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 
51 #include "ocrdma.h"
52 #include "ocrdma_hw.h"
53 #include "ocrdma_verbs.h"
54 #include <rdma/ocrdma-abi.h>
55 
56 int ocrdma_query_pkey(struct ib_device *ibdev, u8 port, u16 index, u16 *pkey)
57 {
58 	if (index > 1)
59 		return -EINVAL;
60 
61 	*pkey = 0xffff;
62 	return 0;
63 }
64 
65 int ocrdma_query_device(struct ib_device *ibdev, struct ib_device_attr *attr,
66 			struct ib_udata *uhw)
67 {
68 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
69 
70 	if (uhw->inlen || uhw->outlen)
71 		return -EINVAL;
72 
73 	memset(attr, 0, sizeof *attr);
74 	memcpy(&attr->fw_ver, &dev->attr.fw_ver[0],
75 	       min(sizeof(dev->attr.fw_ver), sizeof(attr->fw_ver)));
76 	ocrdma_get_guid(dev, (u8 *)&attr->sys_image_guid);
77 	attr->max_mr_size = dev->attr.max_mr_size;
78 	attr->page_size_cap = 0xffff000;
79 	attr->vendor_id = dev->nic_info.pdev->vendor;
80 	attr->vendor_part_id = dev->nic_info.pdev->device;
81 	attr->hw_ver = dev->asic_id;
82 	attr->max_qp = dev->attr.max_qp;
83 	attr->max_ah = OCRDMA_MAX_AH;
84 	attr->max_qp_wr = dev->attr.max_wqe;
85 
86 	attr->device_cap_flags = IB_DEVICE_CURR_QP_STATE_MOD |
87 					IB_DEVICE_RC_RNR_NAK_GEN |
88 					IB_DEVICE_SHUTDOWN_PORT |
89 					IB_DEVICE_SYS_IMAGE_GUID |
90 					IB_DEVICE_LOCAL_DMA_LKEY |
91 					IB_DEVICE_MEM_MGT_EXTENSIONS;
92 	attr->max_send_sge = dev->attr.max_send_sge;
93 	attr->max_recv_sge = dev->attr.max_recv_sge;
94 	attr->max_sge_rd = dev->attr.max_rdma_sge;
95 	attr->max_cq = dev->attr.max_cq;
96 	attr->max_cqe = dev->attr.max_cqe;
97 	attr->max_mr = dev->attr.max_mr;
98 	attr->max_mw = dev->attr.max_mw;
99 	attr->max_pd = dev->attr.max_pd;
100 	attr->atomic_cap = 0;
101 	attr->max_fmr = 0;
102 	attr->max_map_per_fmr = 0;
103 	attr->max_qp_rd_atom =
104 	    min(dev->attr.max_ord_per_qp, dev->attr.max_ird_per_qp);
105 	attr->max_qp_init_rd_atom = dev->attr.max_ord_per_qp;
106 	attr->max_srq = dev->attr.max_srq;
107 	attr->max_srq_sge = dev->attr.max_srq_sge;
108 	attr->max_srq_wr = dev->attr.max_rqe;
109 	attr->local_ca_ack_delay = dev->attr.local_ca_ack_delay;
110 	attr->max_fast_reg_page_list_len = dev->attr.max_pages_per_frmr;
111 	attr->max_pkeys = 1;
112 	return 0;
113 }
114 
115 struct net_device *ocrdma_get_netdev(struct ib_device *ibdev, u8 port_num)
116 {
117 	struct ocrdma_dev *dev;
118 	struct net_device *ndev = NULL;
119 
120 	rcu_read_lock();
121 
122 	dev = get_ocrdma_dev(ibdev);
123 	if (dev)
124 		ndev = dev->nic_info.netdev;
125 	if (ndev)
126 		dev_hold(ndev);
127 
128 	rcu_read_unlock();
129 
130 	return ndev;
131 }
132 
133 static inline void get_link_speed_and_width(struct ocrdma_dev *dev,
134 					    u8 *ib_speed, u8 *ib_width)
135 {
136 	int status;
137 	u8 speed;
138 
139 	status = ocrdma_mbx_get_link_speed(dev, &speed, NULL);
140 	if (status)
141 		speed = OCRDMA_PHYS_LINK_SPEED_ZERO;
142 
143 	switch (speed) {
144 	case OCRDMA_PHYS_LINK_SPEED_1GBPS:
145 		*ib_speed = IB_SPEED_SDR;
146 		*ib_width = IB_WIDTH_1X;
147 		break;
148 
149 	case OCRDMA_PHYS_LINK_SPEED_10GBPS:
150 		*ib_speed = IB_SPEED_QDR;
151 		*ib_width = IB_WIDTH_1X;
152 		break;
153 
154 	case OCRDMA_PHYS_LINK_SPEED_20GBPS:
155 		*ib_speed = IB_SPEED_DDR;
156 		*ib_width = IB_WIDTH_4X;
157 		break;
158 
159 	case OCRDMA_PHYS_LINK_SPEED_40GBPS:
160 		*ib_speed = IB_SPEED_QDR;
161 		*ib_width = IB_WIDTH_4X;
162 		break;
163 
164 	default:
165 		/* Unsupported */
166 		*ib_speed = IB_SPEED_SDR;
167 		*ib_width = IB_WIDTH_1X;
168 	}
169 }
170 
171 int ocrdma_query_port(struct ib_device *ibdev,
172 		      u8 port, struct ib_port_attr *props)
173 {
174 	enum ib_port_state port_state;
175 	struct ocrdma_dev *dev;
176 	struct net_device *netdev;
177 
178 	/* props being zeroed by the caller, avoid zeroing it here */
179 	dev = get_ocrdma_dev(ibdev);
180 	netdev = dev->nic_info.netdev;
181 	if (netif_running(netdev) && netif_oper_up(netdev)) {
182 		port_state = IB_PORT_ACTIVE;
183 		props->phys_state = 5;
184 	} else {
185 		port_state = IB_PORT_DOWN;
186 		props->phys_state = 3;
187 	}
188 	props->max_mtu = IB_MTU_4096;
189 	props->active_mtu = iboe_get_mtu(netdev->mtu);
190 	props->lid = 0;
191 	props->lmc = 0;
192 	props->sm_lid = 0;
193 	props->sm_sl = 0;
194 	props->state = port_state;
195 	props->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_REINIT_SUP |
196 				IB_PORT_DEVICE_MGMT_SUP |
197 				IB_PORT_VENDOR_CLASS_SUP;
198 	props->ip_gids = true;
199 	props->gid_tbl_len = OCRDMA_MAX_SGID;
200 	props->pkey_tbl_len = 1;
201 	props->bad_pkey_cntr = 0;
202 	props->qkey_viol_cntr = 0;
203 	get_link_speed_and_width(dev, &props->active_speed,
204 				 &props->active_width);
205 	props->max_msg_sz = 0x80000000;
206 	props->max_vl_num = 4;
207 	return 0;
208 }
209 
210 int ocrdma_modify_port(struct ib_device *ibdev, u8 port, int mask,
211 		       struct ib_port_modify *props)
212 {
213 	return 0;
214 }
215 
216 static int ocrdma_add_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
217 			   unsigned long len)
218 {
219 	struct ocrdma_mm *mm;
220 
221 	mm = kzalloc(sizeof(*mm), GFP_KERNEL);
222 	if (mm == NULL)
223 		return -ENOMEM;
224 	mm->key.phy_addr = phy_addr;
225 	mm->key.len = len;
226 	INIT_LIST_HEAD(&mm->entry);
227 
228 	mutex_lock(&uctx->mm_list_lock);
229 	list_add_tail(&mm->entry, &uctx->mm_head);
230 	mutex_unlock(&uctx->mm_list_lock);
231 	return 0;
232 }
233 
234 static void ocrdma_del_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
235 			    unsigned long len)
236 {
237 	struct ocrdma_mm *mm, *tmp;
238 
239 	mutex_lock(&uctx->mm_list_lock);
240 	list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
241 		if (len != mm->key.len && phy_addr != mm->key.phy_addr)
242 			continue;
243 
244 		list_del(&mm->entry);
245 		kfree(mm);
246 		break;
247 	}
248 	mutex_unlock(&uctx->mm_list_lock);
249 }
250 
251 static bool ocrdma_search_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
252 			      unsigned long len)
253 {
254 	bool found = false;
255 	struct ocrdma_mm *mm;
256 
257 	mutex_lock(&uctx->mm_list_lock);
258 	list_for_each_entry(mm, &uctx->mm_head, entry) {
259 		if (len != mm->key.len && phy_addr != mm->key.phy_addr)
260 			continue;
261 
262 		found = true;
263 		break;
264 	}
265 	mutex_unlock(&uctx->mm_list_lock);
266 	return found;
267 }
268 
269 
270 static u16 _ocrdma_pd_mgr_get_bitmap(struct ocrdma_dev *dev, bool dpp_pool)
271 {
272 	u16 pd_bitmap_idx = 0;
273 	const unsigned long *pd_bitmap;
274 
275 	if (dpp_pool) {
276 		pd_bitmap = dev->pd_mgr->pd_dpp_bitmap;
277 		pd_bitmap_idx = find_first_zero_bit(pd_bitmap,
278 						    dev->pd_mgr->max_dpp_pd);
279 		__set_bit(pd_bitmap_idx, dev->pd_mgr->pd_dpp_bitmap);
280 		dev->pd_mgr->pd_dpp_count++;
281 		if (dev->pd_mgr->pd_dpp_count > dev->pd_mgr->pd_dpp_thrsh)
282 			dev->pd_mgr->pd_dpp_thrsh = dev->pd_mgr->pd_dpp_count;
283 	} else {
284 		pd_bitmap = dev->pd_mgr->pd_norm_bitmap;
285 		pd_bitmap_idx = find_first_zero_bit(pd_bitmap,
286 						    dev->pd_mgr->max_normal_pd);
287 		__set_bit(pd_bitmap_idx, dev->pd_mgr->pd_norm_bitmap);
288 		dev->pd_mgr->pd_norm_count++;
289 		if (dev->pd_mgr->pd_norm_count > dev->pd_mgr->pd_norm_thrsh)
290 			dev->pd_mgr->pd_norm_thrsh = dev->pd_mgr->pd_norm_count;
291 	}
292 	return pd_bitmap_idx;
293 }
294 
295 static int _ocrdma_pd_mgr_put_bitmap(struct ocrdma_dev *dev, u16 pd_id,
296 					bool dpp_pool)
297 {
298 	u16 pd_count;
299 	u16 pd_bit_index;
300 
301 	pd_count = dpp_pool ? dev->pd_mgr->pd_dpp_count :
302 			      dev->pd_mgr->pd_norm_count;
303 	if (pd_count == 0)
304 		return -EINVAL;
305 
306 	if (dpp_pool) {
307 		pd_bit_index = pd_id - dev->pd_mgr->pd_dpp_start;
308 		if (pd_bit_index >= dev->pd_mgr->max_dpp_pd) {
309 			return -EINVAL;
310 		} else {
311 			__clear_bit(pd_bit_index, dev->pd_mgr->pd_dpp_bitmap);
312 			dev->pd_mgr->pd_dpp_count--;
313 		}
314 	} else {
315 		pd_bit_index = pd_id - dev->pd_mgr->pd_norm_start;
316 		if (pd_bit_index >= dev->pd_mgr->max_normal_pd) {
317 			return -EINVAL;
318 		} else {
319 			__clear_bit(pd_bit_index, dev->pd_mgr->pd_norm_bitmap);
320 			dev->pd_mgr->pd_norm_count--;
321 		}
322 	}
323 
324 	return 0;
325 }
326 
327 static int ocrdma_put_pd_num(struct ocrdma_dev *dev, u16 pd_id,
328 				   bool dpp_pool)
329 {
330 	int status;
331 
332 	mutex_lock(&dev->dev_lock);
333 	status = _ocrdma_pd_mgr_put_bitmap(dev, pd_id, dpp_pool);
334 	mutex_unlock(&dev->dev_lock);
335 	return status;
336 }
337 
338 static int ocrdma_get_pd_num(struct ocrdma_dev *dev, struct ocrdma_pd *pd)
339 {
340 	u16 pd_idx = 0;
341 	int status = 0;
342 
343 	mutex_lock(&dev->dev_lock);
344 	if (pd->dpp_enabled) {
345 		/* try allocating DPP PD, if not available then normal PD */
346 		if (dev->pd_mgr->pd_dpp_count < dev->pd_mgr->max_dpp_pd) {
347 			pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, true);
348 			pd->id = dev->pd_mgr->pd_dpp_start + pd_idx;
349 			pd->dpp_page = dev->pd_mgr->dpp_page_index + pd_idx;
350 		} else if (dev->pd_mgr->pd_norm_count <
351 			   dev->pd_mgr->max_normal_pd) {
352 			pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, false);
353 			pd->id = dev->pd_mgr->pd_norm_start + pd_idx;
354 			pd->dpp_enabled = false;
355 		} else {
356 			status = -EINVAL;
357 		}
358 	} else {
359 		if (dev->pd_mgr->pd_norm_count < dev->pd_mgr->max_normal_pd) {
360 			pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, false);
361 			pd->id = dev->pd_mgr->pd_norm_start + pd_idx;
362 		} else {
363 			status = -EINVAL;
364 		}
365 	}
366 	mutex_unlock(&dev->dev_lock);
367 	return status;
368 }
369 
370 static struct ocrdma_pd *_ocrdma_alloc_pd(struct ocrdma_dev *dev,
371 					  struct ocrdma_ucontext *uctx,
372 					  struct ib_udata *udata)
373 {
374 	struct ocrdma_pd *pd = NULL;
375 	int status;
376 
377 	pd = kzalloc(sizeof(*pd), GFP_KERNEL);
378 	if (!pd)
379 		return ERR_PTR(-ENOMEM);
380 
381 	if (udata && uctx && dev->attr.max_dpp_pds) {
382 		pd->dpp_enabled =
383 			ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R;
384 		pd->num_dpp_qp =
385 			pd->dpp_enabled ? (dev->nic_info.db_page_size /
386 					   dev->attr.wqe_size) : 0;
387 	}
388 
389 	if (dev->pd_mgr->pd_prealloc_valid) {
390 		status = ocrdma_get_pd_num(dev, pd);
391 		if (status == 0) {
392 			return pd;
393 		} else {
394 			kfree(pd);
395 			return ERR_PTR(status);
396 		}
397 	}
398 
399 retry:
400 	status = ocrdma_mbx_alloc_pd(dev, pd);
401 	if (status) {
402 		if (pd->dpp_enabled) {
403 			pd->dpp_enabled = false;
404 			pd->num_dpp_qp = 0;
405 			goto retry;
406 		} else {
407 			kfree(pd);
408 			return ERR_PTR(status);
409 		}
410 	}
411 
412 	return pd;
413 }
414 
415 static inline int is_ucontext_pd(struct ocrdma_ucontext *uctx,
416 				 struct ocrdma_pd *pd)
417 {
418 	return (uctx->cntxt_pd == pd);
419 }
420 
421 static int _ocrdma_dealloc_pd(struct ocrdma_dev *dev,
422 			      struct ocrdma_pd *pd)
423 {
424 	int status;
425 
426 	if (dev->pd_mgr->pd_prealloc_valid)
427 		status = ocrdma_put_pd_num(dev, pd->id, pd->dpp_enabled);
428 	else
429 		status = ocrdma_mbx_dealloc_pd(dev, pd);
430 
431 	kfree(pd);
432 	return status;
433 }
434 
435 static int ocrdma_alloc_ucontext_pd(struct ocrdma_dev *dev,
436 				    struct ocrdma_ucontext *uctx,
437 				    struct ib_udata *udata)
438 {
439 	int status = 0;
440 
441 	uctx->cntxt_pd = _ocrdma_alloc_pd(dev, uctx, udata);
442 	if (IS_ERR(uctx->cntxt_pd)) {
443 		status = PTR_ERR(uctx->cntxt_pd);
444 		uctx->cntxt_pd = NULL;
445 		goto err;
446 	}
447 
448 	uctx->cntxt_pd->uctx = uctx;
449 	uctx->cntxt_pd->ibpd.device = &dev->ibdev;
450 err:
451 	return status;
452 }
453 
454 static int ocrdma_dealloc_ucontext_pd(struct ocrdma_ucontext *uctx)
455 {
456 	struct ocrdma_pd *pd = uctx->cntxt_pd;
457 	struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
458 
459 	if (uctx->pd_in_use) {
460 		pr_err("%s(%d) Freeing in use pdid=0x%x.\n",
461 		       __func__, dev->id, pd->id);
462 	}
463 	uctx->cntxt_pd = NULL;
464 	(void)_ocrdma_dealloc_pd(dev, pd);
465 	return 0;
466 }
467 
468 static struct ocrdma_pd *ocrdma_get_ucontext_pd(struct ocrdma_ucontext *uctx)
469 {
470 	struct ocrdma_pd *pd = NULL;
471 
472 	mutex_lock(&uctx->mm_list_lock);
473 	if (!uctx->pd_in_use) {
474 		uctx->pd_in_use = true;
475 		pd = uctx->cntxt_pd;
476 	}
477 	mutex_unlock(&uctx->mm_list_lock);
478 
479 	return pd;
480 }
481 
482 static void ocrdma_release_ucontext_pd(struct ocrdma_ucontext *uctx)
483 {
484 	mutex_lock(&uctx->mm_list_lock);
485 	uctx->pd_in_use = false;
486 	mutex_unlock(&uctx->mm_list_lock);
487 }
488 
489 struct ib_ucontext *ocrdma_alloc_ucontext(struct ib_device *ibdev,
490 					  struct ib_udata *udata)
491 {
492 	int status;
493 	struct ocrdma_ucontext *ctx;
494 	struct ocrdma_alloc_ucontext_resp resp;
495 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
496 	struct pci_dev *pdev = dev->nic_info.pdev;
497 	u32 map_len = roundup(sizeof(u32) * 2048, PAGE_SIZE);
498 
499 	if (!udata)
500 		return ERR_PTR(-EFAULT);
501 	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
502 	if (!ctx)
503 		return ERR_PTR(-ENOMEM);
504 	INIT_LIST_HEAD(&ctx->mm_head);
505 	mutex_init(&ctx->mm_list_lock);
506 
507 	ctx->ah_tbl.va = dma_zalloc_coherent(&pdev->dev, map_len,
508 					     &ctx->ah_tbl.pa, GFP_KERNEL);
509 	if (!ctx->ah_tbl.va) {
510 		kfree(ctx);
511 		return ERR_PTR(-ENOMEM);
512 	}
513 	ctx->ah_tbl.len = map_len;
514 
515 	memset(&resp, 0, sizeof(resp));
516 	resp.ah_tbl_len = ctx->ah_tbl.len;
517 	resp.ah_tbl_page = virt_to_phys(ctx->ah_tbl.va);
518 
519 	status = ocrdma_add_mmap(ctx, resp.ah_tbl_page, resp.ah_tbl_len);
520 	if (status)
521 		goto map_err;
522 
523 	status = ocrdma_alloc_ucontext_pd(dev, ctx, udata);
524 	if (status)
525 		goto pd_err;
526 
527 	resp.dev_id = dev->id;
528 	resp.max_inline_data = dev->attr.max_inline_data;
529 	resp.wqe_size = dev->attr.wqe_size;
530 	resp.rqe_size = dev->attr.rqe_size;
531 	resp.dpp_wqe_size = dev->attr.wqe_size;
532 
533 	memcpy(resp.fw_ver, dev->attr.fw_ver, sizeof(resp.fw_ver));
534 	status = ib_copy_to_udata(udata, &resp, sizeof(resp));
535 	if (status)
536 		goto cpy_err;
537 	return &ctx->ibucontext;
538 
539 cpy_err:
540 pd_err:
541 	ocrdma_del_mmap(ctx, ctx->ah_tbl.pa, ctx->ah_tbl.len);
542 map_err:
543 	dma_free_coherent(&pdev->dev, ctx->ah_tbl.len, ctx->ah_tbl.va,
544 			  ctx->ah_tbl.pa);
545 	kfree(ctx);
546 	return ERR_PTR(status);
547 }
548 
549 int ocrdma_dealloc_ucontext(struct ib_ucontext *ibctx)
550 {
551 	int status;
552 	struct ocrdma_mm *mm, *tmp;
553 	struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ibctx);
554 	struct ocrdma_dev *dev = get_ocrdma_dev(ibctx->device);
555 	struct pci_dev *pdev = dev->nic_info.pdev;
556 
557 	status = ocrdma_dealloc_ucontext_pd(uctx);
558 
559 	ocrdma_del_mmap(uctx, uctx->ah_tbl.pa, uctx->ah_tbl.len);
560 	dma_free_coherent(&pdev->dev, uctx->ah_tbl.len, uctx->ah_tbl.va,
561 			  uctx->ah_tbl.pa);
562 
563 	list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
564 		list_del(&mm->entry);
565 		kfree(mm);
566 	}
567 	kfree(uctx);
568 	return status;
569 }
570 
571 int ocrdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
572 {
573 	struct ocrdma_ucontext *ucontext = get_ocrdma_ucontext(context);
574 	struct ocrdma_dev *dev = get_ocrdma_dev(context->device);
575 	unsigned long vm_page = vma->vm_pgoff << PAGE_SHIFT;
576 	u64 unmapped_db = (u64) dev->nic_info.unmapped_db;
577 	unsigned long len = (vma->vm_end - vma->vm_start);
578 	int status;
579 	bool found;
580 
581 	if (vma->vm_start & (PAGE_SIZE - 1))
582 		return -EINVAL;
583 	found = ocrdma_search_mmap(ucontext, vma->vm_pgoff << PAGE_SHIFT, len);
584 	if (!found)
585 		return -EINVAL;
586 
587 	if ((vm_page >= unmapped_db) && (vm_page <= (unmapped_db +
588 		dev->nic_info.db_total_size)) &&
589 		(len <=	dev->nic_info.db_page_size)) {
590 		if (vma->vm_flags & VM_READ)
591 			return -EPERM;
592 
593 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
594 		status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
595 					    len, vma->vm_page_prot);
596 	} else if (dev->nic_info.dpp_unmapped_len &&
597 		(vm_page >= (u64) dev->nic_info.dpp_unmapped_addr) &&
598 		(vm_page <= (u64) (dev->nic_info.dpp_unmapped_addr +
599 			dev->nic_info.dpp_unmapped_len)) &&
600 		(len <= dev->nic_info.dpp_unmapped_len)) {
601 		if (vma->vm_flags & VM_READ)
602 			return -EPERM;
603 
604 		vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
605 		status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
606 					    len, vma->vm_page_prot);
607 	} else {
608 		status = remap_pfn_range(vma, vma->vm_start,
609 					 vma->vm_pgoff, len, vma->vm_page_prot);
610 	}
611 	return status;
612 }
613 
614 static int ocrdma_copy_pd_uresp(struct ocrdma_dev *dev, struct ocrdma_pd *pd,
615 				struct ib_ucontext *ib_ctx,
616 				struct ib_udata *udata)
617 {
618 	int status;
619 	u64 db_page_addr;
620 	u64 dpp_page_addr = 0;
621 	u32 db_page_size;
622 	struct ocrdma_alloc_pd_uresp rsp;
623 	struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ib_ctx);
624 
625 	memset(&rsp, 0, sizeof(rsp));
626 	rsp.id = pd->id;
627 	rsp.dpp_enabled = pd->dpp_enabled;
628 	db_page_addr = ocrdma_get_db_addr(dev, pd->id);
629 	db_page_size = dev->nic_info.db_page_size;
630 
631 	status = ocrdma_add_mmap(uctx, db_page_addr, db_page_size);
632 	if (status)
633 		return status;
634 
635 	if (pd->dpp_enabled) {
636 		dpp_page_addr = dev->nic_info.dpp_unmapped_addr +
637 				(pd->id * PAGE_SIZE);
638 		status = ocrdma_add_mmap(uctx, dpp_page_addr,
639 				 PAGE_SIZE);
640 		if (status)
641 			goto dpp_map_err;
642 		rsp.dpp_page_addr_hi = upper_32_bits(dpp_page_addr);
643 		rsp.dpp_page_addr_lo = dpp_page_addr;
644 	}
645 
646 	status = ib_copy_to_udata(udata, &rsp, sizeof(rsp));
647 	if (status)
648 		goto ucopy_err;
649 
650 	pd->uctx = uctx;
651 	return 0;
652 
653 ucopy_err:
654 	if (pd->dpp_enabled)
655 		ocrdma_del_mmap(pd->uctx, dpp_page_addr, PAGE_SIZE);
656 dpp_map_err:
657 	ocrdma_del_mmap(pd->uctx, db_page_addr, db_page_size);
658 	return status;
659 }
660 
661 struct ib_pd *ocrdma_alloc_pd(struct ib_device *ibdev,
662 			      struct ib_ucontext *context,
663 			      struct ib_udata *udata)
664 {
665 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
666 	struct ocrdma_pd *pd;
667 	struct ocrdma_ucontext *uctx = NULL;
668 	int status;
669 	u8 is_uctx_pd = false;
670 
671 	if (udata && context) {
672 		uctx = get_ocrdma_ucontext(context);
673 		pd = ocrdma_get_ucontext_pd(uctx);
674 		if (pd) {
675 			is_uctx_pd = true;
676 			goto pd_mapping;
677 		}
678 	}
679 
680 	pd = _ocrdma_alloc_pd(dev, uctx, udata);
681 	if (IS_ERR(pd)) {
682 		status = PTR_ERR(pd);
683 		goto exit;
684 	}
685 
686 pd_mapping:
687 	if (udata && context) {
688 		status = ocrdma_copy_pd_uresp(dev, pd, context, udata);
689 		if (status)
690 			goto err;
691 	}
692 	return &pd->ibpd;
693 
694 err:
695 	if (is_uctx_pd) {
696 		ocrdma_release_ucontext_pd(uctx);
697 	} else {
698 		if (_ocrdma_dealloc_pd(dev, pd))
699 			pr_err("%s: _ocrdma_dealloc_pd() failed\n", __func__);
700 	}
701 exit:
702 	return ERR_PTR(status);
703 }
704 
705 int ocrdma_dealloc_pd(struct ib_pd *ibpd)
706 {
707 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
708 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
709 	struct ocrdma_ucontext *uctx = NULL;
710 	int status = 0;
711 	u64 usr_db;
712 
713 	uctx = pd->uctx;
714 	if (uctx) {
715 		u64 dpp_db = dev->nic_info.dpp_unmapped_addr +
716 			(pd->id * PAGE_SIZE);
717 		if (pd->dpp_enabled)
718 			ocrdma_del_mmap(pd->uctx, dpp_db, PAGE_SIZE);
719 		usr_db = ocrdma_get_db_addr(dev, pd->id);
720 		ocrdma_del_mmap(pd->uctx, usr_db, dev->nic_info.db_page_size);
721 
722 		if (is_ucontext_pd(uctx, pd)) {
723 			ocrdma_release_ucontext_pd(uctx);
724 			return status;
725 		}
726 	}
727 	status = _ocrdma_dealloc_pd(dev, pd);
728 	return status;
729 }
730 
731 static int ocrdma_alloc_lkey(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
732 			    u32 pdid, int acc, u32 num_pbls, u32 addr_check)
733 {
734 	int status;
735 
736 	mr->hwmr.fr_mr = 0;
737 	mr->hwmr.local_rd = 1;
738 	mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
739 	mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
740 	mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
741 	mr->hwmr.mw_bind = (acc & IB_ACCESS_MW_BIND) ? 1 : 0;
742 	mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
743 	mr->hwmr.num_pbls = num_pbls;
744 
745 	status = ocrdma_mbx_alloc_lkey(dev, &mr->hwmr, pdid, addr_check);
746 	if (status)
747 		return status;
748 
749 	mr->ibmr.lkey = mr->hwmr.lkey;
750 	if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
751 		mr->ibmr.rkey = mr->hwmr.lkey;
752 	return 0;
753 }
754 
755 struct ib_mr *ocrdma_get_dma_mr(struct ib_pd *ibpd, int acc)
756 {
757 	int status;
758 	struct ocrdma_mr *mr;
759 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
760 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
761 
762 	if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE)) {
763 		pr_err("%s err, invalid access rights\n", __func__);
764 		return ERR_PTR(-EINVAL);
765 	}
766 
767 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
768 	if (!mr)
769 		return ERR_PTR(-ENOMEM);
770 
771 	status = ocrdma_alloc_lkey(dev, mr, pd->id, acc, 0,
772 				   OCRDMA_ADDR_CHECK_DISABLE);
773 	if (status) {
774 		kfree(mr);
775 		return ERR_PTR(status);
776 	}
777 
778 	return &mr->ibmr;
779 }
780 
781 static void ocrdma_free_mr_pbl_tbl(struct ocrdma_dev *dev,
782 				   struct ocrdma_hw_mr *mr)
783 {
784 	struct pci_dev *pdev = dev->nic_info.pdev;
785 	int i = 0;
786 
787 	if (mr->pbl_table) {
788 		for (i = 0; i < mr->num_pbls; i++) {
789 			if (!mr->pbl_table[i].va)
790 				continue;
791 			dma_free_coherent(&pdev->dev, mr->pbl_size,
792 					  mr->pbl_table[i].va,
793 					  mr->pbl_table[i].pa);
794 		}
795 		kfree(mr->pbl_table);
796 		mr->pbl_table = NULL;
797 	}
798 }
799 
800 static int ocrdma_get_pbl_info(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
801 			      u32 num_pbes)
802 {
803 	u32 num_pbls = 0;
804 	u32 idx = 0;
805 	int status = 0;
806 	u32 pbl_size;
807 
808 	do {
809 		pbl_size = OCRDMA_MIN_HPAGE_SIZE * (1 << idx);
810 		if (pbl_size > MAX_OCRDMA_PBL_SIZE) {
811 			status = -EFAULT;
812 			break;
813 		}
814 		num_pbls = roundup(num_pbes, (pbl_size / sizeof(u64)));
815 		num_pbls = num_pbls / (pbl_size / sizeof(u64));
816 		idx++;
817 	} while (num_pbls >= dev->attr.max_num_mr_pbl);
818 
819 	mr->hwmr.num_pbes = num_pbes;
820 	mr->hwmr.num_pbls = num_pbls;
821 	mr->hwmr.pbl_size = pbl_size;
822 	return status;
823 }
824 
825 static int ocrdma_build_pbl_tbl(struct ocrdma_dev *dev, struct ocrdma_hw_mr *mr)
826 {
827 	int status = 0;
828 	int i;
829 	u32 dma_len = mr->pbl_size;
830 	struct pci_dev *pdev = dev->nic_info.pdev;
831 	void *va;
832 	dma_addr_t pa;
833 
834 	mr->pbl_table = kcalloc(mr->num_pbls, sizeof(struct ocrdma_pbl),
835 				GFP_KERNEL);
836 
837 	if (!mr->pbl_table)
838 		return -ENOMEM;
839 
840 	for (i = 0; i < mr->num_pbls; i++) {
841 		va = dma_zalloc_coherent(&pdev->dev, dma_len, &pa, GFP_KERNEL);
842 		if (!va) {
843 			ocrdma_free_mr_pbl_tbl(dev, mr);
844 			status = -ENOMEM;
845 			break;
846 		}
847 		mr->pbl_table[i].va = va;
848 		mr->pbl_table[i].pa = pa;
849 	}
850 	return status;
851 }
852 
853 static void build_user_pbes(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
854 			    u32 num_pbes)
855 {
856 	struct ocrdma_pbe *pbe;
857 	struct scatterlist *sg;
858 	struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
859 	struct ib_umem *umem = mr->umem;
860 	int shift, pg_cnt, pages, pbe_cnt, entry, total_num_pbes = 0;
861 
862 	if (!mr->hwmr.num_pbes)
863 		return;
864 
865 	pbe = (struct ocrdma_pbe *)pbl_tbl->va;
866 	pbe_cnt = 0;
867 
868 	shift = umem->page_shift;
869 
870 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
871 		pages = sg_dma_len(sg) >> shift;
872 		for (pg_cnt = 0; pg_cnt < pages; pg_cnt++) {
873 			/* store the page address in pbe */
874 			pbe->pa_lo =
875 			    cpu_to_le32(sg_dma_address(sg) +
876 					(pg_cnt << shift));
877 			pbe->pa_hi =
878 			    cpu_to_le32(upper_32_bits(sg_dma_address(sg) +
879 					 (pg_cnt << shift)));
880 			pbe_cnt += 1;
881 			total_num_pbes += 1;
882 			pbe++;
883 
884 			/* if done building pbes, issue the mbx cmd. */
885 			if (total_num_pbes == num_pbes)
886 				return;
887 
888 			/* if the given pbl is full storing the pbes,
889 			 * move to next pbl.
890 			 */
891 			if (pbe_cnt ==
892 				(mr->hwmr.pbl_size / sizeof(u64))) {
893 				pbl_tbl++;
894 				pbe = (struct ocrdma_pbe *)pbl_tbl->va;
895 				pbe_cnt = 0;
896 			}
897 
898 		}
899 	}
900 }
901 
902 struct ib_mr *ocrdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
903 				 u64 usr_addr, int acc, struct ib_udata *udata)
904 {
905 	int status = -ENOMEM;
906 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
907 	struct ocrdma_mr *mr;
908 	struct ocrdma_pd *pd;
909 	u32 num_pbes;
910 
911 	pd = get_ocrdma_pd(ibpd);
912 
913 	if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE))
914 		return ERR_PTR(-EINVAL);
915 
916 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
917 	if (!mr)
918 		return ERR_PTR(status);
919 	mr->umem = ib_umem_get(ibpd->uobject->context, start, len, acc, 0);
920 	if (IS_ERR(mr->umem)) {
921 		status = -EFAULT;
922 		goto umem_err;
923 	}
924 	num_pbes = ib_umem_page_count(mr->umem);
925 	status = ocrdma_get_pbl_info(dev, mr, num_pbes);
926 	if (status)
927 		goto umem_err;
928 
929 	mr->hwmr.pbe_size = BIT(mr->umem->page_shift);
930 	mr->hwmr.fbo = ib_umem_offset(mr->umem);
931 	mr->hwmr.va = usr_addr;
932 	mr->hwmr.len = len;
933 	mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
934 	mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
935 	mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
936 	mr->hwmr.local_rd = 1;
937 	mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
938 	status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
939 	if (status)
940 		goto umem_err;
941 	build_user_pbes(dev, mr, num_pbes);
942 	status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
943 	if (status)
944 		goto mbx_err;
945 	mr->ibmr.lkey = mr->hwmr.lkey;
946 	if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
947 		mr->ibmr.rkey = mr->hwmr.lkey;
948 
949 	return &mr->ibmr;
950 
951 mbx_err:
952 	ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
953 umem_err:
954 	kfree(mr);
955 	return ERR_PTR(status);
956 }
957 
958 int ocrdma_dereg_mr(struct ib_mr *ib_mr)
959 {
960 	struct ocrdma_mr *mr = get_ocrdma_mr(ib_mr);
961 	struct ocrdma_dev *dev = get_ocrdma_dev(ib_mr->device);
962 
963 	(void) ocrdma_mbx_dealloc_lkey(dev, mr->hwmr.fr_mr, mr->hwmr.lkey);
964 
965 	kfree(mr->pages);
966 	ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
967 
968 	/* it could be user registered memory. */
969 	if (mr->umem)
970 		ib_umem_release(mr->umem);
971 	kfree(mr);
972 
973 	/* Don't stop cleanup, in case FW is unresponsive */
974 	if (dev->mqe_ctx.fw_error_state) {
975 		pr_err("%s(%d) fw not responding.\n",
976 		       __func__, dev->id);
977 	}
978 	return 0;
979 }
980 
981 static int ocrdma_copy_cq_uresp(struct ocrdma_dev *dev, struct ocrdma_cq *cq,
982 				struct ib_udata *udata,
983 				struct ib_ucontext *ib_ctx)
984 {
985 	int status;
986 	struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ib_ctx);
987 	struct ocrdma_create_cq_uresp uresp;
988 
989 	memset(&uresp, 0, sizeof(uresp));
990 	uresp.cq_id = cq->id;
991 	uresp.page_size = PAGE_ALIGN(cq->len);
992 	uresp.num_pages = 1;
993 	uresp.max_hw_cqe = cq->max_hw_cqe;
994 	uresp.page_addr[0] = virt_to_phys(cq->va);
995 	uresp.db_page_addr =  ocrdma_get_db_addr(dev, uctx->cntxt_pd->id);
996 	uresp.db_page_size = dev->nic_info.db_page_size;
997 	uresp.phase_change = cq->phase_change ? 1 : 0;
998 	status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
999 	if (status) {
1000 		pr_err("%s(%d) copy error cqid=0x%x.\n",
1001 		       __func__, dev->id, cq->id);
1002 		goto err;
1003 	}
1004 	status = ocrdma_add_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
1005 	if (status)
1006 		goto err;
1007 	status = ocrdma_add_mmap(uctx, uresp.page_addr[0], uresp.page_size);
1008 	if (status) {
1009 		ocrdma_del_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
1010 		goto err;
1011 	}
1012 	cq->ucontext = uctx;
1013 err:
1014 	return status;
1015 }
1016 
1017 struct ib_cq *ocrdma_create_cq(struct ib_device *ibdev,
1018 			       const struct ib_cq_init_attr *attr,
1019 			       struct ib_ucontext *ib_ctx,
1020 			       struct ib_udata *udata)
1021 {
1022 	int entries = attr->cqe;
1023 	struct ocrdma_cq *cq;
1024 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
1025 	struct ocrdma_ucontext *uctx = NULL;
1026 	u16 pd_id = 0;
1027 	int status;
1028 	struct ocrdma_create_cq_ureq ureq;
1029 
1030 	if (attr->flags)
1031 		return ERR_PTR(-EINVAL);
1032 
1033 	if (udata) {
1034 		if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1035 			return ERR_PTR(-EFAULT);
1036 	} else
1037 		ureq.dpp_cq = 0;
1038 	cq = kzalloc(sizeof(*cq), GFP_KERNEL);
1039 	if (!cq)
1040 		return ERR_PTR(-ENOMEM);
1041 
1042 	spin_lock_init(&cq->cq_lock);
1043 	spin_lock_init(&cq->comp_handler_lock);
1044 	INIT_LIST_HEAD(&cq->sq_head);
1045 	INIT_LIST_HEAD(&cq->rq_head);
1046 
1047 	if (ib_ctx) {
1048 		uctx = get_ocrdma_ucontext(ib_ctx);
1049 		pd_id = uctx->cntxt_pd->id;
1050 	}
1051 
1052 	status = ocrdma_mbx_create_cq(dev, cq, entries, ureq.dpp_cq, pd_id);
1053 	if (status) {
1054 		kfree(cq);
1055 		return ERR_PTR(status);
1056 	}
1057 	if (ib_ctx) {
1058 		status = ocrdma_copy_cq_uresp(dev, cq, udata, ib_ctx);
1059 		if (status)
1060 			goto ctx_err;
1061 	}
1062 	cq->phase = OCRDMA_CQE_VALID;
1063 	dev->cq_tbl[cq->id] = cq;
1064 	return &cq->ibcq;
1065 
1066 ctx_err:
1067 	ocrdma_mbx_destroy_cq(dev, cq);
1068 	kfree(cq);
1069 	return ERR_PTR(status);
1070 }
1071 
1072 int ocrdma_resize_cq(struct ib_cq *ibcq, int new_cnt,
1073 		     struct ib_udata *udata)
1074 {
1075 	int status = 0;
1076 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
1077 
1078 	if (new_cnt < 1 || new_cnt > cq->max_hw_cqe) {
1079 		status = -EINVAL;
1080 		return status;
1081 	}
1082 	ibcq->cqe = new_cnt;
1083 	return status;
1084 }
1085 
1086 static void ocrdma_flush_cq(struct ocrdma_cq *cq)
1087 {
1088 	int cqe_cnt;
1089 	int valid_count = 0;
1090 	unsigned long flags;
1091 
1092 	struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
1093 	struct ocrdma_cqe *cqe = NULL;
1094 
1095 	cqe = cq->va;
1096 	cqe_cnt = cq->cqe_cnt;
1097 
1098 	/* Last irq might have scheduled a polling thread
1099 	 * sync-up with it before hard flushing.
1100 	 */
1101 	spin_lock_irqsave(&cq->cq_lock, flags);
1102 	while (cqe_cnt) {
1103 		if (is_cqe_valid(cq, cqe))
1104 			valid_count++;
1105 		cqe++;
1106 		cqe_cnt--;
1107 	}
1108 	ocrdma_ring_cq_db(dev, cq->id, false, false, valid_count);
1109 	spin_unlock_irqrestore(&cq->cq_lock, flags);
1110 }
1111 
1112 int ocrdma_destroy_cq(struct ib_cq *ibcq)
1113 {
1114 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
1115 	struct ocrdma_eq *eq = NULL;
1116 	struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
1117 	int pdid = 0;
1118 	u32 irq, indx;
1119 
1120 	dev->cq_tbl[cq->id] = NULL;
1121 	indx = ocrdma_get_eq_table_index(dev, cq->eqn);
1122 	BUG_ON(indx == -EINVAL);
1123 
1124 	eq = &dev->eq_tbl[indx];
1125 	irq = ocrdma_get_irq(dev, eq);
1126 	synchronize_irq(irq);
1127 	ocrdma_flush_cq(cq);
1128 
1129 	(void)ocrdma_mbx_destroy_cq(dev, cq);
1130 	if (cq->ucontext) {
1131 		pdid = cq->ucontext->cntxt_pd->id;
1132 		ocrdma_del_mmap(cq->ucontext, (u64) cq->pa,
1133 				PAGE_ALIGN(cq->len));
1134 		ocrdma_del_mmap(cq->ucontext,
1135 				ocrdma_get_db_addr(dev, pdid),
1136 				dev->nic_info.db_page_size);
1137 	}
1138 
1139 	kfree(cq);
1140 	return 0;
1141 }
1142 
1143 static int ocrdma_add_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1144 {
1145 	int status = -EINVAL;
1146 
1147 	if (qp->id < OCRDMA_MAX_QP && dev->qp_tbl[qp->id] == NULL) {
1148 		dev->qp_tbl[qp->id] = qp;
1149 		status = 0;
1150 	}
1151 	return status;
1152 }
1153 
1154 static void ocrdma_del_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1155 {
1156 	dev->qp_tbl[qp->id] = NULL;
1157 }
1158 
1159 static int ocrdma_check_qp_params(struct ib_pd *ibpd, struct ocrdma_dev *dev,
1160 				  struct ib_qp_init_attr *attrs,
1161 				  struct ib_udata *udata)
1162 {
1163 	if ((attrs->qp_type != IB_QPT_GSI) &&
1164 	    (attrs->qp_type != IB_QPT_RC) &&
1165 	    (attrs->qp_type != IB_QPT_UC) &&
1166 	    (attrs->qp_type != IB_QPT_UD)) {
1167 		pr_err("%s(%d) unsupported qp type=0x%x requested\n",
1168 		       __func__, dev->id, attrs->qp_type);
1169 		return -EINVAL;
1170 	}
1171 	/* Skip the check for QP1 to support CM size of 128 */
1172 	if ((attrs->qp_type != IB_QPT_GSI) &&
1173 	    (attrs->cap.max_send_wr > dev->attr.max_wqe)) {
1174 		pr_err("%s(%d) unsupported send_wr=0x%x requested\n",
1175 		       __func__, dev->id, attrs->cap.max_send_wr);
1176 		pr_err("%s(%d) supported send_wr=0x%x\n",
1177 		       __func__, dev->id, dev->attr.max_wqe);
1178 		return -EINVAL;
1179 	}
1180 	if (!attrs->srq && (attrs->cap.max_recv_wr > dev->attr.max_rqe)) {
1181 		pr_err("%s(%d) unsupported recv_wr=0x%x requested\n",
1182 		       __func__, dev->id, attrs->cap.max_recv_wr);
1183 		pr_err("%s(%d) supported recv_wr=0x%x\n",
1184 		       __func__, dev->id, dev->attr.max_rqe);
1185 		return -EINVAL;
1186 	}
1187 	if (attrs->cap.max_inline_data > dev->attr.max_inline_data) {
1188 		pr_err("%s(%d) unsupported inline data size=0x%x requested\n",
1189 		       __func__, dev->id, attrs->cap.max_inline_data);
1190 		pr_err("%s(%d) supported inline data size=0x%x\n",
1191 		       __func__, dev->id, dev->attr.max_inline_data);
1192 		return -EINVAL;
1193 	}
1194 	if (attrs->cap.max_send_sge > dev->attr.max_send_sge) {
1195 		pr_err("%s(%d) unsupported send_sge=0x%x requested\n",
1196 		       __func__, dev->id, attrs->cap.max_send_sge);
1197 		pr_err("%s(%d) supported send_sge=0x%x\n",
1198 		       __func__, dev->id, dev->attr.max_send_sge);
1199 		return -EINVAL;
1200 	}
1201 	if (attrs->cap.max_recv_sge > dev->attr.max_recv_sge) {
1202 		pr_err("%s(%d) unsupported recv_sge=0x%x requested\n",
1203 		       __func__, dev->id, attrs->cap.max_recv_sge);
1204 		pr_err("%s(%d) supported recv_sge=0x%x\n",
1205 		       __func__, dev->id, dev->attr.max_recv_sge);
1206 		return -EINVAL;
1207 	}
1208 	/* unprivileged user space cannot create special QP */
1209 	if (udata && attrs->qp_type == IB_QPT_GSI) {
1210 		pr_err
1211 		    ("%s(%d) Userspace can't create special QPs of type=0x%x\n",
1212 		     __func__, dev->id, attrs->qp_type);
1213 		return -EINVAL;
1214 	}
1215 	/* allow creating only one GSI type of QP */
1216 	if (attrs->qp_type == IB_QPT_GSI && dev->gsi_qp_created) {
1217 		pr_err("%s(%d) GSI special QPs already created.\n",
1218 		       __func__, dev->id);
1219 		return -EINVAL;
1220 	}
1221 	/* verify consumer QPs are not trying to use GSI QP's CQ */
1222 	if ((attrs->qp_type != IB_QPT_GSI) && (dev->gsi_qp_created)) {
1223 		if ((dev->gsi_sqcq == get_ocrdma_cq(attrs->send_cq)) ||
1224 			(dev->gsi_rqcq == get_ocrdma_cq(attrs->recv_cq))) {
1225 			pr_err("%s(%d) Consumer QP cannot use GSI CQs.\n",
1226 				__func__, dev->id);
1227 			return -EINVAL;
1228 		}
1229 	}
1230 	return 0;
1231 }
1232 
1233 static int ocrdma_copy_qp_uresp(struct ocrdma_qp *qp,
1234 				struct ib_udata *udata, int dpp_offset,
1235 				int dpp_credit_lmt, int srq)
1236 {
1237 	int status;
1238 	u64 usr_db;
1239 	struct ocrdma_create_qp_uresp uresp;
1240 	struct ocrdma_pd *pd = qp->pd;
1241 	struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
1242 
1243 	memset(&uresp, 0, sizeof(uresp));
1244 	usr_db = dev->nic_info.unmapped_db +
1245 			(pd->id * dev->nic_info.db_page_size);
1246 	uresp.qp_id = qp->id;
1247 	uresp.sq_dbid = qp->sq.dbid;
1248 	uresp.num_sq_pages = 1;
1249 	uresp.sq_page_size = PAGE_ALIGN(qp->sq.len);
1250 	uresp.sq_page_addr[0] = virt_to_phys(qp->sq.va);
1251 	uresp.num_wqe_allocated = qp->sq.max_cnt;
1252 	if (!srq) {
1253 		uresp.rq_dbid = qp->rq.dbid;
1254 		uresp.num_rq_pages = 1;
1255 		uresp.rq_page_size = PAGE_ALIGN(qp->rq.len);
1256 		uresp.rq_page_addr[0] = virt_to_phys(qp->rq.va);
1257 		uresp.num_rqe_allocated = qp->rq.max_cnt;
1258 	}
1259 	uresp.db_page_addr = usr_db;
1260 	uresp.db_page_size = dev->nic_info.db_page_size;
1261 	uresp.db_sq_offset = OCRDMA_DB_GEN2_SQ_OFFSET;
1262 	uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1263 	uresp.db_shift = OCRDMA_DB_RQ_SHIFT;
1264 
1265 	if (qp->dpp_enabled) {
1266 		uresp.dpp_credit = dpp_credit_lmt;
1267 		uresp.dpp_offset = dpp_offset;
1268 	}
1269 	status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1270 	if (status) {
1271 		pr_err("%s(%d) user copy error.\n", __func__, dev->id);
1272 		goto err;
1273 	}
1274 	status = ocrdma_add_mmap(pd->uctx, uresp.sq_page_addr[0],
1275 				 uresp.sq_page_size);
1276 	if (status)
1277 		goto err;
1278 
1279 	if (!srq) {
1280 		status = ocrdma_add_mmap(pd->uctx, uresp.rq_page_addr[0],
1281 					 uresp.rq_page_size);
1282 		if (status)
1283 			goto rq_map_err;
1284 	}
1285 	return status;
1286 rq_map_err:
1287 	ocrdma_del_mmap(pd->uctx, uresp.sq_page_addr[0], uresp.sq_page_size);
1288 err:
1289 	return status;
1290 }
1291 
1292 static void ocrdma_set_qp_db(struct ocrdma_dev *dev, struct ocrdma_qp *qp,
1293 			     struct ocrdma_pd *pd)
1294 {
1295 	if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1296 		qp->sq_db = dev->nic_info.db +
1297 			(pd->id * dev->nic_info.db_page_size) +
1298 			OCRDMA_DB_GEN2_SQ_OFFSET;
1299 		qp->rq_db = dev->nic_info.db +
1300 			(pd->id * dev->nic_info.db_page_size) +
1301 			OCRDMA_DB_GEN2_RQ_OFFSET;
1302 	} else {
1303 		qp->sq_db = dev->nic_info.db +
1304 			(pd->id * dev->nic_info.db_page_size) +
1305 			OCRDMA_DB_SQ_OFFSET;
1306 		qp->rq_db = dev->nic_info.db +
1307 			(pd->id * dev->nic_info.db_page_size) +
1308 			OCRDMA_DB_RQ_OFFSET;
1309 	}
1310 }
1311 
1312 static int ocrdma_alloc_wr_id_tbl(struct ocrdma_qp *qp)
1313 {
1314 	qp->wqe_wr_id_tbl =
1315 	    kcalloc(qp->sq.max_cnt, sizeof(*(qp->wqe_wr_id_tbl)),
1316 		    GFP_KERNEL);
1317 	if (qp->wqe_wr_id_tbl == NULL)
1318 		return -ENOMEM;
1319 	qp->rqe_wr_id_tbl =
1320 	    kcalloc(qp->rq.max_cnt, sizeof(u64), GFP_KERNEL);
1321 	if (qp->rqe_wr_id_tbl == NULL)
1322 		return -ENOMEM;
1323 
1324 	return 0;
1325 }
1326 
1327 static void ocrdma_set_qp_init_params(struct ocrdma_qp *qp,
1328 				      struct ocrdma_pd *pd,
1329 				      struct ib_qp_init_attr *attrs)
1330 {
1331 	qp->pd = pd;
1332 	spin_lock_init(&qp->q_lock);
1333 	INIT_LIST_HEAD(&qp->sq_entry);
1334 	INIT_LIST_HEAD(&qp->rq_entry);
1335 
1336 	qp->qp_type = attrs->qp_type;
1337 	qp->cap_flags = OCRDMA_QP_INB_RD | OCRDMA_QP_INB_WR;
1338 	qp->max_inline_data = attrs->cap.max_inline_data;
1339 	qp->sq.max_sges = attrs->cap.max_send_sge;
1340 	qp->rq.max_sges = attrs->cap.max_recv_sge;
1341 	qp->state = OCRDMA_QPS_RST;
1342 	qp->signaled = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR) ? true : false;
1343 }
1344 
1345 static void ocrdma_store_gsi_qp_cq(struct ocrdma_dev *dev,
1346 				   struct ib_qp_init_attr *attrs)
1347 {
1348 	if (attrs->qp_type == IB_QPT_GSI) {
1349 		dev->gsi_qp_created = 1;
1350 		dev->gsi_sqcq = get_ocrdma_cq(attrs->send_cq);
1351 		dev->gsi_rqcq = get_ocrdma_cq(attrs->recv_cq);
1352 	}
1353 }
1354 
1355 struct ib_qp *ocrdma_create_qp(struct ib_pd *ibpd,
1356 			       struct ib_qp_init_attr *attrs,
1357 			       struct ib_udata *udata)
1358 {
1359 	int status;
1360 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1361 	struct ocrdma_qp *qp;
1362 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1363 	struct ocrdma_create_qp_ureq ureq;
1364 	u16 dpp_credit_lmt, dpp_offset;
1365 
1366 	status = ocrdma_check_qp_params(ibpd, dev, attrs, udata);
1367 	if (status)
1368 		goto gen_err;
1369 
1370 	memset(&ureq, 0, sizeof(ureq));
1371 	if (udata) {
1372 		if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1373 			return ERR_PTR(-EFAULT);
1374 	}
1375 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1376 	if (!qp) {
1377 		status = -ENOMEM;
1378 		goto gen_err;
1379 	}
1380 	ocrdma_set_qp_init_params(qp, pd, attrs);
1381 	if (udata == NULL)
1382 		qp->cap_flags |= (OCRDMA_QP_MW_BIND | OCRDMA_QP_LKEY0 |
1383 					OCRDMA_QP_FAST_REG);
1384 
1385 	mutex_lock(&dev->dev_lock);
1386 	status = ocrdma_mbx_create_qp(qp, attrs, ureq.enable_dpp_cq,
1387 					ureq.dpp_cq_id,
1388 					&dpp_offset, &dpp_credit_lmt);
1389 	if (status)
1390 		goto mbx_err;
1391 
1392 	/* user space QP's wr_id table are managed in library */
1393 	if (udata == NULL) {
1394 		status = ocrdma_alloc_wr_id_tbl(qp);
1395 		if (status)
1396 			goto map_err;
1397 	}
1398 
1399 	status = ocrdma_add_qpn_map(dev, qp);
1400 	if (status)
1401 		goto map_err;
1402 	ocrdma_set_qp_db(dev, qp, pd);
1403 	if (udata) {
1404 		status = ocrdma_copy_qp_uresp(qp, udata, dpp_offset,
1405 					      dpp_credit_lmt,
1406 					      (attrs->srq != NULL));
1407 		if (status)
1408 			goto cpy_err;
1409 	}
1410 	ocrdma_store_gsi_qp_cq(dev, attrs);
1411 	qp->ibqp.qp_num = qp->id;
1412 	mutex_unlock(&dev->dev_lock);
1413 	return &qp->ibqp;
1414 
1415 cpy_err:
1416 	ocrdma_del_qpn_map(dev, qp);
1417 map_err:
1418 	ocrdma_mbx_destroy_qp(dev, qp);
1419 mbx_err:
1420 	mutex_unlock(&dev->dev_lock);
1421 	kfree(qp->wqe_wr_id_tbl);
1422 	kfree(qp->rqe_wr_id_tbl);
1423 	kfree(qp);
1424 	pr_err("%s(%d) error=%d\n", __func__, dev->id, status);
1425 gen_err:
1426 	return ERR_PTR(status);
1427 }
1428 
1429 int _ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1430 		      int attr_mask)
1431 {
1432 	int status = 0;
1433 	struct ocrdma_qp *qp;
1434 	struct ocrdma_dev *dev;
1435 	enum ib_qp_state old_qps;
1436 
1437 	qp = get_ocrdma_qp(ibqp);
1438 	dev = get_ocrdma_dev(ibqp->device);
1439 	if (attr_mask & IB_QP_STATE)
1440 		status = ocrdma_qp_state_change(qp, attr->qp_state, &old_qps);
1441 	/* if new and previous states are same hw doesn't need to
1442 	 * know about it.
1443 	 */
1444 	if (status < 0)
1445 		return status;
1446 	return ocrdma_mbx_modify_qp(dev, qp, attr, attr_mask);
1447 }
1448 
1449 int ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1450 		     int attr_mask, struct ib_udata *udata)
1451 {
1452 	unsigned long flags;
1453 	int status = -EINVAL;
1454 	struct ocrdma_qp *qp;
1455 	struct ocrdma_dev *dev;
1456 	enum ib_qp_state old_qps, new_qps;
1457 
1458 	qp = get_ocrdma_qp(ibqp);
1459 	dev = get_ocrdma_dev(ibqp->device);
1460 
1461 	/* syncronize with multiple context trying to change, retrive qps */
1462 	mutex_lock(&dev->dev_lock);
1463 	/* syncronize with wqe, rqe posting and cqe processing contexts */
1464 	spin_lock_irqsave(&qp->q_lock, flags);
1465 	old_qps = get_ibqp_state(qp->state);
1466 	if (attr_mask & IB_QP_STATE)
1467 		new_qps = attr->qp_state;
1468 	else
1469 		new_qps = old_qps;
1470 	spin_unlock_irqrestore(&qp->q_lock, flags);
1471 
1472 	if (!ib_modify_qp_is_ok(old_qps, new_qps, ibqp->qp_type, attr_mask)) {
1473 		pr_err("%s(%d) invalid attribute mask=0x%x specified for\n"
1474 		       "qpn=0x%x of type=0x%x old_qps=0x%x, new_qps=0x%x\n",
1475 		       __func__, dev->id, attr_mask, qp->id, ibqp->qp_type,
1476 		       old_qps, new_qps);
1477 		goto param_err;
1478 	}
1479 
1480 	status = _ocrdma_modify_qp(ibqp, attr, attr_mask);
1481 	if (status > 0)
1482 		status = 0;
1483 param_err:
1484 	mutex_unlock(&dev->dev_lock);
1485 	return status;
1486 }
1487 
1488 static enum ib_mtu ocrdma_mtu_int_to_enum(u16 mtu)
1489 {
1490 	switch (mtu) {
1491 	case 256:
1492 		return IB_MTU_256;
1493 	case 512:
1494 		return IB_MTU_512;
1495 	case 1024:
1496 		return IB_MTU_1024;
1497 	case 2048:
1498 		return IB_MTU_2048;
1499 	case 4096:
1500 		return IB_MTU_4096;
1501 	default:
1502 		return IB_MTU_1024;
1503 	}
1504 }
1505 
1506 static int ocrdma_to_ib_qp_acc_flags(int qp_cap_flags)
1507 {
1508 	int ib_qp_acc_flags = 0;
1509 
1510 	if (qp_cap_flags & OCRDMA_QP_INB_WR)
1511 		ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE;
1512 	if (qp_cap_flags & OCRDMA_QP_INB_RD)
1513 		ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE;
1514 	return ib_qp_acc_flags;
1515 }
1516 
1517 int ocrdma_query_qp(struct ib_qp *ibqp,
1518 		    struct ib_qp_attr *qp_attr,
1519 		    int attr_mask, struct ib_qp_init_attr *qp_init_attr)
1520 {
1521 	int status;
1522 	u32 qp_state;
1523 	struct ocrdma_qp_params params;
1524 	struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
1525 	struct ocrdma_dev *dev = get_ocrdma_dev(ibqp->device);
1526 
1527 	memset(&params, 0, sizeof(params));
1528 	mutex_lock(&dev->dev_lock);
1529 	status = ocrdma_mbx_query_qp(dev, qp, &params);
1530 	mutex_unlock(&dev->dev_lock);
1531 	if (status)
1532 		goto mbx_err;
1533 	if (qp->qp_type == IB_QPT_UD)
1534 		qp_attr->qkey = params.qkey;
1535 	qp_attr->path_mtu =
1536 		ocrdma_mtu_int_to_enum(params.path_mtu_pkey_indx &
1537 				OCRDMA_QP_PARAMS_PATH_MTU_MASK) >>
1538 				OCRDMA_QP_PARAMS_PATH_MTU_SHIFT;
1539 	qp_attr->path_mig_state = IB_MIG_MIGRATED;
1540 	qp_attr->rq_psn = params.hop_lmt_rq_psn & OCRDMA_QP_PARAMS_RQ_PSN_MASK;
1541 	qp_attr->sq_psn = params.tclass_sq_psn & OCRDMA_QP_PARAMS_SQ_PSN_MASK;
1542 	qp_attr->dest_qp_num =
1543 	    params.ack_to_rnr_rtc_dest_qpn & OCRDMA_QP_PARAMS_DEST_QPN_MASK;
1544 
1545 	qp_attr->qp_access_flags = ocrdma_to_ib_qp_acc_flags(qp->cap_flags);
1546 	qp_attr->cap.max_send_wr = qp->sq.max_cnt - 1;
1547 	qp_attr->cap.max_recv_wr = qp->rq.max_cnt - 1;
1548 	qp_attr->cap.max_send_sge = qp->sq.max_sges;
1549 	qp_attr->cap.max_recv_sge = qp->rq.max_sges;
1550 	qp_attr->cap.max_inline_data = qp->max_inline_data;
1551 	qp_init_attr->cap = qp_attr->cap;
1552 	qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
1553 
1554 	rdma_ah_set_grh(&qp_attr->ah_attr, NULL,
1555 			params.rnt_rc_sl_fl &
1556 			  OCRDMA_QP_PARAMS_FLOW_LABEL_MASK,
1557 			qp->sgid_idx,
1558 			(params.hop_lmt_rq_psn &
1559 			 OCRDMA_QP_PARAMS_HOP_LMT_MASK) >>
1560 			 OCRDMA_QP_PARAMS_HOP_LMT_SHIFT,
1561 			(params.tclass_sq_psn &
1562 			 OCRDMA_QP_PARAMS_TCLASS_MASK) >>
1563 			 OCRDMA_QP_PARAMS_TCLASS_SHIFT);
1564 	rdma_ah_set_dgid_raw(&qp_attr->ah_attr, &params.dgid[0]);
1565 
1566 	rdma_ah_set_port_num(&qp_attr->ah_attr, 1);
1567 	rdma_ah_set_sl(&qp_attr->ah_attr, (params.rnt_rc_sl_fl &
1568 					   OCRDMA_QP_PARAMS_SL_MASK) >>
1569 					   OCRDMA_QP_PARAMS_SL_SHIFT);
1570 	qp_attr->timeout = (params.ack_to_rnr_rtc_dest_qpn &
1571 			    OCRDMA_QP_PARAMS_ACK_TIMEOUT_MASK) >>
1572 				OCRDMA_QP_PARAMS_ACK_TIMEOUT_SHIFT;
1573 	qp_attr->rnr_retry = (params.ack_to_rnr_rtc_dest_qpn &
1574 			      OCRDMA_QP_PARAMS_RNR_RETRY_CNT_MASK) >>
1575 				OCRDMA_QP_PARAMS_RNR_RETRY_CNT_SHIFT;
1576 	qp_attr->retry_cnt =
1577 	    (params.rnt_rc_sl_fl & OCRDMA_QP_PARAMS_RETRY_CNT_MASK) >>
1578 		OCRDMA_QP_PARAMS_RETRY_CNT_SHIFT;
1579 	qp_attr->min_rnr_timer = 0;
1580 	qp_attr->pkey_index = 0;
1581 	qp_attr->port_num = 1;
1582 	rdma_ah_set_path_bits(&qp_attr->ah_attr, 0);
1583 	rdma_ah_set_static_rate(&qp_attr->ah_attr, 0);
1584 	qp_attr->alt_pkey_index = 0;
1585 	qp_attr->alt_port_num = 0;
1586 	qp_attr->alt_timeout = 0;
1587 	memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr));
1588 	qp_state = (params.max_sge_recv_flags & OCRDMA_QP_PARAMS_STATE_MASK) >>
1589 		    OCRDMA_QP_PARAMS_STATE_SHIFT;
1590 	qp_attr->qp_state = get_ibqp_state(qp_state);
1591 	qp_attr->cur_qp_state = qp_attr->qp_state;
1592 	qp_attr->sq_draining = (qp_state == OCRDMA_QPS_SQ_DRAINING) ? 1 : 0;
1593 	qp_attr->max_dest_rd_atomic =
1594 	    params.max_ord_ird >> OCRDMA_QP_PARAMS_MAX_ORD_SHIFT;
1595 	qp_attr->max_rd_atomic =
1596 	    params.max_ord_ird & OCRDMA_QP_PARAMS_MAX_IRD_MASK;
1597 	qp_attr->en_sqd_async_notify = (params.max_sge_recv_flags &
1598 				OCRDMA_QP_PARAMS_FLAGS_SQD_ASYNC) ? 1 : 0;
1599 	/* Sync driver QP state with FW */
1600 	ocrdma_qp_state_change(qp, qp_attr->qp_state, NULL);
1601 mbx_err:
1602 	return status;
1603 }
1604 
1605 static void ocrdma_srq_toggle_bit(struct ocrdma_srq *srq, unsigned int idx)
1606 {
1607 	unsigned int i = idx / 32;
1608 	u32 mask = (1U << (idx % 32));
1609 
1610 	srq->idx_bit_fields[i] ^= mask;
1611 }
1612 
1613 static int ocrdma_hwq_free_cnt(struct ocrdma_qp_hwq_info *q)
1614 {
1615 	return ((q->max_wqe_idx - q->head) + q->tail) % q->max_cnt;
1616 }
1617 
1618 static int is_hw_sq_empty(struct ocrdma_qp *qp)
1619 {
1620 	return (qp->sq.tail == qp->sq.head);
1621 }
1622 
1623 static int is_hw_rq_empty(struct ocrdma_qp *qp)
1624 {
1625 	return (qp->rq.tail == qp->rq.head);
1626 }
1627 
1628 static void *ocrdma_hwq_head(struct ocrdma_qp_hwq_info *q)
1629 {
1630 	return q->va + (q->head * q->entry_size);
1631 }
1632 
1633 static void *ocrdma_hwq_head_from_idx(struct ocrdma_qp_hwq_info *q,
1634 				      u32 idx)
1635 {
1636 	return q->va + (idx * q->entry_size);
1637 }
1638 
1639 static void ocrdma_hwq_inc_head(struct ocrdma_qp_hwq_info *q)
1640 {
1641 	q->head = (q->head + 1) & q->max_wqe_idx;
1642 }
1643 
1644 static void ocrdma_hwq_inc_tail(struct ocrdma_qp_hwq_info *q)
1645 {
1646 	q->tail = (q->tail + 1) & q->max_wqe_idx;
1647 }
1648 
1649 /* discard the cqe for a given QP */
1650 static void ocrdma_discard_cqes(struct ocrdma_qp *qp, struct ocrdma_cq *cq)
1651 {
1652 	unsigned long cq_flags;
1653 	unsigned long flags;
1654 	int discard_cnt = 0;
1655 	u32 cur_getp, stop_getp;
1656 	struct ocrdma_cqe *cqe;
1657 	u32 qpn = 0, wqe_idx = 0;
1658 
1659 	spin_lock_irqsave(&cq->cq_lock, cq_flags);
1660 
1661 	/* traverse through the CQEs in the hw CQ,
1662 	 * find the matching CQE for a given qp,
1663 	 * mark the matching one discarded by clearing qpn.
1664 	 * ring the doorbell in the poll_cq() as
1665 	 * we don't complete out of order cqe.
1666 	 */
1667 
1668 	cur_getp = cq->getp;
1669 	/* find upto when do we reap the cq. */
1670 	stop_getp = cur_getp;
1671 	do {
1672 		if (is_hw_sq_empty(qp) && (!qp->srq && is_hw_rq_empty(qp)))
1673 			break;
1674 
1675 		cqe = cq->va + cur_getp;
1676 		/* if (a) done reaping whole hw cq, or
1677 		 *    (b) qp_xq becomes empty.
1678 		 * then exit
1679 		 */
1680 		qpn = cqe->cmn.qpn & OCRDMA_CQE_QPN_MASK;
1681 		/* if previously discarded cqe found, skip that too. */
1682 		/* check for matching qp */
1683 		if (qpn == 0 || qpn != qp->id)
1684 			goto skip_cqe;
1685 
1686 		if (is_cqe_for_sq(cqe)) {
1687 			ocrdma_hwq_inc_tail(&qp->sq);
1688 		} else {
1689 			if (qp->srq) {
1690 				wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
1691 					OCRDMA_CQE_BUFTAG_SHIFT) &
1692 					qp->srq->rq.max_wqe_idx;
1693 				BUG_ON(wqe_idx < 1);
1694 				spin_lock_irqsave(&qp->srq->q_lock, flags);
1695 				ocrdma_hwq_inc_tail(&qp->srq->rq);
1696 				ocrdma_srq_toggle_bit(qp->srq, wqe_idx - 1);
1697 				spin_unlock_irqrestore(&qp->srq->q_lock, flags);
1698 
1699 			} else {
1700 				ocrdma_hwq_inc_tail(&qp->rq);
1701 			}
1702 		}
1703 		/* mark cqe discarded so that it is not picked up later
1704 		 * in the poll_cq().
1705 		 */
1706 		discard_cnt += 1;
1707 		cqe->cmn.qpn = 0;
1708 skip_cqe:
1709 		cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
1710 	} while (cur_getp != stop_getp);
1711 	spin_unlock_irqrestore(&cq->cq_lock, cq_flags);
1712 }
1713 
1714 void ocrdma_del_flush_qp(struct ocrdma_qp *qp)
1715 {
1716 	int found = false;
1717 	unsigned long flags;
1718 	struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
1719 	/* sync with any active CQ poll */
1720 
1721 	spin_lock_irqsave(&dev->flush_q_lock, flags);
1722 	found = ocrdma_is_qp_in_sq_flushlist(qp->sq_cq, qp);
1723 	if (found)
1724 		list_del(&qp->sq_entry);
1725 	if (!qp->srq) {
1726 		found = ocrdma_is_qp_in_rq_flushlist(qp->rq_cq, qp);
1727 		if (found)
1728 			list_del(&qp->rq_entry);
1729 	}
1730 	spin_unlock_irqrestore(&dev->flush_q_lock, flags);
1731 }
1732 
1733 int ocrdma_destroy_qp(struct ib_qp *ibqp)
1734 {
1735 	struct ocrdma_pd *pd;
1736 	struct ocrdma_qp *qp;
1737 	struct ocrdma_dev *dev;
1738 	struct ib_qp_attr attrs;
1739 	int attr_mask;
1740 	unsigned long flags;
1741 
1742 	qp = get_ocrdma_qp(ibqp);
1743 	dev = get_ocrdma_dev(ibqp->device);
1744 
1745 	pd = qp->pd;
1746 
1747 	/* change the QP state to ERROR */
1748 	if (qp->state != OCRDMA_QPS_RST) {
1749 		attrs.qp_state = IB_QPS_ERR;
1750 		attr_mask = IB_QP_STATE;
1751 		_ocrdma_modify_qp(ibqp, &attrs, attr_mask);
1752 	}
1753 	/* ensure that CQEs for newly created QP (whose id may be same with
1754 	 * one which just getting destroyed are same), dont get
1755 	 * discarded until the old CQEs are discarded.
1756 	 */
1757 	mutex_lock(&dev->dev_lock);
1758 	(void) ocrdma_mbx_destroy_qp(dev, qp);
1759 
1760 	/*
1761 	 * acquire CQ lock while destroy is in progress, in order to
1762 	 * protect against proessing in-flight CQEs for this QP.
1763 	 */
1764 	spin_lock_irqsave(&qp->sq_cq->cq_lock, flags);
1765 	if (qp->rq_cq && (qp->rq_cq != qp->sq_cq)) {
1766 		spin_lock(&qp->rq_cq->cq_lock);
1767 		ocrdma_del_qpn_map(dev, qp);
1768 		spin_unlock(&qp->rq_cq->cq_lock);
1769 	} else {
1770 		ocrdma_del_qpn_map(dev, qp);
1771 	}
1772 	spin_unlock_irqrestore(&qp->sq_cq->cq_lock, flags);
1773 
1774 	if (!pd->uctx) {
1775 		ocrdma_discard_cqes(qp, qp->sq_cq);
1776 		ocrdma_discard_cqes(qp, qp->rq_cq);
1777 	}
1778 	mutex_unlock(&dev->dev_lock);
1779 
1780 	if (pd->uctx) {
1781 		ocrdma_del_mmap(pd->uctx, (u64) qp->sq.pa,
1782 				PAGE_ALIGN(qp->sq.len));
1783 		if (!qp->srq)
1784 			ocrdma_del_mmap(pd->uctx, (u64) qp->rq.pa,
1785 					PAGE_ALIGN(qp->rq.len));
1786 	}
1787 
1788 	ocrdma_del_flush_qp(qp);
1789 
1790 	kfree(qp->wqe_wr_id_tbl);
1791 	kfree(qp->rqe_wr_id_tbl);
1792 	kfree(qp);
1793 	return 0;
1794 }
1795 
1796 static int ocrdma_copy_srq_uresp(struct ocrdma_dev *dev, struct ocrdma_srq *srq,
1797 				struct ib_udata *udata)
1798 {
1799 	int status;
1800 	struct ocrdma_create_srq_uresp uresp;
1801 
1802 	memset(&uresp, 0, sizeof(uresp));
1803 	uresp.rq_dbid = srq->rq.dbid;
1804 	uresp.num_rq_pages = 1;
1805 	uresp.rq_page_addr[0] = virt_to_phys(srq->rq.va);
1806 	uresp.rq_page_size = srq->rq.len;
1807 	uresp.db_page_addr = dev->nic_info.unmapped_db +
1808 	    (srq->pd->id * dev->nic_info.db_page_size);
1809 	uresp.db_page_size = dev->nic_info.db_page_size;
1810 	uresp.num_rqe_allocated = srq->rq.max_cnt;
1811 	if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1812 		uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1813 		uresp.db_shift = 24;
1814 	} else {
1815 		uresp.db_rq_offset = OCRDMA_DB_RQ_OFFSET;
1816 		uresp.db_shift = 16;
1817 	}
1818 
1819 	status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1820 	if (status)
1821 		return status;
1822 	status = ocrdma_add_mmap(srq->pd->uctx, uresp.rq_page_addr[0],
1823 				 uresp.rq_page_size);
1824 	if (status)
1825 		return status;
1826 	return status;
1827 }
1828 
1829 struct ib_srq *ocrdma_create_srq(struct ib_pd *ibpd,
1830 				 struct ib_srq_init_attr *init_attr,
1831 				 struct ib_udata *udata)
1832 {
1833 	int status = -ENOMEM;
1834 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1835 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1836 	struct ocrdma_srq *srq;
1837 
1838 	if (init_attr->attr.max_sge > dev->attr.max_recv_sge)
1839 		return ERR_PTR(-EINVAL);
1840 	if (init_attr->attr.max_wr > dev->attr.max_rqe)
1841 		return ERR_PTR(-EINVAL);
1842 
1843 	srq = kzalloc(sizeof(*srq), GFP_KERNEL);
1844 	if (!srq)
1845 		return ERR_PTR(status);
1846 
1847 	spin_lock_init(&srq->q_lock);
1848 	srq->pd = pd;
1849 	srq->db = dev->nic_info.db + (pd->id * dev->nic_info.db_page_size);
1850 	status = ocrdma_mbx_create_srq(dev, srq, init_attr, pd);
1851 	if (status)
1852 		goto err;
1853 
1854 	if (udata == NULL) {
1855 		status = -ENOMEM;
1856 		srq->rqe_wr_id_tbl = kcalloc(srq->rq.max_cnt, sizeof(u64),
1857 					     GFP_KERNEL);
1858 		if (srq->rqe_wr_id_tbl == NULL)
1859 			goto arm_err;
1860 
1861 		srq->bit_fields_len = (srq->rq.max_cnt / 32) +
1862 		    (srq->rq.max_cnt % 32 ? 1 : 0);
1863 		srq->idx_bit_fields =
1864 		    kmalloc_array(srq->bit_fields_len, sizeof(u32),
1865 				  GFP_KERNEL);
1866 		if (srq->idx_bit_fields == NULL)
1867 			goto arm_err;
1868 		memset(srq->idx_bit_fields, 0xff,
1869 		       srq->bit_fields_len * sizeof(u32));
1870 	}
1871 
1872 	if (init_attr->attr.srq_limit) {
1873 		status = ocrdma_mbx_modify_srq(srq, &init_attr->attr);
1874 		if (status)
1875 			goto arm_err;
1876 	}
1877 
1878 	if (udata) {
1879 		status = ocrdma_copy_srq_uresp(dev, srq, udata);
1880 		if (status)
1881 			goto arm_err;
1882 	}
1883 
1884 	return &srq->ibsrq;
1885 
1886 arm_err:
1887 	ocrdma_mbx_destroy_srq(dev, srq);
1888 err:
1889 	kfree(srq->rqe_wr_id_tbl);
1890 	kfree(srq->idx_bit_fields);
1891 	kfree(srq);
1892 	return ERR_PTR(status);
1893 }
1894 
1895 int ocrdma_modify_srq(struct ib_srq *ibsrq,
1896 		      struct ib_srq_attr *srq_attr,
1897 		      enum ib_srq_attr_mask srq_attr_mask,
1898 		      struct ib_udata *udata)
1899 {
1900 	int status;
1901 	struct ocrdma_srq *srq;
1902 
1903 	srq = get_ocrdma_srq(ibsrq);
1904 	if (srq_attr_mask & IB_SRQ_MAX_WR)
1905 		status = -EINVAL;
1906 	else
1907 		status = ocrdma_mbx_modify_srq(srq, srq_attr);
1908 	return status;
1909 }
1910 
1911 int ocrdma_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
1912 {
1913 	int status;
1914 	struct ocrdma_srq *srq;
1915 
1916 	srq = get_ocrdma_srq(ibsrq);
1917 	status = ocrdma_mbx_query_srq(srq, srq_attr);
1918 	return status;
1919 }
1920 
1921 int ocrdma_destroy_srq(struct ib_srq *ibsrq)
1922 {
1923 	int status;
1924 	struct ocrdma_srq *srq;
1925 	struct ocrdma_dev *dev = get_ocrdma_dev(ibsrq->device);
1926 
1927 	srq = get_ocrdma_srq(ibsrq);
1928 
1929 	status = ocrdma_mbx_destroy_srq(dev, srq);
1930 
1931 	if (srq->pd->uctx)
1932 		ocrdma_del_mmap(srq->pd->uctx, (u64) srq->rq.pa,
1933 				PAGE_ALIGN(srq->rq.len));
1934 
1935 	kfree(srq->idx_bit_fields);
1936 	kfree(srq->rqe_wr_id_tbl);
1937 	kfree(srq);
1938 	return status;
1939 }
1940 
1941 /* unprivileged verbs and their support functions. */
1942 static void ocrdma_build_ud_hdr(struct ocrdma_qp *qp,
1943 				struct ocrdma_hdr_wqe *hdr,
1944 				const struct ib_send_wr *wr)
1945 {
1946 	struct ocrdma_ewqe_ud_hdr *ud_hdr =
1947 		(struct ocrdma_ewqe_ud_hdr *)(hdr + 1);
1948 	struct ocrdma_ah *ah = get_ocrdma_ah(ud_wr(wr)->ah);
1949 
1950 	ud_hdr->rsvd_dest_qpn = ud_wr(wr)->remote_qpn;
1951 	if (qp->qp_type == IB_QPT_GSI)
1952 		ud_hdr->qkey = qp->qkey;
1953 	else
1954 		ud_hdr->qkey = ud_wr(wr)->remote_qkey;
1955 	ud_hdr->rsvd_ahid = ah->id;
1956 	ud_hdr->hdr_type = ah->hdr_type;
1957 	if (ah->av->valid & OCRDMA_AV_VLAN_VALID)
1958 		hdr->cw |= (OCRDMA_FLAG_AH_VLAN_PR << OCRDMA_WQE_FLAGS_SHIFT);
1959 }
1960 
1961 static void ocrdma_build_sges(struct ocrdma_hdr_wqe *hdr,
1962 			      struct ocrdma_sge *sge, int num_sge,
1963 			      struct ib_sge *sg_list)
1964 {
1965 	int i;
1966 
1967 	for (i = 0; i < num_sge; i++) {
1968 		sge[i].lrkey = sg_list[i].lkey;
1969 		sge[i].addr_lo = sg_list[i].addr;
1970 		sge[i].addr_hi = upper_32_bits(sg_list[i].addr);
1971 		sge[i].len = sg_list[i].length;
1972 		hdr->total_len += sg_list[i].length;
1973 	}
1974 	if (num_sge == 0)
1975 		memset(sge, 0, sizeof(*sge));
1976 }
1977 
1978 static inline uint32_t ocrdma_sglist_len(struct ib_sge *sg_list, int num_sge)
1979 {
1980 	uint32_t total_len = 0, i;
1981 
1982 	for (i = 0; i < num_sge; i++)
1983 		total_len += sg_list[i].length;
1984 	return total_len;
1985 }
1986 
1987 
1988 static int ocrdma_build_inline_sges(struct ocrdma_qp *qp,
1989 				    struct ocrdma_hdr_wqe *hdr,
1990 				    struct ocrdma_sge *sge,
1991 				    const struct ib_send_wr *wr, u32 wqe_size)
1992 {
1993 	int i;
1994 	char *dpp_addr;
1995 
1996 	if (wr->send_flags & IB_SEND_INLINE && qp->qp_type != IB_QPT_UD) {
1997 		hdr->total_len = ocrdma_sglist_len(wr->sg_list, wr->num_sge);
1998 		if (unlikely(hdr->total_len > qp->max_inline_data)) {
1999 			pr_err("%s() supported_len=0x%x,\n"
2000 			       " unsupported len req=0x%x\n", __func__,
2001 				qp->max_inline_data, hdr->total_len);
2002 			return -EINVAL;
2003 		}
2004 		dpp_addr = (char *)sge;
2005 		for (i = 0; i < wr->num_sge; i++) {
2006 			memcpy(dpp_addr,
2007 			       (void *)(unsigned long)wr->sg_list[i].addr,
2008 			       wr->sg_list[i].length);
2009 			dpp_addr += wr->sg_list[i].length;
2010 		}
2011 
2012 		wqe_size += roundup(hdr->total_len, OCRDMA_WQE_ALIGN_BYTES);
2013 		if (0 == hdr->total_len)
2014 			wqe_size += sizeof(struct ocrdma_sge);
2015 		hdr->cw |= (OCRDMA_TYPE_INLINE << OCRDMA_WQE_TYPE_SHIFT);
2016 	} else {
2017 		ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
2018 		if (wr->num_sge)
2019 			wqe_size += (wr->num_sge * sizeof(struct ocrdma_sge));
2020 		else
2021 			wqe_size += sizeof(struct ocrdma_sge);
2022 		hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2023 	}
2024 	hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2025 	return 0;
2026 }
2027 
2028 static int ocrdma_build_send(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2029 			     const struct ib_send_wr *wr)
2030 {
2031 	int status;
2032 	struct ocrdma_sge *sge;
2033 	u32 wqe_size = sizeof(*hdr);
2034 
2035 	if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2036 		ocrdma_build_ud_hdr(qp, hdr, wr);
2037 		sge = (struct ocrdma_sge *)(hdr + 2);
2038 		wqe_size += sizeof(struct ocrdma_ewqe_ud_hdr);
2039 	} else {
2040 		sge = (struct ocrdma_sge *)(hdr + 1);
2041 	}
2042 
2043 	status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
2044 	return status;
2045 }
2046 
2047 static int ocrdma_build_write(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2048 			      const struct ib_send_wr *wr)
2049 {
2050 	int status;
2051 	struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
2052 	struct ocrdma_sge *sge = ext_rw + 1;
2053 	u32 wqe_size = sizeof(*hdr) + sizeof(*ext_rw);
2054 
2055 	status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
2056 	if (status)
2057 		return status;
2058 	ext_rw->addr_lo = rdma_wr(wr)->remote_addr;
2059 	ext_rw->addr_hi = upper_32_bits(rdma_wr(wr)->remote_addr);
2060 	ext_rw->lrkey = rdma_wr(wr)->rkey;
2061 	ext_rw->len = hdr->total_len;
2062 	return 0;
2063 }
2064 
2065 static void ocrdma_build_read(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2066 			      const struct ib_send_wr *wr)
2067 {
2068 	struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
2069 	struct ocrdma_sge *sge = ext_rw + 1;
2070 	u32 wqe_size = ((wr->num_sge + 1) * sizeof(struct ocrdma_sge)) +
2071 	    sizeof(struct ocrdma_hdr_wqe);
2072 
2073 	ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
2074 	hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2075 	hdr->cw |= (OCRDMA_READ << OCRDMA_WQE_OPCODE_SHIFT);
2076 	hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2077 
2078 	ext_rw->addr_lo = rdma_wr(wr)->remote_addr;
2079 	ext_rw->addr_hi = upper_32_bits(rdma_wr(wr)->remote_addr);
2080 	ext_rw->lrkey = rdma_wr(wr)->rkey;
2081 	ext_rw->len = hdr->total_len;
2082 }
2083 
2084 static int get_encoded_page_size(int pg_sz)
2085 {
2086 	/* Max size is 256M 4096 << 16 */
2087 	int i = 0;
2088 	for (; i < 17; i++)
2089 		if (pg_sz == (4096 << i))
2090 			break;
2091 	return i;
2092 }
2093 
2094 static int ocrdma_build_reg(struct ocrdma_qp *qp,
2095 			    struct ocrdma_hdr_wqe *hdr,
2096 			    const struct ib_reg_wr *wr)
2097 {
2098 	u64 fbo;
2099 	struct ocrdma_ewqe_fr *fast_reg = (struct ocrdma_ewqe_fr *)(hdr + 1);
2100 	struct ocrdma_mr *mr = get_ocrdma_mr(wr->mr);
2101 	struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
2102 	struct ocrdma_pbe *pbe;
2103 	u32 wqe_size = sizeof(*fast_reg) + sizeof(*hdr);
2104 	int num_pbes = 0, i;
2105 
2106 	wqe_size = roundup(wqe_size, OCRDMA_WQE_ALIGN_BYTES);
2107 
2108 	hdr->cw |= (OCRDMA_FR_MR << OCRDMA_WQE_OPCODE_SHIFT);
2109 	hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2110 
2111 	if (wr->access & IB_ACCESS_LOCAL_WRITE)
2112 		hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_LOCAL_WR;
2113 	if (wr->access & IB_ACCESS_REMOTE_WRITE)
2114 		hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_WR;
2115 	if (wr->access & IB_ACCESS_REMOTE_READ)
2116 		hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_RD;
2117 	hdr->lkey = wr->key;
2118 	hdr->total_len = mr->ibmr.length;
2119 
2120 	fbo = mr->ibmr.iova - mr->pages[0];
2121 
2122 	fast_reg->va_hi = upper_32_bits(mr->ibmr.iova);
2123 	fast_reg->va_lo = (u32) (mr->ibmr.iova & 0xffffffff);
2124 	fast_reg->fbo_hi = upper_32_bits(fbo);
2125 	fast_reg->fbo_lo = (u32) fbo & 0xffffffff;
2126 	fast_reg->num_sges = mr->npages;
2127 	fast_reg->size_sge = get_encoded_page_size(mr->ibmr.page_size);
2128 
2129 	pbe = pbl_tbl->va;
2130 	for (i = 0; i < mr->npages; i++) {
2131 		u64 buf_addr = mr->pages[i];
2132 
2133 		pbe->pa_lo = cpu_to_le32((u32) (buf_addr & PAGE_MASK));
2134 		pbe->pa_hi = cpu_to_le32((u32) upper_32_bits(buf_addr));
2135 		num_pbes += 1;
2136 		pbe++;
2137 
2138 		/* if the pbl is full storing the pbes,
2139 		 * move to next pbl.
2140 		*/
2141 		if (num_pbes == (mr->hwmr.pbl_size/sizeof(u64))) {
2142 			pbl_tbl++;
2143 			pbe = (struct ocrdma_pbe *)pbl_tbl->va;
2144 		}
2145 	}
2146 
2147 	return 0;
2148 }
2149 
2150 static void ocrdma_ring_sq_db(struct ocrdma_qp *qp)
2151 {
2152 	u32 val = qp->sq.dbid | (1 << OCRDMA_DB_SQ_SHIFT);
2153 
2154 	iowrite32(val, qp->sq_db);
2155 }
2156 
2157 int ocrdma_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
2158 		     const struct ib_send_wr **bad_wr)
2159 {
2160 	int status = 0;
2161 	struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2162 	struct ocrdma_hdr_wqe *hdr;
2163 	unsigned long flags;
2164 
2165 	spin_lock_irqsave(&qp->q_lock, flags);
2166 	if (qp->state != OCRDMA_QPS_RTS && qp->state != OCRDMA_QPS_SQD) {
2167 		spin_unlock_irqrestore(&qp->q_lock, flags);
2168 		*bad_wr = wr;
2169 		return -EINVAL;
2170 	}
2171 
2172 	while (wr) {
2173 		if (qp->qp_type == IB_QPT_UD &&
2174 		    (wr->opcode != IB_WR_SEND &&
2175 		     wr->opcode != IB_WR_SEND_WITH_IMM)) {
2176 			*bad_wr = wr;
2177 			status = -EINVAL;
2178 			break;
2179 		}
2180 		if (ocrdma_hwq_free_cnt(&qp->sq) == 0 ||
2181 		    wr->num_sge > qp->sq.max_sges) {
2182 			*bad_wr = wr;
2183 			status = -ENOMEM;
2184 			break;
2185 		}
2186 		hdr = ocrdma_hwq_head(&qp->sq);
2187 		hdr->cw = 0;
2188 		if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2189 			hdr->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2190 		if (wr->send_flags & IB_SEND_FENCE)
2191 			hdr->cw |=
2192 			    (OCRDMA_FLAG_FENCE_L << OCRDMA_WQE_FLAGS_SHIFT);
2193 		if (wr->send_flags & IB_SEND_SOLICITED)
2194 			hdr->cw |=
2195 			    (OCRDMA_FLAG_SOLICIT << OCRDMA_WQE_FLAGS_SHIFT);
2196 		hdr->total_len = 0;
2197 		switch (wr->opcode) {
2198 		case IB_WR_SEND_WITH_IMM:
2199 			hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2200 			hdr->immdt = ntohl(wr->ex.imm_data);
2201 			/* fall through */
2202 		case IB_WR_SEND:
2203 			hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2204 			ocrdma_build_send(qp, hdr, wr);
2205 			break;
2206 		case IB_WR_SEND_WITH_INV:
2207 			hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
2208 			hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2209 			hdr->lkey = wr->ex.invalidate_rkey;
2210 			status = ocrdma_build_send(qp, hdr, wr);
2211 			break;
2212 		case IB_WR_RDMA_WRITE_WITH_IMM:
2213 			hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2214 			hdr->immdt = ntohl(wr->ex.imm_data);
2215 			/* fall through */
2216 		case IB_WR_RDMA_WRITE:
2217 			hdr->cw |= (OCRDMA_WRITE << OCRDMA_WQE_OPCODE_SHIFT);
2218 			status = ocrdma_build_write(qp, hdr, wr);
2219 			break;
2220 		case IB_WR_RDMA_READ:
2221 			ocrdma_build_read(qp, hdr, wr);
2222 			break;
2223 		case IB_WR_LOCAL_INV:
2224 			hdr->cw |=
2225 			    (OCRDMA_LKEY_INV << OCRDMA_WQE_OPCODE_SHIFT);
2226 			hdr->cw |= ((sizeof(struct ocrdma_hdr_wqe) +
2227 					sizeof(struct ocrdma_sge)) /
2228 				OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT;
2229 			hdr->lkey = wr->ex.invalidate_rkey;
2230 			break;
2231 		case IB_WR_REG_MR:
2232 			status = ocrdma_build_reg(qp, hdr, reg_wr(wr));
2233 			break;
2234 		default:
2235 			status = -EINVAL;
2236 			break;
2237 		}
2238 		if (status) {
2239 			*bad_wr = wr;
2240 			break;
2241 		}
2242 		if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2243 			qp->wqe_wr_id_tbl[qp->sq.head].signaled = 1;
2244 		else
2245 			qp->wqe_wr_id_tbl[qp->sq.head].signaled = 0;
2246 		qp->wqe_wr_id_tbl[qp->sq.head].wrid = wr->wr_id;
2247 		ocrdma_cpu_to_le32(hdr, ((hdr->cw >> OCRDMA_WQE_SIZE_SHIFT) &
2248 				   OCRDMA_WQE_SIZE_MASK) * OCRDMA_WQE_STRIDE);
2249 		/* make sure wqe is written before adapter can access it */
2250 		wmb();
2251 		/* inform hw to start processing it */
2252 		ocrdma_ring_sq_db(qp);
2253 
2254 		/* update pointer, counter for next wr */
2255 		ocrdma_hwq_inc_head(&qp->sq);
2256 		wr = wr->next;
2257 	}
2258 	spin_unlock_irqrestore(&qp->q_lock, flags);
2259 	return status;
2260 }
2261 
2262 static void ocrdma_ring_rq_db(struct ocrdma_qp *qp)
2263 {
2264 	u32 val = qp->rq.dbid | (1 << OCRDMA_DB_RQ_SHIFT);
2265 
2266 	iowrite32(val, qp->rq_db);
2267 }
2268 
2269 static void ocrdma_build_rqe(struct ocrdma_hdr_wqe *rqe,
2270 			     const struct ib_recv_wr *wr, u16 tag)
2271 {
2272 	u32 wqe_size = 0;
2273 	struct ocrdma_sge *sge;
2274 	if (wr->num_sge)
2275 		wqe_size = (wr->num_sge * sizeof(*sge)) + sizeof(*rqe);
2276 	else
2277 		wqe_size = sizeof(*sge) + sizeof(*rqe);
2278 
2279 	rqe->cw = ((wqe_size / OCRDMA_WQE_STRIDE) <<
2280 				OCRDMA_WQE_SIZE_SHIFT);
2281 	rqe->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2282 	rqe->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2283 	rqe->total_len = 0;
2284 	rqe->rsvd_tag = tag;
2285 	sge = (struct ocrdma_sge *)(rqe + 1);
2286 	ocrdma_build_sges(rqe, sge, wr->num_sge, wr->sg_list);
2287 	ocrdma_cpu_to_le32(rqe, wqe_size);
2288 }
2289 
2290 int ocrdma_post_recv(struct ib_qp *ibqp, 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_qp *qp = get_ocrdma_qp(ibqp);
2296 	struct ocrdma_hdr_wqe *rqe;
2297 
2298 	spin_lock_irqsave(&qp->q_lock, flags);
2299 	if (qp->state == OCRDMA_QPS_RST || qp->state == OCRDMA_QPS_ERR) {
2300 		spin_unlock_irqrestore(&qp->q_lock, flags);
2301 		*bad_wr = wr;
2302 		return -EINVAL;
2303 	}
2304 	while (wr) {
2305 		if (ocrdma_hwq_free_cnt(&qp->rq) == 0 ||
2306 		    wr->num_sge > qp->rq.max_sges) {
2307 			*bad_wr = wr;
2308 			status = -ENOMEM;
2309 			break;
2310 		}
2311 		rqe = ocrdma_hwq_head(&qp->rq);
2312 		ocrdma_build_rqe(rqe, wr, 0);
2313 
2314 		qp->rqe_wr_id_tbl[qp->rq.head] = wr->wr_id;
2315 		/* make sure rqe is written before adapter can access it */
2316 		wmb();
2317 
2318 		/* inform hw to start processing it */
2319 		ocrdma_ring_rq_db(qp);
2320 
2321 		/* update pointer, counter for next wr */
2322 		ocrdma_hwq_inc_head(&qp->rq);
2323 		wr = wr->next;
2324 	}
2325 	spin_unlock_irqrestore(&qp->q_lock, flags);
2326 	return status;
2327 }
2328 
2329 /* cqe for srq's rqe can potentially arrive out of order.
2330  * index gives the entry in the shadow table where to store
2331  * the wr_id. tag/index is returned in cqe to reference back
2332  * for a given rqe.
2333  */
2334 static int ocrdma_srq_get_idx(struct ocrdma_srq *srq)
2335 {
2336 	int row = 0;
2337 	int indx = 0;
2338 
2339 	for (row = 0; row < srq->bit_fields_len; row++) {
2340 		if (srq->idx_bit_fields[row]) {
2341 			indx = ffs(srq->idx_bit_fields[row]);
2342 			indx = (row * 32) + (indx - 1);
2343 			BUG_ON(indx >= srq->rq.max_cnt);
2344 			ocrdma_srq_toggle_bit(srq, indx);
2345 			break;
2346 		}
2347 	}
2348 
2349 	BUG_ON(row == srq->bit_fields_len);
2350 	return indx + 1; /* Use from index 1 */
2351 }
2352 
2353 static void ocrdma_ring_srq_db(struct ocrdma_srq *srq)
2354 {
2355 	u32 val = srq->rq.dbid | (1 << 16);
2356 
2357 	iowrite32(val, srq->db + OCRDMA_DB_GEN2_SRQ_OFFSET);
2358 }
2359 
2360 int ocrdma_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
2361 			 const struct ib_recv_wr **bad_wr)
2362 {
2363 	int status = 0;
2364 	unsigned long flags;
2365 	struct ocrdma_srq *srq;
2366 	struct ocrdma_hdr_wqe *rqe;
2367 	u16 tag;
2368 
2369 	srq = get_ocrdma_srq(ibsrq);
2370 
2371 	spin_lock_irqsave(&srq->q_lock, flags);
2372 	while (wr) {
2373 		if (ocrdma_hwq_free_cnt(&srq->rq) == 0 ||
2374 		    wr->num_sge > srq->rq.max_sges) {
2375 			status = -ENOMEM;
2376 			*bad_wr = wr;
2377 			break;
2378 		}
2379 		tag = ocrdma_srq_get_idx(srq);
2380 		rqe = ocrdma_hwq_head(&srq->rq);
2381 		ocrdma_build_rqe(rqe, wr, tag);
2382 
2383 		srq->rqe_wr_id_tbl[tag] = wr->wr_id;
2384 		/* make sure rqe is written before adapter can perform DMA */
2385 		wmb();
2386 		/* inform hw to start processing it */
2387 		ocrdma_ring_srq_db(srq);
2388 		/* update pointer, counter for next wr */
2389 		ocrdma_hwq_inc_head(&srq->rq);
2390 		wr = wr->next;
2391 	}
2392 	spin_unlock_irqrestore(&srq->q_lock, flags);
2393 	return status;
2394 }
2395 
2396 static enum ib_wc_status ocrdma_to_ibwc_err(u16 status)
2397 {
2398 	enum ib_wc_status ibwc_status;
2399 
2400 	switch (status) {
2401 	case OCRDMA_CQE_GENERAL_ERR:
2402 		ibwc_status = IB_WC_GENERAL_ERR;
2403 		break;
2404 	case OCRDMA_CQE_LOC_LEN_ERR:
2405 		ibwc_status = IB_WC_LOC_LEN_ERR;
2406 		break;
2407 	case OCRDMA_CQE_LOC_QP_OP_ERR:
2408 		ibwc_status = IB_WC_LOC_QP_OP_ERR;
2409 		break;
2410 	case OCRDMA_CQE_LOC_EEC_OP_ERR:
2411 		ibwc_status = IB_WC_LOC_EEC_OP_ERR;
2412 		break;
2413 	case OCRDMA_CQE_LOC_PROT_ERR:
2414 		ibwc_status = IB_WC_LOC_PROT_ERR;
2415 		break;
2416 	case OCRDMA_CQE_WR_FLUSH_ERR:
2417 		ibwc_status = IB_WC_WR_FLUSH_ERR;
2418 		break;
2419 	case OCRDMA_CQE_MW_BIND_ERR:
2420 		ibwc_status = IB_WC_MW_BIND_ERR;
2421 		break;
2422 	case OCRDMA_CQE_BAD_RESP_ERR:
2423 		ibwc_status = IB_WC_BAD_RESP_ERR;
2424 		break;
2425 	case OCRDMA_CQE_LOC_ACCESS_ERR:
2426 		ibwc_status = IB_WC_LOC_ACCESS_ERR;
2427 		break;
2428 	case OCRDMA_CQE_REM_INV_REQ_ERR:
2429 		ibwc_status = IB_WC_REM_INV_REQ_ERR;
2430 		break;
2431 	case OCRDMA_CQE_REM_ACCESS_ERR:
2432 		ibwc_status = IB_WC_REM_ACCESS_ERR;
2433 		break;
2434 	case OCRDMA_CQE_REM_OP_ERR:
2435 		ibwc_status = IB_WC_REM_OP_ERR;
2436 		break;
2437 	case OCRDMA_CQE_RETRY_EXC_ERR:
2438 		ibwc_status = IB_WC_RETRY_EXC_ERR;
2439 		break;
2440 	case OCRDMA_CQE_RNR_RETRY_EXC_ERR:
2441 		ibwc_status = IB_WC_RNR_RETRY_EXC_ERR;
2442 		break;
2443 	case OCRDMA_CQE_LOC_RDD_VIOL_ERR:
2444 		ibwc_status = IB_WC_LOC_RDD_VIOL_ERR;
2445 		break;
2446 	case OCRDMA_CQE_REM_INV_RD_REQ_ERR:
2447 		ibwc_status = IB_WC_REM_INV_RD_REQ_ERR;
2448 		break;
2449 	case OCRDMA_CQE_REM_ABORT_ERR:
2450 		ibwc_status = IB_WC_REM_ABORT_ERR;
2451 		break;
2452 	case OCRDMA_CQE_INV_EECN_ERR:
2453 		ibwc_status = IB_WC_INV_EECN_ERR;
2454 		break;
2455 	case OCRDMA_CQE_INV_EEC_STATE_ERR:
2456 		ibwc_status = IB_WC_INV_EEC_STATE_ERR;
2457 		break;
2458 	case OCRDMA_CQE_FATAL_ERR:
2459 		ibwc_status = IB_WC_FATAL_ERR;
2460 		break;
2461 	case OCRDMA_CQE_RESP_TIMEOUT_ERR:
2462 		ibwc_status = IB_WC_RESP_TIMEOUT_ERR;
2463 		break;
2464 	default:
2465 		ibwc_status = IB_WC_GENERAL_ERR;
2466 		break;
2467 	}
2468 	return ibwc_status;
2469 }
2470 
2471 static void ocrdma_update_wc(struct ocrdma_qp *qp, struct ib_wc *ibwc,
2472 		      u32 wqe_idx)
2473 {
2474 	struct ocrdma_hdr_wqe *hdr;
2475 	struct ocrdma_sge *rw;
2476 	int opcode;
2477 
2478 	hdr = ocrdma_hwq_head_from_idx(&qp->sq, wqe_idx);
2479 
2480 	ibwc->wr_id = qp->wqe_wr_id_tbl[wqe_idx].wrid;
2481 	/* Undo the hdr->cw swap */
2482 	opcode = le32_to_cpu(hdr->cw) & OCRDMA_WQE_OPCODE_MASK;
2483 	switch (opcode) {
2484 	case OCRDMA_WRITE:
2485 		ibwc->opcode = IB_WC_RDMA_WRITE;
2486 		break;
2487 	case OCRDMA_READ:
2488 		rw = (struct ocrdma_sge *)(hdr + 1);
2489 		ibwc->opcode = IB_WC_RDMA_READ;
2490 		ibwc->byte_len = rw->len;
2491 		break;
2492 	case OCRDMA_SEND:
2493 		ibwc->opcode = IB_WC_SEND;
2494 		break;
2495 	case OCRDMA_FR_MR:
2496 		ibwc->opcode = IB_WC_REG_MR;
2497 		break;
2498 	case OCRDMA_LKEY_INV:
2499 		ibwc->opcode = IB_WC_LOCAL_INV;
2500 		break;
2501 	default:
2502 		ibwc->status = IB_WC_GENERAL_ERR;
2503 		pr_err("%s() invalid opcode received = 0x%x\n",
2504 		       __func__, hdr->cw & OCRDMA_WQE_OPCODE_MASK);
2505 		break;
2506 	}
2507 }
2508 
2509 static void ocrdma_set_cqe_status_flushed(struct ocrdma_qp *qp,
2510 						struct ocrdma_cqe *cqe)
2511 {
2512 	if (is_cqe_for_sq(cqe)) {
2513 		cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2514 				cqe->flags_status_srcqpn) &
2515 					~OCRDMA_CQE_STATUS_MASK);
2516 		cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2517 				cqe->flags_status_srcqpn) |
2518 				(OCRDMA_CQE_WR_FLUSH_ERR <<
2519 					OCRDMA_CQE_STATUS_SHIFT));
2520 	} else {
2521 		if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2522 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2523 					cqe->flags_status_srcqpn) &
2524 						~OCRDMA_CQE_UD_STATUS_MASK);
2525 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2526 					cqe->flags_status_srcqpn) |
2527 					(OCRDMA_CQE_WR_FLUSH_ERR <<
2528 						OCRDMA_CQE_UD_STATUS_SHIFT));
2529 		} else {
2530 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2531 					cqe->flags_status_srcqpn) &
2532 						~OCRDMA_CQE_STATUS_MASK);
2533 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2534 					cqe->flags_status_srcqpn) |
2535 					(OCRDMA_CQE_WR_FLUSH_ERR <<
2536 						OCRDMA_CQE_STATUS_SHIFT));
2537 		}
2538 	}
2539 }
2540 
2541 static bool ocrdma_update_err_cqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2542 				  struct ocrdma_qp *qp, int status)
2543 {
2544 	bool expand = false;
2545 
2546 	ibwc->byte_len = 0;
2547 	ibwc->qp = &qp->ibqp;
2548 	ibwc->status = ocrdma_to_ibwc_err(status);
2549 
2550 	ocrdma_flush_qp(qp);
2551 	ocrdma_qp_state_change(qp, IB_QPS_ERR, NULL);
2552 
2553 	/* if wqe/rqe pending for which cqe needs to be returned,
2554 	 * trigger inflating it.
2555 	 */
2556 	if (!is_hw_rq_empty(qp) || !is_hw_sq_empty(qp)) {
2557 		expand = true;
2558 		ocrdma_set_cqe_status_flushed(qp, cqe);
2559 	}
2560 	return expand;
2561 }
2562 
2563 static int ocrdma_update_err_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2564 				  struct ocrdma_qp *qp, int status)
2565 {
2566 	ibwc->opcode = IB_WC_RECV;
2567 	ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2568 	ocrdma_hwq_inc_tail(&qp->rq);
2569 
2570 	return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2571 }
2572 
2573 static int ocrdma_update_err_scqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2574 				  struct ocrdma_qp *qp, int status)
2575 {
2576 	ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2577 	ocrdma_hwq_inc_tail(&qp->sq);
2578 
2579 	return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2580 }
2581 
2582 
2583 static bool ocrdma_poll_err_scqe(struct ocrdma_qp *qp,
2584 				 struct ocrdma_cqe *cqe, struct ib_wc *ibwc,
2585 				 bool *polled, bool *stop)
2586 {
2587 	bool expand;
2588 	struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2589 	int status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2590 		OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2591 	if (status < OCRDMA_MAX_CQE_ERR)
2592 		atomic_inc(&dev->cqe_err_stats[status]);
2593 
2594 	/* when hw sq is empty, but rq is not empty, so we continue
2595 	 * to keep the cqe in order to get the cq event again.
2596 	 */
2597 	if (is_hw_sq_empty(qp) && !is_hw_rq_empty(qp)) {
2598 		/* when cq for rq and sq is same, it is safe to return
2599 		 * flush cqe for RQEs.
2600 		 */
2601 		if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2602 			*polled = true;
2603 			status = OCRDMA_CQE_WR_FLUSH_ERR;
2604 			expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2605 		} else {
2606 			/* stop processing further cqe as this cqe is used for
2607 			 * triggering cq event on buddy cq of RQ.
2608 			 * When QP is destroyed, this cqe will be removed
2609 			 * from the cq's hardware q.
2610 			 */
2611 			*polled = false;
2612 			*stop = true;
2613 			expand = false;
2614 		}
2615 	} else if (is_hw_sq_empty(qp)) {
2616 		/* Do nothing */
2617 		expand = false;
2618 		*polled = false;
2619 		*stop = false;
2620 	} else {
2621 		*polled = true;
2622 		expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2623 	}
2624 	return expand;
2625 }
2626 
2627 static bool ocrdma_poll_success_scqe(struct ocrdma_qp *qp,
2628 				     struct ocrdma_cqe *cqe,
2629 				     struct ib_wc *ibwc, bool *polled)
2630 {
2631 	bool expand = false;
2632 	int tail = qp->sq.tail;
2633 	u32 wqe_idx;
2634 
2635 	if (!qp->wqe_wr_id_tbl[tail].signaled) {
2636 		*polled = false;    /* WC cannot be consumed yet */
2637 	} else {
2638 		ibwc->status = IB_WC_SUCCESS;
2639 		ibwc->wc_flags = 0;
2640 		ibwc->qp = &qp->ibqp;
2641 		ocrdma_update_wc(qp, ibwc, tail);
2642 		*polled = true;
2643 	}
2644 	wqe_idx = (le32_to_cpu(cqe->wq.wqeidx) &
2645 			OCRDMA_CQE_WQEIDX_MASK) & qp->sq.max_wqe_idx;
2646 	if (tail != wqe_idx)
2647 		expand = true; /* Coalesced CQE can't be consumed yet */
2648 
2649 	ocrdma_hwq_inc_tail(&qp->sq);
2650 	return expand;
2651 }
2652 
2653 static bool ocrdma_poll_scqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2654 			     struct ib_wc *ibwc, bool *polled, bool *stop)
2655 {
2656 	int status;
2657 	bool expand;
2658 
2659 	status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2660 		OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2661 
2662 	if (status == OCRDMA_CQE_SUCCESS)
2663 		expand = ocrdma_poll_success_scqe(qp, cqe, ibwc, polled);
2664 	else
2665 		expand = ocrdma_poll_err_scqe(qp, cqe, ibwc, polled, stop);
2666 	return expand;
2667 }
2668 
2669 static int ocrdma_update_ud_rcqe(struct ocrdma_dev *dev, struct ib_wc *ibwc,
2670 				 struct ocrdma_cqe *cqe)
2671 {
2672 	int status;
2673 	u16 hdr_type = 0;
2674 
2675 	status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2676 		OCRDMA_CQE_UD_STATUS_MASK) >> OCRDMA_CQE_UD_STATUS_SHIFT;
2677 	ibwc->src_qp = le32_to_cpu(cqe->flags_status_srcqpn) &
2678 						OCRDMA_CQE_SRCQP_MASK;
2679 	ibwc->pkey_index = 0;
2680 	ibwc->wc_flags = IB_WC_GRH;
2681 	ibwc->byte_len = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2682 			  OCRDMA_CQE_UD_XFER_LEN_SHIFT) &
2683 			  OCRDMA_CQE_UD_XFER_LEN_MASK;
2684 
2685 	if (ocrdma_is_udp_encap_supported(dev)) {
2686 		hdr_type = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2687 			    OCRDMA_CQE_UD_L3TYPE_SHIFT) &
2688 			    OCRDMA_CQE_UD_L3TYPE_MASK;
2689 		ibwc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
2690 		ibwc->network_hdr_type = hdr_type;
2691 	}
2692 
2693 	return status;
2694 }
2695 
2696 static void ocrdma_update_free_srq_cqe(struct ib_wc *ibwc,
2697 				       struct ocrdma_cqe *cqe,
2698 				       struct ocrdma_qp *qp)
2699 {
2700 	unsigned long flags;
2701 	struct ocrdma_srq *srq;
2702 	u32 wqe_idx;
2703 
2704 	srq = get_ocrdma_srq(qp->ibqp.srq);
2705 	wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
2706 		OCRDMA_CQE_BUFTAG_SHIFT) & srq->rq.max_wqe_idx;
2707 	BUG_ON(wqe_idx < 1);
2708 
2709 	ibwc->wr_id = srq->rqe_wr_id_tbl[wqe_idx];
2710 	spin_lock_irqsave(&srq->q_lock, flags);
2711 	ocrdma_srq_toggle_bit(srq, wqe_idx - 1);
2712 	spin_unlock_irqrestore(&srq->q_lock, flags);
2713 	ocrdma_hwq_inc_tail(&srq->rq);
2714 }
2715 
2716 static bool ocrdma_poll_err_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2717 				struct ib_wc *ibwc, bool *polled, bool *stop,
2718 				int status)
2719 {
2720 	bool expand;
2721 	struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2722 
2723 	if (status < OCRDMA_MAX_CQE_ERR)
2724 		atomic_inc(&dev->cqe_err_stats[status]);
2725 
2726 	/* when hw_rq is empty, but wq is not empty, so continue
2727 	 * to keep the cqe to get the cq event again.
2728 	 */
2729 	if (is_hw_rq_empty(qp) && !is_hw_sq_empty(qp)) {
2730 		if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2731 			*polled = true;
2732 			status = OCRDMA_CQE_WR_FLUSH_ERR;
2733 			expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2734 		} else {
2735 			*polled = false;
2736 			*stop = true;
2737 			expand = false;
2738 		}
2739 	} else if (is_hw_rq_empty(qp)) {
2740 		/* Do nothing */
2741 		expand = false;
2742 		*polled = false;
2743 		*stop = false;
2744 	} else {
2745 		*polled = true;
2746 		expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2747 	}
2748 	return expand;
2749 }
2750 
2751 static void ocrdma_poll_success_rcqe(struct ocrdma_qp *qp,
2752 				     struct ocrdma_cqe *cqe, struct ib_wc *ibwc)
2753 {
2754 	struct ocrdma_dev *dev;
2755 
2756 	dev = get_ocrdma_dev(qp->ibqp.device);
2757 	ibwc->opcode = IB_WC_RECV;
2758 	ibwc->qp = &qp->ibqp;
2759 	ibwc->status = IB_WC_SUCCESS;
2760 
2761 	if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI)
2762 		ocrdma_update_ud_rcqe(dev, ibwc, cqe);
2763 	else
2764 		ibwc->byte_len = le32_to_cpu(cqe->rq.rxlen);
2765 
2766 	if (is_cqe_imm(cqe)) {
2767 		ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2768 		ibwc->wc_flags |= IB_WC_WITH_IMM;
2769 	} else if (is_cqe_wr_imm(cqe)) {
2770 		ibwc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
2771 		ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2772 		ibwc->wc_flags |= IB_WC_WITH_IMM;
2773 	} else if (is_cqe_invalidated(cqe)) {
2774 		ibwc->ex.invalidate_rkey = le32_to_cpu(cqe->rq.lkey_immdt);
2775 		ibwc->wc_flags |= IB_WC_WITH_INVALIDATE;
2776 	}
2777 	if (qp->ibqp.srq) {
2778 		ocrdma_update_free_srq_cqe(ibwc, cqe, qp);
2779 	} else {
2780 		ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2781 		ocrdma_hwq_inc_tail(&qp->rq);
2782 	}
2783 }
2784 
2785 static bool ocrdma_poll_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2786 			     struct ib_wc *ibwc, bool *polled, bool *stop)
2787 {
2788 	int status;
2789 	bool expand = false;
2790 
2791 	ibwc->wc_flags = 0;
2792 	if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2793 		status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2794 					OCRDMA_CQE_UD_STATUS_MASK) >>
2795 					OCRDMA_CQE_UD_STATUS_SHIFT;
2796 	} else {
2797 		status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2798 			     OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2799 	}
2800 
2801 	if (status == OCRDMA_CQE_SUCCESS) {
2802 		*polled = true;
2803 		ocrdma_poll_success_rcqe(qp, cqe, ibwc);
2804 	} else {
2805 		expand = ocrdma_poll_err_rcqe(qp, cqe, ibwc, polled, stop,
2806 					      status);
2807 	}
2808 	return expand;
2809 }
2810 
2811 static void ocrdma_change_cq_phase(struct ocrdma_cq *cq, struct ocrdma_cqe *cqe,
2812 				   u16 cur_getp)
2813 {
2814 	if (cq->phase_change) {
2815 		if (cur_getp == 0)
2816 			cq->phase = (~cq->phase & OCRDMA_CQE_VALID);
2817 	} else {
2818 		/* clear valid bit */
2819 		cqe->flags_status_srcqpn = 0;
2820 	}
2821 }
2822 
2823 static int ocrdma_poll_hwcq(struct ocrdma_cq *cq, int num_entries,
2824 			    struct ib_wc *ibwc)
2825 {
2826 	u16 qpn = 0;
2827 	int i = 0;
2828 	bool expand = false;
2829 	int polled_hw_cqes = 0;
2830 	struct ocrdma_qp *qp = NULL;
2831 	struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
2832 	struct ocrdma_cqe *cqe;
2833 	u16 cur_getp; bool polled = false; bool stop = false;
2834 
2835 	cur_getp = cq->getp;
2836 	while (num_entries) {
2837 		cqe = cq->va + cur_getp;
2838 		/* check whether valid cqe or not */
2839 		if (!is_cqe_valid(cq, cqe))
2840 			break;
2841 		qpn = (le32_to_cpu(cqe->cmn.qpn) & OCRDMA_CQE_QPN_MASK);
2842 		/* ignore discarded cqe */
2843 		if (qpn == 0)
2844 			goto skip_cqe;
2845 		qp = dev->qp_tbl[qpn];
2846 		BUG_ON(qp == NULL);
2847 
2848 		if (is_cqe_for_sq(cqe)) {
2849 			expand = ocrdma_poll_scqe(qp, cqe, ibwc, &polled,
2850 						  &stop);
2851 		} else {
2852 			expand = ocrdma_poll_rcqe(qp, cqe, ibwc, &polled,
2853 						  &stop);
2854 		}
2855 		if (expand)
2856 			goto expand_cqe;
2857 		if (stop)
2858 			goto stop_cqe;
2859 		/* clear qpn to avoid duplicate processing by discard_cqe() */
2860 		cqe->cmn.qpn = 0;
2861 skip_cqe:
2862 		polled_hw_cqes += 1;
2863 		cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
2864 		ocrdma_change_cq_phase(cq, cqe, cur_getp);
2865 expand_cqe:
2866 		if (polled) {
2867 			num_entries -= 1;
2868 			i += 1;
2869 			ibwc = ibwc + 1;
2870 			polled = false;
2871 		}
2872 	}
2873 stop_cqe:
2874 	cq->getp = cur_getp;
2875 
2876 	if (polled_hw_cqes)
2877 		ocrdma_ring_cq_db(dev, cq->id, false, false, polled_hw_cqes);
2878 
2879 	return i;
2880 }
2881 
2882 /* insert error cqe if the QP's SQ or RQ's CQ matches the CQ under poll. */
2883 static int ocrdma_add_err_cqe(struct ocrdma_cq *cq, int num_entries,
2884 			      struct ocrdma_qp *qp, struct ib_wc *ibwc)
2885 {
2886 	int err_cqes = 0;
2887 
2888 	while (num_entries) {
2889 		if (is_hw_sq_empty(qp) && is_hw_rq_empty(qp))
2890 			break;
2891 		if (!is_hw_sq_empty(qp) && qp->sq_cq == cq) {
2892 			ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2893 			ocrdma_hwq_inc_tail(&qp->sq);
2894 		} else if (!is_hw_rq_empty(qp) && qp->rq_cq == cq) {
2895 			ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2896 			ocrdma_hwq_inc_tail(&qp->rq);
2897 		} else {
2898 			return err_cqes;
2899 		}
2900 		ibwc->byte_len = 0;
2901 		ibwc->status = IB_WC_WR_FLUSH_ERR;
2902 		ibwc = ibwc + 1;
2903 		err_cqes += 1;
2904 		num_entries -= 1;
2905 	}
2906 	return err_cqes;
2907 }
2908 
2909 int ocrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
2910 {
2911 	int cqes_to_poll = num_entries;
2912 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2913 	struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2914 	int num_os_cqe = 0, err_cqes = 0;
2915 	struct ocrdma_qp *qp;
2916 	unsigned long flags;
2917 
2918 	/* poll cqes from adapter CQ */
2919 	spin_lock_irqsave(&cq->cq_lock, flags);
2920 	num_os_cqe = ocrdma_poll_hwcq(cq, cqes_to_poll, wc);
2921 	spin_unlock_irqrestore(&cq->cq_lock, flags);
2922 	cqes_to_poll -= num_os_cqe;
2923 
2924 	if (cqes_to_poll) {
2925 		wc = wc + num_os_cqe;
2926 		/* adapter returns single error cqe when qp moves to
2927 		 * error state. So insert error cqes with wc_status as
2928 		 * FLUSHED for pending WQEs and RQEs of QP's SQ and RQ
2929 		 * respectively which uses this CQ.
2930 		 */
2931 		spin_lock_irqsave(&dev->flush_q_lock, flags);
2932 		list_for_each_entry(qp, &cq->sq_head, sq_entry) {
2933 			if (cqes_to_poll == 0)
2934 				break;
2935 			err_cqes = ocrdma_add_err_cqe(cq, cqes_to_poll, qp, wc);
2936 			cqes_to_poll -= err_cqes;
2937 			num_os_cqe += err_cqes;
2938 			wc = wc + err_cqes;
2939 		}
2940 		spin_unlock_irqrestore(&dev->flush_q_lock, flags);
2941 	}
2942 	return num_os_cqe;
2943 }
2944 
2945 int ocrdma_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags cq_flags)
2946 {
2947 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2948 	struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2949 	u16 cq_id;
2950 	unsigned long flags;
2951 	bool arm_needed = false, sol_needed = false;
2952 
2953 	cq_id = cq->id;
2954 
2955 	spin_lock_irqsave(&cq->cq_lock, flags);
2956 	if (cq_flags & IB_CQ_NEXT_COMP || cq_flags & IB_CQ_SOLICITED)
2957 		arm_needed = true;
2958 	if (cq_flags & IB_CQ_SOLICITED)
2959 		sol_needed = true;
2960 
2961 	ocrdma_ring_cq_db(dev, cq_id, arm_needed, sol_needed, 0);
2962 	spin_unlock_irqrestore(&cq->cq_lock, flags);
2963 
2964 	return 0;
2965 }
2966 
2967 struct ib_mr *ocrdma_alloc_mr(struct ib_pd *ibpd,
2968 			      enum ib_mr_type mr_type,
2969 			      u32 max_num_sg)
2970 {
2971 	int status;
2972 	struct ocrdma_mr *mr;
2973 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
2974 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
2975 
2976 	if (mr_type != IB_MR_TYPE_MEM_REG)
2977 		return ERR_PTR(-EINVAL);
2978 
2979 	if (max_num_sg > dev->attr.max_pages_per_frmr)
2980 		return ERR_PTR(-EINVAL);
2981 
2982 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
2983 	if (!mr)
2984 		return ERR_PTR(-ENOMEM);
2985 
2986 	mr->pages = kcalloc(max_num_sg, sizeof(u64), GFP_KERNEL);
2987 	if (!mr->pages) {
2988 		status = -ENOMEM;
2989 		goto pl_err;
2990 	}
2991 
2992 	status = ocrdma_get_pbl_info(dev, mr, max_num_sg);
2993 	if (status)
2994 		goto pbl_err;
2995 	mr->hwmr.fr_mr = 1;
2996 	mr->hwmr.remote_rd = 0;
2997 	mr->hwmr.remote_wr = 0;
2998 	mr->hwmr.local_rd = 0;
2999 	mr->hwmr.local_wr = 0;
3000 	mr->hwmr.mw_bind = 0;
3001 	status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
3002 	if (status)
3003 		goto pbl_err;
3004 	status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, 0);
3005 	if (status)
3006 		goto mbx_err;
3007 	mr->ibmr.rkey = mr->hwmr.lkey;
3008 	mr->ibmr.lkey = mr->hwmr.lkey;
3009 	dev->stag_arr[(mr->hwmr.lkey >> 8) & (OCRDMA_MAX_STAG - 1)] =
3010 		(unsigned long) mr;
3011 	return &mr->ibmr;
3012 mbx_err:
3013 	ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
3014 pbl_err:
3015 	kfree(mr->pages);
3016 pl_err:
3017 	kfree(mr);
3018 	return ERR_PTR(-ENOMEM);
3019 }
3020 
3021 static int ocrdma_set_page(struct ib_mr *ibmr, u64 addr)
3022 {
3023 	struct ocrdma_mr *mr = get_ocrdma_mr(ibmr);
3024 
3025 	if (unlikely(mr->npages == mr->hwmr.num_pbes))
3026 		return -ENOMEM;
3027 
3028 	mr->pages[mr->npages++] = addr;
3029 
3030 	return 0;
3031 }
3032 
3033 int ocrdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
3034 		     unsigned int *sg_offset)
3035 {
3036 	struct ocrdma_mr *mr = get_ocrdma_mr(ibmr);
3037 
3038 	mr->npages = 0;
3039 
3040 	return ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, ocrdma_set_page);
3041 }
3042