1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2020 Mellanox Technologies, Ltd
5 #include <rte_malloc.h>
6 #include <rte_cycles.h>
8 #include <mlx5_malloc.h>
9 #include <mlx5_common_os.h>
12 #include "mlx5_flow.h"
15 * Destroy Completion Queue used for ASO access.
21 mlx5_aso_cq_destroy(struct mlx5_aso_cq *cq)
24 claim_zero(mlx5_devx_cmd_destroy(cq->cq));
26 claim_zero(mlx5_glue->devx_umem_dereg(cq->umem_obj));
28 mlx5_free((void *)(uintptr_t)cq->umem_buf);
29 memset(cq, 0, sizeof(*cq));
33 * Create Completion Queue used for ASO access.
36 * Context returned from mlx5 open_device() glue function.
38 * Pointer to CQ to create.
39 * @param[in] log_desc_n
40 * Log of number of descriptors in queue.
42 * Socket to use for allocation.
43 * @param[in] uar_page_id
49 * 0 on success, a negative errno value otherwise and rte_errno is set.
52 mlx5_aso_cq_create(void *ctx, struct mlx5_aso_cq *cq, uint16_t log_desc_n,
53 int socket, int uar_page_id, uint32_t eqn)
55 struct mlx5_devx_cq_attr attr = { 0 };
56 size_t pgsize = sysconf(_SC_PAGESIZE);
58 uint16_t cq_size = 1 << log_desc_n;
60 cq->log_desc_n = log_desc_n;
61 umem_size = sizeof(struct mlx5_cqe) * cq_size + sizeof(*cq->db_rec) * 2;
62 cq->umem_buf = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, umem_size,
65 DRV_LOG(ERR, "Failed to allocate memory for CQ.");
69 cq->umem_obj = mlx5_glue->devx_umem_reg(ctx,
70 (void *)(uintptr_t)cq->umem_buf,
72 IBV_ACCESS_LOCAL_WRITE);
74 DRV_LOG(ERR, "Failed to register umem for aso CQ.");
77 attr.q_umem_valid = 1;
78 attr.db_umem_valid = 1;
79 attr.use_first_only = 0;
80 attr.overrun_ignore = 0;
81 attr.uar_page_id = uar_page_id;
82 attr.q_umem_id = mlx5_os_get_umem_id(cq->umem_obj);
83 attr.q_umem_offset = 0;
84 attr.db_umem_id = attr.q_umem_id;
85 attr.db_umem_offset = sizeof(struct mlx5_cqe) * cq_size;
87 attr.log_cq_size = log_desc_n;
88 attr.log_page_size = rte_log2_u32(pgsize);
89 cq->cq = mlx5_devx_cmd_create_cq(ctx, &attr);
92 cq->db_rec = RTE_PTR_ADD(cq->umem_buf, (uintptr_t)attr.db_umem_offset);
94 memset((void *)(uintptr_t)cq->umem_buf, 0xFF, attr.db_umem_offset);
97 mlx5_aso_cq_destroy(cq);
108 mlx5_aso_devx_dereg_mr(struct mlx5_aso_devx_mr *mr)
110 claim_zero(mlx5_devx_cmd_destroy(mr->mkey));
111 if (!mr->is_indirect && mr->umem)
112 claim_zero(mlx5_glue->devx_umem_dereg(mr->umem));
114 memset(mr, 0, sizeof(*mr));
118 * Register Memory Region.
121 * Context returned from mlx5 open_device() glue function.
125 * Pointer to MR to create.
127 * Socket to use for allocation.
129 * Protection Domain number to use.
132 * 0 on success, a negative errno value otherwise and rte_errno is set.
135 mlx5_aso_devx_reg_mr(void *ctx, size_t length, struct mlx5_aso_devx_mr *mr,
138 struct mlx5_devx_mkey_attr mkey_attr;
140 mr->buf = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, length, 4096,
143 DRV_LOG(ERR, "Failed to create ASO bits mem for MR by Devx.");
146 mr->umem = mlx5_glue->devx_umem_reg(ctx, mr->buf, length,
147 IBV_ACCESS_LOCAL_WRITE);
149 DRV_LOG(ERR, "Failed to register Umem for MR by Devx.");
152 mkey_attr.addr = (uintptr_t)mr->buf;
153 mkey_attr.size = length;
154 mkey_attr.umem_id = mlx5_os_get_umem_id(mr->umem);
156 mkey_attr.pg_access = 1;
157 mkey_attr.klm_array = NULL;
158 mkey_attr.klm_num = 0;
159 mkey_attr.relaxed_ordering_read = 0;
160 mkey_attr.relaxed_ordering_write = 0;
161 mr->mkey = mlx5_devx_cmd_mkey_create(ctx, &mkey_attr);
163 DRV_LOG(ERR, "Failed to create direct Mkey.");
167 mr->is_indirect = false;
171 claim_zero(mlx5_glue->devx_umem_dereg(mr->umem));
177 * Destroy Send Queue used for ASO access.
183 mlx5_aso_destroy_sq(struct mlx5_aso_sq *sq)
186 mlx5_glue->devx_umem_dereg(sq->wqe_umem);
190 mlx5_free((void *)(uintptr_t)sq->umem_buf);
194 mlx5_devx_cmd_destroy(sq->sq);
198 mlx5_aso_cq_destroy(&sq->cq);
199 mlx5_aso_devx_dereg_mr(&sq->mr);
200 memset(sq, 0, sizeof(*sq));
204 * Initialize Send Queue used for ASO access.
207 * ASO SQ to initialize.
210 mlx5_aso_init_sq(struct mlx5_aso_sq *sq)
212 volatile struct mlx5_aso_wqe *restrict wqe;
214 int size = 1 << sq->log_desc_n;
217 /* All the next fields state should stay constant. */
218 for (i = 0, wqe = &sq->wqes[0]; i < size; ++i, ++wqe) {
219 wqe->general_cseg.sq_ds = rte_cpu_to_be_32((sq->sqn << 8) |
220 (sizeof(*wqe) >> 4));
221 wqe->aso_cseg.lkey = rte_cpu_to_be_32(sq->mr.mkey->id);
222 addr = (uint64_t)((uint64_t *)sq->mr.buf + i *
223 MLX5_ASO_AGE_ACTIONS_PER_POOL / 64);
224 wqe->aso_cseg.va_h = rte_cpu_to_be_32((uint32_t)(addr >> 32));
225 wqe->aso_cseg.va_l_r = rte_cpu_to_be_32((uint32_t)addr | 1u);
226 wqe->aso_cseg.operand_masks = rte_cpu_to_be_32
228 (ASO_OPER_LOGICAL_OR << ASO_CSEG_COND_OPER_OFFSET) |
229 (ASO_OP_ALWAYS_TRUE << ASO_CSEG_COND_1_OPER_OFFSET) |
230 (ASO_OP_ALWAYS_TRUE << ASO_CSEG_COND_0_OPER_OFFSET) |
231 (BYTEWISE_64BYTE << ASO_CSEG_DATA_MASK_MODE_OFFSET));
232 wqe->aso_cseg.data_mask = RTE_BE64(UINT64_MAX);
237 * Create Send Queue used for ASO access.
240 * Context returned from mlx5 open_device() glue function.
242 * Pointer to SQ to create.
244 * Socket to use for allocation.
246 * User Access Region object.
248 * Protection Domain number to use.
251 * @param[in] log_desc_n
252 * Log of number of descriptors in queue.
255 * 0 on success, a negative errno value otherwise and rte_errno is set.
258 mlx5_aso_sq_create(void *ctx, struct mlx5_aso_sq *sq, int socket,
259 struct mlx5dv_devx_uar *uar, uint32_t pdn,
260 uint32_t eqn, uint16_t log_desc_n)
262 struct mlx5_devx_create_sq_attr attr = { 0 };
263 struct mlx5_devx_modify_sq_attr modify_attr = { 0 };
264 size_t pgsize = sysconf(_SC_PAGESIZE);
265 struct mlx5_devx_wq_attr *wq_attr = &attr.wq_attr;
266 uint32_t sq_desc_n = 1 << log_desc_n;
267 uint32_t wq_size = sizeof(struct mlx5_aso_wqe) * sq_desc_n;
270 if (mlx5_aso_devx_reg_mr(ctx, (MLX5_ASO_AGE_ACTIONS_PER_POOL / 8) *
271 sq_desc_n, &sq->mr, socket, pdn))
273 if (mlx5_aso_cq_create(ctx, &sq->cq, log_desc_n, socket,
274 mlx5_os_get_devx_uar_page_id(uar), eqn))
276 sq->log_desc_n = log_desc_n;
277 sq->umem_buf = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, wq_size +
278 sizeof(*sq->db_rec) * 2, 4096, socket);
280 DRV_LOG(ERR, "Can't allocate wqe buffer.");
283 sq->wqe_umem = mlx5_glue->devx_umem_reg(ctx,
284 (void *)(uintptr_t)sq->umem_buf,
286 sizeof(*sq->db_rec) * 2,
287 IBV_ACCESS_LOCAL_WRITE);
289 DRV_LOG(ERR, "Failed to register umem for SQ.");
293 attr.state = MLX5_SQC_STATE_RST;
296 attr.user_index = 0xFFFF;
297 attr.cqn = sq->cq.cq->id;
298 wq_attr->uar_page = mlx5_os_get_devx_uar_page_id(uar);
300 wq_attr->wq_type = MLX5_WQ_TYPE_CYCLIC;
301 wq_attr->log_wq_pg_sz = rte_log2_u32(pgsize);
302 wq_attr->wq_umem_id = mlx5_os_get_umem_id(sq->wqe_umem);
303 wq_attr->wq_umem_offset = 0;
304 wq_attr->wq_umem_valid = 1;
305 wq_attr->log_wq_stride = 6;
306 wq_attr->log_wq_sz = rte_log2_u32(wq_size) - 6;
307 wq_attr->dbr_umem_id = wq_attr->wq_umem_id;
308 wq_attr->dbr_addr = wq_size;
309 wq_attr->dbr_umem_valid = 1;
310 sq->sq = mlx5_devx_cmd_create_sq(ctx, &attr);
312 DRV_LOG(ERR, "Can't create sq object.");
316 modify_attr.state = MLX5_SQC_STATE_RDY;
317 ret = mlx5_devx_cmd_modify_sq(sq->sq, &modify_attr);
319 DRV_LOG(ERR, "Can't change sq state to ready.");
326 sq->sqn = sq->sq->id;
327 sq->db_rec = RTE_PTR_ADD(sq->umem_buf, (uintptr_t)(wq_attr->dbr_addr));
328 sq->uar_addr = (volatile uint64_t *)((uint8_t *)uar->base_addr + 0x800);
329 mlx5_aso_init_sq(sq);
332 mlx5_aso_destroy_sq(sq);
337 * API to create and initialize Send Queue used for ASO access.
340 * Pointer to shared device context.
343 * 0 on success, a negative errno value otherwise and rte_errno is set.
346 mlx5_aso_queue_init(struct mlx5_dev_ctx_shared *sh)
348 return mlx5_aso_sq_create(sh->ctx, &sh->aso_age_mng->aso_sq, 0,
349 sh->tx_uar, sh->pdn, sh->eqn,
350 MLX5_ASO_QUEUE_LOG_DESC);
354 * API to destroy Send Queue used for ASO access.
357 * Pointer to shared device context.
360 mlx5_aso_queue_uninit(struct mlx5_dev_ctx_shared *sh)
362 mlx5_aso_destroy_sq(&sh->aso_age_mng->aso_sq);
366 * Write a burst of WQEs to ASO SQ.
369 * ASO management data, contains the SQ.
371 * Index of the last valid pool.
374 * Number of WQEs in burst.
377 mlx5_aso_sq_enqueue_burst(struct mlx5_aso_age_mng *mng, uint16_t n)
379 volatile struct mlx5_aso_wqe *wqe;
380 struct mlx5_aso_sq *sq = &mng->aso_sq;
381 struct mlx5_aso_age_pool *pool;
382 uint16_t size = 1 << sq->log_desc_n;
383 uint16_t mask = size - 1;
385 uint16_t start_head = sq->head;
387 max = RTE_MIN(size - (uint16_t)(sq->head - sq->tail), n - sq->next);
390 sq->elts[start_head & mask].burst_size = max;
392 wqe = &sq->wqes[sq->head & mask];
393 rte_prefetch0(&sq->wqes[(sq->head + 1) & mask]);
395 rte_spinlock_lock(&mng->resize_sl);
396 pool = mng->pools[sq->next];
397 rte_spinlock_unlock(&mng->resize_sl);
398 sq->elts[sq->head & mask].pool = pool;
399 wqe->general_cseg.misc =
400 rte_cpu_to_be_32(((struct mlx5_devx_obj *)
401 (pool->flow_hit_aso_obj))->id);
402 wqe->general_cseg.flags = RTE_BE32(MLX5_COMP_ONLY_FIRST_ERR <<
403 MLX5_COMP_MODE_OFFSET);
404 wqe->general_cseg.opcode = rte_cpu_to_be_32
405 (MLX5_OPCODE_ACCESS_ASO |
406 (ASO_OPC_MOD_FLOW_HIT <<
407 WQE_CSEG_OPC_MOD_OFFSET) |
409 WQE_CSEG_WQE_INDEX_OFFSET));
410 sq->pi += 2; /* Each WQE contains 2 WQEBB's. */
415 wqe->general_cseg.flags = RTE_BE32(MLX5_COMP_ALWAYS <<
416 MLX5_COMP_MODE_OFFSET);
418 sq->db_rec[MLX5_SND_DBR] = rte_cpu_to_be_32(sq->pi);
420 *sq->uar_addr = *(volatile uint64_t *)wqe; /* Assume 64 bit ARCH.*/
422 return sq->elts[start_head & mask].burst_size;
426 * Debug utility function. Dump contents of error CQE and WQE.
434 mlx5_aso_dump_err_objs(volatile uint32_t *cqe, volatile uint32_t *wqe)
438 DRV_LOG(ERR, "Error cqe:");
439 for (i = 0; i < 16; i += 4)
440 DRV_LOG(ERR, "%08X %08X %08X %08X", cqe[i], cqe[i + 1],
441 cqe[i + 2], cqe[i + 3]);
442 DRV_LOG(ERR, "\nError wqe:");
443 for (i = 0; i < (int)sizeof(struct mlx5_aso_wqe) / 4; i += 4)
444 DRV_LOG(ERR, "%08X %08X %08X %08X", wqe[i], wqe[i + 1],
445 wqe[i + 2], wqe[i + 3]);
449 * Handle case of error CQE.
455 mlx5_aso_cqe_err_handle(struct mlx5_aso_sq *sq)
457 struct mlx5_aso_cq *cq = &sq->cq;
458 uint32_t idx = cq->cq_ci & ((1 << cq->log_desc_n) - 1);
459 volatile struct mlx5_err_cqe *cqe =
460 (volatile struct mlx5_err_cqe *)&cq->cqes[idx];
463 idx = rte_be_to_cpu_16(cqe->wqe_counter) & (1u << sq->log_desc_n);
464 mlx5_aso_dump_err_objs((volatile uint32_t *)cqe,
465 (volatile uint32_t *)&sq->wqes[idx]);
469 * Update ASO objects upon completion.
472 * Shared device context.
474 * Number of completed ASO objects.
477 mlx5_aso_age_action_update(struct mlx5_dev_ctx_shared *sh, uint16_t n)
479 struct mlx5_aso_age_mng *mng = sh->aso_age_mng;
480 struct mlx5_aso_sq *sq = &mng->aso_sq;
481 struct mlx5_age_info *age_info;
482 const uint16_t size = 1 << sq->log_desc_n;
483 const uint16_t mask = size - 1;
484 const uint64_t curr = MLX5_CURR_TIME_SEC;
485 uint16_t expected = AGE_CANDIDATE;
488 for (i = 0; i < n; ++i) {
489 uint16_t idx = (sq->tail + i) & mask;
490 struct mlx5_aso_age_pool *pool = sq->elts[idx].pool;
491 uint64_t diff = curr - pool->time_of_last_age_check;
492 uint64_t *addr = sq->mr.buf;
495 addr += idx * MLX5_ASO_AGE_ACTIONS_PER_POOL / 64;
496 pool->time_of_last_age_check = curr;
497 for (j = 0; j < MLX5_ASO_AGE_ACTIONS_PER_POOL; j++) {
498 struct mlx5_aso_age_action *act = &pool->actions[j];
499 struct mlx5_age_param *ap = &act->age_params;
505 if (__atomic_load_n(&ap->state, __ATOMIC_RELAXED) !=
510 u8addr = (uint8_t *)addr;
511 hit = (u8addr[byte] >> offset) & 0x1;
513 __atomic_store_n(&ap->sec_since_last_hit, 0,
516 struct mlx5_priv *priv;
518 __atomic_fetch_add(&ap->sec_since_last_hit,
519 diff, __ATOMIC_RELAXED);
520 /* If timeout passed add to aged-out list. */
521 if (ap->sec_since_last_hit <= ap->timeout)
524 rte_eth_devices[ap->port_id].data->dev_private;
525 age_info = GET_PORT_AGE_INFO(priv);
526 rte_spinlock_lock(&age_info->aged_sl);
527 if (__atomic_compare_exchange_n(&ap->state,
533 LIST_INSERT_HEAD(&age_info->aged_aso,
535 MLX5_AGE_SET(age_info,
538 rte_spinlock_unlock(&age_info->aged_sl);
542 mlx5_age_event_prepare(sh);
546 * Handle completions from WQEs sent to ASO SQ.
549 * Shared device context.
552 * Number of CQEs handled.
555 mlx5_aso_completion_handle(struct mlx5_dev_ctx_shared *sh)
557 struct mlx5_aso_age_mng *mng = sh->aso_age_mng;
558 struct mlx5_aso_sq *sq = &mng->aso_sq;
559 struct mlx5_aso_cq *cq = &sq->cq;
560 volatile struct mlx5_cqe *restrict cqe;
561 const unsigned int cq_size = 1 << cq->log_desc_n;
562 const unsigned int mask = cq_size - 1;
564 uint32_t next_idx = cq->cq_ci & mask;
565 const uint16_t max = (uint16_t)(sq->head - sq->tail);
572 next_idx = (cq->cq_ci + 1) & mask;
573 rte_prefetch0(&cq->cqes[next_idx]);
574 cqe = &cq->cqes[idx];
575 ret = check_cqe(cqe, cq_size, cq->cq_ci);
577 * Be sure owner read is done before any other cookie field or
581 if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
582 if (likely(ret == MLX5_CQE_STATUS_HW_OWN))
584 mlx5_aso_cqe_err_handle(sq);
586 i += sq->elts[(sq->tail + i) & mask].burst_size;
591 mlx5_aso_age_action_update(sh, i);
594 cq->db_rec[0] = rte_cpu_to_be_32(cq->cq_ci);
600 * Periodically read CQEs and send WQEs to ASO SQ.
603 * Shared device context containing the ASO SQ.
606 mlx5_flow_aso_alarm(void *arg)
608 struct mlx5_dev_ctx_shared *sh = arg;
609 struct mlx5_aso_sq *sq = &sh->aso_age_mng->aso_sq;
613 rte_spinlock_lock(&sh->aso_age_mng->resize_sl);
614 n = sh->aso_age_mng->next;
615 rte_spinlock_unlock(&sh->aso_age_mng->resize_sl);
616 mlx5_aso_completion_handle(sh);
618 /* End of loop: wait 1 second. */
622 mlx5_aso_sq_enqueue_burst(sh->aso_age_mng, n);
623 if (rte_eal_alarm_set(us, mlx5_flow_aso_alarm, sh))
624 DRV_LOG(ERR, "Cannot reinitialize aso alarm.");
628 * API to start ASO access using ASO SQ.
631 * Pointer to shared device context.
634 * 0 on success, a negative errno value otherwise and rte_errno is set.
637 mlx5_aso_queue_start(struct mlx5_dev_ctx_shared *sh)
639 if (rte_eal_alarm_set(US_PER_S, mlx5_flow_aso_alarm, sh)) {
640 DRV_LOG(ERR, "Cannot reinitialize ASO age alarm.");
647 * API to stop ASO access using ASO SQ.
650 * Pointer to shared device context.
653 * 0 on success, a negative errno value otherwise and rte_errno is set.
656 mlx5_aso_queue_stop(struct mlx5_dev_ctx_shared *sh)
660 if (!sh->aso_age_mng->aso_sq.sq)
664 rte_eal_alarm_cancel(mlx5_flow_aso_alarm, sh);
665 if (rte_errno != EINPROGRESS)