1 /* This file is part of the Emulex RoCE Device Driver for
2  * RoCE (RDMA over Converged Ethernet) adapters.
3  * Copyright (C) 2012-2015 Emulex. All rights reserved.
4  * EMULEX and SLI are trademarks of Emulex.
5  * www.emulex.com
6  *
7  * This software is available to you under a choice of one of two licenses.
8  * You may choose to be licensed under the terms of the GNU General Public
9  * License (GPL) Version 2, available from the file COPYING in the main
10  * directory of this source tree, or the BSD license below:
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  *
16  * - Redistributions of source code must retain the above copyright notice,
17  *   this list of conditions and the following disclaimer.
18  *
19  * - Redistributions in binary form must reproduce the above copyright
20  *   notice, this list of conditions and the following disclaimer in
21  *   the documentation and/or other materials provided with the distribution.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
24  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
27  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
28  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
29  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
30  * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
31  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
32  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
33  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
34  *
35  * Contact Information:
36  * linux-drivers@emulex.com
37  *
38  * Emulex
39  * 3333 Susan Street
40  * Costa Mesa, CA 92626
41  */
42 
43 #include <linux/dma-mapping.h>
44 #include <rdma/ib_verbs.h>
45 #include <rdma/ib_user_verbs.h>
46 #include <rdma/iw_cm.h>
47 #include <rdma/ib_umem.h>
48 #include <rdma/ib_addr.h>
49 #include <rdma/ib_cache.h>
50 #include <rdma/uverbs_ioctl.h>
51 
52 #include "ocrdma.h"
53 #include "ocrdma_hw.h"
54 #include "ocrdma_verbs.h"
55 #include <rdma/ocrdma-abi.h>
56 
57 int ocrdma_query_pkey(struct ib_device *ibdev, u8 port, u16 index, u16 *pkey)
58 {
59 	if (index > 0)
60 		return -EINVAL;
61 
62 	*pkey = 0xffff;
63 	return 0;
64 }
65 
66 int ocrdma_query_device(struct ib_device *ibdev, struct ib_device_attr *attr,
67 			struct ib_udata *uhw)
68 {
69 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
70 
71 	if (uhw->inlen || uhw->outlen)
72 		return -EINVAL;
73 
74 	memset(attr, 0, sizeof *attr);
75 	memcpy(&attr->fw_ver, &dev->attr.fw_ver[0],
76 	       min(sizeof(dev->attr.fw_ver), sizeof(attr->fw_ver)));
77 	ocrdma_get_guid(dev, (u8 *)&attr->sys_image_guid);
78 	attr->max_mr_size = dev->attr.max_mr_size;
79 	attr->page_size_cap = 0xffff000;
80 	attr->vendor_id = dev->nic_info.pdev->vendor;
81 	attr->vendor_part_id = dev->nic_info.pdev->device;
82 	attr->hw_ver = dev->asic_id;
83 	attr->max_qp = dev->attr.max_qp;
84 	attr->max_ah = OCRDMA_MAX_AH;
85 	attr->max_qp_wr = dev->attr.max_wqe;
86 
87 	attr->device_cap_flags = IB_DEVICE_CURR_QP_STATE_MOD |
88 					IB_DEVICE_RC_RNR_NAK_GEN |
89 					IB_DEVICE_SHUTDOWN_PORT |
90 					IB_DEVICE_SYS_IMAGE_GUID |
91 					IB_DEVICE_LOCAL_DMA_LKEY |
92 					IB_DEVICE_MEM_MGT_EXTENSIONS;
93 	attr->max_send_sge = dev->attr.max_send_sge;
94 	attr->max_recv_sge = dev->attr.max_recv_sge;
95 	attr->max_sge_rd = dev->attr.max_rdma_sge;
96 	attr->max_cq = dev->attr.max_cq;
97 	attr->max_cqe = dev->attr.max_cqe;
98 	attr->max_mr = dev->attr.max_mr;
99 	attr->max_mw = dev->attr.max_mw;
100 	attr->max_pd = dev->attr.max_pd;
101 	attr->atomic_cap = 0;
102 	attr->max_qp_rd_atom =
103 	    min(dev->attr.max_ord_per_qp, dev->attr.max_ird_per_qp);
104 	attr->max_qp_init_rd_atom = dev->attr.max_ord_per_qp;
105 	attr->max_srq = dev->attr.max_srq;
106 	attr->max_srq_sge = dev->attr.max_srq_sge;
107 	attr->max_srq_wr = dev->attr.max_rqe;
108 	attr->local_ca_ack_delay = dev->attr.local_ca_ack_delay;
109 	attr->max_fast_reg_page_list_len = dev->attr.max_pages_per_frmr;
110 	attr->max_pkeys = 1;
111 	return 0;
112 }
113 
114 static inline void get_link_speed_and_width(struct ocrdma_dev *dev,
115 					    u16 *ib_speed, u8 *ib_width)
116 {
117 	int status;
118 	u8 speed;
119 
120 	status = ocrdma_mbx_get_link_speed(dev, &speed, NULL);
121 	if (status)
122 		speed = OCRDMA_PHYS_LINK_SPEED_ZERO;
123 
124 	switch (speed) {
125 	case OCRDMA_PHYS_LINK_SPEED_1GBPS:
126 		*ib_speed = IB_SPEED_SDR;
127 		*ib_width = IB_WIDTH_1X;
128 		break;
129 
130 	case OCRDMA_PHYS_LINK_SPEED_10GBPS:
131 		*ib_speed = IB_SPEED_QDR;
132 		*ib_width = IB_WIDTH_1X;
133 		break;
134 
135 	case OCRDMA_PHYS_LINK_SPEED_20GBPS:
136 		*ib_speed = IB_SPEED_DDR;
137 		*ib_width = IB_WIDTH_4X;
138 		break;
139 
140 	case OCRDMA_PHYS_LINK_SPEED_40GBPS:
141 		*ib_speed = IB_SPEED_QDR;
142 		*ib_width = IB_WIDTH_4X;
143 		break;
144 
145 	default:
146 		/* Unsupported */
147 		*ib_speed = IB_SPEED_SDR;
148 		*ib_width = IB_WIDTH_1X;
149 	}
150 }
151 
152 int ocrdma_query_port(struct ib_device *ibdev,
153 		      u8 port, struct ib_port_attr *props)
154 {
155 	enum ib_port_state port_state;
156 	struct ocrdma_dev *dev;
157 	struct net_device *netdev;
158 
159 	/* props being zeroed by the caller, avoid zeroing it here */
160 	dev = get_ocrdma_dev(ibdev);
161 	netdev = dev->nic_info.netdev;
162 	if (netif_running(netdev) && netif_oper_up(netdev)) {
163 		port_state = IB_PORT_ACTIVE;
164 		props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
165 	} else {
166 		port_state = IB_PORT_DOWN;
167 		props->phys_state = IB_PORT_PHYS_STATE_DISABLED;
168 	}
169 	props->max_mtu = IB_MTU_4096;
170 	props->active_mtu = iboe_get_mtu(netdev->mtu);
171 	props->lid = 0;
172 	props->lmc = 0;
173 	props->sm_lid = 0;
174 	props->sm_sl = 0;
175 	props->state = port_state;
176 	props->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_REINIT_SUP |
177 				IB_PORT_DEVICE_MGMT_SUP |
178 				IB_PORT_VENDOR_CLASS_SUP;
179 	props->ip_gids = true;
180 	props->gid_tbl_len = OCRDMA_MAX_SGID;
181 	props->pkey_tbl_len = 1;
182 	props->bad_pkey_cntr = 0;
183 	props->qkey_viol_cntr = 0;
184 	get_link_speed_and_width(dev, &props->active_speed,
185 				 &props->active_width);
186 	props->max_msg_sz = 0x80000000;
187 	props->max_vl_num = 4;
188 	return 0;
189 }
190 
191 static int ocrdma_add_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
192 			   unsigned long len)
193 {
194 	struct ocrdma_mm *mm;
195 
196 	mm = kzalloc(sizeof(*mm), GFP_KERNEL);
197 	if (mm == NULL)
198 		return -ENOMEM;
199 	mm->key.phy_addr = phy_addr;
200 	mm->key.len = len;
201 	INIT_LIST_HEAD(&mm->entry);
202 
203 	mutex_lock(&uctx->mm_list_lock);
204 	list_add_tail(&mm->entry, &uctx->mm_head);
205 	mutex_unlock(&uctx->mm_list_lock);
206 	return 0;
207 }
208 
209 static void ocrdma_del_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
210 			    unsigned long len)
211 {
212 	struct ocrdma_mm *mm, *tmp;
213 
214 	mutex_lock(&uctx->mm_list_lock);
215 	list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
216 		if (len != mm->key.len && phy_addr != mm->key.phy_addr)
217 			continue;
218 
219 		list_del(&mm->entry);
220 		kfree(mm);
221 		break;
222 	}
223 	mutex_unlock(&uctx->mm_list_lock);
224 }
225 
226 static bool ocrdma_search_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
227 			      unsigned long len)
228 {
229 	bool found = false;
230 	struct ocrdma_mm *mm;
231 
232 	mutex_lock(&uctx->mm_list_lock);
233 	list_for_each_entry(mm, &uctx->mm_head, entry) {
234 		if (len != mm->key.len && phy_addr != mm->key.phy_addr)
235 			continue;
236 
237 		found = true;
238 		break;
239 	}
240 	mutex_unlock(&uctx->mm_list_lock);
241 	return found;
242 }
243 
244 
245 static u16 _ocrdma_pd_mgr_get_bitmap(struct ocrdma_dev *dev, bool dpp_pool)
246 {
247 	u16 pd_bitmap_idx = 0;
248 	const unsigned long *pd_bitmap;
249 
250 	if (dpp_pool) {
251 		pd_bitmap = dev->pd_mgr->pd_dpp_bitmap;
252 		pd_bitmap_idx = find_first_zero_bit(pd_bitmap,
253 						    dev->pd_mgr->max_dpp_pd);
254 		__set_bit(pd_bitmap_idx, dev->pd_mgr->pd_dpp_bitmap);
255 		dev->pd_mgr->pd_dpp_count++;
256 		if (dev->pd_mgr->pd_dpp_count > dev->pd_mgr->pd_dpp_thrsh)
257 			dev->pd_mgr->pd_dpp_thrsh = dev->pd_mgr->pd_dpp_count;
258 	} else {
259 		pd_bitmap = dev->pd_mgr->pd_norm_bitmap;
260 		pd_bitmap_idx = find_first_zero_bit(pd_bitmap,
261 						    dev->pd_mgr->max_normal_pd);
262 		__set_bit(pd_bitmap_idx, dev->pd_mgr->pd_norm_bitmap);
263 		dev->pd_mgr->pd_norm_count++;
264 		if (dev->pd_mgr->pd_norm_count > dev->pd_mgr->pd_norm_thrsh)
265 			dev->pd_mgr->pd_norm_thrsh = dev->pd_mgr->pd_norm_count;
266 	}
267 	return pd_bitmap_idx;
268 }
269 
270 static int _ocrdma_pd_mgr_put_bitmap(struct ocrdma_dev *dev, u16 pd_id,
271 					bool dpp_pool)
272 {
273 	u16 pd_count;
274 	u16 pd_bit_index;
275 
276 	pd_count = dpp_pool ? dev->pd_mgr->pd_dpp_count :
277 			      dev->pd_mgr->pd_norm_count;
278 	if (pd_count == 0)
279 		return -EINVAL;
280 
281 	if (dpp_pool) {
282 		pd_bit_index = pd_id - dev->pd_mgr->pd_dpp_start;
283 		if (pd_bit_index >= dev->pd_mgr->max_dpp_pd) {
284 			return -EINVAL;
285 		} else {
286 			__clear_bit(pd_bit_index, dev->pd_mgr->pd_dpp_bitmap);
287 			dev->pd_mgr->pd_dpp_count--;
288 		}
289 	} else {
290 		pd_bit_index = pd_id - dev->pd_mgr->pd_norm_start;
291 		if (pd_bit_index >= dev->pd_mgr->max_normal_pd) {
292 			return -EINVAL;
293 		} else {
294 			__clear_bit(pd_bit_index, dev->pd_mgr->pd_norm_bitmap);
295 			dev->pd_mgr->pd_norm_count--;
296 		}
297 	}
298 
299 	return 0;
300 }
301 
302 static int ocrdma_put_pd_num(struct ocrdma_dev *dev, u16 pd_id,
303 				   bool dpp_pool)
304 {
305 	int status;
306 
307 	mutex_lock(&dev->dev_lock);
308 	status = _ocrdma_pd_mgr_put_bitmap(dev, pd_id, dpp_pool);
309 	mutex_unlock(&dev->dev_lock);
310 	return status;
311 }
312 
313 static int ocrdma_get_pd_num(struct ocrdma_dev *dev, struct ocrdma_pd *pd)
314 {
315 	u16 pd_idx = 0;
316 	int status = 0;
317 
318 	mutex_lock(&dev->dev_lock);
319 	if (pd->dpp_enabled) {
320 		/* try allocating DPP PD, if not available then normal PD */
321 		if (dev->pd_mgr->pd_dpp_count < dev->pd_mgr->max_dpp_pd) {
322 			pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, true);
323 			pd->id = dev->pd_mgr->pd_dpp_start + pd_idx;
324 			pd->dpp_page = dev->pd_mgr->dpp_page_index + pd_idx;
325 		} else if (dev->pd_mgr->pd_norm_count <
326 			   dev->pd_mgr->max_normal_pd) {
327 			pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, false);
328 			pd->id = dev->pd_mgr->pd_norm_start + pd_idx;
329 			pd->dpp_enabled = false;
330 		} else {
331 			status = -EINVAL;
332 		}
333 	} else {
334 		if (dev->pd_mgr->pd_norm_count < dev->pd_mgr->max_normal_pd) {
335 			pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, false);
336 			pd->id = dev->pd_mgr->pd_norm_start + pd_idx;
337 		} else {
338 			status = -EINVAL;
339 		}
340 	}
341 	mutex_unlock(&dev->dev_lock);
342 	return status;
343 }
344 
345 /*
346  * NOTE:
347  *
348  * ocrdma_ucontext must be used here because this function is also
349  * called from ocrdma_alloc_ucontext where ib_udata does not have
350  * valid ib_ucontext pointer. ib_uverbs_get_context does not call
351  * uobj_{alloc|get_xxx} helpers which are used to store the
352  * ib_ucontext in uverbs_attr_bundle wrapping the ib_udata. so
353  * ib_udata does NOT imply valid ib_ucontext here!
354  */
355 static int _ocrdma_alloc_pd(struct ocrdma_dev *dev, struct ocrdma_pd *pd,
356 			    struct ocrdma_ucontext *uctx,
357 			    struct ib_udata *udata)
358 {
359 	int status;
360 
361 	if (udata && uctx && dev->attr.max_dpp_pds) {
362 		pd->dpp_enabled =
363 			ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R;
364 		pd->num_dpp_qp =
365 			pd->dpp_enabled ? (dev->nic_info.db_page_size /
366 					   dev->attr.wqe_size) : 0;
367 	}
368 
369 	if (dev->pd_mgr->pd_prealloc_valid)
370 		return ocrdma_get_pd_num(dev, pd);
371 
372 retry:
373 	status = ocrdma_mbx_alloc_pd(dev, pd);
374 	if (status) {
375 		if (pd->dpp_enabled) {
376 			pd->dpp_enabled = false;
377 			pd->num_dpp_qp = 0;
378 			goto retry;
379 		}
380 		return status;
381 	}
382 
383 	return 0;
384 }
385 
386 static inline int is_ucontext_pd(struct ocrdma_ucontext *uctx,
387 				 struct ocrdma_pd *pd)
388 {
389 	return (uctx->cntxt_pd == pd);
390 }
391 
392 static void _ocrdma_dealloc_pd(struct ocrdma_dev *dev,
393 			      struct ocrdma_pd *pd)
394 {
395 	if (dev->pd_mgr->pd_prealloc_valid)
396 		ocrdma_put_pd_num(dev, pd->id, pd->dpp_enabled);
397 	else
398 		ocrdma_mbx_dealloc_pd(dev, pd);
399 }
400 
401 static int ocrdma_alloc_ucontext_pd(struct ocrdma_dev *dev,
402 				    struct ocrdma_ucontext *uctx,
403 				    struct ib_udata *udata)
404 {
405 	struct ib_device *ibdev = &dev->ibdev;
406 	struct ib_pd *pd;
407 	int status;
408 
409 	pd = rdma_zalloc_drv_obj(ibdev, ib_pd);
410 	if (!pd)
411 		return -ENOMEM;
412 
413 	pd->device  = ibdev;
414 	uctx->cntxt_pd = get_ocrdma_pd(pd);
415 
416 	status = _ocrdma_alloc_pd(dev, uctx->cntxt_pd, uctx, udata);
417 	if (status) {
418 		kfree(uctx->cntxt_pd);
419 		goto err;
420 	}
421 
422 	uctx->cntxt_pd->uctx = uctx;
423 	uctx->cntxt_pd->ibpd.device = &dev->ibdev;
424 err:
425 	return status;
426 }
427 
428 static void ocrdma_dealloc_ucontext_pd(struct ocrdma_ucontext *uctx)
429 {
430 	struct ocrdma_pd *pd = uctx->cntxt_pd;
431 	struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
432 
433 	if (uctx->pd_in_use) {
434 		pr_err("%s(%d) Freeing in use pdid=0x%x.\n",
435 		       __func__, dev->id, pd->id);
436 	}
437 	kfree(uctx->cntxt_pd);
438 	uctx->cntxt_pd = NULL;
439 	_ocrdma_dealloc_pd(dev, pd);
440 }
441 
442 static struct ocrdma_pd *ocrdma_get_ucontext_pd(struct ocrdma_ucontext *uctx)
443 {
444 	struct ocrdma_pd *pd = NULL;
445 
446 	mutex_lock(&uctx->mm_list_lock);
447 	if (!uctx->pd_in_use) {
448 		uctx->pd_in_use = true;
449 		pd = uctx->cntxt_pd;
450 	}
451 	mutex_unlock(&uctx->mm_list_lock);
452 
453 	return pd;
454 }
455 
456 static void ocrdma_release_ucontext_pd(struct ocrdma_ucontext *uctx)
457 {
458 	mutex_lock(&uctx->mm_list_lock);
459 	uctx->pd_in_use = false;
460 	mutex_unlock(&uctx->mm_list_lock);
461 }
462 
463 int ocrdma_alloc_ucontext(struct ib_ucontext *uctx, struct ib_udata *udata)
464 {
465 	struct ib_device *ibdev = uctx->device;
466 	int status;
467 	struct ocrdma_ucontext *ctx = get_ocrdma_ucontext(uctx);
468 	struct ocrdma_alloc_ucontext_resp resp = {};
469 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
470 	struct pci_dev *pdev = dev->nic_info.pdev;
471 	u32 map_len = roundup(sizeof(u32) * 2048, PAGE_SIZE);
472 
473 	if (!udata)
474 		return -EFAULT;
475 	INIT_LIST_HEAD(&ctx->mm_head);
476 	mutex_init(&ctx->mm_list_lock);
477 
478 	ctx->ah_tbl.va = dma_alloc_coherent(&pdev->dev, map_len,
479 					    &ctx->ah_tbl.pa, GFP_KERNEL);
480 	if (!ctx->ah_tbl.va)
481 		return -ENOMEM;
482 
483 	ctx->ah_tbl.len = map_len;
484 
485 	resp.ah_tbl_len = ctx->ah_tbl.len;
486 	resp.ah_tbl_page = virt_to_phys(ctx->ah_tbl.va);
487 
488 	status = ocrdma_add_mmap(ctx, resp.ah_tbl_page, resp.ah_tbl_len);
489 	if (status)
490 		goto map_err;
491 
492 	status = ocrdma_alloc_ucontext_pd(dev, ctx, udata);
493 	if (status)
494 		goto pd_err;
495 
496 	resp.dev_id = dev->id;
497 	resp.max_inline_data = dev->attr.max_inline_data;
498 	resp.wqe_size = dev->attr.wqe_size;
499 	resp.rqe_size = dev->attr.rqe_size;
500 	resp.dpp_wqe_size = dev->attr.wqe_size;
501 
502 	memcpy(resp.fw_ver, dev->attr.fw_ver, sizeof(resp.fw_ver));
503 	status = ib_copy_to_udata(udata, &resp, sizeof(resp));
504 	if (status)
505 		goto cpy_err;
506 	return 0;
507 
508 cpy_err:
509 	ocrdma_dealloc_ucontext_pd(ctx);
510 pd_err:
511 	ocrdma_del_mmap(ctx, ctx->ah_tbl.pa, ctx->ah_tbl.len);
512 map_err:
513 	dma_free_coherent(&pdev->dev, ctx->ah_tbl.len, ctx->ah_tbl.va,
514 			  ctx->ah_tbl.pa);
515 	return status;
516 }
517 
518 void ocrdma_dealloc_ucontext(struct ib_ucontext *ibctx)
519 {
520 	struct ocrdma_mm *mm, *tmp;
521 	struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ibctx);
522 	struct ocrdma_dev *dev = get_ocrdma_dev(ibctx->device);
523 	struct pci_dev *pdev = dev->nic_info.pdev;
524 
525 	ocrdma_dealloc_ucontext_pd(uctx);
526 
527 	ocrdma_del_mmap(uctx, uctx->ah_tbl.pa, uctx->ah_tbl.len);
528 	dma_free_coherent(&pdev->dev, uctx->ah_tbl.len, uctx->ah_tbl.va,
529 			  uctx->ah_tbl.pa);
530 
531 	list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
532 		list_del(&mm->entry);
533 		kfree(mm);
534 	}
535 }
536 
537 int ocrdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
538 {
539 	struct ocrdma_ucontext *ucontext = get_ocrdma_ucontext(context);
540 	struct ocrdma_dev *dev = get_ocrdma_dev(context->device);
541 	unsigned long vm_page = vma->vm_pgoff << PAGE_SHIFT;
542 	u64 unmapped_db = (u64) dev->nic_info.unmapped_db;
543 	unsigned long len = (vma->vm_end - vma->vm_start);
544 	int status;
545 	bool found;
546 
547 	if (vma->vm_start & (PAGE_SIZE - 1))
548 		return -EINVAL;
549 	found = ocrdma_search_mmap(ucontext, vma->vm_pgoff << PAGE_SHIFT, len);
550 	if (!found)
551 		return -EINVAL;
552 
553 	if ((vm_page >= unmapped_db) && (vm_page <= (unmapped_db +
554 		dev->nic_info.db_total_size)) &&
555 		(len <=	dev->nic_info.db_page_size)) {
556 		if (vma->vm_flags & VM_READ)
557 			return -EPERM;
558 
559 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
560 		status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
561 					    len, vma->vm_page_prot);
562 	} else if (dev->nic_info.dpp_unmapped_len &&
563 		(vm_page >= (u64) dev->nic_info.dpp_unmapped_addr) &&
564 		(vm_page <= (u64) (dev->nic_info.dpp_unmapped_addr +
565 			dev->nic_info.dpp_unmapped_len)) &&
566 		(len <= dev->nic_info.dpp_unmapped_len)) {
567 		if (vma->vm_flags & VM_READ)
568 			return -EPERM;
569 
570 		vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
571 		status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
572 					    len, vma->vm_page_prot);
573 	} else {
574 		status = remap_pfn_range(vma, vma->vm_start,
575 					 vma->vm_pgoff, len, vma->vm_page_prot);
576 	}
577 	return status;
578 }
579 
580 static int ocrdma_copy_pd_uresp(struct ocrdma_dev *dev, struct ocrdma_pd *pd,
581 				struct ib_udata *udata)
582 {
583 	int status;
584 	u64 db_page_addr;
585 	u64 dpp_page_addr = 0;
586 	u32 db_page_size;
587 	struct ocrdma_alloc_pd_uresp rsp;
588 	struct ocrdma_ucontext *uctx = rdma_udata_to_drv_context(
589 		udata, struct ocrdma_ucontext, ibucontext);
590 
591 	memset(&rsp, 0, sizeof(rsp));
592 	rsp.id = pd->id;
593 	rsp.dpp_enabled = pd->dpp_enabled;
594 	db_page_addr = ocrdma_get_db_addr(dev, pd->id);
595 	db_page_size = dev->nic_info.db_page_size;
596 
597 	status = ocrdma_add_mmap(uctx, db_page_addr, db_page_size);
598 	if (status)
599 		return status;
600 
601 	if (pd->dpp_enabled) {
602 		dpp_page_addr = dev->nic_info.dpp_unmapped_addr +
603 				(pd->id * PAGE_SIZE);
604 		status = ocrdma_add_mmap(uctx, dpp_page_addr,
605 				 PAGE_SIZE);
606 		if (status)
607 			goto dpp_map_err;
608 		rsp.dpp_page_addr_hi = upper_32_bits(dpp_page_addr);
609 		rsp.dpp_page_addr_lo = dpp_page_addr;
610 	}
611 
612 	status = ib_copy_to_udata(udata, &rsp, sizeof(rsp));
613 	if (status)
614 		goto ucopy_err;
615 
616 	pd->uctx = uctx;
617 	return 0;
618 
619 ucopy_err:
620 	if (pd->dpp_enabled)
621 		ocrdma_del_mmap(pd->uctx, dpp_page_addr, PAGE_SIZE);
622 dpp_map_err:
623 	ocrdma_del_mmap(pd->uctx, db_page_addr, db_page_size);
624 	return status;
625 }
626 
627 int ocrdma_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
628 {
629 	struct ib_device *ibdev = ibpd->device;
630 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
631 	struct ocrdma_pd *pd;
632 	int status;
633 	u8 is_uctx_pd = false;
634 	struct ocrdma_ucontext *uctx = rdma_udata_to_drv_context(
635 		udata, struct ocrdma_ucontext, ibucontext);
636 
637 	if (udata) {
638 		pd = ocrdma_get_ucontext_pd(uctx);
639 		if (pd) {
640 			is_uctx_pd = true;
641 			goto pd_mapping;
642 		}
643 	}
644 
645 	pd = get_ocrdma_pd(ibpd);
646 	status = _ocrdma_alloc_pd(dev, pd, uctx, udata);
647 	if (status)
648 		goto exit;
649 
650 pd_mapping:
651 	if (udata) {
652 		status = ocrdma_copy_pd_uresp(dev, pd, udata);
653 		if (status)
654 			goto err;
655 	}
656 	return 0;
657 
658 err:
659 	if (is_uctx_pd)
660 		ocrdma_release_ucontext_pd(uctx);
661 	else
662 		_ocrdma_dealloc_pd(dev, pd);
663 exit:
664 	return status;
665 }
666 
667 int ocrdma_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
668 {
669 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
670 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
671 	struct ocrdma_ucontext *uctx = NULL;
672 	u64 usr_db;
673 
674 	uctx = pd->uctx;
675 	if (uctx) {
676 		u64 dpp_db = dev->nic_info.dpp_unmapped_addr +
677 			(pd->id * PAGE_SIZE);
678 		if (pd->dpp_enabled)
679 			ocrdma_del_mmap(pd->uctx, dpp_db, PAGE_SIZE);
680 		usr_db = ocrdma_get_db_addr(dev, pd->id);
681 		ocrdma_del_mmap(pd->uctx, usr_db, dev->nic_info.db_page_size);
682 
683 		if (is_ucontext_pd(uctx, pd)) {
684 			ocrdma_release_ucontext_pd(uctx);
685 			return 0;
686 		}
687 	}
688 	_ocrdma_dealloc_pd(dev, pd);
689 	return 0;
690 }
691 
692 static int ocrdma_alloc_lkey(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
693 			    u32 pdid, int acc, u32 num_pbls, u32 addr_check)
694 {
695 	int status;
696 
697 	mr->hwmr.fr_mr = 0;
698 	mr->hwmr.local_rd = 1;
699 	mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
700 	mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
701 	mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
702 	mr->hwmr.mw_bind = (acc & IB_ACCESS_MW_BIND) ? 1 : 0;
703 	mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
704 	mr->hwmr.num_pbls = num_pbls;
705 
706 	status = ocrdma_mbx_alloc_lkey(dev, &mr->hwmr, pdid, addr_check);
707 	if (status)
708 		return status;
709 
710 	mr->ibmr.lkey = mr->hwmr.lkey;
711 	if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
712 		mr->ibmr.rkey = mr->hwmr.lkey;
713 	return 0;
714 }
715 
716 struct ib_mr *ocrdma_get_dma_mr(struct ib_pd *ibpd, int acc)
717 {
718 	int status;
719 	struct ocrdma_mr *mr;
720 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
721 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
722 
723 	if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE)) {
724 		pr_err("%s err, invalid access rights\n", __func__);
725 		return ERR_PTR(-EINVAL);
726 	}
727 
728 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
729 	if (!mr)
730 		return ERR_PTR(-ENOMEM);
731 
732 	status = ocrdma_alloc_lkey(dev, mr, pd->id, acc, 0,
733 				   OCRDMA_ADDR_CHECK_DISABLE);
734 	if (status) {
735 		kfree(mr);
736 		return ERR_PTR(status);
737 	}
738 
739 	return &mr->ibmr;
740 }
741 
742 static void ocrdma_free_mr_pbl_tbl(struct ocrdma_dev *dev,
743 				   struct ocrdma_hw_mr *mr)
744 {
745 	struct pci_dev *pdev = dev->nic_info.pdev;
746 	int i = 0;
747 
748 	if (mr->pbl_table) {
749 		for (i = 0; i < mr->num_pbls; i++) {
750 			if (!mr->pbl_table[i].va)
751 				continue;
752 			dma_free_coherent(&pdev->dev, mr->pbl_size,
753 					  mr->pbl_table[i].va,
754 					  mr->pbl_table[i].pa);
755 		}
756 		kfree(mr->pbl_table);
757 		mr->pbl_table = NULL;
758 	}
759 }
760 
761 static int ocrdma_get_pbl_info(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
762 			      u32 num_pbes)
763 {
764 	u32 num_pbls = 0;
765 	u32 idx = 0;
766 	int status = 0;
767 	u32 pbl_size;
768 
769 	do {
770 		pbl_size = OCRDMA_MIN_HPAGE_SIZE * (1 << idx);
771 		if (pbl_size > MAX_OCRDMA_PBL_SIZE) {
772 			status = -EFAULT;
773 			break;
774 		}
775 		num_pbls = roundup(num_pbes, (pbl_size / sizeof(u64)));
776 		num_pbls = num_pbls / (pbl_size / sizeof(u64));
777 		idx++;
778 	} while (num_pbls >= dev->attr.max_num_mr_pbl);
779 
780 	mr->hwmr.num_pbes = num_pbes;
781 	mr->hwmr.num_pbls = num_pbls;
782 	mr->hwmr.pbl_size = pbl_size;
783 	return status;
784 }
785 
786 static int ocrdma_build_pbl_tbl(struct ocrdma_dev *dev, struct ocrdma_hw_mr *mr)
787 {
788 	int status = 0;
789 	int i;
790 	u32 dma_len = mr->pbl_size;
791 	struct pci_dev *pdev = dev->nic_info.pdev;
792 	void *va;
793 	dma_addr_t pa;
794 
795 	mr->pbl_table = kcalloc(mr->num_pbls, sizeof(struct ocrdma_pbl),
796 				GFP_KERNEL);
797 
798 	if (!mr->pbl_table)
799 		return -ENOMEM;
800 
801 	for (i = 0; i < mr->num_pbls; i++) {
802 		va = dma_alloc_coherent(&pdev->dev, dma_len, &pa, GFP_KERNEL);
803 		if (!va) {
804 			ocrdma_free_mr_pbl_tbl(dev, mr);
805 			status = -ENOMEM;
806 			break;
807 		}
808 		mr->pbl_table[i].va = va;
809 		mr->pbl_table[i].pa = pa;
810 	}
811 	return status;
812 }
813 
814 static void build_user_pbes(struct ocrdma_dev *dev, struct ocrdma_mr *mr)
815 {
816 	struct ocrdma_pbe *pbe;
817 	struct ib_block_iter biter;
818 	struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
819 	int pbe_cnt;
820 	u64 pg_addr;
821 
822 	if (!mr->hwmr.num_pbes)
823 		return;
824 
825 	pbe = (struct ocrdma_pbe *)pbl_tbl->va;
826 	pbe_cnt = 0;
827 
828 	rdma_umem_for_each_dma_block (mr->umem, &biter, PAGE_SIZE) {
829 		/* store the page address in pbe */
830 		pg_addr = rdma_block_iter_dma_address(&biter);
831 		pbe->pa_lo = cpu_to_le32(pg_addr);
832 		pbe->pa_hi = cpu_to_le32(upper_32_bits(pg_addr));
833 		pbe_cnt += 1;
834 		pbe++;
835 
836 		/* if the given pbl is full storing the pbes,
837 		 * move to next pbl.
838 		 */
839 		if (pbe_cnt == (mr->hwmr.pbl_size / sizeof(u64))) {
840 			pbl_tbl++;
841 			pbe = (struct ocrdma_pbe *)pbl_tbl->va;
842 			pbe_cnt = 0;
843 		}
844 	}
845 }
846 
847 struct ib_mr *ocrdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
848 				 u64 usr_addr, int acc, struct ib_udata *udata)
849 {
850 	int status = -ENOMEM;
851 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
852 	struct ocrdma_mr *mr;
853 	struct ocrdma_pd *pd;
854 
855 	pd = get_ocrdma_pd(ibpd);
856 
857 	if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE))
858 		return ERR_PTR(-EINVAL);
859 
860 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
861 	if (!mr)
862 		return ERR_PTR(status);
863 	mr->umem = ib_umem_get(ibpd->device, start, len, acc);
864 	if (IS_ERR(mr->umem)) {
865 		status = -EFAULT;
866 		goto umem_err;
867 	}
868 	status = ocrdma_get_pbl_info(
869 		dev, mr, ib_umem_num_dma_blocks(mr->umem, PAGE_SIZE));
870 	if (status)
871 		goto umem_err;
872 
873 	mr->hwmr.pbe_size = PAGE_SIZE;
874 	mr->hwmr.va = usr_addr;
875 	mr->hwmr.len = len;
876 	mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
877 	mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
878 	mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
879 	mr->hwmr.local_rd = 1;
880 	mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
881 	status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
882 	if (status)
883 		goto umem_err;
884 	build_user_pbes(dev, mr);
885 	status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
886 	if (status)
887 		goto mbx_err;
888 	mr->ibmr.lkey = mr->hwmr.lkey;
889 	if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
890 		mr->ibmr.rkey = mr->hwmr.lkey;
891 
892 	return &mr->ibmr;
893 
894 mbx_err:
895 	ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
896 umem_err:
897 	kfree(mr);
898 	return ERR_PTR(status);
899 }
900 
901 int ocrdma_dereg_mr(struct ib_mr *ib_mr, struct ib_udata *udata)
902 {
903 	struct ocrdma_mr *mr = get_ocrdma_mr(ib_mr);
904 	struct ocrdma_dev *dev = get_ocrdma_dev(ib_mr->device);
905 
906 	(void) ocrdma_mbx_dealloc_lkey(dev, mr->hwmr.fr_mr, mr->hwmr.lkey);
907 
908 	kfree(mr->pages);
909 	ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
910 
911 	/* it could be user registered memory. */
912 	ib_umem_release(mr->umem);
913 	kfree(mr);
914 
915 	/* Don't stop cleanup, in case FW is unresponsive */
916 	if (dev->mqe_ctx.fw_error_state) {
917 		pr_err("%s(%d) fw not responding.\n",
918 		       __func__, dev->id);
919 	}
920 	return 0;
921 }
922 
923 static int ocrdma_copy_cq_uresp(struct ocrdma_dev *dev, struct ocrdma_cq *cq,
924 				struct ib_udata *udata)
925 {
926 	int status;
927 	struct ocrdma_ucontext *uctx = rdma_udata_to_drv_context(
928 		udata, struct ocrdma_ucontext, ibucontext);
929 	struct ocrdma_create_cq_uresp uresp;
930 
931 	/* this must be user flow! */
932 	if (!udata)
933 		return -EINVAL;
934 
935 	memset(&uresp, 0, sizeof(uresp));
936 	uresp.cq_id = cq->id;
937 	uresp.page_size = PAGE_ALIGN(cq->len);
938 	uresp.num_pages = 1;
939 	uresp.max_hw_cqe = cq->max_hw_cqe;
940 	uresp.page_addr[0] = virt_to_phys(cq->va);
941 	uresp.db_page_addr =  ocrdma_get_db_addr(dev, uctx->cntxt_pd->id);
942 	uresp.db_page_size = dev->nic_info.db_page_size;
943 	uresp.phase_change = cq->phase_change ? 1 : 0;
944 	status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
945 	if (status) {
946 		pr_err("%s(%d) copy error cqid=0x%x.\n",
947 		       __func__, dev->id, cq->id);
948 		goto err;
949 	}
950 	status = ocrdma_add_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
951 	if (status)
952 		goto err;
953 	status = ocrdma_add_mmap(uctx, uresp.page_addr[0], uresp.page_size);
954 	if (status) {
955 		ocrdma_del_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
956 		goto err;
957 	}
958 	cq->ucontext = uctx;
959 err:
960 	return status;
961 }
962 
963 int ocrdma_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
964 		     struct ib_udata *udata)
965 {
966 	struct ib_device *ibdev = ibcq->device;
967 	int entries = attr->cqe;
968 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
969 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
970 	struct ocrdma_ucontext *uctx = rdma_udata_to_drv_context(
971 		udata, struct ocrdma_ucontext, ibucontext);
972 	u16 pd_id = 0;
973 	int status;
974 	struct ocrdma_create_cq_ureq ureq;
975 
976 	if (attr->flags)
977 		return -EINVAL;
978 
979 	if (udata) {
980 		if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
981 			return -EFAULT;
982 	} else
983 		ureq.dpp_cq = 0;
984 
985 	spin_lock_init(&cq->cq_lock);
986 	spin_lock_init(&cq->comp_handler_lock);
987 	INIT_LIST_HEAD(&cq->sq_head);
988 	INIT_LIST_HEAD(&cq->rq_head);
989 
990 	if (udata)
991 		pd_id = uctx->cntxt_pd->id;
992 
993 	status = ocrdma_mbx_create_cq(dev, cq, entries, ureq.dpp_cq, pd_id);
994 	if (status)
995 		return status;
996 
997 	if (udata) {
998 		status = ocrdma_copy_cq_uresp(dev, cq, udata);
999 		if (status)
1000 			goto ctx_err;
1001 	}
1002 	cq->phase = OCRDMA_CQE_VALID;
1003 	dev->cq_tbl[cq->id] = cq;
1004 	return 0;
1005 
1006 ctx_err:
1007 	ocrdma_mbx_destroy_cq(dev, cq);
1008 	return status;
1009 }
1010 
1011 int ocrdma_resize_cq(struct ib_cq *ibcq, int new_cnt,
1012 		     struct ib_udata *udata)
1013 {
1014 	int status = 0;
1015 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
1016 
1017 	if (new_cnt < 1 || new_cnt > cq->max_hw_cqe) {
1018 		status = -EINVAL;
1019 		return status;
1020 	}
1021 	ibcq->cqe = new_cnt;
1022 	return status;
1023 }
1024 
1025 static void ocrdma_flush_cq(struct ocrdma_cq *cq)
1026 {
1027 	int cqe_cnt;
1028 	int valid_count = 0;
1029 	unsigned long flags;
1030 
1031 	struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
1032 	struct ocrdma_cqe *cqe = NULL;
1033 
1034 	cqe = cq->va;
1035 	cqe_cnt = cq->cqe_cnt;
1036 
1037 	/* Last irq might have scheduled a polling thread
1038 	 * sync-up with it before hard flushing.
1039 	 */
1040 	spin_lock_irqsave(&cq->cq_lock, flags);
1041 	while (cqe_cnt) {
1042 		if (is_cqe_valid(cq, cqe))
1043 			valid_count++;
1044 		cqe++;
1045 		cqe_cnt--;
1046 	}
1047 	ocrdma_ring_cq_db(dev, cq->id, false, false, valid_count);
1048 	spin_unlock_irqrestore(&cq->cq_lock, flags);
1049 }
1050 
1051 int ocrdma_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
1052 {
1053 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
1054 	struct ocrdma_eq *eq = NULL;
1055 	struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
1056 	int pdid = 0;
1057 	u32 irq, indx;
1058 
1059 	dev->cq_tbl[cq->id] = NULL;
1060 	indx = ocrdma_get_eq_table_index(dev, cq->eqn);
1061 
1062 	eq = &dev->eq_tbl[indx];
1063 	irq = ocrdma_get_irq(dev, eq);
1064 	synchronize_irq(irq);
1065 	ocrdma_flush_cq(cq);
1066 
1067 	ocrdma_mbx_destroy_cq(dev, cq);
1068 	if (cq->ucontext) {
1069 		pdid = cq->ucontext->cntxt_pd->id;
1070 		ocrdma_del_mmap(cq->ucontext, (u64) cq->pa,
1071 				PAGE_ALIGN(cq->len));
1072 		ocrdma_del_mmap(cq->ucontext,
1073 				ocrdma_get_db_addr(dev, pdid),
1074 				dev->nic_info.db_page_size);
1075 	}
1076 	return 0;
1077 }
1078 
1079 static int ocrdma_add_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1080 {
1081 	int status = -EINVAL;
1082 
1083 	if (qp->id < OCRDMA_MAX_QP && dev->qp_tbl[qp->id] == NULL) {
1084 		dev->qp_tbl[qp->id] = qp;
1085 		status = 0;
1086 	}
1087 	return status;
1088 }
1089 
1090 static void ocrdma_del_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1091 {
1092 	dev->qp_tbl[qp->id] = NULL;
1093 }
1094 
1095 static int ocrdma_check_qp_params(struct ib_pd *ibpd, struct ocrdma_dev *dev,
1096 				  struct ib_qp_init_attr *attrs,
1097 				  struct ib_udata *udata)
1098 {
1099 	if ((attrs->qp_type != IB_QPT_GSI) &&
1100 	    (attrs->qp_type != IB_QPT_RC) &&
1101 	    (attrs->qp_type != IB_QPT_UC) &&
1102 	    (attrs->qp_type != IB_QPT_UD)) {
1103 		pr_err("%s(%d) unsupported qp type=0x%x requested\n",
1104 		       __func__, dev->id, attrs->qp_type);
1105 		return -EOPNOTSUPP;
1106 	}
1107 	/* Skip the check for QP1 to support CM size of 128 */
1108 	if ((attrs->qp_type != IB_QPT_GSI) &&
1109 	    (attrs->cap.max_send_wr > dev->attr.max_wqe)) {
1110 		pr_err("%s(%d) unsupported send_wr=0x%x requested\n",
1111 		       __func__, dev->id, attrs->cap.max_send_wr);
1112 		pr_err("%s(%d) supported send_wr=0x%x\n",
1113 		       __func__, dev->id, dev->attr.max_wqe);
1114 		return -EINVAL;
1115 	}
1116 	if (!attrs->srq && (attrs->cap.max_recv_wr > dev->attr.max_rqe)) {
1117 		pr_err("%s(%d) unsupported recv_wr=0x%x requested\n",
1118 		       __func__, dev->id, attrs->cap.max_recv_wr);
1119 		pr_err("%s(%d) supported recv_wr=0x%x\n",
1120 		       __func__, dev->id, dev->attr.max_rqe);
1121 		return -EINVAL;
1122 	}
1123 	if (attrs->cap.max_inline_data > dev->attr.max_inline_data) {
1124 		pr_err("%s(%d) unsupported inline data size=0x%x requested\n",
1125 		       __func__, dev->id, attrs->cap.max_inline_data);
1126 		pr_err("%s(%d) supported inline data size=0x%x\n",
1127 		       __func__, dev->id, dev->attr.max_inline_data);
1128 		return -EINVAL;
1129 	}
1130 	if (attrs->cap.max_send_sge > dev->attr.max_send_sge) {
1131 		pr_err("%s(%d) unsupported send_sge=0x%x requested\n",
1132 		       __func__, dev->id, attrs->cap.max_send_sge);
1133 		pr_err("%s(%d) supported send_sge=0x%x\n",
1134 		       __func__, dev->id, dev->attr.max_send_sge);
1135 		return -EINVAL;
1136 	}
1137 	if (attrs->cap.max_recv_sge > dev->attr.max_recv_sge) {
1138 		pr_err("%s(%d) unsupported recv_sge=0x%x requested\n",
1139 		       __func__, dev->id, attrs->cap.max_recv_sge);
1140 		pr_err("%s(%d) supported recv_sge=0x%x\n",
1141 		       __func__, dev->id, dev->attr.max_recv_sge);
1142 		return -EINVAL;
1143 	}
1144 	/* unprivileged user space cannot create special QP */
1145 	if (udata && attrs->qp_type == IB_QPT_GSI) {
1146 		pr_err
1147 		    ("%s(%d) Userspace can't create special QPs of type=0x%x\n",
1148 		     __func__, dev->id, attrs->qp_type);
1149 		return -EINVAL;
1150 	}
1151 	/* allow creating only one GSI type of QP */
1152 	if (attrs->qp_type == IB_QPT_GSI && dev->gsi_qp_created) {
1153 		pr_err("%s(%d) GSI special QPs already created.\n",
1154 		       __func__, dev->id);
1155 		return -EINVAL;
1156 	}
1157 	/* verify consumer QPs are not trying to use GSI QP's CQ */
1158 	if ((attrs->qp_type != IB_QPT_GSI) && (dev->gsi_qp_created)) {
1159 		if ((dev->gsi_sqcq == get_ocrdma_cq(attrs->send_cq)) ||
1160 			(dev->gsi_rqcq == get_ocrdma_cq(attrs->recv_cq))) {
1161 			pr_err("%s(%d) Consumer QP cannot use GSI CQs.\n",
1162 				__func__, dev->id);
1163 			return -EINVAL;
1164 		}
1165 	}
1166 	return 0;
1167 }
1168 
1169 static int ocrdma_copy_qp_uresp(struct ocrdma_qp *qp,
1170 				struct ib_udata *udata, int dpp_offset,
1171 				int dpp_credit_lmt, int srq)
1172 {
1173 	int status;
1174 	u64 usr_db;
1175 	struct ocrdma_create_qp_uresp uresp;
1176 	struct ocrdma_pd *pd = qp->pd;
1177 	struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
1178 
1179 	memset(&uresp, 0, sizeof(uresp));
1180 	usr_db = dev->nic_info.unmapped_db +
1181 			(pd->id * dev->nic_info.db_page_size);
1182 	uresp.qp_id = qp->id;
1183 	uresp.sq_dbid = qp->sq.dbid;
1184 	uresp.num_sq_pages = 1;
1185 	uresp.sq_page_size = PAGE_ALIGN(qp->sq.len);
1186 	uresp.sq_page_addr[0] = virt_to_phys(qp->sq.va);
1187 	uresp.num_wqe_allocated = qp->sq.max_cnt;
1188 	if (!srq) {
1189 		uresp.rq_dbid = qp->rq.dbid;
1190 		uresp.num_rq_pages = 1;
1191 		uresp.rq_page_size = PAGE_ALIGN(qp->rq.len);
1192 		uresp.rq_page_addr[0] = virt_to_phys(qp->rq.va);
1193 		uresp.num_rqe_allocated = qp->rq.max_cnt;
1194 	}
1195 	uresp.db_page_addr = usr_db;
1196 	uresp.db_page_size = dev->nic_info.db_page_size;
1197 	uresp.db_sq_offset = OCRDMA_DB_GEN2_SQ_OFFSET;
1198 	uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1199 	uresp.db_shift = OCRDMA_DB_RQ_SHIFT;
1200 
1201 	if (qp->dpp_enabled) {
1202 		uresp.dpp_credit = dpp_credit_lmt;
1203 		uresp.dpp_offset = dpp_offset;
1204 	}
1205 	status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1206 	if (status) {
1207 		pr_err("%s(%d) user copy error.\n", __func__, dev->id);
1208 		goto err;
1209 	}
1210 	status = ocrdma_add_mmap(pd->uctx, uresp.sq_page_addr[0],
1211 				 uresp.sq_page_size);
1212 	if (status)
1213 		goto err;
1214 
1215 	if (!srq) {
1216 		status = ocrdma_add_mmap(pd->uctx, uresp.rq_page_addr[0],
1217 					 uresp.rq_page_size);
1218 		if (status)
1219 			goto rq_map_err;
1220 	}
1221 	return status;
1222 rq_map_err:
1223 	ocrdma_del_mmap(pd->uctx, uresp.sq_page_addr[0], uresp.sq_page_size);
1224 err:
1225 	return status;
1226 }
1227 
1228 static void ocrdma_set_qp_db(struct ocrdma_dev *dev, struct ocrdma_qp *qp,
1229 			     struct ocrdma_pd *pd)
1230 {
1231 	if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1232 		qp->sq_db = dev->nic_info.db +
1233 			(pd->id * dev->nic_info.db_page_size) +
1234 			OCRDMA_DB_GEN2_SQ_OFFSET;
1235 		qp->rq_db = dev->nic_info.db +
1236 			(pd->id * dev->nic_info.db_page_size) +
1237 			OCRDMA_DB_GEN2_RQ_OFFSET;
1238 	} else {
1239 		qp->sq_db = dev->nic_info.db +
1240 			(pd->id * dev->nic_info.db_page_size) +
1241 			OCRDMA_DB_SQ_OFFSET;
1242 		qp->rq_db = dev->nic_info.db +
1243 			(pd->id * dev->nic_info.db_page_size) +
1244 			OCRDMA_DB_RQ_OFFSET;
1245 	}
1246 }
1247 
1248 static int ocrdma_alloc_wr_id_tbl(struct ocrdma_qp *qp)
1249 {
1250 	qp->wqe_wr_id_tbl =
1251 	    kcalloc(qp->sq.max_cnt, sizeof(*(qp->wqe_wr_id_tbl)),
1252 		    GFP_KERNEL);
1253 	if (qp->wqe_wr_id_tbl == NULL)
1254 		return -ENOMEM;
1255 	qp->rqe_wr_id_tbl =
1256 	    kcalloc(qp->rq.max_cnt, sizeof(u64), GFP_KERNEL);
1257 	if (qp->rqe_wr_id_tbl == NULL)
1258 		return -ENOMEM;
1259 
1260 	return 0;
1261 }
1262 
1263 static void ocrdma_set_qp_init_params(struct ocrdma_qp *qp,
1264 				      struct ocrdma_pd *pd,
1265 				      struct ib_qp_init_attr *attrs)
1266 {
1267 	qp->pd = pd;
1268 	spin_lock_init(&qp->q_lock);
1269 	INIT_LIST_HEAD(&qp->sq_entry);
1270 	INIT_LIST_HEAD(&qp->rq_entry);
1271 
1272 	qp->qp_type = attrs->qp_type;
1273 	qp->cap_flags = OCRDMA_QP_INB_RD | OCRDMA_QP_INB_WR;
1274 	qp->max_inline_data = attrs->cap.max_inline_data;
1275 	qp->sq.max_sges = attrs->cap.max_send_sge;
1276 	qp->rq.max_sges = attrs->cap.max_recv_sge;
1277 	qp->state = OCRDMA_QPS_RST;
1278 	qp->signaled = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR) ? true : false;
1279 }
1280 
1281 static void ocrdma_store_gsi_qp_cq(struct ocrdma_dev *dev,
1282 				   struct ib_qp_init_attr *attrs)
1283 {
1284 	if (attrs->qp_type == IB_QPT_GSI) {
1285 		dev->gsi_qp_created = 1;
1286 		dev->gsi_sqcq = get_ocrdma_cq(attrs->send_cq);
1287 		dev->gsi_rqcq = get_ocrdma_cq(attrs->recv_cq);
1288 	}
1289 }
1290 
1291 struct ib_qp *ocrdma_create_qp(struct ib_pd *ibpd,
1292 			       struct ib_qp_init_attr *attrs,
1293 			       struct ib_udata *udata)
1294 {
1295 	int status;
1296 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1297 	struct ocrdma_qp *qp;
1298 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1299 	struct ocrdma_create_qp_ureq ureq;
1300 	u16 dpp_credit_lmt, dpp_offset;
1301 
1302 	status = ocrdma_check_qp_params(ibpd, dev, attrs, udata);
1303 	if (status)
1304 		goto gen_err;
1305 
1306 	memset(&ureq, 0, sizeof(ureq));
1307 	if (udata) {
1308 		if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1309 			return ERR_PTR(-EFAULT);
1310 	}
1311 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1312 	if (!qp) {
1313 		status = -ENOMEM;
1314 		goto gen_err;
1315 	}
1316 	ocrdma_set_qp_init_params(qp, pd, attrs);
1317 	if (udata == NULL)
1318 		qp->cap_flags |= (OCRDMA_QP_MW_BIND | OCRDMA_QP_LKEY0 |
1319 					OCRDMA_QP_FAST_REG);
1320 
1321 	mutex_lock(&dev->dev_lock);
1322 	status = ocrdma_mbx_create_qp(qp, attrs, ureq.enable_dpp_cq,
1323 					ureq.dpp_cq_id,
1324 					&dpp_offset, &dpp_credit_lmt);
1325 	if (status)
1326 		goto mbx_err;
1327 
1328 	/* user space QP's wr_id table are managed in library */
1329 	if (udata == NULL) {
1330 		status = ocrdma_alloc_wr_id_tbl(qp);
1331 		if (status)
1332 			goto map_err;
1333 	}
1334 
1335 	status = ocrdma_add_qpn_map(dev, qp);
1336 	if (status)
1337 		goto map_err;
1338 	ocrdma_set_qp_db(dev, qp, pd);
1339 	if (udata) {
1340 		status = ocrdma_copy_qp_uresp(qp, udata, dpp_offset,
1341 					      dpp_credit_lmt,
1342 					      (attrs->srq != NULL));
1343 		if (status)
1344 			goto cpy_err;
1345 	}
1346 	ocrdma_store_gsi_qp_cq(dev, attrs);
1347 	qp->ibqp.qp_num = qp->id;
1348 	mutex_unlock(&dev->dev_lock);
1349 	return &qp->ibqp;
1350 
1351 cpy_err:
1352 	ocrdma_del_qpn_map(dev, qp);
1353 map_err:
1354 	ocrdma_mbx_destroy_qp(dev, qp);
1355 mbx_err:
1356 	mutex_unlock(&dev->dev_lock);
1357 	kfree(qp->wqe_wr_id_tbl);
1358 	kfree(qp->rqe_wr_id_tbl);
1359 	kfree(qp);
1360 	pr_err("%s(%d) error=%d\n", __func__, dev->id, status);
1361 gen_err:
1362 	return ERR_PTR(status);
1363 }
1364 
1365 int _ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1366 		      int attr_mask)
1367 {
1368 	int status = 0;
1369 	struct ocrdma_qp *qp;
1370 	struct ocrdma_dev *dev;
1371 	enum ib_qp_state old_qps;
1372 
1373 	qp = get_ocrdma_qp(ibqp);
1374 	dev = get_ocrdma_dev(ibqp->device);
1375 	if (attr_mask & IB_QP_STATE)
1376 		status = ocrdma_qp_state_change(qp, attr->qp_state, &old_qps);
1377 	/* if new and previous states are same hw doesn't need to
1378 	 * know about it.
1379 	 */
1380 	if (status < 0)
1381 		return status;
1382 	return ocrdma_mbx_modify_qp(dev, qp, attr, attr_mask);
1383 }
1384 
1385 int ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1386 		     int attr_mask, struct ib_udata *udata)
1387 {
1388 	unsigned long flags;
1389 	int status = -EINVAL;
1390 	struct ocrdma_qp *qp;
1391 	struct ocrdma_dev *dev;
1392 	enum ib_qp_state old_qps, new_qps;
1393 
1394 	qp = get_ocrdma_qp(ibqp);
1395 	dev = get_ocrdma_dev(ibqp->device);
1396 
1397 	/* syncronize with multiple context trying to change, retrive qps */
1398 	mutex_lock(&dev->dev_lock);
1399 	/* syncronize with wqe, rqe posting and cqe processing contexts */
1400 	spin_lock_irqsave(&qp->q_lock, flags);
1401 	old_qps = get_ibqp_state(qp->state);
1402 	if (attr_mask & IB_QP_STATE)
1403 		new_qps = attr->qp_state;
1404 	else
1405 		new_qps = old_qps;
1406 	spin_unlock_irqrestore(&qp->q_lock, flags);
1407 
1408 	if (!ib_modify_qp_is_ok(old_qps, new_qps, ibqp->qp_type, attr_mask)) {
1409 		pr_err("%s(%d) invalid attribute mask=0x%x specified for\n"
1410 		       "qpn=0x%x of type=0x%x old_qps=0x%x, new_qps=0x%x\n",
1411 		       __func__, dev->id, attr_mask, qp->id, ibqp->qp_type,
1412 		       old_qps, new_qps);
1413 		goto param_err;
1414 	}
1415 
1416 	status = _ocrdma_modify_qp(ibqp, attr, attr_mask);
1417 	if (status > 0)
1418 		status = 0;
1419 param_err:
1420 	mutex_unlock(&dev->dev_lock);
1421 	return status;
1422 }
1423 
1424 static enum ib_mtu ocrdma_mtu_int_to_enum(u16 mtu)
1425 {
1426 	switch (mtu) {
1427 	case 256:
1428 		return IB_MTU_256;
1429 	case 512:
1430 		return IB_MTU_512;
1431 	case 1024:
1432 		return IB_MTU_1024;
1433 	case 2048:
1434 		return IB_MTU_2048;
1435 	case 4096:
1436 		return IB_MTU_4096;
1437 	default:
1438 		return IB_MTU_1024;
1439 	}
1440 }
1441 
1442 static int ocrdma_to_ib_qp_acc_flags(int qp_cap_flags)
1443 {
1444 	int ib_qp_acc_flags = 0;
1445 
1446 	if (qp_cap_flags & OCRDMA_QP_INB_WR)
1447 		ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE;
1448 	if (qp_cap_flags & OCRDMA_QP_INB_RD)
1449 		ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE;
1450 	return ib_qp_acc_flags;
1451 }
1452 
1453 int ocrdma_query_qp(struct ib_qp *ibqp,
1454 		    struct ib_qp_attr *qp_attr,
1455 		    int attr_mask, struct ib_qp_init_attr *qp_init_attr)
1456 {
1457 	int status;
1458 	u32 qp_state;
1459 	struct ocrdma_qp_params params;
1460 	struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
1461 	struct ocrdma_dev *dev = get_ocrdma_dev(ibqp->device);
1462 
1463 	memset(&params, 0, sizeof(params));
1464 	mutex_lock(&dev->dev_lock);
1465 	status = ocrdma_mbx_query_qp(dev, qp, &params);
1466 	mutex_unlock(&dev->dev_lock);
1467 	if (status)
1468 		goto mbx_err;
1469 	if (qp->qp_type == IB_QPT_UD)
1470 		qp_attr->qkey = params.qkey;
1471 	qp_attr->path_mtu =
1472 		ocrdma_mtu_int_to_enum(params.path_mtu_pkey_indx &
1473 				OCRDMA_QP_PARAMS_PATH_MTU_MASK) >>
1474 				OCRDMA_QP_PARAMS_PATH_MTU_SHIFT;
1475 	qp_attr->path_mig_state = IB_MIG_MIGRATED;
1476 	qp_attr->rq_psn = params.hop_lmt_rq_psn & OCRDMA_QP_PARAMS_RQ_PSN_MASK;
1477 	qp_attr->sq_psn = params.tclass_sq_psn & OCRDMA_QP_PARAMS_SQ_PSN_MASK;
1478 	qp_attr->dest_qp_num =
1479 	    params.ack_to_rnr_rtc_dest_qpn & OCRDMA_QP_PARAMS_DEST_QPN_MASK;
1480 
1481 	qp_attr->qp_access_flags = ocrdma_to_ib_qp_acc_flags(qp->cap_flags);
1482 	qp_attr->cap.max_send_wr = qp->sq.max_cnt - 1;
1483 	qp_attr->cap.max_recv_wr = qp->rq.max_cnt - 1;
1484 	qp_attr->cap.max_send_sge = qp->sq.max_sges;
1485 	qp_attr->cap.max_recv_sge = qp->rq.max_sges;
1486 	qp_attr->cap.max_inline_data = qp->max_inline_data;
1487 	qp_init_attr->cap = qp_attr->cap;
1488 	qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
1489 
1490 	rdma_ah_set_grh(&qp_attr->ah_attr, NULL,
1491 			params.rnt_rc_sl_fl &
1492 			  OCRDMA_QP_PARAMS_FLOW_LABEL_MASK,
1493 			qp->sgid_idx,
1494 			(params.hop_lmt_rq_psn &
1495 			 OCRDMA_QP_PARAMS_HOP_LMT_MASK) >>
1496 			 OCRDMA_QP_PARAMS_HOP_LMT_SHIFT,
1497 			(params.tclass_sq_psn &
1498 			 OCRDMA_QP_PARAMS_TCLASS_MASK) >>
1499 			 OCRDMA_QP_PARAMS_TCLASS_SHIFT);
1500 	rdma_ah_set_dgid_raw(&qp_attr->ah_attr, &params.dgid[0]);
1501 
1502 	rdma_ah_set_port_num(&qp_attr->ah_attr, 1);
1503 	rdma_ah_set_sl(&qp_attr->ah_attr, (params.rnt_rc_sl_fl &
1504 					   OCRDMA_QP_PARAMS_SL_MASK) >>
1505 					   OCRDMA_QP_PARAMS_SL_SHIFT);
1506 	qp_attr->timeout = (params.ack_to_rnr_rtc_dest_qpn &
1507 			    OCRDMA_QP_PARAMS_ACK_TIMEOUT_MASK) >>
1508 				OCRDMA_QP_PARAMS_ACK_TIMEOUT_SHIFT;
1509 	qp_attr->rnr_retry = (params.ack_to_rnr_rtc_dest_qpn &
1510 			      OCRDMA_QP_PARAMS_RNR_RETRY_CNT_MASK) >>
1511 				OCRDMA_QP_PARAMS_RNR_RETRY_CNT_SHIFT;
1512 	qp_attr->retry_cnt =
1513 	    (params.rnt_rc_sl_fl & OCRDMA_QP_PARAMS_RETRY_CNT_MASK) >>
1514 		OCRDMA_QP_PARAMS_RETRY_CNT_SHIFT;
1515 	qp_attr->min_rnr_timer = 0;
1516 	qp_attr->pkey_index = 0;
1517 	qp_attr->port_num = 1;
1518 	rdma_ah_set_path_bits(&qp_attr->ah_attr, 0);
1519 	rdma_ah_set_static_rate(&qp_attr->ah_attr, 0);
1520 	qp_attr->alt_pkey_index = 0;
1521 	qp_attr->alt_port_num = 0;
1522 	qp_attr->alt_timeout = 0;
1523 	memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr));
1524 	qp_state = (params.max_sge_recv_flags & OCRDMA_QP_PARAMS_STATE_MASK) >>
1525 		    OCRDMA_QP_PARAMS_STATE_SHIFT;
1526 	qp_attr->qp_state = get_ibqp_state(qp_state);
1527 	qp_attr->cur_qp_state = qp_attr->qp_state;
1528 	qp_attr->sq_draining = (qp_state == OCRDMA_QPS_SQ_DRAINING) ? 1 : 0;
1529 	qp_attr->max_dest_rd_atomic =
1530 	    params.max_ord_ird >> OCRDMA_QP_PARAMS_MAX_ORD_SHIFT;
1531 	qp_attr->max_rd_atomic =
1532 	    params.max_ord_ird & OCRDMA_QP_PARAMS_MAX_IRD_MASK;
1533 	qp_attr->en_sqd_async_notify = (params.max_sge_recv_flags &
1534 				OCRDMA_QP_PARAMS_FLAGS_SQD_ASYNC) ? 1 : 0;
1535 	/* Sync driver QP state with FW */
1536 	ocrdma_qp_state_change(qp, qp_attr->qp_state, NULL);
1537 mbx_err:
1538 	return status;
1539 }
1540 
1541 static void ocrdma_srq_toggle_bit(struct ocrdma_srq *srq, unsigned int idx)
1542 {
1543 	unsigned int i = idx / 32;
1544 	u32 mask = (1U << (idx % 32));
1545 
1546 	srq->idx_bit_fields[i] ^= mask;
1547 }
1548 
1549 static int ocrdma_hwq_free_cnt(struct ocrdma_qp_hwq_info *q)
1550 {
1551 	return ((q->max_wqe_idx - q->head) + q->tail) % q->max_cnt;
1552 }
1553 
1554 static int is_hw_sq_empty(struct ocrdma_qp *qp)
1555 {
1556 	return (qp->sq.tail == qp->sq.head);
1557 }
1558 
1559 static int is_hw_rq_empty(struct ocrdma_qp *qp)
1560 {
1561 	return (qp->rq.tail == qp->rq.head);
1562 }
1563 
1564 static void *ocrdma_hwq_head(struct ocrdma_qp_hwq_info *q)
1565 {
1566 	return q->va + (q->head * q->entry_size);
1567 }
1568 
1569 static void *ocrdma_hwq_head_from_idx(struct ocrdma_qp_hwq_info *q,
1570 				      u32 idx)
1571 {
1572 	return q->va + (idx * q->entry_size);
1573 }
1574 
1575 static void ocrdma_hwq_inc_head(struct ocrdma_qp_hwq_info *q)
1576 {
1577 	q->head = (q->head + 1) & q->max_wqe_idx;
1578 }
1579 
1580 static void ocrdma_hwq_inc_tail(struct ocrdma_qp_hwq_info *q)
1581 {
1582 	q->tail = (q->tail + 1) & q->max_wqe_idx;
1583 }
1584 
1585 /* discard the cqe for a given QP */
1586 static void ocrdma_discard_cqes(struct ocrdma_qp *qp, struct ocrdma_cq *cq)
1587 {
1588 	unsigned long cq_flags;
1589 	unsigned long flags;
1590 	int discard_cnt = 0;
1591 	u32 cur_getp, stop_getp;
1592 	struct ocrdma_cqe *cqe;
1593 	u32 qpn = 0, wqe_idx = 0;
1594 
1595 	spin_lock_irqsave(&cq->cq_lock, cq_flags);
1596 
1597 	/* traverse through the CQEs in the hw CQ,
1598 	 * find the matching CQE for a given qp,
1599 	 * mark the matching one discarded by clearing qpn.
1600 	 * ring the doorbell in the poll_cq() as
1601 	 * we don't complete out of order cqe.
1602 	 */
1603 
1604 	cur_getp = cq->getp;
1605 	/* find upto when do we reap the cq. */
1606 	stop_getp = cur_getp;
1607 	do {
1608 		if (is_hw_sq_empty(qp) && (!qp->srq && is_hw_rq_empty(qp)))
1609 			break;
1610 
1611 		cqe = cq->va + cur_getp;
1612 		/* if (a) done reaping whole hw cq, or
1613 		 *    (b) qp_xq becomes empty.
1614 		 * then exit
1615 		 */
1616 		qpn = cqe->cmn.qpn & OCRDMA_CQE_QPN_MASK;
1617 		/* if previously discarded cqe found, skip that too. */
1618 		/* check for matching qp */
1619 		if (qpn == 0 || qpn != qp->id)
1620 			goto skip_cqe;
1621 
1622 		if (is_cqe_for_sq(cqe)) {
1623 			ocrdma_hwq_inc_tail(&qp->sq);
1624 		} else {
1625 			if (qp->srq) {
1626 				wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
1627 					OCRDMA_CQE_BUFTAG_SHIFT) &
1628 					qp->srq->rq.max_wqe_idx;
1629 				BUG_ON(wqe_idx < 1);
1630 				spin_lock_irqsave(&qp->srq->q_lock, flags);
1631 				ocrdma_hwq_inc_tail(&qp->srq->rq);
1632 				ocrdma_srq_toggle_bit(qp->srq, wqe_idx - 1);
1633 				spin_unlock_irqrestore(&qp->srq->q_lock, flags);
1634 
1635 			} else {
1636 				ocrdma_hwq_inc_tail(&qp->rq);
1637 			}
1638 		}
1639 		/* mark cqe discarded so that it is not picked up later
1640 		 * in the poll_cq().
1641 		 */
1642 		discard_cnt += 1;
1643 		cqe->cmn.qpn = 0;
1644 skip_cqe:
1645 		cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
1646 	} while (cur_getp != stop_getp);
1647 	spin_unlock_irqrestore(&cq->cq_lock, cq_flags);
1648 }
1649 
1650 void ocrdma_del_flush_qp(struct ocrdma_qp *qp)
1651 {
1652 	int found = false;
1653 	unsigned long flags;
1654 	struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
1655 	/* sync with any active CQ poll */
1656 
1657 	spin_lock_irqsave(&dev->flush_q_lock, flags);
1658 	found = ocrdma_is_qp_in_sq_flushlist(qp->sq_cq, qp);
1659 	if (found)
1660 		list_del(&qp->sq_entry);
1661 	if (!qp->srq) {
1662 		found = ocrdma_is_qp_in_rq_flushlist(qp->rq_cq, qp);
1663 		if (found)
1664 			list_del(&qp->rq_entry);
1665 	}
1666 	spin_unlock_irqrestore(&dev->flush_q_lock, flags);
1667 }
1668 
1669 int ocrdma_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
1670 {
1671 	struct ocrdma_pd *pd;
1672 	struct ocrdma_qp *qp;
1673 	struct ocrdma_dev *dev;
1674 	struct ib_qp_attr attrs;
1675 	int attr_mask;
1676 	unsigned long flags;
1677 
1678 	qp = get_ocrdma_qp(ibqp);
1679 	dev = get_ocrdma_dev(ibqp->device);
1680 
1681 	pd = qp->pd;
1682 
1683 	/* change the QP state to ERROR */
1684 	if (qp->state != OCRDMA_QPS_RST) {
1685 		attrs.qp_state = IB_QPS_ERR;
1686 		attr_mask = IB_QP_STATE;
1687 		_ocrdma_modify_qp(ibqp, &attrs, attr_mask);
1688 	}
1689 	/* ensure that CQEs for newly created QP (whose id may be same with
1690 	 * one which just getting destroyed are same), dont get
1691 	 * discarded until the old CQEs are discarded.
1692 	 */
1693 	mutex_lock(&dev->dev_lock);
1694 	(void) ocrdma_mbx_destroy_qp(dev, qp);
1695 
1696 	/*
1697 	 * acquire CQ lock while destroy is in progress, in order to
1698 	 * protect against proessing in-flight CQEs for this QP.
1699 	 */
1700 	spin_lock_irqsave(&qp->sq_cq->cq_lock, flags);
1701 	if (qp->rq_cq && (qp->rq_cq != qp->sq_cq)) {
1702 		spin_lock(&qp->rq_cq->cq_lock);
1703 		ocrdma_del_qpn_map(dev, qp);
1704 		spin_unlock(&qp->rq_cq->cq_lock);
1705 	} else {
1706 		ocrdma_del_qpn_map(dev, qp);
1707 	}
1708 	spin_unlock_irqrestore(&qp->sq_cq->cq_lock, flags);
1709 
1710 	if (!pd->uctx) {
1711 		ocrdma_discard_cqes(qp, qp->sq_cq);
1712 		ocrdma_discard_cqes(qp, qp->rq_cq);
1713 	}
1714 	mutex_unlock(&dev->dev_lock);
1715 
1716 	if (pd->uctx) {
1717 		ocrdma_del_mmap(pd->uctx, (u64) qp->sq.pa,
1718 				PAGE_ALIGN(qp->sq.len));
1719 		if (!qp->srq)
1720 			ocrdma_del_mmap(pd->uctx, (u64) qp->rq.pa,
1721 					PAGE_ALIGN(qp->rq.len));
1722 	}
1723 
1724 	ocrdma_del_flush_qp(qp);
1725 
1726 	kfree(qp->wqe_wr_id_tbl);
1727 	kfree(qp->rqe_wr_id_tbl);
1728 	kfree(qp);
1729 	return 0;
1730 }
1731 
1732 static int ocrdma_copy_srq_uresp(struct ocrdma_dev *dev, struct ocrdma_srq *srq,
1733 				struct ib_udata *udata)
1734 {
1735 	int status;
1736 	struct ocrdma_create_srq_uresp uresp;
1737 
1738 	memset(&uresp, 0, sizeof(uresp));
1739 	uresp.rq_dbid = srq->rq.dbid;
1740 	uresp.num_rq_pages = 1;
1741 	uresp.rq_page_addr[0] = virt_to_phys(srq->rq.va);
1742 	uresp.rq_page_size = srq->rq.len;
1743 	uresp.db_page_addr = dev->nic_info.unmapped_db +
1744 	    (srq->pd->id * dev->nic_info.db_page_size);
1745 	uresp.db_page_size = dev->nic_info.db_page_size;
1746 	uresp.num_rqe_allocated = srq->rq.max_cnt;
1747 	if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1748 		uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1749 		uresp.db_shift = 24;
1750 	} else {
1751 		uresp.db_rq_offset = OCRDMA_DB_RQ_OFFSET;
1752 		uresp.db_shift = 16;
1753 	}
1754 
1755 	status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1756 	if (status)
1757 		return status;
1758 	status = ocrdma_add_mmap(srq->pd->uctx, uresp.rq_page_addr[0],
1759 				 uresp.rq_page_size);
1760 	if (status)
1761 		return status;
1762 	return status;
1763 }
1764 
1765 int ocrdma_create_srq(struct ib_srq *ibsrq, struct ib_srq_init_attr *init_attr,
1766 		      struct ib_udata *udata)
1767 {
1768 	int status;
1769 	struct ocrdma_pd *pd = get_ocrdma_pd(ibsrq->pd);
1770 	struct ocrdma_dev *dev = get_ocrdma_dev(ibsrq->device);
1771 	struct ocrdma_srq *srq = get_ocrdma_srq(ibsrq);
1772 
1773 	if (init_attr->attr.max_sge > dev->attr.max_recv_sge)
1774 		return -EINVAL;
1775 	if (init_attr->attr.max_wr > dev->attr.max_rqe)
1776 		return -EINVAL;
1777 
1778 	spin_lock_init(&srq->q_lock);
1779 	srq->pd = pd;
1780 	srq->db = dev->nic_info.db + (pd->id * dev->nic_info.db_page_size);
1781 	status = ocrdma_mbx_create_srq(dev, srq, init_attr, pd);
1782 	if (status)
1783 		return status;
1784 
1785 	if (!udata) {
1786 		srq->rqe_wr_id_tbl = kcalloc(srq->rq.max_cnt, sizeof(u64),
1787 					     GFP_KERNEL);
1788 		if (!srq->rqe_wr_id_tbl) {
1789 			status = -ENOMEM;
1790 			goto arm_err;
1791 		}
1792 
1793 		srq->bit_fields_len = (srq->rq.max_cnt / 32) +
1794 		    (srq->rq.max_cnt % 32 ? 1 : 0);
1795 		srq->idx_bit_fields =
1796 		    kmalloc_array(srq->bit_fields_len, sizeof(u32),
1797 				  GFP_KERNEL);
1798 		if (!srq->idx_bit_fields) {
1799 			status = -ENOMEM;
1800 			goto arm_err;
1801 		}
1802 		memset(srq->idx_bit_fields, 0xff,
1803 		       srq->bit_fields_len * sizeof(u32));
1804 	}
1805 
1806 	if (init_attr->attr.srq_limit) {
1807 		status = ocrdma_mbx_modify_srq(srq, &init_attr->attr);
1808 		if (status)
1809 			goto arm_err;
1810 	}
1811 
1812 	if (udata) {
1813 		status = ocrdma_copy_srq_uresp(dev, srq, udata);
1814 		if (status)
1815 			goto arm_err;
1816 	}
1817 
1818 	return 0;
1819 
1820 arm_err:
1821 	ocrdma_mbx_destroy_srq(dev, srq);
1822 	kfree(srq->rqe_wr_id_tbl);
1823 	kfree(srq->idx_bit_fields);
1824 	return status;
1825 }
1826 
1827 int ocrdma_modify_srq(struct ib_srq *ibsrq,
1828 		      struct ib_srq_attr *srq_attr,
1829 		      enum ib_srq_attr_mask srq_attr_mask,
1830 		      struct ib_udata *udata)
1831 {
1832 	int status;
1833 	struct ocrdma_srq *srq;
1834 
1835 	srq = get_ocrdma_srq(ibsrq);
1836 	if (srq_attr_mask & IB_SRQ_MAX_WR)
1837 		status = -EINVAL;
1838 	else
1839 		status = ocrdma_mbx_modify_srq(srq, srq_attr);
1840 	return status;
1841 }
1842 
1843 int ocrdma_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
1844 {
1845 	int status;
1846 	struct ocrdma_srq *srq;
1847 
1848 	srq = get_ocrdma_srq(ibsrq);
1849 	status = ocrdma_mbx_query_srq(srq, srq_attr);
1850 	return status;
1851 }
1852 
1853 int ocrdma_destroy_srq(struct ib_srq *ibsrq, struct ib_udata *udata)
1854 {
1855 	struct ocrdma_srq *srq;
1856 	struct ocrdma_dev *dev = get_ocrdma_dev(ibsrq->device);
1857 
1858 	srq = get_ocrdma_srq(ibsrq);
1859 
1860 	ocrdma_mbx_destroy_srq(dev, srq);
1861 
1862 	if (srq->pd->uctx)
1863 		ocrdma_del_mmap(srq->pd->uctx, (u64) srq->rq.pa,
1864 				PAGE_ALIGN(srq->rq.len));
1865 
1866 	kfree(srq->idx_bit_fields);
1867 	kfree(srq->rqe_wr_id_tbl);
1868 	return 0;
1869 }
1870 
1871 /* unprivileged verbs and their support functions. */
1872 static void ocrdma_build_ud_hdr(struct ocrdma_qp *qp,
1873 				struct ocrdma_hdr_wqe *hdr,
1874 				const struct ib_send_wr *wr)
1875 {
1876 	struct ocrdma_ewqe_ud_hdr *ud_hdr =
1877 		(struct ocrdma_ewqe_ud_hdr *)(hdr + 1);
1878 	struct ocrdma_ah *ah = get_ocrdma_ah(ud_wr(wr)->ah);
1879 
1880 	ud_hdr->rsvd_dest_qpn = ud_wr(wr)->remote_qpn;
1881 	if (qp->qp_type == IB_QPT_GSI)
1882 		ud_hdr->qkey = qp->qkey;
1883 	else
1884 		ud_hdr->qkey = ud_wr(wr)->remote_qkey;
1885 	ud_hdr->rsvd_ahid = ah->id;
1886 	ud_hdr->hdr_type = ah->hdr_type;
1887 	if (ah->av->valid & OCRDMA_AV_VLAN_VALID)
1888 		hdr->cw |= (OCRDMA_FLAG_AH_VLAN_PR << OCRDMA_WQE_FLAGS_SHIFT);
1889 }
1890 
1891 static void ocrdma_build_sges(struct ocrdma_hdr_wqe *hdr,
1892 			      struct ocrdma_sge *sge, int num_sge,
1893 			      struct ib_sge *sg_list)
1894 {
1895 	int i;
1896 
1897 	for (i = 0; i < num_sge; i++) {
1898 		sge[i].lrkey = sg_list[i].lkey;
1899 		sge[i].addr_lo = sg_list[i].addr;
1900 		sge[i].addr_hi = upper_32_bits(sg_list[i].addr);
1901 		sge[i].len = sg_list[i].length;
1902 		hdr->total_len += sg_list[i].length;
1903 	}
1904 	if (num_sge == 0)
1905 		memset(sge, 0, sizeof(*sge));
1906 }
1907 
1908 static inline uint32_t ocrdma_sglist_len(struct ib_sge *sg_list, int num_sge)
1909 {
1910 	uint32_t total_len = 0, i;
1911 
1912 	for (i = 0; i < num_sge; i++)
1913 		total_len += sg_list[i].length;
1914 	return total_len;
1915 }
1916 
1917 
1918 static int ocrdma_build_inline_sges(struct ocrdma_qp *qp,
1919 				    struct ocrdma_hdr_wqe *hdr,
1920 				    struct ocrdma_sge *sge,
1921 				    const struct ib_send_wr *wr, u32 wqe_size)
1922 {
1923 	int i;
1924 	char *dpp_addr;
1925 
1926 	if (wr->send_flags & IB_SEND_INLINE && qp->qp_type != IB_QPT_UD) {
1927 		hdr->total_len = ocrdma_sglist_len(wr->sg_list, wr->num_sge);
1928 		if (unlikely(hdr->total_len > qp->max_inline_data)) {
1929 			pr_err("%s() supported_len=0x%x,\n"
1930 			       " unsupported len req=0x%x\n", __func__,
1931 				qp->max_inline_data, hdr->total_len);
1932 			return -EINVAL;
1933 		}
1934 		dpp_addr = (char *)sge;
1935 		for (i = 0; i < wr->num_sge; i++) {
1936 			memcpy(dpp_addr,
1937 			       (void *)(unsigned long)wr->sg_list[i].addr,
1938 			       wr->sg_list[i].length);
1939 			dpp_addr += wr->sg_list[i].length;
1940 		}
1941 
1942 		wqe_size += roundup(hdr->total_len, OCRDMA_WQE_ALIGN_BYTES);
1943 		if (0 == hdr->total_len)
1944 			wqe_size += sizeof(struct ocrdma_sge);
1945 		hdr->cw |= (OCRDMA_TYPE_INLINE << OCRDMA_WQE_TYPE_SHIFT);
1946 	} else {
1947 		ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
1948 		if (wr->num_sge)
1949 			wqe_size += (wr->num_sge * sizeof(struct ocrdma_sge));
1950 		else
1951 			wqe_size += sizeof(struct ocrdma_sge);
1952 		hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
1953 	}
1954 	hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
1955 	return 0;
1956 }
1957 
1958 static int ocrdma_build_send(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1959 			     const struct ib_send_wr *wr)
1960 {
1961 	int status;
1962 	struct ocrdma_sge *sge;
1963 	u32 wqe_size = sizeof(*hdr);
1964 
1965 	if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
1966 		ocrdma_build_ud_hdr(qp, hdr, wr);
1967 		sge = (struct ocrdma_sge *)(hdr + 2);
1968 		wqe_size += sizeof(struct ocrdma_ewqe_ud_hdr);
1969 	} else {
1970 		sge = (struct ocrdma_sge *)(hdr + 1);
1971 	}
1972 
1973 	status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
1974 	return status;
1975 }
1976 
1977 static int ocrdma_build_write(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1978 			      const struct ib_send_wr *wr)
1979 {
1980 	int status;
1981 	struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
1982 	struct ocrdma_sge *sge = ext_rw + 1;
1983 	u32 wqe_size = sizeof(*hdr) + sizeof(*ext_rw);
1984 
1985 	status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
1986 	if (status)
1987 		return status;
1988 	ext_rw->addr_lo = rdma_wr(wr)->remote_addr;
1989 	ext_rw->addr_hi = upper_32_bits(rdma_wr(wr)->remote_addr);
1990 	ext_rw->lrkey = rdma_wr(wr)->rkey;
1991 	ext_rw->len = hdr->total_len;
1992 	return 0;
1993 }
1994 
1995 static void ocrdma_build_read(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1996 			      const struct ib_send_wr *wr)
1997 {
1998 	struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
1999 	struct ocrdma_sge *sge = ext_rw + 1;
2000 	u32 wqe_size = ((wr->num_sge + 1) * sizeof(struct ocrdma_sge)) +
2001 	    sizeof(struct ocrdma_hdr_wqe);
2002 
2003 	ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
2004 	hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2005 	hdr->cw |= (OCRDMA_READ << OCRDMA_WQE_OPCODE_SHIFT);
2006 	hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2007 
2008 	ext_rw->addr_lo = rdma_wr(wr)->remote_addr;
2009 	ext_rw->addr_hi = upper_32_bits(rdma_wr(wr)->remote_addr);
2010 	ext_rw->lrkey = rdma_wr(wr)->rkey;
2011 	ext_rw->len = hdr->total_len;
2012 }
2013 
2014 static int get_encoded_page_size(int pg_sz)
2015 {
2016 	/* Max size is 256M 4096 << 16 */
2017 	int i = 0;
2018 	for (; i < 17; i++)
2019 		if (pg_sz == (4096 << i))
2020 			break;
2021 	return i;
2022 }
2023 
2024 static int ocrdma_build_reg(struct ocrdma_qp *qp,
2025 			    struct ocrdma_hdr_wqe *hdr,
2026 			    const struct ib_reg_wr *wr)
2027 {
2028 	u64 fbo;
2029 	struct ocrdma_ewqe_fr *fast_reg = (struct ocrdma_ewqe_fr *)(hdr + 1);
2030 	struct ocrdma_mr *mr = get_ocrdma_mr(wr->mr);
2031 	struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
2032 	struct ocrdma_pbe *pbe;
2033 	u32 wqe_size = sizeof(*fast_reg) + sizeof(*hdr);
2034 	int num_pbes = 0, i;
2035 
2036 	wqe_size = roundup(wqe_size, OCRDMA_WQE_ALIGN_BYTES);
2037 
2038 	hdr->cw |= (OCRDMA_FR_MR << OCRDMA_WQE_OPCODE_SHIFT);
2039 	hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2040 
2041 	if (wr->access & IB_ACCESS_LOCAL_WRITE)
2042 		hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_LOCAL_WR;
2043 	if (wr->access & IB_ACCESS_REMOTE_WRITE)
2044 		hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_WR;
2045 	if (wr->access & IB_ACCESS_REMOTE_READ)
2046 		hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_RD;
2047 	hdr->lkey = wr->key;
2048 	hdr->total_len = mr->ibmr.length;
2049 
2050 	fbo = mr->ibmr.iova - mr->pages[0];
2051 
2052 	fast_reg->va_hi = upper_32_bits(mr->ibmr.iova);
2053 	fast_reg->va_lo = (u32) (mr->ibmr.iova & 0xffffffff);
2054 	fast_reg->fbo_hi = upper_32_bits(fbo);
2055 	fast_reg->fbo_lo = (u32) fbo & 0xffffffff;
2056 	fast_reg->num_sges = mr->npages;
2057 	fast_reg->size_sge = get_encoded_page_size(mr->ibmr.page_size);
2058 
2059 	pbe = pbl_tbl->va;
2060 	for (i = 0; i < mr->npages; i++) {
2061 		u64 buf_addr = mr->pages[i];
2062 
2063 		pbe->pa_lo = cpu_to_le32((u32) (buf_addr & PAGE_MASK));
2064 		pbe->pa_hi = cpu_to_le32((u32) upper_32_bits(buf_addr));
2065 		num_pbes += 1;
2066 		pbe++;
2067 
2068 		/* if the pbl is full storing the pbes,
2069 		 * move to next pbl.
2070 		*/
2071 		if (num_pbes == (mr->hwmr.pbl_size/sizeof(u64))) {
2072 			pbl_tbl++;
2073 			pbe = (struct ocrdma_pbe *)pbl_tbl->va;
2074 		}
2075 	}
2076 
2077 	return 0;
2078 }
2079 
2080 static void ocrdma_ring_sq_db(struct ocrdma_qp *qp)
2081 {
2082 	u32 val = qp->sq.dbid | (1 << OCRDMA_DB_SQ_SHIFT);
2083 
2084 	iowrite32(val, qp->sq_db);
2085 }
2086 
2087 int ocrdma_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
2088 		     const struct ib_send_wr **bad_wr)
2089 {
2090 	int status = 0;
2091 	struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2092 	struct ocrdma_hdr_wqe *hdr;
2093 	unsigned long flags;
2094 
2095 	spin_lock_irqsave(&qp->q_lock, flags);
2096 	if (qp->state != OCRDMA_QPS_RTS && qp->state != OCRDMA_QPS_SQD) {
2097 		spin_unlock_irqrestore(&qp->q_lock, flags);
2098 		*bad_wr = wr;
2099 		return -EINVAL;
2100 	}
2101 
2102 	while (wr) {
2103 		if (qp->qp_type == IB_QPT_UD &&
2104 		    (wr->opcode != IB_WR_SEND &&
2105 		     wr->opcode != IB_WR_SEND_WITH_IMM)) {
2106 			*bad_wr = wr;
2107 			status = -EINVAL;
2108 			break;
2109 		}
2110 		if (ocrdma_hwq_free_cnt(&qp->sq) == 0 ||
2111 		    wr->num_sge > qp->sq.max_sges) {
2112 			*bad_wr = wr;
2113 			status = -ENOMEM;
2114 			break;
2115 		}
2116 		hdr = ocrdma_hwq_head(&qp->sq);
2117 		hdr->cw = 0;
2118 		if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2119 			hdr->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2120 		if (wr->send_flags & IB_SEND_FENCE)
2121 			hdr->cw |=
2122 			    (OCRDMA_FLAG_FENCE_L << OCRDMA_WQE_FLAGS_SHIFT);
2123 		if (wr->send_flags & IB_SEND_SOLICITED)
2124 			hdr->cw |=
2125 			    (OCRDMA_FLAG_SOLICIT << OCRDMA_WQE_FLAGS_SHIFT);
2126 		hdr->total_len = 0;
2127 		switch (wr->opcode) {
2128 		case IB_WR_SEND_WITH_IMM:
2129 			hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2130 			hdr->immdt = ntohl(wr->ex.imm_data);
2131 			fallthrough;
2132 		case IB_WR_SEND:
2133 			hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2134 			ocrdma_build_send(qp, hdr, wr);
2135 			break;
2136 		case IB_WR_SEND_WITH_INV:
2137 			hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
2138 			hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2139 			hdr->lkey = wr->ex.invalidate_rkey;
2140 			status = ocrdma_build_send(qp, hdr, wr);
2141 			break;
2142 		case IB_WR_RDMA_WRITE_WITH_IMM:
2143 			hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2144 			hdr->immdt = ntohl(wr->ex.imm_data);
2145 			fallthrough;
2146 		case IB_WR_RDMA_WRITE:
2147 			hdr->cw |= (OCRDMA_WRITE << OCRDMA_WQE_OPCODE_SHIFT);
2148 			status = ocrdma_build_write(qp, hdr, wr);
2149 			break;
2150 		case IB_WR_RDMA_READ:
2151 			ocrdma_build_read(qp, hdr, wr);
2152 			break;
2153 		case IB_WR_LOCAL_INV:
2154 			hdr->cw |=
2155 			    (OCRDMA_LKEY_INV << OCRDMA_WQE_OPCODE_SHIFT);
2156 			hdr->cw |= ((sizeof(struct ocrdma_hdr_wqe) +
2157 					sizeof(struct ocrdma_sge)) /
2158 				OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT;
2159 			hdr->lkey = wr->ex.invalidate_rkey;
2160 			break;
2161 		case IB_WR_REG_MR:
2162 			status = ocrdma_build_reg(qp, hdr, reg_wr(wr));
2163 			break;
2164 		default:
2165 			status = -EINVAL;
2166 			break;
2167 		}
2168 		if (status) {
2169 			*bad_wr = wr;
2170 			break;
2171 		}
2172 		if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2173 			qp->wqe_wr_id_tbl[qp->sq.head].signaled = 1;
2174 		else
2175 			qp->wqe_wr_id_tbl[qp->sq.head].signaled = 0;
2176 		qp->wqe_wr_id_tbl[qp->sq.head].wrid = wr->wr_id;
2177 		ocrdma_cpu_to_le32(hdr, ((hdr->cw >> OCRDMA_WQE_SIZE_SHIFT) &
2178 				   OCRDMA_WQE_SIZE_MASK) * OCRDMA_WQE_STRIDE);
2179 		/* make sure wqe is written before adapter can access it */
2180 		wmb();
2181 		/* inform hw to start processing it */
2182 		ocrdma_ring_sq_db(qp);
2183 
2184 		/* update pointer, counter for next wr */
2185 		ocrdma_hwq_inc_head(&qp->sq);
2186 		wr = wr->next;
2187 	}
2188 	spin_unlock_irqrestore(&qp->q_lock, flags);
2189 	return status;
2190 }
2191 
2192 static void ocrdma_ring_rq_db(struct ocrdma_qp *qp)
2193 {
2194 	u32 val = qp->rq.dbid | (1 << OCRDMA_DB_RQ_SHIFT);
2195 
2196 	iowrite32(val, qp->rq_db);
2197 }
2198 
2199 static void ocrdma_build_rqe(struct ocrdma_hdr_wqe *rqe,
2200 			     const struct ib_recv_wr *wr, u16 tag)
2201 {
2202 	u32 wqe_size = 0;
2203 	struct ocrdma_sge *sge;
2204 	if (wr->num_sge)
2205 		wqe_size = (wr->num_sge * sizeof(*sge)) + sizeof(*rqe);
2206 	else
2207 		wqe_size = sizeof(*sge) + sizeof(*rqe);
2208 
2209 	rqe->cw = ((wqe_size / OCRDMA_WQE_STRIDE) <<
2210 				OCRDMA_WQE_SIZE_SHIFT);
2211 	rqe->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2212 	rqe->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2213 	rqe->total_len = 0;
2214 	rqe->rsvd_tag = tag;
2215 	sge = (struct ocrdma_sge *)(rqe + 1);
2216 	ocrdma_build_sges(rqe, sge, wr->num_sge, wr->sg_list);
2217 	ocrdma_cpu_to_le32(rqe, wqe_size);
2218 }
2219 
2220 int ocrdma_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
2221 		     const struct ib_recv_wr **bad_wr)
2222 {
2223 	int status = 0;
2224 	unsigned long flags;
2225 	struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2226 	struct ocrdma_hdr_wqe *rqe;
2227 
2228 	spin_lock_irqsave(&qp->q_lock, flags);
2229 	if (qp->state == OCRDMA_QPS_RST || qp->state == OCRDMA_QPS_ERR) {
2230 		spin_unlock_irqrestore(&qp->q_lock, flags);
2231 		*bad_wr = wr;
2232 		return -EINVAL;
2233 	}
2234 	while (wr) {
2235 		if (ocrdma_hwq_free_cnt(&qp->rq) == 0 ||
2236 		    wr->num_sge > qp->rq.max_sges) {
2237 			*bad_wr = wr;
2238 			status = -ENOMEM;
2239 			break;
2240 		}
2241 		rqe = ocrdma_hwq_head(&qp->rq);
2242 		ocrdma_build_rqe(rqe, wr, 0);
2243 
2244 		qp->rqe_wr_id_tbl[qp->rq.head] = wr->wr_id;
2245 		/* make sure rqe is written before adapter can access it */
2246 		wmb();
2247 
2248 		/* inform hw to start processing it */
2249 		ocrdma_ring_rq_db(qp);
2250 
2251 		/* update pointer, counter for next wr */
2252 		ocrdma_hwq_inc_head(&qp->rq);
2253 		wr = wr->next;
2254 	}
2255 	spin_unlock_irqrestore(&qp->q_lock, flags);
2256 	return status;
2257 }
2258 
2259 /* cqe for srq's rqe can potentially arrive out of order.
2260  * index gives the entry in the shadow table where to store
2261  * the wr_id. tag/index is returned in cqe to reference back
2262  * for a given rqe.
2263  */
2264 static int ocrdma_srq_get_idx(struct ocrdma_srq *srq)
2265 {
2266 	int row = 0;
2267 	int indx = 0;
2268 
2269 	for (row = 0; row < srq->bit_fields_len; row++) {
2270 		if (srq->idx_bit_fields[row]) {
2271 			indx = ffs(srq->idx_bit_fields[row]);
2272 			indx = (row * 32) + (indx - 1);
2273 			BUG_ON(indx >= srq->rq.max_cnt);
2274 			ocrdma_srq_toggle_bit(srq, indx);
2275 			break;
2276 		}
2277 	}
2278 
2279 	BUG_ON(row == srq->bit_fields_len);
2280 	return indx + 1; /* Use from index 1 */
2281 }
2282 
2283 static void ocrdma_ring_srq_db(struct ocrdma_srq *srq)
2284 {
2285 	u32 val = srq->rq.dbid | (1 << 16);
2286 
2287 	iowrite32(val, srq->db + OCRDMA_DB_GEN2_SRQ_OFFSET);
2288 }
2289 
2290 int ocrdma_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
2291 			 const struct ib_recv_wr **bad_wr)
2292 {
2293 	int status = 0;
2294 	unsigned long flags;
2295 	struct ocrdma_srq *srq;
2296 	struct ocrdma_hdr_wqe *rqe;
2297 	u16 tag;
2298 
2299 	srq = get_ocrdma_srq(ibsrq);
2300 
2301 	spin_lock_irqsave(&srq->q_lock, flags);
2302 	while (wr) {
2303 		if (ocrdma_hwq_free_cnt(&srq->rq) == 0 ||
2304 		    wr->num_sge > srq->rq.max_sges) {
2305 			status = -ENOMEM;
2306 			*bad_wr = wr;
2307 			break;
2308 		}
2309 		tag = ocrdma_srq_get_idx(srq);
2310 		rqe = ocrdma_hwq_head(&srq->rq);
2311 		ocrdma_build_rqe(rqe, wr, tag);
2312 
2313 		srq->rqe_wr_id_tbl[tag] = wr->wr_id;
2314 		/* make sure rqe is written before adapter can perform DMA */
2315 		wmb();
2316 		/* inform hw to start processing it */
2317 		ocrdma_ring_srq_db(srq);
2318 		/* update pointer, counter for next wr */
2319 		ocrdma_hwq_inc_head(&srq->rq);
2320 		wr = wr->next;
2321 	}
2322 	spin_unlock_irqrestore(&srq->q_lock, flags);
2323 	return status;
2324 }
2325 
2326 static enum ib_wc_status ocrdma_to_ibwc_err(u16 status)
2327 {
2328 	enum ib_wc_status ibwc_status;
2329 
2330 	switch (status) {
2331 	case OCRDMA_CQE_GENERAL_ERR:
2332 		ibwc_status = IB_WC_GENERAL_ERR;
2333 		break;
2334 	case OCRDMA_CQE_LOC_LEN_ERR:
2335 		ibwc_status = IB_WC_LOC_LEN_ERR;
2336 		break;
2337 	case OCRDMA_CQE_LOC_QP_OP_ERR:
2338 		ibwc_status = IB_WC_LOC_QP_OP_ERR;
2339 		break;
2340 	case OCRDMA_CQE_LOC_EEC_OP_ERR:
2341 		ibwc_status = IB_WC_LOC_EEC_OP_ERR;
2342 		break;
2343 	case OCRDMA_CQE_LOC_PROT_ERR:
2344 		ibwc_status = IB_WC_LOC_PROT_ERR;
2345 		break;
2346 	case OCRDMA_CQE_WR_FLUSH_ERR:
2347 		ibwc_status = IB_WC_WR_FLUSH_ERR;
2348 		break;
2349 	case OCRDMA_CQE_MW_BIND_ERR:
2350 		ibwc_status = IB_WC_MW_BIND_ERR;
2351 		break;
2352 	case OCRDMA_CQE_BAD_RESP_ERR:
2353 		ibwc_status = IB_WC_BAD_RESP_ERR;
2354 		break;
2355 	case OCRDMA_CQE_LOC_ACCESS_ERR:
2356 		ibwc_status = IB_WC_LOC_ACCESS_ERR;
2357 		break;
2358 	case OCRDMA_CQE_REM_INV_REQ_ERR:
2359 		ibwc_status = IB_WC_REM_INV_REQ_ERR;
2360 		break;
2361 	case OCRDMA_CQE_REM_ACCESS_ERR:
2362 		ibwc_status = IB_WC_REM_ACCESS_ERR;
2363 		break;
2364 	case OCRDMA_CQE_REM_OP_ERR:
2365 		ibwc_status = IB_WC_REM_OP_ERR;
2366 		break;
2367 	case OCRDMA_CQE_RETRY_EXC_ERR:
2368 		ibwc_status = IB_WC_RETRY_EXC_ERR;
2369 		break;
2370 	case OCRDMA_CQE_RNR_RETRY_EXC_ERR:
2371 		ibwc_status = IB_WC_RNR_RETRY_EXC_ERR;
2372 		break;
2373 	case OCRDMA_CQE_LOC_RDD_VIOL_ERR:
2374 		ibwc_status = IB_WC_LOC_RDD_VIOL_ERR;
2375 		break;
2376 	case OCRDMA_CQE_REM_INV_RD_REQ_ERR:
2377 		ibwc_status = IB_WC_REM_INV_RD_REQ_ERR;
2378 		break;
2379 	case OCRDMA_CQE_REM_ABORT_ERR:
2380 		ibwc_status = IB_WC_REM_ABORT_ERR;
2381 		break;
2382 	case OCRDMA_CQE_INV_EECN_ERR:
2383 		ibwc_status = IB_WC_INV_EECN_ERR;
2384 		break;
2385 	case OCRDMA_CQE_INV_EEC_STATE_ERR:
2386 		ibwc_status = IB_WC_INV_EEC_STATE_ERR;
2387 		break;
2388 	case OCRDMA_CQE_FATAL_ERR:
2389 		ibwc_status = IB_WC_FATAL_ERR;
2390 		break;
2391 	case OCRDMA_CQE_RESP_TIMEOUT_ERR:
2392 		ibwc_status = IB_WC_RESP_TIMEOUT_ERR;
2393 		break;
2394 	default:
2395 		ibwc_status = IB_WC_GENERAL_ERR;
2396 		break;
2397 	}
2398 	return ibwc_status;
2399 }
2400 
2401 static void ocrdma_update_wc(struct ocrdma_qp *qp, struct ib_wc *ibwc,
2402 		      u32 wqe_idx)
2403 {
2404 	struct ocrdma_hdr_wqe *hdr;
2405 	struct ocrdma_sge *rw;
2406 	int opcode;
2407 
2408 	hdr = ocrdma_hwq_head_from_idx(&qp->sq, wqe_idx);
2409 
2410 	ibwc->wr_id = qp->wqe_wr_id_tbl[wqe_idx].wrid;
2411 	/* Undo the hdr->cw swap */
2412 	opcode = le32_to_cpu(hdr->cw) & OCRDMA_WQE_OPCODE_MASK;
2413 	switch (opcode) {
2414 	case OCRDMA_WRITE:
2415 		ibwc->opcode = IB_WC_RDMA_WRITE;
2416 		break;
2417 	case OCRDMA_READ:
2418 		rw = (struct ocrdma_sge *)(hdr + 1);
2419 		ibwc->opcode = IB_WC_RDMA_READ;
2420 		ibwc->byte_len = rw->len;
2421 		break;
2422 	case OCRDMA_SEND:
2423 		ibwc->opcode = IB_WC_SEND;
2424 		break;
2425 	case OCRDMA_FR_MR:
2426 		ibwc->opcode = IB_WC_REG_MR;
2427 		break;
2428 	case OCRDMA_LKEY_INV:
2429 		ibwc->opcode = IB_WC_LOCAL_INV;
2430 		break;
2431 	default:
2432 		ibwc->status = IB_WC_GENERAL_ERR;
2433 		pr_err("%s() invalid opcode received = 0x%x\n",
2434 		       __func__, hdr->cw & OCRDMA_WQE_OPCODE_MASK);
2435 		break;
2436 	}
2437 }
2438 
2439 static void ocrdma_set_cqe_status_flushed(struct ocrdma_qp *qp,
2440 						struct ocrdma_cqe *cqe)
2441 {
2442 	if (is_cqe_for_sq(cqe)) {
2443 		cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2444 				cqe->flags_status_srcqpn) &
2445 					~OCRDMA_CQE_STATUS_MASK);
2446 		cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2447 				cqe->flags_status_srcqpn) |
2448 				(OCRDMA_CQE_WR_FLUSH_ERR <<
2449 					OCRDMA_CQE_STATUS_SHIFT));
2450 	} else {
2451 		if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2452 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2453 					cqe->flags_status_srcqpn) &
2454 						~OCRDMA_CQE_UD_STATUS_MASK);
2455 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2456 					cqe->flags_status_srcqpn) |
2457 					(OCRDMA_CQE_WR_FLUSH_ERR <<
2458 						OCRDMA_CQE_UD_STATUS_SHIFT));
2459 		} else {
2460 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2461 					cqe->flags_status_srcqpn) &
2462 						~OCRDMA_CQE_STATUS_MASK);
2463 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2464 					cqe->flags_status_srcqpn) |
2465 					(OCRDMA_CQE_WR_FLUSH_ERR <<
2466 						OCRDMA_CQE_STATUS_SHIFT));
2467 		}
2468 	}
2469 }
2470 
2471 static bool ocrdma_update_err_cqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2472 				  struct ocrdma_qp *qp, int status)
2473 {
2474 	bool expand = false;
2475 
2476 	ibwc->byte_len = 0;
2477 	ibwc->qp = &qp->ibqp;
2478 	ibwc->status = ocrdma_to_ibwc_err(status);
2479 
2480 	ocrdma_flush_qp(qp);
2481 	ocrdma_qp_state_change(qp, IB_QPS_ERR, NULL);
2482 
2483 	/* if wqe/rqe pending for which cqe needs to be returned,
2484 	 * trigger inflating it.
2485 	 */
2486 	if (!is_hw_rq_empty(qp) || !is_hw_sq_empty(qp)) {
2487 		expand = true;
2488 		ocrdma_set_cqe_status_flushed(qp, cqe);
2489 	}
2490 	return expand;
2491 }
2492 
2493 static int ocrdma_update_err_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2494 				  struct ocrdma_qp *qp, int status)
2495 {
2496 	ibwc->opcode = IB_WC_RECV;
2497 	ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2498 	ocrdma_hwq_inc_tail(&qp->rq);
2499 
2500 	return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2501 }
2502 
2503 static int ocrdma_update_err_scqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2504 				  struct ocrdma_qp *qp, int status)
2505 {
2506 	ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2507 	ocrdma_hwq_inc_tail(&qp->sq);
2508 
2509 	return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2510 }
2511 
2512 
2513 static bool ocrdma_poll_err_scqe(struct ocrdma_qp *qp,
2514 				 struct ocrdma_cqe *cqe, struct ib_wc *ibwc,
2515 				 bool *polled, bool *stop)
2516 {
2517 	bool expand;
2518 	struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2519 	int status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2520 		OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2521 	if (status < OCRDMA_MAX_CQE_ERR)
2522 		atomic_inc(&dev->cqe_err_stats[status]);
2523 
2524 	/* when hw sq is empty, but rq is not empty, so we continue
2525 	 * to keep the cqe in order to get the cq event again.
2526 	 */
2527 	if (is_hw_sq_empty(qp) && !is_hw_rq_empty(qp)) {
2528 		/* when cq for rq and sq is same, it is safe to return
2529 		 * flush cqe for RQEs.
2530 		 */
2531 		if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2532 			*polled = true;
2533 			status = OCRDMA_CQE_WR_FLUSH_ERR;
2534 			expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2535 		} else {
2536 			/* stop processing further cqe as this cqe is used for
2537 			 * triggering cq event on buddy cq of RQ.
2538 			 * When QP is destroyed, this cqe will be removed
2539 			 * from the cq's hardware q.
2540 			 */
2541 			*polled = false;
2542 			*stop = true;
2543 			expand = false;
2544 		}
2545 	} else if (is_hw_sq_empty(qp)) {
2546 		/* Do nothing */
2547 		expand = false;
2548 		*polled = false;
2549 		*stop = false;
2550 	} else {
2551 		*polled = true;
2552 		expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2553 	}
2554 	return expand;
2555 }
2556 
2557 static bool ocrdma_poll_success_scqe(struct ocrdma_qp *qp,
2558 				     struct ocrdma_cqe *cqe,
2559 				     struct ib_wc *ibwc, bool *polled)
2560 {
2561 	bool expand = false;
2562 	int tail = qp->sq.tail;
2563 	u32 wqe_idx;
2564 
2565 	if (!qp->wqe_wr_id_tbl[tail].signaled) {
2566 		*polled = false;    /* WC cannot be consumed yet */
2567 	} else {
2568 		ibwc->status = IB_WC_SUCCESS;
2569 		ibwc->wc_flags = 0;
2570 		ibwc->qp = &qp->ibqp;
2571 		ocrdma_update_wc(qp, ibwc, tail);
2572 		*polled = true;
2573 	}
2574 	wqe_idx = (le32_to_cpu(cqe->wq.wqeidx) &
2575 			OCRDMA_CQE_WQEIDX_MASK) & qp->sq.max_wqe_idx;
2576 	if (tail != wqe_idx)
2577 		expand = true; /* Coalesced CQE can't be consumed yet */
2578 
2579 	ocrdma_hwq_inc_tail(&qp->sq);
2580 	return expand;
2581 }
2582 
2583 static bool ocrdma_poll_scqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2584 			     struct ib_wc *ibwc, bool *polled, bool *stop)
2585 {
2586 	int status;
2587 	bool expand;
2588 
2589 	status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2590 		OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2591 
2592 	if (status == OCRDMA_CQE_SUCCESS)
2593 		expand = ocrdma_poll_success_scqe(qp, cqe, ibwc, polled);
2594 	else
2595 		expand = ocrdma_poll_err_scqe(qp, cqe, ibwc, polled, stop);
2596 	return expand;
2597 }
2598 
2599 static int ocrdma_update_ud_rcqe(struct ocrdma_dev *dev, struct ib_wc *ibwc,
2600 				 struct ocrdma_cqe *cqe)
2601 {
2602 	int status;
2603 	u16 hdr_type = 0;
2604 
2605 	status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2606 		OCRDMA_CQE_UD_STATUS_MASK) >> OCRDMA_CQE_UD_STATUS_SHIFT;
2607 	ibwc->src_qp = le32_to_cpu(cqe->flags_status_srcqpn) &
2608 						OCRDMA_CQE_SRCQP_MASK;
2609 	ibwc->pkey_index = 0;
2610 	ibwc->wc_flags = IB_WC_GRH;
2611 	ibwc->byte_len = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2612 			  OCRDMA_CQE_UD_XFER_LEN_SHIFT) &
2613 			  OCRDMA_CQE_UD_XFER_LEN_MASK;
2614 
2615 	if (ocrdma_is_udp_encap_supported(dev)) {
2616 		hdr_type = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2617 			    OCRDMA_CQE_UD_L3TYPE_SHIFT) &
2618 			    OCRDMA_CQE_UD_L3TYPE_MASK;
2619 		ibwc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
2620 		ibwc->network_hdr_type = hdr_type;
2621 	}
2622 
2623 	return status;
2624 }
2625 
2626 static void ocrdma_update_free_srq_cqe(struct ib_wc *ibwc,
2627 				       struct ocrdma_cqe *cqe,
2628 				       struct ocrdma_qp *qp)
2629 {
2630 	unsigned long flags;
2631 	struct ocrdma_srq *srq;
2632 	u32 wqe_idx;
2633 
2634 	srq = get_ocrdma_srq(qp->ibqp.srq);
2635 	wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
2636 		OCRDMA_CQE_BUFTAG_SHIFT) & srq->rq.max_wqe_idx;
2637 	BUG_ON(wqe_idx < 1);
2638 
2639 	ibwc->wr_id = srq->rqe_wr_id_tbl[wqe_idx];
2640 	spin_lock_irqsave(&srq->q_lock, flags);
2641 	ocrdma_srq_toggle_bit(srq, wqe_idx - 1);
2642 	spin_unlock_irqrestore(&srq->q_lock, flags);
2643 	ocrdma_hwq_inc_tail(&srq->rq);
2644 }
2645 
2646 static bool ocrdma_poll_err_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2647 				struct ib_wc *ibwc, bool *polled, bool *stop,
2648 				int status)
2649 {
2650 	bool expand;
2651 	struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2652 
2653 	if (status < OCRDMA_MAX_CQE_ERR)
2654 		atomic_inc(&dev->cqe_err_stats[status]);
2655 
2656 	/* when hw_rq is empty, but wq is not empty, so continue
2657 	 * to keep the cqe to get the cq event again.
2658 	 */
2659 	if (is_hw_rq_empty(qp) && !is_hw_sq_empty(qp)) {
2660 		if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2661 			*polled = true;
2662 			status = OCRDMA_CQE_WR_FLUSH_ERR;
2663 			expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2664 		} else {
2665 			*polled = false;
2666 			*stop = true;
2667 			expand = false;
2668 		}
2669 	} else if (is_hw_rq_empty(qp)) {
2670 		/* Do nothing */
2671 		expand = false;
2672 		*polled = false;
2673 		*stop = false;
2674 	} else {
2675 		*polled = true;
2676 		expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2677 	}
2678 	return expand;
2679 }
2680 
2681 static void ocrdma_poll_success_rcqe(struct ocrdma_qp *qp,
2682 				     struct ocrdma_cqe *cqe, struct ib_wc *ibwc)
2683 {
2684 	struct ocrdma_dev *dev;
2685 
2686 	dev = get_ocrdma_dev(qp->ibqp.device);
2687 	ibwc->opcode = IB_WC_RECV;
2688 	ibwc->qp = &qp->ibqp;
2689 	ibwc->status = IB_WC_SUCCESS;
2690 
2691 	if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI)
2692 		ocrdma_update_ud_rcqe(dev, ibwc, cqe);
2693 	else
2694 		ibwc->byte_len = le32_to_cpu(cqe->rq.rxlen);
2695 
2696 	if (is_cqe_imm(cqe)) {
2697 		ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2698 		ibwc->wc_flags |= IB_WC_WITH_IMM;
2699 	} else if (is_cqe_wr_imm(cqe)) {
2700 		ibwc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
2701 		ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2702 		ibwc->wc_flags |= IB_WC_WITH_IMM;
2703 	} else if (is_cqe_invalidated(cqe)) {
2704 		ibwc->ex.invalidate_rkey = le32_to_cpu(cqe->rq.lkey_immdt);
2705 		ibwc->wc_flags |= IB_WC_WITH_INVALIDATE;
2706 	}
2707 	if (qp->ibqp.srq) {
2708 		ocrdma_update_free_srq_cqe(ibwc, cqe, qp);
2709 	} else {
2710 		ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2711 		ocrdma_hwq_inc_tail(&qp->rq);
2712 	}
2713 }
2714 
2715 static bool ocrdma_poll_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2716 			     struct ib_wc *ibwc, bool *polled, bool *stop)
2717 {
2718 	int status;
2719 	bool expand = false;
2720 
2721 	ibwc->wc_flags = 0;
2722 	if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2723 		status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2724 					OCRDMA_CQE_UD_STATUS_MASK) >>
2725 					OCRDMA_CQE_UD_STATUS_SHIFT;
2726 	} else {
2727 		status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2728 			     OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2729 	}
2730 
2731 	if (status == OCRDMA_CQE_SUCCESS) {
2732 		*polled = true;
2733 		ocrdma_poll_success_rcqe(qp, cqe, ibwc);
2734 	} else {
2735 		expand = ocrdma_poll_err_rcqe(qp, cqe, ibwc, polled, stop,
2736 					      status);
2737 	}
2738 	return expand;
2739 }
2740 
2741 static void ocrdma_change_cq_phase(struct ocrdma_cq *cq, struct ocrdma_cqe *cqe,
2742 				   u16 cur_getp)
2743 {
2744 	if (cq->phase_change) {
2745 		if (cur_getp == 0)
2746 			cq->phase = (~cq->phase & OCRDMA_CQE_VALID);
2747 	} else {
2748 		/* clear valid bit */
2749 		cqe->flags_status_srcqpn = 0;
2750 	}
2751 }
2752 
2753 static int ocrdma_poll_hwcq(struct ocrdma_cq *cq, int num_entries,
2754 			    struct ib_wc *ibwc)
2755 {
2756 	u16 qpn = 0;
2757 	int i = 0;
2758 	bool expand = false;
2759 	int polled_hw_cqes = 0;
2760 	struct ocrdma_qp *qp = NULL;
2761 	struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
2762 	struct ocrdma_cqe *cqe;
2763 	u16 cur_getp; bool polled = false; bool stop = false;
2764 
2765 	cur_getp = cq->getp;
2766 	while (num_entries) {
2767 		cqe = cq->va + cur_getp;
2768 		/* check whether valid cqe or not */
2769 		if (!is_cqe_valid(cq, cqe))
2770 			break;
2771 		qpn = (le32_to_cpu(cqe->cmn.qpn) & OCRDMA_CQE_QPN_MASK);
2772 		/* ignore discarded cqe */
2773 		if (qpn == 0)
2774 			goto skip_cqe;
2775 		qp = dev->qp_tbl[qpn];
2776 		BUG_ON(qp == NULL);
2777 
2778 		if (is_cqe_for_sq(cqe)) {
2779 			expand = ocrdma_poll_scqe(qp, cqe, ibwc, &polled,
2780 						  &stop);
2781 		} else {
2782 			expand = ocrdma_poll_rcqe(qp, cqe, ibwc, &polled,
2783 						  &stop);
2784 		}
2785 		if (expand)
2786 			goto expand_cqe;
2787 		if (stop)
2788 			goto stop_cqe;
2789 		/* clear qpn to avoid duplicate processing by discard_cqe() */
2790 		cqe->cmn.qpn = 0;
2791 skip_cqe:
2792 		polled_hw_cqes += 1;
2793 		cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
2794 		ocrdma_change_cq_phase(cq, cqe, cur_getp);
2795 expand_cqe:
2796 		if (polled) {
2797 			num_entries -= 1;
2798 			i += 1;
2799 			ibwc = ibwc + 1;
2800 			polled = false;
2801 		}
2802 	}
2803 stop_cqe:
2804 	cq->getp = cur_getp;
2805 
2806 	if (polled_hw_cqes)
2807 		ocrdma_ring_cq_db(dev, cq->id, false, false, polled_hw_cqes);
2808 
2809 	return i;
2810 }
2811 
2812 /* insert error cqe if the QP's SQ or RQ's CQ matches the CQ under poll. */
2813 static int ocrdma_add_err_cqe(struct ocrdma_cq *cq, int num_entries,
2814 			      struct ocrdma_qp *qp, struct ib_wc *ibwc)
2815 {
2816 	int err_cqes = 0;
2817 
2818 	while (num_entries) {
2819 		if (is_hw_sq_empty(qp) && is_hw_rq_empty(qp))
2820 			break;
2821 		if (!is_hw_sq_empty(qp) && qp->sq_cq == cq) {
2822 			ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2823 			ocrdma_hwq_inc_tail(&qp->sq);
2824 		} else if (!is_hw_rq_empty(qp) && qp->rq_cq == cq) {
2825 			ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2826 			ocrdma_hwq_inc_tail(&qp->rq);
2827 		} else {
2828 			return err_cqes;
2829 		}
2830 		ibwc->byte_len = 0;
2831 		ibwc->status = IB_WC_WR_FLUSH_ERR;
2832 		ibwc = ibwc + 1;
2833 		err_cqes += 1;
2834 		num_entries -= 1;
2835 	}
2836 	return err_cqes;
2837 }
2838 
2839 int ocrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
2840 {
2841 	int cqes_to_poll = num_entries;
2842 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2843 	struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2844 	int num_os_cqe = 0, err_cqes = 0;
2845 	struct ocrdma_qp *qp;
2846 	unsigned long flags;
2847 
2848 	/* poll cqes from adapter CQ */
2849 	spin_lock_irqsave(&cq->cq_lock, flags);
2850 	num_os_cqe = ocrdma_poll_hwcq(cq, cqes_to_poll, wc);
2851 	spin_unlock_irqrestore(&cq->cq_lock, flags);
2852 	cqes_to_poll -= num_os_cqe;
2853 
2854 	if (cqes_to_poll) {
2855 		wc = wc + num_os_cqe;
2856 		/* adapter returns single error cqe when qp moves to
2857 		 * error state. So insert error cqes with wc_status as
2858 		 * FLUSHED for pending WQEs and RQEs of QP's SQ and RQ
2859 		 * respectively which uses this CQ.
2860 		 */
2861 		spin_lock_irqsave(&dev->flush_q_lock, flags);
2862 		list_for_each_entry(qp, &cq->sq_head, sq_entry) {
2863 			if (cqes_to_poll == 0)
2864 				break;
2865 			err_cqes = ocrdma_add_err_cqe(cq, cqes_to_poll, qp, wc);
2866 			cqes_to_poll -= err_cqes;
2867 			num_os_cqe += err_cqes;
2868 			wc = wc + err_cqes;
2869 		}
2870 		spin_unlock_irqrestore(&dev->flush_q_lock, flags);
2871 	}
2872 	return num_os_cqe;
2873 }
2874 
2875 int ocrdma_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags cq_flags)
2876 {
2877 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2878 	struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2879 	u16 cq_id;
2880 	unsigned long flags;
2881 	bool arm_needed = false, sol_needed = false;
2882 
2883 	cq_id = cq->id;
2884 
2885 	spin_lock_irqsave(&cq->cq_lock, flags);
2886 	if (cq_flags & IB_CQ_NEXT_COMP || cq_flags & IB_CQ_SOLICITED)
2887 		arm_needed = true;
2888 	if (cq_flags & IB_CQ_SOLICITED)
2889 		sol_needed = true;
2890 
2891 	ocrdma_ring_cq_db(dev, cq_id, arm_needed, sol_needed, 0);
2892 	spin_unlock_irqrestore(&cq->cq_lock, flags);
2893 
2894 	return 0;
2895 }
2896 
2897 struct ib_mr *ocrdma_alloc_mr(struct ib_pd *ibpd, enum ib_mr_type mr_type,
2898 			      u32 max_num_sg)
2899 {
2900 	int status;
2901 	struct ocrdma_mr *mr;
2902 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
2903 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
2904 
2905 	if (mr_type != IB_MR_TYPE_MEM_REG)
2906 		return ERR_PTR(-EINVAL);
2907 
2908 	if (max_num_sg > dev->attr.max_pages_per_frmr)
2909 		return ERR_PTR(-EINVAL);
2910 
2911 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
2912 	if (!mr)
2913 		return ERR_PTR(-ENOMEM);
2914 
2915 	mr->pages = kcalloc(max_num_sg, sizeof(u64), GFP_KERNEL);
2916 	if (!mr->pages) {
2917 		status = -ENOMEM;
2918 		goto pl_err;
2919 	}
2920 
2921 	status = ocrdma_get_pbl_info(dev, mr, max_num_sg);
2922 	if (status)
2923 		goto pbl_err;
2924 	mr->hwmr.fr_mr = 1;
2925 	mr->hwmr.remote_rd = 0;
2926 	mr->hwmr.remote_wr = 0;
2927 	mr->hwmr.local_rd = 0;
2928 	mr->hwmr.local_wr = 0;
2929 	mr->hwmr.mw_bind = 0;
2930 	status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
2931 	if (status)
2932 		goto pbl_err;
2933 	status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, 0);
2934 	if (status)
2935 		goto mbx_err;
2936 	mr->ibmr.rkey = mr->hwmr.lkey;
2937 	mr->ibmr.lkey = mr->hwmr.lkey;
2938 	dev->stag_arr[(mr->hwmr.lkey >> 8) & (OCRDMA_MAX_STAG - 1)] =
2939 		(unsigned long) mr;
2940 	return &mr->ibmr;
2941 mbx_err:
2942 	ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
2943 pbl_err:
2944 	kfree(mr->pages);
2945 pl_err:
2946 	kfree(mr);
2947 	return ERR_PTR(-ENOMEM);
2948 }
2949 
2950 static int ocrdma_set_page(struct ib_mr *ibmr, u64 addr)
2951 {
2952 	struct ocrdma_mr *mr = get_ocrdma_mr(ibmr);
2953 
2954 	if (unlikely(mr->npages == mr->hwmr.num_pbes))
2955 		return -ENOMEM;
2956 
2957 	mr->pages[mr->npages++] = addr;
2958 
2959 	return 0;
2960 }
2961 
2962 int ocrdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
2963 		     unsigned int *sg_offset)
2964 {
2965 	struct ocrdma_mr *mr = get_ocrdma_mr(ibmr);
2966 
2967 	mr->npages = 0;
2968 
2969 	return ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, ocrdma_set_page);
2970 }
2971