xref: /openbmc/linux/drivers/infiniband/hw/mlx5/wr.c (revision fb574682)
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 {
268 	u64 result;
269 
270 	result = MLX5_MKEY_MASK_LR |
271 		 MLX5_MKEY_MASK_LW |
272 		 MLX5_MKEY_MASK_RR |
273 		 MLX5_MKEY_MASK_RW;
274 
275 	if (atomic)
276 		result |= MLX5_MKEY_MASK_A;
277 
278 	return cpu_to_be64(result);
279 }
280 
281 static __be64 get_umr_update_pd_mask(void)
282 {
283 	u64 result;
284 
285 	result = MLX5_MKEY_MASK_PD;
286 
287 	return cpu_to_be64(result);
288 }
289 
290 static int umr_check_mkey_mask(struct mlx5_ib_dev *dev, u64 mask)
291 {
292 	if ((mask & MLX5_MKEY_MASK_PAGE_SIZE &&
293 	     MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled)) ||
294 	    (mask & MLX5_MKEY_MASK_A &&
295 	     MLX5_CAP_GEN(dev->mdev, umr_modify_atomic_disabled)))
296 		return -EPERM;
297 	return 0;
298 }
299 
300 static int set_reg_umr_segment(struct mlx5_ib_dev *dev,
301 			       struct mlx5_wqe_umr_ctrl_seg *umr,
302 			       const struct ib_send_wr *wr, int atomic)
303 {
304 	const struct mlx5_umr_wr *umrwr = umr_wr(wr);
305 
306 	memset(umr, 0, sizeof(*umr));
307 
308 	if (!umrwr->ignore_free_state) {
309 		if (wr->send_flags & MLX5_IB_SEND_UMR_FAIL_IF_FREE)
310 			 /* fail if free */
311 			umr->flags = MLX5_UMR_CHECK_FREE;
312 		else
313 			/* fail if not free */
314 			umr->flags = MLX5_UMR_CHECK_NOT_FREE;
315 	}
316 
317 	umr->xlt_octowords = cpu_to_be16(get_xlt_octo(umrwr->xlt_size));
318 	if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_XLT) {
319 		u64 offset = get_xlt_octo(umrwr->offset);
320 
321 		umr->xlt_offset = cpu_to_be16(offset & 0xffff);
322 		umr->xlt_offset_47_16 = cpu_to_be32(offset >> 16);
323 		umr->flags |= MLX5_UMR_TRANSLATION_OFFSET_EN;
324 	}
325 	if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_TRANSLATION)
326 		umr->mkey_mask |= get_umr_update_translation_mask();
327 	if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_PD_ACCESS) {
328 		umr->mkey_mask |= get_umr_update_access_mask(atomic);
329 		umr->mkey_mask |= get_umr_update_pd_mask();
330 	}
331 	if (wr->send_flags & MLX5_IB_SEND_UMR_ENABLE_MR)
332 		umr->mkey_mask |= get_umr_enable_mr_mask();
333 	if (wr->send_flags & MLX5_IB_SEND_UMR_DISABLE_MR)
334 		umr->mkey_mask |= get_umr_disable_mr_mask();
335 
336 	if (!wr->num_sge)
337 		umr->flags |= MLX5_UMR_INLINE;
338 
339 	return umr_check_mkey_mask(dev, be64_to_cpu(umr->mkey_mask));
340 }
341 
342 static u8 get_umr_flags(int acc)
343 {
344 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX5_PERM_ATOMIC       : 0) |
345 	       (acc & IB_ACCESS_REMOTE_WRITE  ? MLX5_PERM_REMOTE_WRITE : 0) |
346 	       (acc & IB_ACCESS_REMOTE_READ   ? MLX5_PERM_REMOTE_READ  : 0) |
347 	       (acc & IB_ACCESS_LOCAL_WRITE   ? MLX5_PERM_LOCAL_WRITE  : 0) |
348 		MLX5_PERM_LOCAL_READ | MLX5_PERM_UMR_EN;
349 }
350 
351 static void set_reg_mkey_seg(struct mlx5_mkey_seg *seg,
352 			     struct mlx5_ib_mr *mr,
353 			     u32 key, int access)
354 {
355 	int ndescs = ALIGN(mr->ndescs + mr->meta_ndescs, 8) >> 1;
356 
357 	memset(seg, 0, sizeof(*seg));
358 
359 	if (mr->access_mode == MLX5_MKC_ACCESS_MODE_MTT)
360 		seg->log2_page_size = ilog2(mr->ibmr.page_size);
361 	else if (mr->access_mode == MLX5_MKC_ACCESS_MODE_KLMS)
362 		/* KLMs take twice the size of MTTs */
363 		ndescs *= 2;
364 
365 	seg->flags = get_umr_flags(access) | mr->access_mode;
366 	seg->qpn_mkey7_0 = cpu_to_be32((key & 0xff) | 0xffffff00);
367 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL);
368 	seg->start_addr = cpu_to_be64(mr->ibmr.iova);
369 	seg->len = cpu_to_be64(mr->ibmr.length);
370 	seg->xlt_oct_size = cpu_to_be32(ndescs);
371 }
372 
373 static void set_linv_mkey_seg(struct mlx5_mkey_seg *seg)
374 {
375 	memset(seg, 0, sizeof(*seg));
376 	seg->status = MLX5_MKEY_STATUS_FREE;
377 }
378 
379 static void set_reg_mkey_segment(struct mlx5_mkey_seg *seg,
380 				 const struct ib_send_wr *wr)
381 {
382 	const struct mlx5_umr_wr *umrwr = umr_wr(wr);
383 
384 	memset(seg, 0, sizeof(*seg));
385 	if (wr->send_flags & MLX5_IB_SEND_UMR_DISABLE_MR)
386 		seg->status = MLX5_MKEY_STATUS_FREE;
387 
388 	seg->flags = convert_access(umrwr->access_flags);
389 	if (umrwr->pd)
390 		seg->flags_pd = cpu_to_be32(to_mpd(umrwr->pd)->pdn);
391 	if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_TRANSLATION &&
392 	    !umrwr->length)
393 		seg->flags_pd |= cpu_to_be32(MLX5_MKEY_LEN64);
394 
395 	seg->start_addr = cpu_to_be64(umrwr->virt_addr);
396 	seg->len = cpu_to_be64(umrwr->length);
397 	seg->log2_page_size = umrwr->page_shift;
398 	seg->qpn_mkey7_0 = cpu_to_be32(0xffffff00 |
399 				       mlx5_mkey_variant(umrwr->mkey));
400 }
401 
402 static void set_reg_data_seg(struct mlx5_wqe_data_seg *dseg,
403 			     struct mlx5_ib_mr *mr,
404 			     struct mlx5_ib_pd *pd)
405 {
406 	int bcount = mr->desc_size * (mr->ndescs + mr->meta_ndescs);
407 
408 	dseg->addr = cpu_to_be64(mr->desc_map);
409 	dseg->byte_count = cpu_to_be32(ALIGN(bcount, 64));
410 	dseg->lkey = cpu_to_be32(pd->ibpd.local_dma_lkey);
411 }
412 
413 static __be32 send_ieth(const struct ib_send_wr *wr)
414 {
415 	switch (wr->opcode) {
416 	case IB_WR_SEND_WITH_IMM:
417 	case IB_WR_RDMA_WRITE_WITH_IMM:
418 		return wr->ex.imm_data;
419 
420 	case IB_WR_SEND_WITH_INV:
421 		return cpu_to_be32(wr->ex.invalidate_rkey);
422 
423 	default:
424 		return 0;
425 	}
426 }
427 
428 static u8 calc_sig(void *wqe, int size)
429 {
430 	u8 *p = wqe;
431 	u8 res = 0;
432 	int i;
433 
434 	for (i = 0; i < size; i++)
435 		res ^= p[i];
436 
437 	return ~res;
438 }
439 
440 static u8 wq_sig(void *wqe)
441 {
442 	return calc_sig(wqe, (*((u8 *)wqe + 8) & 0x3f) << 4);
443 }
444 
445 static int set_data_inl_seg(struct mlx5_ib_qp *qp, const struct ib_send_wr *wr,
446 			    void **wqe, int *wqe_sz, void **cur_edge)
447 {
448 	struct mlx5_wqe_inline_seg *seg;
449 	size_t offset;
450 	int inl = 0;
451 	int i;
452 
453 	seg = *wqe;
454 	*wqe += sizeof(*seg);
455 	offset = sizeof(*seg);
456 
457 	for (i = 0; i < wr->num_sge; i++) {
458 		size_t len  = wr->sg_list[i].length;
459 		void *addr = (void *)(unsigned long)(wr->sg_list[i].addr);
460 
461 		inl += len;
462 
463 		if (unlikely(inl > qp->max_inline_data))
464 			return -ENOMEM;
465 
466 		while (likely(len)) {
467 			size_t leftlen;
468 			size_t copysz;
469 
470 			handle_post_send_edge(&qp->sq, wqe,
471 					      *wqe_sz + (offset >> 4),
472 					      cur_edge);
473 
474 			leftlen = *cur_edge - *wqe;
475 			copysz = min_t(size_t, leftlen, len);
476 
477 			memcpy(*wqe, addr, copysz);
478 			len -= copysz;
479 			addr += copysz;
480 			*wqe += copysz;
481 			offset += copysz;
482 		}
483 	}
484 
485 	seg->byte_count = cpu_to_be32(inl | MLX5_INLINE_SEG);
486 
487 	*wqe_sz +=  ALIGN(inl + sizeof(seg->byte_count), 16) / 16;
488 
489 	return 0;
490 }
491 
492 static u16 prot_field_size(enum ib_signature_type type)
493 {
494 	switch (type) {
495 	case IB_SIG_TYPE_T10_DIF:
496 		return MLX5_DIF_SIZE;
497 	default:
498 		return 0;
499 	}
500 }
501 
502 static u8 bs_selector(int block_size)
503 {
504 	switch (block_size) {
505 	case 512:	    return 0x1;
506 	case 520:	    return 0x2;
507 	case 4096:	    return 0x3;
508 	case 4160:	    return 0x4;
509 	case 1073741824:    return 0x5;
510 	default:	    return 0;
511 	}
512 }
513 
514 static void mlx5_fill_inl_bsf(struct ib_sig_domain *domain,
515 			      struct mlx5_bsf_inl *inl)
516 {
517 	/* Valid inline section and allow BSF refresh */
518 	inl->vld_refresh = cpu_to_be16(MLX5_BSF_INL_VALID |
519 				       MLX5_BSF_REFRESH_DIF);
520 	inl->dif_apptag = cpu_to_be16(domain->sig.dif.app_tag);
521 	inl->dif_reftag = cpu_to_be32(domain->sig.dif.ref_tag);
522 	/* repeating block */
523 	inl->rp_inv_seed = MLX5_BSF_REPEAT_BLOCK;
524 	inl->sig_type = domain->sig.dif.bg_type == IB_T10DIF_CRC ?
525 			MLX5_DIF_CRC : MLX5_DIF_IPCS;
526 
527 	if (domain->sig.dif.ref_remap)
528 		inl->dif_inc_ref_guard_check |= MLX5_BSF_INC_REFTAG;
529 
530 	if (domain->sig.dif.app_escape) {
531 		if (domain->sig.dif.ref_escape)
532 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPREF_ESCAPE;
533 		else
534 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPTAG_ESCAPE;
535 	}
536 
537 	inl->dif_app_bitmask_check =
538 		cpu_to_be16(domain->sig.dif.apptag_check_mask);
539 }
540 
541 static int mlx5_set_bsf(struct ib_mr *sig_mr,
542 			struct ib_sig_attrs *sig_attrs,
543 			struct mlx5_bsf *bsf, u32 data_size)
544 {
545 	struct mlx5_core_sig_ctx *msig = to_mmr(sig_mr)->sig;
546 	struct mlx5_bsf_basic *basic = &bsf->basic;
547 	struct ib_sig_domain *mem = &sig_attrs->mem;
548 	struct ib_sig_domain *wire = &sig_attrs->wire;
549 
550 	memset(bsf, 0, sizeof(*bsf));
551 
552 	/* Basic + Extended + Inline */
553 	basic->bsf_size_sbs = 1 << 7;
554 	/* Input domain check byte mask */
555 	basic->check_byte_mask = sig_attrs->check_mask;
556 	basic->raw_data_size = cpu_to_be32(data_size);
557 
558 	/* Memory domain */
559 	switch (sig_attrs->mem.sig_type) {
560 	case IB_SIG_TYPE_NONE:
561 		break;
562 	case IB_SIG_TYPE_T10_DIF:
563 		basic->mem.bs_selector = bs_selector(mem->sig.dif.pi_interval);
564 		basic->m_bfs_psv = cpu_to_be32(msig->psv_memory.psv_idx);
565 		mlx5_fill_inl_bsf(mem, &bsf->m_inl);
566 		break;
567 	default:
568 		return -EINVAL;
569 	}
570 
571 	/* Wire domain */
572 	switch (sig_attrs->wire.sig_type) {
573 	case IB_SIG_TYPE_NONE:
574 		break;
575 	case IB_SIG_TYPE_T10_DIF:
576 		if (mem->sig.dif.pi_interval == wire->sig.dif.pi_interval &&
577 		    mem->sig_type == wire->sig_type) {
578 			/* Same block structure */
579 			basic->bsf_size_sbs |= 1 << 4;
580 			if (mem->sig.dif.bg_type == wire->sig.dif.bg_type)
581 				basic->wire.copy_byte_mask |= MLX5_CPY_GRD_MASK;
582 			if (mem->sig.dif.app_tag == wire->sig.dif.app_tag)
583 				basic->wire.copy_byte_mask |= MLX5_CPY_APP_MASK;
584 			if (mem->sig.dif.ref_tag == wire->sig.dif.ref_tag)
585 				basic->wire.copy_byte_mask |= MLX5_CPY_REF_MASK;
586 		} else
587 			basic->wire.bs_selector =
588 				bs_selector(wire->sig.dif.pi_interval);
589 
590 		basic->w_bfs_psv = cpu_to_be32(msig->psv_wire.psv_idx);
591 		mlx5_fill_inl_bsf(wire, &bsf->w_inl);
592 		break;
593 	default:
594 		return -EINVAL;
595 	}
596 
597 	return 0;
598 }
599 
600 
601 static int set_sig_data_segment(const struct ib_send_wr *send_wr,
602 				struct ib_mr *sig_mr,
603 				struct ib_sig_attrs *sig_attrs,
604 				struct mlx5_ib_qp *qp, void **seg, int *size,
605 				void **cur_edge)
606 {
607 	struct mlx5_bsf *bsf;
608 	u32 data_len;
609 	u32 data_key;
610 	u64 data_va;
611 	u32 prot_len = 0;
612 	u32 prot_key = 0;
613 	u64 prot_va = 0;
614 	bool prot = false;
615 	int ret;
616 	int wqe_size;
617 	struct mlx5_ib_mr *mr = to_mmr(sig_mr);
618 	struct mlx5_ib_mr *pi_mr = mr->pi_mr;
619 
620 	data_len = pi_mr->data_length;
621 	data_key = pi_mr->ibmr.lkey;
622 	data_va = pi_mr->data_iova;
623 	if (pi_mr->meta_ndescs) {
624 		prot_len = pi_mr->meta_length;
625 		prot_key = pi_mr->ibmr.lkey;
626 		prot_va = pi_mr->pi_iova;
627 		prot = true;
628 	}
629 
630 	if (!prot || (data_key == prot_key && data_va == prot_va &&
631 		      data_len == prot_len)) {
632 		/**
633 		 * Source domain doesn't contain signature information
634 		 * or data and protection are interleaved in memory.
635 		 * So need construct:
636 		 *                  ------------------
637 		 *                 |     data_klm     |
638 		 *                  ------------------
639 		 *                 |       BSF        |
640 		 *                  ------------------
641 		 **/
642 		struct mlx5_klm *data_klm = *seg;
643 
644 		data_klm->bcount = cpu_to_be32(data_len);
645 		data_klm->key = cpu_to_be32(data_key);
646 		data_klm->va = cpu_to_be64(data_va);
647 		wqe_size = ALIGN(sizeof(*data_klm), 64);
648 	} else {
649 		/**
650 		 * Source domain contains signature information
651 		 * So need construct a strided block format:
652 		 *               ---------------------------
653 		 *              |     stride_block_ctrl     |
654 		 *               ---------------------------
655 		 *              |          data_klm         |
656 		 *               ---------------------------
657 		 *              |          prot_klm         |
658 		 *               ---------------------------
659 		 *              |             BSF           |
660 		 *               ---------------------------
661 		 **/
662 		struct mlx5_stride_block_ctrl_seg *sblock_ctrl;
663 		struct mlx5_stride_block_entry *data_sentry;
664 		struct mlx5_stride_block_entry *prot_sentry;
665 		u16 block_size = sig_attrs->mem.sig.dif.pi_interval;
666 		int prot_size;
667 
668 		sblock_ctrl = *seg;
669 		data_sentry = (void *)sblock_ctrl + sizeof(*sblock_ctrl);
670 		prot_sentry = (void *)data_sentry + sizeof(*data_sentry);
671 
672 		prot_size = prot_field_size(sig_attrs->mem.sig_type);
673 		if (!prot_size) {
674 			pr_err("Bad block size given: %u\n", block_size);
675 			return -EINVAL;
676 		}
677 		sblock_ctrl->bcount_per_cycle = cpu_to_be32(block_size +
678 							    prot_size);
679 		sblock_ctrl->op = cpu_to_be32(MLX5_STRIDE_BLOCK_OP);
680 		sblock_ctrl->repeat_count = cpu_to_be32(data_len / block_size);
681 		sblock_ctrl->num_entries = cpu_to_be16(2);
682 
683 		data_sentry->bcount = cpu_to_be16(block_size);
684 		data_sentry->key = cpu_to_be32(data_key);
685 		data_sentry->va = cpu_to_be64(data_va);
686 		data_sentry->stride = cpu_to_be16(block_size);
687 
688 		prot_sentry->bcount = cpu_to_be16(prot_size);
689 		prot_sentry->key = cpu_to_be32(prot_key);
690 		prot_sentry->va = cpu_to_be64(prot_va);
691 		prot_sentry->stride = cpu_to_be16(prot_size);
692 
693 		wqe_size = ALIGN(sizeof(*sblock_ctrl) + sizeof(*data_sentry) +
694 				 sizeof(*prot_sentry), 64);
695 	}
696 
697 	*seg += wqe_size;
698 	*size += wqe_size / 16;
699 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
700 
701 	bsf = *seg;
702 	ret = mlx5_set_bsf(sig_mr, sig_attrs, bsf, data_len);
703 	if (ret)
704 		return -EINVAL;
705 
706 	*seg += sizeof(*bsf);
707 	*size += sizeof(*bsf) / 16;
708 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
709 
710 	return 0;
711 }
712 
713 static void set_sig_mkey_segment(struct mlx5_mkey_seg *seg,
714 				 struct ib_mr *sig_mr, int access_flags,
715 				 u32 size, u32 length, u32 pdn)
716 {
717 	u32 sig_key = sig_mr->rkey;
718 	u8 sigerr = to_mmr(sig_mr)->sig->sigerr_count & 1;
719 
720 	memset(seg, 0, sizeof(*seg));
721 
722 	seg->flags = get_umr_flags(access_flags) | MLX5_MKC_ACCESS_MODE_KLMS;
723 	seg->qpn_mkey7_0 = cpu_to_be32((sig_key & 0xff) | 0xffffff00);
724 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL | sigerr << 26 |
725 				    MLX5_MKEY_BSF_EN | pdn);
726 	seg->len = cpu_to_be64(length);
727 	seg->xlt_oct_size = cpu_to_be32(get_xlt_octo(size));
728 	seg->bsfs_octo_size = cpu_to_be32(MLX5_MKEY_BSF_OCTO_SIZE);
729 }
730 
731 static void set_sig_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
732 				u32 size)
733 {
734 	memset(umr, 0, sizeof(*umr));
735 
736 	umr->flags = MLX5_FLAGS_INLINE | MLX5_FLAGS_CHECK_FREE;
737 	umr->xlt_octowords = cpu_to_be16(get_xlt_octo(size));
738 	umr->bsf_octowords = cpu_to_be16(MLX5_MKEY_BSF_OCTO_SIZE);
739 	umr->mkey_mask = sig_mkey_mask();
740 }
741 
742 static int set_pi_umr_wr(const struct ib_send_wr *send_wr,
743 			 struct mlx5_ib_qp *qp, void **seg, int *size,
744 			 void **cur_edge)
745 {
746 	const struct ib_reg_wr *wr = reg_wr(send_wr);
747 	struct mlx5_ib_mr *sig_mr = to_mmr(wr->mr);
748 	struct mlx5_ib_mr *pi_mr = sig_mr->pi_mr;
749 	struct ib_sig_attrs *sig_attrs = sig_mr->ibmr.sig_attrs;
750 	u32 pdn = to_mpd(qp->ibqp.pd)->pdn;
751 	u32 xlt_size;
752 	int region_len, ret;
753 
754 	if (unlikely(send_wr->num_sge != 0) ||
755 	    unlikely(wr->access & IB_ACCESS_REMOTE_ATOMIC) ||
756 	    unlikely(!sig_mr->sig) || unlikely(!qp->ibqp.integrity_en) ||
757 	    unlikely(!sig_mr->sig->sig_status_checked))
758 		return -EINVAL;
759 
760 	/* length of the protected region, data + protection */
761 	region_len = pi_mr->ibmr.length;
762 
763 	/**
764 	 * KLM octoword size - if protection was provided
765 	 * then we use strided block format (3 octowords),
766 	 * else we use single KLM (1 octoword)
767 	 **/
768 	if (sig_attrs->mem.sig_type != IB_SIG_TYPE_NONE)
769 		xlt_size = 0x30;
770 	else
771 		xlt_size = sizeof(struct mlx5_klm);
772 
773 	set_sig_umr_segment(*seg, xlt_size);
774 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
775 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
776 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
777 
778 	set_sig_mkey_segment(*seg, wr->mr, wr->access, xlt_size, region_len,
779 			     pdn);
780 	*seg += sizeof(struct mlx5_mkey_seg);
781 	*size += sizeof(struct mlx5_mkey_seg) / 16;
782 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
783 
784 	ret = set_sig_data_segment(send_wr, wr->mr, sig_attrs, qp, seg, size,
785 				   cur_edge);
786 	if (ret)
787 		return ret;
788 
789 	sig_mr->sig->sig_status_checked = false;
790 	return 0;
791 }
792 
793 static int set_psv_wr(struct ib_sig_domain *domain,
794 		      u32 psv_idx, void **seg, int *size)
795 {
796 	struct mlx5_seg_set_psv *psv_seg = *seg;
797 
798 	memset(psv_seg, 0, sizeof(*psv_seg));
799 	psv_seg->psv_num = cpu_to_be32(psv_idx);
800 	switch (domain->sig_type) {
801 	case IB_SIG_TYPE_NONE:
802 		break;
803 	case IB_SIG_TYPE_T10_DIF:
804 		psv_seg->transient_sig = cpu_to_be32(domain->sig.dif.bg << 16 |
805 						     domain->sig.dif.app_tag);
806 		psv_seg->ref_tag = cpu_to_be32(domain->sig.dif.ref_tag);
807 		break;
808 	default:
809 		pr_err("Bad signature type (%d) is given.\n",
810 		       domain->sig_type);
811 		return -EINVAL;
812 	}
813 
814 	*seg += sizeof(*psv_seg);
815 	*size += sizeof(*psv_seg) / 16;
816 
817 	return 0;
818 }
819 
820 static int set_reg_wr(struct mlx5_ib_qp *qp,
821 		      const struct ib_reg_wr *wr,
822 		      void **seg, int *size, void **cur_edge,
823 		      bool check_not_free)
824 {
825 	struct mlx5_ib_mr *mr = to_mmr(wr->mr);
826 	struct mlx5_ib_pd *pd = to_mpd(qp->ibqp.pd);
827 	struct mlx5_ib_dev *dev = to_mdev(pd->ibpd.device);
828 	int mr_list_size = (mr->ndescs + mr->meta_ndescs) * mr->desc_size;
829 	bool umr_inline = mr_list_size <= MLX5_IB_SQ_UMR_INLINE_THRESHOLD;
830 	bool atomic = wr->access & IB_ACCESS_REMOTE_ATOMIC;
831 	u8 flags = 0;
832 
833 	if (!mlx5_ib_can_use_umr(dev, atomic, wr->access)) {
834 		mlx5_ib_warn(to_mdev(qp->ibqp.device),
835 			     "Fast update of %s for MR is disabled\n",
836 			     (MLX5_CAP_GEN(dev->mdev,
837 					   umr_modify_entity_size_disabled)) ?
838 				     "entity size" :
839 				     "atomic access");
840 		return -EINVAL;
841 	}
842 
843 	if (unlikely(wr->wr.send_flags & IB_SEND_INLINE)) {
844 		mlx5_ib_warn(to_mdev(qp->ibqp.device),
845 			     "Invalid IB_SEND_INLINE send flag\n");
846 		return -EINVAL;
847 	}
848 
849 	if (check_not_free)
850 		flags |= MLX5_UMR_CHECK_NOT_FREE;
851 	if (umr_inline)
852 		flags |= MLX5_UMR_INLINE;
853 
854 	set_reg_umr_seg(*seg, mr, flags, atomic);
855 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
856 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
857 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
858 
859 	set_reg_mkey_seg(*seg, mr, wr->key, wr->access);
860 	*seg += sizeof(struct mlx5_mkey_seg);
861 	*size += sizeof(struct mlx5_mkey_seg) / 16;
862 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
863 
864 	if (umr_inline) {
865 		memcpy_send_wqe(&qp->sq, cur_edge, seg, size, mr->descs,
866 				mr_list_size);
867 		*size = ALIGN(*size, MLX5_SEND_WQE_BB >> 4);
868 	} else {
869 		set_reg_data_seg(*seg, mr, pd);
870 		*seg += sizeof(struct mlx5_wqe_data_seg);
871 		*size += (sizeof(struct mlx5_wqe_data_seg) / 16);
872 	}
873 	return 0;
874 }
875 
876 static void set_linv_wr(struct mlx5_ib_qp *qp, void **seg, int *size,
877 			void **cur_edge)
878 {
879 	set_linv_umr_seg(*seg);
880 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
881 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
882 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
883 	set_linv_mkey_seg(*seg);
884 	*seg += sizeof(struct mlx5_mkey_seg);
885 	*size += sizeof(struct mlx5_mkey_seg) / 16;
886 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
887 }
888 
889 static void dump_wqe(struct mlx5_ib_qp *qp, u32 idx, int size_16)
890 {
891 	__be32 *p = NULL;
892 	int i, j;
893 
894 	pr_debug("dump WQE index %u:\n", idx);
895 	for (i = 0, j = 0; i < size_16 * 4; i += 4, j += 4) {
896 		if ((i & 0xf) == 0) {
897 			p = mlx5_frag_buf_get_wqe(&qp->sq.fbc, idx);
898 			pr_debug("WQBB at %p:\n", (void *)p);
899 			j = 0;
900 			idx = (idx + 1) & (qp->sq.wqe_cnt - 1);
901 		}
902 		pr_debug("%08x %08x %08x %08x\n", be32_to_cpu(p[j]),
903 			 be32_to_cpu(p[j + 1]), be32_to_cpu(p[j + 2]),
904 			 be32_to_cpu(p[j + 3]));
905 	}
906 }
907 
908 static int __begin_wqe(struct mlx5_ib_qp *qp, void **seg,
909 		       struct mlx5_wqe_ctrl_seg **ctrl,
910 		       const struct ib_send_wr *wr, unsigned int *idx,
911 		       int *size, void **cur_edge, int nreq,
912 		       bool send_signaled, bool solicited)
913 {
914 	if (unlikely(mlx5_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)))
915 		return -ENOMEM;
916 
917 	*idx = qp->sq.cur_post & (qp->sq.wqe_cnt - 1);
918 	*seg = mlx5_frag_buf_get_wqe(&qp->sq.fbc, *idx);
919 	*ctrl = *seg;
920 	*(uint32_t *)(*seg + 8) = 0;
921 	(*ctrl)->imm = send_ieth(wr);
922 	(*ctrl)->fm_ce_se = qp->sq_signal_bits |
923 		(send_signaled ? MLX5_WQE_CTRL_CQ_UPDATE : 0) |
924 		(solicited ? MLX5_WQE_CTRL_SOLICITED : 0);
925 
926 	*seg += sizeof(**ctrl);
927 	*size = sizeof(**ctrl) / 16;
928 	*cur_edge = qp->sq.cur_edge;
929 
930 	return 0;
931 }
932 
933 static int begin_wqe(struct mlx5_ib_qp *qp, void **seg,
934 		     struct mlx5_wqe_ctrl_seg **ctrl,
935 		     const struct ib_send_wr *wr, unsigned int *idx, int *size,
936 		     void **cur_edge, int nreq)
937 {
938 	return __begin_wqe(qp, seg, ctrl, wr, idx, size, cur_edge, nreq,
939 			   wr->send_flags & IB_SEND_SIGNALED,
940 			   wr->send_flags & IB_SEND_SOLICITED);
941 }
942 
943 static void finish_wqe(struct mlx5_ib_qp *qp,
944 		       struct mlx5_wqe_ctrl_seg *ctrl,
945 		       void *seg, u8 size, void *cur_edge,
946 		       unsigned int idx, u64 wr_id, int nreq, u8 fence,
947 		       u32 mlx5_opcode)
948 {
949 	u8 opmod = 0;
950 
951 	ctrl->opmod_idx_opcode = cpu_to_be32(((u32)(qp->sq.cur_post) << 8) |
952 					     mlx5_opcode | ((u32)opmod << 24));
953 	ctrl->qpn_ds = cpu_to_be32(size | (qp->trans_qp.base.mqp.qpn << 8));
954 	ctrl->fm_ce_se |= fence;
955 	if (unlikely(qp->flags_en & MLX5_QP_FLAG_SIGNATURE))
956 		ctrl->signature = wq_sig(ctrl);
957 
958 	qp->sq.wrid[idx] = wr_id;
959 	qp->sq.w_list[idx].opcode = mlx5_opcode;
960 	qp->sq.wqe_head[idx] = qp->sq.head + nreq;
961 	qp->sq.cur_post += DIV_ROUND_UP(size * 16, MLX5_SEND_WQE_BB);
962 	qp->sq.w_list[idx].next = qp->sq.cur_post;
963 
964 	/* We save the edge which was possibly updated during the WQE
965 	 * construction, into SQ's cache.
966 	 */
967 	seg = PTR_ALIGN(seg, MLX5_SEND_WQE_BB);
968 	qp->sq.cur_edge = (unlikely(seg == cur_edge)) ?
969 			  get_sq_edge(&qp->sq, qp->sq.cur_post &
970 				      (qp->sq.wqe_cnt - 1)) :
971 			  cur_edge;
972 }
973 
974 static void handle_rdma_op(const struct ib_send_wr *wr, void **seg, int *size)
975 {
976 	set_raddr_seg(*seg, rdma_wr(wr)->remote_addr, rdma_wr(wr)->rkey);
977 	*seg += sizeof(struct mlx5_wqe_raddr_seg);
978 	*size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
979 }
980 
981 static void handle_local_inv(struct mlx5_ib_qp *qp, const struct ib_send_wr *wr,
982 			     struct mlx5_wqe_ctrl_seg **ctrl, void **seg,
983 			     int *size, void **cur_edge, unsigned int idx)
984 {
985 	qp->sq.wr_data[idx] = IB_WR_LOCAL_INV;
986 	(*ctrl)->imm = cpu_to_be32(wr->ex.invalidate_rkey);
987 	set_linv_wr(qp, seg, size, cur_edge);
988 }
989 
990 static int handle_reg_mr(struct mlx5_ib_qp *qp, const struct ib_send_wr *wr,
991 			 struct mlx5_wqe_ctrl_seg **ctrl, void **seg, int *size,
992 			 void **cur_edge, unsigned int idx)
993 {
994 	qp->sq.wr_data[idx] = IB_WR_REG_MR;
995 	(*ctrl)->imm = cpu_to_be32(reg_wr(wr)->key);
996 	return set_reg_wr(qp, reg_wr(wr), seg, size, cur_edge, true);
997 }
998 
999 static int handle_psv(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1000 		      const struct ib_send_wr *wr,
1001 		      struct mlx5_wqe_ctrl_seg **ctrl, void **seg, int *size,
1002 		      void **cur_edge, unsigned int *idx, int nreq,
1003 		      struct ib_sig_domain *domain, u32 psv_index,
1004 		      u8 next_fence)
1005 {
1006 	int err;
1007 
1008 	/*
1009 	 * SET_PSV WQEs are not signaled and solicited on error.
1010 	 */
1011 	err = __begin_wqe(qp, seg, ctrl, wr, idx, size, cur_edge, nreq,
1012 			  false, true);
1013 	if (unlikely(err)) {
1014 		mlx5_ib_warn(dev, "\n");
1015 		err = -ENOMEM;
1016 		goto out;
1017 	}
1018 	err = set_psv_wr(domain, psv_index, seg, size);
1019 	if (unlikely(err)) {
1020 		mlx5_ib_warn(dev, "\n");
1021 		goto out;
1022 	}
1023 	finish_wqe(qp, *ctrl, *seg, *size, *cur_edge, *idx, wr->wr_id, nreq,
1024 		   next_fence, MLX5_OPCODE_SET_PSV);
1025 
1026 out:
1027 	return err;
1028 }
1029 
1030 static int handle_reg_mr_integrity(struct mlx5_ib_dev *dev,
1031 				   struct mlx5_ib_qp *qp,
1032 				   const struct ib_send_wr *wr,
1033 				   struct mlx5_wqe_ctrl_seg **ctrl, void **seg,
1034 				   int *size, void **cur_edge,
1035 				   unsigned int *idx, int nreq, u8 fence,
1036 				   u8 next_fence)
1037 {
1038 	struct mlx5_ib_mr *mr;
1039 	struct mlx5_ib_mr *pi_mr;
1040 	struct mlx5_ib_mr pa_pi_mr;
1041 	struct ib_sig_attrs *sig_attrs;
1042 	struct ib_reg_wr reg_pi_wr;
1043 	int err;
1044 
1045 	qp->sq.wr_data[*idx] = IB_WR_REG_MR_INTEGRITY;
1046 
1047 	mr = to_mmr(reg_wr(wr)->mr);
1048 	pi_mr = mr->pi_mr;
1049 
1050 	if (pi_mr) {
1051 		memset(&reg_pi_wr, 0,
1052 		       sizeof(struct ib_reg_wr));
1053 
1054 		reg_pi_wr.mr = &pi_mr->ibmr;
1055 		reg_pi_wr.access = reg_wr(wr)->access;
1056 		reg_pi_wr.key = pi_mr->ibmr.rkey;
1057 
1058 		(*ctrl)->imm = cpu_to_be32(reg_pi_wr.key);
1059 		/* UMR for data + prot registration */
1060 		err = set_reg_wr(qp, &reg_pi_wr, seg, size, cur_edge, false);
1061 		if (unlikely(err))
1062 			goto out;
1063 
1064 		finish_wqe(qp, *ctrl, *seg, *size, *cur_edge, *idx, wr->wr_id,
1065 			   nreq, fence, MLX5_OPCODE_UMR);
1066 
1067 		err = begin_wqe(qp, seg, ctrl, wr, idx, size, cur_edge, nreq);
1068 		if (unlikely(err)) {
1069 			mlx5_ib_warn(dev, "\n");
1070 			err = -ENOMEM;
1071 			goto out;
1072 		}
1073 	} else {
1074 		memset(&pa_pi_mr, 0, sizeof(struct mlx5_ib_mr));
1075 		/* No UMR, use local_dma_lkey */
1076 		pa_pi_mr.ibmr.lkey = mr->ibmr.pd->local_dma_lkey;
1077 		pa_pi_mr.ndescs = mr->ndescs;
1078 		pa_pi_mr.data_length = mr->data_length;
1079 		pa_pi_mr.data_iova = mr->data_iova;
1080 		if (mr->meta_ndescs) {
1081 			pa_pi_mr.meta_ndescs = mr->meta_ndescs;
1082 			pa_pi_mr.meta_length = mr->meta_length;
1083 			pa_pi_mr.pi_iova = mr->pi_iova;
1084 		}
1085 
1086 		pa_pi_mr.ibmr.length = mr->ibmr.length;
1087 		mr->pi_mr = &pa_pi_mr;
1088 	}
1089 	(*ctrl)->imm = cpu_to_be32(mr->ibmr.rkey);
1090 	/* UMR for sig MR */
1091 	err = set_pi_umr_wr(wr, qp, seg, size, cur_edge);
1092 	if (unlikely(err)) {
1093 		mlx5_ib_warn(dev, "\n");
1094 		goto out;
1095 	}
1096 	finish_wqe(qp, *ctrl, *seg, *size, *cur_edge, *idx, wr->wr_id, nreq,
1097 		   fence, MLX5_OPCODE_UMR);
1098 
1099 	sig_attrs = mr->ibmr.sig_attrs;
1100 	err = handle_psv(dev, qp, wr, ctrl, seg, size, cur_edge, idx, nreq,
1101 			 &sig_attrs->mem, mr->sig->psv_memory.psv_idx,
1102 			 next_fence);
1103 	if (unlikely(err))
1104 		goto out;
1105 
1106 	err = handle_psv(dev, qp, wr, ctrl, seg, size, cur_edge, idx, nreq,
1107 			 &sig_attrs->wire, mr->sig->psv_wire.psv_idx,
1108 			 next_fence);
1109 	if (unlikely(err))
1110 		goto out;
1111 
1112 	qp->next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
1113 
1114 out:
1115 	return err;
1116 }
1117 
1118 static int handle_qpt_rc(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1119 			 const struct ib_send_wr *wr,
1120 			 struct mlx5_wqe_ctrl_seg **ctrl, void **seg, int *size,
1121 			 void **cur_edge, unsigned int *idx, int nreq, u8 fence,
1122 			 u8 next_fence, int *num_sge)
1123 {
1124 	int err = 0;
1125 
1126 	switch (wr->opcode) {
1127 	case IB_WR_RDMA_READ:
1128 	case IB_WR_RDMA_WRITE:
1129 	case IB_WR_RDMA_WRITE_WITH_IMM:
1130 		handle_rdma_op(wr, seg, size);
1131 		break;
1132 
1133 	case IB_WR_ATOMIC_CMP_AND_SWP:
1134 	case IB_WR_ATOMIC_FETCH_AND_ADD:
1135 	case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
1136 		mlx5_ib_warn(dev, "Atomic operations are not supported yet\n");
1137 		err = -EOPNOTSUPP;
1138 		goto out;
1139 
1140 	case IB_WR_LOCAL_INV:
1141 		handle_local_inv(qp, wr, ctrl, seg, size, cur_edge, *idx);
1142 		*num_sge = 0;
1143 		break;
1144 
1145 	case IB_WR_REG_MR:
1146 		err = handle_reg_mr(qp, wr, ctrl, seg, size, cur_edge, *idx);
1147 		if (unlikely(err))
1148 			goto out;
1149 		*num_sge = 0;
1150 		break;
1151 
1152 	case IB_WR_REG_MR_INTEGRITY:
1153 		err = handle_reg_mr_integrity(dev, qp, wr, ctrl, seg, size,
1154 					      cur_edge, idx, nreq, fence,
1155 					      next_fence);
1156 		if (unlikely(err))
1157 			goto out;
1158 		*num_sge = 0;
1159 		break;
1160 
1161 	default:
1162 		break;
1163 	}
1164 
1165 out:
1166 	return err;
1167 }
1168 
1169 static void handle_qpt_uc(const struct ib_send_wr *wr, void **seg, int *size)
1170 {
1171 	switch (wr->opcode) {
1172 	case IB_WR_RDMA_WRITE:
1173 	case IB_WR_RDMA_WRITE_WITH_IMM:
1174 		handle_rdma_op(wr, seg, size);
1175 		break;
1176 	default:
1177 		break;
1178 	}
1179 }
1180 
1181 static void handle_qpt_hw_gsi(struct mlx5_ib_qp *qp,
1182 			      const struct ib_send_wr *wr, void **seg,
1183 			      int *size, void **cur_edge)
1184 {
1185 	set_datagram_seg(*seg, wr);
1186 	*seg += sizeof(struct mlx5_wqe_datagram_seg);
1187 	*size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
1188 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
1189 }
1190 
1191 static void handle_qpt_ud(struct mlx5_ib_qp *qp, const struct ib_send_wr *wr,
1192 			  void **seg, int *size, void **cur_edge)
1193 {
1194 	set_datagram_seg(*seg, wr);
1195 	*seg += sizeof(struct mlx5_wqe_datagram_seg);
1196 	*size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
1197 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
1198 
1199 	/* handle qp that supports ud offload */
1200 	if (qp->flags & IB_QP_CREATE_IPOIB_UD_LSO) {
1201 		struct mlx5_wqe_eth_pad *pad;
1202 
1203 		pad = *seg;
1204 		memset(pad, 0, sizeof(struct mlx5_wqe_eth_pad));
1205 		*seg += sizeof(struct mlx5_wqe_eth_pad);
1206 		*size += sizeof(struct mlx5_wqe_eth_pad) / 16;
1207 		set_eth_seg(wr, qp, seg, size, cur_edge);
1208 		handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
1209 	}
1210 }
1211 
1212 static int handle_qpt_reg_umr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1213 			      const struct ib_send_wr *wr,
1214 			      struct mlx5_wqe_ctrl_seg **ctrl, void **seg,
1215 			      int *size, void **cur_edge, unsigned int idx)
1216 {
1217 	int err = 0;
1218 
1219 	if (unlikely(wr->opcode != MLX5_IB_WR_UMR)) {
1220 		err = -EINVAL;
1221 		mlx5_ib_warn(dev, "bad opcode %d\n", wr->opcode);
1222 		goto out;
1223 	}
1224 
1225 	qp->sq.wr_data[idx] = MLX5_IB_WR_UMR;
1226 	(*ctrl)->imm = cpu_to_be32(umr_wr(wr)->mkey);
1227 	err = set_reg_umr_segment(dev, *seg, wr,
1228 				  !!(MLX5_CAP_GEN(dev->mdev, atomic)));
1229 	if (unlikely(err))
1230 		goto out;
1231 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
1232 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
1233 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
1234 	set_reg_mkey_segment(*seg, wr);
1235 	*seg += sizeof(struct mlx5_mkey_seg);
1236 	*size += sizeof(struct mlx5_mkey_seg) / 16;
1237 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
1238 out:
1239 	return err;
1240 }
1241 
1242 int mlx5_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
1243 		      const struct ib_send_wr **bad_wr, bool drain)
1244 {
1245 	struct mlx5_wqe_ctrl_seg *ctrl = NULL;  /* compiler warning */
1246 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1247 	struct mlx5_core_dev *mdev = dev->mdev;
1248 	struct mlx5_ib_qp *qp;
1249 	struct mlx5_wqe_xrc_seg *xrc;
1250 	struct mlx5_bf *bf;
1251 	void *cur_edge;
1252 	int uninitialized_var(size);
1253 	unsigned long flags;
1254 	unsigned int idx;
1255 	int err = 0;
1256 	int num_sge;
1257 	void *seg;
1258 	int nreq;
1259 	int i;
1260 	u8 next_fence = 0;
1261 	u8 fence;
1262 
1263 	if (unlikely(mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR &&
1264 		     !drain)) {
1265 		*bad_wr = wr;
1266 		return -EIO;
1267 	}
1268 
1269 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
1270 		return mlx5_ib_gsi_post_send(ibqp, wr, bad_wr);
1271 
1272 	qp = to_mqp(ibqp);
1273 	bf = &qp->bf;
1274 
1275 	spin_lock_irqsave(&qp->sq.lock, flags);
1276 
1277 	for (nreq = 0; wr; nreq++, wr = wr->next) {
1278 		if (unlikely(wr->opcode >= ARRAY_SIZE(mlx5_ib_opcode))) {
1279 			mlx5_ib_warn(dev, "\n");
1280 			err = -EINVAL;
1281 			*bad_wr = wr;
1282 			goto out;
1283 		}
1284 
1285 		num_sge = wr->num_sge;
1286 		if (unlikely(num_sge > qp->sq.max_gs)) {
1287 			mlx5_ib_warn(dev, "\n");
1288 			err = -EINVAL;
1289 			*bad_wr = wr;
1290 			goto out;
1291 		}
1292 
1293 		err = begin_wqe(qp, &seg, &ctrl, wr, &idx, &size, &cur_edge,
1294 				nreq);
1295 		if (err) {
1296 			mlx5_ib_warn(dev, "\n");
1297 			err = -ENOMEM;
1298 			*bad_wr = wr;
1299 			goto out;
1300 		}
1301 
1302 		if (wr->opcode == IB_WR_REG_MR ||
1303 		    wr->opcode == IB_WR_REG_MR_INTEGRITY) {
1304 			fence = dev->umr_fence;
1305 			next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
1306 		} else  {
1307 			if (wr->send_flags & IB_SEND_FENCE) {
1308 				if (qp->next_fence)
1309 					fence = MLX5_FENCE_MODE_SMALL_AND_FENCE;
1310 				else
1311 					fence = MLX5_FENCE_MODE_FENCE;
1312 			} else {
1313 				fence = qp->next_fence;
1314 			}
1315 		}
1316 
1317 		switch (ibqp->qp_type) {
1318 		case IB_QPT_XRC_INI:
1319 			xrc = seg;
1320 			seg += sizeof(*xrc);
1321 			size += sizeof(*xrc) / 16;
1322 			fallthrough;
1323 		case IB_QPT_RC:
1324 			err = handle_qpt_rc(dev, qp, wr, &ctrl, &seg, &size,
1325 					    &cur_edge, &idx, nreq, fence,
1326 					    next_fence, &num_sge);
1327 			if (unlikely(err)) {
1328 				*bad_wr = wr;
1329 				goto out;
1330 			} else if (wr->opcode == IB_WR_REG_MR_INTEGRITY) {
1331 				goto skip_psv;
1332 			}
1333 			break;
1334 
1335 		case IB_QPT_UC:
1336 			handle_qpt_uc(wr, &seg, &size);
1337 			break;
1338 		case IB_QPT_SMI:
1339 			if (unlikely(!mdev->port_caps[qp->port - 1].has_smi)) {
1340 				mlx5_ib_warn(dev, "Send SMP MADs is not allowed\n");
1341 				err = -EPERM;
1342 				*bad_wr = wr;
1343 				goto out;
1344 			}
1345 			fallthrough;
1346 		case MLX5_IB_QPT_HW_GSI:
1347 			handle_qpt_hw_gsi(qp, wr, &seg, &size, &cur_edge);
1348 			break;
1349 		case IB_QPT_UD:
1350 			handle_qpt_ud(qp, wr, &seg, &size, &cur_edge);
1351 			break;
1352 		case MLX5_IB_QPT_REG_UMR:
1353 			err = handle_qpt_reg_umr(dev, qp, wr, &ctrl, &seg,
1354 						       &size, &cur_edge, idx);
1355 			if (unlikely(err))
1356 				goto out;
1357 			break;
1358 
1359 		default:
1360 			break;
1361 		}
1362 
1363 		if (wr->send_flags & IB_SEND_INLINE && num_sge) {
1364 			err = set_data_inl_seg(qp, wr, &seg, &size, &cur_edge);
1365 			if (unlikely(err)) {
1366 				mlx5_ib_warn(dev, "\n");
1367 				*bad_wr = wr;
1368 				goto out;
1369 			}
1370 		} else {
1371 			for (i = 0; i < num_sge; i++) {
1372 				handle_post_send_edge(&qp->sq, &seg, size,
1373 						      &cur_edge);
1374 				if (unlikely(!wr->sg_list[i].length))
1375 					continue;
1376 
1377 				set_data_ptr_seg(
1378 					(struct mlx5_wqe_data_seg *)seg,
1379 					wr->sg_list + i);
1380 				size += sizeof(struct mlx5_wqe_data_seg) / 16;
1381 				seg += sizeof(struct mlx5_wqe_data_seg);
1382 			}
1383 		}
1384 
1385 		qp->next_fence = next_fence;
1386 		finish_wqe(qp, ctrl, seg, size, cur_edge, idx, wr->wr_id, nreq,
1387 			   fence, mlx5_ib_opcode[wr->opcode]);
1388 skip_psv:
1389 		if (0)
1390 			dump_wqe(qp, idx, size);
1391 	}
1392 
1393 out:
1394 	if (likely(nreq)) {
1395 		qp->sq.head += nreq;
1396 
1397 		/* Make sure that descriptors are written before
1398 		 * updating doorbell record and ringing the doorbell
1399 		 */
1400 		wmb();
1401 
1402 		qp->db.db[MLX5_SND_DBR] = cpu_to_be32(qp->sq.cur_post);
1403 
1404 		/* Make sure doorbell record is visible to the HCA before
1405 		 * we hit doorbell.
1406 		 */
1407 		wmb();
1408 
1409 		mlx5_write64((__be32 *)ctrl, bf->bfreg->map + bf->offset);
1410 		/* Make sure doorbells don't leak out of SQ spinlock
1411 		 * and reach the HCA out of order.
1412 		 */
1413 		bf->offset ^= bf->buf_size;
1414 	}
1415 
1416 	spin_unlock_irqrestore(&qp->sq.lock, flags);
1417 
1418 	return err;
1419 }
1420 
1421 static void set_sig_seg(struct mlx5_rwqe_sig *sig, int max_gs)
1422 {
1423 	 sig->signature = calc_sig(sig, (max_gs + 1) << 2);
1424 }
1425 
1426 int mlx5_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
1427 		      const struct ib_recv_wr **bad_wr, bool drain)
1428 {
1429 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
1430 	struct mlx5_wqe_data_seg *scat;
1431 	struct mlx5_rwqe_sig *sig;
1432 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1433 	struct mlx5_core_dev *mdev = dev->mdev;
1434 	unsigned long flags;
1435 	int err = 0;
1436 	int nreq;
1437 	int ind;
1438 	int i;
1439 
1440 	if (unlikely(mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR &&
1441 		     !drain)) {
1442 		*bad_wr = wr;
1443 		return -EIO;
1444 	}
1445 
1446 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
1447 		return mlx5_ib_gsi_post_recv(ibqp, wr, bad_wr);
1448 
1449 	spin_lock_irqsave(&qp->rq.lock, flags);
1450 
1451 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
1452 
1453 	for (nreq = 0; wr; nreq++, wr = wr->next) {
1454 		if (mlx5_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1455 			err = -ENOMEM;
1456 			*bad_wr = wr;
1457 			goto out;
1458 		}
1459 
1460 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1461 			err = -EINVAL;
1462 			*bad_wr = wr;
1463 			goto out;
1464 		}
1465 
1466 		scat = mlx5_frag_buf_get_wqe(&qp->rq.fbc, ind);
1467 		if (qp->flags_en & MLX5_QP_FLAG_SIGNATURE)
1468 			scat++;
1469 
1470 		for (i = 0; i < wr->num_sge; i++)
1471 			set_data_ptr_seg(scat + i, wr->sg_list + i);
1472 
1473 		if (i < qp->rq.max_gs) {
1474 			scat[i].byte_count = 0;
1475 			scat[i].lkey       = cpu_to_be32(MLX5_INVALID_LKEY);
1476 			scat[i].addr       = 0;
1477 		}
1478 
1479 		if (qp->flags_en & MLX5_QP_FLAG_SIGNATURE) {
1480 			sig = (struct mlx5_rwqe_sig *)scat;
1481 			set_sig_seg(sig, qp->rq.max_gs);
1482 		}
1483 
1484 		qp->rq.wrid[ind] = wr->wr_id;
1485 
1486 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
1487 	}
1488 
1489 out:
1490 	if (likely(nreq)) {
1491 		qp->rq.head += nreq;
1492 
1493 		/* Make sure that descriptors are written before
1494 		 * doorbell record.
1495 		 */
1496 		wmb();
1497 
1498 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
1499 	}
1500 
1501 	spin_unlock_irqrestore(&qp->rq.lock, flags);
1502 
1503 	return err;
1504 }
1505