xref: /dpdk/drivers/compress/mlx5/mlx5_compress.c (revision d38febb08d57fec29fed27a2d12a507fc6fcdfa1)
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 ibv_context *ctx; /* Device context. */
39 	struct rte_compressdev *cdev;
40 	void *uar;
41 	uint32_t pdn; /* Protection Domain number. */
42 	uint8_t min_block_size;
43 	uint8_t sq_ts_format; /* Whether SQ supports timestamp formats. */
44 	/* Minimum huffman block size supported by the device. */
45 	struct ibv_pd *pd;
46 	struct rte_compressdev_config dev_config;
47 	LIST_HEAD(xform_list, mlx5_compress_xform) xform_list;
48 	rte_spinlock_t xform_sl;
49 	struct mlx5_mr_share_cache mr_scache; /* Global shared MR cache. */
50 	volatile uint64_t *uar_addr;
51 #ifndef RTE_ARCH_64
52 	rte_spinlock_t uar32_sl;
53 #endif /* RTE_ARCH_64 */
54 };
55 
56 struct mlx5_compress_qp {
57 	uint16_t qp_id;
58 	uint16_t entries_n;
59 	uint16_t pi;
60 	uint16_t ci;
61 	struct mlx5_mr_ctrl mr_ctrl;
62 	int socket_id;
63 	struct mlx5_devx_cq cq;
64 	struct mlx5_devx_sq sq;
65 	struct mlx5_pmd_mr opaque_mr;
66 	struct rte_comp_op **ops;
67 	struct mlx5_compress_priv *priv;
68 	struct rte_compressdev_stats stats;
69 };
70 
71 TAILQ_HEAD(mlx5_compress_privs, mlx5_compress_priv) mlx5_compress_priv_list =
72 				TAILQ_HEAD_INITIALIZER(mlx5_compress_priv_list);
73 static pthread_mutex_t priv_list_lock = PTHREAD_MUTEX_INITIALIZER;
74 
75 int mlx5_compress_logtype;
76 
77 static const struct rte_compressdev_capabilities mlx5_caps[] = {
78 	{
79 		.algo = RTE_COMP_ALGO_NULL,
80 		.comp_feature_flags = RTE_COMP_FF_ADLER32_CHECKSUM |
81 				      RTE_COMP_FF_CRC32_CHECKSUM |
82 				      RTE_COMP_FF_CRC32_ADLER32_CHECKSUM |
83 				      RTE_COMP_FF_SHAREABLE_PRIV_XFORM,
84 	},
85 	{
86 		.algo = RTE_COMP_ALGO_DEFLATE,
87 		.comp_feature_flags = RTE_COMP_FF_ADLER32_CHECKSUM |
88 				      RTE_COMP_FF_CRC32_CHECKSUM |
89 				      RTE_COMP_FF_CRC32_ADLER32_CHECKSUM |
90 				      RTE_COMP_FF_SHAREABLE_PRIV_XFORM |
91 				      RTE_COMP_FF_HUFFMAN_FIXED |
92 				      RTE_COMP_FF_HUFFMAN_DYNAMIC,
93 		.window_size = {.min = 10, .max = 15, .increment = 1},
94 	},
95 	{
96 		.algo = RTE_COMP_ALGO_LIST_END,
97 	}
98 };
99 
100 static void
101 mlx5_compress_dev_info_get(struct rte_compressdev *dev,
102 			   struct rte_compressdev_info *info)
103 {
104 	RTE_SET_USED(dev);
105 	if (info != NULL) {
106 		info->max_nb_queue_pairs = MLX5_COMPRESS_MAX_QPS;
107 		info->feature_flags = RTE_COMPDEV_FF_HW_ACCELERATED;
108 		info->capabilities = mlx5_caps;
109 	}
110 }
111 
112 static int
113 mlx5_compress_dev_configure(struct rte_compressdev *dev,
114 			    struct rte_compressdev_config *config)
115 {
116 	struct mlx5_compress_priv *priv;
117 
118 	if (dev == NULL || config == NULL)
119 		return -EINVAL;
120 	priv = dev->data->dev_private;
121 	priv->dev_config = *config;
122 	return 0;
123 }
124 
125 static int
126 mlx5_compress_dev_close(struct rte_compressdev *dev)
127 {
128 	RTE_SET_USED(dev);
129 	return 0;
130 }
131 
132 static int
133 mlx5_compress_qp_release(struct rte_compressdev *dev, uint16_t qp_id)
134 {
135 	struct mlx5_compress_qp *qp = dev->data->queue_pairs[qp_id];
136 
137 	if (qp->sq.sq != NULL)
138 		mlx5_devx_sq_destroy(&qp->sq);
139 	if (qp->cq.cq != NULL)
140 		mlx5_devx_cq_destroy(&qp->cq);
141 	if (qp->opaque_mr.obj != NULL) {
142 		void *opaq = qp->opaque_mr.addr;
143 
144 		mlx5_common_verbs_dereg_mr(&qp->opaque_mr);
145 		if (opaq != NULL)
146 			rte_free(opaq);
147 	}
148 	mlx5_mr_btree_free(&qp->mr_ctrl.cache_bh);
149 	rte_free(qp);
150 	dev->data->queue_pairs[qp_id] = NULL;
151 	return 0;
152 }
153 
154 static void
155 mlx5_compress_init_sq(struct mlx5_compress_qp *qp)
156 {
157 	volatile struct mlx5_gga_wqe *restrict wqe =
158 				    (volatile struct mlx5_gga_wqe *)qp->sq.wqes;
159 	volatile struct mlx5_gga_compress_opaque *opaq = qp->opaque_mr.addr;
160 	const uint32_t sq_ds = rte_cpu_to_be_32((qp->sq.sq->id << 8) | 4u);
161 	const uint32_t flags = RTE_BE32(MLX5_COMP_ALWAYS <<
162 					MLX5_COMP_MODE_OFFSET);
163 	const uint32_t opaq_lkey = rte_cpu_to_be_32(qp->opaque_mr.lkey);
164 	int i;
165 
166 	/* All the next fields state should stay constant. */
167 	for (i = 0; i < qp->entries_n; ++i, ++wqe) {
168 		wqe->sq_ds = sq_ds;
169 		wqe->flags = flags;
170 		wqe->opaque_lkey = opaq_lkey;
171 		wqe->opaque_vaddr = rte_cpu_to_be_64
172 						((uint64_t)(uintptr_t)&opaq[i]);
173 	}
174 }
175 
176 static int
177 mlx5_compress_qp_setup(struct rte_compressdev *dev, uint16_t qp_id,
178 		       uint32_t max_inflight_ops, int socket_id)
179 {
180 	struct mlx5_compress_priv *priv = dev->data->dev_private;
181 	struct mlx5_compress_qp *qp;
182 	struct mlx5_devx_cq_attr cq_attr = {
183 		.uar_page_id = mlx5_os_get_devx_uar_page_id(priv->uar),
184 	};
185 	struct mlx5_devx_create_sq_attr sq_attr = {
186 		.user_index = qp_id,
187 		.wq_attr = (struct mlx5_devx_wq_attr){
188 			.pd = priv->pdn,
189 			.uar_page = mlx5_os_get_devx_uar_page_id(priv->uar),
190 		},
191 	};
192 	struct mlx5_devx_modify_sq_attr modify_attr = {
193 		.state = MLX5_SQC_STATE_RDY,
194 	};
195 	uint32_t log_ops_n = rte_log2_u32(max_inflight_ops);
196 	uint32_t alloc_size = sizeof(*qp);
197 	void *opaq_buf;
198 	int ret;
199 
200 	alloc_size = RTE_ALIGN(alloc_size, RTE_CACHE_LINE_SIZE);
201 	alloc_size += sizeof(struct rte_comp_op *) * (1u << log_ops_n);
202 	qp = rte_zmalloc_socket(__func__, alloc_size, RTE_CACHE_LINE_SIZE,
203 				socket_id);
204 	if (qp == NULL) {
205 		DRV_LOG(ERR, "Failed to allocate qp memory.");
206 		rte_errno = ENOMEM;
207 		return -rte_errno;
208 	}
209 	dev->data->queue_pairs[qp_id] = qp;
210 	if (mlx5_mr_btree_init(&qp->mr_ctrl.cache_bh, MLX5_MR_BTREE_CACHE_N,
211 			       priv->dev_config.socket_id)) {
212 		DRV_LOG(ERR, "Cannot allocate MR Btree for qp %u.",
213 			(uint32_t)qp_id);
214 		rte_errno = ENOMEM;
215 		goto err;
216 	}
217 	opaq_buf = rte_calloc(__func__, (size_t)1 << log_ops_n,
218 			      sizeof(struct mlx5_gga_compress_opaque),
219 			      sizeof(struct mlx5_gga_compress_opaque));
220 	if (opaq_buf == NULL) {
221 		DRV_LOG(ERR, "Failed to allocate opaque memory.");
222 		rte_errno = ENOMEM;
223 		goto err;
224 	}
225 	qp->entries_n = 1 << log_ops_n;
226 	qp->socket_id = socket_id;
227 	qp->qp_id = qp_id;
228 	qp->priv = priv;
229 	qp->ops = (struct rte_comp_op **)RTE_ALIGN((uintptr_t)(qp + 1),
230 						   RTE_CACHE_LINE_SIZE);
231 	if (mlx5_common_verbs_reg_mr(priv->pd, opaq_buf, qp->entries_n *
232 					sizeof(struct mlx5_gga_compress_opaque),
233 							 &qp->opaque_mr) != 0) {
234 		rte_free(opaq_buf);
235 		DRV_LOG(ERR, "Failed to register opaque MR.");
236 		rte_errno = ENOMEM;
237 		goto err;
238 	}
239 	ret = mlx5_devx_cq_create(priv->ctx, &qp->cq, log_ops_n, &cq_attr,
240 				  socket_id);
241 	if (ret != 0) {
242 		DRV_LOG(ERR, "Failed to create CQ.");
243 		goto err;
244 	}
245 	sq_attr.cqn = qp->cq.cq->id;
246 	sq_attr.ts_format = mlx5_ts_format_conv(priv->sq_ts_format);
247 	ret = mlx5_devx_sq_create(priv->ctx, &qp->sq, log_ops_n, &sq_attr,
248 				  socket_id);
249 	if (ret != 0) {
250 		DRV_LOG(ERR, "Failed to create SQ.");
251 		goto err;
252 	}
253 	mlx5_compress_init_sq(qp);
254 	ret = mlx5_devx_cmd_modify_sq(qp->sq.sq, &modify_attr);
255 	if (ret != 0) {
256 		DRV_LOG(ERR, "Can't change SQ state to ready.");
257 		goto err;
258 	}
259 	/* Save pointer of global generation number to check memory event. */
260 	qp->mr_ctrl.dev_gen_ptr = &priv->mr_scache.dev_gen;
261 	DRV_LOG(INFO, "QP %u: SQN=0x%X CQN=0x%X entries num = %u",
262 		(uint32_t)qp_id, qp->sq.sq->id, qp->cq.cq->id, qp->entries_n);
263 	return 0;
264 err:
265 	mlx5_compress_qp_release(dev, qp_id);
266 	return -1;
267 }
268 
269 static int
270 mlx5_compress_xform_free(struct rte_compressdev *dev, void *xform)
271 {
272 	struct mlx5_compress_priv *priv = dev->data->dev_private;
273 
274 	rte_spinlock_lock(&priv->xform_sl);
275 	LIST_REMOVE((struct mlx5_compress_xform *)xform, next);
276 	rte_spinlock_unlock(&priv->xform_sl);
277 	rte_free(xform);
278 	return 0;
279 }
280 
281 static int
282 mlx5_compress_xform_create(struct rte_compressdev *dev,
283 			   const struct rte_comp_xform *xform,
284 			   void **private_xform)
285 {
286 	struct mlx5_compress_priv *priv = dev->data->dev_private;
287 	struct mlx5_compress_xform *xfrm;
288 	uint32_t size;
289 
290 	if (xform->type == RTE_COMP_COMPRESS && xform->compress.level ==
291 							  RTE_COMP_LEVEL_NONE) {
292 		DRV_LOG(ERR, "Non-compressed block is not supported.");
293 		return -ENOTSUP;
294 	}
295 	if ((xform->type == RTE_COMP_COMPRESS && xform->compress.hash_algo !=
296 	     RTE_COMP_HASH_ALGO_NONE) || (xform->type == RTE_COMP_DECOMPRESS &&
297 		      xform->decompress.hash_algo != RTE_COMP_HASH_ALGO_NONE)) {
298 		DRV_LOG(ERR, "SHA is not supported.");
299 		return -ENOTSUP;
300 	}
301 	xfrm = rte_zmalloc_socket(__func__, sizeof(*xfrm), 0,
302 						    priv->dev_config.socket_id);
303 	if (xfrm == NULL)
304 		return -ENOMEM;
305 	xfrm->opcode = MLX5_OPCODE_MMO;
306 	xfrm->type = xform->type;
307 	switch (xform->type) {
308 	case RTE_COMP_COMPRESS:
309 		switch (xform->compress.algo) {
310 		case RTE_COMP_ALGO_NULL:
311 			xfrm->opcode += MLX5_OPC_MOD_MMO_DMA <<
312 							WQE_CSEG_OPC_MOD_OFFSET;
313 			break;
314 		case RTE_COMP_ALGO_DEFLATE:
315 			size = 1 << xform->compress.window_size;
316 			size /= MLX5_GGA_COMP_WIN_SIZE_UNITS;
317 			xfrm->gga_ctrl1 += RTE_MIN(rte_log2_u32(size),
318 					 MLX5_COMP_MAX_WIN_SIZE_CONF) <<
319 						WQE_GGA_COMP_WIN_SIZE_OFFSET;
320 			switch (xform->compress.level) {
321 			case RTE_COMP_LEVEL_PMD_DEFAULT:
322 				size = MLX5_GGA_COMP_LOG_BLOCK_SIZE_MAX;
323 				break;
324 			case RTE_COMP_LEVEL_MAX:
325 				size = priv->min_block_size;
326 				break;
327 			default:
328 				size = RTE_MAX(MLX5_GGA_COMP_LOG_BLOCK_SIZE_MAX
329 					+ 1 - xform->compress.level,
330 					priv->min_block_size);
331 			}
332 			xfrm->gga_ctrl1 += RTE_MIN(size,
333 					    MLX5_GGA_COMP_LOG_BLOCK_SIZE_MAX) <<
334 						 WQE_GGA_COMP_BLOCK_SIZE_OFFSET;
335 			xfrm->opcode += MLX5_OPC_MOD_MMO_COMP <<
336 							WQE_CSEG_OPC_MOD_OFFSET;
337 			size = xform->compress.deflate.huffman ==
338 						      RTE_COMP_HUFFMAN_DYNAMIC ?
339 					    MLX5_GGA_COMP_LOG_DYNAMIC_SIZE_MAX :
340 					     MLX5_GGA_COMP_LOG_DYNAMIC_SIZE_MIN;
341 			xfrm->gga_ctrl1 += size <<
342 					       WQE_GGA_COMP_DYNAMIC_SIZE_OFFSET;
343 			break;
344 		default:
345 			goto err;
346 		}
347 		xfrm->csum_type = xform->compress.chksum;
348 		break;
349 	case RTE_COMP_DECOMPRESS:
350 		switch (xform->decompress.algo) {
351 		case RTE_COMP_ALGO_NULL:
352 			xfrm->opcode += MLX5_OPC_MOD_MMO_DMA <<
353 							WQE_CSEG_OPC_MOD_OFFSET;
354 			break;
355 		case RTE_COMP_ALGO_DEFLATE:
356 			xfrm->opcode += MLX5_OPC_MOD_MMO_DECOMP <<
357 							WQE_CSEG_OPC_MOD_OFFSET;
358 			break;
359 		default:
360 			goto err;
361 		}
362 		xfrm->csum_type = xform->decompress.chksum;
363 		break;
364 	default:
365 		DRV_LOG(ERR, "Algorithm %u is not supported.", xform->type);
366 		goto err;
367 	}
368 	DRV_LOG(DEBUG, "New xform: gga ctrl1 = 0x%08X opcode = 0x%08X csum "
369 		"type = %d.", xfrm->gga_ctrl1, xfrm->opcode, xfrm->csum_type);
370 	xfrm->gga_ctrl1 = rte_cpu_to_be_32(xfrm->gga_ctrl1);
371 	rte_spinlock_lock(&priv->xform_sl);
372 	LIST_INSERT_HEAD(&priv->xform_list, xfrm, next);
373 	rte_spinlock_unlock(&priv->xform_sl);
374 	*private_xform = xfrm;
375 	return 0;
376 err:
377 	rte_free(xfrm);
378 	return -ENOTSUP;
379 }
380 
381 static void
382 mlx5_compress_dev_stop(struct rte_compressdev *dev)
383 {
384 	RTE_SET_USED(dev);
385 }
386 
387 static int
388 mlx5_compress_dev_start(struct rte_compressdev *dev)
389 {
390 	RTE_SET_USED(dev);
391 	return 0;
392 }
393 
394 static void
395 mlx5_compress_stats_get(struct rte_compressdev *dev,
396 		struct rte_compressdev_stats *stats)
397 {
398 	int qp_id;
399 
400 	for (qp_id = 0; qp_id < dev->data->nb_queue_pairs; qp_id++) {
401 		struct mlx5_compress_qp *qp = dev->data->queue_pairs[qp_id];
402 
403 		stats->enqueued_count += qp->stats.enqueued_count;
404 		stats->dequeued_count += qp->stats.dequeued_count;
405 		stats->enqueue_err_count += qp->stats.enqueue_err_count;
406 		stats->dequeue_err_count += qp->stats.dequeue_err_count;
407 	}
408 }
409 
410 static void
411 mlx5_compress_stats_reset(struct rte_compressdev *dev)
412 {
413 	int qp_id;
414 
415 	for (qp_id = 0; qp_id < dev->data->nb_queue_pairs; qp_id++) {
416 		struct mlx5_compress_qp *qp = dev->data->queue_pairs[qp_id];
417 
418 		memset(&qp->stats, 0, sizeof(qp->stats));
419 	}
420 }
421 
422 static struct rte_compressdev_ops mlx5_compress_ops = {
423 	.dev_configure		= mlx5_compress_dev_configure,
424 	.dev_start		= mlx5_compress_dev_start,
425 	.dev_stop		= mlx5_compress_dev_stop,
426 	.dev_close		= mlx5_compress_dev_close,
427 	.dev_infos_get		= mlx5_compress_dev_info_get,
428 	.stats_get		= mlx5_compress_stats_get,
429 	.stats_reset		= mlx5_compress_stats_reset,
430 	.queue_pair_setup	= mlx5_compress_qp_setup,
431 	.queue_pair_release	= mlx5_compress_qp_release,
432 	.private_xform_create	= mlx5_compress_xform_create,
433 	.private_xform_free	= mlx5_compress_xform_free,
434 	.stream_create		= NULL,
435 	.stream_free		= NULL,
436 };
437 
438 /**
439  * Query LKey from a packet buffer for QP. If not found, add the mempool.
440  *
441  * @param priv
442  *   Pointer to the priv object.
443  * @param addr
444  *   Search key.
445  * @param mr_ctrl
446  *   Pointer to per-queue MR control structure.
447  * @param ol_flags
448  *   Mbuf offload features.
449  *
450  * @return
451  *   Searched LKey on success, UINT32_MAX on no match.
452  */
453 static __rte_always_inline uint32_t
454 mlx5_compress_addr2mr(struct mlx5_compress_priv *priv, uintptr_t addr,
455 		      struct mlx5_mr_ctrl *mr_ctrl, uint64_t ol_flags)
456 {
457 	uint32_t lkey;
458 
459 	/* Check generation bit to see if there's any change on existing MRs. */
460 	if (unlikely(*mr_ctrl->dev_gen_ptr != mr_ctrl->cur_gen))
461 		mlx5_mr_flush_local_cache(mr_ctrl);
462 	/* Linear search on MR cache array. */
463 	lkey = mlx5_mr_lookup_lkey(mr_ctrl->cache, &mr_ctrl->mru,
464 				   MLX5_MR_CACHE_N, addr);
465 	if (likely(lkey != UINT32_MAX))
466 		return lkey;
467 	/* Take slower bottom-half on miss. */
468 	return mlx5_mr_addr2mr_bh(priv->pd, 0, &priv->mr_scache, mr_ctrl, addr,
469 				  !!(ol_flags & EXT_ATTACHED_MBUF));
470 }
471 
472 static __rte_always_inline uint32_t
473 mlx5_compress_dseg_set(struct mlx5_compress_qp *qp,
474 		       volatile struct mlx5_wqe_dseg *restrict dseg,
475 		       struct rte_mbuf *restrict mbuf,
476 		       uint32_t offset, uint32_t len)
477 {
478 	uintptr_t addr = rte_pktmbuf_mtod_offset(mbuf, uintptr_t, offset);
479 
480 	dseg->bcount = rte_cpu_to_be_32(len);
481 	dseg->lkey = mlx5_compress_addr2mr(qp->priv, addr, &qp->mr_ctrl,
482 					   mbuf->ol_flags);
483 	dseg->pbuf = rte_cpu_to_be_64(addr);
484 	return dseg->lkey;
485 }
486 
487 /*
488  * Provide safe 64bit store operation to mlx5 UAR region for both 32bit and
489  * 64bit architectures.
490  */
491 static __rte_always_inline void
492 mlx5_compress_uar_write(uint64_t val, struct mlx5_compress_priv *priv)
493 {
494 #ifdef RTE_ARCH_64
495 	*priv->uar_addr = val;
496 #else /* !RTE_ARCH_64 */
497 	rte_spinlock_lock(&priv->uar32_sl);
498 	*(volatile uint32_t *)priv->uar_addr = val;
499 	rte_io_wmb();
500 	*((volatile uint32_t *)priv->uar_addr + 1) = val >> 32;
501 	rte_spinlock_unlock(&priv->uar32_sl);
502 #endif
503 }
504 
505 static uint16_t
506 mlx5_compress_enqueue_burst(void *queue_pair, struct rte_comp_op **ops,
507 			    uint16_t nb_ops)
508 {
509 	struct mlx5_compress_qp *qp = queue_pair;
510 	volatile struct mlx5_gga_wqe *wqes = (volatile struct mlx5_gga_wqe *)
511 							      qp->sq.wqes, *wqe;
512 	struct mlx5_compress_xform *xform;
513 	struct rte_comp_op *op;
514 	uint16_t mask = qp->entries_n - 1;
515 	uint16_t remain = qp->entries_n - (qp->pi - qp->ci);
516 	uint16_t idx;
517 	bool invalid;
518 
519 	if (remain < nb_ops)
520 		nb_ops = remain;
521 	else
522 		remain = nb_ops;
523 	if (unlikely(remain == 0))
524 		return 0;
525 	do {
526 		idx = qp->pi & mask;
527 		wqe = &wqes[idx];
528 		rte_prefetch0(&wqes[(qp->pi + 1) & mask]);
529 		op = *ops++;
530 		xform = op->private_xform;
531 		/*
532 		 * Check operation arguments and error cases:
533 		 *   - Operation type must be state-less.
534 		 *   - Compress operation flush flag must be FULL or FINAL.
535 		 *   - Source and destination buffers must be mapped internally.
536 		 */
537 		invalid = op->op_type != RTE_COMP_OP_STATELESS ||
538 					    (xform->type == RTE_COMP_COMPRESS &&
539 					  op->flush_flag < RTE_COMP_FLUSH_FULL);
540 		if (unlikely(invalid ||
541 			     (mlx5_compress_dseg_set(qp, &wqe->gather,
542 						     op->m_src,
543 						     op->src.offset,
544 						     op->src.length) ==
545 								  UINT32_MAX) ||
546 			     (mlx5_compress_dseg_set(qp, &wqe->scatter,
547 						op->m_dst,
548 						op->dst.offset,
549 						rte_pktmbuf_pkt_len(op->m_dst) -
550 							      op->dst.offset) ==
551 								 UINT32_MAX))) {
552 			op->status = invalid ? RTE_COMP_OP_STATUS_INVALID_ARGS :
553 						       RTE_COMP_OP_STATUS_ERROR;
554 			nb_ops -= remain;
555 			if (unlikely(nb_ops == 0))
556 				return 0;
557 			break;
558 		}
559 		wqe->gga_ctrl1 = xform->gga_ctrl1;
560 		wqe->opcode = rte_cpu_to_be_32(xform->opcode + (qp->pi << 8));
561 		qp->ops[idx] = op;
562 		qp->pi++;
563 	} while (--remain);
564 	qp->stats.enqueued_count += nb_ops;
565 	rte_io_wmb();
566 	qp->sq.db_rec[MLX5_SND_DBR] = rte_cpu_to_be_32(qp->pi);
567 	rte_wmb();
568 	mlx5_compress_uar_write(*(volatile uint64_t *)wqe, qp->priv);
569 	rte_wmb();
570 	return nb_ops;
571 }
572 
573 static void
574 mlx5_compress_dump_err_objs(volatile uint32_t *cqe, volatile uint32_t *wqe,
575 			     volatile uint32_t *opaq)
576 {
577 	size_t i;
578 
579 	DRV_LOG(ERR, "Error cqe:");
580 	for (i = 0; i < sizeof(struct mlx5_err_cqe) >> 2; i += 4)
581 		DRV_LOG(ERR, "%08X %08X %08X %08X", cqe[i], cqe[i + 1],
582 			cqe[i + 2], cqe[i + 3]);
583 	DRV_LOG(ERR, "\nError wqe:");
584 	for (i = 0; i < sizeof(struct mlx5_gga_wqe) >> 2; i += 4)
585 		DRV_LOG(ERR, "%08X %08X %08X %08X", wqe[i], wqe[i + 1],
586 			wqe[i + 2], wqe[i + 3]);
587 	DRV_LOG(ERR, "\nError opaq:");
588 	for (i = 0; i < sizeof(struct mlx5_gga_compress_opaque) >> 2; i += 4)
589 		DRV_LOG(ERR, "%08X %08X %08X %08X", opaq[i], opaq[i + 1],
590 			opaq[i + 2], opaq[i + 3]);
591 }
592 
593 static void
594 mlx5_compress_cqe_err_handle(struct mlx5_compress_qp *qp,
595 			     struct rte_comp_op *op)
596 {
597 	const uint32_t idx = qp->ci & (qp->entries_n - 1);
598 	volatile struct mlx5_err_cqe *cqe = (volatile struct mlx5_err_cqe *)
599 							      &qp->cq.cqes[idx];
600 	volatile struct mlx5_gga_wqe *wqes = (volatile struct mlx5_gga_wqe *)
601 								    qp->sq.wqes;
602 	volatile struct mlx5_gga_compress_opaque *opaq = qp->opaque_mr.addr;
603 
604 	op->status = RTE_COMP_OP_STATUS_ERROR;
605 	op->consumed = 0;
606 	op->produced = 0;
607 	op->output_chksum = 0;
608 	op->debug_status = rte_be_to_cpu_32(opaq[idx].syndrom) |
609 			      ((uint64_t)rte_be_to_cpu_32(cqe->syndrome) << 32);
610 	mlx5_compress_dump_err_objs((volatile uint32_t *)cqe,
611 				 (volatile uint32_t *)&wqes[idx],
612 				 (volatile uint32_t *)&opaq[idx]);
613 	qp->stats.dequeue_err_count++;
614 }
615 
616 static uint16_t
617 mlx5_compress_dequeue_burst(void *queue_pair, struct rte_comp_op **ops,
618 			    uint16_t nb_ops)
619 {
620 	struct mlx5_compress_qp *qp = queue_pair;
621 	volatile struct mlx5_compress_xform *restrict xform;
622 	volatile struct mlx5_cqe *restrict cqe;
623 	volatile struct mlx5_gga_compress_opaque *opaq = qp->opaque_mr.addr;
624 	struct rte_comp_op *restrict op;
625 	const unsigned int cq_size = qp->entries_n;
626 	const unsigned int mask = cq_size - 1;
627 	uint32_t idx;
628 	uint32_t next_idx = qp->ci & mask;
629 	const uint16_t max = RTE_MIN((uint16_t)(qp->pi - qp->ci), nb_ops);
630 	uint16_t i = 0;
631 	int ret;
632 
633 	if (unlikely(max == 0))
634 		return 0;
635 	do {
636 		idx = next_idx;
637 		next_idx = (qp->ci + 1) & mask;
638 		rte_prefetch0(&qp->cq.cqes[next_idx]);
639 		rte_prefetch0(qp->ops[next_idx]);
640 		op = qp->ops[idx];
641 		cqe = &qp->cq.cqes[idx];
642 		ret = check_cqe(cqe, cq_size, qp->ci);
643 		/*
644 		 * Be sure owner read is done before any other cookie field or
645 		 * opaque field.
646 		 */
647 		rte_io_rmb();
648 		if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
649 			if (likely(ret == MLX5_CQE_STATUS_HW_OWN))
650 				break;
651 			mlx5_compress_cqe_err_handle(qp, op);
652 		} else {
653 			xform = op->private_xform;
654 			op->status = RTE_COMP_OP_STATUS_SUCCESS;
655 			op->consumed = op->src.length;
656 			op->produced = rte_be_to_cpu_32(cqe->byte_cnt);
657 			MLX5_ASSERT(cqe->byte_cnt ==
658 				    opaq[idx].scattered_length);
659 			switch (xform->csum_type) {
660 			case RTE_COMP_CHECKSUM_CRC32:
661 				op->output_chksum = (uint64_t)rte_be_to_cpu_32
662 						    (opaq[idx].crc32);
663 				break;
664 			case RTE_COMP_CHECKSUM_ADLER32:
665 				op->output_chksum = (uint64_t)rte_be_to_cpu_32
666 					    (opaq[idx].adler32) << 32;
667 				break;
668 			case RTE_COMP_CHECKSUM_CRC32_ADLER32:
669 				op->output_chksum = (uint64_t)rte_be_to_cpu_32
670 							     (opaq[idx].crc32) |
671 						     ((uint64_t)rte_be_to_cpu_32
672 						     (opaq[idx].adler32) << 32);
673 				break;
674 			default:
675 				break;
676 			}
677 		}
678 		ops[i++] = op;
679 		qp->ci++;
680 	} while (i < max);
681 	if (likely(i != 0)) {
682 		rte_io_wmb();
683 		qp->cq.db_rec[0] = rte_cpu_to_be_32(qp->ci);
684 		qp->stats.dequeued_count += i;
685 	}
686 	return i;
687 }
688 
689 static void
690 mlx5_compress_hw_global_release(struct mlx5_compress_priv *priv)
691 {
692 	if (priv->pd != NULL) {
693 		claim_zero(mlx5_glue->dealloc_pd(priv->pd));
694 		priv->pd = NULL;
695 	}
696 	if (priv->uar != NULL) {
697 		mlx5_glue->devx_free_uar(priv->uar);
698 		priv->uar = NULL;
699 	}
700 }
701 
702 static int
703 mlx5_compress_pd_create(struct mlx5_compress_priv *priv)
704 {
705 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
706 	struct mlx5dv_obj obj;
707 	struct mlx5dv_pd pd_info;
708 	int ret;
709 
710 	priv->pd = mlx5_glue->alloc_pd(priv->ctx);
711 	if (priv->pd == NULL) {
712 		DRV_LOG(ERR, "Failed to allocate PD.");
713 		return errno ? -errno : -ENOMEM;
714 	}
715 	obj.pd.in = priv->pd;
716 	obj.pd.out = &pd_info;
717 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
718 	if (ret != 0) {
719 		DRV_LOG(ERR, "Fail to get PD object info.");
720 		mlx5_glue->dealloc_pd(priv->pd);
721 		priv->pd = NULL;
722 		return -errno;
723 	}
724 	priv->pdn = pd_info.pdn;
725 	return 0;
726 #else
727 	(void)priv;
728 	DRV_LOG(ERR, "Cannot get pdn - no DV support.");
729 	return -ENOTSUP;
730 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
731 }
732 
733 static int
734 mlx5_compress_hw_global_prepare(struct mlx5_compress_priv *priv)
735 {
736 	if (mlx5_compress_pd_create(priv) != 0)
737 		return -1;
738 	priv->uar = mlx5_devx_alloc_uar(priv->ctx, -1);
739 	if (priv->uar == NULL || mlx5_os_get_devx_uar_reg_addr(priv->uar) ==
740 	    NULL) {
741 		rte_errno = errno;
742 		claim_zero(mlx5_glue->dealloc_pd(priv->pd));
743 		DRV_LOG(ERR, "Failed to allocate UAR.");
744 		return -1;
745 	}
746 	priv->uar_addr = mlx5_os_get_devx_uar_reg_addr(priv->uar);
747 	MLX5_ASSERT(priv->uar_addr);
748 #ifndef RTE_ARCH_64
749 	rte_spinlock_init(&priv->uar32_sl);
750 #endif /* RTE_ARCH_64 */
751 	return 0;
752 }
753 
754 /**
755  * Callback for memory event.
756  *
757  * @param event_type
758  *   Memory event type.
759  * @param addr
760  *   Address of memory.
761  * @param len
762  *   Size of memory.
763  */
764 static void
765 mlx5_compress_mr_mem_event_cb(enum rte_mem_event event_type, const void *addr,
766 			      size_t len, void *arg __rte_unused)
767 {
768 	struct mlx5_compress_priv *priv;
769 
770 	/* Must be called from the primary process. */
771 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
772 	switch (event_type) {
773 	case RTE_MEM_EVENT_FREE:
774 		pthread_mutex_lock(&priv_list_lock);
775 		/* Iterate all the existing mlx5 devices. */
776 		TAILQ_FOREACH(priv, &mlx5_compress_priv_list, next)
777 			mlx5_free_mr_by_addr(&priv->mr_scache,
778 					     priv->ctx->device->name,
779 					     addr, len);
780 		pthread_mutex_unlock(&priv_list_lock);
781 		break;
782 	case RTE_MEM_EVENT_ALLOC:
783 	default:
784 		break;
785 	}
786 }
787 
788 static int
789 mlx5_compress_dev_probe(struct rte_device *dev)
790 {
791 	struct ibv_device *ibv;
792 	struct rte_compressdev *cdev;
793 	struct ibv_context *ctx;
794 	struct mlx5_compress_priv *priv;
795 	struct mlx5_hca_attr att = { 0 };
796 	struct rte_compressdev_pmd_init_params init_params = {
797 		.name = "",
798 		.socket_id = dev->numa_node,
799 	};
800 
801 	if (rte_eal_process_type() != RTE_PROC_PRIMARY) {
802 		DRV_LOG(ERR, "Non-primary process type is not supported.");
803 		rte_errno = ENOTSUP;
804 		return -rte_errno;
805 	}
806 	ibv = mlx5_os_get_ibv_dev(dev);
807 	if (ibv == NULL)
808 		return -rte_errno;
809 	ctx = mlx5_glue->dv_open_device(ibv);
810 	if (ctx == NULL) {
811 		DRV_LOG(ERR, "Failed to open IB device \"%s\".", ibv->name);
812 		rte_errno = ENODEV;
813 		return -rte_errno;
814 	}
815 	if (mlx5_devx_cmd_query_hca_attr(ctx, &att) != 0 ||
816 	    att.mmo_compress_en == 0 || att.mmo_decompress_en == 0 ||
817 	    att.mmo_dma_en == 0) {
818 		DRV_LOG(ERR, "Not enough capabilities to support compress "
819 			"operations, maybe old FW/OFED version?");
820 		claim_zero(mlx5_glue->close_device(ctx));
821 		rte_errno = ENOTSUP;
822 		return -ENOTSUP;
823 	}
824 	cdev = rte_compressdev_pmd_create(ibv->name, dev,
825 					  sizeof(*priv), &init_params);
826 	if (cdev == NULL) {
827 		DRV_LOG(ERR, "Failed to create device \"%s\".", ibv->name);
828 		claim_zero(mlx5_glue->close_device(ctx));
829 		return -ENODEV;
830 	}
831 	DRV_LOG(INFO,
832 		"Compress device %s was created successfully.", ibv->name);
833 	cdev->dev_ops = &mlx5_compress_ops;
834 	cdev->dequeue_burst = mlx5_compress_dequeue_burst;
835 	cdev->enqueue_burst = mlx5_compress_enqueue_burst;
836 	cdev->feature_flags = RTE_COMPDEV_FF_HW_ACCELERATED;
837 	priv = cdev->data->dev_private;
838 	priv->ctx = ctx;
839 	priv->cdev = cdev;
840 	priv->min_block_size = att.compress_min_block_size;
841 	priv->sq_ts_format = att.sq_ts_format;
842 	if (mlx5_compress_hw_global_prepare(priv) != 0) {
843 		rte_compressdev_pmd_destroy(priv->cdev);
844 		claim_zero(mlx5_glue->close_device(priv->ctx));
845 		return -1;
846 	}
847 	if (mlx5_mr_btree_init(&priv->mr_scache.cache,
848 			     MLX5_MR_BTREE_CACHE_N * 2, rte_socket_id()) != 0) {
849 		DRV_LOG(ERR, "Failed to allocate shared cache MR memory.");
850 		mlx5_compress_hw_global_release(priv);
851 		rte_compressdev_pmd_destroy(priv->cdev);
852 		claim_zero(mlx5_glue->close_device(priv->ctx));
853 		rte_errno = ENOMEM;
854 		return -rte_errno;
855 	}
856 	priv->mr_scache.reg_mr_cb = mlx5_common_verbs_reg_mr;
857 	priv->mr_scache.dereg_mr_cb = mlx5_common_verbs_dereg_mr;
858 	/* Register callback function for global shared MR cache management. */
859 	if (TAILQ_EMPTY(&mlx5_compress_priv_list))
860 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
861 						mlx5_compress_mr_mem_event_cb,
862 						NULL);
863 	pthread_mutex_lock(&priv_list_lock);
864 	TAILQ_INSERT_TAIL(&mlx5_compress_priv_list, priv, next);
865 	pthread_mutex_unlock(&priv_list_lock);
866 	return 0;
867 }
868 
869 static int
870 mlx5_compress_dev_remove(struct rte_device *dev)
871 {
872 	struct mlx5_compress_priv *priv = NULL;
873 
874 	pthread_mutex_lock(&priv_list_lock);
875 	TAILQ_FOREACH(priv, &mlx5_compress_priv_list, next)
876 		if (priv->cdev->device == dev)
877 			break;
878 	if (priv)
879 		TAILQ_REMOVE(&mlx5_compress_priv_list, priv, next);
880 	pthread_mutex_unlock(&priv_list_lock);
881 	if (priv) {
882 		if (TAILQ_EMPTY(&mlx5_compress_priv_list))
883 			rte_mem_event_callback_unregister("MLX5_MEM_EVENT_CB",
884 							  NULL);
885 		mlx5_mr_release_cache(&priv->mr_scache);
886 		mlx5_compress_hw_global_release(priv);
887 		rte_compressdev_pmd_destroy(priv->cdev);
888 		claim_zero(mlx5_glue->close_device(priv->ctx));
889 	}
890 	return 0;
891 }
892 
893 static const struct rte_pci_id mlx5_compress_pci_id_map[] = {
894 	{
895 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
896 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
897 	},
898 	{
899 		.vendor_id = 0
900 	}
901 };
902 
903 static struct mlx5_class_driver mlx5_compress_driver = {
904 	.drv_class = MLX5_CLASS_COMPRESS,
905 	.name = RTE_STR(MLX5_COMPRESS_DRIVER_NAME),
906 	.id_table = mlx5_compress_pci_id_map,
907 	.probe = mlx5_compress_dev_probe,
908 	.remove = mlx5_compress_dev_remove,
909 };
910 
911 RTE_INIT(rte_mlx5_compress_init)
912 {
913 	mlx5_common_init();
914 	if (mlx5_glue != NULL)
915 		mlx5_class_driver_register(&mlx5_compress_driver);
916 }
917 
918 RTE_LOG_REGISTER_DEFAULT(mlx5_compress_logtype, NOTICE)
919 RTE_PMD_EXPORT_NAME(MLX5_COMPRESS_DRIVER_NAME, __COUNTER__);
920 RTE_PMD_REGISTER_PCI_TABLE(MLX5_COMPRESS_DRIVER_NAME, mlx5_compress_pci_id_map);
921 RTE_PMD_REGISTER_KMOD_DEP(MLX5_COMPRESS_DRIVER_NAME, "* ib_uverbs & mlx5_core & mlx5_ib");
922