1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright 2021 Mellanox Technologies, Ltd 3 */ 4 5 #include <rte_malloc.h> 6 #include <rte_log.h> 7 #include <rte_errno.h> 8 #include <rte_bus_pci.h> 9 #include <rte_spinlock.h> 10 #include <rte_comp.h> 11 #include <rte_compressdev.h> 12 #include <rte_compressdev_pmd.h> 13 14 #include <mlx5_glue.h> 15 #include <mlx5_common.h> 16 #include <mlx5_devx_cmds.h> 17 #include <mlx5_common_os.h> 18 #include <mlx5_common_devx.h> 19 #include <mlx5_common_mr.h> 20 #include <mlx5_prm.h> 21 22 #include "mlx5_compress_utils.h" 23 24 #define MLX5_COMPRESS_DRIVER_NAME mlx5_compress 25 #define MLX5_COMPRESS_MAX_QPS 1024 26 #define MLX5_COMP_MAX_WIN_SIZE_CONF 6u 27 28 struct mlx5_compress_xform { 29 LIST_ENTRY(mlx5_compress_xform) next; 30 enum rte_comp_xform_type type; 31 enum rte_comp_checksum_type csum_type; 32 uint32_t opcode; 33 uint32_t gga_ctrl1; /* BE. */ 34 }; 35 36 struct mlx5_compress_priv { 37 TAILQ_ENTRY(mlx5_compress_priv) next; 38 struct rte_compressdev *compressdev; 39 struct mlx5_common_device *cdev; /* Backend mlx5 device. */ 40 void *uar; 41 uint8_t min_block_size; 42 /* Minimum huffman block size supported by the device. */ 43 struct rte_compressdev_config dev_config; 44 LIST_HEAD(xform_list, mlx5_compress_xform) xform_list; 45 rte_spinlock_t xform_sl; 46 volatile uint64_t *uar_addr; 47 /* HCA caps*/ 48 uint32_t mmo_decomp_sq:1; 49 uint32_t mmo_decomp_qp:1; 50 uint32_t mmo_comp_sq:1; 51 uint32_t mmo_comp_qp:1; 52 uint32_t mmo_dma_sq:1; 53 uint32_t mmo_dma_qp:1; 54 #ifndef RTE_ARCH_64 55 rte_spinlock_t uar32_sl; 56 #endif /* RTE_ARCH_64 */ 57 }; 58 59 struct mlx5_compress_qp { 60 uint16_t qp_id; 61 uint16_t entries_n; 62 uint16_t pi; 63 uint16_t ci; 64 struct mlx5_mr_ctrl mr_ctrl; 65 int socket_id; 66 struct mlx5_devx_cq cq; 67 struct mlx5_devx_qp qp; 68 struct mlx5_pmd_mr opaque_mr; 69 struct rte_comp_op **ops; 70 struct mlx5_compress_priv *priv; 71 struct rte_compressdev_stats stats; 72 }; 73 74 TAILQ_HEAD(mlx5_compress_privs, mlx5_compress_priv) mlx5_compress_priv_list = 75 TAILQ_HEAD_INITIALIZER(mlx5_compress_priv_list); 76 static pthread_mutex_t priv_list_lock = PTHREAD_MUTEX_INITIALIZER; 77 78 int mlx5_compress_logtype; 79 80 static const struct rte_compressdev_capabilities mlx5_caps[] = { 81 { 82 .algo = RTE_COMP_ALGO_NULL, 83 .comp_feature_flags = RTE_COMP_FF_ADLER32_CHECKSUM | 84 RTE_COMP_FF_CRC32_CHECKSUM | 85 RTE_COMP_FF_CRC32_ADLER32_CHECKSUM | 86 RTE_COMP_FF_SHAREABLE_PRIV_XFORM, 87 }, 88 { 89 .algo = RTE_COMP_ALGO_DEFLATE, 90 .comp_feature_flags = RTE_COMP_FF_ADLER32_CHECKSUM | 91 RTE_COMP_FF_CRC32_CHECKSUM | 92 RTE_COMP_FF_CRC32_ADLER32_CHECKSUM | 93 RTE_COMP_FF_SHAREABLE_PRIV_XFORM | 94 RTE_COMP_FF_HUFFMAN_FIXED | 95 RTE_COMP_FF_HUFFMAN_DYNAMIC, 96 .window_size = {.min = 10, .max = 15, .increment = 1}, 97 }, 98 { 99 .algo = RTE_COMP_ALGO_LIST_END, 100 } 101 }; 102 103 static void 104 mlx5_compress_dev_info_get(struct rte_compressdev *dev, 105 struct rte_compressdev_info *info) 106 { 107 RTE_SET_USED(dev); 108 if (info != NULL) { 109 info->max_nb_queue_pairs = MLX5_COMPRESS_MAX_QPS; 110 info->feature_flags = RTE_COMPDEV_FF_HW_ACCELERATED; 111 info->capabilities = mlx5_caps; 112 } 113 } 114 115 static int 116 mlx5_compress_dev_configure(struct rte_compressdev *dev, 117 struct rte_compressdev_config *config) 118 { 119 struct mlx5_compress_priv *priv; 120 121 if (dev == NULL || config == NULL) 122 return -EINVAL; 123 priv = dev->data->dev_private; 124 priv->dev_config = *config; 125 return 0; 126 } 127 128 static int 129 mlx5_compress_dev_close(struct rte_compressdev *dev) 130 { 131 RTE_SET_USED(dev); 132 return 0; 133 } 134 135 static int 136 mlx5_compress_qp_release(struct rte_compressdev *dev, uint16_t qp_id) 137 { 138 struct mlx5_compress_qp *qp = dev->data->queue_pairs[qp_id]; 139 140 if (qp->qp.qp != NULL) 141 mlx5_devx_qp_destroy(&qp->qp); 142 if (qp->cq.cq != NULL) 143 mlx5_devx_cq_destroy(&qp->cq); 144 if (qp->opaque_mr.obj != NULL) { 145 void *opaq = qp->opaque_mr.addr; 146 147 mlx5_common_verbs_dereg_mr(&qp->opaque_mr); 148 if (opaq != NULL) 149 rte_free(opaq); 150 } 151 mlx5_mr_btree_free(&qp->mr_ctrl.cache_bh); 152 rte_free(qp); 153 dev->data->queue_pairs[qp_id] = NULL; 154 return 0; 155 } 156 157 static void 158 mlx5_compress_init_qp(struct mlx5_compress_qp *qp) 159 { 160 volatile struct mlx5_gga_wqe *restrict wqe = 161 (volatile struct mlx5_gga_wqe *)qp->qp.wqes; 162 volatile struct mlx5_gga_compress_opaque *opaq = qp->opaque_mr.addr; 163 const uint32_t sq_ds = rte_cpu_to_be_32((qp->qp.qp->id << 8) | 4u); 164 const uint32_t flags = RTE_BE32(MLX5_COMP_ALWAYS << 165 MLX5_COMP_MODE_OFFSET); 166 const uint32_t opaq_lkey = rte_cpu_to_be_32(qp->opaque_mr.lkey); 167 int i; 168 169 /* All the next fields state should stay constant. */ 170 for (i = 0; i < qp->entries_n; ++i, ++wqe) { 171 wqe->sq_ds = sq_ds; 172 wqe->flags = flags; 173 wqe->opaque_lkey = opaq_lkey; 174 wqe->opaque_vaddr = rte_cpu_to_be_64 175 ((uint64_t)(uintptr_t)&opaq[i]); 176 } 177 } 178 179 static int 180 mlx5_compress_qp_setup(struct rte_compressdev *dev, uint16_t qp_id, 181 uint32_t max_inflight_ops, int socket_id) 182 { 183 struct mlx5_compress_priv *priv = dev->data->dev_private; 184 struct mlx5_compress_qp *qp; 185 struct mlx5_devx_cq_attr cq_attr = { 186 .uar_page_id = mlx5_os_get_devx_uar_page_id(priv->uar), 187 }; 188 struct mlx5_devx_qp_attr qp_attr = { 189 .pd = priv->cdev->pdn, 190 .uar_index = mlx5_os_get_devx_uar_page_id(priv->uar), 191 .user_index = qp_id, 192 }; 193 uint32_t log_ops_n = rte_log2_u32(max_inflight_ops); 194 uint32_t alloc_size = sizeof(*qp); 195 void *opaq_buf; 196 int ret; 197 198 alloc_size = RTE_ALIGN(alloc_size, RTE_CACHE_LINE_SIZE); 199 alloc_size += sizeof(struct rte_comp_op *) * (1u << log_ops_n); 200 qp = rte_zmalloc_socket(__func__, alloc_size, RTE_CACHE_LINE_SIZE, 201 socket_id); 202 if (qp == NULL) { 203 DRV_LOG(ERR, "Failed to allocate qp memory."); 204 rte_errno = ENOMEM; 205 return -rte_errno; 206 } 207 dev->data->queue_pairs[qp_id] = qp; 208 if (mlx5_mr_ctrl_init(&qp->mr_ctrl, &priv->cdev->mr_scache.dev_gen, 209 priv->dev_config.socket_id)) { 210 DRV_LOG(ERR, "Cannot allocate MR Btree for qp %u.", 211 (uint32_t)qp_id); 212 rte_errno = ENOMEM; 213 goto err; 214 } 215 opaq_buf = rte_calloc(__func__, (size_t)1 << log_ops_n, 216 sizeof(struct mlx5_gga_compress_opaque), 217 sizeof(struct mlx5_gga_compress_opaque)); 218 if (opaq_buf == NULL) { 219 DRV_LOG(ERR, "Failed to allocate opaque memory."); 220 rte_errno = ENOMEM; 221 goto err; 222 } 223 qp->entries_n = 1 << log_ops_n; 224 qp->socket_id = socket_id; 225 qp->qp_id = qp_id; 226 qp->priv = priv; 227 qp->ops = (struct rte_comp_op **)RTE_ALIGN((uintptr_t)(qp + 1), 228 RTE_CACHE_LINE_SIZE); 229 if (mlx5_common_verbs_reg_mr(priv->cdev->pd, opaq_buf, qp->entries_n * 230 sizeof(struct mlx5_gga_compress_opaque), 231 &qp->opaque_mr) != 0) { 232 rte_free(opaq_buf); 233 DRV_LOG(ERR, "Failed to register opaque MR."); 234 rte_errno = ENOMEM; 235 goto err; 236 } 237 ret = mlx5_devx_cq_create(priv->cdev->ctx, &qp->cq, log_ops_n, &cq_attr, 238 socket_id); 239 if (ret != 0) { 240 DRV_LOG(ERR, "Failed to create CQ."); 241 goto err; 242 } 243 qp_attr.cqn = qp->cq.cq->id; 244 qp_attr.ts_format = 245 mlx5_ts_format_conv(priv->cdev->config.hca_attr.qp_ts_format); 246 qp_attr.rq_size = 0; 247 qp_attr.sq_size = RTE_BIT32(log_ops_n); 248 qp_attr.mmo = priv->mmo_decomp_qp && priv->mmo_comp_qp 249 && priv->mmo_dma_qp; 250 ret = mlx5_devx_qp_create(priv->cdev->ctx, &qp->qp, log_ops_n, &qp_attr, 251 socket_id); 252 if (ret != 0) { 253 DRV_LOG(ERR, "Failed to create QP."); 254 goto err; 255 } 256 mlx5_compress_init_qp(qp); 257 ret = mlx5_devx_qp2rts(&qp->qp, 0); 258 if (ret) 259 goto err; 260 DRV_LOG(INFO, "QP %u: SQN=0x%X CQN=0x%X entries num = %u", 261 (uint32_t)qp_id, qp->qp.qp->id, qp->cq.cq->id, qp->entries_n); 262 return 0; 263 err: 264 mlx5_compress_qp_release(dev, qp_id); 265 return -1; 266 } 267 268 static int 269 mlx5_compress_xform_free(struct rte_compressdev *dev, void *xform) 270 { 271 struct mlx5_compress_priv *priv = dev->data->dev_private; 272 273 rte_spinlock_lock(&priv->xform_sl); 274 LIST_REMOVE((struct mlx5_compress_xform *)xform, next); 275 rte_spinlock_unlock(&priv->xform_sl); 276 rte_free(xform); 277 return 0; 278 } 279 280 static int 281 mlx5_compress_xform_create(struct rte_compressdev *dev, 282 const struct rte_comp_xform *xform, 283 void **private_xform) 284 { 285 struct mlx5_compress_priv *priv = dev->data->dev_private; 286 struct mlx5_compress_xform *xfrm; 287 uint32_t size; 288 289 if (xform->type == RTE_COMP_COMPRESS && xform->compress.level == 290 RTE_COMP_LEVEL_NONE) { 291 DRV_LOG(ERR, "Non-compressed block is not supported."); 292 return -ENOTSUP; 293 } 294 if ((xform->type == RTE_COMP_COMPRESS && xform->compress.hash_algo != 295 RTE_COMP_HASH_ALGO_NONE) || (xform->type == RTE_COMP_DECOMPRESS && 296 xform->decompress.hash_algo != RTE_COMP_HASH_ALGO_NONE)) { 297 DRV_LOG(ERR, "SHA is not supported."); 298 return -ENOTSUP; 299 } 300 xfrm = rte_zmalloc_socket(__func__, sizeof(*xfrm), 0, 301 priv->dev_config.socket_id); 302 if (xfrm == NULL) 303 return -ENOMEM; 304 xfrm->opcode = MLX5_OPCODE_MMO; 305 xfrm->type = xform->type; 306 switch (xform->type) { 307 case RTE_COMP_COMPRESS: 308 switch (xform->compress.algo) { 309 case RTE_COMP_ALGO_NULL: 310 xfrm->opcode += MLX5_OPC_MOD_MMO_DMA << 311 WQE_CSEG_OPC_MOD_OFFSET; 312 break; 313 case RTE_COMP_ALGO_DEFLATE: 314 size = 1 << xform->compress.window_size; 315 size /= MLX5_GGA_COMP_WIN_SIZE_UNITS; 316 xfrm->gga_ctrl1 += RTE_MIN(rte_log2_u32(size), 317 MLX5_COMP_MAX_WIN_SIZE_CONF) << 318 WQE_GGA_COMP_WIN_SIZE_OFFSET; 319 switch (xform->compress.level) { 320 case RTE_COMP_LEVEL_PMD_DEFAULT: 321 size = MLX5_GGA_COMP_LOG_BLOCK_SIZE_MAX; 322 break; 323 case RTE_COMP_LEVEL_MAX: 324 size = priv->min_block_size; 325 break; 326 default: 327 size = RTE_MAX(MLX5_GGA_COMP_LOG_BLOCK_SIZE_MAX 328 + 1 - xform->compress.level, 329 priv->min_block_size); 330 } 331 xfrm->gga_ctrl1 += RTE_MIN(size, 332 MLX5_GGA_COMP_LOG_BLOCK_SIZE_MAX) << 333 WQE_GGA_COMP_BLOCK_SIZE_OFFSET; 334 xfrm->opcode += MLX5_OPC_MOD_MMO_COMP << 335 WQE_CSEG_OPC_MOD_OFFSET; 336 size = xform->compress.deflate.huffman == 337 RTE_COMP_HUFFMAN_DYNAMIC ? 338 MLX5_GGA_COMP_LOG_DYNAMIC_SIZE_MAX : 339 MLX5_GGA_COMP_LOG_DYNAMIC_SIZE_MIN; 340 xfrm->gga_ctrl1 += size << 341 WQE_GGA_COMP_DYNAMIC_SIZE_OFFSET; 342 break; 343 default: 344 goto err; 345 } 346 xfrm->csum_type = xform->compress.chksum; 347 break; 348 case RTE_COMP_DECOMPRESS: 349 switch (xform->decompress.algo) { 350 case RTE_COMP_ALGO_NULL: 351 xfrm->opcode += MLX5_OPC_MOD_MMO_DMA << 352 WQE_CSEG_OPC_MOD_OFFSET; 353 break; 354 case RTE_COMP_ALGO_DEFLATE: 355 xfrm->opcode += MLX5_OPC_MOD_MMO_DECOMP << 356 WQE_CSEG_OPC_MOD_OFFSET; 357 break; 358 default: 359 goto err; 360 } 361 xfrm->csum_type = xform->decompress.chksum; 362 break; 363 default: 364 DRV_LOG(ERR, "Algorithm %u is not supported.", xform->type); 365 goto err; 366 } 367 DRV_LOG(DEBUG, "New xform: gga ctrl1 = 0x%08X opcode = 0x%08X csum " 368 "type = %d.", xfrm->gga_ctrl1, xfrm->opcode, xfrm->csum_type); 369 xfrm->gga_ctrl1 = rte_cpu_to_be_32(xfrm->gga_ctrl1); 370 rte_spinlock_lock(&priv->xform_sl); 371 LIST_INSERT_HEAD(&priv->xform_list, xfrm, next); 372 rte_spinlock_unlock(&priv->xform_sl); 373 *private_xform = xfrm; 374 return 0; 375 err: 376 rte_free(xfrm); 377 return -ENOTSUP; 378 } 379 380 static void 381 mlx5_compress_dev_stop(struct rte_compressdev *dev) 382 { 383 RTE_SET_USED(dev); 384 } 385 386 static int 387 mlx5_compress_dev_start(struct rte_compressdev *dev) 388 { 389 RTE_SET_USED(dev); 390 return 0; 391 } 392 393 static void 394 mlx5_compress_stats_get(struct rte_compressdev *dev, 395 struct rte_compressdev_stats *stats) 396 { 397 int qp_id; 398 399 for (qp_id = 0; qp_id < dev->data->nb_queue_pairs; qp_id++) { 400 struct mlx5_compress_qp *qp = dev->data->queue_pairs[qp_id]; 401 402 stats->enqueued_count += qp->stats.enqueued_count; 403 stats->dequeued_count += qp->stats.dequeued_count; 404 stats->enqueue_err_count += qp->stats.enqueue_err_count; 405 stats->dequeue_err_count += qp->stats.dequeue_err_count; 406 } 407 } 408 409 static void 410 mlx5_compress_stats_reset(struct rte_compressdev *dev) 411 { 412 int qp_id; 413 414 for (qp_id = 0; qp_id < dev->data->nb_queue_pairs; qp_id++) { 415 struct mlx5_compress_qp *qp = dev->data->queue_pairs[qp_id]; 416 417 memset(&qp->stats, 0, sizeof(qp->stats)); 418 } 419 } 420 421 static struct rte_compressdev_ops mlx5_compress_ops = { 422 .dev_configure = mlx5_compress_dev_configure, 423 .dev_start = mlx5_compress_dev_start, 424 .dev_stop = mlx5_compress_dev_stop, 425 .dev_close = mlx5_compress_dev_close, 426 .dev_infos_get = mlx5_compress_dev_info_get, 427 .stats_get = mlx5_compress_stats_get, 428 .stats_reset = mlx5_compress_stats_reset, 429 .queue_pair_setup = mlx5_compress_qp_setup, 430 .queue_pair_release = mlx5_compress_qp_release, 431 .private_xform_create = mlx5_compress_xform_create, 432 .private_xform_free = mlx5_compress_xform_free, 433 .stream_create = NULL, 434 .stream_free = NULL, 435 }; 436 437 static __rte_always_inline uint32_t 438 mlx5_compress_dseg_set(struct mlx5_compress_qp *qp, 439 volatile struct mlx5_wqe_dseg *restrict dseg, 440 struct rte_mbuf *restrict mbuf, 441 uint32_t offset, uint32_t len) 442 { 443 uintptr_t addr = rte_pktmbuf_mtod_offset(mbuf, uintptr_t, offset); 444 445 dseg->bcount = rte_cpu_to_be_32(len); 446 dseg->lkey = mlx5_mr_mb2mr(qp->priv->cdev, 0, &qp->mr_ctrl, mbuf); 447 dseg->pbuf = rte_cpu_to_be_64(addr); 448 return dseg->lkey; 449 } 450 451 /* 452 * Provide safe 64bit store operation to mlx5 UAR region for both 32bit and 453 * 64bit architectures. 454 */ 455 static __rte_always_inline void 456 mlx5_compress_uar_write(uint64_t val, struct mlx5_compress_priv *priv) 457 { 458 #ifdef RTE_ARCH_64 459 *priv->uar_addr = val; 460 #else /* !RTE_ARCH_64 */ 461 rte_spinlock_lock(&priv->uar32_sl); 462 *(volatile uint32_t *)priv->uar_addr = val; 463 rte_io_wmb(); 464 *((volatile uint32_t *)priv->uar_addr + 1) = val >> 32; 465 rte_spinlock_unlock(&priv->uar32_sl); 466 #endif 467 } 468 469 static uint16_t 470 mlx5_compress_enqueue_burst(void *queue_pair, struct rte_comp_op **ops, 471 uint16_t nb_ops) 472 { 473 struct mlx5_compress_qp *qp = queue_pair; 474 volatile struct mlx5_gga_wqe *wqes = (volatile struct mlx5_gga_wqe *) 475 qp->qp.wqes, *wqe; 476 struct mlx5_compress_xform *xform; 477 struct rte_comp_op *op; 478 uint16_t mask = qp->entries_n - 1; 479 uint16_t remain = qp->entries_n - (qp->pi - qp->ci); 480 uint16_t idx; 481 bool invalid; 482 483 if (remain < nb_ops) 484 nb_ops = remain; 485 else 486 remain = nb_ops; 487 if (unlikely(remain == 0)) 488 return 0; 489 do { 490 idx = qp->pi & mask; 491 wqe = &wqes[idx]; 492 rte_prefetch0(&wqes[(qp->pi + 1) & mask]); 493 op = *ops++; 494 xform = op->private_xform; 495 /* 496 * Check operation arguments and error cases: 497 * - Operation type must be state-less. 498 * - Compress operation flush flag must be FULL or FINAL. 499 * - Source and destination buffers must be mapped internally. 500 */ 501 invalid = op->op_type != RTE_COMP_OP_STATELESS || 502 (xform->type == RTE_COMP_COMPRESS && 503 op->flush_flag < RTE_COMP_FLUSH_FULL); 504 if (unlikely(invalid || 505 (mlx5_compress_dseg_set(qp, &wqe->gather, 506 op->m_src, 507 op->src.offset, 508 op->src.length) == 509 UINT32_MAX) || 510 (mlx5_compress_dseg_set(qp, &wqe->scatter, 511 op->m_dst, 512 op->dst.offset, 513 rte_pktmbuf_pkt_len(op->m_dst) - 514 op->dst.offset) == 515 UINT32_MAX))) { 516 op->status = invalid ? RTE_COMP_OP_STATUS_INVALID_ARGS : 517 RTE_COMP_OP_STATUS_ERROR; 518 nb_ops -= remain; 519 if (unlikely(nb_ops == 0)) 520 return 0; 521 break; 522 } 523 wqe->gga_ctrl1 = xform->gga_ctrl1; 524 wqe->opcode = rte_cpu_to_be_32(xform->opcode + (qp->pi << 8)); 525 qp->ops[idx] = op; 526 qp->pi++; 527 } while (--remain); 528 qp->stats.enqueued_count += nb_ops; 529 rte_io_wmb(); 530 qp->qp.db_rec[MLX5_SND_DBR] = rte_cpu_to_be_32(qp->pi); 531 rte_wmb(); 532 mlx5_compress_uar_write(*(volatile uint64_t *)wqe, qp->priv); 533 rte_wmb(); 534 return nb_ops; 535 } 536 537 static void 538 mlx5_compress_dump_err_objs(volatile uint32_t *cqe, volatile uint32_t *wqe, 539 volatile uint32_t *opaq) 540 { 541 size_t i; 542 543 DRV_LOG(ERR, "Error cqe:"); 544 for (i = 0; i < sizeof(struct mlx5_err_cqe) >> 2; i += 4) 545 DRV_LOG(ERR, "%08X %08X %08X %08X", cqe[i], cqe[i + 1], 546 cqe[i + 2], cqe[i + 3]); 547 DRV_LOG(ERR, "\nError wqe:"); 548 for (i = 0; i < sizeof(struct mlx5_gga_wqe) >> 2; i += 4) 549 DRV_LOG(ERR, "%08X %08X %08X %08X", wqe[i], wqe[i + 1], 550 wqe[i + 2], wqe[i + 3]); 551 DRV_LOG(ERR, "\nError opaq:"); 552 for (i = 0; i < sizeof(struct mlx5_gga_compress_opaque) >> 2; i += 4) 553 DRV_LOG(ERR, "%08X %08X %08X %08X", opaq[i], opaq[i + 1], 554 opaq[i + 2], opaq[i + 3]); 555 } 556 557 static void 558 mlx5_compress_cqe_err_handle(struct mlx5_compress_qp *qp, 559 struct rte_comp_op *op) 560 { 561 const uint32_t idx = qp->ci & (qp->entries_n - 1); 562 volatile struct mlx5_err_cqe *cqe = (volatile struct mlx5_err_cqe *) 563 &qp->cq.cqes[idx]; 564 volatile struct mlx5_gga_wqe *wqes = (volatile struct mlx5_gga_wqe *) 565 qp->qp.wqes; 566 volatile struct mlx5_gga_compress_opaque *opaq = qp->opaque_mr.addr; 567 568 op->status = RTE_COMP_OP_STATUS_ERROR; 569 op->consumed = 0; 570 op->produced = 0; 571 op->output_chksum = 0; 572 op->debug_status = rte_be_to_cpu_32(opaq[idx].syndrom) | 573 ((uint64_t)rte_be_to_cpu_32(cqe->syndrome) << 32); 574 mlx5_compress_dump_err_objs((volatile uint32_t *)cqe, 575 (volatile uint32_t *)&wqes[idx], 576 (volatile uint32_t *)&opaq[idx]); 577 qp->stats.dequeue_err_count++; 578 } 579 580 static uint16_t 581 mlx5_compress_dequeue_burst(void *queue_pair, struct rte_comp_op **ops, 582 uint16_t nb_ops) 583 { 584 struct mlx5_compress_qp *qp = queue_pair; 585 volatile struct mlx5_compress_xform *restrict xform; 586 volatile struct mlx5_cqe *restrict cqe; 587 volatile struct mlx5_gga_compress_opaque *opaq = qp->opaque_mr.addr; 588 struct rte_comp_op *restrict op; 589 const unsigned int cq_size = qp->entries_n; 590 const unsigned int mask = cq_size - 1; 591 uint32_t idx; 592 uint32_t next_idx = qp->ci & mask; 593 const uint16_t max = RTE_MIN((uint16_t)(qp->pi - qp->ci), nb_ops); 594 uint16_t i = 0; 595 int ret; 596 597 if (unlikely(max == 0)) 598 return 0; 599 do { 600 idx = next_idx; 601 next_idx = (qp->ci + 1) & mask; 602 rte_prefetch0(&qp->cq.cqes[next_idx]); 603 rte_prefetch0(qp->ops[next_idx]); 604 op = qp->ops[idx]; 605 cqe = &qp->cq.cqes[idx]; 606 ret = check_cqe(cqe, cq_size, qp->ci); 607 /* 608 * Be sure owner read is done before any other cookie field or 609 * opaque field. 610 */ 611 rte_io_rmb(); 612 if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) { 613 if (likely(ret == MLX5_CQE_STATUS_HW_OWN)) 614 break; 615 mlx5_compress_cqe_err_handle(qp, op); 616 } else { 617 xform = op->private_xform; 618 op->status = RTE_COMP_OP_STATUS_SUCCESS; 619 op->consumed = op->src.length; 620 op->produced = rte_be_to_cpu_32(cqe->byte_cnt); 621 MLX5_ASSERT(cqe->byte_cnt == 622 opaq[idx].scattered_length); 623 switch (xform->csum_type) { 624 case RTE_COMP_CHECKSUM_CRC32: 625 op->output_chksum = (uint64_t)rte_be_to_cpu_32 626 (opaq[idx].crc32); 627 break; 628 case RTE_COMP_CHECKSUM_ADLER32: 629 op->output_chksum = (uint64_t)rte_be_to_cpu_32 630 (opaq[idx].adler32) << 32; 631 break; 632 case RTE_COMP_CHECKSUM_CRC32_ADLER32: 633 op->output_chksum = (uint64_t)rte_be_to_cpu_32 634 (opaq[idx].crc32) | 635 ((uint64_t)rte_be_to_cpu_32 636 (opaq[idx].adler32) << 32); 637 break; 638 default: 639 break; 640 } 641 } 642 ops[i++] = op; 643 qp->ci++; 644 } while (i < max); 645 if (likely(i != 0)) { 646 rte_io_wmb(); 647 qp->cq.db_rec[0] = rte_cpu_to_be_32(qp->ci); 648 qp->stats.dequeued_count += i; 649 } 650 return i; 651 } 652 653 static void 654 mlx5_compress_uar_release(struct mlx5_compress_priv *priv) 655 { 656 if (priv->uar != NULL) { 657 mlx5_glue->devx_free_uar(priv->uar); 658 priv->uar = NULL; 659 } 660 } 661 662 static int 663 mlx5_compress_uar_prepare(struct mlx5_compress_priv *priv) 664 { 665 priv->uar = mlx5_devx_alloc_uar(priv->cdev->ctx, -1); 666 if (priv->uar == NULL || mlx5_os_get_devx_uar_reg_addr(priv->uar) == 667 NULL) { 668 rte_errno = errno; 669 DRV_LOG(ERR, "Failed to allocate UAR."); 670 return -1; 671 } 672 priv->uar_addr = mlx5_os_get_devx_uar_reg_addr(priv->uar); 673 MLX5_ASSERT(priv->uar_addr); 674 #ifndef RTE_ARCH_64 675 rte_spinlock_init(&priv->uar32_sl); 676 #endif /* RTE_ARCH_64 */ 677 return 0; 678 } 679 680 static int 681 mlx5_compress_dev_probe(struct mlx5_common_device *cdev) 682 { 683 struct rte_compressdev *compressdev; 684 struct mlx5_compress_priv *priv; 685 struct mlx5_hca_attr *attr = &cdev->config.hca_attr; 686 struct rte_compressdev_pmd_init_params init_params = { 687 .name = "", 688 .socket_id = cdev->dev->numa_node, 689 }; 690 const char *ibdev_name = mlx5_os_get_ctx_device_name(cdev->ctx); 691 692 if (rte_eal_process_type() != RTE_PROC_PRIMARY) { 693 DRV_LOG(ERR, "Non-primary process type is not supported."); 694 rte_errno = ENOTSUP; 695 return -rte_errno; 696 } 697 if ((attr->mmo_compress_sq_en == 0 || attr->mmo_decompress_sq_en == 0 || 698 attr->mmo_dma_sq_en == 0) && (attr->mmo_compress_qp_en == 0 || 699 attr->mmo_decompress_qp_en == 0 || attr->mmo_dma_qp_en == 0)) { 700 DRV_LOG(ERR, "Not enough capabilities to support compress " 701 "operations, maybe old FW/OFED version?"); 702 rte_errno = ENOTSUP; 703 return -ENOTSUP; 704 } 705 compressdev = rte_compressdev_pmd_create(ibdev_name, cdev->dev, 706 sizeof(*priv), &init_params); 707 if (compressdev == NULL) { 708 DRV_LOG(ERR, "Failed to create device \"%s\".", ibdev_name); 709 return -ENODEV; 710 } 711 DRV_LOG(INFO, 712 "Compress device %s was created successfully.", ibdev_name); 713 compressdev->dev_ops = &mlx5_compress_ops; 714 compressdev->dequeue_burst = mlx5_compress_dequeue_burst; 715 compressdev->enqueue_burst = mlx5_compress_enqueue_burst; 716 compressdev->feature_flags = RTE_COMPDEV_FF_HW_ACCELERATED; 717 priv = compressdev->data->dev_private; 718 priv->mmo_decomp_sq = attr->mmo_decompress_sq_en; 719 priv->mmo_decomp_qp = attr->mmo_decompress_qp_en; 720 priv->mmo_comp_sq = attr->mmo_compress_sq_en; 721 priv->mmo_comp_qp = attr->mmo_compress_qp_en; 722 priv->mmo_dma_sq = attr->mmo_dma_sq_en; 723 priv->mmo_dma_qp = attr->mmo_dma_qp_en; 724 priv->cdev = cdev; 725 priv->compressdev = compressdev; 726 priv->min_block_size = attr->compress_min_block_size; 727 if (mlx5_compress_uar_prepare(priv) != 0) { 728 rte_compressdev_pmd_destroy(priv->compressdev); 729 return -1; 730 } 731 pthread_mutex_lock(&priv_list_lock); 732 TAILQ_INSERT_TAIL(&mlx5_compress_priv_list, priv, next); 733 pthread_mutex_unlock(&priv_list_lock); 734 return 0; 735 } 736 737 static int 738 mlx5_compress_dev_remove(struct mlx5_common_device *cdev) 739 { 740 struct mlx5_compress_priv *priv = NULL; 741 742 pthread_mutex_lock(&priv_list_lock); 743 TAILQ_FOREACH(priv, &mlx5_compress_priv_list, next) 744 if (priv->compressdev->device == cdev->dev) 745 break; 746 if (priv) 747 TAILQ_REMOVE(&mlx5_compress_priv_list, priv, next); 748 pthread_mutex_unlock(&priv_list_lock); 749 if (priv) { 750 mlx5_compress_uar_release(priv); 751 rte_compressdev_pmd_destroy(priv->compressdev); 752 } 753 return 0; 754 } 755 756 static const struct rte_pci_id mlx5_compress_pci_id_map[] = { 757 { 758 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, 759 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF) 760 }, 761 { 762 .vendor_id = 0 763 } 764 }; 765 766 static struct mlx5_class_driver mlx5_compress_driver = { 767 .drv_class = MLX5_CLASS_COMPRESS, 768 .name = RTE_STR(MLX5_COMPRESS_DRIVER_NAME), 769 .id_table = mlx5_compress_pci_id_map, 770 .probe = mlx5_compress_dev_probe, 771 .remove = mlx5_compress_dev_remove, 772 }; 773 774 RTE_INIT(rte_mlx5_compress_init) 775 { 776 mlx5_common_init(); 777 if (mlx5_glue != NULL) 778 mlx5_class_driver_register(&mlx5_compress_driver); 779 } 780 781 RTE_LOG_REGISTER_DEFAULT(mlx5_compress_logtype, NOTICE) 782 RTE_PMD_EXPORT_NAME(MLX5_COMPRESS_DRIVER_NAME, __COUNTER__); 783 RTE_PMD_REGISTER_PCI_TABLE(MLX5_COMPRESS_DRIVER_NAME, mlx5_compress_pci_id_map); 784 RTE_PMD_REGISTER_KMOD_DEP(MLX5_COMPRESS_DRIVER_NAME, "* ib_uverbs & mlx5_core & mlx5_ib"); 785