xref: /dpdk/drivers/compress/mlx5/mlx5_compress.c (revision c87bc83a33d2cd0f159b175336553d5b3293aef0)
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 	opaq_buf = rte_calloc(__func__, (size_t)1 << log_ops_n,
211 			      sizeof(struct mlx5_gga_compress_opaque),
212 			      sizeof(struct mlx5_gga_compress_opaque));
213 	if (opaq_buf == NULL) {
214 		DRV_LOG(ERR, "Failed to allocate opaque memory.");
215 		rte_errno = ENOMEM;
216 		goto err;
217 	}
218 	if (mlx5_mr_btree_init(&qp->mr_ctrl.cache_bh, MLX5_MR_BTREE_CACHE_N,
219 			       priv->dev_config.socket_id)) {
220 		DRV_LOG(ERR, "Cannot allocate MR Btree for qp %u.",
221 			(uint32_t)qp_id);
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 			if (xform->compress.level == RTE_COMP_LEVEL_PMD_DEFAULT)
321 				size = MLX5_GGA_COMP_LOG_BLOCK_SIZE_MAX;
322 			else
323 				size = priv->min_block_size - 1 +
324 							  xform->compress.level;
325 			xfrm->gga_ctrl1 += RTE_MIN(size,
326 					    MLX5_GGA_COMP_LOG_BLOCK_SIZE_MAX) <<
327 						 WQE_GGA_COMP_BLOCK_SIZE_OFFSET;
328 			xfrm->opcode += MLX5_OPC_MOD_MMO_COMP <<
329 							WQE_CSEG_OPC_MOD_OFFSET;
330 			size = xform->compress.deflate.huffman ==
331 						      RTE_COMP_HUFFMAN_DYNAMIC ?
332 					    MLX5_GGA_COMP_LOG_DYNAMIC_SIZE_MAX :
333 					     MLX5_GGA_COMP_LOG_DYNAMIC_SIZE_MIN;
334 			xfrm->gga_ctrl1 += size <<
335 					       WQE_GGA_COMP_DYNAMIC_SIZE_OFFSET;
336 			break;
337 		default:
338 			goto err;
339 		}
340 		xfrm->csum_type = xform->compress.chksum;
341 		break;
342 	case RTE_COMP_DECOMPRESS:
343 		switch (xform->decompress.algo) {
344 		case RTE_COMP_ALGO_NULL:
345 			xfrm->opcode += MLX5_OPC_MOD_MMO_DMA <<
346 							WQE_CSEG_OPC_MOD_OFFSET;
347 			break;
348 		case RTE_COMP_ALGO_DEFLATE:
349 			xfrm->opcode += MLX5_OPC_MOD_MMO_DECOMP <<
350 							WQE_CSEG_OPC_MOD_OFFSET;
351 			break;
352 		default:
353 			goto err;
354 		}
355 		xfrm->csum_type = xform->decompress.chksum;
356 		break;
357 	default:
358 		DRV_LOG(ERR, "Algorithm %u is not supported.", xform->type);
359 		goto err;
360 	}
361 	DRV_LOG(DEBUG, "New xform: gga ctrl1 = 0x%08X opcode = 0x%08X csum "
362 		"type = %d.", xfrm->gga_ctrl1, xfrm->opcode, xfrm->csum_type);
363 	xfrm->gga_ctrl1 = rte_cpu_to_be_32(xfrm->gga_ctrl1);
364 	rte_spinlock_lock(&priv->xform_sl);
365 	LIST_INSERT_HEAD(&priv->xform_list, xfrm, next);
366 	rte_spinlock_unlock(&priv->xform_sl);
367 	*private_xform = xfrm;
368 	return 0;
369 err:
370 	rte_free(xfrm);
371 	return -ENOTSUP;
372 }
373 
374 static void
375 mlx5_compress_dev_stop(struct rte_compressdev *dev)
376 {
377 	RTE_SET_USED(dev);
378 }
379 
380 static int
381 mlx5_compress_dev_start(struct rte_compressdev *dev)
382 {
383 	RTE_SET_USED(dev);
384 	return 0;
385 }
386 
387 static void
388 mlx5_compress_stats_get(struct rte_compressdev *dev,
389 		struct rte_compressdev_stats *stats)
390 {
391 	int qp_id;
392 
393 	for (qp_id = 0; qp_id < dev->data->nb_queue_pairs; qp_id++) {
394 		struct mlx5_compress_qp *qp = dev->data->queue_pairs[qp_id];
395 
396 		stats->enqueued_count += qp->stats.enqueued_count;
397 		stats->dequeued_count += qp->stats.dequeued_count;
398 		stats->enqueue_err_count += qp->stats.enqueue_err_count;
399 		stats->dequeue_err_count += qp->stats.dequeue_err_count;
400 	}
401 }
402 
403 static void
404 mlx5_compress_stats_reset(struct rte_compressdev *dev)
405 {
406 	int qp_id;
407 
408 	for (qp_id = 0; qp_id < dev->data->nb_queue_pairs; qp_id++) {
409 		struct mlx5_compress_qp *qp = dev->data->queue_pairs[qp_id];
410 
411 		memset(&qp->stats, 0, sizeof(qp->stats));
412 	}
413 }
414 
415 static struct rte_compressdev_ops mlx5_compress_ops = {
416 	.dev_configure		= mlx5_compress_dev_configure,
417 	.dev_start		= mlx5_compress_dev_start,
418 	.dev_stop		= mlx5_compress_dev_stop,
419 	.dev_close		= mlx5_compress_dev_close,
420 	.dev_infos_get		= mlx5_compress_dev_info_get,
421 	.stats_get		= mlx5_compress_stats_get,
422 	.stats_reset		= mlx5_compress_stats_reset,
423 	.queue_pair_setup	= mlx5_compress_qp_setup,
424 	.queue_pair_release	= mlx5_compress_qp_release,
425 	.private_xform_create	= mlx5_compress_xform_create,
426 	.private_xform_free	= mlx5_compress_xform_free,
427 	.stream_create		= NULL,
428 	.stream_free		= NULL,
429 };
430 
431 /**
432  * Query LKey from a packet buffer for QP. If not found, add the mempool.
433  *
434  * @param priv
435  *   Pointer to the priv object.
436  * @param addr
437  *   Search key.
438  * @param mr_ctrl
439  *   Pointer to per-queue MR control structure.
440  * @param ol_flags
441  *   Mbuf offload features.
442  *
443  * @return
444  *   Searched LKey on success, UINT32_MAX on no match.
445  */
446 static __rte_always_inline uint32_t
447 mlx5_compress_addr2mr(struct mlx5_compress_priv *priv, uintptr_t addr,
448 		      struct mlx5_mr_ctrl *mr_ctrl, uint64_t ol_flags)
449 {
450 	uint32_t lkey;
451 
452 	/* Check generation bit to see if there's any change on existing MRs. */
453 	if (unlikely(*mr_ctrl->dev_gen_ptr != mr_ctrl->cur_gen))
454 		mlx5_mr_flush_local_cache(mr_ctrl);
455 	/* Linear search on MR cache array. */
456 	lkey = mlx5_mr_lookup_lkey(mr_ctrl->cache, &mr_ctrl->mru,
457 				   MLX5_MR_CACHE_N, addr);
458 	if (likely(lkey != UINT32_MAX))
459 		return lkey;
460 	/* Take slower bottom-half on miss. */
461 	return mlx5_mr_addr2mr_bh(priv->pd, 0, &priv->mr_scache, mr_ctrl, addr,
462 				  !!(ol_flags & EXT_ATTACHED_MBUF));
463 }
464 
465 static __rte_always_inline uint32_t
466 mlx5_compress_dseg_set(struct mlx5_compress_qp *qp,
467 		       volatile struct mlx5_wqe_dseg *restrict dseg,
468 		       struct rte_mbuf *restrict mbuf,
469 		       uint32_t offset, uint32_t len)
470 {
471 	uintptr_t addr = rte_pktmbuf_mtod_offset(mbuf, uintptr_t, offset);
472 
473 	dseg->bcount = rte_cpu_to_be_32(len);
474 	dseg->lkey = mlx5_compress_addr2mr(qp->priv, addr, &qp->mr_ctrl,
475 					   mbuf->ol_flags);
476 	dseg->pbuf = rte_cpu_to_be_64(addr);
477 	return dseg->lkey;
478 }
479 
480 /*
481  * Provide safe 64bit store operation to mlx5 UAR region for both 32bit and
482  * 64bit architectures.
483  */
484 static __rte_always_inline void
485 mlx5_compress_uar_write(uint64_t val, struct mlx5_compress_priv *priv)
486 {
487 #ifdef RTE_ARCH_64
488 	*priv->uar_addr = val;
489 #else /* !RTE_ARCH_64 */
490 	rte_spinlock_lock(&priv->uar32_sl);
491 	*(volatile uint32_t *)priv->uar_addr = val;
492 	rte_io_wmb();
493 	*((volatile uint32_t *)priv->uar_addr + 1) = val >> 32;
494 	rte_spinlock_unlock(&priv->uar32_sl);
495 #endif
496 }
497 
498 static uint16_t
499 mlx5_compress_enqueue_burst(void *queue_pair, struct rte_comp_op **ops,
500 			    uint16_t nb_ops)
501 {
502 	struct mlx5_compress_qp *qp = queue_pair;
503 	volatile struct mlx5_gga_wqe *wqes = (volatile struct mlx5_gga_wqe *)
504 							      qp->sq.wqes, *wqe;
505 	struct mlx5_compress_xform *xform;
506 	struct rte_comp_op *op;
507 	uint16_t mask = qp->entries_n - 1;
508 	uint16_t remain = qp->entries_n - (qp->pi - qp->ci);
509 	uint16_t idx;
510 	bool invalid;
511 
512 	if (remain < nb_ops)
513 		nb_ops = remain;
514 	else
515 		remain = nb_ops;
516 	if (unlikely(remain == 0))
517 		return 0;
518 	do {
519 		idx = qp->pi & mask;
520 		wqe = &wqes[idx];
521 		rte_prefetch0(&wqes[(qp->pi + 1) & mask]);
522 		op = *ops++;
523 		xform = op->private_xform;
524 		/*
525 		 * Check operation arguments and error cases:
526 		 *   - Operation type must be state-less.
527 		 *   - Compress operation flush flag must be FULL or FINAL.
528 		 *   - Source and destination buffers must be mapped internally.
529 		 */
530 		invalid = op->op_type != RTE_COMP_OP_STATELESS ||
531 					    (xform->type == RTE_COMP_COMPRESS &&
532 					  op->flush_flag < RTE_COMP_FLUSH_FULL);
533 		if (unlikely(invalid ||
534 			     (mlx5_compress_dseg_set(qp, &wqe->gather,
535 						     op->m_src,
536 						     op->src.offset,
537 						     op->src.length) ==
538 								  UINT32_MAX) ||
539 			     (mlx5_compress_dseg_set(qp, &wqe->scatter,
540 						op->m_dst,
541 						op->dst.offset,
542 						rte_pktmbuf_pkt_len(op->m_dst) -
543 							      op->dst.offset) ==
544 								 UINT32_MAX))) {
545 			op->status = invalid ? RTE_COMP_OP_STATUS_INVALID_ARGS :
546 						       RTE_COMP_OP_STATUS_ERROR;
547 			nb_ops -= remain;
548 			if (unlikely(nb_ops == 0))
549 				return 0;
550 			break;
551 		}
552 		wqe->gga_ctrl1 = xform->gga_ctrl1;
553 		wqe->opcode = rte_cpu_to_be_32(xform->opcode + (qp->pi << 8));
554 		qp->ops[idx] = op;
555 		qp->pi++;
556 	} while (--remain);
557 	qp->stats.enqueued_count += nb_ops;
558 	rte_io_wmb();
559 	qp->sq.db_rec[MLX5_SND_DBR] = rte_cpu_to_be_32(qp->pi);
560 	rte_wmb();
561 	mlx5_compress_uar_write(*(volatile uint64_t *)wqe, qp->priv);
562 	rte_wmb();
563 	return nb_ops;
564 }
565 
566 static void
567 mlx5_compress_dump_err_objs(volatile uint32_t *cqe, volatile uint32_t *wqe,
568 			     volatile uint32_t *opaq)
569 {
570 	size_t i;
571 
572 	DRV_LOG(ERR, "Error cqe:");
573 	for (i = 0; i < sizeof(struct mlx5_err_cqe) >> 2; i += 4)
574 		DRV_LOG(ERR, "%08X %08X %08X %08X", cqe[i], cqe[i + 1],
575 			cqe[i + 2], cqe[i + 3]);
576 	DRV_LOG(ERR, "\nError wqe:");
577 	for (i = 0; i < sizeof(struct mlx5_gga_wqe) >> 2; i += 4)
578 		DRV_LOG(ERR, "%08X %08X %08X %08X", wqe[i], wqe[i + 1],
579 			wqe[i + 2], wqe[i + 3]);
580 	DRV_LOG(ERR, "\nError opaq:");
581 	for (i = 0; i < sizeof(struct mlx5_gga_compress_opaque) >> 2; i += 4)
582 		DRV_LOG(ERR, "%08X %08X %08X %08X", opaq[i], opaq[i + 1],
583 			opaq[i + 2], opaq[i + 3]);
584 }
585 
586 static void
587 mlx5_compress_cqe_err_handle(struct mlx5_compress_qp *qp,
588 			     struct rte_comp_op *op)
589 {
590 	const uint32_t idx = qp->ci & (qp->entries_n - 1);
591 	volatile struct mlx5_err_cqe *cqe = (volatile struct mlx5_err_cqe *)
592 							      &qp->cq.cqes[idx];
593 	volatile struct mlx5_gga_wqe *wqes = (volatile struct mlx5_gga_wqe *)
594 								    qp->sq.wqes;
595 	volatile struct mlx5_gga_compress_opaque *opaq = qp->opaque_mr.addr;
596 
597 	op->status = RTE_COMP_OP_STATUS_ERROR;
598 	op->consumed = 0;
599 	op->produced = 0;
600 	op->output_chksum = 0;
601 	op->debug_status = rte_be_to_cpu_32(opaq[idx].syndrom) |
602 			      ((uint64_t)rte_be_to_cpu_32(cqe->syndrome) << 32);
603 	mlx5_compress_dump_err_objs((volatile uint32_t *)cqe,
604 				 (volatile uint32_t *)&wqes[idx],
605 				 (volatile uint32_t *)&opaq[idx]);
606 	qp->stats.dequeue_err_count++;
607 }
608 
609 static uint16_t
610 mlx5_compress_dequeue_burst(void *queue_pair, struct rte_comp_op **ops,
611 			    uint16_t nb_ops)
612 {
613 	struct mlx5_compress_qp *qp = queue_pair;
614 	volatile struct mlx5_compress_xform *restrict xform;
615 	volatile struct mlx5_cqe *restrict cqe;
616 	volatile struct mlx5_gga_compress_opaque *opaq = qp->opaque_mr.addr;
617 	struct rte_comp_op *restrict op;
618 	const unsigned int cq_size = qp->entries_n;
619 	const unsigned int mask = cq_size - 1;
620 	uint32_t idx;
621 	uint32_t next_idx = qp->ci & mask;
622 	const uint16_t max = RTE_MIN((uint16_t)(qp->pi - qp->ci), nb_ops);
623 	uint16_t i = 0;
624 	int ret;
625 
626 	if (unlikely(max == 0))
627 		return 0;
628 	do {
629 		idx = next_idx;
630 		next_idx = (qp->ci + 1) & mask;
631 		rte_prefetch0(&qp->cq.cqes[next_idx]);
632 		rte_prefetch0(qp->ops[next_idx]);
633 		op = qp->ops[idx];
634 		cqe = &qp->cq.cqes[idx];
635 		ret = check_cqe(cqe, cq_size, qp->ci);
636 		/*
637 		 * Be sure owner read is done before any other cookie field or
638 		 * opaque field.
639 		 */
640 		rte_io_rmb();
641 		if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
642 			if (likely(ret == MLX5_CQE_STATUS_HW_OWN))
643 				break;
644 			mlx5_compress_cqe_err_handle(qp, op);
645 		} else {
646 			xform = op->private_xform;
647 			op->status = RTE_COMP_OP_STATUS_SUCCESS;
648 			op->consumed = op->src.length;
649 			op->produced = rte_be_to_cpu_32(cqe->byte_cnt);
650 			MLX5_ASSERT(cqe->byte_cnt ==
651 				    opaq[idx].scattered_length);
652 			switch (xform->csum_type) {
653 			case RTE_COMP_CHECKSUM_CRC32:
654 				op->output_chksum = (uint64_t)rte_be_to_cpu_32
655 						    (opaq[idx].crc32);
656 				break;
657 			case RTE_COMP_CHECKSUM_ADLER32:
658 				op->output_chksum = (uint64_t)rte_be_to_cpu_32
659 					    (opaq[idx].adler32) << 32;
660 				break;
661 			case RTE_COMP_CHECKSUM_CRC32_ADLER32:
662 				op->output_chksum = (uint64_t)rte_be_to_cpu_32
663 							     (opaq[idx].crc32) |
664 						     ((uint64_t)rte_be_to_cpu_32
665 						     (opaq[idx].adler32) << 32);
666 				break;
667 			default:
668 				break;
669 			}
670 		}
671 		ops[i++] = op;
672 		qp->ci++;
673 	} while (i < max);
674 	if (likely(i != 0)) {
675 		rte_io_wmb();
676 		qp->cq.db_rec[0] = rte_cpu_to_be_32(qp->ci);
677 		qp->stats.dequeued_count += i;
678 	}
679 	return i;
680 }
681 
682 static void
683 mlx5_compress_hw_global_release(struct mlx5_compress_priv *priv)
684 {
685 	if (priv->pd != NULL) {
686 		claim_zero(mlx5_glue->dealloc_pd(priv->pd));
687 		priv->pd = NULL;
688 	}
689 	if (priv->uar != NULL) {
690 		mlx5_glue->devx_free_uar(priv->uar);
691 		priv->uar = NULL;
692 	}
693 }
694 
695 static int
696 mlx5_compress_pd_create(struct mlx5_compress_priv *priv)
697 {
698 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
699 	struct mlx5dv_obj obj;
700 	struct mlx5dv_pd pd_info;
701 	int ret;
702 
703 	priv->pd = mlx5_glue->alloc_pd(priv->ctx);
704 	if (priv->pd == NULL) {
705 		DRV_LOG(ERR, "Failed to allocate PD.");
706 		return errno ? -errno : -ENOMEM;
707 	}
708 	obj.pd.in = priv->pd;
709 	obj.pd.out = &pd_info;
710 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
711 	if (ret != 0) {
712 		DRV_LOG(ERR, "Fail to get PD object info.");
713 		mlx5_glue->dealloc_pd(priv->pd);
714 		priv->pd = NULL;
715 		return -errno;
716 	}
717 	priv->pdn = pd_info.pdn;
718 	return 0;
719 #else
720 	(void)priv;
721 	DRV_LOG(ERR, "Cannot get pdn - no DV support.");
722 	return -ENOTSUP;
723 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
724 }
725 
726 static int
727 mlx5_compress_hw_global_prepare(struct mlx5_compress_priv *priv)
728 {
729 	if (mlx5_compress_pd_create(priv) != 0)
730 		return -1;
731 	priv->uar = mlx5_devx_alloc_uar(priv->ctx, -1);
732 	if (priv->uar == NULL || mlx5_os_get_devx_uar_reg_addr(priv->uar) ==
733 	    NULL) {
734 		rte_errno = errno;
735 		claim_zero(mlx5_glue->dealloc_pd(priv->pd));
736 		DRV_LOG(ERR, "Failed to allocate UAR.");
737 		return -1;
738 	}
739 	priv->uar_addr = mlx5_os_get_devx_uar_reg_addr(priv->uar);
740 	MLX5_ASSERT(priv->uar_addr);
741 #ifndef RTE_ARCH_64
742 	rte_spinlock_init(&priv->uar32_sl);
743 #endif /* RTE_ARCH_64 */
744 	return 0;
745 }
746 
747 /**
748  * Callback for memory event.
749  *
750  * @param event_type
751  *   Memory event type.
752  * @param addr
753  *   Address of memory.
754  * @param len
755  *   Size of memory.
756  */
757 static void
758 mlx5_compress_mr_mem_event_cb(enum rte_mem_event event_type, const void *addr,
759 			      size_t len, void *arg __rte_unused)
760 {
761 	struct mlx5_compress_priv *priv;
762 
763 	/* Must be called from the primary process. */
764 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
765 	switch (event_type) {
766 	case RTE_MEM_EVENT_FREE:
767 		pthread_mutex_lock(&priv_list_lock);
768 		/* Iterate all the existing mlx5 devices. */
769 		TAILQ_FOREACH(priv, &mlx5_compress_priv_list, next)
770 			mlx5_free_mr_by_addr(&priv->mr_scache,
771 					     priv->ctx->device->name,
772 					     addr, len);
773 		pthread_mutex_unlock(&priv_list_lock);
774 		break;
775 	case RTE_MEM_EVENT_ALLOC:
776 	default:
777 		break;
778 	}
779 }
780 
781 static int
782 mlx5_compress_dev_probe(struct rte_device *dev)
783 {
784 	struct ibv_device *ibv;
785 	struct rte_compressdev *cdev;
786 	struct ibv_context *ctx;
787 	struct mlx5_compress_priv *priv;
788 	struct mlx5_hca_attr att = { 0 };
789 	struct rte_compressdev_pmd_init_params init_params = {
790 		.name = "",
791 		.socket_id = dev->numa_node,
792 	};
793 
794 	if (rte_eal_process_type() != RTE_PROC_PRIMARY) {
795 		DRV_LOG(ERR, "Non-primary process type is not supported.");
796 		rte_errno = ENOTSUP;
797 		return -rte_errno;
798 	}
799 	ibv = mlx5_os_get_ibv_dev(dev);
800 	if (ibv == NULL)
801 		return -rte_errno;
802 	ctx = mlx5_glue->dv_open_device(ibv);
803 	if (ctx == NULL) {
804 		DRV_LOG(ERR, "Failed to open IB device \"%s\".", ibv->name);
805 		rte_errno = ENODEV;
806 		return -rte_errno;
807 	}
808 	if (mlx5_devx_cmd_query_hca_attr(ctx, &att) != 0 ||
809 	    att.mmo_compress_en == 0 || att.mmo_decompress_en == 0 ||
810 	    att.mmo_dma_en == 0) {
811 		DRV_LOG(ERR, "Not enough capabilities to support compress "
812 			"operations, maybe old FW/OFED version?");
813 		claim_zero(mlx5_glue->close_device(ctx));
814 		rte_errno = ENOTSUP;
815 		return -ENOTSUP;
816 	}
817 	cdev = rte_compressdev_pmd_create(ibv->name, dev,
818 					  sizeof(*priv), &init_params);
819 	if (cdev == NULL) {
820 		DRV_LOG(ERR, "Failed to create device \"%s\".", ibv->name);
821 		claim_zero(mlx5_glue->close_device(ctx));
822 		return -ENODEV;
823 	}
824 	DRV_LOG(INFO,
825 		"Compress device %s was created successfully.", ibv->name);
826 	cdev->dev_ops = &mlx5_compress_ops;
827 	cdev->dequeue_burst = mlx5_compress_dequeue_burst;
828 	cdev->enqueue_burst = mlx5_compress_enqueue_burst;
829 	cdev->feature_flags = RTE_COMPDEV_FF_HW_ACCELERATED;
830 	priv = cdev->data->dev_private;
831 	priv->ctx = ctx;
832 	priv->cdev = cdev;
833 	priv->min_block_size = att.compress_min_block_size;
834 	priv->sq_ts_format = att.sq_ts_format;
835 	if (mlx5_compress_hw_global_prepare(priv) != 0) {
836 		rte_compressdev_pmd_destroy(priv->cdev);
837 		claim_zero(mlx5_glue->close_device(priv->ctx));
838 		return -1;
839 	}
840 	if (mlx5_mr_btree_init(&priv->mr_scache.cache,
841 			     MLX5_MR_BTREE_CACHE_N * 2, rte_socket_id()) != 0) {
842 		DRV_LOG(ERR, "Failed to allocate shared cache MR memory.");
843 		mlx5_compress_hw_global_release(priv);
844 		rte_compressdev_pmd_destroy(priv->cdev);
845 		claim_zero(mlx5_glue->close_device(priv->ctx));
846 		rte_errno = ENOMEM;
847 		return -rte_errno;
848 	}
849 	priv->mr_scache.reg_mr_cb = mlx5_common_verbs_reg_mr;
850 	priv->mr_scache.dereg_mr_cb = mlx5_common_verbs_dereg_mr;
851 	/* Register callback function for global shared MR cache management. */
852 	if (TAILQ_EMPTY(&mlx5_compress_priv_list))
853 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
854 						mlx5_compress_mr_mem_event_cb,
855 						NULL);
856 	pthread_mutex_lock(&priv_list_lock);
857 	TAILQ_INSERT_TAIL(&mlx5_compress_priv_list, priv, next);
858 	pthread_mutex_unlock(&priv_list_lock);
859 	return 0;
860 }
861 
862 static int
863 mlx5_compress_dev_remove(struct rte_device *dev)
864 {
865 	struct mlx5_compress_priv *priv = NULL;
866 
867 	pthread_mutex_lock(&priv_list_lock);
868 	TAILQ_FOREACH(priv, &mlx5_compress_priv_list, next)
869 		if (priv->cdev->device == dev)
870 			break;
871 	if (priv)
872 		TAILQ_REMOVE(&mlx5_compress_priv_list, priv, next);
873 	pthread_mutex_unlock(&priv_list_lock);
874 	if (priv) {
875 		if (TAILQ_EMPTY(&mlx5_compress_priv_list))
876 			rte_mem_event_callback_unregister("MLX5_MEM_EVENT_CB",
877 							  NULL);
878 		mlx5_mr_release_cache(&priv->mr_scache);
879 		mlx5_compress_hw_global_release(priv);
880 		rte_compressdev_pmd_destroy(priv->cdev);
881 		claim_zero(mlx5_glue->close_device(priv->ctx));
882 	}
883 	return 0;
884 }
885 
886 static const struct rte_pci_id mlx5_compress_pci_id_map[] = {
887 	{
888 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
889 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
890 	},
891 	{
892 		.vendor_id = 0
893 	}
894 };
895 
896 static struct mlx5_class_driver mlx5_compress_driver = {
897 	.drv_class = MLX5_CLASS_COMPRESS,
898 	.name = RTE_STR(MLX5_COMPRESS_DRIVER_NAME),
899 	.id_table = mlx5_compress_pci_id_map,
900 	.probe = mlx5_compress_dev_probe,
901 	.remove = mlx5_compress_dev_remove,
902 };
903 
904 RTE_INIT(rte_mlx5_compress_init)
905 {
906 	mlx5_common_init();
907 	if (mlx5_glue != NULL)
908 		mlx5_class_driver_register(&mlx5_compress_driver);
909 }
910 
911 RTE_LOG_REGISTER_DEFAULT(mlx5_compress_logtype, NOTICE)
912 RTE_PMD_EXPORT_NAME(MLX5_COMPRESS_DRIVER_NAME, __COUNTER__);
913 RTE_PMD_REGISTER_PCI_TABLE(MLX5_COMPRESS_DRIVER_NAME, mlx5_compress_pci_id_map);
914 RTE_PMD_REGISTER_KMOD_DEP(MLX5_COMPRESS_DRIVER_NAME, "* ib_uverbs & mlx5_core & mlx5_ib");
915