * CHAR_BIT)))) {
179 if (zero_one == 1 && last_bit != 1)
183 if (zero_one == 0 && last_bit != 0)
193 * Count number of ranges spanned by the node's pointers
196 acl_count_fanout(struct rte_acl_node *node)
201 if (node->fanout != 0)
204 ranges = acl_count_sequential_groups(&node->values, 0);
206 for (n = 0; n < node->num_ptrs; n++) {
207 if (node->ptrs[n].ptr != NULL)
208 ranges += acl_count_sequential_groups(
209 &node->ptrs[n].values, 1);
212 node->fanout = ranges;
217 * Determine the type of nodes and count each type
220 acl_count_trie_types(struct acl_node_counters *counts,
221 struct rte_acl_node *node, uint64_t no_match, int force_dfa)
225 uint64_t dfa[RTE_ACL_DFA_SIZE];
227 /* skip if this node has been counted */
228 if (node->node_type != (uint32_t)RTE_ACL_NODE_UNDEFINED)
231 if (node->match_flag != 0 || node->num_ptrs == 0) {
233 node->node_type = RTE_ACL_NODE_MATCH;
237 num_ptrs = acl_count_fanout(node);
239 /* Force type to dfa */
241 num_ptrs = RTE_ACL_DFA_SIZE;
243 /* determine node type based on number of ranges */
246 node->node_type = RTE_ACL_NODE_SINGLE;
247 } else if (num_ptrs <= RTE_ACL_QUAD_MAX) {
249 counts->quad_vectors += node->fanout;
250 node->node_type = RTE_ACL_NODE_QRANGE;
253 node->node_type = RTE_ACL_NODE_DFA;
254 if (force_dfa != 0) {
255 /* always expand to a max number of nodes. */
256 for (n = 0; n != RTE_DIM(node->dfa_gr64); n++)
257 node->dfa_gr64[n] = n;
260 acl_node_fill_dfa(node, dfa, no_match, 0);
261 node->fanout = acl_dfa_count_gr64(dfa, node->dfa_gr64);
263 counts->dfa_gr64 += node->fanout;
267 * recursively count the types of all children
269 for (n = 0; n < node->num_ptrs; n++) {
270 if (node->ptrs[n].ptr != NULL)
271 acl_count_trie_types(counts, node->ptrs[n].ptr,
277 acl_add_ptrs(struct rte_acl_node *node, uint64_t *node_array, uint64_t no_match,
282 uint64_t *node_a, index, dfa[RTE_ACL_DFA_SIZE];
284 acl_node_fill_dfa(node, dfa, no_match, resolved);
287 * Rather than going from 0 to 256, the range count and
288 * the layout are from 80-ff then 0-7f due to signed compare
291 if (node->node_type == RTE_ACL_NODE_QRANGE) {
295 index = dfa[QRANGE_MIN];
298 for (x = QRANGE_MIN + 1; x < UINT8_MAX + 1; x++) {
299 if (dfa[x] != index) {
302 node->transitions[m++] = (uint8_t)(x - 1);
306 for (x = 0; x < INT8_MAX + 1; x++) {
307 if (dfa[x] != index) {
310 node->transitions[m++] = (uint8_t)(x - 1);
314 /* fill unused locations with max value - nothing is greater */
315 for (; m < RTE_ACL_QUAD_SIZE; m++)
316 node->transitions[m] = INT8_MAX;
318 RTE_ACL_VERIFY(m <= RTE_ACL_QUAD_SIZE);
320 } else if (node->node_type == RTE_ACL_NODE_DFA && resolved) {
321 acl_dfa_fill_gr64(node, dfa, node_array);
326 * Routine that allocates space for this node and recursively calls
327 * to allocate space for each child. Once all the children are allocated,
328 * then resolve all transitions for this node.
331 acl_gen_node(struct rte_acl_node *node, uint64_t *node_array,
332 uint64_t no_match, struct rte_acl_indices *index, int num_categories)
334 uint32_t n, sz, *qtrp;
336 struct rte_acl_match_results *match;
338 if (node->node_index != RTE_ACL_NODE_UNDEFINED)
343 switch (node->node_type) {
344 case RTE_ACL_NODE_DFA:
345 array_ptr = &node_array[index->dfa_index];
346 node->node_index = acl_dfa_gen_idx(node, index->dfa_index);
347 sz = node->fanout * RTE_ACL_DFA_GR64_SIZE;
348 index->dfa_index += sz;
349 for (n = 0; n < sz; n++)
350 array_ptr[n] = no_match;
352 case RTE_ACL_NODE_SINGLE:
353 node->node_index = RTE_ACL_QUAD_SINGLE | index->single_index |
355 array_ptr = &node_array[index->single_index];
356 index->single_index += 1;
357 array_ptr[0] = no_match;
359 case RTE_ACL_NODE_QRANGE:
360 array_ptr = &node_array[index->quad_index];
361 acl_add_ptrs(node, array_ptr, no_match, 0);
362 qtrp = (uint32_t *)node->transitions;
363 node->node_index = qtrp[0];
364 node->node_index <<= sizeof(index->quad_index) * CHAR_BIT;
365 node->node_index |= index->quad_index | node->node_type;
366 index->quad_index += node->fanout;
368 case RTE_ACL_NODE_MATCH:
369 match = ((struct rte_acl_match_results *)
370 (node_array + index->match_start));
371 for (n = 0; n != RTE_DIM(match->results); n++)
372 RTE_ACL_VERIFY(match->results[0] == 0);
373 memcpy(match + index->match_index, node->mrt,
375 node->node_index = index->match_index | node->node_type;
376 index->match_index += 1;
378 case RTE_ACL_NODE_UNDEFINED:
379 RTE_ACL_VERIFY(node->node_type !=
380 (uint32_t)RTE_ACL_NODE_UNDEFINED);
384 /* recursively allocate space for all children */
385 for (n = 0; n < node->num_ptrs; n++) {
386 if (node->ptrs[n].ptr != NULL)
387 acl_gen_node(node->ptrs[n].ptr,
394 /* All children are resolved, resolve this node's pointers */
395 switch (node->node_type) {
396 case RTE_ACL_NODE_DFA:
397 acl_add_ptrs(node, array_ptr, no_match, 1);
399 case RTE_ACL_NODE_SINGLE:
400 for (n = 0; n < node->num_ptrs; n++) {
401 if (node->ptrs[n].ptr != NULL)
402 array_ptr[0] = node->ptrs[n].ptr->node_index;
405 case RTE_ACL_NODE_QRANGE:
406 acl_add_ptrs(node, array_ptr, no_match, 1);
408 case RTE_ACL_NODE_MATCH:
410 case RTE_ACL_NODE_UNDEFINED:
411 RTE_ACL_VERIFY(node->node_type !=
412 (uint32_t)RTE_ACL_NODE_UNDEFINED);
418 acl_calc_counts_indices(struct acl_node_counters *counts,
419 struct rte_acl_indices *indices,
420 struct rte_acl_bld_trie *node_bld_trie, uint32_t num_tries,
425 memset(indices, 0, sizeof(*indices));
426 memset(counts, 0, sizeof(*counts));
428 /* Get stats on nodes */
429 for (n = 0; n < num_tries; n++) {
430 acl_count_trie_types(counts, node_bld_trie[n].trie,
434 indices->dfa_index = RTE_ACL_DFA_SIZE + 1;
435 indices->quad_index = indices->dfa_index +
436 counts->dfa_gr64 * RTE_ACL_DFA_GR64_SIZE;
437 indices->single_index = indices->quad_index + counts->quad_vectors;
438 indices->match_start = indices->single_index + counts->single + 1;
439 indices->match_start = RTE_ALIGN(indices->match_start,
440 (XMM_SIZE / sizeof(uint64_t)));
441 indices->match_index = 1;
445 * Generate the runtime structure using build structure
448 rte_acl_gen(struct rte_acl_ctx *ctx, struct rte_acl_trie *trie,
449 struct rte_acl_bld_trie *node_bld_trie, uint32_t num_tries,
450 uint32_t num_categories, uint32_t data_index_sz, size_t max_size)
454 uint64_t *node_array, no_match;
455 uint32_t n, match_index;
456 struct rte_acl_match_results *match;
457 struct acl_node_counters counts;
458 struct rte_acl_indices indices;
460 no_match = RTE_ACL_NODE_MATCH;
462 /* Fill counts and indices arrays from the nodes. */
463 acl_calc_counts_indices(&counts, &indices,
464 node_bld_trie, num_tries, no_match);
466 /* Allocate runtime memory (align to cache boundary) */
467 total_size = RTE_ALIGN(data_index_sz, RTE_CACHE_LINE_SIZE) +
468 indices.match_start * sizeof(uint64_t) +
469 (counts.match + 1) * sizeof(struct rte_acl_match_results) +
472 if (total_size > max_size) {
474 "Gen phase for ACL ctx \"%s\" exceeds max_size limit, "
475 "bytes required: %zu, allowed: %zu\n",
476 ctx->name, total_size, max_size);
480 mem = rte_zmalloc_socket(ctx->name, total_size, RTE_CACHE_LINE_SIZE,
484 "allocation of %zu bytes on socket %d for %s failed\n",
485 total_size, ctx->socket_id, ctx->name);
489 /* Fill the runtime structure */
490 match_index = indices.match_start;
491 node_array = (uint64_t *)((uintptr_t)mem +
492 RTE_ALIGN(data_index_sz, RTE_CACHE_LINE_SIZE));
495 * Setup the NOMATCH node (a SINGLE at the
496 * highest index, that points to itself)
499 node_array[RTE_ACL_DFA_SIZE] = RTE_ACL_IDLE_NODE;
501 for (n = 0; n < RTE_ACL_DFA_SIZE; n++)
502 node_array[n] = no_match;
504 /* NOMATCH result at index 0 */
505 match = ((struct rte_acl_match_results *)(node_array + match_index));
506 memset(match, 0, sizeof(*match));
508 for (n = 0; n < num_tries; n++) {
510 acl_gen_node(node_bld_trie[n].trie, node_array, no_match,
511 &indices, num_categories);
513 if (node_bld_trie[n].trie->node_index == no_match)
514 trie[n].root_index = 0;
516 trie[n].root_index = node_bld_trie[n].trie->node_index;
520 ctx->mem_sz = total_size;
521 ctx->data_indexes = mem;
522 ctx->num_tries = num_tries;
523 ctx->num_categories = num_categories;
524 ctx->match_index = match_index;
525 ctx->no_match = no_match;
526 ctx->idle = node_array[RTE_ACL_DFA_SIZE];
527 ctx->trans_table = node_array;
528 memcpy(ctx->trie, trie, sizeof(ctx->trie));
530 acl_gen_log_stats(ctx, &counts, &indices, max_size);