xref: /openbmc/linux/drivers/infiniband/hw/mlx5/wr.c (revision 11788d9b)
1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3  * Copyright (c) 2020, Mellanox Technologies inc. All rights reserved.
4  */
5 
6 #include <linux/gfp.h>
7 #include <linux/mlx5/qp.h>
8 #include <linux/mlx5/driver.h>
9 #include "wr.h"
10 
11 static const u32 mlx5_ib_opcode[] = {
12 	[IB_WR_SEND]				= MLX5_OPCODE_SEND,
13 	[IB_WR_LSO]				= MLX5_OPCODE_LSO,
14 	[IB_WR_SEND_WITH_IMM]			= MLX5_OPCODE_SEND_IMM,
15 	[IB_WR_RDMA_WRITE]			= MLX5_OPCODE_RDMA_WRITE,
16 	[IB_WR_RDMA_WRITE_WITH_IMM]		= MLX5_OPCODE_RDMA_WRITE_IMM,
17 	[IB_WR_RDMA_READ]			= MLX5_OPCODE_RDMA_READ,
18 	[IB_WR_ATOMIC_CMP_AND_SWP]		= MLX5_OPCODE_ATOMIC_CS,
19 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= MLX5_OPCODE_ATOMIC_FA,
20 	[IB_WR_SEND_WITH_INV]			= MLX5_OPCODE_SEND_INVAL,
21 	[IB_WR_LOCAL_INV]			= MLX5_OPCODE_UMR,
22 	[IB_WR_REG_MR]				= MLX5_OPCODE_UMR,
23 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= MLX5_OPCODE_ATOMIC_MASKED_CS,
24 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= MLX5_OPCODE_ATOMIC_MASKED_FA,
25 	[MLX5_IB_WR_UMR]			= MLX5_OPCODE_UMR,
26 };
27 
28 /* handle_post_send_edge - Check if we get to SQ edge. If yes, update to the
29  * next nearby edge and get new address translation for current WQE position.
30  * @sq - SQ buffer.
31  * @seg: Current WQE position (16B aligned).
32  * @wqe_sz: Total current WQE size [16B].
33  * @cur_edge: Updated current edge.
34  */
35 static inline void handle_post_send_edge(struct mlx5_ib_wq *sq, void **seg,
36 					 u32 wqe_sz, void **cur_edge)
37 {
38 	u32 idx;
39 
40 	if (likely(*seg != *cur_edge))
41 		return;
42 
43 	idx = (sq->cur_post + (wqe_sz >> 2)) & (sq->wqe_cnt - 1);
44 	*cur_edge = get_sq_edge(sq, idx);
45 
46 	*seg = mlx5_frag_buf_get_wqe(&sq->fbc, idx);
47 }
48 
49 /* memcpy_send_wqe - copy data from src to WQE and update the relevant WQ's
50  * pointers. At the end @seg is aligned to 16B regardless the copied size.
51  * @sq - SQ buffer.
52  * @cur_edge: Updated current edge.
53  * @seg: Current WQE position (16B aligned).
54  * @wqe_sz: Total current WQE size [16B].
55  * @src: Pointer to copy from.
56  * @n: Number of bytes to copy.
57  */
58 static inline void memcpy_send_wqe(struct mlx5_ib_wq *sq, void **cur_edge,
59 				   void **seg, u32 *wqe_sz, const void *src,
60 				   size_t n)
61 {
62 	while (likely(n)) {
63 		size_t leftlen = *cur_edge - *seg;
64 		size_t copysz = min_t(size_t, leftlen, n);
65 		size_t stride;
66 
67 		memcpy(*seg, src, copysz);
68 
69 		n -= copysz;
70 		src += copysz;
71 		stride = !n ? ALIGN(copysz, 16) : copysz;
72 		*seg += stride;
73 		*wqe_sz += stride >> 4;
74 		handle_post_send_edge(sq, seg, *wqe_sz, cur_edge);
75 	}
76 }
77 
78 static int mlx5_wq_overflow(struct mlx5_ib_wq *wq, int nreq,
79 			    struct ib_cq *ib_cq)
80 {
81 	struct mlx5_ib_cq *cq;
82 	unsigned int cur;
83 
84 	cur = wq->head - wq->tail;
85 	if (likely(cur + nreq < wq->max_post))
86 		return 0;
87 
88 	cq = to_mcq(ib_cq);
89 	spin_lock(&cq->lock);
90 	cur = wq->head - wq->tail;
91 	spin_unlock(&cq->lock);
92 
93 	return cur + nreq >= wq->max_post;
94 }
95 
96 static __always_inline void set_raddr_seg(struct mlx5_wqe_raddr_seg *rseg,
97 					  u64 remote_addr, u32 rkey)
98 {
99 	rseg->raddr    = cpu_to_be64(remote_addr);
100 	rseg->rkey     = cpu_to_be32(rkey);
101 	rseg->reserved = 0;
102 }
103 
104 static void set_eth_seg(const struct ib_send_wr *wr, struct mlx5_ib_qp *qp,
105 			void **seg, int *size, void **cur_edge)
106 {
107 	struct mlx5_wqe_eth_seg *eseg = *seg;
108 
109 	memset(eseg, 0, sizeof(struct mlx5_wqe_eth_seg));
110 
111 	if (wr->send_flags & IB_SEND_IP_CSUM)
112 		eseg->cs_flags = MLX5_ETH_WQE_L3_CSUM |
113 				 MLX5_ETH_WQE_L4_CSUM;
114 
115 	if (wr->opcode == IB_WR_LSO) {
116 		struct ib_ud_wr *ud_wr = container_of(wr, struct ib_ud_wr, wr);
117 		size_t left, copysz;
118 		void *pdata = ud_wr->header;
119 		size_t stride;
120 
121 		left = ud_wr->hlen;
122 		eseg->mss = cpu_to_be16(ud_wr->mss);
123 		eseg->inline_hdr.sz = cpu_to_be16(left);
124 
125 		/* memcpy_send_wqe should get a 16B align address. Hence, we
126 		 * first copy up to the current edge and then, if needed,
127 		 * continue to memcpy_send_wqe.
128 		 */
129 		copysz = min_t(u64, *cur_edge - (void *)eseg->inline_hdr.start,
130 			       left);
131 		memcpy(eseg->inline_hdr.start, pdata, copysz);
132 		stride = ALIGN(sizeof(struct mlx5_wqe_eth_seg) -
133 			       sizeof(eseg->inline_hdr.start) + copysz, 16);
134 		*size += stride / 16;
135 		*seg += stride;
136 
137 		if (copysz < left) {
138 			handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
139 			left -= copysz;
140 			pdata += copysz;
141 			memcpy_send_wqe(&qp->sq, cur_edge, seg, size, pdata,
142 					left);
143 		}
144 
145 		return;
146 	}
147 
148 	*seg += sizeof(struct mlx5_wqe_eth_seg);
149 	*size += sizeof(struct mlx5_wqe_eth_seg) / 16;
150 }
151 
152 static void set_datagram_seg(struct mlx5_wqe_datagram_seg *dseg,
153 			     const struct ib_send_wr *wr)
154 {
155 	memcpy(&dseg->av, &to_mah(ud_wr(wr)->ah)->av, sizeof(struct mlx5_av));
156 	dseg->av.dqp_dct =
157 		cpu_to_be32(ud_wr(wr)->remote_qpn | MLX5_EXTENDED_UD_AV);
158 	dseg->av.key.qkey.qkey = cpu_to_be32(ud_wr(wr)->remote_qkey);
159 }
160 
161 static void set_data_ptr_seg(struct mlx5_wqe_data_seg *dseg, struct ib_sge *sg)
162 {
163 	dseg->byte_count = cpu_to_be32(sg->length);
164 	dseg->lkey       = cpu_to_be32(sg->lkey);
165 	dseg->addr       = cpu_to_be64(sg->addr);
166 }
167 
168 static u64 get_xlt_octo(u64 bytes)
169 {
170 	return ALIGN(bytes, MLX5_IB_UMR_XLT_ALIGNMENT) /
171 	       MLX5_IB_UMR_OCTOWORD;
172 }
173 
174 static __be64 frwr_mkey_mask(bool atomic)
175 {
176 	u64 result;
177 
178 	result = MLX5_MKEY_MASK_LEN		|
179 		MLX5_MKEY_MASK_PAGE_SIZE	|
180 		MLX5_MKEY_MASK_START_ADDR	|
181 		MLX5_MKEY_MASK_EN_RINVAL	|
182 		MLX5_MKEY_MASK_KEY		|
183 		MLX5_MKEY_MASK_LR		|
184 		MLX5_MKEY_MASK_LW		|
185 		MLX5_MKEY_MASK_RR		|
186 		MLX5_MKEY_MASK_RW		|
187 		MLX5_MKEY_MASK_SMALL_FENCE	|
188 		MLX5_MKEY_MASK_FREE;
189 
190 	if (atomic)
191 		result |= MLX5_MKEY_MASK_A;
192 
193 	return cpu_to_be64(result);
194 }
195 
196 static __be64 sig_mkey_mask(void)
197 {
198 	u64 result;
199 
200 	result = MLX5_MKEY_MASK_LEN		|
201 		MLX5_MKEY_MASK_PAGE_SIZE	|
202 		MLX5_MKEY_MASK_START_ADDR	|
203 		MLX5_MKEY_MASK_EN_SIGERR	|
204 		MLX5_MKEY_MASK_EN_RINVAL	|
205 		MLX5_MKEY_MASK_KEY		|
206 		MLX5_MKEY_MASK_LR		|
207 		MLX5_MKEY_MASK_LW		|
208 		MLX5_MKEY_MASK_RR		|
209 		MLX5_MKEY_MASK_RW		|
210 		MLX5_MKEY_MASK_SMALL_FENCE	|
211 		MLX5_MKEY_MASK_FREE		|
212 		MLX5_MKEY_MASK_BSF_EN;
213 
214 	return cpu_to_be64(result);
215 }
216 
217 static void set_reg_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr,
218 			    struct mlx5_ib_mr *mr, u8 flags, bool atomic)
219 {
220 	int size = (mr->ndescs + mr->meta_ndescs) * mr->desc_size;
221 
222 	memset(umr, 0, sizeof(*umr));
223 
224 	umr->flags = flags;
225 	umr->xlt_octowords = cpu_to_be16(get_xlt_octo(size));
226 	umr->mkey_mask = frwr_mkey_mask(atomic);
227 }
228 
229 static void set_linv_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr)
230 {
231 	memset(umr, 0, sizeof(*umr));
232 	umr->mkey_mask = cpu_to_be64(MLX5_MKEY_MASK_FREE);
233 	umr->flags = MLX5_UMR_INLINE;
234 }
235 
236 static __be64 get_umr_enable_mr_mask(void)
237 {
238 	u64 result;
239 
240 	result = MLX5_MKEY_MASK_KEY |
241 		 MLX5_MKEY_MASK_FREE;
242 
243 	return cpu_to_be64(result);
244 }
245 
246 static __be64 get_umr_disable_mr_mask(void)
247 {
248 	u64 result;
249 
250 	result = MLX5_MKEY_MASK_FREE;
251 
252 	return cpu_to_be64(result);
253 }
254 
255 static __be64 get_umr_update_translation_mask(void)
256 {
257 	u64 result;
258 
259 	result = MLX5_MKEY_MASK_LEN |
260 		 MLX5_MKEY_MASK_PAGE_SIZE |
261 		 MLX5_MKEY_MASK_START_ADDR;
262 
263 	return cpu_to_be64(result);
264 }
265 
266 static __be64 get_umr_update_access_mask(int atomic,
267 					 int relaxed_ordering_write,
268 					 int relaxed_ordering_read)
269 {
270 	u64 result;
271 
272 	result = MLX5_MKEY_MASK_LR |
273 		 MLX5_MKEY_MASK_LW |
274 		 MLX5_MKEY_MASK_RR |
275 		 MLX5_MKEY_MASK_RW;
276 
277 	if (atomic)
278 		result |= MLX5_MKEY_MASK_A;
279 
280 	if (relaxed_ordering_write)
281 		result |= MLX5_MKEY_MASK_RELAXED_ORDERING_WRITE;
282 
283 	if (relaxed_ordering_read)
284 		result |= MLX5_MKEY_MASK_RELAXED_ORDERING_READ;
285 
286 	return cpu_to_be64(result);
287 }
288 
289 static __be64 get_umr_update_pd_mask(void)
290 {
291 	u64 result;
292 
293 	result = MLX5_MKEY_MASK_PD;
294 
295 	return cpu_to_be64(result);
296 }
297 
298 static int umr_check_mkey_mask(struct mlx5_ib_dev *dev, u64 mask)
299 {
300 	if (mask & MLX5_MKEY_MASK_PAGE_SIZE &&
301 	    MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled))
302 		return -EPERM;
303 
304 	if (mask & MLX5_MKEY_MASK_A &&
305 	    MLX5_CAP_GEN(dev->mdev, umr_modify_atomic_disabled))
306 		return -EPERM;
307 
308 	if (mask & MLX5_MKEY_MASK_RELAXED_ORDERING_WRITE &&
309 	    !MLX5_CAP_GEN(dev->mdev, relaxed_ordering_write_umr))
310 		return -EPERM;
311 
312 	if (mask & MLX5_MKEY_MASK_RELAXED_ORDERING_READ &&
313 	    !MLX5_CAP_GEN(dev->mdev, relaxed_ordering_read_umr))
314 		return -EPERM;
315 
316 	return 0;
317 }
318 
319 static int set_reg_umr_segment(struct mlx5_ib_dev *dev,
320 			       struct mlx5_wqe_umr_ctrl_seg *umr,
321 			       const struct ib_send_wr *wr)
322 {
323 	const struct mlx5_umr_wr *umrwr = umr_wr(wr);
324 
325 	memset(umr, 0, sizeof(*umr));
326 
327 	if (!umrwr->ignore_free_state) {
328 		if (wr->send_flags & MLX5_IB_SEND_UMR_FAIL_IF_FREE)
329 			 /* fail if free */
330 			umr->flags = MLX5_UMR_CHECK_FREE;
331 		else
332 			/* fail if not free */
333 			umr->flags = MLX5_UMR_CHECK_NOT_FREE;
334 	}
335 
336 	umr->xlt_octowords = cpu_to_be16(get_xlt_octo(umrwr->xlt_size));
337 	if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_XLT) {
338 		u64 offset = get_xlt_octo(umrwr->offset);
339 
340 		umr->xlt_offset = cpu_to_be16(offset & 0xffff);
341 		umr->xlt_offset_47_16 = cpu_to_be32(offset >> 16);
342 		umr->flags |= MLX5_UMR_TRANSLATION_OFFSET_EN;
343 	}
344 	if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_TRANSLATION)
345 		umr->mkey_mask |= get_umr_update_translation_mask();
346 	if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_PD_ACCESS) {
347 		umr->mkey_mask |= get_umr_update_access_mask(
348 			!!(MLX5_CAP_GEN(dev->mdev, atomic)),
349 			!!(MLX5_CAP_GEN(dev->mdev, relaxed_ordering_write_umr)),
350 			!!(MLX5_CAP_GEN(dev->mdev, relaxed_ordering_read_umr)));
351 		umr->mkey_mask |= get_umr_update_pd_mask();
352 	}
353 	if (wr->send_flags & MLX5_IB_SEND_UMR_ENABLE_MR)
354 		umr->mkey_mask |= get_umr_enable_mr_mask();
355 	if (wr->send_flags & MLX5_IB_SEND_UMR_DISABLE_MR)
356 		umr->mkey_mask |= get_umr_disable_mr_mask();
357 
358 	if (!wr->num_sge)
359 		umr->flags |= MLX5_UMR_INLINE;
360 
361 	return umr_check_mkey_mask(dev, be64_to_cpu(umr->mkey_mask));
362 }
363 
364 static u8 get_umr_flags(int acc)
365 {
366 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX5_PERM_ATOMIC       : 0) |
367 	       (acc & IB_ACCESS_REMOTE_WRITE  ? MLX5_PERM_REMOTE_WRITE : 0) |
368 	       (acc & IB_ACCESS_REMOTE_READ   ? MLX5_PERM_REMOTE_READ  : 0) |
369 	       (acc & IB_ACCESS_LOCAL_WRITE   ? MLX5_PERM_LOCAL_WRITE  : 0) |
370 		MLX5_PERM_LOCAL_READ | MLX5_PERM_UMR_EN;
371 }
372 
373 static void set_reg_mkey_seg(struct mlx5_mkey_seg *seg,
374 			     struct mlx5_ib_mr *mr,
375 			     u32 key, int access)
376 {
377 	int ndescs = ALIGN(mr->ndescs + mr->meta_ndescs, 8) >> 1;
378 
379 	memset(seg, 0, sizeof(*seg));
380 
381 	if (mr->access_mode == MLX5_MKC_ACCESS_MODE_MTT)
382 		seg->log2_page_size = ilog2(mr->ibmr.page_size);
383 	else if (mr->access_mode == MLX5_MKC_ACCESS_MODE_KLMS)
384 		/* KLMs take twice the size of MTTs */
385 		ndescs *= 2;
386 
387 	seg->flags = get_umr_flags(access) | mr->access_mode;
388 	seg->qpn_mkey7_0 = cpu_to_be32((key & 0xff) | 0xffffff00);
389 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL);
390 	seg->start_addr = cpu_to_be64(mr->ibmr.iova);
391 	seg->len = cpu_to_be64(mr->ibmr.length);
392 	seg->xlt_oct_size = cpu_to_be32(ndescs);
393 }
394 
395 static void set_linv_mkey_seg(struct mlx5_mkey_seg *seg)
396 {
397 	memset(seg, 0, sizeof(*seg));
398 	seg->status = MLX5_MKEY_STATUS_FREE;
399 }
400 
401 static void set_reg_mkey_segment(struct mlx5_mkey_seg *seg,
402 				 const struct ib_send_wr *wr)
403 {
404 	const struct mlx5_umr_wr *umrwr = umr_wr(wr);
405 
406 	memset(seg, 0, sizeof(*seg));
407 	if (wr->send_flags & MLX5_IB_SEND_UMR_DISABLE_MR)
408 		MLX5_SET(mkc, seg, free, 1);
409 
410 	MLX5_SET(mkc, seg, a,
411 		 !!(umrwr->access_flags & IB_ACCESS_REMOTE_ATOMIC));
412 	MLX5_SET(mkc, seg, rw,
413 		 !!(umrwr->access_flags & IB_ACCESS_REMOTE_WRITE));
414 	MLX5_SET(mkc, seg, rr, !!(umrwr->access_flags & IB_ACCESS_REMOTE_READ));
415 	MLX5_SET(mkc, seg, lw, !!(umrwr->access_flags & IB_ACCESS_LOCAL_WRITE));
416 	MLX5_SET(mkc, seg, lr, 1);
417 	MLX5_SET(mkc, seg, relaxed_ordering_write,
418 		 !!(umrwr->access_flags & IB_ACCESS_RELAXED_ORDERING));
419 	MLX5_SET(mkc, seg, relaxed_ordering_read,
420 		 !!(umrwr->access_flags & IB_ACCESS_RELAXED_ORDERING));
421 
422 	if (umrwr->pd)
423 		MLX5_SET(mkc, seg, pd, to_mpd(umrwr->pd)->pdn);
424 	if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_TRANSLATION &&
425 	    !umrwr->length)
426 		MLX5_SET(mkc, seg, length64, 1);
427 
428 	MLX5_SET64(mkc, seg, start_addr, umrwr->virt_addr);
429 	MLX5_SET64(mkc, seg, len, umrwr->length);
430 	MLX5_SET(mkc, seg, log_page_size, umrwr->page_shift);
431 	MLX5_SET(mkc, seg, qpn, 0xffffff);
432 	MLX5_SET(mkc, seg, mkey_7_0, mlx5_mkey_variant(umrwr->mkey));
433 }
434 
435 static void set_reg_data_seg(struct mlx5_wqe_data_seg *dseg,
436 			     struct mlx5_ib_mr *mr,
437 			     struct mlx5_ib_pd *pd)
438 {
439 	int bcount = mr->desc_size * (mr->ndescs + mr->meta_ndescs);
440 
441 	dseg->addr = cpu_to_be64(mr->desc_map);
442 	dseg->byte_count = cpu_to_be32(ALIGN(bcount, 64));
443 	dseg->lkey = cpu_to_be32(pd->ibpd.local_dma_lkey);
444 }
445 
446 static __be32 send_ieth(const struct ib_send_wr *wr)
447 {
448 	switch (wr->opcode) {
449 	case IB_WR_SEND_WITH_IMM:
450 	case IB_WR_RDMA_WRITE_WITH_IMM:
451 		return wr->ex.imm_data;
452 
453 	case IB_WR_SEND_WITH_INV:
454 		return cpu_to_be32(wr->ex.invalidate_rkey);
455 
456 	default:
457 		return 0;
458 	}
459 }
460 
461 static u8 calc_sig(void *wqe, int size)
462 {
463 	u8 *p = wqe;
464 	u8 res = 0;
465 	int i;
466 
467 	for (i = 0; i < size; i++)
468 		res ^= p[i];
469 
470 	return ~res;
471 }
472 
473 static u8 wq_sig(void *wqe)
474 {
475 	return calc_sig(wqe, (*((u8 *)wqe + 8) & 0x3f) << 4);
476 }
477 
478 static int set_data_inl_seg(struct mlx5_ib_qp *qp, const struct ib_send_wr *wr,
479 			    void **wqe, int *wqe_sz, void **cur_edge)
480 {
481 	struct mlx5_wqe_inline_seg *seg;
482 	size_t offset;
483 	int inl = 0;
484 	int i;
485 
486 	seg = *wqe;
487 	*wqe += sizeof(*seg);
488 	offset = sizeof(*seg);
489 
490 	for (i = 0; i < wr->num_sge; i++) {
491 		size_t len  = wr->sg_list[i].length;
492 		void *addr = (void *)(unsigned long)(wr->sg_list[i].addr);
493 
494 		inl += len;
495 
496 		if (unlikely(inl > qp->max_inline_data))
497 			return -ENOMEM;
498 
499 		while (likely(len)) {
500 			size_t leftlen;
501 			size_t copysz;
502 
503 			handle_post_send_edge(&qp->sq, wqe,
504 					      *wqe_sz + (offset >> 4),
505 					      cur_edge);
506 
507 			leftlen = *cur_edge - *wqe;
508 			copysz = min_t(size_t, leftlen, len);
509 
510 			memcpy(*wqe, addr, copysz);
511 			len -= copysz;
512 			addr += copysz;
513 			*wqe += copysz;
514 			offset += copysz;
515 		}
516 	}
517 
518 	seg->byte_count = cpu_to_be32(inl | MLX5_INLINE_SEG);
519 
520 	*wqe_sz +=  ALIGN(inl + sizeof(seg->byte_count), 16) / 16;
521 
522 	return 0;
523 }
524 
525 static u16 prot_field_size(enum ib_signature_type type)
526 {
527 	switch (type) {
528 	case IB_SIG_TYPE_T10_DIF:
529 		return MLX5_DIF_SIZE;
530 	default:
531 		return 0;
532 	}
533 }
534 
535 static u8 bs_selector(int block_size)
536 {
537 	switch (block_size) {
538 	case 512:	    return 0x1;
539 	case 520:	    return 0x2;
540 	case 4096:	    return 0x3;
541 	case 4160:	    return 0x4;
542 	case 1073741824:    return 0x5;
543 	default:	    return 0;
544 	}
545 }
546 
547 static void mlx5_fill_inl_bsf(struct ib_sig_domain *domain,
548 			      struct mlx5_bsf_inl *inl)
549 {
550 	/* Valid inline section and allow BSF refresh */
551 	inl->vld_refresh = cpu_to_be16(MLX5_BSF_INL_VALID |
552 				       MLX5_BSF_REFRESH_DIF);
553 	inl->dif_apptag = cpu_to_be16(domain->sig.dif.app_tag);
554 	inl->dif_reftag = cpu_to_be32(domain->sig.dif.ref_tag);
555 	/* repeating block */
556 	inl->rp_inv_seed = MLX5_BSF_REPEAT_BLOCK;
557 	inl->sig_type = domain->sig.dif.bg_type == IB_T10DIF_CRC ?
558 			MLX5_DIF_CRC : MLX5_DIF_IPCS;
559 
560 	if (domain->sig.dif.ref_remap)
561 		inl->dif_inc_ref_guard_check |= MLX5_BSF_INC_REFTAG;
562 
563 	if (domain->sig.dif.app_escape) {
564 		if (domain->sig.dif.ref_escape)
565 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPREF_ESCAPE;
566 		else
567 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPTAG_ESCAPE;
568 	}
569 
570 	inl->dif_app_bitmask_check =
571 		cpu_to_be16(domain->sig.dif.apptag_check_mask);
572 }
573 
574 static int mlx5_set_bsf(struct ib_mr *sig_mr,
575 			struct ib_sig_attrs *sig_attrs,
576 			struct mlx5_bsf *bsf, u32 data_size)
577 {
578 	struct mlx5_core_sig_ctx *msig = to_mmr(sig_mr)->sig;
579 	struct mlx5_bsf_basic *basic = &bsf->basic;
580 	struct ib_sig_domain *mem = &sig_attrs->mem;
581 	struct ib_sig_domain *wire = &sig_attrs->wire;
582 
583 	memset(bsf, 0, sizeof(*bsf));
584 
585 	/* Basic + Extended + Inline */
586 	basic->bsf_size_sbs = 1 << 7;
587 	/* Input domain check byte mask */
588 	basic->check_byte_mask = sig_attrs->check_mask;
589 	basic->raw_data_size = cpu_to_be32(data_size);
590 
591 	/* Memory domain */
592 	switch (sig_attrs->mem.sig_type) {
593 	case IB_SIG_TYPE_NONE:
594 		break;
595 	case IB_SIG_TYPE_T10_DIF:
596 		basic->mem.bs_selector = bs_selector(mem->sig.dif.pi_interval);
597 		basic->m_bfs_psv = cpu_to_be32(msig->psv_memory.psv_idx);
598 		mlx5_fill_inl_bsf(mem, &bsf->m_inl);
599 		break;
600 	default:
601 		return -EINVAL;
602 	}
603 
604 	/* Wire domain */
605 	switch (sig_attrs->wire.sig_type) {
606 	case IB_SIG_TYPE_NONE:
607 		break;
608 	case IB_SIG_TYPE_T10_DIF:
609 		if (mem->sig.dif.pi_interval == wire->sig.dif.pi_interval &&
610 		    mem->sig_type == wire->sig_type) {
611 			/* Same block structure */
612 			basic->bsf_size_sbs |= 1 << 4;
613 			if (mem->sig.dif.bg_type == wire->sig.dif.bg_type)
614 				basic->wire.copy_byte_mask |= MLX5_CPY_GRD_MASK;
615 			if (mem->sig.dif.app_tag == wire->sig.dif.app_tag)
616 				basic->wire.copy_byte_mask |= MLX5_CPY_APP_MASK;
617 			if (mem->sig.dif.ref_tag == wire->sig.dif.ref_tag)
618 				basic->wire.copy_byte_mask |= MLX5_CPY_REF_MASK;
619 		} else
620 			basic->wire.bs_selector =
621 				bs_selector(wire->sig.dif.pi_interval);
622 
623 		basic->w_bfs_psv = cpu_to_be32(msig->psv_wire.psv_idx);
624 		mlx5_fill_inl_bsf(wire, &bsf->w_inl);
625 		break;
626 	default:
627 		return -EINVAL;
628 	}
629 
630 	return 0;
631 }
632 
633 
634 static int set_sig_data_segment(const struct ib_send_wr *send_wr,
635 				struct ib_mr *sig_mr,
636 				struct ib_sig_attrs *sig_attrs,
637 				struct mlx5_ib_qp *qp, void **seg, int *size,
638 				void **cur_edge)
639 {
640 	struct mlx5_bsf *bsf;
641 	u32 data_len;
642 	u32 data_key;
643 	u64 data_va;
644 	u32 prot_len = 0;
645 	u32 prot_key = 0;
646 	u64 prot_va = 0;
647 	bool prot = false;
648 	int ret;
649 	int wqe_size;
650 	struct mlx5_ib_mr *mr = to_mmr(sig_mr);
651 	struct mlx5_ib_mr *pi_mr = mr->pi_mr;
652 
653 	data_len = pi_mr->data_length;
654 	data_key = pi_mr->ibmr.lkey;
655 	data_va = pi_mr->data_iova;
656 	if (pi_mr->meta_ndescs) {
657 		prot_len = pi_mr->meta_length;
658 		prot_key = pi_mr->ibmr.lkey;
659 		prot_va = pi_mr->pi_iova;
660 		prot = true;
661 	}
662 
663 	if (!prot || (data_key == prot_key && data_va == prot_va &&
664 		      data_len == prot_len)) {
665 		/**
666 		 * Source domain doesn't contain signature information
667 		 * or data and protection are interleaved in memory.
668 		 * So need construct:
669 		 *                  ------------------
670 		 *                 |     data_klm     |
671 		 *                  ------------------
672 		 *                 |       BSF        |
673 		 *                  ------------------
674 		 **/
675 		struct mlx5_klm *data_klm = *seg;
676 
677 		data_klm->bcount = cpu_to_be32(data_len);
678 		data_klm->key = cpu_to_be32(data_key);
679 		data_klm->va = cpu_to_be64(data_va);
680 		wqe_size = ALIGN(sizeof(*data_klm), 64);
681 	} else {
682 		/**
683 		 * Source domain contains signature information
684 		 * So need construct a strided block format:
685 		 *               ---------------------------
686 		 *              |     stride_block_ctrl     |
687 		 *               ---------------------------
688 		 *              |          data_klm         |
689 		 *               ---------------------------
690 		 *              |          prot_klm         |
691 		 *               ---------------------------
692 		 *              |             BSF           |
693 		 *               ---------------------------
694 		 **/
695 		struct mlx5_stride_block_ctrl_seg *sblock_ctrl;
696 		struct mlx5_stride_block_entry *data_sentry;
697 		struct mlx5_stride_block_entry *prot_sentry;
698 		u16 block_size = sig_attrs->mem.sig.dif.pi_interval;
699 		int prot_size;
700 
701 		sblock_ctrl = *seg;
702 		data_sentry = (void *)sblock_ctrl + sizeof(*sblock_ctrl);
703 		prot_sentry = (void *)data_sentry + sizeof(*data_sentry);
704 
705 		prot_size = prot_field_size(sig_attrs->mem.sig_type);
706 		if (!prot_size) {
707 			pr_err("Bad block size given: %u\n", block_size);
708 			return -EINVAL;
709 		}
710 		sblock_ctrl->bcount_per_cycle = cpu_to_be32(block_size +
711 							    prot_size);
712 		sblock_ctrl->op = cpu_to_be32(MLX5_STRIDE_BLOCK_OP);
713 		sblock_ctrl->repeat_count = cpu_to_be32(data_len / block_size);
714 		sblock_ctrl->num_entries = cpu_to_be16(2);
715 
716 		data_sentry->bcount = cpu_to_be16(block_size);
717 		data_sentry->key = cpu_to_be32(data_key);
718 		data_sentry->va = cpu_to_be64(data_va);
719 		data_sentry->stride = cpu_to_be16(block_size);
720 
721 		prot_sentry->bcount = cpu_to_be16(prot_size);
722 		prot_sentry->key = cpu_to_be32(prot_key);
723 		prot_sentry->va = cpu_to_be64(prot_va);
724 		prot_sentry->stride = cpu_to_be16(prot_size);
725 
726 		wqe_size = ALIGN(sizeof(*sblock_ctrl) + sizeof(*data_sentry) +
727 				 sizeof(*prot_sentry), 64);
728 	}
729 
730 	*seg += wqe_size;
731 	*size += wqe_size / 16;
732 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
733 
734 	bsf = *seg;
735 	ret = mlx5_set_bsf(sig_mr, sig_attrs, bsf, data_len);
736 	if (ret)
737 		return -EINVAL;
738 
739 	*seg += sizeof(*bsf);
740 	*size += sizeof(*bsf) / 16;
741 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
742 
743 	return 0;
744 }
745 
746 static void set_sig_mkey_segment(struct mlx5_mkey_seg *seg,
747 				 struct ib_mr *sig_mr, int access_flags,
748 				 u32 size, u32 length, u32 pdn)
749 {
750 	u32 sig_key = sig_mr->rkey;
751 	u8 sigerr = to_mmr(sig_mr)->sig->sigerr_count & 1;
752 
753 	memset(seg, 0, sizeof(*seg));
754 
755 	seg->flags = get_umr_flags(access_flags) | MLX5_MKC_ACCESS_MODE_KLMS;
756 	seg->qpn_mkey7_0 = cpu_to_be32((sig_key & 0xff) | 0xffffff00);
757 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL | sigerr << 26 |
758 				    MLX5_MKEY_BSF_EN | pdn);
759 	seg->len = cpu_to_be64(length);
760 	seg->xlt_oct_size = cpu_to_be32(get_xlt_octo(size));
761 	seg->bsfs_octo_size = cpu_to_be32(MLX5_MKEY_BSF_OCTO_SIZE);
762 }
763 
764 static void set_sig_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
765 				u32 size)
766 {
767 	memset(umr, 0, sizeof(*umr));
768 
769 	umr->flags = MLX5_FLAGS_INLINE | MLX5_FLAGS_CHECK_FREE;
770 	umr->xlt_octowords = cpu_to_be16(get_xlt_octo(size));
771 	umr->bsf_octowords = cpu_to_be16(MLX5_MKEY_BSF_OCTO_SIZE);
772 	umr->mkey_mask = sig_mkey_mask();
773 }
774 
775 static int set_pi_umr_wr(const struct ib_send_wr *send_wr,
776 			 struct mlx5_ib_qp *qp, void **seg, int *size,
777 			 void **cur_edge)
778 {
779 	const struct ib_reg_wr *wr = reg_wr(send_wr);
780 	struct mlx5_ib_mr *sig_mr = to_mmr(wr->mr);
781 	struct mlx5_ib_mr *pi_mr = sig_mr->pi_mr;
782 	struct ib_sig_attrs *sig_attrs = sig_mr->ibmr.sig_attrs;
783 	u32 pdn = to_mpd(qp->ibqp.pd)->pdn;
784 	u32 xlt_size;
785 	int region_len, ret;
786 
787 	if (unlikely(send_wr->num_sge != 0) ||
788 	    unlikely(wr->access & IB_ACCESS_REMOTE_ATOMIC) ||
789 	    unlikely(!sig_mr->sig) || unlikely(!qp->ibqp.integrity_en) ||
790 	    unlikely(!sig_mr->sig->sig_status_checked))
791 		return -EINVAL;
792 
793 	/* length of the protected region, data + protection */
794 	region_len = pi_mr->ibmr.length;
795 
796 	/**
797 	 * KLM octoword size - if protection was provided
798 	 * then we use strided block format (3 octowords),
799 	 * else we use single KLM (1 octoword)
800 	 **/
801 	if (sig_attrs->mem.sig_type != IB_SIG_TYPE_NONE)
802 		xlt_size = 0x30;
803 	else
804 		xlt_size = sizeof(struct mlx5_klm);
805 
806 	set_sig_umr_segment(*seg, xlt_size);
807 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
808 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
809 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
810 
811 	set_sig_mkey_segment(*seg, wr->mr, wr->access, xlt_size, region_len,
812 			     pdn);
813 	*seg += sizeof(struct mlx5_mkey_seg);
814 	*size += sizeof(struct mlx5_mkey_seg) / 16;
815 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
816 
817 	ret = set_sig_data_segment(send_wr, wr->mr, sig_attrs, qp, seg, size,
818 				   cur_edge);
819 	if (ret)
820 		return ret;
821 
822 	sig_mr->sig->sig_status_checked = false;
823 	return 0;
824 }
825 
826 static int set_psv_wr(struct ib_sig_domain *domain,
827 		      u32 psv_idx, void **seg, int *size)
828 {
829 	struct mlx5_seg_set_psv *psv_seg = *seg;
830 
831 	memset(psv_seg, 0, sizeof(*psv_seg));
832 	psv_seg->psv_num = cpu_to_be32(psv_idx);
833 	switch (domain->sig_type) {
834 	case IB_SIG_TYPE_NONE:
835 		break;
836 	case IB_SIG_TYPE_T10_DIF:
837 		psv_seg->transient_sig = cpu_to_be32(domain->sig.dif.bg << 16 |
838 						     domain->sig.dif.app_tag);
839 		psv_seg->ref_tag = cpu_to_be32(domain->sig.dif.ref_tag);
840 		break;
841 	default:
842 		pr_err("Bad signature type (%d) is given.\n",
843 		       domain->sig_type);
844 		return -EINVAL;
845 	}
846 
847 	*seg += sizeof(*psv_seg);
848 	*size += sizeof(*psv_seg) / 16;
849 
850 	return 0;
851 }
852 
853 static int set_reg_wr(struct mlx5_ib_qp *qp,
854 		      const struct ib_reg_wr *wr,
855 		      void **seg, int *size, void **cur_edge,
856 		      bool check_not_free)
857 {
858 	struct mlx5_ib_mr *mr = to_mmr(wr->mr);
859 	struct mlx5_ib_pd *pd = to_mpd(qp->ibqp.pd);
860 	struct mlx5_ib_dev *dev = to_mdev(pd->ibpd.device);
861 	int mr_list_size = (mr->ndescs + mr->meta_ndescs) * mr->desc_size;
862 	bool umr_inline = mr_list_size <= MLX5_IB_SQ_UMR_INLINE_THRESHOLD;
863 	bool atomic = wr->access & IB_ACCESS_REMOTE_ATOMIC;
864 	u8 flags = 0;
865 
866 	if (!mlx5_ib_can_use_umr(dev, atomic, wr->access)) {
867 		mlx5_ib_warn(to_mdev(qp->ibqp.device),
868 			     "Fast update of %s for MR is disabled\n",
869 			     (MLX5_CAP_GEN(dev->mdev,
870 					   umr_modify_entity_size_disabled)) ?
871 				     "entity size" :
872 				     "atomic access");
873 		return -EINVAL;
874 	}
875 
876 	if (unlikely(wr->wr.send_flags & IB_SEND_INLINE)) {
877 		mlx5_ib_warn(to_mdev(qp->ibqp.device),
878 			     "Invalid IB_SEND_INLINE send flag\n");
879 		return -EINVAL;
880 	}
881 
882 	if (check_not_free)
883 		flags |= MLX5_UMR_CHECK_NOT_FREE;
884 	if (umr_inline)
885 		flags |= MLX5_UMR_INLINE;
886 
887 	set_reg_umr_seg(*seg, mr, flags, atomic);
888 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
889 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
890 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
891 
892 	set_reg_mkey_seg(*seg, mr, wr->key, wr->access);
893 	*seg += sizeof(struct mlx5_mkey_seg);
894 	*size += sizeof(struct mlx5_mkey_seg) / 16;
895 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
896 
897 	if (umr_inline) {
898 		memcpy_send_wqe(&qp->sq, cur_edge, seg, size, mr->descs,
899 				mr_list_size);
900 		*size = ALIGN(*size, MLX5_SEND_WQE_BB >> 4);
901 	} else {
902 		set_reg_data_seg(*seg, mr, pd);
903 		*seg += sizeof(struct mlx5_wqe_data_seg);
904 		*size += (sizeof(struct mlx5_wqe_data_seg) / 16);
905 	}
906 	return 0;
907 }
908 
909 static void set_linv_wr(struct mlx5_ib_qp *qp, void **seg, int *size,
910 			void **cur_edge)
911 {
912 	set_linv_umr_seg(*seg);
913 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
914 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
915 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
916 	set_linv_mkey_seg(*seg);
917 	*seg += sizeof(struct mlx5_mkey_seg);
918 	*size += sizeof(struct mlx5_mkey_seg) / 16;
919 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
920 }
921 
922 static void dump_wqe(struct mlx5_ib_qp *qp, u32 idx, int size_16)
923 {
924 	__be32 *p = NULL;
925 	int i, j;
926 
927 	pr_debug("dump WQE index %u:\n", idx);
928 	for (i = 0, j = 0; i < size_16 * 4; i += 4, j += 4) {
929 		if ((i & 0xf) == 0) {
930 			p = mlx5_frag_buf_get_wqe(&qp->sq.fbc, idx);
931 			pr_debug("WQBB at %p:\n", (void *)p);
932 			j = 0;
933 			idx = (idx + 1) & (qp->sq.wqe_cnt - 1);
934 		}
935 		pr_debug("%08x %08x %08x %08x\n", be32_to_cpu(p[j]),
936 			 be32_to_cpu(p[j + 1]), be32_to_cpu(p[j + 2]),
937 			 be32_to_cpu(p[j + 3]));
938 	}
939 }
940 
941 static int __begin_wqe(struct mlx5_ib_qp *qp, void **seg,
942 		       struct mlx5_wqe_ctrl_seg **ctrl,
943 		       const struct ib_send_wr *wr, unsigned int *idx,
944 		       int *size, void **cur_edge, int nreq,
945 		       bool send_signaled, bool solicited)
946 {
947 	if (unlikely(mlx5_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)))
948 		return -ENOMEM;
949 
950 	*idx = qp->sq.cur_post & (qp->sq.wqe_cnt - 1);
951 	*seg = mlx5_frag_buf_get_wqe(&qp->sq.fbc, *idx);
952 	*ctrl = *seg;
953 	*(uint32_t *)(*seg + 8) = 0;
954 	(*ctrl)->imm = send_ieth(wr);
955 	(*ctrl)->fm_ce_se = qp->sq_signal_bits |
956 		(send_signaled ? MLX5_WQE_CTRL_CQ_UPDATE : 0) |
957 		(solicited ? MLX5_WQE_CTRL_SOLICITED : 0);
958 
959 	*seg += sizeof(**ctrl);
960 	*size = sizeof(**ctrl) / 16;
961 	*cur_edge = qp->sq.cur_edge;
962 
963 	return 0;
964 }
965 
966 static int begin_wqe(struct mlx5_ib_qp *qp, void **seg,
967 		     struct mlx5_wqe_ctrl_seg **ctrl,
968 		     const struct ib_send_wr *wr, unsigned int *idx, int *size,
969 		     void **cur_edge, int nreq)
970 {
971 	return __begin_wqe(qp, seg, ctrl, wr, idx, size, cur_edge, nreq,
972 			   wr->send_flags & IB_SEND_SIGNALED,
973 			   wr->send_flags & IB_SEND_SOLICITED);
974 }
975 
976 static void finish_wqe(struct mlx5_ib_qp *qp,
977 		       struct mlx5_wqe_ctrl_seg *ctrl,
978 		       void *seg, u8 size, void *cur_edge,
979 		       unsigned int idx, u64 wr_id, int nreq, u8 fence,
980 		       u32 mlx5_opcode)
981 {
982 	u8 opmod = 0;
983 
984 	ctrl->opmod_idx_opcode = cpu_to_be32(((u32)(qp->sq.cur_post) << 8) |
985 					     mlx5_opcode | ((u32)opmod << 24));
986 	ctrl->qpn_ds = cpu_to_be32(size | (qp->trans_qp.base.mqp.qpn << 8));
987 	ctrl->fm_ce_se |= fence;
988 	if (unlikely(qp->flags_en & MLX5_QP_FLAG_SIGNATURE))
989 		ctrl->signature = wq_sig(ctrl);
990 
991 	qp->sq.wrid[idx] = wr_id;
992 	qp->sq.w_list[idx].opcode = mlx5_opcode;
993 	qp->sq.wqe_head[idx] = qp->sq.head + nreq;
994 	qp->sq.cur_post += DIV_ROUND_UP(size * 16, MLX5_SEND_WQE_BB);
995 	qp->sq.w_list[idx].next = qp->sq.cur_post;
996 
997 	/* We save the edge which was possibly updated during the WQE
998 	 * construction, into SQ's cache.
999 	 */
1000 	seg = PTR_ALIGN(seg, MLX5_SEND_WQE_BB);
1001 	qp->sq.cur_edge = (unlikely(seg == cur_edge)) ?
1002 			  get_sq_edge(&qp->sq, qp->sq.cur_post &
1003 				      (qp->sq.wqe_cnt - 1)) :
1004 			  cur_edge;
1005 }
1006 
1007 static void handle_rdma_op(const struct ib_send_wr *wr, void **seg, int *size)
1008 {
1009 	set_raddr_seg(*seg, rdma_wr(wr)->remote_addr, rdma_wr(wr)->rkey);
1010 	*seg += sizeof(struct mlx5_wqe_raddr_seg);
1011 	*size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
1012 }
1013 
1014 static void handle_local_inv(struct mlx5_ib_qp *qp, const struct ib_send_wr *wr,
1015 			     struct mlx5_wqe_ctrl_seg **ctrl, void **seg,
1016 			     int *size, void **cur_edge, unsigned int idx)
1017 {
1018 	qp->sq.wr_data[idx] = IB_WR_LOCAL_INV;
1019 	(*ctrl)->imm = cpu_to_be32(wr->ex.invalidate_rkey);
1020 	set_linv_wr(qp, seg, size, cur_edge);
1021 }
1022 
1023 static int handle_reg_mr(struct mlx5_ib_qp *qp, const struct ib_send_wr *wr,
1024 			 struct mlx5_wqe_ctrl_seg **ctrl, void **seg, int *size,
1025 			 void **cur_edge, unsigned int idx)
1026 {
1027 	qp->sq.wr_data[idx] = IB_WR_REG_MR;
1028 	(*ctrl)->imm = cpu_to_be32(reg_wr(wr)->key);
1029 	return set_reg_wr(qp, reg_wr(wr), seg, size, cur_edge, true);
1030 }
1031 
1032 static int handle_psv(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1033 		      const struct ib_send_wr *wr,
1034 		      struct mlx5_wqe_ctrl_seg **ctrl, void **seg, int *size,
1035 		      void **cur_edge, unsigned int *idx, int nreq,
1036 		      struct ib_sig_domain *domain, u32 psv_index,
1037 		      u8 next_fence)
1038 {
1039 	int err;
1040 
1041 	/*
1042 	 * SET_PSV WQEs are not signaled and solicited on error.
1043 	 */
1044 	err = __begin_wqe(qp, seg, ctrl, wr, idx, size, cur_edge, nreq,
1045 			  false, true);
1046 	if (unlikely(err)) {
1047 		mlx5_ib_warn(dev, "\n");
1048 		err = -ENOMEM;
1049 		goto out;
1050 	}
1051 	err = set_psv_wr(domain, psv_index, seg, size);
1052 	if (unlikely(err)) {
1053 		mlx5_ib_warn(dev, "\n");
1054 		goto out;
1055 	}
1056 	finish_wqe(qp, *ctrl, *seg, *size, *cur_edge, *idx, wr->wr_id, nreq,
1057 		   next_fence, MLX5_OPCODE_SET_PSV);
1058 
1059 out:
1060 	return err;
1061 }
1062 
1063 static int handle_reg_mr_integrity(struct mlx5_ib_dev *dev,
1064 				   struct mlx5_ib_qp *qp,
1065 				   const struct ib_send_wr *wr,
1066 				   struct mlx5_wqe_ctrl_seg **ctrl, void **seg,
1067 				   int *size, void **cur_edge,
1068 				   unsigned int *idx, int nreq, u8 fence,
1069 				   u8 next_fence)
1070 {
1071 	struct mlx5_ib_mr *mr;
1072 	struct mlx5_ib_mr *pi_mr;
1073 	struct mlx5_ib_mr pa_pi_mr;
1074 	struct ib_sig_attrs *sig_attrs;
1075 	struct ib_reg_wr reg_pi_wr;
1076 	int err;
1077 
1078 	qp->sq.wr_data[*idx] = IB_WR_REG_MR_INTEGRITY;
1079 
1080 	mr = to_mmr(reg_wr(wr)->mr);
1081 	pi_mr = mr->pi_mr;
1082 
1083 	if (pi_mr) {
1084 		memset(&reg_pi_wr, 0,
1085 		       sizeof(struct ib_reg_wr));
1086 
1087 		reg_pi_wr.mr = &pi_mr->ibmr;
1088 		reg_pi_wr.access = reg_wr(wr)->access;
1089 		reg_pi_wr.key = pi_mr->ibmr.rkey;
1090 
1091 		(*ctrl)->imm = cpu_to_be32(reg_pi_wr.key);
1092 		/* UMR for data + prot registration */
1093 		err = set_reg_wr(qp, &reg_pi_wr, seg, size, cur_edge, false);
1094 		if (unlikely(err))
1095 			goto out;
1096 
1097 		finish_wqe(qp, *ctrl, *seg, *size, *cur_edge, *idx, wr->wr_id,
1098 			   nreq, fence, MLX5_OPCODE_UMR);
1099 
1100 		err = begin_wqe(qp, seg, ctrl, wr, idx, size, cur_edge, nreq);
1101 		if (unlikely(err)) {
1102 			mlx5_ib_warn(dev, "\n");
1103 			err = -ENOMEM;
1104 			goto out;
1105 		}
1106 	} else {
1107 		memset(&pa_pi_mr, 0, sizeof(struct mlx5_ib_mr));
1108 		/* No UMR, use local_dma_lkey */
1109 		pa_pi_mr.ibmr.lkey = mr->ibmr.pd->local_dma_lkey;
1110 		pa_pi_mr.ndescs = mr->ndescs;
1111 		pa_pi_mr.data_length = mr->data_length;
1112 		pa_pi_mr.data_iova = mr->data_iova;
1113 		if (mr->meta_ndescs) {
1114 			pa_pi_mr.meta_ndescs = mr->meta_ndescs;
1115 			pa_pi_mr.meta_length = mr->meta_length;
1116 			pa_pi_mr.pi_iova = mr->pi_iova;
1117 		}
1118 
1119 		pa_pi_mr.ibmr.length = mr->ibmr.length;
1120 		mr->pi_mr = &pa_pi_mr;
1121 	}
1122 	(*ctrl)->imm = cpu_to_be32(mr->ibmr.rkey);
1123 	/* UMR for sig MR */
1124 	err = set_pi_umr_wr(wr, qp, seg, size, cur_edge);
1125 	if (unlikely(err)) {
1126 		mlx5_ib_warn(dev, "\n");
1127 		goto out;
1128 	}
1129 	finish_wqe(qp, *ctrl, *seg, *size, *cur_edge, *idx, wr->wr_id, nreq,
1130 		   fence, MLX5_OPCODE_UMR);
1131 
1132 	sig_attrs = mr->ibmr.sig_attrs;
1133 	err = handle_psv(dev, qp, wr, ctrl, seg, size, cur_edge, idx, nreq,
1134 			 &sig_attrs->mem, mr->sig->psv_memory.psv_idx,
1135 			 next_fence);
1136 	if (unlikely(err))
1137 		goto out;
1138 
1139 	err = handle_psv(dev, qp, wr, ctrl, seg, size, cur_edge, idx, nreq,
1140 			 &sig_attrs->wire, mr->sig->psv_wire.psv_idx,
1141 			 next_fence);
1142 	if (unlikely(err))
1143 		goto out;
1144 
1145 	qp->next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
1146 
1147 out:
1148 	return err;
1149 }
1150 
1151 static int handle_qpt_rc(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1152 			 const struct ib_send_wr *wr,
1153 			 struct mlx5_wqe_ctrl_seg **ctrl, void **seg, int *size,
1154 			 void **cur_edge, unsigned int *idx, int nreq, u8 fence,
1155 			 u8 next_fence, int *num_sge)
1156 {
1157 	int err = 0;
1158 
1159 	switch (wr->opcode) {
1160 	case IB_WR_RDMA_READ:
1161 	case IB_WR_RDMA_WRITE:
1162 	case IB_WR_RDMA_WRITE_WITH_IMM:
1163 		handle_rdma_op(wr, seg, size);
1164 		break;
1165 
1166 	case IB_WR_ATOMIC_CMP_AND_SWP:
1167 	case IB_WR_ATOMIC_FETCH_AND_ADD:
1168 	case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
1169 		mlx5_ib_warn(dev, "Atomic operations are not supported yet\n");
1170 		err = -EOPNOTSUPP;
1171 		goto out;
1172 
1173 	case IB_WR_LOCAL_INV:
1174 		handle_local_inv(qp, wr, ctrl, seg, size, cur_edge, *idx);
1175 		*num_sge = 0;
1176 		break;
1177 
1178 	case IB_WR_REG_MR:
1179 		err = handle_reg_mr(qp, wr, ctrl, seg, size, cur_edge, *idx);
1180 		if (unlikely(err))
1181 			goto out;
1182 		*num_sge = 0;
1183 		break;
1184 
1185 	case IB_WR_REG_MR_INTEGRITY:
1186 		err = handle_reg_mr_integrity(dev, qp, wr, ctrl, seg, size,
1187 					      cur_edge, idx, nreq, fence,
1188 					      next_fence);
1189 		if (unlikely(err))
1190 			goto out;
1191 		*num_sge = 0;
1192 		break;
1193 
1194 	default:
1195 		break;
1196 	}
1197 
1198 out:
1199 	return err;
1200 }
1201 
1202 static void handle_qpt_uc(const struct ib_send_wr *wr, void **seg, int *size)
1203 {
1204 	switch (wr->opcode) {
1205 	case IB_WR_RDMA_WRITE:
1206 	case IB_WR_RDMA_WRITE_WITH_IMM:
1207 		handle_rdma_op(wr, seg, size);
1208 		break;
1209 	default:
1210 		break;
1211 	}
1212 }
1213 
1214 static void handle_qpt_hw_gsi(struct mlx5_ib_qp *qp,
1215 			      const struct ib_send_wr *wr, void **seg,
1216 			      int *size, void **cur_edge)
1217 {
1218 	set_datagram_seg(*seg, wr);
1219 	*seg += sizeof(struct mlx5_wqe_datagram_seg);
1220 	*size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
1221 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
1222 }
1223 
1224 static void handle_qpt_ud(struct mlx5_ib_qp *qp, const struct ib_send_wr *wr,
1225 			  void **seg, int *size, void **cur_edge)
1226 {
1227 	set_datagram_seg(*seg, wr);
1228 	*seg += sizeof(struct mlx5_wqe_datagram_seg);
1229 	*size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
1230 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
1231 
1232 	/* handle qp that supports ud offload */
1233 	if (qp->flags & IB_QP_CREATE_IPOIB_UD_LSO) {
1234 		struct mlx5_wqe_eth_pad *pad;
1235 
1236 		pad = *seg;
1237 		memset(pad, 0, sizeof(struct mlx5_wqe_eth_pad));
1238 		*seg += sizeof(struct mlx5_wqe_eth_pad);
1239 		*size += sizeof(struct mlx5_wqe_eth_pad) / 16;
1240 		set_eth_seg(wr, qp, seg, size, cur_edge);
1241 		handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
1242 	}
1243 }
1244 
1245 static int handle_qpt_reg_umr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1246 			      const struct ib_send_wr *wr,
1247 			      struct mlx5_wqe_ctrl_seg **ctrl, void **seg,
1248 			      int *size, void **cur_edge, unsigned int idx)
1249 {
1250 	int err = 0;
1251 
1252 	if (unlikely(wr->opcode != MLX5_IB_WR_UMR)) {
1253 		err = -EINVAL;
1254 		mlx5_ib_warn(dev, "bad opcode %d\n", wr->opcode);
1255 		goto out;
1256 	}
1257 
1258 	qp->sq.wr_data[idx] = MLX5_IB_WR_UMR;
1259 	(*ctrl)->imm = cpu_to_be32(umr_wr(wr)->mkey);
1260 	err = set_reg_umr_segment(dev, *seg, wr);
1261 	if (unlikely(err))
1262 		goto out;
1263 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
1264 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
1265 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
1266 	set_reg_mkey_segment(*seg, wr);
1267 	*seg += sizeof(struct mlx5_mkey_seg);
1268 	*size += sizeof(struct mlx5_mkey_seg) / 16;
1269 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
1270 out:
1271 	return err;
1272 }
1273 
1274 int mlx5_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
1275 		      const struct ib_send_wr **bad_wr, bool drain)
1276 {
1277 	struct mlx5_wqe_ctrl_seg *ctrl = NULL;  /* compiler warning */
1278 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1279 	struct mlx5_core_dev *mdev = dev->mdev;
1280 	struct mlx5_ib_qp *qp;
1281 	struct mlx5_wqe_xrc_seg *xrc;
1282 	struct mlx5_bf *bf;
1283 	void *cur_edge;
1284 	int size;
1285 	unsigned long flags;
1286 	unsigned int idx;
1287 	int err = 0;
1288 	int num_sge;
1289 	void *seg;
1290 	int nreq;
1291 	int i;
1292 	u8 next_fence = 0;
1293 	u8 fence;
1294 
1295 	if (unlikely(mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR &&
1296 		     !drain)) {
1297 		*bad_wr = wr;
1298 		return -EIO;
1299 	}
1300 
1301 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
1302 		return mlx5_ib_gsi_post_send(ibqp, wr, bad_wr);
1303 
1304 	qp = to_mqp(ibqp);
1305 	bf = &qp->bf;
1306 
1307 	spin_lock_irqsave(&qp->sq.lock, flags);
1308 
1309 	for (nreq = 0; wr; nreq++, wr = wr->next) {
1310 		if (unlikely(wr->opcode >= ARRAY_SIZE(mlx5_ib_opcode))) {
1311 			mlx5_ib_warn(dev, "\n");
1312 			err = -EINVAL;
1313 			*bad_wr = wr;
1314 			goto out;
1315 		}
1316 
1317 		num_sge = wr->num_sge;
1318 		if (unlikely(num_sge > qp->sq.max_gs)) {
1319 			mlx5_ib_warn(dev, "\n");
1320 			err = -EINVAL;
1321 			*bad_wr = wr;
1322 			goto out;
1323 		}
1324 
1325 		err = begin_wqe(qp, &seg, &ctrl, wr, &idx, &size, &cur_edge,
1326 				nreq);
1327 		if (err) {
1328 			mlx5_ib_warn(dev, "\n");
1329 			err = -ENOMEM;
1330 			*bad_wr = wr;
1331 			goto out;
1332 		}
1333 
1334 		if (wr->opcode == IB_WR_REG_MR ||
1335 		    wr->opcode == IB_WR_REG_MR_INTEGRITY) {
1336 			fence = dev->umr_fence;
1337 			next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
1338 		} else  {
1339 			if (wr->send_flags & IB_SEND_FENCE) {
1340 				if (qp->next_fence)
1341 					fence = MLX5_FENCE_MODE_SMALL_AND_FENCE;
1342 				else
1343 					fence = MLX5_FENCE_MODE_FENCE;
1344 			} else {
1345 				fence = qp->next_fence;
1346 			}
1347 		}
1348 
1349 		switch (ibqp->qp_type) {
1350 		case IB_QPT_XRC_INI:
1351 			xrc = seg;
1352 			seg += sizeof(*xrc);
1353 			size += sizeof(*xrc) / 16;
1354 			fallthrough;
1355 		case IB_QPT_RC:
1356 			err = handle_qpt_rc(dev, qp, wr, &ctrl, &seg, &size,
1357 					    &cur_edge, &idx, nreq, fence,
1358 					    next_fence, &num_sge);
1359 			if (unlikely(err)) {
1360 				*bad_wr = wr;
1361 				goto out;
1362 			} else if (wr->opcode == IB_WR_REG_MR_INTEGRITY) {
1363 				goto skip_psv;
1364 			}
1365 			break;
1366 
1367 		case IB_QPT_UC:
1368 			handle_qpt_uc(wr, &seg, &size);
1369 			break;
1370 		case IB_QPT_SMI:
1371 			if (unlikely(!mdev->port_caps[qp->port - 1].has_smi)) {
1372 				mlx5_ib_warn(dev, "Send SMP MADs is not allowed\n");
1373 				err = -EPERM;
1374 				*bad_wr = wr;
1375 				goto out;
1376 			}
1377 			fallthrough;
1378 		case MLX5_IB_QPT_HW_GSI:
1379 			handle_qpt_hw_gsi(qp, wr, &seg, &size, &cur_edge);
1380 			break;
1381 		case IB_QPT_UD:
1382 			handle_qpt_ud(qp, wr, &seg, &size, &cur_edge);
1383 			break;
1384 		case MLX5_IB_QPT_REG_UMR:
1385 			err = handle_qpt_reg_umr(dev, qp, wr, &ctrl, &seg,
1386 						       &size, &cur_edge, idx);
1387 			if (unlikely(err))
1388 				goto out;
1389 			break;
1390 
1391 		default:
1392 			break;
1393 		}
1394 
1395 		if (wr->send_flags & IB_SEND_INLINE && num_sge) {
1396 			err = set_data_inl_seg(qp, wr, &seg, &size, &cur_edge);
1397 			if (unlikely(err)) {
1398 				mlx5_ib_warn(dev, "\n");
1399 				*bad_wr = wr;
1400 				goto out;
1401 			}
1402 		} else {
1403 			for (i = 0; i < num_sge; i++) {
1404 				handle_post_send_edge(&qp->sq, &seg, size,
1405 						      &cur_edge);
1406 				if (unlikely(!wr->sg_list[i].length))
1407 					continue;
1408 
1409 				set_data_ptr_seg(
1410 					(struct mlx5_wqe_data_seg *)seg,
1411 					wr->sg_list + i);
1412 				size += sizeof(struct mlx5_wqe_data_seg) / 16;
1413 				seg += sizeof(struct mlx5_wqe_data_seg);
1414 			}
1415 		}
1416 
1417 		qp->next_fence = next_fence;
1418 		finish_wqe(qp, ctrl, seg, size, cur_edge, idx, wr->wr_id, nreq,
1419 			   fence, mlx5_ib_opcode[wr->opcode]);
1420 skip_psv:
1421 		if (0)
1422 			dump_wqe(qp, idx, size);
1423 	}
1424 
1425 out:
1426 	if (likely(nreq)) {
1427 		qp->sq.head += nreq;
1428 
1429 		/* Make sure that descriptors are written before
1430 		 * updating doorbell record and ringing the doorbell
1431 		 */
1432 		wmb();
1433 
1434 		qp->db.db[MLX5_SND_DBR] = cpu_to_be32(qp->sq.cur_post);
1435 
1436 		/* Make sure doorbell record is visible to the HCA before
1437 		 * we hit doorbell.
1438 		 */
1439 		wmb();
1440 
1441 		mlx5_write64((__be32 *)ctrl, bf->bfreg->map + bf->offset);
1442 		/* Make sure doorbells don't leak out of SQ spinlock
1443 		 * and reach the HCA out of order.
1444 		 */
1445 		bf->offset ^= bf->buf_size;
1446 	}
1447 
1448 	spin_unlock_irqrestore(&qp->sq.lock, flags);
1449 
1450 	return err;
1451 }
1452 
1453 static void set_sig_seg(struct mlx5_rwqe_sig *sig, int max_gs)
1454 {
1455 	 sig->signature = calc_sig(sig, (max_gs + 1) << 2);
1456 }
1457 
1458 int mlx5_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
1459 		      const struct ib_recv_wr **bad_wr, bool drain)
1460 {
1461 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
1462 	struct mlx5_wqe_data_seg *scat;
1463 	struct mlx5_rwqe_sig *sig;
1464 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1465 	struct mlx5_core_dev *mdev = dev->mdev;
1466 	unsigned long flags;
1467 	int err = 0;
1468 	int nreq;
1469 	int ind;
1470 	int i;
1471 
1472 	if (unlikely(mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR &&
1473 		     !drain)) {
1474 		*bad_wr = wr;
1475 		return -EIO;
1476 	}
1477 
1478 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
1479 		return mlx5_ib_gsi_post_recv(ibqp, wr, bad_wr);
1480 
1481 	spin_lock_irqsave(&qp->rq.lock, flags);
1482 
1483 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
1484 
1485 	for (nreq = 0; wr; nreq++, wr = wr->next) {
1486 		if (mlx5_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1487 			err = -ENOMEM;
1488 			*bad_wr = wr;
1489 			goto out;
1490 		}
1491 
1492 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1493 			err = -EINVAL;
1494 			*bad_wr = wr;
1495 			goto out;
1496 		}
1497 
1498 		scat = mlx5_frag_buf_get_wqe(&qp->rq.fbc, ind);
1499 		if (qp->flags_en & MLX5_QP_FLAG_SIGNATURE)
1500 			scat++;
1501 
1502 		for (i = 0; i < wr->num_sge; i++)
1503 			set_data_ptr_seg(scat + i, wr->sg_list + i);
1504 
1505 		if (i < qp->rq.max_gs) {
1506 			scat[i].byte_count = 0;
1507 			scat[i].lkey       = cpu_to_be32(MLX5_INVALID_LKEY);
1508 			scat[i].addr       = 0;
1509 		}
1510 
1511 		if (qp->flags_en & MLX5_QP_FLAG_SIGNATURE) {
1512 			sig = (struct mlx5_rwqe_sig *)scat;
1513 			set_sig_seg(sig, qp->rq.max_gs);
1514 		}
1515 
1516 		qp->rq.wrid[ind] = wr->wr_id;
1517 
1518 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
1519 	}
1520 
1521 out:
1522 	if (likely(nreq)) {
1523 		qp->rq.head += nreq;
1524 
1525 		/* Make sure that descriptors are written before
1526 		 * doorbell record.
1527 		 */
1528 		wmb();
1529 
1530 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
1531 	}
1532 
1533 	spin_unlock_irqrestore(&qp->rq.lock, flags);
1534 
1535 	return err;
1536 }
1537