xref: /dpdk/drivers/compress/mlx5/mlx5_compress.c (revision 334ed198ab4d22f249d7b8274568253df33914a6)
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_devarg_params {
29 	uint32_t log_block_sz;
30 };
31 
32 struct mlx5_compress_xform {
33 	LIST_ENTRY(mlx5_compress_xform) next;
34 	enum rte_comp_xform_type type;
35 	enum rte_comp_checksum_type csum_type;
36 	uint32_t opcode;
37 	uint32_t gga_ctrl1; /* BE. */
38 };
39 
40 struct mlx5_compress_priv {
41 	TAILQ_ENTRY(mlx5_compress_priv) next;
42 	struct rte_compressdev *compressdev;
43 	struct mlx5_common_device *cdev; /* Backend mlx5 device. */
44 	void *uar;
45 	uint8_t min_block_size;
46 	/* Minimum huffman block size supported by the device. */
47 	struct rte_compressdev_config dev_config;
48 	LIST_HEAD(xform_list, mlx5_compress_xform) xform_list;
49 	rte_spinlock_t xform_sl;
50 	volatile uint64_t *uar_addr;
51 	/* HCA caps*/
52 	uint32_t mmo_decomp_sq:1;
53 	uint32_t mmo_decomp_qp:1;
54 	uint32_t mmo_comp_sq:1;
55 	uint32_t mmo_comp_qp:1;
56 	uint32_t mmo_dma_sq:1;
57 	uint32_t mmo_dma_qp:1;
58 	uint32_t log_block_sz;
59 #ifndef RTE_ARCH_64
60 	rte_spinlock_t uar32_sl;
61 #endif /* RTE_ARCH_64 */
62 };
63 
64 struct mlx5_compress_qp {
65 	uint16_t qp_id;
66 	uint16_t entries_n;
67 	uint16_t pi;
68 	uint16_t ci;
69 	struct mlx5_mr_ctrl mr_ctrl;
70 	int socket_id;
71 	struct mlx5_devx_cq cq;
72 	struct mlx5_devx_qp qp;
73 	struct mlx5_pmd_mr opaque_mr;
74 	struct rte_comp_op **ops;
75 	struct mlx5_compress_priv *priv;
76 	struct rte_compressdev_stats stats;
77 };
78 
79 TAILQ_HEAD(mlx5_compress_privs, mlx5_compress_priv) mlx5_compress_priv_list =
80 				TAILQ_HEAD_INITIALIZER(mlx5_compress_priv_list);
81 static pthread_mutex_t priv_list_lock = PTHREAD_MUTEX_INITIALIZER;
82 
83 int mlx5_compress_logtype;
84 
85 static const struct rte_compressdev_capabilities mlx5_caps[] = {
86 	{
87 		.algo = RTE_COMP_ALGO_NULL,
88 		.comp_feature_flags = RTE_COMP_FF_ADLER32_CHECKSUM |
89 				      RTE_COMP_FF_CRC32_CHECKSUM |
90 				      RTE_COMP_FF_CRC32_ADLER32_CHECKSUM |
91 				      RTE_COMP_FF_SHAREABLE_PRIV_XFORM,
92 	},
93 	{
94 		.algo = RTE_COMP_ALGO_DEFLATE,
95 		.comp_feature_flags = RTE_COMP_FF_ADLER32_CHECKSUM |
96 				      RTE_COMP_FF_CRC32_CHECKSUM |
97 				      RTE_COMP_FF_CRC32_ADLER32_CHECKSUM |
98 				      RTE_COMP_FF_SHAREABLE_PRIV_XFORM |
99 				      RTE_COMP_FF_HUFFMAN_FIXED |
100 				      RTE_COMP_FF_HUFFMAN_DYNAMIC,
101 		.window_size = {.min = 10, .max = 15, .increment = 1},
102 	},
103 	{
104 		.algo = RTE_COMP_ALGO_LIST_END,
105 	}
106 };
107 
108 static void
109 mlx5_compress_dev_info_get(struct rte_compressdev *dev,
110 			   struct rte_compressdev_info *info)
111 {
112 	RTE_SET_USED(dev);
113 	if (info != NULL) {
114 		info->max_nb_queue_pairs = MLX5_COMPRESS_MAX_QPS;
115 		info->feature_flags = RTE_COMPDEV_FF_HW_ACCELERATED;
116 		info->capabilities = mlx5_caps;
117 	}
118 }
119 
120 static int
121 mlx5_compress_dev_configure(struct rte_compressdev *dev,
122 			    struct rte_compressdev_config *config)
123 {
124 	struct mlx5_compress_priv *priv;
125 
126 	if (dev == NULL || config == NULL)
127 		return -EINVAL;
128 	priv = dev->data->dev_private;
129 	priv->dev_config = *config;
130 	return 0;
131 }
132 
133 static int
134 mlx5_compress_dev_close(struct rte_compressdev *dev)
135 {
136 	RTE_SET_USED(dev);
137 	return 0;
138 }
139 
140 static int
141 mlx5_compress_qp_release(struct rte_compressdev *dev, uint16_t qp_id)
142 {
143 	struct mlx5_compress_qp *qp = dev->data->queue_pairs[qp_id];
144 
145 	if (qp->qp.qp != NULL)
146 		mlx5_devx_qp_destroy(&qp->qp);
147 	if (qp->cq.cq != NULL)
148 		mlx5_devx_cq_destroy(&qp->cq);
149 	if (qp->opaque_mr.obj != NULL) {
150 		void *opaq = qp->opaque_mr.addr;
151 
152 		mlx5_common_verbs_dereg_mr(&qp->opaque_mr);
153 		if (opaq != NULL)
154 			rte_free(opaq);
155 	}
156 	mlx5_mr_btree_free(&qp->mr_ctrl.cache_bh);
157 	rte_free(qp);
158 	dev->data->queue_pairs[qp_id] = NULL;
159 	return 0;
160 }
161 
162 static void
163 mlx5_compress_init_qp(struct mlx5_compress_qp *qp)
164 {
165 	volatile struct mlx5_gga_wqe *restrict wqe =
166 				    (volatile struct mlx5_gga_wqe *)qp->qp.wqes;
167 	volatile struct mlx5_gga_compress_opaque *opaq = qp->opaque_mr.addr;
168 	const uint32_t sq_ds = rte_cpu_to_be_32((qp->qp.qp->id << 8) | 4u);
169 	const uint32_t flags = RTE_BE32(MLX5_COMP_ALWAYS <<
170 					MLX5_COMP_MODE_OFFSET);
171 	const uint32_t opaq_lkey = rte_cpu_to_be_32(qp->opaque_mr.lkey);
172 	int i;
173 
174 	/* All the next fields state should stay constant. */
175 	for (i = 0; i < qp->entries_n; ++i, ++wqe) {
176 		wqe->sq_ds = sq_ds;
177 		wqe->flags = flags;
178 		wqe->opaque_lkey = opaq_lkey;
179 		wqe->opaque_vaddr = rte_cpu_to_be_64
180 						((uint64_t)(uintptr_t)&opaq[i]);
181 	}
182 }
183 
184 static int
185 mlx5_compress_qp_setup(struct rte_compressdev *dev, uint16_t qp_id,
186 		       uint32_t max_inflight_ops, int socket_id)
187 {
188 	struct mlx5_compress_priv *priv = dev->data->dev_private;
189 	struct mlx5_compress_qp *qp;
190 	struct mlx5_devx_cq_attr cq_attr = {
191 		.uar_page_id = mlx5_os_get_devx_uar_page_id(priv->uar),
192 	};
193 	struct mlx5_devx_qp_attr qp_attr = {
194 		.pd = priv->cdev->pdn,
195 		.uar_index = mlx5_os_get_devx_uar_page_id(priv->uar),
196 		.user_index = qp_id,
197 	};
198 	uint32_t log_ops_n = rte_log2_u32(max_inflight_ops);
199 	uint32_t alloc_size = sizeof(*qp);
200 	void *opaq_buf;
201 	int ret;
202 
203 	alloc_size = RTE_ALIGN(alloc_size, RTE_CACHE_LINE_SIZE);
204 	alloc_size += sizeof(struct rte_comp_op *) * (1u << log_ops_n);
205 	qp = rte_zmalloc_socket(__func__, alloc_size, RTE_CACHE_LINE_SIZE,
206 				socket_id);
207 	if (qp == NULL) {
208 		DRV_LOG(ERR, "Failed to allocate qp memory.");
209 		rte_errno = ENOMEM;
210 		return -rte_errno;
211 	}
212 	dev->data->queue_pairs[qp_id] = qp;
213 	if (mlx5_mr_ctrl_init(&qp->mr_ctrl, priv->cdev,
214 			      priv->dev_config.socket_id)) {
215 		DRV_LOG(ERR, "Cannot allocate MR Btree for qp %u.",
216 			(uint32_t)qp_id);
217 		rte_errno = ENOMEM;
218 		goto err;
219 	}
220 	opaq_buf = rte_calloc(__func__, (size_t)1 << log_ops_n,
221 			      sizeof(struct mlx5_gga_compress_opaque),
222 			      sizeof(struct mlx5_gga_compress_opaque));
223 	if (opaq_buf == NULL) {
224 		DRV_LOG(ERR, "Failed to allocate opaque memory.");
225 		rte_errno = ENOMEM;
226 		goto err;
227 	}
228 	qp->entries_n = 1 << log_ops_n;
229 	qp->socket_id = socket_id;
230 	qp->qp_id = qp_id;
231 	qp->priv = priv;
232 	qp->ops = (struct rte_comp_op **)RTE_ALIGN((uintptr_t)(qp + 1),
233 						   RTE_CACHE_LINE_SIZE);
234 	if (mlx5_common_verbs_reg_mr(priv->cdev->pd, opaq_buf, qp->entries_n *
235 					sizeof(struct mlx5_gga_compress_opaque),
236 							 &qp->opaque_mr) != 0) {
237 		rte_free(opaq_buf);
238 		DRV_LOG(ERR, "Failed to register opaque MR.");
239 		rte_errno = ENOMEM;
240 		goto err;
241 	}
242 	ret = mlx5_devx_cq_create(priv->cdev->ctx, &qp->cq, log_ops_n, &cq_attr,
243 				  socket_id);
244 	if (ret != 0) {
245 		DRV_LOG(ERR, "Failed to create CQ.");
246 		goto err;
247 	}
248 	qp_attr.cqn = qp->cq.cq->id;
249 	qp_attr.ts_format =
250 		mlx5_ts_format_conv(priv->cdev->config.hca_attr.qp_ts_format);
251 	qp_attr.rq_size = 0;
252 	qp_attr.sq_size = RTE_BIT32(log_ops_n);
253 	qp_attr.mmo = priv->mmo_decomp_qp && priv->mmo_comp_qp
254 			&& priv->mmo_dma_qp;
255 	ret = mlx5_devx_qp_create(priv->cdev->ctx, &qp->qp, log_ops_n, &qp_attr,
256 				  socket_id);
257 	if (ret != 0) {
258 		DRV_LOG(ERR, "Failed to create QP.");
259 		goto err;
260 	}
261 	mlx5_compress_init_qp(qp);
262 	ret = mlx5_devx_qp2rts(&qp->qp, 0);
263 	if (ret)
264 		goto err;
265 	DRV_LOG(INFO, "QP %u: SQN=0x%X CQN=0x%X entries num = %u",
266 		(uint32_t)qp_id, qp->qp.qp->id, qp->cq.cq->id, qp->entries_n);
267 	return 0;
268 err:
269 	mlx5_compress_qp_release(dev, qp_id);
270 	return -1;
271 }
272 
273 static int
274 mlx5_compress_xform_free(struct rte_compressdev *dev, void *xform)
275 {
276 	struct mlx5_compress_priv *priv = dev->data->dev_private;
277 
278 	rte_spinlock_lock(&priv->xform_sl);
279 	LIST_REMOVE((struct mlx5_compress_xform *)xform, next);
280 	rte_spinlock_unlock(&priv->xform_sl);
281 	rte_free(xform);
282 	return 0;
283 }
284 
285 static int
286 mlx5_compress_xform_create(struct rte_compressdev *dev,
287 			   const struct rte_comp_xform *xform,
288 			   void **private_xform)
289 {
290 	struct mlx5_compress_priv *priv = dev->data->dev_private;
291 	struct mlx5_compress_xform *xfrm;
292 	uint32_t size;
293 
294 	switch (xform->type) {
295 	case RTE_COMP_COMPRESS:
296 		if (xform->compress.algo == RTE_COMP_ALGO_NULL &&
297 				!priv->mmo_dma_qp && !priv->mmo_dma_sq) {
298 			DRV_LOG(ERR, "Not enough capabilities to support DMA operation, maybe old FW/OFED version?");
299 			return -ENOTSUP;
300 		} else if (!priv->mmo_comp_qp && !priv->mmo_comp_sq) {
301 			DRV_LOG(ERR, "Not enough capabilities to support compress operation, maybe old FW/OFED version?");
302 			return -ENOTSUP;
303 		}
304 		if (xform->compress.level == RTE_COMP_LEVEL_NONE) {
305 			DRV_LOG(ERR, "Non-compressed block is not supported.");
306 			return -ENOTSUP;
307 		}
308 		if (xform->compress.hash_algo != RTE_COMP_HASH_ALGO_NONE) {
309 			DRV_LOG(ERR, "SHA is not supported.");
310 			return -ENOTSUP;
311 		}
312 		break;
313 	case RTE_COMP_DECOMPRESS:
314 		if (xform->decompress.algo == RTE_COMP_ALGO_NULL &&
315 				!priv->mmo_dma_qp && !priv->mmo_dma_sq) {
316 			DRV_LOG(ERR, "Not enough capabilities to support DMA operation, maybe old FW/OFED version?");
317 			return -ENOTSUP;
318 		} else if (!priv->mmo_decomp_qp && !priv->mmo_decomp_sq) {
319 			DRV_LOG(ERR, "Not enough capabilities to support decompress operation, maybe old FW/OFED version?");
320 			return -ENOTSUP;
321 		}
322 		if (xform->compress.hash_algo != RTE_COMP_HASH_ALGO_NONE) {
323 			DRV_LOG(ERR, "SHA is not supported.");
324 			return -ENOTSUP;
325 		}
326 		break;
327 	default:
328 		DRV_LOG(ERR, "Xform type should be compress/decompress");
329 		return -ENOTSUP;
330 	}
331 
332 	xfrm = rte_zmalloc_socket(__func__, sizeof(*xfrm), 0,
333 						    priv->dev_config.socket_id);
334 	if (xfrm == NULL)
335 		return -ENOMEM;
336 	xfrm->opcode = MLX5_OPCODE_MMO;
337 	xfrm->type = xform->type;
338 	switch (xform->type) {
339 	case RTE_COMP_COMPRESS:
340 		switch (xform->compress.algo) {
341 		case RTE_COMP_ALGO_NULL:
342 			xfrm->opcode += MLX5_OPC_MOD_MMO_DMA <<
343 							WQE_CSEG_OPC_MOD_OFFSET;
344 			break;
345 		case RTE_COMP_ALGO_DEFLATE:
346 			size = 1 << xform->compress.window_size;
347 			size /= MLX5_GGA_COMP_WIN_SIZE_UNITS;
348 			xfrm->gga_ctrl1 += RTE_MIN(rte_log2_u32(size),
349 					 MLX5_COMP_MAX_WIN_SIZE_CONF) <<
350 						WQE_GGA_COMP_WIN_SIZE_OFFSET;
351 			size = priv->log_block_sz;
352 			xfrm->gga_ctrl1 += size <<
353 						WQE_GGA_COMP_BLOCK_SIZE_OFFSET;
354 			xfrm->opcode += MLX5_OPC_MOD_MMO_COMP <<
355 							WQE_CSEG_OPC_MOD_OFFSET;
356 			size = xform->compress.deflate.huffman ==
357 						      RTE_COMP_HUFFMAN_DYNAMIC ?
358 					    MLX5_GGA_COMP_LOG_DYNAMIC_SIZE_MAX :
359 					     MLX5_GGA_COMP_LOG_DYNAMIC_SIZE_MIN;
360 			xfrm->gga_ctrl1 += size <<
361 					       WQE_GGA_COMP_DYNAMIC_SIZE_OFFSET;
362 			break;
363 		default:
364 			goto err;
365 		}
366 		xfrm->csum_type = xform->compress.chksum;
367 		break;
368 	case RTE_COMP_DECOMPRESS:
369 		switch (xform->decompress.algo) {
370 		case RTE_COMP_ALGO_NULL:
371 			xfrm->opcode += MLX5_OPC_MOD_MMO_DMA <<
372 							WQE_CSEG_OPC_MOD_OFFSET;
373 			break;
374 		case RTE_COMP_ALGO_DEFLATE:
375 			xfrm->opcode += MLX5_OPC_MOD_MMO_DECOMP <<
376 							WQE_CSEG_OPC_MOD_OFFSET;
377 			break;
378 		default:
379 			goto err;
380 		}
381 		xfrm->csum_type = xform->decompress.chksum;
382 		break;
383 	default:
384 		DRV_LOG(ERR, "Algorithm %u is not supported.", xform->type);
385 		goto err;
386 	}
387 	DRV_LOG(DEBUG, "New xform: gga ctrl1 = 0x%08X opcode = 0x%08X csum "
388 		"type = %d.", xfrm->gga_ctrl1, xfrm->opcode, xfrm->csum_type);
389 	xfrm->gga_ctrl1 = rte_cpu_to_be_32(xfrm->gga_ctrl1);
390 	rte_spinlock_lock(&priv->xform_sl);
391 	LIST_INSERT_HEAD(&priv->xform_list, xfrm, next);
392 	rte_spinlock_unlock(&priv->xform_sl);
393 	*private_xform = xfrm;
394 	return 0;
395 err:
396 	rte_free(xfrm);
397 	return -ENOTSUP;
398 }
399 
400 static void
401 mlx5_compress_dev_stop(struct rte_compressdev *dev)
402 {
403 	RTE_SET_USED(dev);
404 }
405 
406 static int
407 mlx5_compress_dev_start(struct rte_compressdev *dev)
408 {
409 	struct mlx5_compress_priv *priv = dev->data->dev_private;
410 
411 	return mlx5_dev_mempool_subscribe(priv->cdev);
412 }
413 
414 static void
415 mlx5_compress_stats_get(struct rte_compressdev *dev,
416 		struct rte_compressdev_stats *stats)
417 {
418 	int qp_id;
419 
420 	for (qp_id = 0; qp_id < dev->data->nb_queue_pairs; qp_id++) {
421 		struct mlx5_compress_qp *qp = dev->data->queue_pairs[qp_id];
422 
423 		stats->enqueued_count += qp->stats.enqueued_count;
424 		stats->dequeued_count += qp->stats.dequeued_count;
425 		stats->enqueue_err_count += qp->stats.enqueue_err_count;
426 		stats->dequeue_err_count += qp->stats.dequeue_err_count;
427 	}
428 }
429 
430 static void
431 mlx5_compress_stats_reset(struct rte_compressdev *dev)
432 {
433 	int qp_id;
434 
435 	for (qp_id = 0; qp_id < dev->data->nb_queue_pairs; qp_id++) {
436 		struct mlx5_compress_qp *qp = dev->data->queue_pairs[qp_id];
437 
438 		memset(&qp->stats, 0, sizeof(qp->stats));
439 	}
440 }
441 
442 static struct rte_compressdev_ops mlx5_compress_ops = {
443 	.dev_configure		= mlx5_compress_dev_configure,
444 	.dev_start		= mlx5_compress_dev_start,
445 	.dev_stop		= mlx5_compress_dev_stop,
446 	.dev_close		= mlx5_compress_dev_close,
447 	.dev_infos_get		= mlx5_compress_dev_info_get,
448 	.stats_get		= mlx5_compress_stats_get,
449 	.stats_reset		= mlx5_compress_stats_reset,
450 	.queue_pair_setup	= mlx5_compress_qp_setup,
451 	.queue_pair_release	= mlx5_compress_qp_release,
452 	.private_xform_create	= mlx5_compress_xform_create,
453 	.private_xform_free	= mlx5_compress_xform_free,
454 	.stream_create		= NULL,
455 	.stream_free		= NULL,
456 };
457 
458 static __rte_always_inline uint32_t
459 mlx5_compress_dseg_set(struct mlx5_compress_qp *qp,
460 		       volatile struct mlx5_wqe_dseg *restrict dseg,
461 		       struct rte_mbuf *restrict mbuf,
462 		       uint32_t offset, uint32_t len)
463 {
464 	uintptr_t addr = rte_pktmbuf_mtod_offset(mbuf, uintptr_t, offset);
465 
466 	dseg->bcount = rte_cpu_to_be_32(len);
467 	dseg->lkey = mlx5_mr_mb2mr(&qp->mr_ctrl, mbuf, 0);
468 	dseg->pbuf = rte_cpu_to_be_64(addr);
469 	return dseg->lkey;
470 }
471 
472 /*
473  * Provide safe 64bit store operation to mlx5 UAR region for both 32bit and
474  * 64bit architectures.
475  */
476 static __rte_always_inline void
477 mlx5_compress_uar_write(uint64_t val, struct mlx5_compress_priv *priv)
478 {
479 #ifdef RTE_ARCH_64
480 	*priv->uar_addr = val;
481 #else /* !RTE_ARCH_64 */
482 	rte_spinlock_lock(&priv->uar32_sl);
483 	*(volatile uint32_t *)priv->uar_addr = val;
484 	rte_io_wmb();
485 	*((volatile uint32_t *)priv->uar_addr + 1) = val >> 32;
486 	rte_spinlock_unlock(&priv->uar32_sl);
487 #endif
488 }
489 
490 static uint16_t
491 mlx5_compress_enqueue_burst(void *queue_pair, struct rte_comp_op **ops,
492 			    uint16_t nb_ops)
493 {
494 	struct mlx5_compress_qp *qp = queue_pair;
495 	volatile struct mlx5_gga_wqe *wqes = (volatile struct mlx5_gga_wqe *)
496 							      qp->qp.wqes, *wqe;
497 	struct mlx5_compress_xform *xform;
498 	struct rte_comp_op *op;
499 	uint16_t mask = qp->entries_n - 1;
500 	uint16_t remain = qp->entries_n - (qp->pi - qp->ci);
501 	uint16_t idx;
502 	bool invalid;
503 
504 	if (remain < nb_ops)
505 		nb_ops = remain;
506 	else
507 		remain = nb_ops;
508 	if (unlikely(remain == 0))
509 		return 0;
510 	do {
511 		idx = qp->pi & mask;
512 		wqe = &wqes[idx];
513 		rte_prefetch0(&wqes[(qp->pi + 1) & mask]);
514 		op = *ops++;
515 		xform = op->private_xform;
516 		/*
517 		 * Check operation arguments and error cases:
518 		 *   - Operation type must be state-less.
519 		 *   - Compress operation flush flag must be FULL or FINAL.
520 		 *   - Source and destination buffers must be mapped internally.
521 		 */
522 		invalid = op->op_type != RTE_COMP_OP_STATELESS ||
523 					    (xform->type == RTE_COMP_COMPRESS &&
524 					  op->flush_flag < RTE_COMP_FLUSH_FULL);
525 		if (unlikely(invalid ||
526 			     (mlx5_compress_dseg_set(qp, &wqe->gather,
527 						     op->m_src,
528 						     op->src.offset,
529 						     op->src.length) ==
530 								  UINT32_MAX) ||
531 			     (mlx5_compress_dseg_set(qp, &wqe->scatter,
532 						op->m_dst,
533 						op->dst.offset,
534 						rte_pktmbuf_pkt_len(op->m_dst) -
535 							      op->dst.offset) ==
536 								 UINT32_MAX))) {
537 			op->status = invalid ? RTE_COMP_OP_STATUS_INVALID_ARGS :
538 						       RTE_COMP_OP_STATUS_ERROR;
539 			nb_ops -= remain;
540 			if (unlikely(nb_ops == 0))
541 				return 0;
542 			break;
543 		}
544 		wqe->gga_ctrl1 = xform->gga_ctrl1;
545 		wqe->opcode = rte_cpu_to_be_32(xform->opcode + (qp->pi << 8));
546 		qp->ops[idx] = op;
547 		qp->pi++;
548 	} while (--remain);
549 	qp->stats.enqueued_count += nb_ops;
550 	rte_io_wmb();
551 	qp->qp.db_rec[MLX5_SND_DBR] = rte_cpu_to_be_32(qp->pi);
552 	rte_wmb();
553 	mlx5_compress_uar_write(*(volatile uint64_t *)wqe, qp->priv);
554 	rte_wmb();
555 	return nb_ops;
556 }
557 
558 static void
559 mlx5_compress_dump_err_objs(volatile uint32_t *cqe, volatile uint32_t *wqe,
560 			     volatile uint32_t *opaq)
561 {
562 	size_t i;
563 
564 	DRV_LOG(ERR, "Error cqe:");
565 	for (i = 0; i < sizeof(struct mlx5_err_cqe) >> 2; i += 4)
566 		DRV_LOG(ERR, "%08X %08X %08X %08X", cqe[i], cqe[i + 1],
567 			cqe[i + 2], cqe[i + 3]);
568 	DRV_LOG(ERR, "\nError wqe:");
569 	for (i = 0; i < sizeof(struct mlx5_gga_wqe) >> 2; i += 4)
570 		DRV_LOG(ERR, "%08X %08X %08X %08X", wqe[i], wqe[i + 1],
571 			wqe[i + 2], wqe[i + 3]);
572 	DRV_LOG(ERR, "\nError opaq:");
573 	for (i = 0; i < sizeof(struct mlx5_gga_compress_opaque) >> 2; i += 4)
574 		DRV_LOG(ERR, "%08X %08X %08X %08X", opaq[i], opaq[i + 1],
575 			opaq[i + 2], opaq[i + 3]);
576 }
577 
578 static void
579 mlx5_compress_cqe_err_handle(struct mlx5_compress_qp *qp,
580 			     struct rte_comp_op *op)
581 {
582 	const uint32_t idx = qp->ci & (qp->entries_n - 1);
583 	volatile struct mlx5_err_cqe *cqe = (volatile struct mlx5_err_cqe *)
584 							      &qp->cq.cqes[idx];
585 	volatile struct mlx5_gga_wqe *wqes = (volatile struct mlx5_gga_wqe *)
586 								    qp->qp.wqes;
587 	volatile struct mlx5_gga_compress_opaque *opaq = qp->opaque_mr.addr;
588 
589 	op->status = RTE_COMP_OP_STATUS_ERROR;
590 	op->consumed = 0;
591 	op->produced = 0;
592 	op->output_chksum = 0;
593 	op->debug_status = rte_be_to_cpu_32(opaq[idx].syndrom) |
594 			      ((uint64_t)rte_be_to_cpu_32(cqe->syndrome) << 32);
595 	mlx5_compress_dump_err_objs((volatile uint32_t *)cqe,
596 				 (volatile uint32_t *)&wqes[idx],
597 				 (volatile uint32_t *)&opaq[idx]);
598 	qp->stats.dequeue_err_count++;
599 }
600 
601 static uint16_t
602 mlx5_compress_dequeue_burst(void *queue_pair, struct rte_comp_op **ops,
603 			    uint16_t nb_ops)
604 {
605 	struct mlx5_compress_qp *qp = queue_pair;
606 	volatile struct mlx5_compress_xform *restrict xform;
607 	volatile struct mlx5_cqe *restrict cqe;
608 	volatile struct mlx5_gga_compress_opaque *opaq = qp->opaque_mr.addr;
609 	struct rte_comp_op *restrict op;
610 	const unsigned int cq_size = qp->entries_n;
611 	const unsigned int mask = cq_size - 1;
612 	uint32_t idx;
613 	uint32_t next_idx = qp->ci & mask;
614 	const uint16_t max = RTE_MIN((uint16_t)(qp->pi - qp->ci), nb_ops);
615 	uint16_t i = 0;
616 	int ret;
617 
618 	if (unlikely(max == 0))
619 		return 0;
620 	do {
621 		idx = next_idx;
622 		next_idx = (qp->ci + 1) & mask;
623 		rte_prefetch0(&qp->cq.cqes[next_idx]);
624 		rte_prefetch0(qp->ops[next_idx]);
625 		op = qp->ops[idx];
626 		cqe = &qp->cq.cqes[idx];
627 		ret = check_cqe(cqe, cq_size, qp->ci);
628 		/*
629 		 * Be sure owner read is done before any other cookie field or
630 		 * opaque field.
631 		 */
632 		rte_io_rmb();
633 		if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
634 			if (likely(ret == MLX5_CQE_STATUS_HW_OWN))
635 				break;
636 			mlx5_compress_cqe_err_handle(qp, op);
637 		} else {
638 			xform = op->private_xform;
639 			op->status = RTE_COMP_OP_STATUS_SUCCESS;
640 			op->consumed = op->src.length;
641 			op->produced = rte_be_to_cpu_32(cqe->byte_cnt);
642 			MLX5_ASSERT(cqe->byte_cnt ==
643 				    opaq[idx].scattered_length);
644 			switch (xform->csum_type) {
645 			case RTE_COMP_CHECKSUM_CRC32:
646 				op->output_chksum = (uint64_t)rte_be_to_cpu_32
647 						    (opaq[idx].crc32);
648 				break;
649 			case RTE_COMP_CHECKSUM_ADLER32:
650 				op->output_chksum = (uint64_t)rte_be_to_cpu_32
651 					    (opaq[idx].adler32) << 32;
652 				break;
653 			case RTE_COMP_CHECKSUM_CRC32_ADLER32:
654 				op->output_chksum = (uint64_t)rte_be_to_cpu_32
655 							     (opaq[idx].crc32) |
656 						     ((uint64_t)rte_be_to_cpu_32
657 						     (opaq[idx].adler32) << 32);
658 				break;
659 			default:
660 				break;
661 			}
662 		}
663 		ops[i++] = op;
664 		qp->ci++;
665 	} while (i < max);
666 	if (likely(i != 0)) {
667 		rte_io_wmb();
668 		qp->cq.db_rec[0] = rte_cpu_to_be_32(qp->ci);
669 		qp->stats.dequeued_count += i;
670 	}
671 	return i;
672 }
673 
674 static void
675 mlx5_compress_uar_release(struct mlx5_compress_priv *priv)
676 {
677 	if (priv->uar != NULL) {
678 		mlx5_glue->devx_free_uar(priv->uar);
679 		priv->uar = NULL;
680 	}
681 }
682 
683 static int
684 mlx5_compress_uar_prepare(struct mlx5_compress_priv *priv)
685 {
686 	priv->uar = mlx5_devx_alloc_uar(priv->cdev->ctx, -1);
687 	if (priv->uar == NULL || mlx5_os_get_devx_uar_reg_addr(priv->uar) ==
688 	    NULL) {
689 		rte_errno = errno;
690 		DRV_LOG(ERR, "Failed to allocate UAR.");
691 		return -1;
692 	}
693 	priv->uar_addr = mlx5_os_get_devx_uar_reg_addr(priv->uar);
694 	MLX5_ASSERT(priv->uar_addr);
695 #ifndef RTE_ARCH_64
696 	rte_spinlock_init(&priv->uar32_sl);
697 #endif /* RTE_ARCH_64 */
698 	return 0;
699 }
700 
701 static int
702 mlx5_compress_args_check_handler(const char *key, const char *val, void *opaque)
703 {
704 	struct mlx5_compress_devarg_params *devarg_prms = opaque;
705 
706 	if (strcmp(key, "log-block-size") == 0) {
707 		errno = 0;
708 		devarg_prms->log_block_sz = (uint32_t)strtoul(val, NULL, 10);
709 		if (errno) {
710 			DRV_LOG(WARNING, "%s: \"%s\" is an invalid integer."
711 				, key, val);
712 			return -errno;
713 		}
714 		return 0;
715 	}
716 	return 0;
717 }
718 
719 static int
720 mlx5_compress_handle_devargs(struct rte_devargs *devargs,
721 			  struct mlx5_compress_devarg_params *devarg_prms,
722 			  struct mlx5_hca_attr *att)
723 {
724 	struct rte_kvargs *kvlist;
725 
726 	devarg_prms->log_block_sz = MLX5_GGA_COMP_LOG_BLOCK_SIZE_MAX;
727 	if (devargs == NULL)
728 		return 0;
729 	kvlist = rte_kvargs_parse(devargs->args, NULL);
730 	if (kvlist == NULL) {
731 		DRV_LOG(ERR, "Failed to parse devargs.");
732 		rte_errno = EINVAL;
733 		return -1;
734 	}
735 	if (rte_kvargs_process(kvlist, NULL, mlx5_compress_args_check_handler,
736 			   devarg_prms) != 0) {
737 		DRV_LOG(ERR, "Devargs handler function Failed.");
738 		rte_kvargs_free(kvlist);
739 		rte_errno = EINVAL;
740 		return -1;
741 	}
742 	rte_kvargs_free(kvlist);
743 	if (devarg_prms->log_block_sz > MLX5_GGA_COMP_LOG_BLOCK_SIZE_MAX ||
744 		devarg_prms->log_block_sz < att->compress_min_block_size) {
745 		DRV_LOG(WARNING, "Log block size provided is out of range("
746 			"%u); default it to %u.",
747 			devarg_prms->log_block_sz,
748 			MLX5_GGA_COMP_LOG_BLOCK_SIZE_MAX);
749 		devarg_prms->log_block_sz = MLX5_GGA_COMP_LOG_BLOCK_SIZE_MAX;
750 	}
751 	return 0;
752 }
753 
754 static int
755 mlx5_compress_dev_probe(struct mlx5_common_device *cdev)
756 {
757 	struct rte_compressdev *compressdev;
758 	struct mlx5_compress_priv *priv;
759 	struct mlx5_hca_attr *attr = &cdev->config.hca_attr;
760 	struct mlx5_compress_devarg_params devarg_prms = {0};
761 	struct rte_compressdev_pmd_init_params init_params = {
762 		.name = "",
763 		.socket_id = cdev->dev->numa_node,
764 	};
765 	const char *ibdev_name = mlx5_os_get_ctx_device_name(cdev->ctx);
766 
767 	if (rte_eal_process_type() != RTE_PROC_PRIMARY) {
768 		DRV_LOG(ERR, "Non-primary process type is not supported.");
769 		rte_errno = ENOTSUP;
770 		return -rte_errno;
771 	}
772 	if (!attr->mmo_decompress_qp_en && !attr->mmo_decompress_sq_en
773 		&& !attr->mmo_compress_qp_en && !attr->mmo_compress_sq_en
774 		&& !attr->mmo_dma_qp_en && !attr->mmo_dma_sq_en) {
775 		DRV_LOG(ERR, "Not enough capabilities to support compress operations, maybe old FW/OFED version?");
776 		claim_zero(mlx5_glue->close_device(cdev->ctx));
777 		rte_errno = ENOTSUP;
778 		return -ENOTSUP;
779 	}
780 	mlx5_compress_handle_devargs(cdev->dev->devargs, &devarg_prms, attr);
781 	compressdev = rte_compressdev_pmd_create(ibdev_name, cdev->dev,
782 						 sizeof(*priv), &init_params);
783 	if (compressdev == NULL) {
784 		DRV_LOG(ERR, "Failed to create device \"%s\".", ibdev_name);
785 		return -ENODEV;
786 	}
787 	DRV_LOG(INFO,
788 		"Compress device %s was created successfully.", ibdev_name);
789 	compressdev->dev_ops = &mlx5_compress_ops;
790 	compressdev->dequeue_burst = mlx5_compress_dequeue_burst;
791 	compressdev->enqueue_burst = mlx5_compress_enqueue_burst;
792 	compressdev->feature_flags = RTE_COMPDEV_FF_HW_ACCELERATED;
793 	priv = compressdev->data->dev_private;
794 	priv->log_block_sz = devarg_prms.log_block_sz;
795 	priv->mmo_decomp_sq = attr->mmo_decompress_sq_en;
796 	priv->mmo_decomp_qp = attr->mmo_decompress_qp_en;
797 	priv->mmo_comp_sq = attr->mmo_compress_sq_en;
798 	priv->mmo_comp_qp = attr->mmo_compress_qp_en;
799 	priv->mmo_dma_sq = attr->mmo_dma_sq_en;
800 	priv->mmo_dma_qp = attr->mmo_dma_qp_en;
801 	priv->cdev = cdev;
802 	priv->compressdev = compressdev;
803 	priv->min_block_size = attr->compress_min_block_size;
804 	if (mlx5_compress_uar_prepare(priv) != 0) {
805 		rte_compressdev_pmd_destroy(priv->compressdev);
806 		return -1;
807 	}
808 	pthread_mutex_lock(&priv_list_lock);
809 	TAILQ_INSERT_TAIL(&mlx5_compress_priv_list, priv, next);
810 	pthread_mutex_unlock(&priv_list_lock);
811 	return 0;
812 }
813 
814 static int
815 mlx5_compress_dev_remove(struct mlx5_common_device *cdev)
816 {
817 	struct mlx5_compress_priv *priv = NULL;
818 
819 	pthread_mutex_lock(&priv_list_lock);
820 	TAILQ_FOREACH(priv, &mlx5_compress_priv_list, next)
821 		if (priv->compressdev->device == cdev->dev)
822 			break;
823 	if (priv)
824 		TAILQ_REMOVE(&mlx5_compress_priv_list, priv, next);
825 	pthread_mutex_unlock(&priv_list_lock);
826 	if (priv) {
827 		mlx5_compress_uar_release(priv);
828 		rte_compressdev_pmd_destroy(priv->compressdev);
829 	}
830 	return 0;
831 }
832 
833 static const struct rte_pci_id mlx5_compress_pci_id_map[] = {
834 	{
835 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
836 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
837 	},
838 	{
839 		.vendor_id = 0
840 	}
841 };
842 
843 static struct mlx5_class_driver mlx5_compress_driver = {
844 	.drv_class = MLX5_CLASS_COMPRESS,
845 	.name = RTE_STR(MLX5_COMPRESS_DRIVER_NAME),
846 	.id_table = mlx5_compress_pci_id_map,
847 	.probe = mlx5_compress_dev_probe,
848 	.remove = mlx5_compress_dev_remove,
849 };
850 
851 RTE_INIT(rte_mlx5_compress_init)
852 {
853 	mlx5_common_init();
854 	if (mlx5_glue != NULL)
855 		mlx5_class_driver_register(&mlx5_compress_driver);
856 }
857 
858 RTE_LOG_REGISTER_DEFAULT(mlx5_compress_logtype, NOTICE)
859 RTE_PMD_EXPORT_NAME(MLX5_COMPRESS_DRIVER_NAME, __COUNTER__);
860 RTE_PMD_REGISTER_PCI_TABLE(MLX5_COMPRESS_DRIVER_NAME, mlx5_compress_pci_id_map);
861 RTE_PMD_REGISTER_KMOD_DEP(MLX5_COMPRESS_DRIVER_NAME, "* ib_uverbs & mlx5_core & mlx5_ib");
862