xref: /openbmc/linux/drivers/infiniband/hw/qedr/verbs.c (revision 61bf3293)
1 /* QLogic qedr NIC Driver
2  * Copyright (c) 2015-2016  QLogic Corporation
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and /or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 #include <linux/dma-mapping.h>
33 #include <linux/crc32.h>
34 #include <net/ip.h>
35 #include <net/ipv6.h>
36 #include <net/udp.h>
37 #include <linux/iommu.h>
38 
39 #include <rdma/ib_verbs.h>
40 #include <rdma/ib_user_verbs.h>
41 #include <rdma/iw_cm.h>
42 #include <rdma/ib_umem.h>
43 #include <rdma/ib_addr.h>
44 #include <rdma/ib_cache.h>
45 #include <rdma/uverbs_ioctl.h>
46 
47 #include <linux/qed/common_hsi.h>
48 #include "qedr_hsi_rdma.h"
49 #include <linux/qed/qed_if.h>
50 #include "qedr.h"
51 #include "verbs.h"
52 #include <rdma/qedr-abi.h>
53 #include "qedr_roce_cm.h"
54 #include "qedr_iw_cm.h"
55 
56 #define QEDR_SRQ_WQE_ELEM_SIZE	sizeof(union rdma_srq_elm)
57 #define	RDMA_MAX_SGE_PER_SRQ	(4)
58 #define RDMA_MAX_SRQ_WQE_SIZE	(RDMA_MAX_SGE_PER_SRQ + 1)
59 
60 #define DB_ADDR_SHIFT(addr)		((addr) << DB_PWM_ADDR_OFFSET_SHIFT)
61 
62 enum {
63 	QEDR_USER_MMAP_IO_WC = 0,
64 	QEDR_USER_MMAP_PHYS_PAGE,
65 };
66 
67 static inline int qedr_ib_copy_to_udata(struct ib_udata *udata, void *src,
68 					size_t len)
69 {
70 	size_t min_len = min_t(size_t, len, udata->outlen);
71 
72 	return ib_copy_to_udata(udata, src, min_len);
73 }
74 
75 int qedr_query_pkey(struct ib_device *ibdev, u8 port, u16 index, u16 *pkey)
76 {
77 	if (index >= QEDR_ROCE_PKEY_TABLE_LEN)
78 		return -EINVAL;
79 
80 	*pkey = QEDR_ROCE_PKEY_DEFAULT;
81 	return 0;
82 }
83 
84 int qedr_iw_query_gid(struct ib_device *ibdev, u8 port,
85 		      int index, union ib_gid *sgid)
86 {
87 	struct qedr_dev *dev = get_qedr_dev(ibdev);
88 
89 	memset(sgid->raw, 0, sizeof(sgid->raw));
90 	ether_addr_copy(sgid->raw, dev->ndev->dev_addr);
91 
92 	DP_DEBUG(dev, QEDR_MSG_INIT, "QUERY sgid[%d]=%llx:%llx\n", index,
93 		 sgid->global.interface_id, sgid->global.subnet_prefix);
94 
95 	return 0;
96 }
97 
98 int qedr_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
99 {
100 	struct qedr_dev *dev = get_qedr_dev(ibsrq->device);
101 	struct qedr_device_attr *qattr = &dev->attr;
102 	struct qedr_srq *srq = get_qedr_srq(ibsrq);
103 
104 	srq_attr->srq_limit = srq->srq_limit;
105 	srq_attr->max_wr = qattr->max_srq_wr;
106 	srq_attr->max_sge = qattr->max_sge;
107 
108 	return 0;
109 }
110 
111 int qedr_query_device(struct ib_device *ibdev,
112 		      struct ib_device_attr *attr, struct ib_udata *udata)
113 {
114 	struct qedr_dev *dev = get_qedr_dev(ibdev);
115 	struct qedr_device_attr *qattr = &dev->attr;
116 
117 	if (!dev->rdma_ctx) {
118 		DP_ERR(dev,
119 		       "qedr_query_device called with invalid params rdma_ctx=%p\n",
120 		       dev->rdma_ctx);
121 		return -EINVAL;
122 	}
123 
124 	memset(attr, 0, sizeof(*attr));
125 
126 	attr->fw_ver = qattr->fw_ver;
127 	attr->sys_image_guid = qattr->sys_image_guid;
128 	attr->max_mr_size = qattr->max_mr_size;
129 	attr->page_size_cap = qattr->page_size_caps;
130 	attr->vendor_id = qattr->vendor_id;
131 	attr->vendor_part_id = qattr->vendor_part_id;
132 	attr->hw_ver = qattr->hw_ver;
133 	attr->max_qp = qattr->max_qp;
134 	attr->max_qp_wr = max_t(u32, qattr->max_sqe, qattr->max_rqe);
135 	attr->device_cap_flags = IB_DEVICE_CURR_QP_STATE_MOD |
136 	    IB_DEVICE_RC_RNR_NAK_GEN |
137 	    IB_DEVICE_LOCAL_DMA_LKEY | IB_DEVICE_MEM_MGT_EXTENSIONS;
138 
139 	attr->max_send_sge = qattr->max_sge;
140 	attr->max_recv_sge = qattr->max_sge;
141 	attr->max_sge_rd = qattr->max_sge;
142 	attr->max_cq = qattr->max_cq;
143 	attr->max_cqe = qattr->max_cqe;
144 	attr->max_mr = qattr->max_mr;
145 	attr->max_mw = qattr->max_mw;
146 	attr->max_pd = qattr->max_pd;
147 	attr->atomic_cap = dev->atomic_cap;
148 	attr->max_qp_init_rd_atom =
149 	    1 << (fls(qattr->max_qp_req_rd_atomic_resc) - 1);
150 	attr->max_qp_rd_atom =
151 	    min(1 << (fls(qattr->max_qp_resp_rd_atomic_resc) - 1),
152 		attr->max_qp_init_rd_atom);
153 
154 	attr->max_srq = qattr->max_srq;
155 	attr->max_srq_sge = qattr->max_srq_sge;
156 	attr->max_srq_wr = qattr->max_srq_wr;
157 
158 	attr->local_ca_ack_delay = qattr->dev_ack_delay;
159 	attr->max_fast_reg_page_list_len = qattr->max_mr / 8;
160 	attr->max_pkeys = QEDR_ROCE_PKEY_MAX;
161 	attr->max_ah = qattr->max_ah;
162 
163 	return 0;
164 }
165 
166 static inline void get_link_speed_and_width(int speed, u8 *ib_speed,
167 					    u8 *ib_width)
168 {
169 	switch (speed) {
170 	case 1000:
171 		*ib_speed = IB_SPEED_SDR;
172 		*ib_width = IB_WIDTH_1X;
173 		break;
174 	case 10000:
175 		*ib_speed = IB_SPEED_QDR;
176 		*ib_width = IB_WIDTH_1X;
177 		break;
178 
179 	case 20000:
180 		*ib_speed = IB_SPEED_DDR;
181 		*ib_width = IB_WIDTH_4X;
182 		break;
183 
184 	case 25000:
185 		*ib_speed = IB_SPEED_EDR;
186 		*ib_width = IB_WIDTH_1X;
187 		break;
188 
189 	case 40000:
190 		*ib_speed = IB_SPEED_QDR;
191 		*ib_width = IB_WIDTH_4X;
192 		break;
193 
194 	case 50000:
195 		*ib_speed = IB_SPEED_HDR;
196 		*ib_width = IB_WIDTH_1X;
197 		break;
198 
199 	case 100000:
200 		*ib_speed = IB_SPEED_EDR;
201 		*ib_width = IB_WIDTH_4X;
202 		break;
203 
204 	default:
205 		/* Unsupported */
206 		*ib_speed = IB_SPEED_SDR;
207 		*ib_width = IB_WIDTH_1X;
208 	}
209 }
210 
211 int qedr_query_port(struct ib_device *ibdev, u8 port, struct ib_port_attr *attr)
212 {
213 	struct qedr_dev *dev;
214 	struct qed_rdma_port *rdma_port;
215 
216 	dev = get_qedr_dev(ibdev);
217 
218 	if (!dev->rdma_ctx) {
219 		DP_ERR(dev, "rdma_ctx is NULL\n");
220 		return -EINVAL;
221 	}
222 
223 	rdma_port = dev->ops->rdma_query_port(dev->rdma_ctx);
224 
225 	/* *attr being zeroed by the caller, avoid zeroing it here */
226 	if (rdma_port->port_state == QED_RDMA_PORT_UP) {
227 		attr->state = IB_PORT_ACTIVE;
228 		attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
229 	} else {
230 		attr->state = IB_PORT_DOWN;
231 		attr->phys_state = IB_PORT_PHYS_STATE_DISABLED;
232 	}
233 	attr->max_mtu = IB_MTU_4096;
234 	attr->active_mtu = iboe_get_mtu(dev->ndev->mtu);
235 	attr->lid = 0;
236 	attr->lmc = 0;
237 	attr->sm_lid = 0;
238 	attr->sm_sl = 0;
239 	attr->ip_gids = true;
240 	if (rdma_protocol_iwarp(&dev->ibdev, 1)) {
241 		attr->gid_tbl_len = 1;
242 		attr->pkey_tbl_len = 1;
243 	} else {
244 		attr->gid_tbl_len = QEDR_MAX_SGID;
245 		attr->pkey_tbl_len = QEDR_ROCE_PKEY_TABLE_LEN;
246 	}
247 	attr->bad_pkey_cntr = rdma_port->pkey_bad_counter;
248 	attr->qkey_viol_cntr = 0;
249 	get_link_speed_and_width(rdma_port->link_speed,
250 				 &attr->active_speed, &attr->active_width);
251 	attr->max_msg_sz = rdma_port->max_msg_size;
252 	attr->max_vl_num = 4;
253 
254 	return 0;
255 }
256 
257 int qedr_alloc_ucontext(struct ib_ucontext *uctx, struct ib_udata *udata)
258 {
259 	struct ib_device *ibdev = uctx->device;
260 	int rc;
261 	struct qedr_ucontext *ctx = get_qedr_ucontext(uctx);
262 	struct qedr_alloc_ucontext_resp uresp = {};
263 	struct qedr_alloc_ucontext_req ureq = {};
264 	struct qedr_dev *dev = get_qedr_dev(ibdev);
265 	struct qed_rdma_add_user_out_params oparams;
266 	struct qedr_user_mmap_entry *entry;
267 
268 	if (!udata)
269 		return -EFAULT;
270 
271 	if (udata->inlen) {
272 		rc = ib_copy_from_udata(&ureq, udata,
273 					min(sizeof(ureq), udata->inlen));
274 		if (rc) {
275 			DP_ERR(dev, "Problem copying data from user space\n");
276 			return -EFAULT;
277 		}
278 
279 		ctx->db_rec = !!(ureq.context_flags & QEDR_ALLOC_UCTX_DB_REC);
280 	}
281 
282 	rc = dev->ops->rdma_add_user(dev->rdma_ctx, &oparams);
283 	if (rc) {
284 		DP_ERR(dev,
285 		       "failed to allocate a DPI for a new RoCE application, rc=%d. To overcome this consider to increase the number of DPIs, increase the doorbell BAR size or just close unnecessary RoCE applications. In order to increase the number of DPIs consult the qedr readme\n",
286 		       rc);
287 		return rc;
288 	}
289 
290 	ctx->dpi = oparams.dpi;
291 	ctx->dpi_addr = oparams.dpi_addr;
292 	ctx->dpi_phys_addr = oparams.dpi_phys_addr;
293 	ctx->dpi_size = oparams.dpi_size;
294 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
295 	if (!entry) {
296 		rc = -ENOMEM;
297 		goto err;
298 	}
299 
300 	entry->io_address = ctx->dpi_phys_addr;
301 	entry->length = ctx->dpi_size;
302 	entry->mmap_flag = QEDR_USER_MMAP_IO_WC;
303 	entry->dpi = ctx->dpi;
304 	entry->dev = dev;
305 	rc = rdma_user_mmap_entry_insert(uctx, &entry->rdma_entry,
306 					 ctx->dpi_size);
307 	if (rc) {
308 		kfree(entry);
309 		goto err;
310 	}
311 	ctx->db_mmap_entry = &entry->rdma_entry;
312 
313 	if (!dev->user_dpm_enabled)
314 		uresp.dpm_flags = 0;
315 	else if (rdma_protocol_iwarp(&dev->ibdev, 1))
316 		uresp.dpm_flags = QEDR_DPM_TYPE_IWARP_LEGACY;
317 	else
318 		uresp.dpm_flags = QEDR_DPM_TYPE_ROCE_ENHANCED |
319 				  QEDR_DPM_TYPE_ROCE_LEGACY;
320 
321 	uresp.dpm_flags |= QEDR_DPM_SIZES_SET;
322 	uresp.ldpm_limit_size = QEDR_LDPM_MAX_SIZE;
323 	uresp.edpm_trans_size = QEDR_EDPM_TRANS_SIZE;
324 
325 	uresp.wids_enabled = 1;
326 	uresp.wid_count = oparams.wid_count;
327 	uresp.db_pa = rdma_user_mmap_get_offset(ctx->db_mmap_entry);
328 	uresp.db_size = ctx->dpi_size;
329 	uresp.max_send_wr = dev->attr.max_sqe;
330 	uresp.max_recv_wr = dev->attr.max_rqe;
331 	uresp.max_srq_wr = dev->attr.max_srq_wr;
332 	uresp.sges_per_send_wr = QEDR_MAX_SQE_ELEMENTS_PER_SQE;
333 	uresp.sges_per_recv_wr = QEDR_MAX_RQE_ELEMENTS_PER_RQE;
334 	uresp.sges_per_srq_wr = dev->attr.max_srq_sge;
335 	uresp.max_cqes = QEDR_MAX_CQES;
336 
337 	rc = qedr_ib_copy_to_udata(udata, &uresp, sizeof(uresp));
338 	if (rc)
339 		goto err;
340 
341 	ctx->dev = dev;
342 
343 	DP_DEBUG(dev, QEDR_MSG_INIT, "Allocating user context %p\n",
344 		 &ctx->ibucontext);
345 	return 0;
346 
347 err:
348 	if (!ctx->db_mmap_entry)
349 		dev->ops->rdma_remove_user(dev->rdma_ctx, ctx->dpi);
350 	else
351 		rdma_user_mmap_entry_remove(ctx->db_mmap_entry);
352 
353 	return rc;
354 }
355 
356 void qedr_dealloc_ucontext(struct ib_ucontext *ibctx)
357 {
358 	struct qedr_ucontext *uctx = get_qedr_ucontext(ibctx);
359 
360 	DP_DEBUG(uctx->dev, QEDR_MSG_INIT, "Deallocating user context %p\n",
361 		 uctx);
362 
363 	rdma_user_mmap_entry_remove(uctx->db_mmap_entry);
364 }
365 
366 void qedr_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
367 {
368 	struct qedr_user_mmap_entry *entry = get_qedr_mmap_entry(rdma_entry);
369 	struct qedr_dev *dev = entry->dev;
370 
371 	if (entry->mmap_flag == QEDR_USER_MMAP_PHYS_PAGE)
372 		free_page((unsigned long)entry->address);
373 	else if (entry->mmap_flag == QEDR_USER_MMAP_IO_WC)
374 		dev->ops->rdma_remove_user(dev->rdma_ctx, entry->dpi);
375 
376 	kfree(entry);
377 }
378 
379 int qedr_mmap(struct ib_ucontext *ucontext, struct vm_area_struct *vma)
380 {
381 	struct ib_device *dev = ucontext->device;
382 	size_t length = vma->vm_end - vma->vm_start;
383 	struct rdma_user_mmap_entry *rdma_entry;
384 	struct qedr_user_mmap_entry *entry;
385 	int rc = 0;
386 	u64 pfn;
387 
388 	ibdev_dbg(dev,
389 		  "start %#lx, end %#lx, length = %#zx, pgoff = %#lx\n",
390 		  vma->vm_start, vma->vm_end, length, vma->vm_pgoff);
391 
392 	rdma_entry = rdma_user_mmap_entry_get(ucontext, vma);
393 	if (!rdma_entry) {
394 		ibdev_dbg(dev, "pgoff[%#lx] does not have valid entry\n",
395 			  vma->vm_pgoff);
396 		return -EINVAL;
397 	}
398 	entry = get_qedr_mmap_entry(rdma_entry);
399 	ibdev_dbg(dev,
400 		  "Mapping address[%#llx], length[%#zx], mmap_flag[%d]\n",
401 		  entry->io_address, length, entry->mmap_flag);
402 
403 	switch (entry->mmap_flag) {
404 	case QEDR_USER_MMAP_IO_WC:
405 		pfn = entry->io_address >> PAGE_SHIFT;
406 		rc = rdma_user_mmap_io(ucontext, vma, pfn, length,
407 				       pgprot_writecombine(vma->vm_page_prot),
408 				       rdma_entry);
409 		break;
410 	case QEDR_USER_MMAP_PHYS_PAGE:
411 		rc = vm_insert_page(vma, vma->vm_start,
412 				    virt_to_page(entry->address));
413 		break;
414 	default:
415 		rc = -EINVAL;
416 	}
417 
418 	if (rc)
419 		ibdev_dbg(dev,
420 			  "Couldn't mmap address[%#llx] length[%#zx] mmap_flag[%d] err[%d]\n",
421 			  entry->io_address, length, entry->mmap_flag, rc);
422 
423 	rdma_user_mmap_entry_put(rdma_entry);
424 	return rc;
425 }
426 
427 int qedr_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
428 {
429 	struct ib_device *ibdev = ibpd->device;
430 	struct qedr_dev *dev = get_qedr_dev(ibdev);
431 	struct qedr_pd *pd = get_qedr_pd(ibpd);
432 	u16 pd_id;
433 	int rc;
434 
435 	DP_DEBUG(dev, QEDR_MSG_INIT, "Function called from: %s\n",
436 		 udata ? "User Lib" : "Kernel");
437 
438 	if (!dev->rdma_ctx) {
439 		DP_ERR(dev, "invalid RDMA context\n");
440 		return -EINVAL;
441 	}
442 
443 	rc = dev->ops->rdma_alloc_pd(dev->rdma_ctx, &pd_id);
444 	if (rc)
445 		return rc;
446 
447 	pd->pd_id = pd_id;
448 
449 	if (udata) {
450 		struct qedr_alloc_pd_uresp uresp = {
451 			.pd_id = pd_id,
452 		};
453 		struct qedr_ucontext *context = rdma_udata_to_drv_context(
454 			udata, struct qedr_ucontext, ibucontext);
455 
456 		rc = qedr_ib_copy_to_udata(udata, &uresp, sizeof(uresp));
457 		if (rc) {
458 			DP_ERR(dev, "copy error pd_id=0x%x.\n", pd_id);
459 			dev->ops->rdma_dealloc_pd(dev->rdma_ctx, pd_id);
460 			return rc;
461 		}
462 
463 		pd->uctx = context;
464 		pd->uctx->pd = pd;
465 	}
466 
467 	return 0;
468 }
469 
470 void qedr_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
471 {
472 	struct qedr_dev *dev = get_qedr_dev(ibpd->device);
473 	struct qedr_pd *pd = get_qedr_pd(ibpd);
474 
475 	DP_DEBUG(dev, QEDR_MSG_INIT, "Deallocating PD %d\n", pd->pd_id);
476 	dev->ops->rdma_dealloc_pd(dev->rdma_ctx, pd->pd_id);
477 }
478 
479 static void qedr_free_pbl(struct qedr_dev *dev,
480 			  struct qedr_pbl_info *pbl_info, struct qedr_pbl *pbl)
481 {
482 	struct pci_dev *pdev = dev->pdev;
483 	int i;
484 
485 	for (i = 0; i < pbl_info->num_pbls; i++) {
486 		if (!pbl[i].va)
487 			continue;
488 		dma_free_coherent(&pdev->dev, pbl_info->pbl_size,
489 				  pbl[i].va, pbl[i].pa);
490 	}
491 
492 	kfree(pbl);
493 }
494 
495 #define MIN_FW_PBL_PAGE_SIZE (4 * 1024)
496 #define MAX_FW_PBL_PAGE_SIZE (64 * 1024)
497 
498 #define NUM_PBES_ON_PAGE(_page_size) (_page_size / sizeof(u64))
499 #define MAX_PBES_ON_PAGE NUM_PBES_ON_PAGE(MAX_FW_PBL_PAGE_SIZE)
500 #define MAX_PBES_TWO_LAYER (MAX_PBES_ON_PAGE * MAX_PBES_ON_PAGE)
501 
502 static struct qedr_pbl *qedr_alloc_pbl_tbl(struct qedr_dev *dev,
503 					   struct qedr_pbl_info *pbl_info,
504 					   gfp_t flags)
505 {
506 	struct pci_dev *pdev = dev->pdev;
507 	struct qedr_pbl *pbl_table;
508 	dma_addr_t *pbl_main_tbl;
509 	dma_addr_t pa;
510 	void *va;
511 	int i;
512 
513 	pbl_table = kcalloc(pbl_info->num_pbls, sizeof(*pbl_table), flags);
514 	if (!pbl_table)
515 		return ERR_PTR(-ENOMEM);
516 
517 	for (i = 0; i < pbl_info->num_pbls; i++) {
518 		va = dma_alloc_coherent(&pdev->dev, pbl_info->pbl_size, &pa,
519 					flags);
520 		if (!va)
521 			goto err;
522 
523 		pbl_table[i].va = va;
524 		pbl_table[i].pa = pa;
525 	}
526 
527 	/* Two-Layer PBLs, if we have more than one pbl we need to initialize
528 	 * the first one with physical pointers to all of the rest
529 	 */
530 	pbl_main_tbl = (dma_addr_t *)pbl_table[0].va;
531 	for (i = 0; i < pbl_info->num_pbls - 1; i++)
532 		pbl_main_tbl[i] = pbl_table[i + 1].pa;
533 
534 	return pbl_table;
535 
536 err:
537 	for (i--; i >= 0; i--)
538 		dma_free_coherent(&pdev->dev, pbl_info->pbl_size,
539 				  pbl_table[i].va, pbl_table[i].pa);
540 
541 	qedr_free_pbl(dev, pbl_info, pbl_table);
542 
543 	return ERR_PTR(-ENOMEM);
544 }
545 
546 static int qedr_prepare_pbl_tbl(struct qedr_dev *dev,
547 				struct qedr_pbl_info *pbl_info,
548 				u32 num_pbes, int two_layer_capable)
549 {
550 	u32 pbl_capacity;
551 	u32 pbl_size;
552 	u32 num_pbls;
553 
554 	if ((num_pbes > MAX_PBES_ON_PAGE) && two_layer_capable) {
555 		if (num_pbes > MAX_PBES_TWO_LAYER) {
556 			DP_ERR(dev, "prepare pbl table: too many pages %d\n",
557 			       num_pbes);
558 			return -EINVAL;
559 		}
560 
561 		/* calculate required pbl page size */
562 		pbl_size = MIN_FW_PBL_PAGE_SIZE;
563 		pbl_capacity = NUM_PBES_ON_PAGE(pbl_size) *
564 			       NUM_PBES_ON_PAGE(pbl_size);
565 
566 		while (pbl_capacity < num_pbes) {
567 			pbl_size *= 2;
568 			pbl_capacity = pbl_size / sizeof(u64);
569 			pbl_capacity = pbl_capacity * pbl_capacity;
570 		}
571 
572 		num_pbls = DIV_ROUND_UP(num_pbes, NUM_PBES_ON_PAGE(pbl_size));
573 		num_pbls++;	/* One for the layer0 ( points to the pbls) */
574 		pbl_info->two_layered = true;
575 	} else {
576 		/* One layered PBL */
577 		num_pbls = 1;
578 		pbl_size = max_t(u32, MIN_FW_PBL_PAGE_SIZE,
579 				 roundup_pow_of_two((num_pbes * sizeof(u64))));
580 		pbl_info->two_layered = false;
581 	}
582 
583 	pbl_info->num_pbls = num_pbls;
584 	pbl_info->pbl_size = pbl_size;
585 	pbl_info->num_pbes = num_pbes;
586 
587 	DP_DEBUG(dev, QEDR_MSG_MR,
588 		 "prepare pbl table: num_pbes=%d, num_pbls=%d, pbl_size=%d\n",
589 		 pbl_info->num_pbes, pbl_info->num_pbls, pbl_info->pbl_size);
590 
591 	return 0;
592 }
593 
594 static void qedr_populate_pbls(struct qedr_dev *dev, struct ib_umem *umem,
595 			       struct qedr_pbl *pbl,
596 			       struct qedr_pbl_info *pbl_info, u32 pg_shift)
597 {
598 	int pbe_cnt, total_num_pbes = 0;
599 	u32 fw_pg_cnt, fw_pg_per_umem_pg;
600 	struct qedr_pbl *pbl_tbl;
601 	struct sg_dma_page_iter sg_iter;
602 	struct regpair *pbe;
603 	u64 pg_addr;
604 
605 	if (!pbl_info->num_pbes)
606 		return;
607 
608 	/* If we have a two layered pbl, the first pbl points to the rest
609 	 * of the pbls and the first entry lays on the second pbl in the table
610 	 */
611 	if (pbl_info->two_layered)
612 		pbl_tbl = &pbl[1];
613 	else
614 		pbl_tbl = pbl;
615 
616 	pbe = (struct regpair *)pbl_tbl->va;
617 	if (!pbe) {
618 		DP_ERR(dev, "cannot populate PBL due to a NULL PBE\n");
619 		return;
620 	}
621 
622 	pbe_cnt = 0;
623 
624 	fw_pg_per_umem_pg = BIT(PAGE_SHIFT - pg_shift);
625 
626 	for_each_sg_dma_page (umem->sg_head.sgl, &sg_iter, umem->nmap, 0) {
627 		pg_addr = sg_page_iter_dma_address(&sg_iter);
628 		for (fw_pg_cnt = 0; fw_pg_cnt < fw_pg_per_umem_pg;) {
629 			pbe->lo = cpu_to_le32(pg_addr);
630 			pbe->hi = cpu_to_le32(upper_32_bits(pg_addr));
631 
632 			pg_addr += BIT(pg_shift);
633 			pbe_cnt++;
634 			total_num_pbes++;
635 			pbe++;
636 
637 			if (total_num_pbes == pbl_info->num_pbes)
638 				return;
639 
640 			/* If the given pbl is full storing the pbes,
641 			 * move to next pbl.
642 			 */
643 			if (pbe_cnt == (pbl_info->pbl_size / sizeof(u64))) {
644 				pbl_tbl++;
645 				pbe = (struct regpair *)pbl_tbl->va;
646 				pbe_cnt = 0;
647 			}
648 
649 			fw_pg_cnt++;
650 		}
651 	}
652 }
653 
654 static int qedr_db_recovery_add(struct qedr_dev *dev,
655 				void __iomem *db_addr,
656 				void *db_data,
657 				enum qed_db_rec_width db_width,
658 				enum qed_db_rec_space db_space)
659 {
660 	if (!db_data) {
661 		DP_DEBUG(dev, QEDR_MSG_INIT, "avoiding db rec since old lib\n");
662 		return 0;
663 	}
664 
665 	return dev->ops->common->db_recovery_add(dev->cdev, db_addr, db_data,
666 						 db_width, db_space);
667 }
668 
669 static void qedr_db_recovery_del(struct qedr_dev *dev,
670 				 void __iomem *db_addr,
671 				 void *db_data)
672 {
673 	if (!db_data) {
674 		DP_DEBUG(dev, QEDR_MSG_INIT, "avoiding db rec since old lib\n");
675 		return;
676 	}
677 
678 	/* Ignore return code as there is not much we can do about it. Error
679 	 * log will be printed inside.
680 	 */
681 	dev->ops->common->db_recovery_del(dev->cdev, db_addr, db_data);
682 }
683 
684 static int qedr_copy_cq_uresp(struct qedr_dev *dev,
685 			      struct qedr_cq *cq, struct ib_udata *udata,
686 			      u32 db_offset)
687 {
688 	struct qedr_create_cq_uresp uresp;
689 	int rc;
690 
691 	memset(&uresp, 0, sizeof(uresp));
692 
693 	uresp.db_offset = db_offset;
694 	uresp.icid = cq->icid;
695 	if (cq->q.db_mmap_entry)
696 		uresp.db_rec_addr =
697 			rdma_user_mmap_get_offset(cq->q.db_mmap_entry);
698 
699 	rc = qedr_ib_copy_to_udata(udata, &uresp, sizeof(uresp));
700 	if (rc)
701 		DP_ERR(dev, "copy error cqid=0x%x.\n", cq->icid);
702 
703 	return rc;
704 }
705 
706 static void consume_cqe(struct qedr_cq *cq)
707 {
708 	if (cq->latest_cqe == cq->toggle_cqe)
709 		cq->pbl_toggle ^= RDMA_CQE_REQUESTER_TOGGLE_BIT_MASK;
710 
711 	cq->latest_cqe = qed_chain_consume(&cq->pbl);
712 }
713 
714 static inline int qedr_align_cq_entries(int entries)
715 {
716 	u64 size, aligned_size;
717 
718 	/* We allocate an extra entry that we don't report to the FW. */
719 	size = (entries + 1) * QEDR_CQE_SIZE;
720 	aligned_size = ALIGN(size, PAGE_SIZE);
721 
722 	return aligned_size / QEDR_CQE_SIZE;
723 }
724 
725 static int qedr_init_user_db_rec(struct ib_udata *udata,
726 				 struct qedr_dev *dev, struct qedr_userq *q,
727 				 bool requires_db_rec)
728 {
729 	struct qedr_ucontext *uctx =
730 		rdma_udata_to_drv_context(udata, struct qedr_ucontext,
731 					  ibucontext);
732 	struct qedr_user_mmap_entry *entry;
733 	int rc;
734 
735 	/* Aborting for non doorbell userqueue (SRQ) or non-supporting lib */
736 	if (requires_db_rec == 0 || !uctx->db_rec)
737 		return 0;
738 
739 	/* Allocate a page for doorbell recovery, add to mmap */
740 	q->db_rec_data = (void *)get_zeroed_page(GFP_USER);
741 	if (!q->db_rec_data) {
742 		DP_ERR(dev, "get_zeroed_page failed\n");
743 		return -ENOMEM;
744 	}
745 
746 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
747 	if (!entry)
748 		goto err_free_db_data;
749 
750 	entry->address = q->db_rec_data;
751 	entry->length = PAGE_SIZE;
752 	entry->mmap_flag = QEDR_USER_MMAP_PHYS_PAGE;
753 	rc = rdma_user_mmap_entry_insert(&uctx->ibucontext,
754 					 &entry->rdma_entry,
755 					 PAGE_SIZE);
756 	if (rc)
757 		goto err_free_entry;
758 
759 	q->db_mmap_entry = &entry->rdma_entry;
760 
761 	return 0;
762 
763 err_free_entry:
764 	kfree(entry);
765 
766 err_free_db_data:
767 	free_page((unsigned long)q->db_rec_data);
768 	q->db_rec_data = NULL;
769 	return -ENOMEM;
770 }
771 
772 static inline int qedr_init_user_queue(struct ib_udata *udata,
773 				       struct qedr_dev *dev,
774 				       struct qedr_userq *q, u64 buf_addr,
775 				       size_t buf_len, bool requires_db_rec,
776 				       int access,
777 				       int alloc_and_init)
778 {
779 	u32 fw_pages;
780 	int rc;
781 
782 	q->buf_addr = buf_addr;
783 	q->buf_len = buf_len;
784 	q->umem = ib_umem_get(&dev->ibdev, q->buf_addr, q->buf_len, access);
785 	if (IS_ERR(q->umem)) {
786 		DP_ERR(dev, "create user queue: failed ib_umem_get, got %ld\n",
787 		       PTR_ERR(q->umem));
788 		return PTR_ERR(q->umem);
789 	}
790 
791 	fw_pages = ib_umem_page_count(q->umem) <<
792 	    (PAGE_SHIFT - FW_PAGE_SHIFT);
793 
794 	rc = qedr_prepare_pbl_tbl(dev, &q->pbl_info, fw_pages, 0);
795 	if (rc)
796 		goto err0;
797 
798 	if (alloc_and_init) {
799 		q->pbl_tbl = qedr_alloc_pbl_tbl(dev, &q->pbl_info, GFP_KERNEL);
800 		if (IS_ERR(q->pbl_tbl)) {
801 			rc = PTR_ERR(q->pbl_tbl);
802 			goto err0;
803 		}
804 		qedr_populate_pbls(dev, q->umem, q->pbl_tbl, &q->pbl_info,
805 				   FW_PAGE_SHIFT);
806 	} else {
807 		q->pbl_tbl = kzalloc(sizeof(*q->pbl_tbl), GFP_KERNEL);
808 		if (!q->pbl_tbl) {
809 			rc = -ENOMEM;
810 			goto err0;
811 		}
812 	}
813 
814 	/* mmap the user address used to store doorbell data for recovery */
815 	return qedr_init_user_db_rec(udata, dev, q, requires_db_rec);
816 
817 err0:
818 	ib_umem_release(q->umem);
819 	q->umem = NULL;
820 
821 	return rc;
822 }
823 
824 static inline void qedr_init_cq_params(struct qedr_cq *cq,
825 				       struct qedr_ucontext *ctx,
826 				       struct qedr_dev *dev, int vector,
827 				       int chain_entries, int page_cnt,
828 				       u64 pbl_ptr,
829 				       struct qed_rdma_create_cq_in_params
830 				       *params)
831 {
832 	memset(params, 0, sizeof(*params));
833 	params->cq_handle_hi = upper_32_bits((uintptr_t)cq);
834 	params->cq_handle_lo = lower_32_bits((uintptr_t)cq);
835 	params->cnq_id = vector;
836 	params->cq_size = chain_entries - 1;
837 	params->dpi = (ctx) ? ctx->dpi : dev->dpi;
838 	params->pbl_num_pages = page_cnt;
839 	params->pbl_ptr = pbl_ptr;
840 	params->pbl_two_level = 0;
841 }
842 
843 static void doorbell_cq(struct qedr_cq *cq, u32 cons, u8 flags)
844 {
845 	cq->db.data.agg_flags = flags;
846 	cq->db.data.value = cpu_to_le32(cons);
847 	writeq(cq->db.raw, cq->db_addr);
848 }
849 
850 int qedr_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
851 {
852 	struct qedr_cq *cq = get_qedr_cq(ibcq);
853 	unsigned long sflags;
854 	struct qedr_dev *dev;
855 
856 	dev = get_qedr_dev(ibcq->device);
857 
858 	if (cq->destroyed) {
859 		DP_ERR(dev,
860 		       "warning: arm was invoked after destroy for cq %p (icid=%d)\n",
861 		       cq, cq->icid);
862 		return -EINVAL;
863 	}
864 
865 
866 	if (cq->cq_type == QEDR_CQ_TYPE_GSI)
867 		return 0;
868 
869 	spin_lock_irqsave(&cq->cq_lock, sflags);
870 
871 	cq->arm_flags = 0;
872 
873 	if (flags & IB_CQ_SOLICITED)
874 		cq->arm_flags |= DQ_UCM_ROCE_CQ_ARM_SE_CF_CMD;
875 
876 	if (flags & IB_CQ_NEXT_COMP)
877 		cq->arm_flags |= DQ_UCM_ROCE_CQ_ARM_CF_CMD;
878 
879 	doorbell_cq(cq, cq->cq_cons - 1, cq->arm_flags);
880 
881 	spin_unlock_irqrestore(&cq->cq_lock, sflags);
882 
883 	return 0;
884 }
885 
886 int qedr_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
887 		   struct ib_udata *udata)
888 {
889 	struct ib_device *ibdev = ibcq->device;
890 	struct qedr_ucontext *ctx = rdma_udata_to_drv_context(
891 		udata, struct qedr_ucontext, ibucontext);
892 	struct qed_rdma_destroy_cq_out_params destroy_oparams;
893 	struct qed_rdma_destroy_cq_in_params destroy_iparams;
894 	struct qedr_dev *dev = get_qedr_dev(ibdev);
895 	struct qed_rdma_create_cq_in_params params;
896 	struct qedr_create_cq_ureq ureq = {};
897 	int vector = attr->comp_vector;
898 	int entries = attr->cqe;
899 	struct qedr_cq *cq = get_qedr_cq(ibcq);
900 	int chain_entries;
901 	u32 db_offset;
902 	int page_cnt;
903 	u64 pbl_ptr;
904 	u16 icid;
905 	int rc;
906 
907 	DP_DEBUG(dev, QEDR_MSG_INIT,
908 		 "create_cq: called from %s. entries=%d, vector=%d\n",
909 		 udata ? "User Lib" : "Kernel", entries, vector);
910 
911 	if (entries > QEDR_MAX_CQES) {
912 		DP_ERR(dev,
913 		       "create cq: the number of entries %d is too high. Must be equal or below %d.\n",
914 		       entries, QEDR_MAX_CQES);
915 		return -EINVAL;
916 	}
917 
918 	chain_entries = qedr_align_cq_entries(entries);
919 	chain_entries = min_t(int, chain_entries, QEDR_MAX_CQES);
920 
921 	/* calc db offset. user will add DPI base, kernel will add db addr */
922 	db_offset = DB_ADDR_SHIFT(DQ_PWM_OFFSET_UCM_RDMA_CQ_CONS_32BIT);
923 
924 	if (udata) {
925 		if (ib_copy_from_udata(&ureq, udata, min(sizeof(ureq),
926 							 udata->inlen))) {
927 			DP_ERR(dev,
928 			       "create cq: problem copying data from user space\n");
929 			goto err0;
930 		}
931 
932 		if (!ureq.len) {
933 			DP_ERR(dev,
934 			       "create cq: cannot create a cq with 0 entries\n");
935 			goto err0;
936 		}
937 
938 		cq->cq_type = QEDR_CQ_TYPE_USER;
939 
940 		rc = qedr_init_user_queue(udata, dev, &cq->q, ureq.addr,
941 					  ureq.len, true, IB_ACCESS_LOCAL_WRITE,
942 					  1);
943 		if (rc)
944 			goto err0;
945 
946 		pbl_ptr = cq->q.pbl_tbl->pa;
947 		page_cnt = cq->q.pbl_info.num_pbes;
948 
949 		cq->ibcq.cqe = chain_entries;
950 		cq->q.db_addr = ctx->dpi_addr + db_offset;
951 	} else {
952 		cq->cq_type = QEDR_CQ_TYPE_KERNEL;
953 
954 		rc = dev->ops->common->chain_alloc(dev->cdev,
955 						   QED_CHAIN_USE_TO_CONSUME,
956 						   QED_CHAIN_MODE_PBL,
957 						   QED_CHAIN_CNT_TYPE_U32,
958 						   chain_entries,
959 						   sizeof(union rdma_cqe),
960 						   &cq->pbl, NULL);
961 		if (rc)
962 			goto err0;
963 
964 		page_cnt = qed_chain_get_page_cnt(&cq->pbl);
965 		pbl_ptr = qed_chain_get_pbl_phys(&cq->pbl);
966 		cq->ibcq.cqe = cq->pbl.capacity;
967 	}
968 
969 	qedr_init_cq_params(cq, ctx, dev, vector, chain_entries, page_cnt,
970 			    pbl_ptr, &params);
971 
972 	rc = dev->ops->rdma_create_cq(dev->rdma_ctx, &params, &icid);
973 	if (rc)
974 		goto err1;
975 
976 	cq->icid = icid;
977 	cq->sig = QEDR_CQ_MAGIC_NUMBER;
978 	spin_lock_init(&cq->cq_lock);
979 
980 	if (udata) {
981 		rc = qedr_copy_cq_uresp(dev, cq, udata, db_offset);
982 		if (rc)
983 			goto err2;
984 
985 		rc = qedr_db_recovery_add(dev, cq->q.db_addr,
986 					  &cq->q.db_rec_data->db_data,
987 					  DB_REC_WIDTH_64B,
988 					  DB_REC_USER);
989 		if (rc)
990 			goto err2;
991 
992 	} else {
993 		/* Generate doorbell address. */
994 		cq->db.data.icid = cq->icid;
995 		cq->db_addr = dev->db_addr + db_offset;
996 		cq->db.data.params = DB_AGG_CMD_SET <<
997 		    RDMA_PWM_VAL32_DATA_AGG_CMD_SHIFT;
998 
999 		/* point to the very last element, passing it we will toggle */
1000 		cq->toggle_cqe = qed_chain_get_last_elem(&cq->pbl);
1001 		cq->pbl_toggle = RDMA_CQE_REQUESTER_TOGGLE_BIT_MASK;
1002 		cq->latest_cqe = NULL;
1003 		consume_cqe(cq);
1004 		cq->cq_cons = qed_chain_get_cons_idx_u32(&cq->pbl);
1005 
1006 		rc = qedr_db_recovery_add(dev, cq->db_addr, &cq->db.data,
1007 					  DB_REC_WIDTH_64B, DB_REC_KERNEL);
1008 		if (rc)
1009 			goto err2;
1010 	}
1011 
1012 	DP_DEBUG(dev, QEDR_MSG_CQ,
1013 		 "create cq: icid=0x%0x, addr=%p, size(entries)=0x%0x\n",
1014 		 cq->icid, cq, params.cq_size);
1015 
1016 	return 0;
1017 
1018 err2:
1019 	destroy_iparams.icid = cq->icid;
1020 	dev->ops->rdma_destroy_cq(dev->rdma_ctx, &destroy_iparams,
1021 				  &destroy_oparams);
1022 err1:
1023 	if (udata) {
1024 		qedr_free_pbl(dev, &cq->q.pbl_info, cq->q.pbl_tbl);
1025 		ib_umem_release(cq->q.umem);
1026 		if (cq->q.db_mmap_entry)
1027 			rdma_user_mmap_entry_remove(cq->q.db_mmap_entry);
1028 	} else {
1029 		dev->ops->common->chain_free(dev->cdev, &cq->pbl);
1030 	}
1031 err0:
1032 	return -EINVAL;
1033 }
1034 
1035 int qedr_resize_cq(struct ib_cq *ibcq, int new_cnt, struct ib_udata *udata)
1036 {
1037 	struct qedr_dev *dev = get_qedr_dev(ibcq->device);
1038 	struct qedr_cq *cq = get_qedr_cq(ibcq);
1039 
1040 	DP_ERR(dev, "cq %p RESIZE NOT SUPPORTED\n", cq);
1041 
1042 	return 0;
1043 }
1044 
1045 #define QEDR_DESTROY_CQ_MAX_ITERATIONS		(10)
1046 #define QEDR_DESTROY_CQ_ITER_DURATION		(10)
1047 
1048 void qedr_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
1049 {
1050 	struct qedr_dev *dev = get_qedr_dev(ibcq->device);
1051 	struct qed_rdma_destroy_cq_out_params oparams;
1052 	struct qed_rdma_destroy_cq_in_params iparams;
1053 	struct qedr_cq *cq = get_qedr_cq(ibcq);
1054 	int iter;
1055 
1056 	DP_DEBUG(dev, QEDR_MSG_CQ, "destroy cq %p (icid=%d)\n", cq, cq->icid);
1057 
1058 	cq->destroyed = 1;
1059 
1060 	/* GSIs CQs are handled by driver, so they don't exist in the FW */
1061 	if (cq->cq_type == QEDR_CQ_TYPE_GSI) {
1062 		qedr_db_recovery_del(dev, cq->db_addr, &cq->db.data);
1063 		return;
1064 	}
1065 
1066 	iparams.icid = cq->icid;
1067 	dev->ops->rdma_destroy_cq(dev->rdma_ctx, &iparams, &oparams);
1068 	dev->ops->common->chain_free(dev->cdev, &cq->pbl);
1069 
1070 	if (udata) {
1071 		qedr_free_pbl(dev, &cq->q.pbl_info, cq->q.pbl_tbl);
1072 		ib_umem_release(cq->q.umem);
1073 
1074 		if (cq->q.db_rec_data) {
1075 			qedr_db_recovery_del(dev, cq->q.db_addr,
1076 					     &cq->q.db_rec_data->db_data);
1077 			rdma_user_mmap_entry_remove(cq->q.db_mmap_entry);
1078 		}
1079 	} else {
1080 		qedr_db_recovery_del(dev, cq->db_addr, &cq->db.data);
1081 	}
1082 
1083 	/* We don't want the IRQ handler to handle a non-existing CQ so we
1084 	 * wait until all CNQ interrupts, if any, are received. This will always
1085 	 * happen and will always happen very fast. If not, then a serious error
1086 	 * has occured. That is why we can use a long delay.
1087 	 * We spin for a short time so we don’t lose time on context switching
1088 	 * in case all the completions are handled in that span. Otherwise
1089 	 * we sleep for a while and check again. Since the CNQ may be
1090 	 * associated with (only) the current CPU we use msleep to allow the
1091 	 * current CPU to be freed.
1092 	 * The CNQ notification is increased in qedr_irq_handler().
1093 	 */
1094 	iter = QEDR_DESTROY_CQ_MAX_ITERATIONS;
1095 	while (oparams.num_cq_notif != READ_ONCE(cq->cnq_notif) && iter) {
1096 		udelay(QEDR_DESTROY_CQ_ITER_DURATION);
1097 		iter--;
1098 	}
1099 
1100 	iter = QEDR_DESTROY_CQ_MAX_ITERATIONS;
1101 	while (oparams.num_cq_notif != READ_ONCE(cq->cnq_notif) && iter) {
1102 		msleep(QEDR_DESTROY_CQ_ITER_DURATION);
1103 		iter--;
1104 	}
1105 
1106 	/* Note that we don't need to have explicit code to wait for the
1107 	 * completion of the event handler because it is invoked from the EQ.
1108 	 * Since the destroy CQ ramrod has also been received on the EQ we can
1109 	 * be certain that there's no event handler in process.
1110 	 */
1111 }
1112 
1113 static inline int get_gid_info_from_table(struct ib_qp *ibqp,
1114 					  struct ib_qp_attr *attr,
1115 					  int attr_mask,
1116 					  struct qed_rdma_modify_qp_in_params
1117 					  *qp_params)
1118 {
1119 	const struct ib_gid_attr *gid_attr;
1120 	enum rdma_network_type nw_type;
1121 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
1122 	u32 ipv4_addr;
1123 	int ret;
1124 	int i;
1125 
1126 	gid_attr = grh->sgid_attr;
1127 	ret = rdma_read_gid_l2_fields(gid_attr, &qp_params->vlan_id, NULL);
1128 	if (ret)
1129 		return ret;
1130 
1131 	nw_type = rdma_gid_attr_network_type(gid_attr);
1132 	switch (nw_type) {
1133 	case RDMA_NETWORK_IPV6:
1134 		memcpy(&qp_params->sgid.bytes[0], &gid_attr->gid.raw[0],
1135 		       sizeof(qp_params->sgid));
1136 		memcpy(&qp_params->dgid.bytes[0],
1137 		       &grh->dgid,
1138 		       sizeof(qp_params->dgid));
1139 		qp_params->roce_mode = ROCE_V2_IPV6;
1140 		SET_FIELD(qp_params->modify_flags,
1141 			  QED_ROCE_MODIFY_QP_VALID_ROCE_MODE, 1);
1142 		break;
1143 	case RDMA_NETWORK_IB:
1144 		memcpy(&qp_params->sgid.bytes[0], &gid_attr->gid.raw[0],
1145 		       sizeof(qp_params->sgid));
1146 		memcpy(&qp_params->dgid.bytes[0],
1147 		       &grh->dgid,
1148 		       sizeof(qp_params->dgid));
1149 		qp_params->roce_mode = ROCE_V1;
1150 		break;
1151 	case RDMA_NETWORK_IPV4:
1152 		memset(&qp_params->sgid, 0, sizeof(qp_params->sgid));
1153 		memset(&qp_params->dgid, 0, sizeof(qp_params->dgid));
1154 		ipv4_addr = qedr_get_ipv4_from_gid(gid_attr->gid.raw);
1155 		qp_params->sgid.ipv4_addr = ipv4_addr;
1156 		ipv4_addr =
1157 		    qedr_get_ipv4_from_gid(grh->dgid.raw);
1158 		qp_params->dgid.ipv4_addr = ipv4_addr;
1159 		SET_FIELD(qp_params->modify_flags,
1160 			  QED_ROCE_MODIFY_QP_VALID_ROCE_MODE, 1);
1161 		qp_params->roce_mode = ROCE_V2_IPV4;
1162 		break;
1163 	}
1164 
1165 	for (i = 0; i < 4; i++) {
1166 		qp_params->sgid.dwords[i] = ntohl(qp_params->sgid.dwords[i]);
1167 		qp_params->dgid.dwords[i] = ntohl(qp_params->dgid.dwords[i]);
1168 	}
1169 
1170 	if (qp_params->vlan_id >= VLAN_CFI_MASK)
1171 		qp_params->vlan_id = 0;
1172 
1173 	return 0;
1174 }
1175 
1176 static int qedr_check_qp_attrs(struct ib_pd *ibpd, struct qedr_dev *dev,
1177 			       struct ib_qp_init_attr *attrs,
1178 			       struct ib_udata *udata)
1179 {
1180 	struct qedr_device_attr *qattr = &dev->attr;
1181 
1182 	/* QP0... attrs->qp_type == IB_QPT_GSI */
1183 	if (attrs->qp_type != IB_QPT_RC && attrs->qp_type != IB_QPT_GSI) {
1184 		DP_DEBUG(dev, QEDR_MSG_QP,
1185 			 "create qp: unsupported qp type=0x%x requested\n",
1186 			 attrs->qp_type);
1187 		return -EOPNOTSUPP;
1188 	}
1189 
1190 	if (attrs->cap.max_send_wr > qattr->max_sqe) {
1191 		DP_ERR(dev,
1192 		       "create qp: cannot create a SQ with %d elements (max_send_wr=0x%x)\n",
1193 		       attrs->cap.max_send_wr, qattr->max_sqe);
1194 		return -EINVAL;
1195 	}
1196 
1197 	if (attrs->cap.max_inline_data > qattr->max_inline) {
1198 		DP_ERR(dev,
1199 		       "create qp: unsupported inline data size=0x%x requested (max_inline=0x%x)\n",
1200 		       attrs->cap.max_inline_data, qattr->max_inline);
1201 		return -EINVAL;
1202 	}
1203 
1204 	if (attrs->cap.max_send_sge > qattr->max_sge) {
1205 		DP_ERR(dev,
1206 		       "create qp: unsupported send_sge=0x%x requested (max_send_sge=0x%x)\n",
1207 		       attrs->cap.max_send_sge, qattr->max_sge);
1208 		return -EINVAL;
1209 	}
1210 
1211 	if (attrs->cap.max_recv_sge > qattr->max_sge) {
1212 		DP_ERR(dev,
1213 		       "create qp: unsupported recv_sge=0x%x requested (max_recv_sge=0x%x)\n",
1214 		       attrs->cap.max_recv_sge, qattr->max_sge);
1215 		return -EINVAL;
1216 	}
1217 
1218 	/* Unprivileged user space cannot create special QP */
1219 	if (udata && attrs->qp_type == IB_QPT_GSI) {
1220 		DP_ERR(dev,
1221 		       "create qp: userspace can't create special QPs of type=0x%x\n",
1222 		       attrs->qp_type);
1223 		return -EINVAL;
1224 	}
1225 
1226 	return 0;
1227 }
1228 
1229 static int qedr_copy_srq_uresp(struct qedr_dev *dev,
1230 			       struct qedr_srq *srq, struct ib_udata *udata)
1231 {
1232 	struct qedr_create_srq_uresp uresp = {};
1233 	int rc;
1234 
1235 	uresp.srq_id = srq->srq_id;
1236 
1237 	rc = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1238 	if (rc)
1239 		DP_ERR(dev, "create srq: problem copying data to user space\n");
1240 
1241 	return rc;
1242 }
1243 
1244 static void qedr_copy_rq_uresp(struct qedr_dev *dev,
1245 			      struct qedr_create_qp_uresp *uresp,
1246 			      struct qedr_qp *qp)
1247 {
1248 	/* iWARP requires two doorbells per RQ. */
1249 	if (rdma_protocol_iwarp(&dev->ibdev, 1)) {
1250 		uresp->rq_db_offset =
1251 		    DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_IWARP_RQ_PROD);
1252 		uresp->rq_db2_offset = DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_FLAGS);
1253 	} else {
1254 		uresp->rq_db_offset =
1255 		    DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_ROCE_RQ_PROD);
1256 	}
1257 
1258 	uresp->rq_icid = qp->icid;
1259 	if (qp->urq.db_mmap_entry)
1260 		uresp->rq_db_rec_addr =
1261 			rdma_user_mmap_get_offset(qp->urq.db_mmap_entry);
1262 }
1263 
1264 static void qedr_copy_sq_uresp(struct qedr_dev *dev,
1265 			       struct qedr_create_qp_uresp *uresp,
1266 			       struct qedr_qp *qp)
1267 {
1268 	uresp->sq_db_offset = DB_ADDR_SHIFT(DQ_PWM_OFFSET_XCM_RDMA_SQ_PROD);
1269 
1270 	/* iWARP uses the same cid for rq and sq */
1271 	if (rdma_protocol_iwarp(&dev->ibdev, 1))
1272 		uresp->sq_icid = qp->icid;
1273 	else
1274 		uresp->sq_icid = qp->icid + 1;
1275 
1276 	if (qp->usq.db_mmap_entry)
1277 		uresp->sq_db_rec_addr =
1278 			rdma_user_mmap_get_offset(qp->usq.db_mmap_entry);
1279 }
1280 
1281 static int qedr_copy_qp_uresp(struct qedr_dev *dev,
1282 			      struct qedr_qp *qp, struct ib_udata *udata,
1283 			      struct qedr_create_qp_uresp *uresp)
1284 {
1285 	int rc;
1286 
1287 	memset(uresp, 0, sizeof(*uresp));
1288 	qedr_copy_sq_uresp(dev, uresp, qp);
1289 	qedr_copy_rq_uresp(dev, uresp, qp);
1290 
1291 	uresp->atomic_supported = dev->atomic_cap != IB_ATOMIC_NONE;
1292 	uresp->qp_id = qp->qp_id;
1293 
1294 	rc = qedr_ib_copy_to_udata(udata, uresp, sizeof(*uresp));
1295 	if (rc)
1296 		DP_ERR(dev,
1297 		       "create qp: failed a copy to user space with qp icid=0x%x.\n",
1298 		       qp->icid);
1299 
1300 	return rc;
1301 }
1302 
1303 static void qedr_set_common_qp_params(struct qedr_dev *dev,
1304 				      struct qedr_qp *qp,
1305 				      struct qedr_pd *pd,
1306 				      struct ib_qp_init_attr *attrs)
1307 {
1308 	spin_lock_init(&qp->q_lock);
1309 	if (rdma_protocol_iwarp(&dev->ibdev, 1)) {
1310 		kref_init(&qp->refcnt);
1311 		init_completion(&qp->iwarp_cm_comp);
1312 	}
1313 	qp->pd = pd;
1314 	qp->qp_type = attrs->qp_type;
1315 	qp->max_inline_data = attrs->cap.max_inline_data;
1316 	qp->sq.max_sges = attrs->cap.max_send_sge;
1317 	qp->state = QED_ROCE_QP_STATE_RESET;
1318 	qp->signaled = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR) ? true : false;
1319 	qp->sq_cq = get_qedr_cq(attrs->send_cq);
1320 	qp->dev = dev;
1321 
1322 	if (attrs->srq) {
1323 		qp->srq = get_qedr_srq(attrs->srq);
1324 	} else {
1325 		qp->rq_cq = get_qedr_cq(attrs->recv_cq);
1326 		qp->rq.max_sges = attrs->cap.max_recv_sge;
1327 		DP_DEBUG(dev, QEDR_MSG_QP,
1328 			 "RQ params:\trq_max_sges = %d, rq_cq_id = %d\n",
1329 			 qp->rq.max_sges, qp->rq_cq->icid);
1330 	}
1331 
1332 	DP_DEBUG(dev, QEDR_MSG_QP,
1333 		 "QP params:\tpd = %d, qp_type = %d, max_inline_data = %d, state = %d, signaled = %d, use_srq=%d\n",
1334 		 pd->pd_id, qp->qp_type, qp->max_inline_data,
1335 		 qp->state, qp->signaled, (attrs->srq) ? 1 : 0);
1336 	DP_DEBUG(dev, QEDR_MSG_QP,
1337 		 "SQ params:\tsq_max_sges = %d, sq_cq_id = %d\n",
1338 		 qp->sq.max_sges, qp->sq_cq->icid);
1339 }
1340 
1341 static int qedr_set_roce_db_info(struct qedr_dev *dev, struct qedr_qp *qp)
1342 {
1343 	int rc;
1344 
1345 	qp->sq.db = dev->db_addr +
1346 		    DB_ADDR_SHIFT(DQ_PWM_OFFSET_XCM_RDMA_SQ_PROD);
1347 	qp->sq.db_data.data.icid = qp->icid + 1;
1348 	rc = qedr_db_recovery_add(dev, qp->sq.db,
1349 				  &qp->sq.db_data,
1350 				  DB_REC_WIDTH_32B,
1351 				  DB_REC_KERNEL);
1352 	if (rc)
1353 		return rc;
1354 
1355 	if (!qp->srq) {
1356 		qp->rq.db = dev->db_addr +
1357 			    DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_ROCE_RQ_PROD);
1358 		qp->rq.db_data.data.icid = qp->icid;
1359 
1360 		rc = qedr_db_recovery_add(dev, qp->rq.db,
1361 					  &qp->rq.db_data,
1362 					  DB_REC_WIDTH_32B,
1363 					  DB_REC_KERNEL);
1364 		if (rc)
1365 			qedr_db_recovery_del(dev, qp->sq.db,
1366 					     &qp->sq.db_data);
1367 	}
1368 
1369 	return rc;
1370 }
1371 
1372 static int qedr_check_srq_params(struct qedr_dev *dev,
1373 				 struct ib_srq_init_attr *attrs,
1374 				 struct ib_udata *udata)
1375 {
1376 	struct qedr_device_attr *qattr = &dev->attr;
1377 
1378 	if (attrs->attr.max_wr > qattr->max_srq_wr) {
1379 		DP_ERR(dev,
1380 		       "create srq: unsupported srq_wr=0x%x requested (max_srq_wr=0x%x)\n",
1381 		       attrs->attr.max_wr, qattr->max_srq_wr);
1382 		return -EINVAL;
1383 	}
1384 
1385 	if (attrs->attr.max_sge > qattr->max_sge) {
1386 		DP_ERR(dev,
1387 		       "create srq: unsupported sge=0x%x requested (max_srq_sge=0x%x)\n",
1388 		       attrs->attr.max_sge, qattr->max_sge);
1389 		return -EINVAL;
1390 	}
1391 
1392 	return 0;
1393 }
1394 
1395 static void qedr_free_srq_user_params(struct qedr_srq *srq)
1396 {
1397 	qedr_free_pbl(srq->dev, &srq->usrq.pbl_info, srq->usrq.pbl_tbl);
1398 	ib_umem_release(srq->usrq.umem);
1399 	ib_umem_release(srq->prod_umem);
1400 }
1401 
1402 static void qedr_free_srq_kernel_params(struct qedr_srq *srq)
1403 {
1404 	struct qedr_srq_hwq_info *hw_srq = &srq->hw_srq;
1405 	struct qedr_dev *dev = srq->dev;
1406 
1407 	dev->ops->common->chain_free(dev->cdev, &hw_srq->pbl);
1408 
1409 	dma_free_coherent(&dev->pdev->dev, sizeof(struct rdma_srq_producers),
1410 			  hw_srq->virt_prod_pair_addr,
1411 			  hw_srq->phy_prod_pair_addr);
1412 }
1413 
1414 static int qedr_init_srq_user_params(struct ib_udata *udata,
1415 				     struct qedr_srq *srq,
1416 				     struct qedr_create_srq_ureq *ureq,
1417 				     int access)
1418 {
1419 	struct scatterlist *sg;
1420 	int rc;
1421 
1422 	rc = qedr_init_user_queue(udata, srq->dev, &srq->usrq, ureq->srq_addr,
1423 				  ureq->srq_len, false, access, 1);
1424 	if (rc)
1425 		return rc;
1426 
1427 	srq->prod_umem = ib_umem_get(srq->ibsrq.device, ureq->prod_pair_addr,
1428 				     sizeof(struct rdma_srq_producers), access);
1429 	if (IS_ERR(srq->prod_umem)) {
1430 		qedr_free_pbl(srq->dev, &srq->usrq.pbl_info, srq->usrq.pbl_tbl);
1431 		ib_umem_release(srq->usrq.umem);
1432 		DP_ERR(srq->dev,
1433 		       "create srq: failed ib_umem_get for producer, got %ld\n",
1434 		       PTR_ERR(srq->prod_umem));
1435 		return PTR_ERR(srq->prod_umem);
1436 	}
1437 
1438 	sg = srq->prod_umem->sg_head.sgl;
1439 	srq->hw_srq.phy_prod_pair_addr = sg_dma_address(sg);
1440 
1441 	return 0;
1442 }
1443 
1444 static int qedr_alloc_srq_kernel_params(struct qedr_srq *srq,
1445 					struct qedr_dev *dev,
1446 					struct ib_srq_init_attr *init_attr)
1447 {
1448 	struct qedr_srq_hwq_info *hw_srq = &srq->hw_srq;
1449 	dma_addr_t phy_prod_pair_addr;
1450 	u32 num_elems;
1451 	void *va;
1452 	int rc;
1453 
1454 	va = dma_alloc_coherent(&dev->pdev->dev,
1455 				sizeof(struct rdma_srq_producers),
1456 				&phy_prod_pair_addr, GFP_KERNEL);
1457 	if (!va) {
1458 		DP_ERR(dev,
1459 		       "create srq: failed to allocate dma memory for producer\n");
1460 		return -ENOMEM;
1461 	}
1462 
1463 	hw_srq->phy_prod_pair_addr = phy_prod_pair_addr;
1464 	hw_srq->virt_prod_pair_addr = va;
1465 
1466 	num_elems = init_attr->attr.max_wr * RDMA_MAX_SRQ_WQE_SIZE;
1467 	rc = dev->ops->common->chain_alloc(dev->cdev,
1468 					   QED_CHAIN_USE_TO_CONSUME_PRODUCE,
1469 					   QED_CHAIN_MODE_PBL,
1470 					   QED_CHAIN_CNT_TYPE_U32,
1471 					   num_elems,
1472 					   QEDR_SRQ_WQE_ELEM_SIZE,
1473 					   &hw_srq->pbl, NULL);
1474 	if (rc)
1475 		goto err0;
1476 
1477 	hw_srq->num_elems = num_elems;
1478 
1479 	return 0;
1480 
1481 err0:
1482 	dma_free_coherent(&dev->pdev->dev, sizeof(struct rdma_srq_producers),
1483 			  va, phy_prod_pair_addr);
1484 	return rc;
1485 }
1486 
1487 int qedr_create_srq(struct ib_srq *ibsrq, struct ib_srq_init_attr *init_attr,
1488 		    struct ib_udata *udata)
1489 {
1490 	struct qed_rdma_destroy_srq_in_params destroy_in_params;
1491 	struct qed_rdma_create_srq_in_params in_params = {};
1492 	struct qedr_dev *dev = get_qedr_dev(ibsrq->device);
1493 	struct qed_rdma_create_srq_out_params out_params;
1494 	struct qedr_pd *pd = get_qedr_pd(ibsrq->pd);
1495 	struct qedr_create_srq_ureq ureq = {};
1496 	u64 pbl_base_addr, phy_prod_pair_addr;
1497 	struct qedr_srq_hwq_info *hw_srq;
1498 	u32 page_cnt, page_size;
1499 	struct qedr_srq *srq = get_qedr_srq(ibsrq);
1500 	int rc = 0;
1501 
1502 	DP_DEBUG(dev, QEDR_MSG_QP,
1503 		 "create SRQ called from %s (pd %p)\n",
1504 		 (udata) ? "User lib" : "kernel", pd);
1505 
1506 	rc = qedr_check_srq_params(dev, init_attr, udata);
1507 	if (rc)
1508 		return -EINVAL;
1509 
1510 	srq->dev = dev;
1511 	hw_srq = &srq->hw_srq;
1512 	spin_lock_init(&srq->lock);
1513 
1514 	hw_srq->max_wr = init_attr->attr.max_wr;
1515 	hw_srq->max_sges = init_attr->attr.max_sge;
1516 
1517 	if (udata) {
1518 		if (ib_copy_from_udata(&ureq, udata, min(sizeof(ureq),
1519 							 udata->inlen))) {
1520 			DP_ERR(dev,
1521 			       "create srq: problem copying data from user space\n");
1522 			goto err0;
1523 		}
1524 
1525 		rc = qedr_init_srq_user_params(udata, srq, &ureq, 0);
1526 		if (rc)
1527 			goto err0;
1528 
1529 		page_cnt = srq->usrq.pbl_info.num_pbes;
1530 		pbl_base_addr = srq->usrq.pbl_tbl->pa;
1531 		phy_prod_pair_addr = hw_srq->phy_prod_pair_addr;
1532 		page_size = PAGE_SIZE;
1533 	} else {
1534 		struct qed_chain *pbl;
1535 
1536 		rc = qedr_alloc_srq_kernel_params(srq, dev, init_attr);
1537 		if (rc)
1538 			goto err0;
1539 
1540 		pbl = &hw_srq->pbl;
1541 		page_cnt = qed_chain_get_page_cnt(pbl);
1542 		pbl_base_addr = qed_chain_get_pbl_phys(pbl);
1543 		phy_prod_pair_addr = hw_srq->phy_prod_pair_addr;
1544 		page_size = QED_CHAIN_PAGE_SIZE;
1545 	}
1546 
1547 	in_params.pd_id = pd->pd_id;
1548 	in_params.pbl_base_addr = pbl_base_addr;
1549 	in_params.prod_pair_addr = phy_prod_pair_addr;
1550 	in_params.num_pages = page_cnt;
1551 	in_params.page_size = page_size;
1552 
1553 	rc = dev->ops->rdma_create_srq(dev->rdma_ctx, &in_params, &out_params);
1554 	if (rc)
1555 		goto err1;
1556 
1557 	srq->srq_id = out_params.srq_id;
1558 
1559 	if (udata) {
1560 		rc = qedr_copy_srq_uresp(dev, srq, udata);
1561 		if (rc)
1562 			goto err2;
1563 	}
1564 
1565 	rc = xa_insert_irq(&dev->srqs, srq->srq_id, srq, GFP_KERNEL);
1566 	if (rc)
1567 		goto err2;
1568 
1569 	DP_DEBUG(dev, QEDR_MSG_SRQ,
1570 		 "create srq: created srq with srq_id=0x%0x\n", srq->srq_id);
1571 	return 0;
1572 
1573 err2:
1574 	destroy_in_params.srq_id = srq->srq_id;
1575 
1576 	dev->ops->rdma_destroy_srq(dev->rdma_ctx, &destroy_in_params);
1577 err1:
1578 	if (udata)
1579 		qedr_free_srq_user_params(srq);
1580 	else
1581 		qedr_free_srq_kernel_params(srq);
1582 err0:
1583 	return -EFAULT;
1584 }
1585 
1586 void qedr_destroy_srq(struct ib_srq *ibsrq, struct ib_udata *udata)
1587 {
1588 	struct qed_rdma_destroy_srq_in_params in_params = {};
1589 	struct qedr_dev *dev = get_qedr_dev(ibsrq->device);
1590 	struct qedr_srq *srq = get_qedr_srq(ibsrq);
1591 
1592 	xa_erase_irq(&dev->srqs, srq->srq_id);
1593 	in_params.srq_id = srq->srq_id;
1594 	dev->ops->rdma_destroy_srq(dev->rdma_ctx, &in_params);
1595 
1596 	if (ibsrq->uobject)
1597 		qedr_free_srq_user_params(srq);
1598 	else
1599 		qedr_free_srq_kernel_params(srq);
1600 
1601 	DP_DEBUG(dev, QEDR_MSG_SRQ,
1602 		 "destroy srq: destroyed srq with srq_id=0x%0x\n",
1603 		 srq->srq_id);
1604 }
1605 
1606 int qedr_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
1607 		    enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
1608 {
1609 	struct qed_rdma_modify_srq_in_params in_params = {};
1610 	struct qedr_dev *dev = get_qedr_dev(ibsrq->device);
1611 	struct qedr_srq *srq = get_qedr_srq(ibsrq);
1612 	int rc;
1613 
1614 	if (attr_mask & IB_SRQ_MAX_WR) {
1615 		DP_ERR(dev,
1616 		       "modify srq: invalid attribute mask=0x%x specified for %p\n",
1617 		       attr_mask, srq);
1618 		return -EINVAL;
1619 	}
1620 
1621 	if (attr_mask & IB_SRQ_LIMIT) {
1622 		if (attr->srq_limit >= srq->hw_srq.max_wr) {
1623 			DP_ERR(dev,
1624 			       "modify srq: invalid srq_limit=0x%x (max_srq_limit=0x%x)\n",
1625 			       attr->srq_limit, srq->hw_srq.max_wr);
1626 			return -EINVAL;
1627 		}
1628 
1629 		in_params.srq_id = srq->srq_id;
1630 		in_params.wqe_limit = attr->srq_limit;
1631 		rc = dev->ops->rdma_modify_srq(dev->rdma_ctx, &in_params);
1632 		if (rc)
1633 			return rc;
1634 	}
1635 
1636 	srq->srq_limit = attr->srq_limit;
1637 
1638 	DP_DEBUG(dev, QEDR_MSG_SRQ,
1639 		 "modify srq: modified srq with srq_id=0x%0x\n", srq->srq_id);
1640 
1641 	return 0;
1642 }
1643 
1644 static inline void
1645 qedr_init_common_qp_in_params(struct qedr_dev *dev,
1646 			      struct qedr_pd *pd,
1647 			      struct qedr_qp *qp,
1648 			      struct ib_qp_init_attr *attrs,
1649 			      bool fmr_and_reserved_lkey,
1650 			      struct qed_rdma_create_qp_in_params *params)
1651 {
1652 	/* QP handle to be written in an async event */
1653 	params->qp_handle_async_lo = lower_32_bits((uintptr_t) qp);
1654 	params->qp_handle_async_hi = upper_32_bits((uintptr_t) qp);
1655 
1656 	params->signal_all = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR);
1657 	params->fmr_and_reserved_lkey = fmr_and_reserved_lkey;
1658 	params->pd = pd->pd_id;
1659 	params->dpi = pd->uctx ? pd->uctx->dpi : dev->dpi;
1660 	params->sq_cq_id = get_qedr_cq(attrs->send_cq)->icid;
1661 	params->stats_queue = 0;
1662 	params->srq_id = 0;
1663 	params->use_srq = false;
1664 
1665 	if (!qp->srq) {
1666 		params->rq_cq_id = get_qedr_cq(attrs->recv_cq)->icid;
1667 
1668 	} else {
1669 		params->rq_cq_id = get_qedr_cq(attrs->recv_cq)->icid;
1670 		params->srq_id = qp->srq->srq_id;
1671 		params->use_srq = true;
1672 	}
1673 }
1674 
1675 static inline void qedr_qp_user_print(struct qedr_dev *dev, struct qedr_qp *qp)
1676 {
1677 	DP_DEBUG(dev, QEDR_MSG_QP, "create qp: successfully created user QP. "
1678 		 "qp=%p. "
1679 		 "sq_addr=0x%llx, "
1680 		 "sq_len=%zd, "
1681 		 "rq_addr=0x%llx, "
1682 		 "rq_len=%zd"
1683 		 "\n",
1684 		 qp,
1685 		 qp->usq.buf_addr,
1686 		 qp->usq.buf_len, qp->urq.buf_addr, qp->urq.buf_len);
1687 }
1688 
1689 static inline void
1690 qedr_iwarp_populate_user_qp(struct qedr_dev *dev,
1691 			    struct qedr_qp *qp,
1692 			    struct qed_rdma_create_qp_out_params *out_params)
1693 {
1694 	qp->usq.pbl_tbl->va = out_params->sq_pbl_virt;
1695 	qp->usq.pbl_tbl->pa = out_params->sq_pbl_phys;
1696 
1697 	qedr_populate_pbls(dev, qp->usq.umem, qp->usq.pbl_tbl,
1698 			   &qp->usq.pbl_info, FW_PAGE_SHIFT);
1699 	if (!qp->srq) {
1700 		qp->urq.pbl_tbl->va = out_params->rq_pbl_virt;
1701 		qp->urq.pbl_tbl->pa = out_params->rq_pbl_phys;
1702 	}
1703 
1704 	qedr_populate_pbls(dev, qp->urq.umem, qp->urq.pbl_tbl,
1705 			   &qp->urq.pbl_info, FW_PAGE_SHIFT);
1706 }
1707 
1708 static void qedr_cleanup_user(struct qedr_dev *dev,
1709 			      struct qedr_ucontext *ctx,
1710 			      struct qedr_qp *qp)
1711 {
1712 	ib_umem_release(qp->usq.umem);
1713 	qp->usq.umem = NULL;
1714 
1715 	ib_umem_release(qp->urq.umem);
1716 	qp->urq.umem = NULL;
1717 
1718 	if (rdma_protocol_roce(&dev->ibdev, 1)) {
1719 		qedr_free_pbl(dev, &qp->usq.pbl_info, qp->usq.pbl_tbl);
1720 		qedr_free_pbl(dev, &qp->urq.pbl_info, qp->urq.pbl_tbl);
1721 	} else {
1722 		kfree(qp->usq.pbl_tbl);
1723 		kfree(qp->urq.pbl_tbl);
1724 	}
1725 
1726 	if (qp->usq.db_rec_data) {
1727 		qedr_db_recovery_del(dev, qp->usq.db_addr,
1728 				     &qp->usq.db_rec_data->db_data);
1729 		rdma_user_mmap_entry_remove(qp->usq.db_mmap_entry);
1730 	}
1731 
1732 	if (qp->urq.db_rec_data) {
1733 		qedr_db_recovery_del(dev, qp->urq.db_addr,
1734 				     &qp->urq.db_rec_data->db_data);
1735 		rdma_user_mmap_entry_remove(qp->urq.db_mmap_entry);
1736 	}
1737 
1738 	if (rdma_protocol_iwarp(&dev->ibdev, 1))
1739 		qedr_db_recovery_del(dev, qp->urq.db_rec_db2_addr,
1740 				     &qp->urq.db_rec_db2_data);
1741 }
1742 
1743 static int qedr_create_user_qp(struct qedr_dev *dev,
1744 			       struct qedr_qp *qp,
1745 			       struct ib_pd *ibpd,
1746 			       struct ib_udata *udata,
1747 			       struct ib_qp_init_attr *attrs)
1748 {
1749 	struct qed_rdma_create_qp_in_params in_params;
1750 	struct qed_rdma_create_qp_out_params out_params;
1751 	struct qedr_pd *pd = get_qedr_pd(ibpd);
1752 	struct qedr_create_qp_uresp uresp;
1753 	struct qedr_ucontext *ctx = NULL;
1754 	struct qedr_create_qp_ureq ureq;
1755 	int alloc_and_init = rdma_protocol_roce(&dev->ibdev, 1);
1756 	int rc = -EINVAL;
1757 
1758 	qp->create_type = QEDR_QP_CREATE_USER;
1759 	memset(&ureq, 0, sizeof(ureq));
1760 	rc = ib_copy_from_udata(&ureq, udata, min(sizeof(ureq), udata->inlen));
1761 	if (rc) {
1762 		DP_ERR(dev, "Problem copying data from user space\n");
1763 		return rc;
1764 	}
1765 
1766 	/* SQ - read access only (0) */
1767 	rc = qedr_init_user_queue(udata, dev, &qp->usq, ureq.sq_addr,
1768 				  ureq.sq_len, true, 0, alloc_and_init);
1769 	if (rc)
1770 		return rc;
1771 
1772 	if (!qp->srq) {
1773 		/* RQ - read access only (0) */
1774 		rc = qedr_init_user_queue(udata, dev, &qp->urq, ureq.rq_addr,
1775 					  ureq.rq_len, true, 0, alloc_and_init);
1776 		if (rc)
1777 			return rc;
1778 	}
1779 
1780 	memset(&in_params, 0, sizeof(in_params));
1781 	qedr_init_common_qp_in_params(dev, pd, qp, attrs, false, &in_params);
1782 	in_params.qp_handle_lo = ureq.qp_handle_lo;
1783 	in_params.qp_handle_hi = ureq.qp_handle_hi;
1784 	in_params.sq_num_pages = qp->usq.pbl_info.num_pbes;
1785 	in_params.sq_pbl_ptr = qp->usq.pbl_tbl->pa;
1786 	if (!qp->srq) {
1787 		in_params.rq_num_pages = qp->urq.pbl_info.num_pbes;
1788 		in_params.rq_pbl_ptr = qp->urq.pbl_tbl->pa;
1789 	}
1790 
1791 	qp->qed_qp = dev->ops->rdma_create_qp(dev->rdma_ctx,
1792 					      &in_params, &out_params);
1793 
1794 	if (!qp->qed_qp) {
1795 		rc = -ENOMEM;
1796 		goto err1;
1797 	}
1798 
1799 	if (rdma_protocol_iwarp(&dev->ibdev, 1))
1800 		qedr_iwarp_populate_user_qp(dev, qp, &out_params);
1801 
1802 	qp->qp_id = out_params.qp_id;
1803 	qp->icid = out_params.icid;
1804 
1805 	rc = qedr_copy_qp_uresp(dev, qp, udata, &uresp);
1806 	if (rc)
1807 		goto err;
1808 
1809 	/* db offset was calculated in copy_qp_uresp, now set in the user q */
1810 	ctx = pd->uctx;
1811 	qp->usq.db_addr = ctx->dpi_addr + uresp.sq_db_offset;
1812 	qp->urq.db_addr = ctx->dpi_addr + uresp.rq_db_offset;
1813 
1814 	if (rdma_protocol_iwarp(&dev->ibdev, 1)) {
1815 		qp->urq.db_rec_db2_addr = ctx->dpi_addr + uresp.rq_db2_offset;
1816 
1817 		/* calculate the db_rec_db2 data since it is constant so no
1818 		 *  need to reflect from user
1819 		 */
1820 		qp->urq.db_rec_db2_data.data.icid = cpu_to_le16(qp->icid);
1821 		qp->urq.db_rec_db2_data.data.value =
1822 			cpu_to_le16(DQ_TCM_IWARP_POST_RQ_CF_CMD);
1823 	}
1824 
1825 	rc = qedr_db_recovery_add(dev, qp->usq.db_addr,
1826 				  &qp->usq.db_rec_data->db_data,
1827 				  DB_REC_WIDTH_32B,
1828 				  DB_REC_USER);
1829 	if (rc)
1830 		goto err;
1831 
1832 	rc = qedr_db_recovery_add(dev, qp->urq.db_addr,
1833 				  &qp->urq.db_rec_data->db_data,
1834 				  DB_REC_WIDTH_32B,
1835 				  DB_REC_USER);
1836 	if (rc)
1837 		goto err;
1838 
1839 	if (rdma_protocol_iwarp(&dev->ibdev, 1)) {
1840 		rc = qedr_db_recovery_add(dev, qp->urq.db_rec_db2_addr,
1841 					  &qp->urq.db_rec_db2_data,
1842 					  DB_REC_WIDTH_32B,
1843 					  DB_REC_USER);
1844 		if (rc)
1845 			goto err;
1846 	}
1847 	qedr_qp_user_print(dev, qp);
1848 
1849 	return rc;
1850 err:
1851 	rc = dev->ops->rdma_destroy_qp(dev->rdma_ctx, qp->qed_qp);
1852 	if (rc)
1853 		DP_ERR(dev, "create qp: fatal fault. rc=%d", rc);
1854 
1855 err1:
1856 	qedr_cleanup_user(dev, ctx, qp);
1857 	return rc;
1858 }
1859 
1860 static int qedr_set_iwarp_db_info(struct qedr_dev *dev, struct qedr_qp *qp)
1861 {
1862 	int rc;
1863 
1864 	qp->sq.db = dev->db_addr +
1865 	    DB_ADDR_SHIFT(DQ_PWM_OFFSET_XCM_RDMA_SQ_PROD);
1866 	qp->sq.db_data.data.icid = qp->icid;
1867 
1868 	rc = qedr_db_recovery_add(dev, qp->sq.db,
1869 				  &qp->sq.db_data,
1870 				  DB_REC_WIDTH_32B,
1871 				  DB_REC_KERNEL);
1872 	if (rc)
1873 		return rc;
1874 
1875 	qp->rq.db = dev->db_addr +
1876 		    DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_IWARP_RQ_PROD);
1877 	qp->rq.db_data.data.icid = qp->icid;
1878 	qp->rq.iwarp_db2 = dev->db_addr +
1879 			   DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_FLAGS);
1880 	qp->rq.iwarp_db2_data.data.icid = qp->icid;
1881 	qp->rq.iwarp_db2_data.data.value = DQ_TCM_IWARP_POST_RQ_CF_CMD;
1882 
1883 	rc = qedr_db_recovery_add(dev, qp->rq.db,
1884 				  &qp->rq.db_data,
1885 				  DB_REC_WIDTH_32B,
1886 				  DB_REC_KERNEL);
1887 	if (rc)
1888 		return rc;
1889 
1890 	rc = qedr_db_recovery_add(dev, qp->rq.iwarp_db2,
1891 				  &qp->rq.iwarp_db2_data,
1892 				  DB_REC_WIDTH_32B,
1893 				  DB_REC_KERNEL);
1894 	return rc;
1895 }
1896 
1897 static int
1898 qedr_roce_create_kernel_qp(struct qedr_dev *dev,
1899 			   struct qedr_qp *qp,
1900 			   struct qed_rdma_create_qp_in_params *in_params,
1901 			   u32 n_sq_elems, u32 n_rq_elems)
1902 {
1903 	struct qed_rdma_create_qp_out_params out_params;
1904 	int rc;
1905 
1906 	rc = dev->ops->common->chain_alloc(dev->cdev,
1907 					   QED_CHAIN_USE_TO_PRODUCE,
1908 					   QED_CHAIN_MODE_PBL,
1909 					   QED_CHAIN_CNT_TYPE_U32,
1910 					   n_sq_elems,
1911 					   QEDR_SQE_ELEMENT_SIZE,
1912 					   &qp->sq.pbl, NULL);
1913 
1914 	if (rc)
1915 		return rc;
1916 
1917 	in_params->sq_num_pages = qed_chain_get_page_cnt(&qp->sq.pbl);
1918 	in_params->sq_pbl_ptr = qed_chain_get_pbl_phys(&qp->sq.pbl);
1919 
1920 	rc = dev->ops->common->chain_alloc(dev->cdev,
1921 					   QED_CHAIN_USE_TO_CONSUME_PRODUCE,
1922 					   QED_CHAIN_MODE_PBL,
1923 					   QED_CHAIN_CNT_TYPE_U32,
1924 					   n_rq_elems,
1925 					   QEDR_RQE_ELEMENT_SIZE,
1926 					   &qp->rq.pbl, NULL);
1927 	if (rc)
1928 		return rc;
1929 
1930 	in_params->rq_num_pages = qed_chain_get_page_cnt(&qp->rq.pbl);
1931 	in_params->rq_pbl_ptr = qed_chain_get_pbl_phys(&qp->rq.pbl);
1932 
1933 	qp->qed_qp = dev->ops->rdma_create_qp(dev->rdma_ctx,
1934 					      in_params, &out_params);
1935 
1936 	if (!qp->qed_qp)
1937 		return -EINVAL;
1938 
1939 	qp->qp_id = out_params.qp_id;
1940 	qp->icid = out_params.icid;
1941 
1942 	return qedr_set_roce_db_info(dev, qp);
1943 }
1944 
1945 static int
1946 qedr_iwarp_create_kernel_qp(struct qedr_dev *dev,
1947 			    struct qedr_qp *qp,
1948 			    struct qed_rdma_create_qp_in_params *in_params,
1949 			    u32 n_sq_elems, u32 n_rq_elems)
1950 {
1951 	struct qed_rdma_create_qp_out_params out_params;
1952 	struct qed_chain_ext_pbl ext_pbl;
1953 	int rc;
1954 
1955 	in_params->sq_num_pages = QED_CHAIN_PAGE_CNT(n_sq_elems,
1956 						     QEDR_SQE_ELEMENT_SIZE,
1957 						     QED_CHAIN_MODE_PBL);
1958 	in_params->rq_num_pages = QED_CHAIN_PAGE_CNT(n_rq_elems,
1959 						     QEDR_RQE_ELEMENT_SIZE,
1960 						     QED_CHAIN_MODE_PBL);
1961 
1962 	qp->qed_qp = dev->ops->rdma_create_qp(dev->rdma_ctx,
1963 					      in_params, &out_params);
1964 
1965 	if (!qp->qed_qp)
1966 		return -EINVAL;
1967 
1968 	/* Now we allocate the chain */
1969 	ext_pbl.p_pbl_virt = out_params.sq_pbl_virt;
1970 	ext_pbl.p_pbl_phys = out_params.sq_pbl_phys;
1971 
1972 	rc = dev->ops->common->chain_alloc(dev->cdev,
1973 					   QED_CHAIN_USE_TO_PRODUCE,
1974 					   QED_CHAIN_MODE_PBL,
1975 					   QED_CHAIN_CNT_TYPE_U32,
1976 					   n_sq_elems,
1977 					   QEDR_SQE_ELEMENT_SIZE,
1978 					   &qp->sq.pbl, &ext_pbl);
1979 
1980 	if (rc)
1981 		goto err;
1982 
1983 	ext_pbl.p_pbl_virt = out_params.rq_pbl_virt;
1984 	ext_pbl.p_pbl_phys = out_params.rq_pbl_phys;
1985 
1986 	rc = dev->ops->common->chain_alloc(dev->cdev,
1987 					   QED_CHAIN_USE_TO_CONSUME_PRODUCE,
1988 					   QED_CHAIN_MODE_PBL,
1989 					   QED_CHAIN_CNT_TYPE_U32,
1990 					   n_rq_elems,
1991 					   QEDR_RQE_ELEMENT_SIZE,
1992 					   &qp->rq.pbl, &ext_pbl);
1993 
1994 	if (rc)
1995 		goto err;
1996 
1997 	qp->qp_id = out_params.qp_id;
1998 	qp->icid = out_params.icid;
1999 
2000 	return qedr_set_iwarp_db_info(dev, qp);
2001 
2002 err:
2003 	dev->ops->rdma_destroy_qp(dev->rdma_ctx, qp->qed_qp);
2004 
2005 	return rc;
2006 }
2007 
2008 static void qedr_cleanup_kernel(struct qedr_dev *dev, struct qedr_qp *qp)
2009 {
2010 	dev->ops->common->chain_free(dev->cdev, &qp->sq.pbl);
2011 	kfree(qp->wqe_wr_id);
2012 
2013 	dev->ops->common->chain_free(dev->cdev, &qp->rq.pbl);
2014 	kfree(qp->rqe_wr_id);
2015 
2016 	/* GSI qp is not registered to db mechanism so no need to delete */
2017 	if (qp->qp_type == IB_QPT_GSI)
2018 		return;
2019 
2020 	qedr_db_recovery_del(dev, qp->sq.db, &qp->sq.db_data);
2021 
2022 	if (!qp->srq) {
2023 		qedr_db_recovery_del(dev, qp->rq.db, &qp->rq.db_data);
2024 
2025 		if (rdma_protocol_iwarp(&dev->ibdev, 1))
2026 			qedr_db_recovery_del(dev, qp->rq.iwarp_db2,
2027 					     &qp->rq.iwarp_db2_data);
2028 	}
2029 }
2030 
2031 static int qedr_create_kernel_qp(struct qedr_dev *dev,
2032 				 struct qedr_qp *qp,
2033 				 struct ib_pd *ibpd,
2034 				 struct ib_qp_init_attr *attrs)
2035 {
2036 	struct qed_rdma_create_qp_in_params in_params;
2037 	struct qedr_pd *pd = get_qedr_pd(ibpd);
2038 	int rc = -EINVAL;
2039 	u32 n_rq_elems;
2040 	u32 n_sq_elems;
2041 	u32 n_sq_entries;
2042 
2043 	memset(&in_params, 0, sizeof(in_params));
2044 	qp->create_type = QEDR_QP_CREATE_KERNEL;
2045 
2046 	/* A single work request may take up to QEDR_MAX_SQ_WQE_SIZE elements in
2047 	 * the ring. The ring should allow at least a single WR, even if the
2048 	 * user requested none, due to allocation issues.
2049 	 * We should add an extra WR since the prod and cons indices of
2050 	 * wqe_wr_id are managed in such a way that the WQ is considered full
2051 	 * when (prod+1)%max_wr==cons. We currently don't do that because we
2052 	 * double the number of entries due an iSER issue that pushes far more
2053 	 * WRs than indicated. If we decline its ib_post_send() then we get
2054 	 * error prints in the dmesg we'd like to avoid.
2055 	 */
2056 	qp->sq.max_wr = min_t(u32, attrs->cap.max_send_wr * dev->wq_multiplier,
2057 			      dev->attr.max_sqe);
2058 
2059 	qp->wqe_wr_id = kcalloc(qp->sq.max_wr, sizeof(*qp->wqe_wr_id),
2060 				GFP_KERNEL);
2061 	if (!qp->wqe_wr_id) {
2062 		DP_ERR(dev, "create qp: failed SQ shadow memory allocation\n");
2063 		return -ENOMEM;
2064 	}
2065 
2066 	/* QP handle to be written in CQE */
2067 	in_params.qp_handle_lo = lower_32_bits((uintptr_t) qp);
2068 	in_params.qp_handle_hi = upper_32_bits((uintptr_t) qp);
2069 
2070 	/* A single work request may take up to QEDR_MAX_RQ_WQE_SIZE elements in
2071 	 * the ring. There ring should allow at least a single WR, even if the
2072 	 * user requested none, due to allocation issues.
2073 	 */
2074 	qp->rq.max_wr = (u16) max_t(u32, attrs->cap.max_recv_wr, 1);
2075 
2076 	/* Allocate driver internal RQ array */
2077 	qp->rqe_wr_id = kcalloc(qp->rq.max_wr, sizeof(*qp->rqe_wr_id),
2078 				GFP_KERNEL);
2079 	if (!qp->rqe_wr_id) {
2080 		DP_ERR(dev,
2081 		       "create qp: failed RQ shadow memory allocation\n");
2082 		kfree(qp->wqe_wr_id);
2083 		return -ENOMEM;
2084 	}
2085 
2086 	qedr_init_common_qp_in_params(dev, pd, qp, attrs, true, &in_params);
2087 
2088 	n_sq_entries = attrs->cap.max_send_wr;
2089 	n_sq_entries = min_t(u32, n_sq_entries, dev->attr.max_sqe);
2090 	n_sq_entries = max_t(u32, n_sq_entries, 1);
2091 	n_sq_elems = n_sq_entries * QEDR_MAX_SQE_ELEMENTS_PER_SQE;
2092 
2093 	n_rq_elems = qp->rq.max_wr * QEDR_MAX_RQE_ELEMENTS_PER_RQE;
2094 
2095 	if (rdma_protocol_iwarp(&dev->ibdev, 1))
2096 		rc = qedr_iwarp_create_kernel_qp(dev, qp, &in_params,
2097 						 n_sq_elems, n_rq_elems);
2098 	else
2099 		rc = qedr_roce_create_kernel_qp(dev, qp, &in_params,
2100 						n_sq_elems, n_rq_elems);
2101 	if (rc)
2102 		qedr_cleanup_kernel(dev, qp);
2103 
2104 	return rc;
2105 }
2106 
2107 struct ib_qp *qedr_create_qp(struct ib_pd *ibpd,
2108 			     struct ib_qp_init_attr *attrs,
2109 			     struct ib_udata *udata)
2110 {
2111 	struct qedr_dev *dev = get_qedr_dev(ibpd->device);
2112 	struct qedr_pd *pd = get_qedr_pd(ibpd);
2113 	struct qedr_qp *qp;
2114 	struct ib_qp *ibqp;
2115 	int rc = 0;
2116 
2117 	DP_DEBUG(dev, QEDR_MSG_QP, "create qp: called from %s, pd=%p\n",
2118 		 udata ? "user library" : "kernel", pd);
2119 
2120 	rc = qedr_check_qp_attrs(ibpd, dev, attrs, udata);
2121 	if (rc)
2122 		return ERR_PTR(rc);
2123 
2124 	DP_DEBUG(dev, QEDR_MSG_QP,
2125 		 "create qp: called from %s, event_handler=%p, eepd=%p sq_cq=%p, sq_icid=%d, rq_cq=%p, rq_icid=%d\n",
2126 		 udata ? "user library" : "kernel", attrs->event_handler, pd,
2127 		 get_qedr_cq(attrs->send_cq),
2128 		 get_qedr_cq(attrs->send_cq)->icid,
2129 		 get_qedr_cq(attrs->recv_cq),
2130 		 attrs->recv_cq ? get_qedr_cq(attrs->recv_cq)->icid : 0);
2131 
2132 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
2133 	if (!qp) {
2134 		DP_ERR(dev, "create qp: failed allocating memory\n");
2135 		return ERR_PTR(-ENOMEM);
2136 	}
2137 
2138 	qedr_set_common_qp_params(dev, qp, pd, attrs);
2139 
2140 	if (attrs->qp_type == IB_QPT_GSI) {
2141 		ibqp = qedr_create_gsi_qp(dev, attrs, qp);
2142 		if (IS_ERR(ibqp))
2143 			kfree(qp);
2144 		return ibqp;
2145 	}
2146 
2147 	if (udata)
2148 		rc = qedr_create_user_qp(dev, qp, ibpd, udata, attrs);
2149 	else
2150 		rc = qedr_create_kernel_qp(dev, qp, ibpd, attrs);
2151 
2152 	if (rc)
2153 		goto err;
2154 
2155 	qp->ibqp.qp_num = qp->qp_id;
2156 
2157 	if (rdma_protocol_iwarp(&dev->ibdev, 1)) {
2158 		rc = xa_insert(&dev->qps, qp->qp_id, qp, GFP_KERNEL);
2159 		if (rc)
2160 			goto err;
2161 	}
2162 
2163 	return &qp->ibqp;
2164 
2165 err:
2166 	kfree(qp);
2167 
2168 	return ERR_PTR(-EFAULT);
2169 }
2170 
2171 static enum ib_qp_state qedr_get_ibqp_state(enum qed_roce_qp_state qp_state)
2172 {
2173 	switch (qp_state) {
2174 	case QED_ROCE_QP_STATE_RESET:
2175 		return IB_QPS_RESET;
2176 	case QED_ROCE_QP_STATE_INIT:
2177 		return IB_QPS_INIT;
2178 	case QED_ROCE_QP_STATE_RTR:
2179 		return IB_QPS_RTR;
2180 	case QED_ROCE_QP_STATE_RTS:
2181 		return IB_QPS_RTS;
2182 	case QED_ROCE_QP_STATE_SQD:
2183 		return IB_QPS_SQD;
2184 	case QED_ROCE_QP_STATE_ERR:
2185 		return IB_QPS_ERR;
2186 	case QED_ROCE_QP_STATE_SQE:
2187 		return IB_QPS_SQE;
2188 	}
2189 	return IB_QPS_ERR;
2190 }
2191 
2192 static enum qed_roce_qp_state qedr_get_state_from_ibqp(
2193 					enum ib_qp_state qp_state)
2194 {
2195 	switch (qp_state) {
2196 	case IB_QPS_RESET:
2197 		return QED_ROCE_QP_STATE_RESET;
2198 	case IB_QPS_INIT:
2199 		return QED_ROCE_QP_STATE_INIT;
2200 	case IB_QPS_RTR:
2201 		return QED_ROCE_QP_STATE_RTR;
2202 	case IB_QPS_RTS:
2203 		return QED_ROCE_QP_STATE_RTS;
2204 	case IB_QPS_SQD:
2205 		return QED_ROCE_QP_STATE_SQD;
2206 	case IB_QPS_ERR:
2207 		return QED_ROCE_QP_STATE_ERR;
2208 	default:
2209 		return QED_ROCE_QP_STATE_ERR;
2210 	}
2211 }
2212 
2213 static void qedr_reset_qp_hwq_info(struct qedr_qp_hwq_info *qph)
2214 {
2215 	qed_chain_reset(&qph->pbl);
2216 	qph->prod = 0;
2217 	qph->cons = 0;
2218 	qph->wqe_cons = 0;
2219 	qph->db_data.data.value = cpu_to_le16(0);
2220 }
2221 
2222 static int qedr_update_qp_state(struct qedr_dev *dev,
2223 				struct qedr_qp *qp,
2224 				enum qed_roce_qp_state cur_state,
2225 				enum qed_roce_qp_state new_state)
2226 {
2227 	int status = 0;
2228 
2229 	if (new_state == cur_state)
2230 		return 0;
2231 
2232 	switch (cur_state) {
2233 	case QED_ROCE_QP_STATE_RESET:
2234 		switch (new_state) {
2235 		case QED_ROCE_QP_STATE_INIT:
2236 			qp->prev_wqe_size = 0;
2237 			qedr_reset_qp_hwq_info(&qp->sq);
2238 			qedr_reset_qp_hwq_info(&qp->rq);
2239 			break;
2240 		default:
2241 			status = -EINVAL;
2242 			break;
2243 		}
2244 		break;
2245 	case QED_ROCE_QP_STATE_INIT:
2246 		switch (new_state) {
2247 		case QED_ROCE_QP_STATE_RTR:
2248 			/* Update doorbell (in case post_recv was
2249 			 * done before move to RTR)
2250 			 */
2251 
2252 			if (rdma_protocol_roce(&dev->ibdev, 1)) {
2253 				writel(qp->rq.db_data.raw, qp->rq.db);
2254 			}
2255 			break;
2256 		case QED_ROCE_QP_STATE_ERR:
2257 			break;
2258 		default:
2259 			/* Invalid state change. */
2260 			status = -EINVAL;
2261 			break;
2262 		}
2263 		break;
2264 	case QED_ROCE_QP_STATE_RTR:
2265 		/* RTR->XXX */
2266 		switch (new_state) {
2267 		case QED_ROCE_QP_STATE_RTS:
2268 			break;
2269 		case QED_ROCE_QP_STATE_ERR:
2270 			break;
2271 		default:
2272 			/* Invalid state change. */
2273 			status = -EINVAL;
2274 			break;
2275 		}
2276 		break;
2277 	case QED_ROCE_QP_STATE_RTS:
2278 		/* RTS->XXX */
2279 		switch (new_state) {
2280 		case QED_ROCE_QP_STATE_SQD:
2281 			break;
2282 		case QED_ROCE_QP_STATE_ERR:
2283 			break;
2284 		default:
2285 			/* Invalid state change. */
2286 			status = -EINVAL;
2287 			break;
2288 		}
2289 		break;
2290 	case QED_ROCE_QP_STATE_SQD:
2291 		/* SQD->XXX */
2292 		switch (new_state) {
2293 		case QED_ROCE_QP_STATE_RTS:
2294 		case QED_ROCE_QP_STATE_ERR:
2295 			break;
2296 		default:
2297 			/* Invalid state change. */
2298 			status = -EINVAL;
2299 			break;
2300 		}
2301 		break;
2302 	case QED_ROCE_QP_STATE_ERR:
2303 		/* ERR->XXX */
2304 		switch (new_state) {
2305 		case QED_ROCE_QP_STATE_RESET:
2306 			if ((qp->rq.prod != qp->rq.cons) ||
2307 			    (qp->sq.prod != qp->sq.cons)) {
2308 				DP_NOTICE(dev,
2309 					  "Error->Reset with rq/sq not empty rq.prod=%x rq.cons=%x sq.prod=%x sq.cons=%x\n",
2310 					  qp->rq.prod, qp->rq.cons, qp->sq.prod,
2311 					  qp->sq.cons);
2312 				status = -EINVAL;
2313 			}
2314 			break;
2315 		default:
2316 			status = -EINVAL;
2317 			break;
2318 		}
2319 		break;
2320 	default:
2321 		status = -EINVAL;
2322 		break;
2323 	}
2324 
2325 	return status;
2326 }
2327 
2328 int qedr_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2329 		   int attr_mask, struct ib_udata *udata)
2330 {
2331 	struct qedr_qp *qp = get_qedr_qp(ibqp);
2332 	struct qed_rdma_modify_qp_in_params qp_params = { 0 };
2333 	struct qedr_dev *dev = get_qedr_dev(&qp->dev->ibdev);
2334 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
2335 	enum ib_qp_state old_qp_state, new_qp_state;
2336 	enum qed_roce_qp_state cur_state;
2337 	int rc = 0;
2338 
2339 	DP_DEBUG(dev, QEDR_MSG_QP,
2340 		 "modify qp: qp %p attr_mask=0x%x, state=%d", qp, attr_mask,
2341 		 attr->qp_state);
2342 
2343 	old_qp_state = qedr_get_ibqp_state(qp->state);
2344 	if (attr_mask & IB_QP_STATE)
2345 		new_qp_state = attr->qp_state;
2346 	else
2347 		new_qp_state = old_qp_state;
2348 
2349 	if (rdma_protocol_roce(&dev->ibdev, 1)) {
2350 		if (!ib_modify_qp_is_ok(old_qp_state, new_qp_state,
2351 					ibqp->qp_type, attr_mask)) {
2352 			DP_ERR(dev,
2353 			       "modify qp: invalid attribute mask=0x%x specified for\n"
2354 			       "qpn=0x%x of type=0x%x old_qp_state=0x%x, new_qp_state=0x%x\n",
2355 			       attr_mask, qp->qp_id, ibqp->qp_type,
2356 			       old_qp_state, new_qp_state);
2357 			rc = -EINVAL;
2358 			goto err;
2359 		}
2360 	}
2361 
2362 	/* Translate the masks... */
2363 	if (attr_mask & IB_QP_STATE) {
2364 		SET_FIELD(qp_params.modify_flags,
2365 			  QED_RDMA_MODIFY_QP_VALID_NEW_STATE, 1);
2366 		qp_params.new_state = qedr_get_state_from_ibqp(attr->qp_state);
2367 	}
2368 
2369 	if (attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY)
2370 		qp_params.sqd_async = true;
2371 
2372 	if (attr_mask & IB_QP_PKEY_INDEX) {
2373 		SET_FIELD(qp_params.modify_flags,
2374 			  QED_ROCE_MODIFY_QP_VALID_PKEY, 1);
2375 		if (attr->pkey_index >= QEDR_ROCE_PKEY_TABLE_LEN) {
2376 			rc = -EINVAL;
2377 			goto err;
2378 		}
2379 
2380 		qp_params.pkey = QEDR_ROCE_PKEY_DEFAULT;
2381 	}
2382 
2383 	if (attr_mask & IB_QP_QKEY)
2384 		qp->qkey = attr->qkey;
2385 
2386 	if (attr_mask & IB_QP_ACCESS_FLAGS) {
2387 		SET_FIELD(qp_params.modify_flags,
2388 			  QED_RDMA_MODIFY_QP_VALID_RDMA_OPS_EN, 1);
2389 		qp_params.incoming_rdma_read_en = attr->qp_access_flags &
2390 						  IB_ACCESS_REMOTE_READ;
2391 		qp_params.incoming_rdma_write_en = attr->qp_access_flags &
2392 						   IB_ACCESS_REMOTE_WRITE;
2393 		qp_params.incoming_atomic_en = attr->qp_access_flags &
2394 					       IB_ACCESS_REMOTE_ATOMIC;
2395 	}
2396 
2397 	if (attr_mask & (IB_QP_AV | IB_QP_PATH_MTU)) {
2398 		if (rdma_protocol_iwarp(&dev->ibdev, 1))
2399 			return -EINVAL;
2400 
2401 		if (attr_mask & IB_QP_PATH_MTU) {
2402 			if (attr->path_mtu < IB_MTU_256 ||
2403 			    attr->path_mtu > IB_MTU_4096) {
2404 				pr_err("error: Only MTU sizes of 256, 512, 1024, 2048 and 4096 are supported by RoCE\n");
2405 				rc = -EINVAL;
2406 				goto err;
2407 			}
2408 			qp->mtu = min(ib_mtu_enum_to_int(attr->path_mtu),
2409 				      ib_mtu_enum_to_int(iboe_get_mtu
2410 							 (dev->ndev->mtu)));
2411 		}
2412 
2413 		if (!qp->mtu) {
2414 			qp->mtu =
2415 			ib_mtu_enum_to_int(iboe_get_mtu(dev->ndev->mtu));
2416 			pr_err("Fixing zeroed MTU to qp->mtu = %d\n", qp->mtu);
2417 		}
2418 
2419 		SET_FIELD(qp_params.modify_flags,
2420 			  QED_ROCE_MODIFY_QP_VALID_ADDRESS_VECTOR, 1);
2421 
2422 		qp_params.traffic_class_tos = grh->traffic_class;
2423 		qp_params.flow_label = grh->flow_label;
2424 		qp_params.hop_limit_ttl = grh->hop_limit;
2425 
2426 		qp->sgid_idx = grh->sgid_index;
2427 
2428 		rc = get_gid_info_from_table(ibqp, attr, attr_mask, &qp_params);
2429 		if (rc) {
2430 			DP_ERR(dev,
2431 			       "modify qp: problems with GID index %d (rc=%d)\n",
2432 			       grh->sgid_index, rc);
2433 			return rc;
2434 		}
2435 
2436 		rc = qedr_get_dmac(dev, &attr->ah_attr,
2437 				   qp_params.remote_mac_addr);
2438 		if (rc)
2439 			return rc;
2440 
2441 		qp_params.use_local_mac = true;
2442 		ether_addr_copy(qp_params.local_mac_addr, dev->ndev->dev_addr);
2443 
2444 		DP_DEBUG(dev, QEDR_MSG_QP, "dgid=%x:%x:%x:%x\n",
2445 			 qp_params.dgid.dwords[0], qp_params.dgid.dwords[1],
2446 			 qp_params.dgid.dwords[2], qp_params.dgid.dwords[3]);
2447 		DP_DEBUG(dev, QEDR_MSG_QP, "sgid=%x:%x:%x:%x\n",
2448 			 qp_params.sgid.dwords[0], qp_params.sgid.dwords[1],
2449 			 qp_params.sgid.dwords[2], qp_params.sgid.dwords[3]);
2450 		DP_DEBUG(dev, QEDR_MSG_QP, "remote_mac=[%pM]\n",
2451 			 qp_params.remote_mac_addr);
2452 
2453 		qp_params.mtu = qp->mtu;
2454 		qp_params.lb_indication = false;
2455 	}
2456 
2457 	if (!qp_params.mtu) {
2458 		/* Stay with current MTU */
2459 		if (qp->mtu)
2460 			qp_params.mtu = qp->mtu;
2461 		else
2462 			qp_params.mtu =
2463 			    ib_mtu_enum_to_int(iboe_get_mtu(dev->ndev->mtu));
2464 	}
2465 
2466 	if (attr_mask & IB_QP_TIMEOUT) {
2467 		SET_FIELD(qp_params.modify_flags,
2468 			  QED_ROCE_MODIFY_QP_VALID_ACK_TIMEOUT, 1);
2469 
2470 		/* The received timeout value is an exponent used like this:
2471 		 *    "12.7.34 LOCAL ACK TIMEOUT
2472 		 *    Value representing the transport (ACK) timeout for use by
2473 		 *    the remote, expressed as: 4.096 * 2^timeout [usec]"
2474 		 * The FW expects timeout in msec so we need to divide the usec
2475 		 * result by 1000. We'll approximate 1000~2^10, and 4.096 ~ 2^2,
2476 		 * so we get: 2^2 * 2^timeout / 2^10 = 2^(timeout - 8).
2477 		 * The value of zero means infinite so we use a 'max_t' to make
2478 		 * sure that sub 1 msec values will be configured as 1 msec.
2479 		 */
2480 		if (attr->timeout)
2481 			qp_params.ack_timeout =
2482 					1 << max_t(int, attr->timeout - 8, 0);
2483 		else
2484 			qp_params.ack_timeout = 0;
2485 	}
2486 
2487 	if (attr_mask & IB_QP_RETRY_CNT) {
2488 		SET_FIELD(qp_params.modify_flags,
2489 			  QED_ROCE_MODIFY_QP_VALID_RETRY_CNT, 1);
2490 		qp_params.retry_cnt = attr->retry_cnt;
2491 	}
2492 
2493 	if (attr_mask & IB_QP_RNR_RETRY) {
2494 		SET_FIELD(qp_params.modify_flags,
2495 			  QED_ROCE_MODIFY_QP_VALID_RNR_RETRY_CNT, 1);
2496 		qp_params.rnr_retry_cnt = attr->rnr_retry;
2497 	}
2498 
2499 	if (attr_mask & IB_QP_RQ_PSN) {
2500 		SET_FIELD(qp_params.modify_flags,
2501 			  QED_ROCE_MODIFY_QP_VALID_RQ_PSN, 1);
2502 		qp_params.rq_psn = attr->rq_psn;
2503 		qp->rq_psn = attr->rq_psn;
2504 	}
2505 
2506 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2507 		if (attr->max_rd_atomic > dev->attr.max_qp_req_rd_atomic_resc) {
2508 			rc = -EINVAL;
2509 			DP_ERR(dev,
2510 			       "unsupported max_rd_atomic=%d, supported=%d\n",
2511 			       attr->max_rd_atomic,
2512 			       dev->attr.max_qp_req_rd_atomic_resc);
2513 			goto err;
2514 		}
2515 
2516 		SET_FIELD(qp_params.modify_flags,
2517 			  QED_RDMA_MODIFY_QP_VALID_MAX_RD_ATOMIC_REQ, 1);
2518 		qp_params.max_rd_atomic_req = attr->max_rd_atomic;
2519 	}
2520 
2521 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
2522 		SET_FIELD(qp_params.modify_flags,
2523 			  QED_ROCE_MODIFY_QP_VALID_MIN_RNR_NAK_TIMER, 1);
2524 		qp_params.min_rnr_nak_timer = attr->min_rnr_timer;
2525 	}
2526 
2527 	if (attr_mask & IB_QP_SQ_PSN) {
2528 		SET_FIELD(qp_params.modify_flags,
2529 			  QED_ROCE_MODIFY_QP_VALID_SQ_PSN, 1);
2530 		qp_params.sq_psn = attr->sq_psn;
2531 		qp->sq_psn = attr->sq_psn;
2532 	}
2533 
2534 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2535 		if (attr->max_dest_rd_atomic >
2536 		    dev->attr.max_qp_resp_rd_atomic_resc) {
2537 			DP_ERR(dev,
2538 			       "unsupported max_dest_rd_atomic=%d, supported=%d\n",
2539 			       attr->max_dest_rd_atomic,
2540 			       dev->attr.max_qp_resp_rd_atomic_resc);
2541 
2542 			rc = -EINVAL;
2543 			goto err;
2544 		}
2545 
2546 		SET_FIELD(qp_params.modify_flags,
2547 			  QED_RDMA_MODIFY_QP_VALID_MAX_RD_ATOMIC_RESP, 1);
2548 		qp_params.max_rd_atomic_resp = attr->max_dest_rd_atomic;
2549 	}
2550 
2551 	if (attr_mask & IB_QP_DEST_QPN) {
2552 		SET_FIELD(qp_params.modify_flags,
2553 			  QED_ROCE_MODIFY_QP_VALID_DEST_QP, 1);
2554 
2555 		qp_params.dest_qp = attr->dest_qp_num;
2556 		qp->dest_qp_num = attr->dest_qp_num;
2557 	}
2558 
2559 	cur_state = qp->state;
2560 
2561 	/* Update the QP state before the actual ramrod to prevent a race with
2562 	 * fast path. Modifying the QP state to error will cause the device to
2563 	 * flush the CQEs and while polling the flushed CQEs will considered as
2564 	 * a potential issue if the QP isn't in error state.
2565 	 */
2566 	if ((attr_mask & IB_QP_STATE) && qp->qp_type != IB_QPT_GSI &&
2567 	    !udata && qp_params.new_state == QED_ROCE_QP_STATE_ERR)
2568 		qp->state = QED_ROCE_QP_STATE_ERR;
2569 
2570 	if (qp->qp_type != IB_QPT_GSI)
2571 		rc = dev->ops->rdma_modify_qp(dev->rdma_ctx,
2572 					      qp->qed_qp, &qp_params);
2573 
2574 	if (attr_mask & IB_QP_STATE) {
2575 		if ((qp->qp_type != IB_QPT_GSI) && (!udata))
2576 			rc = qedr_update_qp_state(dev, qp, cur_state,
2577 						  qp_params.new_state);
2578 		qp->state = qp_params.new_state;
2579 	}
2580 
2581 err:
2582 	return rc;
2583 }
2584 
2585 static int qedr_to_ib_qp_acc_flags(struct qed_rdma_query_qp_out_params *params)
2586 {
2587 	int ib_qp_acc_flags = 0;
2588 
2589 	if (params->incoming_rdma_write_en)
2590 		ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE;
2591 	if (params->incoming_rdma_read_en)
2592 		ib_qp_acc_flags |= IB_ACCESS_REMOTE_READ;
2593 	if (params->incoming_atomic_en)
2594 		ib_qp_acc_flags |= IB_ACCESS_REMOTE_ATOMIC;
2595 	ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE;
2596 	return ib_qp_acc_flags;
2597 }
2598 
2599 int qedr_query_qp(struct ib_qp *ibqp,
2600 		  struct ib_qp_attr *qp_attr,
2601 		  int attr_mask, struct ib_qp_init_attr *qp_init_attr)
2602 {
2603 	struct qed_rdma_query_qp_out_params params;
2604 	struct qedr_qp *qp = get_qedr_qp(ibqp);
2605 	struct qedr_dev *dev = qp->dev;
2606 	int rc = 0;
2607 
2608 	memset(&params, 0, sizeof(params));
2609 
2610 	rc = dev->ops->rdma_query_qp(dev->rdma_ctx, qp->qed_qp, &params);
2611 	if (rc)
2612 		goto err;
2613 
2614 	memset(qp_attr, 0, sizeof(*qp_attr));
2615 	memset(qp_init_attr, 0, sizeof(*qp_init_attr));
2616 
2617 	qp_attr->qp_state = qedr_get_ibqp_state(params.state);
2618 	qp_attr->cur_qp_state = qedr_get_ibqp_state(params.state);
2619 	qp_attr->path_mtu = ib_mtu_int_to_enum(params.mtu);
2620 	qp_attr->path_mig_state = IB_MIG_MIGRATED;
2621 	qp_attr->rq_psn = params.rq_psn;
2622 	qp_attr->sq_psn = params.sq_psn;
2623 	qp_attr->dest_qp_num = params.dest_qp;
2624 
2625 	qp_attr->qp_access_flags = qedr_to_ib_qp_acc_flags(&params);
2626 
2627 	qp_attr->cap.max_send_wr = qp->sq.max_wr;
2628 	qp_attr->cap.max_recv_wr = qp->rq.max_wr;
2629 	qp_attr->cap.max_send_sge = qp->sq.max_sges;
2630 	qp_attr->cap.max_recv_sge = qp->rq.max_sges;
2631 	qp_attr->cap.max_inline_data = ROCE_REQ_MAX_INLINE_DATA_SIZE;
2632 	qp_init_attr->cap = qp_attr->cap;
2633 
2634 	qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
2635 	rdma_ah_set_grh(&qp_attr->ah_attr, NULL,
2636 			params.flow_label, qp->sgid_idx,
2637 			params.hop_limit_ttl, params.traffic_class_tos);
2638 	rdma_ah_set_dgid_raw(&qp_attr->ah_attr, &params.dgid.bytes[0]);
2639 	rdma_ah_set_port_num(&qp_attr->ah_attr, 1);
2640 	rdma_ah_set_sl(&qp_attr->ah_attr, 0);
2641 	qp_attr->timeout = params.timeout;
2642 	qp_attr->rnr_retry = params.rnr_retry;
2643 	qp_attr->retry_cnt = params.retry_cnt;
2644 	qp_attr->min_rnr_timer = params.min_rnr_nak_timer;
2645 	qp_attr->pkey_index = params.pkey_index;
2646 	qp_attr->port_num = 1;
2647 	rdma_ah_set_path_bits(&qp_attr->ah_attr, 0);
2648 	rdma_ah_set_static_rate(&qp_attr->ah_attr, 0);
2649 	qp_attr->alt_pkey_index = 0;
2650 	qp_attr->alt_port_num = 0;
2651 	qp_attr->alt_timeout = 0;
2652 	memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr));
2653 
2654 	qp_attr->sq_draining = (params.state == QED_ROCE_QP_STATE_SQD) ? 1 : 0;
2655 	qp_attr->max_dest_rd_atomic = params.max_dest_rd_atomic;
2656 	qp_attr->max_rd_atomic = params.max_rd_atomic;
2657 	qp_attr->en_sqd_async_notify = (params.sqd_async) ? 1 : 0;
2658 
2659 	DP_DEBUG(dev, QEDR_MSG_QP, "QEDR_QUERY_QP: max_inline_data=%d\n",
2660 		 qp_attr->cap.max_inline_data);
2661 
2662 err:
2663 	return rc;
2664 }
2665 
2666 static int qedr_free_qp_resources(struct qedr_dev *dev, struct qedr_qp *qp,
2667 				  struct ib_udata *udata)
2668 {
2669 	struct qedr_ucontext *ctx =
2670 		rdma_udata_to_drv_context(udata, struct qedr_ucontext,
2671 					  ibucontext);
2672 	int rc;
2673 
2674 	if (qp->qp_type != IB_QPT_GSI) {
2675 		rc = dev->ops->rdma_destroy_qp(dev->rdma_ctx, qp->qed_qp);
2676 		if (rc)
2677 			return rc;
2678 	}
2679 
2680 	if (qp->create_type == QEDR_QP_CREATE_USER)
2681 		qedr_cleanup_user(dev, ctx, qp);
2682 	else
2683 		qedr_cleanup_kernel(dev, qp);
2684 
2685 	return 0;
2686 }
2687 
2688 int qedr_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
2689 {
2690 	struct qedr_qp *qp = get_qedr_qp(ibqp);
2691 	struct qedr_dev *dev = qp->dev;
2692 	struct ib_qp_attr attr;
2693 	int attr_mask = 0;
2694 
2695 	DP_DEBUG(dev, QEDR_MSG_QP, "destroy qp: destroying %p, qp type=%d\n",
2696 		 qp, qp->qp_type);
2697 
2698 	if (rdma_protocol_roce(&dev->ibdev, 1)) {
2699 		if ((qp->state != QED_ROCE_QP_STATE_RESET) &&
2700 		    (qp->state != QED_ROCE_QP_STATE_ERR) &&
2701 		    (qp->state != QED_ROCE_QP_STATE_INIT)) {
2702 
2703 			attr.qp_state = IB_QPS_ERR;
2704 			attr_mask |= IB_QP_STATE;
2705 
2706 			/* Change the QP state to ERROR */
2707 			qedr_modify_qp(ibqp, &attr, attr_mask, NULL);
2708 		}
2709 	} else {
2710 		/* If connection establishment started the WAIT_FOR_CONNECT
2711 		 * bit will be on and we need to Wait for the establishment
2712 		 * to complete before destroying the qp.
2713 		 */
2714 		if (test_and_set_bit(QEDR_IWARP_CM_WAIT_FOR_CONNECT,
2715 				     &qp->iwarp_cm_flags))
2716 			wait_for_completion(&qp->iwarp_cm_comp);
2717 
2718 		/* If graceful disconnect started, the WAIT_FOR_DISCONNECT
2719 		 * bit will be on, and we need to wait for the disconnect to
2720 		 * complete before continuing. We can use the same completion,
2721 		 * iwarp_cm_comp, since this is the only place that waits for
2722 		 * this completion and it is sequential. In addition,
2723 		 * disconnect can't occur before the connection is fully
2724 		 * established, therefore if WAIT_FOR_DISCONNECT is on it
2725 		 * means WAIT_FOR_CONNECT is also on and the completion for
2726 		 * CONNECT already occurred.
2727 		 */
2728 		if (test_and_set_bit(QEDR_IWARP_CM_WAIT_FOR_DISCONNECT,
2729 				     &qp->iwarp_cm_flags))
2730 			wait_for_completion(&qp->iwarp_cm_comp);
2731 	}
2732 
2733 	if (qp->qp_type == IB_QPT_GSI)
2734 		qedr_destroy_gsi_qp(dev);
2735 
2736 	/* We need to remove the entry from the xarray before we release the
2737 	 * qp_id to avoid a race of the qp_id being reallocated and failing
2738 	 * on xa_insert
2739 	 */
2740 	if (rdma_protocol_iwarp(&dev->ibdev, 1))
2741 		xa_erase(&dev->qps, qp->qp_id);
2742 
2743 	qedr_free_qp_resources(dev, qp, udata);
2744 
2745 	if (rdma_protocol_iwarp(&dev->ibdev, 1))
2746 		qedr_iw_qp_rem_ref(&qp->ibqp);
2747 
2748 	return 0;
2749 }
2750 
2751 int qedr_create_ah(struct ib_ah *ibah, struct rdma_ah_init_attr *init_attr,
2752 		   struct ib_udata *udata)
2753 {
2754 	struct qedr_ah *ah = get_qedr_ah(ibah);
2755 
2756 	rdma_copy_ah_attr(&ah->attr, init_attr->ah_attr);
2757 
2758 	return 0;
2759 }
2760 
2761 void qedr_destroy_ah(struct ib_ah *ibah, u32 flags)
2762 {
2763 	struct qedr_ah *ah = get_qedr_ah(ibah);
2764 
2765 	rdma_destroy_ah_attr(&ah->attr);
2766 }
2767 
2768 static void free_mr_info(struct qedr_dev *dev, struct mr_info *info)
2769 {
2770 	struct qedr_pbl *pbl, *tmp;
2771 
2772 	if (info->pbl_table)
2773 		list_add_tail(&info->pbl_table->list_entry,
2774 			      &info->free_pbl_list);
2775 
2776 	if (!list_empty(&info->inuse_pbl_list))
2777 		list_splice(&info->inuse_pbl_list, &info->free_pbl_list);
2778 
2779 	list_for_each_entry_safe(pbl, tmp, &info->free_pbl_list, list_entry) {
2780 		list_del(&pbl->list_entry);
2781 		qedr_free_pbl(dev, &info->pbl_info, pbl);
2782 	}
2783 }
2784 
2785 static int init_mr_info(struct qedr_dev *dev, struct mr_info *info,
2786 			size_t page_list_len, bool two_layered)
2787 {
2788 	struct qedr_pbl *tmp;
2789 	int rc;
2790 
2791 	INIT_LIST_HEAD(&info->free_pbl_list);
2792 	INIT_LIST_HEAD(&info->inuse_pbl_list);
2793 
2794 	rc = qedr_prepare_pbl_tbl(dev, &info->pbl_info,
2795 				  page_list_len, two_layered);
2796 	if (rc)
2797 		goto done;
2798 
2799 	info->pbl_table = qedr_alloc_pbl_tbl(dev, &info->pbl_info, GFP_KERNEL);
2800 	if (IS_ERR(info->pbl_table)) {
2801 		rc = PTR_ERR(info->pbl_table);
2802 		goto done;
2803 	}
2804 
2805 	DP_DEBUG(dev, QEDR_MSG_MR, "pbl_table_pa = %pa\n",
2806 		 &info->pbl_table->pa);
2807 
2808 	/* in usual case we use 2 PBLs, so we add one to free
2809 	 * list and allocating another one
2810 	 */
2811 	tmp = qedr_alloc_pbl_tbl(dev, &info->pbl_info, GFP_KERNEL);
2812 	if (IS_ERR(tmp)) {
2813 		DP_DEBUG(dev, QEDR_MSG_MR, "Extra PBL is not allocated\n");
2814 		goto done;
2815 	}
2816 
2817 	list_add_tail(&tmp->list_entry, &info->free_pbl_list);
2818 
2819 	DP_DEBUG(dev, QEDR_MSG_MR, "extra pbl_table_pa = %pa\n", &tmp->pa);
2820 
2821 done:
2822 	if (rc)
2823 		free_mr_info(dev, info);
2824 
2825 	return rc;
2826 }
2827 
2828 struct ib_mr *qedr_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
2829 			       u64 usr_addr, int acc, struct ib_udata *udata)
2830 {
2831 	struct qedr_dev *dev = get_qedr_dev(ibpd->device);
2832 	struct qedr_mr *mr;
2833 	struct qedr_pd *pd;
2834 	int rc = -ENOMEM;
2835 
2836 	pd = get_qedr_pd(ibpd);
2837 	DP_DEBUG(dev, QEDR_MSG_MR,
2838 		 "qedr_register user mr pd = %d start = %lld, len = %lld, usr_addr = %lld, acc = %d\n",
2839 		 pd->pd_id, start, len, usr_addr, acc);
2840 
2841 	if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE))
2842 		return ERR_PTR(-EINVAL);
2843 
2844 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
2845 	if (!mr)
2846 		return ERR_PTR(rc);
2847 
2848 	mr->type = QEDR_MR_USER;
2849 
2850 	mr->umem = ib_umem_get(ibpd->device, start, len, acc);
2851 	if (IS_ERR(mr->umem)) {
2852 		rc = -EFAULT;
2853 		goto err0;
2854 	}
2855 
2856 	rc = init_mr_info(dev, &mr->info, ib_umem_page_count(mr->umem), 1);
2857 	if (rc)
2858 		goto err1;
2859 
2860 	qedr_populate_pbls(dev, mr->umem, mr->info.pbl_table,
2861 			   &mr->info.pbl_info, PAGE_SHIFT);
2862 
2863 	rc = dev->ops->rdma_alloc_tid(dev->rdma_ctx, &mr->hw_mr.itid);
2864 	if (rc) {
2865 		DP_ERR(dev, "roce alloc tid returned an error %d\n", rc);
2866 		goto err1;
2867 	}
2868 
2869 	/* Index only, 18 bit long, lkey = itid << 8 | key */
2870 	mr->hw_mr.tid_type = QED_RDMA_TID_REGISTERED_MR;
2871 	mr->hw_mr.key = 0;
2872 	mr->hw_mr.pd = pd->pd_id;
2873 	mr->hw_mr.local_read = 1;
2874 	mr->hw_mr.local_write = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
2875 	mr->hw_mr.remote_read = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
2876 	mr->hw_mr.remote_write = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
2877 	mr->hw_mr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
2878 	mr->hw_mr.mw_bind = false;
2879 	mr->hw_mr.pbl_ptr = mr->info.pbl_table[0].pa;
2880 	mr->hw_mr.pbl_two_level = mr->info.pbl_info.two_layered;
2881 	mr->hw_mr.pbl_page_size_log = ilog2(mr->info.pbl_info.pbl_size);
2882 	mr->hw_mr.page_size_log = PAGE_SHIFT;
2883 	mr->hw_mr.fbo = ib_umem_offset(mr->umem);
2884 	mr->hw_mr.length = len;
2885 	mr->hw_mr.vaddr = usr_addr;
2886 	mr->hw_mr.zbva = false;
2887 	mr->hw_mr.phy_mr = false;
2888 	mr->hw_mr.dma_mr = false;
2889 
2890 	rc = dev->ops->rdma_register_tid(dev->rdma_ctx, &mr->hw_mr);
2891 	if (rc) {
2892 		DP_ERR(dev, "roce register tid returned an error %d\n", rc);
2893 		goto err2;
2894 	}
2895 
2896 	mr->ibmr.lkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
2897 	if (mr->hw_mr.remote_write || mr->hw_mr.remote_read ||
2898 	    mr->hw_mr.remote_atomic)
2899 		mr->ibmr.rkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
2900 
2901 	DP_DEBUG(dev, QEDR_MSG_MR, "register user mr lkey: %x\n",
2902 		 mr->ibmr.lkey);
2903 	return &mr->ibmr;
2904 
2905 err2:
2906 	dev->ops->rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid);
2907 err1:
2908 	qedr_free_pbl(dev, &mr->info.pbl_info, mr->info.pbl_table);
2909 err0:
2910 	kfree(mr);
2911 	return ERR_PTR(rc);
2912 }
2913 
2914 int qedr_dereg_mr(struct ib_mr *ib_mr, struct ib_udata *udata)
2915 {
2916 	struct qedr_mr *mr = get_qedr_mr(ib_mr);
2917 	struct qedr_dev *dev = get_qedr_dev(ib_mr->device);
2918 	int rc = 0;
2919 
2920 	rc = dev->ops->rdma_deregister_tid(dev->rdma_ctx, mr->hw_mr.itid);
2921 	if (rc)
2922 		return rc;
2923 
2924 	dev->ops->rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid);
2925 
2926 	if (mr->type != QEDR_MR_DMA)
2927 		free_mr_info(dev, &mr->info);
2928 
2929 	/* it could be user registered memory. */
2930 	ib_umem_release(mr->umem);
2931 
2932 	kfree(mr);
2933 
2934 	return rc;
2935 }
2936 
2937 static struct qedr_mr *__qedr_alloc_mr(struct ib_pd *ibpd,
2938 				       int max_page_list_len)
2939 {
2940 	struct qedr_pd *pd = get_qedr_pd(ibpd);
2941 	struct qedr_dev *dev = get_qedr_dev(ibpd->device);
2942 	struct qedr_mr *mr;
2943 	int rc = -ENOMEM;
2944 
2945 	DP_DEBUG(dev, QEDR_MSG_MR,
2946 		 "qedr_alloc_frmr pd = %d max_page_list_len= %d\n", pd->pd_id,
2947 		 max_page_list_len);
2948 
2949 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
2950 	if (!mr)
2951 		return ERR_PTR(rc);
2952 
2953 	mr->dev = dev;
2954 	mr->type = QEDR_MR_FRMR;
2955 
2956 	rc = init_mr_info(dev, &mr->info, max_page_list_len, 1);
2957 	if (rc)
2958 		goto err0;
2959 
2960 	rc = dev->ops->rdma_alloc_tid(dev->rdma_ctx, &mr->hw_mr.itid);
2961 	if (rc) {
2962 		DP_ERR(dev, "roce alloc tid returned an error %d\n", rc);
2963 		goto err0;
2964 	}
2965 
2966 	/* Index only, 18 bit long, lkey = itid << 8 | key */
2967 	mr->hw_mr.tid_type = QED_RDMA_TID_FMR;
2968 	mr->hw_mr.key = 0;
2969 	mr->hw_mr.pd = pd->pd_id;
2970 	mr->hw_mr.local_read = 1;
2971 	mr->hw_mr.local_write = 0;
2972 	mr->hw_mr.remote_read = 0;
2973 	mr->hw_mr.remote_write = 0;
2974 	mr->hw_mr.remote_atomic = 0;
2975 	mr->hw_mr.mw_bind = false;
2976 	mr->hw_mr.pbl_ptr = 0;
2977 	mr->hw_mr.pbl_two_level = mr->info.pbl_info.two_layered;
2978 	mr->hw_mr.pbl_page_size_log = ilog2(mr->info.pbl_info.pbl_size);
2979 	mr->hw_mr.fbo = 0;
2980 	mr->hw_mr.length = 0;
2981 	mr->hw_mr.vaddr = 0;
2982 	mr->hw_mr.zbva = false;
2983 	mr->hw_mr.phy_mr = true;
2984 	mr->hw_mr.dma_mr = false;
2985 
2986 	rc = dev->ops->rdma_register_tid(dev->rdma_ctx, &mr->hw_mr);
2987 	if (rc) {
2988 		DP_ERR(dev, "roce register tid returned an error %d\n", rc);
2989 		goto err1;
2990 	}
2991 
2992 	mr->ibmr.lkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
2993 	mr->ibmr.rkey = mr->ibmr.lkey;
2994 
2995 	DP_DEBUG(dev, QEDR_MSG_MR, "alloc frmr: %x\n", mr->ibmr.lkey);
2996 	return mr;
2997 
2998 err1:
2999 	dev->ops->rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid);
3000 err0:
3001 	kfree(mr);
3002 	return ERR_PTR(rc);
3003 }
3004 
3005 struct ib_mr *qedr_alloc_mr(struct ib_pd *ibpd, enum ib_mr_type mr_type,
3006 			    u32 max_num_sg, struct ib_udata *udata)
3007 {
3008 	struct qedr_mr *mr;
3009 
3010 	if (mr_type != IB_MR_TYPE_MEM_REG)
3011 		return ERR_PTR(-EINVAL);
3012 
3013 	mr = __qedr_alloc_mr(ibpd, max_num_sg);
3014 
3015 	if (IS_ERR(mr))
3016 		return ERR_PTR(-EINVAL);
3017 
3018 	return &mr->ibmr;
3019 }
3020 
3021 static int qedr_set_page(struct ib_mr *ibmr, u64 addr)
3022 {
3023 	struct qedr_mr *mr = get_qedr_mr(ibmr);
3024 	struct qedr_pbl *pbl_table;
3025 	struct regpair *pbe;
3026 	u32 pbes_in_page;
3027 
3028 	if (unlikely(mr->npages == mr->info.pbl_info.num_pbes)) {
3029 		DP_ERR(mr->dev, "qedr_set_page fails when %d\n", mr->npages);
3030 		return -ENOMEM;
3031 	}
3032 
3033 	DP_DEBUG(mr->dev, QEDR_MSG_MR, "qedr_set_page pages[%d] = 0x%llx\n",
3034 		 mr->npages, addr);
3035 
3036 	pbes_in_page = mr->info.pbl_info.pbl_size / sizeof(u64);
3037 	pbl_table = mr->info.pbl_table + (mr->npages / pbes_in_page);
3038 	pbe = (struct regpair *)pbl_table->va;
3039 	pbe +=  mr->npages % pbes_in_page;
3040 	pbe->lo = cpu_to_le32((u32)addr);
3041 	pbe->hi = cpu_to_le32((u32)upper_32_bits(addr));
3042 
3043 	mr->npages++;
3044 
3045 	return 0;
3046 }
3047 
3048 static void handle_completed_mrs(struct qedr_dev *dev, struct mr_info *info)
3049 {
3050 	int work = info->completed - info->completed_handled - 1;
3051 
3052 	DP_DEBUG(dev, QEDR_MSG_MR, "Special FMR work = %d\n", work);
3053 	while (work-- > 0 && !list_empty(&info->inuse_pbl_list)) {
3054 		struct qedr_pbl *pbl;
3055 
3056 		/* Free all the page list that are possible to be freed
3057 		 * (all the ones that were invalidated), under the assumption
3058 		 * that if an FMR was completed successfully that means that
3059 		 * if there was an invalidate operation before it also ended
3060 		 */
3061 		pbl = list_first_entry(&info->inuse_pbl_list,
3062 				       struct qedr_pbl, list_entry);
3063 		list_move_tail(&pbl->list_entry, &info->free_pbl_list);
3064 		info->completed_handled++;
3065 	}
3066 }
3067 
3068 int qedr_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
3069 		   int sg_nents, unsigned int *sg_offset)
3070 {
3071 	struct qedr_mr *mr = get_qedr_mr(ibmr);
3072 
3073 	mr->npages = 0;
3074 
3075 	handle_completed_mrs(mr->dev, &mr->info);
3076 	return ib_sg_to_pages(ibmr, sg, sg_nents, NULL, qedr_set_page);
3077 }
3078 
3079 struct ib_mr *qedr_get_dma_mr(struct ib_pd *ibpd, int acc)
3080 {
3081 	struct qedr_dev *dev = get_qedr_dev(ibpd->device);
3082 	struct qedr_pd *pd = get_qedr_pd(ibpd);
3083 	struct qedr_mr *mr;
3084 	int rc;
3085 
3086 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3087 	if (!mr)
3088 		return ERR_PTR(-ENOMEM);
3089 
3090 	mr->type = QEDR_MR_DMA;
3091 
3092 	rc = dev->ops->rdma_alloc_tid(dev->rdma_ctx, &mr->hw_mr.itid);
3093 	if (rc) {
3094 		DP_ERR(dev, "roce alloc tid returned an error %d\n", rc);
3095 		goto err1;
3096 	}
3097 
3098 	/* index only, 18 bit long, lkey = itid << 8 | key */
3099 	mr->hw_mr.tid_type = QED_RDMA_TID_REGISTERED_MR;
3100 	mr->hw_mr.pd = pd->pd_id;
3101 	mr->hw_mr.local_read = 1;
3102 	mr->hw_mr.local_write = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
3103 	mr->hw_mr.remote_read = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
3104 	mr->hw_mr.remote_write = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
3105 	mr->hw_mr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
3106 	mr->hw_mr.dma_mr = true;
3107 
3108 	rc = dev->ops->rdma_register_tid(dev->rdma_ctx, &mr->hw_mr);
3109 	if (rc) {
3110 		DP_ERR(dev, "roce register tid returned an error %d\n", rc);
3111 		goto err2;
3112 	}
3113 
3114 	mr->ibmr.lkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
3115 	if (mr->hw_mr.remote_write || mr->hw_mr.remote_read ||
3116 	    mr->hw_mr.remote_atomic)
3117 		mr->ibmr.rkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
3118 
3119 	DP_DEBUG(dev, QEDR_MSG_MR, "get dma mr: lkey = %x\n", mr->ibmr.lkey);
3120 	return &mr->ibmr;
3121 
3122 err2:
3123 	dev->ops->rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid);
3124 err1:
3125 	kfree(mr);
3126 	return ERR_PTR(rc);
3127 }
3128 
3129 static inline int qedr_wq_is_full(struct qedr_qp_hwq_info *wq)
3130 {
3131 	return (((wq->prod + 1) % wq->max_wr) == wq->cons);
3132 }
3133 
3134 static int sge_data_len(struct ib_sge *sg_list, int num_sge)
3135 {
3136 	int i, len = 0;
3137 
3138 	for (i = 0; i < num_sge; i++)
3139 		len += sg_list[i].length;
3140 
3141 	return len;
3142 }
3143 
3144 static void swap_wqe_data64(u64 *p)
3145 {
3146 	int i;
3147 
3148 	for (i = 0; i < QEDR_SQE_ELEMENT_SIZE / sizeof(u64); i++, p++)
3149 		*p = cpu_to_be64(cpu_to_le64(*p));
3150 }
3151 
3152 static u32 qedr_prepare_sq_inline_data(struct qedr_dev *dev,
3153 				       struct qedr_qp *qp, u8 *wqe_size,
3154 				       const struct ib_send_wr *wr,
3155 				       const struct ib_send_wr **bad_wr,
3156 				       u8 *bits, u8 bit)
3157 {
3158 	u32 data_size = sge_data_len(wr->sg_list, wr->num_sge);
3159 	char *seg_prt, *wqe;
3160 	int i, seg_siz;
3161 
3162 	if (data_size > ROCE_REQ_MAX_INLINE_DATA_SIZE) {
3163 		DP_ERR(dev, "Too much inline data in WR: %d\n", data_size);
3164 		*bad_wr = wr;
3165 		return 0;
3166 	}
3167 
3168 	if (!data_size)
3169 		return data_size;
3170 
3171 	*bits |= bit;
3172 
3173 	seg_prt = NULL;
3174 	wqe = NULL;
3175 	seg_siz = 0;
3176 
3177 	/* Copy data inline */
3178 	for (i = 0; i < wr->num_sge; i++) {
3179 		u32 len = wr->sg_list[i].length;
3180 		void *src = (void *)(uintptr_t)wr->sg_list[i].addr;
3181 
3182 		while (len > 0) {
3183 			u32 cur;
3184 
3185 			/* New segment required */
3186 			if (!seg_siz) {
3187 				wqe = (char *)qed_chain_produce(&qp->sq.pbl);
3188 				seg_prt = wqe;
3189 				seg_siz = sizeof(struct rdma_sq_common_wqe);
3190 				(*wqe_size)++;
3191 			}
3192 
3193 			/* Calculate currently allowed length */
3194 			cur = min_t(u32, len, seg_siz);
3195 			memcpy(seg_prt, src, cur);
3196 
3197 			/* Update segment variables */
3198 			seg_prt += cur;
3199 			seg_siz -= cur;
3200 
3201 			/* Update sge variables */
3202 			src += cur;
3203 			len -= cur;
3204 
3205 			/* Swap fully-completed segments */
3206 			if (!seg_siz)
3207 				swap_wqe_data64((u64 *)wqe);
3208 		}
3209 	}
3210 
3211 	/* swap last not completed segment */
3212 	if (seg_siz)
3213 		swap_wqe_data64((u64 *)wqe);
3214 
3215 	return data_size;
3216 }
3217 
3218 #define RQ_SGE_SET(sge, vaddr, vlength, vflags)			\
3219 	do {							\
3220 		DMA_REGPAIR_LE(sge->addr, vaddr);		\
3221 		(sge)->length = cpu_to_le32(vlength);		\
3222 		(sge)->flags = cpu_to_le32(vflags);		\
3223 	} while (0)
3224 
3225 #define SRQ_HDR_SET(hdr, vwr_id, num_sge)			\
3226 	do {							\
3227 		DMA_REGPAIR_LE(hdr->wr_id, vwr_id);		\
3228 		(hdr)->num_sges = num_sge;			\
3229 	} while (0)
3230 
3231 #define SRQ_SGE_SET(sge, vaddr, vlength, vlkey)			\
3232 	do {							\
3233 		DMA_REGPAIR_LE(sge->addr, vaddr);		\
3234 		(sge)->length = cpu_to_le32(vlength);		\
3235 		(sge)->l_key = cpu_to_le32(vlkey);		\
3236 	} while (0)
3237 
3238 static u32 qedr_prepare_sq_sges(struct qedr_qp *qp, u8 *wqe_size,
3239 				const struct ib_send_wr *wr)
3240 {
3241 	u32 data_size = 0;
3242 	int i;
3243 
3244 	for (i = 0; i < wr->num_sge; i++) {
3245 		struct rdma_sq_sge *sge = qed_chain_produce(&qp->sq.pbl);
3246 
3247 		DMA_REGPAIR_LE(sge->addr, wr->sg_list[i].addr);
3248 		sge->l_key = cpu_to_le32(wr->sg_list[i].lkey);
3249 		sge->length = cpu_to_le32(wr->sg_list[i].length);
3250 		data_size += wr->sg_list[i].length;
3251 	}
3252 
3253 	if (wqe_size)
3254 		*wqe_size += wr->num_sge;
3255 
3256 	return data_size;
3257 }
3258 
3259 static u32 qedr_prepare_sq_rdma_data(struct qedr_dev *dev,
3260 				     struct qedr_qp *qp,
3261 				     struct rdma_sq_rdma_wqe_1st *rwqe,
3262 				     struct rdma_sq_rdma_wqe_2nd *rwqe2,
3263 				     const struct ib_send_wr *wr,
3264 				     const struct ib_send_wr **bad_wr)
3265 {
3266 	rwqe2->r_key = cpu_to_le32(rdma_wr(wr)->rkey);
3267 	DMA_REGPAIR_LE(rwqe2->remote_va, rdma_wr(wr)->remote_addr);
3268 
3269 	if (wr->send_flags & IB_SEND_INLINE &&
3270 	    (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
3271 	     wr->opcode == IB_WR_RDMA_WRITE)) {
3272 		u8 flags = 0;
3273 
3274 		SET_FIELD2(flags, RDMA_SQ_RDMA_WQE_1ST_INLINE_FLG, 1);
3275 		return qedr_prepare_sq_inline_data(dev, qp, &rwqe->wqe_size, wr,
3276 						   bad_wr, &rwqe->flags, flags);
3277 	}
3278 
3279 	return qedr_prepare_sq_sges(qp, &rwqe->wqe_size, wr);
3280 }
3281 
3282 static u32 qedr_prepare_sq_send_data(struct qedr_dev *dev,
3283 				     struct qedr_qp *qp,
3284 				     struct rdma_sq_send_wqe_1st *swqe,
3285 				     struct rdma_sq_send_wqe_2st *swqe2,
3286 				     const struct ib_send_wr *wr,
3287 				     const struct ib_send_wr **bad_wr)
3288 {
3289 	memset(swqe2, 0, sizeof(*swqe2));
3290 	if (wr->send_flags & IB_SEND_INLINE) {
3291 		u8 flags = 0;
3292 
3293 		SET_FIELD2(flags, RDMA_SQ_SEND_WQE_INLINE_FLG, 1);
3294 		return qedr_prepare_sq_inline_data(dev, qp, &swqe->wqe_size, wr,
3295 						   bad_wr, &swqe->flags, flags);
3296 	}
3297 
3298 	return qedr_prepare_sq_sges(qp, &swqe->wqe_size, wr);
3299 }
3300 
3301 static int qedr_prepare_reg(struct qedr_qp *qp,
3302 			    struct rdma_sq_fmr_wqe_1st *fwqe1,
3303 			    const struct ib_reg_wr *wr)
3304 {
3305 	struct qedr_mr *mr = get_qedr_mr(wr->mr);
3306 	struct rdma_sq_fmr_wqe_2nd *fwqe2;
3307 
3308 	fwqe2 = (struct rdma_sq_fmr_wqe_2nd *)qed_chain_produce(&qp->sq.pbl);
3309 	fwqe1->addr.hi = upper_32_bits(mr->ibmr.iova);
3310 	fwqe1->addr.lo = lower_32_bits(mr->ibmr.iova);
3311 	fwqe1->l_key = wr->key;
3312 
3313 	fwqe2->access_ctrl = 0;
3314 
3315 	SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_REMOTE_READ,
3316 		   !!(wr->access & IB_ACCESS_REMOTE_READ));
3317 	SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_REMOTE_WRITE,
3318 		   !!(wr->access & IB_ACCESS_REMOTE_WRITE));
3319 	SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_ENABLE_ATOMIC,
3320 		   !!(wr->access & IB_ACCESS_REMOTE_ATOMIC));
3321 	SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_LOCAL_READ, 1);
3322 	SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_LOCAL_WRITE,
3323 		   !!(wr->access & IB_ACCESS_LOCAL_WRITE));
3324 	fwqe2->fmr_ctrl = 0;
3325 
3326 	SET_FIELD2(fwqe2->fmr_ctrl, RDMA_SQ_FMR_WQE_2ND_PAGE_SIZE_LOG,
3327 		   ilog2(mr->ibmr.page_size) - 12);
3328 
3329 	fwqe2->length_hi = 0;
3330 	fwqe2->length_lo = mr->ibmr.length;
3331 	fwqe2->pbl_addr.hi = upper_32_bits(mr->info.pbl_table->pa);
3332 	fwqe2->pbl_addr.lo = lower_32_bits(mr->info.pbl_table->pa);
3333 
3334 	qp->wqe_wr_id[qp->sq.prod].mr = mr;
3335 
3336 	return 0;
3337 }
3338 
3339 static enum ib_wc_opcode qedr_ib_to_wc_opcode(enum ib_wr_opcode opcode)
3340 {
3341 	switch (opcode) {
3342 	case IB_WR_RDMA_WRITE:
3343 	case IB_WR_RDMA_WRITE_WITH_IMM:
3344 		return IB_WC_RDMA_WRITE;
3345 	case IB_WR_SEND_WITH_IMM:
3346 	case IB_WR_SEND:
3347 	case IB_WR_SEND_WITH_INV:
3348 		return IB_WC_SEND;
3349 	case IB_WR_RDMA_READ:
3350 	case IB_WR_RDMA_READ_WITH_INV:
3351 		return IB_WC_RDMA_READ;
3352 	case IB_WR_ATOMIC_CMP_AND_SWP:
3353 		return IB_WC_COMP_SWAP;
3354 	case IB_WR_ATOMIC_FETCH_AND_ADD:
3355 		return IB_WC_FETCH_ADD;
3356 	case IB_WR_REG_MR:
3357 		return IB_WC_REG_MR;
3358 	case IB_WR_LOCAL_INV:
3359 		return IB_WC_LOCAL_INV;
3360 	default:
3361 		return IB_WC_SEND;
3362 	}
3363 }
3364 
3365 static inline bool qedr_can_post_send(struct qedr_qp *qp,
3366 				      const struct ib_send_wr *wr)
3367 {
3368 	int wq_is_full, err_wr, pbl_is_full;
3369 	struct qedr_dev *dev = qp->dev;
3370 
3371 	/* prevent SQ overflow and/or processing of a bad WR */
3372 	err_wr = wr->num_sge > qp->sq.max_sges;
3373 	wq_is_full = qedr_wq_is_full(&qp->sq);
3374 	pbl_is_full = qed_chain_get_elem_left_u32(&qp->sq.pbl) <
3375 		      QEDR_MAX_SQE_ELEMENTS_PER_SQE;
3376 	if (wq_is_full || err_wr || pbl_is_full) {
3377 		if (wq_is_full && !(qp->err_bitmap & QEDR_QP_ERR_SQ_FULL)) {
3378 			DP_ERR(dev,
3379 			       "error: WQ is full. Post send on QP %p failed (this error appears only once)\n",
3380 			       qp);
3381 			qp->err_bitmap |= QEDR_QP_ERR_SQ_FULL;
3382 		}
3383 
3384 		if (err_wr && !(qp->err_bitmap & QEDR_QP_ERR_BAD_SR)) {
3385 			DP_ERR(dev,
3386 			       "error: WR is bad. Post send on QP %p failed (this error appears only once)\n",
3387 			       qp);
3388 			qp->err_bitmap |= QEDR_QP_ERR_BAD_SR;
3389 		}
3390 
3391 		if (pbl_is_full &&
3392 		    !(qp->err_bitmap & QEDR_QP_ERR_SQ_PBL_FULL)) {
3393 			DP_ERR(dev,
3394 			       "error: WQ PBL is full. Post send on QP %p failed (this error appears only once)\n",
3395 			       qp);
3396 			qp->err_bitmap |= QEDR_QP_ERR_SQ_PBL_FULL;
3397 		}
3398 		return false;
3399 	}
3400 	return true;
3401 }
3402 
3403 static int __qedr_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
3404 			    const struct ib_send_wr **bad_wr)
3405 {
3406 	struct qedr_dev *dev = get_qedr_dev(ibqp->device);
3407 	struct qedr_qp *qp = get_qedr_qp(ibqp);
3408 	struct rdma_sq_atomic_wqe_1st *awqe1;
3409 	struct rdma_sq_atomic_wqe_2nd *awqe2;
3410 	struct rdma_sq_atomic_wqe_3rd *awqe3;
3411 	struct rdma_sq_send_wqe_2st *swqe2;
3412 	struct rdma_sq_local_inv_wqe *iwqe;
3413 	struct rdma_sq_rdma_wqe_2nd *rwqe2;
3414 	struct rdma_sq_send_wqe_1st *swqe;
3415 	struct rdma_sq_rdma_wqe_1st *rwqe;
3416 	struct rdma_sq_fmr_wqe_1st *fwqe1;
3417 	struct rdma_sq_common_wqe *wqe;
3418 	u32 length;
3419 	int rc = 0;
3420 	bool comp;
3421 
3422 	if (!qedr_can_post_send(qp, wr)) {
3423 		*bad_wr = wr;
3424 		return -ENOMEM;
3425 	}
3426 
3427 	wqe = qed_chain_produce(&qp->sq.pbl);
3428 	qp->wqe_wr_id[qp->sq.prod].signaled =
3429 		!!(wr->send_flags & IB_SEND_SIGNALED) || qp->signaled;
3430 
3431 	wqe->flags = 0;
3432 	SET_FIELD2(wqe->flags, RDMA_SQ_SEND_WQE_SE_FLG,
3433 		   !!(wr->send_flags & IB_SEND_SOLICITED));
3434 	comp = (!!(wr->send_flags & IB_SEND_SIGNALED)) || qp->signaled;
3435 	SET_FIELD2(wqe->flags, RDMA_SQ_SEND_WQE_COMP_FLG, comp);
3436 	SET_FIELD2(wqe->flags, RDMA_SQ_SEND_WQE_RD_FENCE_FLG,
3437 		   !!(wr->send_flags & IB_SEND_FENCE));
3438 	wqe->prev_wqe_size = qp->prev_wqe_size;
3439 
3440 	qp->wqe_wr_id[qp->sq.prod].opcode = qedr_ib_to_wc_opcode(wr->opcode);
3441 
3442 	switch (wr->opcode) {
3443 	case IB_WR_SEND_WITH_IMM:
3444 		if (unlikely(rdma_protocol_iwarp(&dev->ibdev, 1))) {
3445 			rc = -EINVAL;
3446 			*bad_wr = wr;
3447 			break;
3448 		}
3449 		wqe->req_type = RDMA_SQ_REQ_TYPE_SEND_WITH_IMM;
3450 		swqe = (struct rdma_sq_send_wqe_1st *)wqe;
3451 		swqe->wqe_size = 2;
3452 		swqe2 = qed_chain_produce(&qp->sq.pbl);
3453 
3454 		swqe->inv_key_or_imm_data = cpu_to_le32(be32_to_cpu(wr->ex.imm_data));
3455 		length = qedr_prepare_sq_send_data(dev, qp, swqe, swqe2,
3456 						   wr, bad_wr);
3457 		swqe->length = cpu_to_le32(length);
3458 		qp->wqe_wr_id[qp->sq.prod].wqe_size = swqe->wqe_size;
3459 		qp->prev_wqe_size = swqe->wqe_size;
3460 		qp->wqe_wr_id[qp->sq.prod].bytes_len = swqe->length;
3461 		break;
3462 	case IB_WR_SEND:
3463 		wqe->req_type = RDMA_SQ_REQ_TYPE_SEND;
3464 		swqe = (struct rdma_sq_send_wqe_1st *)wqe;
3465 
3466 		swqe->wqe_size = 2;
3467 		swqe2 = qed_chain_produce(&qp->sq.pbl);
3468 		length = qedr_prepare_sq_send_data(dev, qp, swqe, swqe2,
3469 						   wr, bad_wr);
3470 		swqe->length = cpu_to_le32(length);
3471 		qp->wqe_wr_id[qp->sq.prod].wqe_size = swqe->wqe_size;
3472 		qp->prev_wqe_size = swqe->wqe_size;
3473 		qp->wqe_wr_id[qp->sq.prod].bytes_len = swqe->length;
3474 		break;
3475 	case IB_WR_SEND_WITH_INV:
3476 		wqe->req_type = RDMA_SQ_REQ_TYPE_SEND_WITH_INVALIDATE;
3477 		swqe = (struct rdma_sq_send_wqe_1st *)wqe;
3478 		swqe2 = qed_chain_produce(&qp->sq.pbl);
3479 		swqe->wqe_size = 2;
3480 		swqe->inv_key_or_imm_data = cpu_to_le32(wr->ex.invalidate_rkey);
3481 		length = qedr_prepare_sq_send_data(dev, qp, swqe, swqe2,
3482 						   wr, bad_wr);
3483 		swqe->length = cpu_to_le32(length);
3484 		qp->wqe_wr_id[qp->sq.prod].wqe_size = swqe->wqe_size;
3485 		qp->prev_wqe_size = swqe->wqe_size;
3486 		qp->wqe_wr_id[qp->sq.prod].bytes_len = swqe->length;
3487 		break;
3488 
3489 	case IB_WR_RDMA_WRITE_WITH_IMM:
3490 		if (unlikely(rdma_protocol_iwarp(&dev->ibdev, 1))) {
3491 			rc = -EINVAL;
3492 			*bad_wr = wr;
3493 			break;
3494 		}
3495 		wqe->req_type = RDMA_SQ_REQ_TYPE_RDMA_WR_WITH_IMM;
3496 		rwqe = (struct rdma_sq_rdma_wqe_1st *)wqe;
3497 
3498 		rwqe->wqe_size = 2;
3499 		rwqe->imm_data = htonl(cpu_to_le32(wr->ex.imm_data));
3500 		rwqe2 = qed_chain_produce(&qp->sq.pbl);
3501 		length = qedr_prepare_sq_rdma_data(dev, qp, rwqe, rwqe2,
3502 						   wr, bad_wr);
3503 		rwqe->length = cpu_to_le32(length);
3504 		qp->wqe_wr_id[qp->sq.prod].wqe_size = rwqe->wqe_size;
3505 		qp->prev_wqe_size = rwqe->wqe_size;
3506 		qp->wqe_wr_id[qp->sq.prod].bytes_len = rwqe->length;
3507 		break;
3508 	case IB_WR_RDMA_WRITE:
3509 		wqe->req_type = RDMA_SQ_REQ_TYPE_RDMA_WR;
3510 		rwqe = (struct rdma_sq_rdma_wqe_1st *)wqe;
3511 
3512 		rwqe->wqe_size = 2;
3513 		rwqe2 = qed_chain_produce(&qp->sq.pbl);
3514 		length = qedr_prepare_sq_rdma_data(dev, qp, rwqe, rwqe2,
3515 						   wr, bad_wr);
3516 		rwqe->length = cpu_to_le32(length);
3517 		qp->wqe_wr_id[qp->sq.prod].wqe_size = rwqe->wqe_size;
3518 		qp->prev_wqe_size = rwqe->wqe_size;
3519 		qp->wqe_wr_id[qp->sq.prod].bytes_len = rwqe->length;
3520 		break;
3521 	case IB_WR_RDMA_READ_WITH_INV:
3522 		SET_FIELD2(wqe->flags, RDMA_SQ_RDMA_WQE_1ST_READ_INV_FLG, 1);
3523 		/* fallthrough -- same is identical to RDMA READ */
3524 
3525 	case IB_WR_RDMA_READ:
3526 		wqe->req_type = RDMA_SQ_REQ_TYPE_RDMA_RD;
3527 		rwqe = (struct rdma_sq_rdma_wqe_1st *)wqe;
3528 
3529 		rwqe->wqe_size = 2;
3530 		rwqe2 = qed_chain_produce(&qp->sq.pbl);
3531 		length = qedr_prepare_sq_rdma_data(dev, qp, rwqe, rwqe2,
3532 						   wr, bad_wr);
3533 		rwqe->length = cpu_to_le32(length);
3534 		qp->wqe_wr_id[qp->sq.prod].wqe_size = rwqe->wqe_size;
3535 		qp->prev_wqe_size = rwqe->wqe_size;
3536 		qp->wqe_wr_id[qp->sq.prod].bytes_len = rwqe->length;
3537 		break;
3538 
3539 	case IB_WR_ATOMIC_CMP_AND_SWP:
3540 	case IB_WR_ATOMIC_FETCH_AND_ADD:
3541 		awqe1 = (struct rdma_sq_atomic_wqe_1st *)wqe;
3542 		awqe1->wqe_size = 4;
3543 
3544 		awqe2 = qed_chain_produce(&qp->sq.pbl);
3545 		DMA_REGPAIR_LE(awqe2->remote_va, atomic_wr(wr)->remote_addr);
3546 		awqe2->r_key = cpu_to_le32(atomic_wr(wr)->rkey);
3547 
3548 		awqe3 = qed_chain_produce(&qp->sq.pbl);
3549 
3550 		if (wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
3551 			wqe->req_type = RDMA_SQ_REQ_TYPE_ATOMIC_ADD;
3552 			DMA_REGPAIR_LE(awqe3->swap_data,
3553 				       atomic_wr(wr)->compare_add);
3554 		} else {
3555 			wqe->req_type = RDMA_SQ_REQ_TYPE_ATOMIC_CMP_AND_SWAP;
3556 			DMA_REGPAIR_LE(awqe3->swap_data,
3557 				       atomic_wr(wr)->swap);
3558 			DMA_REGPAIR_LE(awqe3->cmp_data,
3559 				       atomic_wr(wr)->compare_add);
3560 		}
3561 
3562 		qedr_prepare_sq_sges(qp, NULL, wr);
3563 
3564 		qp->wqe_wr_id[qp->sq.prod].wqe_size = awqe1->wqe_size;
3565 		qp->prev_wqe_size = awqe1->wqe_size;
3566 		break;
3567 
3568 	case IB_WR_LOCAL_INV:
3569 		iwqe = (struct rdma_sq_local_inv_wqe *)wqe;
3570 		iwqe->wqe_size = 1;
3571 
3572 		iwqe->req_type = RDMA_SQ_REQ_TYPE_LOCAL_INVALIDATE;
3573 		iwqe->inv_l_key = wr->ex.invalidate_rkey;
3574 		qp->wqe_wr_id[qp->sq.prod].wqe_size = iwqe->wqe_size;
3575 		qp->prev_wqe_size = iwqe->wqe_size;
3576 		break;
3577 	case IB_WR_REG_MR:
3578 		DP_DEBUG(dev, QEDR_MSG_CQ, "REG_MR\n");
3579 		wqe->req_type = RDMA_SQ_REQ_TYPE_FAST_MR;
3580 		fwqe1 = (struct rdma_sq_fmr_wqe_1st *)wqe;
3581 		fwqe1->wqe_size = 2;
3582 
3583 		rc = qedr_prepare_reg(qp, fwqe1, reg_wr(wr));
3584 		if (rc) {
3585 			DP_ERR(dev, "IB_REG_MR failed rc=%d\n", rc);
3586 			*bad_wr = wr;
3587 			break;
3588 		}
3589 
3590 		qp->wqe_wr_id[qp->sq.prod].wqe_size = fwqe1->wqe_size;
3591 		qp->prev_wqe_size = fwqe1->wqe_size;
3592 		break;
3593 	default:
3594 		DP_ERR(dev, "invalid opcode 0x%x!\n", wr->opcode);
3595 		rc = -EINVAL;
3596 		*bad_wr = wr;
3597 		break;
3598 	}
3599 
3600 	if (*bad_wr) {
3601 		u16 value;
3602 
3603 		/* Restore prod to its position before
3604 		 * this WR was processed
3605 		 */
3606 		value = le16_to_cpu(qp->sq.db_data.data.value);
3607 		qed_chain_set_prod(&qp->sq.pbl, value, wqe);
3608 
3609 		/* Restore prev_wqe_size */
3610 		qp->prev_wqe_size = wqe->prev_wqe_size;
3611 		rc = -EINVAL;
3612 		DP_ERR(dev, "POST SEND FAILED\n");
3613 	}
3614 
3615 	return rc;
3616 }
3617 
3618 int qedr_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
3619 		   const struct ib_send_wr **bad_wr)
3620 {
3621 	struct qedr_dev *dev = get_qedr_dev(ibqp->device);
3622 	struct qedr_qp *qp = get_qedr_qp(ibqp);
3623 	unsigned long flags;
3624 	int rc = 0;
3625 
3626 	*bad_wr = NULL;
3627 
3628 	if (qp->qp_type == IB_QPT_GSI)
3629 		return qedr_gsi_post_send(ibqp, wr, bad_wr);
3630 
3631 	spin_lock_irqsave(&qp->q_lock, flags);
3632 
3633 	if (rdma_protocol_roce(&dev->ibdev, 1)) {
3634 		if ((qp->state != QED_ROCE_QP_STATE_RTS) &&
3635 		    (qp->state != QED_ROCE_QP_STATE_ERR) &&
3636 		    (qp->state != QED_ROCE_QP_STATE_SQD)) {
3637 			spin_unlock_irqrestore(&qp->q_lock, flags);
3638 			*bad_wr = wr;
3639 			DP_DEBUG(dev, QEDR_MSG_CQ,
3640 				 "QP in wrong state! QP icid=0x%x state %d\n",
3641 				 qp->icid, qp->state);
3642 			return -EINVAL;
3643 		}
3644 	}
3645 
3646 	while (wr) {
3647 		rc = __qedr_post_send(ibqp, wr, bad_wr);
3648 		if (rc)
3649 			break;
3650 
3651 		qp->wqe_wr_id[qp->sq.prod].wr_id = wr->wr_id;
3652 
3653 		qedr_inc_sw_prod(&qp->sq);
3654 
3655 		qp->sq.db_data.data.value++;
3656 
3657 		wr = wr->next;
3658 	}
3659 
3660 	/* Trigger doorbell
3661 	 * If there was a failure in the first WR then it will be triggered in
3662 	 * vane. However this is not harmful (as long as the producer value is
3663 	 * unchanged). For performance reasons we avoid checking for this
3664 	 * redundant doorbell.
3665 	 *
3666 	 * qp->wqe_wr_id is accessed during qedr_poll_cq, as
3667 	 * soon as we give the doorbell, we could get a completion
3668 	 * for this wr, therefore we need to make sure that the
3669 	 * memory is updated before giving the doorbell.
3670 	 * During qedr_poll_cq, rmb is called before accessing the
3671 	 * cqe. This covers for the smp_rmb as well.
3672 	 */
3673 	smp_wmb();
3674 	writel(qp->sq.db_data.raw, qp->sq.db);
3675 
3676 	spin_unlock_irqrestore(&qp->q_lock, flags);
3677 
3678 	return rc;
3679 }
3680 
3681 static u32 qedr_srq_elem_left(struct qedr_srq_hwq_info *hw_srq)
3682 {
3683 	u32 used;
3684 
3685 	/* Calculate number of elements used based on producer
3686 	 * count and consumer count and subtract it from max
3687 	 * work request supported so that we get elements left.
3688 	 */
3689 	used = hw_srq->wr_prod_cnt - hw_srq->wr_cons_cnt;
3690 
3691 	return hw_srq->max_wr - used;
3692 }
3693 
3694 int qedr_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
3695 		       const struct ib_recv_wr **bad_wr)
3696 {
3697 	struct qedr_srq *srq = get_qedr_srq(ibsrq);
3698 	struct qedr_srq_hwq_info *hw_srq;
3699 	struct qedr_dev *dev = srq->dev;
3700 	struct qed_chain *pbl;
3701 	unsigned long flags;
3702 	int status = 0;
3703 	u32 num_sge;
3704 	u32 offset;
3705 
3706 	spin_lock_irqsave(&srq->lock, flags);
3707 
3708 	hw_srq = &srq->hw_srq;
3709 	pbl = &srq->hw_srq.pbl;
3710 	while (wr) {
3711 		struct rdma_srq_wqe_header *hdr;
3712 		int i;
3713 
3714 		if (!qedr_srq_elem_left(hw_srq) ||
3715 		    wr->num_sge > srq->hw_srq.max_sges) {
3716 			DP_ERR(dev, "Can't post WR  (%d,%d) || (%d > %d)\n",
3717 			       hw_srq->wr_prod_cnt, hw_srq->wr_cons_cnt,
3718 			       wr->num_sge, srq->hw_srq.max_sges);
3719 			status = -ENOMEM;
3720 			*bad_wr = wr;
3721 			break;
3722 		}
3723 
3724 		hdr = qed_chain_produce(pbl);
3725 		num_sge = wr->num_sge;
3726 		/* Set number of sge and work request id in header */
3727 		SRQ_HDR_SET(hdr, wr->wr_id, num_sge);
3728 
3729 		srq->hw_srq.wr_prod_cnt++;
3730 		hw_srq->wqe_prod++;
3731 		hw_srq->sge_prod++;
3732 
3733 		DP_DEBUG(dev, QEDR_MSG_SRQ,
3734 			 "SRQ WR: SGEs: %d with wr_id[%d] = %llx\n",
3735 			 wr->num_sge, hw_srq->wqe_prod, wr->wr_id);
3736 
3737 		for (i = 0; i < wr->num_sge; i++) {
3738 			struct rdma_srq_sge *srq_sge = qed_chain_produce(pbl);
3739 
3740 			/* Set SGE length, lkey and address */
3741 			SRQ_SGE_SET(srq_sge, wr->sg_list[i].addr,
3742 				    wr->sg_list[i].length, wr->sg_list[i].lkey);
3743 
3744 			DP_DEBUG(dev, QEDR_MSG_SRQ,
3745 				 "[%d]: len %d key %x addr %x:%x\n",
3746 				 i, srq_sge->length, srq_sge->l_key,
3747 				 srq_sge->addr.hi, srq_sge->addr.lo);
3748 			hw_srq->sge_prod++;
3749 		}
3750 
3751 		/* Flush WQE and SGE information before
3752 		 * updating producer.
3753 		 */
3754 		wmb();
3755 
3756 		/* SRQ producer is 8 bytes. Need to update SGE producer index
3757 		 * in first 4 bytes and need to update WQE producer in
3758 		 * next 4 bytes.
3759 		 */
3760 		*srq->hw_srq.virt_prod_pair_addr = hw_srq->sge_prod;
3761 		offset = offsetof(struct rdma_srq_producers, wqe_prod);
3762 		*((u8 *)srq->hw_srq.virt_prod_pair_addr + offset) =
3763 			hw_srq->wqe_prod;
3764 
3765 		/* Flush producer after updating it. */
3766 		wmb();
3767 		wr = wr->next;
3768 	}
3769 
3770 	DP_DEBUG(dev, QEDR_MSG_SRQ, "POST: Elements in S-RQ: %d\n",
3771 		 qed_chain_get_elem_left(pbl));
3772 	spin_unlock_irqrestore(&srq->lock, flags);
3773 
3774 	return status;
3775 }
3776 
3777 int qedr_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
3778 		   const struct ib_recv_wr **bad_wr)
3779 {
3780 	struct qedr_qp *qp = get_qedr_qp(ibqp);
3781 	struct qedr_dev *dev = qp->dev;
3782 	unsigned long flags;
3783 	int status = 0;
3784 
3785 	if (qp->qp_type == IB_QPT_GSI)
3786 		return qedr_gsi_post_recv(ibqp, wr, bad_wr);
3787 
3788 	spin_lock_irqsave(&qp->q_lock, flags);
3789 
3790 	if (qp->state == QED_ROCE_QP_STATE_RESET) {
3791 		spin_unlock_irqrestore(&qp->q_lock, flags);
3792 		*bad_wr = wr;
3793 		return -EINVAL;
3794 	}
3795 
3796 	while (wr) {
3797 		int i;
3798 
3799 		if (qed_chain_get_elem_left_u32(&qp->rq.pbl) <
3800 		    QEDR_MAX_RQE_ELEMENTS_PER_RQE ||
3801 		    wr->num_sge > qp->rq.max_sges) {
3802 			DP_ERR(dev, "Can't post WR  (%d < %d) || (%d > %d)\n",
3803 			       qed_chain_get_elem_left_u32(&qp->rq.pbl),
3804 			       QEDR_MAX_RQE_ELEMENTS_PER_RQE, wr->num_sge,
3805 			       qp->rq.max_sges);
3806 			status = -ENOMEM;
3807 			*bad_wr = wr;
3808 			break;
3809 		}
3810 		for (i = 0; i < wr->num_sge; i++) {
3811 			u32 flags = 0;
3812 			struct rdma_rq_sge *rqe =
3813 			    qed_chain_produce(&qp->rq.pbl);
3814 
3815 			/* First one must include the number
3816 			 * of SGE in the list
3817 			 */
3818 			if (!i)
3819 				SET_FIELD(flags, RDMA_RQ_SGE_NUM_SGES,
3820 					  wr->num_sge);
3821 
3822 			SET_FIELD(flags, RDMA_RQ_SGE_L_KEY_LO,
3823 				  wr->sg_list[i].lkey);
3824 
3825 			RQ_SGE_SET(rqe, wr->sg_list[i].addr,
3826 				   wr->sg_list[i].length, flags);
3827 		}
3828 
3829 		/* Special case of no sges. FW requires between 1-4 sges...
3830 		 * in this case we need to post 1 sge with length zero. this is
3831 		 * because rdma write with immediate consumes an RQ.
3832 		 */
3833 		if (!wr->num_sge) {
3834 			u32 flags = 0;
3835 			struct rdma_rq_sge *rqe =
3836 			    qed_chain_produce(&qp->rq.pbl);
3837 
3838 			/* First one must include the number
3839 			 * of SGE in the list
3840 			 */
3841 			SET_FIELD(flags, RDMA_RQ_SGE_L_KEY_LO, 0);
3842 			SET_FIELD(flags, RDMA_RQ_SGE_NUM_SGES, 1);
3843 
3844 			RQ_SGE_SET(rqe, 0, 0, flags);
3845 			i = 1;
3846 		}
3847 
3848 		qp->rqe_wr_id[qp->rq.prod].wr_id = wr->wr_id;
3849 		qp->rqe_wr_id[qp->rq.prod].wqe_size = i;
3850 
3851 		qedr_inc_sw_prod(&qp->rq);
3852 
3853 		/* qp->rqe_wr_id is accessed during qedr_poll_cq, as
3854 		 * soon as we give the doorbell, we could get a completion
3855 		 * for this wr, therefore we need to make sure that the
3856 		 * memory is update before giving the doorbell.
3857 		 * During qedr_poll_cq, rmb is called before accessing the
3858 		 * cqe. This covers for the smp_rmb as well.
3859 		 */
3860 		smp_wmb();
3861 
3862 		qp->rq.db_data.data.value++;
3863 
3864 		writel(qp->rq.db_data.raw, qp->rq.db);
3865 
3866 		if (rdma_protocol_iwarp(&dev->ibdev, 1)) {
3867 			writel(qp->rq.iwarp_db2_data.raw, qp->rq.iwarp_db2);
3868 		}
3869 
3870 		wr = wr->next;
3871 	}
3872 
3873 	spin_unlock_irqrestore(&qp->q_lock, flags);
3874 
3875 	return status;
3876 }
3877 
3878 static int is_valid_cqe(struct qedr_cq *cq, union rdma_cqe *cqe)
3879 {
3880 	struct rdma_cqe_requester *resp_cqe = &cqe->req;
3881 
3882 	return (resp_cqe->flags & RDMA_CQE_REQUESTER_TOGGLE_BIT_MASK) ==
3883 		cq->pbl_toggle;
3884 }
3885 
3886 static struct qedr_qp *cqe_get_qp(union rdma_cqe *cqe)
3887 {
3888 	struct rdma_cqe_requester *resp_cqe = &cqe->req;
3889 	struct qedr_qp *qp;
3890 
3891 	qp = (struct qedr_qp *)(uintptr_t)HILO_GEN(resp_cqe->qp_handle.hi,
3892 						   resp_cqe->qp_handle.lo,
3893 						   u64);
3894 	return qp;
3895 }
3896 
3897 static enum rdma_cqe_type cqe_get_type(union rdma_cqe *cqe)
3898 {
3899 	struct rdma_cqe_requester *resp_cqe = &cqe->req;
3900 
3901 	return GET_FIELD(resp_cqe->flags, RDMA_CQE_REQUESTER_TYPE);
3902 }
3903 
3904 /* Return latest CQE (needs processing) */
3905 static union rdma_cqe *get_cqe(struct qedr_cq *cq)
3906 {
3907 	return cq->latest_cqe;
3908 }
3909 
3910 /* In fmr we need to increase the number of fmr completed counter for the fmr
3911  * algorithm determining whether we can free a pbl or not.
3912  * we need to perform this whether the work request was signaled or not. for
3913  * this purpose we call this function from the condition that checks if a wr
3914  * should be skipped, to make sure we don't miss it ( possibly this fmr
3915  * operation was not signalted)
3916  */
3917 static inline void qedr_chk_if_fmr(struct qedr_qp *qp)
3918 {
3919 	if (qp->wqe_wr_id[qp->sq.cons].opcode == IB_WC_REG_MR)
3920 		qp->wqe_wr_id[qp->sq.cons].mr->info.completed++;
3921 }
3922 
3923 static int process_req(struct qedr_dev *dev, struct qedr_qp *qp,
3924 		       struct qedr_cq *cq, int num_entries,
3925 		       struct ib_wc *wc, u16 hw_cons, enum ib_wc_status status,
3926 		       int force)
3927 {
3928 	u16 cnt = 0;
3929 
3930 	while (num_entries && qp->sq.wqe_cons != hw_cons) {
3931 		if (!qp->wqe_wr_id[qp->sq.cons].signaled && !force) {
3932 			qedr_chk_if_fmr(qp);
3933 			/* skip WC */
3934 			goto next_cqe;
3935 		}
3936 
3937 		/* fill WC */
3938 		wc->status = status;
3939 		wc->vendor_err = 0;
3940 		wc->wc_flags = 0;
3941 		wc->src_qp = qp->id;
3942 		wc->qp = &qp->ibqp;
3943 
3944 		wc->wr_id = qp->wqe_wr_id[qp->sq.cons].wr_id;
3945 		wc->opcode = qp->wqe_wr_id[qp->sq.cons].opcode;
3946 
3947 		switch (wc->opcode) {
3948 		case IB_WC_RDMA_WRITE:
3949 			wc->byte_len = qp->wqe_wr_id[qp->sq.cons].bytes_len;
3950 			break;
3951 		case IB_WC_COMP_SWAP:
3952 		case IB_WC_FETCH_ADD:
3953 			wc->byte_len = 8;
3954 			break;
3955 		case IB_WC_REG_MR:
3956 			qp->wqe_wr_id[qp->sq.cons].mr->info.completed++;
3957 			break;
3958 		case IB_WC_RDMA_READ:
3959 		case IB_WC_SEND:
3960 			wc->byte_len = qp->wqe_wr_id[qp->sq.cons].bytes_len;
3961 			break;
3962 		default:
3963 			break;
3964 		}
3965 
3966 		num_entries--;
3967 		wc++;
3968 		cnt++;
3969 next_cqe:
3970 		while (qp->wqe_wr_id[qp->sq.cons].wqe_size--)
3971 			qed_chain_consume(&qp->sq.pbl);
3972 		qedr_inc_sw_cons(&qp->sq);
3973 	}
3974 
3975 	return cnt;
3976 }
3977 
3978 static int qedr_poll_cq_req(struct qedr_dev *dev,
3979 			    struct qedr_qp *qp, struct qedr_cq *cq,
3980 			    int num_entries, struct ib_wc *wc,
3981 			    struct rdma_cqe_requester *req)
3982 {
3983 	int cnt = 0;
3984 
3985 	switch (req->status) {
3986 	case RDMA_CQE_REQ_STS_OK:
3987 		cnt = process_req(dev, qp, cq, num_entries, wc, req->sq_cons,
3988 				  IB_WC_SUCCESS, 0);
3989 		break;
3990 	case RDMA_CQE_REQ_STS_WORK_REQUEST_FLUSHED_ERR:
3991 		if (qp->state != QED_ROCE_QP_STATE_ERR)
3992 			DP_DEBUG(dev, QEDR_MSG_CQ,
3993 				 "Error: POLL CQ with RDMA_CQE_REQ_STS_WORK_REQUEST_FLUSHED_ERR. CQ icid=0x%x, QP icid=0x%x\n",
3994 				 cq->icid, qp->icid);
3995 		cnt = process_req(dev, qp, cq, num_entries, wc, req->sq_cons,
3996 				  IB_WC_WR_FLUSH_ERR, 1);
3997 		break;
3998 	default:
3999 		/* process all WQE before the cosumer */
4000 		qp->state = QED_ROCE_QP_STATE_ERR;
4001 		cnt = process_req(dev, qp, cq, num_entries, wc,
4002 				  req->sq_cons - 1, IB_WC_SUCCESS, 0);
4003 		wc += cnt;
4004 		/* if we have extra WC fill it with actual error info */
4005 		if (cnt < num_entries) {
4006 			enum ib_wc_status wc_status;
4007 
4008 			switch (req->status) {
4009 			case RDMA_CQE_REQ_STS_BAD_RESPONSE_ERR:
4010 				DP_ERR(dev,
4011 				       "Error: POLL CQ with RDMA_CQE_REQ_STS_BAD_RESPONSE_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4012 				       cq->icid, qp->icid);
4013 				wc_status = IB_WC_BAD_RESP_ERR;
4014 				break;
4015 			case RDMA_CQE_REQ_STS_LOCAL_LENGTH_ERR:
4016 				DP_ERR(dev,
4017 				       "Error: POLL CQ with RDMA_CQE_REQ_STS_LOCAL_LENGTH_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4018 				       cq->icid, qp->icid);
4019 				wc_status = IB_WC_LOC_LEN_ERR;
4020 				break;
4021 			case RDMA_CQE_REQ_STS_LOCAL_QP_OPERATION_ERR:
4022 				DP_ERR(dev,
4023 				       "Error: POLL CQ with RDMA_CQE_REQ_STS_LOCAL_QP_OPERATION_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4024 				       cq->icid, qp->icid);
4025 				wc_status = IB_WC_LOC_QP_OP_ERR;
4026 				break;
4027 			case RDMA_CQE_REQ_STS_LOCAL_PROTECTION_ERR:
4028 				DP_ERR(dev,
4029 				       "Error: POLL CQ with RDMA_CQE_REQ_STS_LOCAL_PROTECTION_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4030 				       cq->icid, qp->icid);
4031 				wc_status = IB_WC_LOC_PROT_ERR;
4032 				break;
4033 			case RDMA_CQE_REQ_STS_MEMORY_MGT_OPERATION_ERR:
4034 				DP_ERR(dev,
4035 				       "Error: POLL CQ with RDMA_CQE_REQ_STS_MEMORY_MGT_OPERATION_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4036 				       cq->icid, qp->icid);
4037 				wc_status = IB_WC_MW_BIND_ERR;
4038 				break;
4039 			case RDMA_CQE_REQ_STS_REMOTE_INVALID_REQUEST_ERR:
4040 				DP_ERR(dev,
4041 				       "Error: POLL CQ with RDMA_CQE_REQ_STS_REMOTE_INVALID_REQUEST_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4042 				       cq->icid, qp->icid);
4043 				wc_status = IB_WC_REM_INV_REQ_ERR;
4044 				break;
4045 			case RDMA_CQE_REQ_STS_REMOTE_ACCESS_ERR:
4046 				DP_ERR(dev,
4047 				       "Error: POLL CQ with RDMA_CQE_REQ_STS_REMOTE_ACCESS_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4048 				       cq->icid, qp->icid);
4049 				wc_status = IB_WC_REM_ACCESS_ERR;
4050 				break;
4051 			case RDMA_CQE_REQ_STS_REMOTE_OPERATION_ERR:
4052 				DP_ERR(dev,
4053 				       "Error: POLL CQ with RDMA_CQE_REQ_STS_REMOTE_OPERATION_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4054 				       cq->icid, qp->icid);
4055 				wc_status = IB_WC_REM_OP_ERR;
4056 				break;
4057 			case RDMA_CQE_REQ_STS_RNR_NAK_RETRY_CNT_ERR:
4058 				DP_ERR(dev,
4059 				       "Error: POLL CQ with RDMA_CQE_REQ_STS_RNR_NAK_RETRY_CNT_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4060 				       cq->icid, qp->icid);
4061 				wc_status = IB_WC_RNR_RETRY_EXC_ERR;
4062 				break;
4063 			case RDMA_CQE_REQ_STS_TRANSPORT_RETRY_CNT_ERR:
4064 				DP_ERR(dev,
4065 				       "Error: POLL CQ with ROCE_CQE_REQ_STS_TRANSPORT_RETRY_CNT_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4066 				       cq->icid, qp->icid);
4067 				wc_status = IB_WC_RETRY_EXC_ERR;
4068 				break;
4069 			default:
4070 				DP_ERR(dev,
4071 				       "Error: POLL CQ with IB_WC_GENERAL_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4072 				       cq->icid, qp->icid);
4073 				wc_status = IB_WC_GENERAL_ERR;
4074 			}
4075 			cnt += process_req(dev, qp, cq, 1, wc, req->sq_cons,
4076 					   wc_status, 1);
4077 		}
4078 	}
4079 
4080 	return cnt;
4081 }
4082 
4083 static inline int qedr_cqe_resp_status_to_ib(u8 status)
4084 {
4085 	switch (status) {
4086 	case RDMA_CQE_RESP_STS_LOCAL_ACCESS_ERR:
4087 		return IB_WC_LOC_ACCESS_ERR;
4088 	case RDMA_CQE_RESP_STS_LOCAL_LENGTH_ERR:
4089 		return IB_WC_LOC_LEN_ERR;
4090 	case RDMA_CQE_RESP_STS_LOCAL_QP_OPERATION_ERR:
4091 		return IB_WC_LOC_QP_OP_ERR;
4092 	case RDMA_CQE_RESP_STS_LOCAL_PROTECTION_ERR:
4093 		return IB_WC_LOC_PROT_ERR;
4094 	case RDMA_CQE_RESP_STS_MEMORY_MGT_OPERATION_ERR:
4095 		return IB_WC_MW_BIND_ERR;
4096 	case RDMA_CQE_RESP_STS_REMOTE_INVALID_REQUEST_ERR:
4097 		return IB_WC_REM_INV_RD_REQ_ERR;
4098 	case RDMA_CQE_RESP_STS_OK:
4099 		return IB_WC_SUCCESS;
4100 	default:
4101 		return IB_WC_GENERAL_ERR;
4102 	}
4103 }
4104 
4105 static inline int qedr_set_ok_cqe_resp_wc(struct rdma_cqe_responder *resp,
4106 					  struct ib_wc *wc)
4107 {
4108 	wc->status = IB_WC_SUCCESS;
4109 	wc->byte_len = le32_to_cpu(resp->length);
4110 
4111 	if (resp->flags & QEDR_RESP_IMM) {
4112 		wc->ex.imm_data = cpu_to_be32(le32_to_cpu(resp->imm_data_or_inv_r_Key));
4113 		wc->wc_flags |= IB_WC_WITH_IMM;
4114 
4115 		if (resp->flags & QEDR_RESP_RDMA)
4116 			wc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
4117 
4118 		if (resp->flags & QEDR_RESP_INV)
4119 			return -EINVAL;
4120 
4121 	} else if (resp->flags & QEDR_RESP_INV) {
4122 		wc->ex.imm_data = le32_to_cpu(resp->imm_data_or_inv_r_Key);
4123 		wc->wc_flags |= IB_WC_WITH_INVALIDATE;
4124 
4125 		if (resp->flags & QEDR_RESP_RDMA)
4126 			return -EINVAL;
4127 
4128 	} else if (resp->flags & QEDR_RESP_RDMA) {
4129 		return -EINVAL;
4130 	}
4131 
4132 	return 0;
4133 }
4134 
4135 static void __process_resp_one(struct qedr_dev *dev, struct qedr_qp *qp,
4136 			       struct qedr_cq *cq, struct ib_wc *wc,
4137 			       struct rdma_cqe_responder *resp, u64 wr_id)
4138 {
4139 	/* Must fill fields before qedr_set_ok_cqe_resp_wc() */
4140 	wc->opcode = IB_WC_RECV;
4141 	wc->wc_flags = 0;
4142 
4143 	if (likely(resp->status == RDMA_CQE_RESP_STS_OK)) {
4144 		if (qedr_set_ok_cqe_resp_wc(resp, wc))
4145 			DP_ERR(dev,
4146 			       "CQ %p (icid=%d) has invalid CQE responder flags=0x%x\n",
4147 			       cq, cq->icid, resp->flags);
4148 
4149 	} else {
4150 		wc->status = qedr_cqe_resp_status_to_ib(resp->status);
4151 		if (wc->status == IB_WC_GENERAL_ERR)
4152 			DP_ERR(dev,
4153 			       "CQ %p (icid=%d) contains an invalid CQE status %d\n",
4154 			       cq, cq->icid, resp->status);
4155 	}
4156 
4157 	/* Fill the rest of the WC */
4158 	wc->vendor_err = 0;
4159 	wc->src_qp = qp->id;
4160 	wc->qp = &qp->ibqp;
4161 	wc->wr_id = wr_id;
4162 }
4163 
4164 static int process_resp_one_srq(struct qedr_dev *dev, struct qedr_qp *qp,
4165 				struct qedr_cq *cq, struct ib_wc *wc,
4166 				struct rdma_cqe_responder *resp)
4167 {
4168 	struct qedr_srq *srq = qp->srq;
4169 	u64 wr_id;
4170 
4171 	wr_id = HILO_GEN(le32_to_cpu(resp->srq_wr_id.hi),
4172 			 le32_to_cpu(resp->srq_wr_id.lo), u64);
4173 
4174 	if (resp->status == RDMA_CQE_RESP_STS_WORK_REQUEST_FLUSHED_ERR) {
4175 		wc->status = IB_WC_WR_FLUSH_ERR;
4176 		wc->vendor_err = 0;
4177 		wc->wr_id = wr_id;
4178 		wc->byte_len = 0;
4179 		wc->src_qp = qp->id;
4180 		wc->qp = &qp->ibqp;
4181 		wc->wr_id = wr_id;
4182 	} else {
4183 		__process_resp_one(dev, qp, cq, wc, resp, wr_id);
4184 	}
4185 	srq->hw_srq.wr_cons_cnt++;
4186 
4187 	return 1;
4188 }
4189 static int process_resp_one(struct qedr_dev *dev, struct qedr_qp *qp,
4190 			    struct qedr_cq *cq, struct ib_wc *wc,
4191 			    struct rdma_cqe_responder *resp)
4192 {
4193 	u64 wr_id = qp->rqe_wr_id[qp->rq.cons].wr_id;
4194 
4195 	__process_resp_one(dev, qp, cq, wc, resp, wr_id);
4196 
4197 	while (qp->rqe_wr_id[qp->rq.cons].wqe_size--)
4198 		qed_chain_consume(&qp->rq.pbl);
4199 	qedr_inc_sw_cons(&qp->rq);
4200 
4201 	return 1;
4202 }
4203 
4204 static int process_resp_flush(struct qedr_qp *qp, struct qedr_cq *cq,
4205 			      int num_entries, struct ib_wc *wc, u16 hw_cons)
4206 {
4207 	u16 cnt = 0;
4208 
4209 	while (num_entries && qp->rq.wqe_cons != hw_cons) {
4210 		/* fill WC */
4211 		wc->status = IB_WC_WR_FLUSH_ERR;
4212 		wc->vendor_err = 0;
4213 		wc->wc_flags = 0;
4214 		wc->src_qp = qp->id;
4215 		wc->byte_len = 0;
4216 		wc->wr_id = qp->rqe_wr_id[qp->rq.cons].wr_id;
4217 		wc->qp = &qp->ibqp;
4218 		num_entries--;
4219 		wc++;
4220 		cnt++;
4221 		while (qp->rqe_wr_id[qp->rq.cons].wqe_size--)
4222 			qed_chain_consume(&qp->rq.pbl);
4223 		qedr_inc_sw_cons(&qp->rq);
4224 	}
4225 
4226 	return cnt;
4227 }
4228 
4229 static void try_consume_resp_cqe(struct qedr_cq *cq, struct qedr_qp *qp,
4230 				 struct rdma_cqe_responder *resp, int *update)
4231 {
4232 	if (le16_to_cpu(resp->rq_cons_or_srq_id) == qp->rq.wqe_cons) {
4233 		consume_cqe(cq);
4234 		*update |= 1;
4235 	}
4236 }
4237 
4238 static int qedr_poll_cq_resp_srq(struct qedr_dev *dev, struct qedr_qp *qp,
4239 				 struct qedr_cq *cq, int num_entries,
4240 				 struct ib_wc *wc,
4241 				 struct rdma_cqe_responder *resp)
4242 {
4243 	int cnt;
4244 
4245 	cnt = process_resp_one_srq(dev, qp, cq, wc, resp);
4246 	consume_cqe(cq);
4247 
4248 	return cnt;
4249 }
4250 
4251 static int qedr_poll_cq_resp(struct qedr_dev *dev, struct qedr_qp *qp,
4252 			     struct qedr_cq *cq, int num_entries,
4253 			     struct ib_wc *wc, struct rdma_cqe_responder *resp,
4254 			     int *update)
4255 {
4256 	int cnt;
4257 
4258 	if (resp->status == RDMA_CQE_RESP_STS_WORK_REQUEST_FLUSHED_ERR) {
4259 		cnt = process_resp_flush(qp, cq, num_entries, wc,
4260 					 resp->rq_cons_or_srq_id);
4261 		try_consume_resp_cqe(cq, qp, resp, update);
4262 	} else {
4263 		cnt = process_resp_one(dev, qp, cq, wc, resp);
4264 		consume_cqe(cq);
4265 		*update |= 1;
4266 	}
4267 
4268 	return cnt;
4269 }
4270 
4271 static void try_consume_req_cqe(struct qedr_cq *cq, struct qedr_qp *qp,
4272 				struct rdma_cqe_requester *req, int *update)
4273 {
4274 	if (le16_to_cpu(req->sq_cons) == qp->sq.wqe_cons) {
4275 		consume_cqe(cq);
4276 		*update |= 1;
4277 	}
4278 }
4279 
4280 int qedr_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
4281 {
4282 	struct qedr_dev *dev = get_qedr_dev(ibcq->device);
4283 	struct qedr_cq *cq = get_qedr_cq(ibcq);
4284 	union rdma_cqe *cqe;
4285 	u32 old_cons, new_cons;
4286 	unsigned long flags;
4287 	int update = 0;
4288 	int done = 0;
4289 
4290 	if (cq->destroyed) {
4291 		DP_ERR(dev,
4292 		       "warning: poll was invoked after destroy for cq %p (icid=%d)\n",
4293 		       cq, cq->icid);
4294 		return 0;
4295 	}
4296 
4297 	if (cq->cq_type == QEDR_CQ_TYPE_GSI)
4298 		return qedr_gsi_poll_cq(ibcq, num_entries, wc);
4299 
4300 	spin_lock_irqsave(&cq->cq_lock, flags);
4301 	cqe = cq->latest_cqe;
4302 	old_cons = qed_chain_get_cons_idx_u32(&cq->pbl);
4303 	while (num_entries && is_valid_cqe(cq, cqe)) {
4304 		struct qedr_qp *qp;
4305 		int cnt = 0;
4306 
4307 		/* prevent speculative reads of any field of CQE */
4308 		rmb();
4309 
4310 		qp = cqe_get_qp(cqe);
4311 		if (!qp) {
4312 			WARN(1, "Error: CQE QP pointer is NULL. CQE=%p\n", cqe);
4313 			break;
4314 		}
4315 
4316 		wc->qp = &qp->ibqp;
4317 
4318 		switch (cqe_get_type(cqe)) {
4319 		case RDMA_CQE_TYPE_REQUESTER:
4320 			cnt = qedr_poll_cq_req(dev, qp, cq, num_entries, wc,
4321 					       &cqe->req);
4322 			try_consume_req_cqe(cq, qp, &cqe->req, &update);
4323 			break;
4324 		case RDMA_CQE_TYPE_RESPONDER_RQ:
4325 			cnt = qedr_poll_cq_resp(dev, qp, cq, num_entries, wc,
4326 						&cqe->resp, &update);
4327 			break;
4328 		case RDMA_CQE_TYPE_RESPONDER_SRQ:
4329 			cnt = qedr_poll_cq_resp_srq(dev, qp, cq, num_entries,
4330 						    wc, &cqe->resp);
4331 			update = 1;
4332 			break;
4333 		case RDMA_CQE_TYPE_INVALID:
4334 		default:
4335 			DP_ERR(dev, "Error: invalid CQE type = %d\n",
4336 			       cqe_get_type(cqe));
4337 		}
4338 		num_entries -= cnt;
4339 		wc += cnt;
4340 		done += cnt;
4341 
4342 		cqe = get_cqe(cq);
4343 	}
4344 	new_cons = qed_chain_get_cons_idx_u32(&cq->pbl);
4345 
4346 	cq->cq_cons += new_cons - old_cons;
4347 
4348 	if (update)
4349 		/* doorbell notifies abount latest VALID entry,
4350 		 * but chain already point to the next INVALID one
4351 		 */
4352 		doorbell_cq(cq, cq->cq_cons - 1, cq->arm_flags);
4353 
4354 	spin_unlock_irqrestore(&cq->cq_lock, flags);
4355 	return done;
4356 }
4357 
4358 int qedr_process_mad(struct ib_device *ibdev, int process_mad_flags,
4359 		     u8 port_num, const struct ib_wc *in_wc,
4360 		     const struct ib_grh *in_grh, const struct ib_mad *in,
4361 		     struct ib_mad *out_mad, size_t *out_mad_size,
4362 		     u16 *out_mad_pkey_index)
4363 {
4364 	return IB_MAD_RESULT_SUCCESS;
4365 }
4366