xref: /linux/drivers/infiniband/hw/ocrdma/ocrdma_verbs.c (revision e5c86679d5e864947a52fb31e45a425dea3e7fa9)
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_gid(struct ib_device *ibdev, u8 port,
66 		     int index, union ib_gid *sgid)
67 {
68 	int ret;
69 	struct ocrdma_dev *dev;
70 
71 	dev = get_ocrdma_dev(ibdev);
72 	memset(sgid, 0, sizeof(*sgid));
73 	if (index >= OCRDMA_MAX_SGID)
74 		return -EINVAL;
75 
76 	ret = ib_get_cached_gid(ibdev, port, index, sgid, NULL);
77 	if (ret == -EAGAIN) {
78 		memcpy(sgid, &zgid, sizeof(*sgid));
79 		return 0;
80 	}
81 
82 	return ret;
83 }
84 
85 int ocrdma_add_gid(struct ib_device *device,
86 		   u8 port_num,
87 		   unsigned int index,
88 		   const union ib_gid *gid,
89 		   const struct ib_gid_attr *attr,
90 		   void **context) {
91 	return  0;
92 }
93 
94 int  ocrdma_del_gid(struct ib_device *device,
95 		    u8 port_num,
96 		    unsigned int index,
97 		    void **context) {
98 	return 0;
99 }
100 
101 int ocrdma_query_device(struct ib_device *ibdev, struct ib_device_attr *attr,
102 			struct ib_udata *uhw)
103 {
104 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
105 
106 	if (uhw->inlen || uhw->outlen)
107 		return -EINVAL;
108 
109 	memset(attr, 0, sizeof *attr);
110 	memcpy(&attr->fw_ver, &dev->attr.fw_ver[0],
111 	       min(sizeof(dev->attr.fw_ver), sizeof(attr->fw_ver)));
112 	ocrdma_get_guid(dev, (u8 *)&attr->sys_image_guid);
113 	attr->max_mr_size = dev->attr.max_mr_size;
114 	attr->page_size_cap = 0xffff000;
115 	attr->vendor_id = dev->nic_info.pdev->vendor;
116 	attr->vendor_part_id = dev->nic_info.pdev->device;
117 	attr->hw_ver = dev->asic_id;
118 	attr->max_qp = dev->attr.max_qp;
119 	attr->max_ah = OCRDMA_MAX_AH;
120 	attr->max_qp_wr = dev->attr.max_wqe;
121 
122 	attr->device_cap_flags = IB_DEVICE_CURR_QP_STATE_MOD |
123 					IB_DEVICE_RC_RNR_NAK_GEN |
124 					IB_DEVICE_SHUTDOWN_PORT |
125 					IB_DEVICE_SYS_IMAGE_GUID |
126 					IB_DEVICE_LOCAL_DMA_LKEY |
127 					IB_DEVICE_MEM_MGT_EXTENSIONS;
128 	attr->max_sge = min(dev->attr.max_send_sge, dev->attr.max_recv_sge);
129 	attr->max_sge_rd = dev->attr.max_rdma_sge;
130 	attr->max_cq = dev->attr.max_cq;
131 	attr->max_cqe = dev->attr.max_cqe;
132 	attr->max_mr = dev->attr.max_mr;
133 	attr->max_mw = dev->attr.max_mw;
134 	attr->max_pd = dev->attr.max_pd;
135 	attr->atomic_cap = 0;
136 	attr->max_fmr = 0;
137 	attr->max_map_per_fmr = 0;
138 	attr->max_qp_rd_atom =
139 	    min(dev->attr.max_ord_per_qp, dev->attr.max_ird_per_qp);
140 	attr->max_qp_init_rd_atom = dev->attr.max_ord_per_qp;
141 	attr->max_srq = dev->attr.max_srq;
142 	attr->max_srq_sge = dev->attr.max_srq_sge;
143 	attr->max_srq_wr = dev->attr.max_rqe;
144 	attr->local_ca_ack_delay = dev->attr.local_ca_ack_delay;
145 	attr->max_fast_reg_page_list_len = dev->attr.max_pages_per_frmr;
146 	attr->max_pkeys = 1;
147 	return 0;
148 }
149 
150 struct net_device *ocrdma_get_netdev(struct ib_device *ibdev, u8 port_num)
151 {
152 	struct ocrdma_dev *dev;
153 	struct net_device *ndev = NULL;
154 
155 	rcu_read_lock();
156 
157 	dev = get_ocrdma_dev(ibdev);
158 	if (dev)
159 		ndev = dev->nic_info.netdev;
160 	if (ndev)
161 		dev_hold(ndev);
162 
163 	rcu_read_unlock();
164 
165 	return ndev;
166 }
167 
168 static inline void get_link_speed_and_width(struct ocrdma_dev *dev,
169 					    u8 *ib_speed, u8 *ib_width)
170 {
171 	int status;
172 	u8 speed;
173 
174 	status = ocrdma_mbx_get_link_speed(dev, &speed, NULL);
175 	if (status)
176 		speed = OCRDMA_PHYS_LINK_SPEED_ZERO;
177 
178 	switch (speed) {
179 	case OCRDMA_PHYS_LINK_SPEED_1GBPS:
180 		*ib_speed = IB_SPEED_SDR;
181 		*ib_width = IB_WIDTH_1X;
182 		break;
183 
184 	case OCRDMA_PHYS_LINK_SPEED_10GBPS:
185 		*ib_speed = IB_SPEED_QDR;
186 		*ib_width = IB_WIDTH_1X;
187 		break;
188 
189 	case OCRDMA_PHYS_LINK_SPEED_20GBPS:
190 		*ib_speed = IB_SPEED_DDR;
191 		*ib_width = IB_WIDTH_4X;
192 		break;
193 
194 	case OCRDMA_PHYS_LINK_SPEED_40GBPS:
195 		*ib_speed = IB_SPEED_QDR;
196 		*ib_width = IB_WIDTH_4X;
197 		break;
198 
199 	default:
200 		/* Unsupported */
201 		*ib_speed = IB_SPEED_SDR;
202 		*ib_width = IB_WIDTH_1X;
203 	}
204 }
205 
206 int ocrdma_query_port(struct ib_device *ibdev,
207 		      u8 port, struct ib_port_attr *props)
208 {
209 	enum ib_port_state port_state;
210 	struct ocrdma_dev *dev;
211 	struct net_device *netdev;
212 
213 	/* props being zeroed by the caller, avoid zeroing it here */
214 	dev = get_ocrdma_dev(ibdev);
215 	if (port > 1) {
216 		pr_err("%s(%d) invalid_port=0x%x\n", __func__,
217 		       dev->id, port);
218 		return -EINVAL;
219 	}
220 	netdev = dev->nic_info.netdev;
221 	if (netif_running(netdev) && netif_oper_up(netdev)) {
222 		port_state = IB_PORT_ACTIVE;
223 		props->phys_state = 5;
224 	} else {
225 		port_state = IB_PORT_DOWN;
226 		props->phys_state = 3;
227 	}
228 	props->max_mtu = IB_MTU_4096;
229 	props->active_mtu = iboe_get_mtu(netdev->mtu);
230 	props->lid = 0;
231 	props->lmc = 0;
232 	props->sm_lid = 0;
233 	props->sm_sl = 0;
234 	props->state = port_state;
235 	props->port_cap_flags =
236 	    IB_PORT_CM_SUP |
237 	    IB_PORT_REINIT_SUP |
238 	    IB_PORT_DEVICE_MGMT_SUP | IB_PORT_VENDOR_CLASS_SUP |
239 	    IB_PORT_IP_BASED_GIDS;
240 	props->gid_tbl_len = OCRDMA_MAX_SGID;
241 	props->pkey_tbl_len = 1;
242 	props->bad_pkey_cntr = 0;
243 	props->qkey_viol_cntr = 0;
244 	get_link_speed_and_width(dev, &props->active_speed,
245 				 &props->active_width);
246 	props->max_msg_sz = 0x80000000;
247 	props->max_vl_num = 4;
248 	return 0;
249 }
250 
251 int ocrdma_modify_port(struct ib_device *ibdev, u8 port, int mask,
252 		       struct ib_port_modify *props)
253 {
254 	struct ocrdma_dev *dev;
255 
256 	dev = get_ocrdma_dev(ibdev);
257 	if (port > 1) {
258 		pr_err("%s(%d) invalid_port=0x%x\n", __func__, dev->id, port);
259 		return -EINVAL;
260 	}
261 	return 0;
262 }
263 
264 static int ocrdma_add_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
265 			   unsigned long len)
266 {
267 	struct ocrdma_mm *mm;
268 
269 	mm = kzalloc(sizeof(*mm), GFP_KERNEL);
270 	if (mm == NULL)
271 		return -ENOMEM;
272 	mm->key.phy_addr = phy_addr;
273 	mm->key.len = len;
274 	INIT_LIST_HEAD(&mm->entry);
275 
276 	mutex_lock(&uctx->mm_list_lock);
277 	list_add_tail(&mm->entry, &uctx->mm_head);
278 	mutex_unlock(&uctx->mm_list_lock);
279 	return 0;
280 }
281 
282 static void ocrdma_del_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
283 			    unsigned long len)
284 {
285 	struct ocrdma_mm *mm, *tmp;
286 
287 	mutex_lock(&uctx->mm_list_lock);
288 	list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
289 		if (len != mm->key.len && phy_addr != mm->key.phy_addr)
290 			continue;
291 
292 		list_del(&mm->entry);
293 		kfree(mm);
294 		break;
295 	}
296 	mutex_unlock(&uctx->mm_list_lock);
297 }
298 
299 static bool ocrdma_search_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
300 			      unsigned long len)
301 {
302 	bool found = false;
303 	struct ocrdma_mm *mm;
304 
305 	mutex_lock(&uctx->mm_list_lock);
306 	list_for_each_entry(mm, &uctx->mm_head, entry) {
307 		if (len != mm->key.len && phy_addr != mm->key.phy_addr)
308 			continue;
309 
310 		found = true;
311 		break;
312 	}
313 	mutex_unlock(&uctx->mm_list_lock);
314 	return found;
315 }
316 
317 
318 static u16 _ocrdma_pd_mgr_get_bitmap(struct ocrdma_dev *dev, bool dpp_pool)
319 {
320 	u16 pd_bitmap_idx = 0;
321 	const unsigned long *pd_bitmap;
322 
323 	if (dpp_pool) {
324 		pd_bitmap = dev->pd_mgr->pd_dpp_bitmap;
325 		pd_bitmap_idx = find_first_zero_bit(pd_bitmap,
326 						    dev->pd_mgr->max_dpp_pd);
327 		__set_bit(pd_bitmap_idx, dev->pd_mgr->pd_dpp_bitmap);
328 		dev->pd_mgr->pd_dpp_count++;
329 		if (dev->pd_mgr->pd_dpp_count > dev->pd_mgr->pd_dpp_thrsh)
330 			dev->pd_mgr->pd_dpp_thrsh = dev->pd_mgr->pd_dpp_count;
331 	} else {
332 		pd_bitmap = dev->pd_mgr->pd_norm_bitmap;
333 		pd_bitmap_idx = find_first_zero_bit(pd_bitmap,
334 						    dev->pd_mgr->max_normal_pd);
335 		__set_bit(pd_bitmap_idx, dev->pd_mgr->pd_norm_bitmap);
336 		dev->pd_mgr->pd_norm_count++;
337 		if (dev->pd_mgr->pd_norm_count > dev->pd_mgr->pd_norm_thrsh)
338 			dev->pd_mgr->pd_norm_thrsh = dev->pd_mgr->pd_norm_count;
339 	}
340 	return pd_bitmap_idx;
341 }
342 
343 static int _ocrdma_pd_mgr_put_bitmap(struct ocrdma_dev *dev, u16 pd_id,
344 					bool dpp_pool)
345 {
346 	u16 pd_count;
347 	u16 pd_bit_index;
348 
349 	pd_count = dpp_pool ? dev->pd_mgr->pd_dpp_count :
350 			      dev->pd_mgr->pd_norm_count;
351 	if (pd_count == 0)
352 		return -EINVAL;
353 
354 	if (dpp_pool) {
355 		pd_bit_index = pd_id - dev->pd_mgr->pd_dpp_start;
356 		if (pd_bit_index >= dev->pd_mgr->max_dpp_pd) {
357 			return -EINVAL;
358 		} else {
359 			__clear_bit(pd_bit_index, dev->pd_mgr->pd_dpp_bitmap);
360 			dev->pd_mgr->pd_dpp_count--;
361 		}
362 	} else {
363 		pd_bit_index = pd_id - dev->pd_mgr->pd_norm_start;
364 		if (pd_bit_index >= dev->pd_mgr->max_normal_pd) {
365 			return -EINVAL;
366 		} else {
367 			__clear_bit(pd_bit_index, dev->pd_mgr->pd_norm_bitmap);
368 			dev->pd_mgr->pd_norm_count--;
369 		}
370 	}
371 
372 	return 0;
373 }
374 
375 static int ocrdma_put_pd_num(struct ocrdma_dev *dev, u16 pd_id,
376 				   bool dpp_pool)
377 {
378 	int status;
379 
380 	mutex_lock(&dev->dev_lock);
381 	status = _ocrdma_pd_mgr_put_bitmap(dev, pd_id, dpp_pool);
382 	mutex_unlock(&dev->dev_lock);
383 	return status;
384 }
385 
386 static int ocrdma_get_pd_num(struct ocrdma_dev *dev, struct ocrdma_pd *pd)
387 {
388 	u16 pd_idx = 0;
389 	int status = 0;
390 
391 	mutex_lock(&dev->dev_lock);
392 	if (pd->dpp_enabled) {
393 		/* try allocating DPP PD, if not available then normal PD */
394 		if (dev->pd_mgr->pd_dpp_count < dev->pd_mgr->max_dpp_pd) {
395 			pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, true);
396 			pd->id = dev->pd_mgr->pd_dpp_start + pd_idx;
397 			pd->dpp_page = dev->pd_mgr->dpp_page_index + pd_idx;
398 		} else if (dev->pd_mgr->pd_norm_count <
399 			   dev->pd_mgr->max_normal_pd) {
400 			pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, false);
401 			pd->id = dev->pd_mgr->pd_norm_start + pd_idx;
402 			pd->dpp_enabled = false;
403 		} else {
404 			status = -EINVAL;
405 		}
406 	} else {
407 		if (dev->pd_mgr->pd_norm_count < dev->pd_mgr->max_normal_pd) {
408 			pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, false);
409 			pd->id = dev->pd_mgr->pd_norm_start + pd_idx;
410 		} else {
411 			status = -EINVAL;
412 		}
413 	}
414 	mutex_unlock(&dev->dev_lock);
415 	return status;
416 }
417 
418 static struct ocrdma_pd *_ocrdma_alloc_pd(struct ocrdma_dev *dev,
419 					  struct ocrdma_ucontext *uctx,
420 					  struct ib_udata *udata)
421 {
422 	struct ocrdma_pd *pd = NULL;
423 	int status;
424 
425 	pd = kzalloc(sizeof(*pd), GFP_KERNEL);
426 	if (!pd)
427 		return ERR_PTR(-ENOMEM);
428 
429 	if (udata && uctx && dev->attr.max_dpp_pds) {
430 		pd->dpp_enabled =
431 			ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R;
432 		pd->num_dpp_qp =
433 			pd->dpp_enabled ? (dev->nic_info.db_page_size /
434 					   dev->attr.wqe_size) : 0;
435 	}
436 
437 	if (dev->pd_mgr->pd_prealloc_valid) {
438 		status = ocrdma_get_pd_num(dev, pd);
439 		if (status == 0) {
440 			return pd;
441 		} else {
442 			kfree(pd);
443 			return ERR_PTR(status);
444 		}
445 	}
446 
447 retry:
448 	status = ocrdma_mbx_alloc_pd(dev, pd);
449 	if (status) {
450 		if (pd->dpp_enabled) {
451 			pd->dpp_enabled = false;
452 			pd->num_dpp_qp = 0;
453 			goto retry;
454 		} else {
455 			kfree(pd);
456 			return ERR_PTR(status);
457 		}
458 	}
459 
460 	return pd;
461 }
462 
463 static inline int is_ucontext_pd(struct ocrdma_ucontext *uctx,
464 				 struct ocrdma_pd *pd)
465 {
466 	return (uctx->cntxt_pd == pd ? true : false);
467 }
468 
469 static int _ocrdma_dealloc_pd(struct ocrdma_dev *dev,
470 			      struct ocrdma_pd *pd)
471 {
472 	int status;
473 
474 	if (dev->pd_mgr->pd_prealloc_valid)
475 		status = ocrdma_put_pd_num(dev, pd->id, pd->dpp_enabled);
476 	else
477 		status = ocrdma_mbx_dealloc_pd(dev, pd);
478 
479 	kfree(pd);
480 	return status;
481 }
482 
483 static int ocrdma_alloc_ucontext_pd(struct ocrdma_dev *dev,
484 				    struct ocrdma_ucontext *uctx,
485 				    struct ib_udata *udata)
486 {
487 	int status = 0;
488 
489 	uctx->cntxt_pd = _ocrdma_alloc_pd(dev, uctx, udata);
490 	if (IS_ERR(uctx->cntxt_pd)) {
491 		status = PTR_ERR(uctx->cntxt_pd);
492 		uctx->cntxt_pd = NULL;
493 		goto err;
494 	}
495 
496 	uctx->cntxt_pd->uctx = uctx;
497 	uctx->cntxt_pd->ibpd.device = &dev->ibdev;
498 err:
499 	return status;
500 }
501 
502 static int ocrdma_dealloc_ucontext_pd(struct ocrdma_ucontext *uctx)
503 {
504 	struct ocrdma_pd *pd = uctx->cntxt_pd;
505 	struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
506 
507 	if (uctx->pd_in_use) {
508 		pr_err("%s(%d) Freeing in use pdid=0x%x.\n",
509 		       __func__, dev->id, pd->id);
510 	}
511 	uctx->cntxt_pd = NULL;
512 	(void)_ocrdma_dealloc_pd(dev, pd);
513 	return 0;
514 }
515 
516 static struct ocrdma_pd *ocrdma_get_ucontext_pd(struct ocrdma_ucontext *uctx)
517 {
518 	struct ocrdma_pd *pd = NULL;
519 
520 	mutex_lock(&uctx->mm_list_lock);
521 	if (!uctx->pd_in_use) {
522 		uctx->pd_in_use = true;
523 		pd = uctx->cntxt_pd;
524 	}
525 	mutex_unlock(&uctx->mm_list_lock);
526 
527 	return pd;
528 }
529 
530 static void ocrdma_release_ucontext_pd(struct ocrdma_ucontext *uctx)
531 {
532 	mutex_lock(&uctx->mm_list_lock);
533 	uctx->pd_in_use = false;
534 	mutex_unlock(&uctx->mm_list_lock);
535 }
536 
537 struct ib_ucontext *ocrdma_alloc_ucontext(struct ib_device *ibdev,
538 					  struct ib_udata *udata)
539 {
540 	int status;
541 	struct ocrdma_ucontext *ctx;
542 	struct ocrdma_alloc_ucontext_resp resp;
543 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
544 	struct pci_dev *pdev = dev->nic_info.pdev;
545 	u32 map_len = roundup(sizeof(u32) * 2048, PAGE_SIZE);
546 
547 	if (!udata)
548 		return ERR_PTR(-EFAULT);
549 	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
550 	if (!ctx)
551 		return ERR_PTR(-ENOMEM);
552 	INIT_LIST_HEAD(&ctx->mm_head);
553 	mutex_init(&ctx->mm_list_lock);
554 
555 	ctx->ah_tbl.va = dma_alloc_coherent(&pdev->dev, map_len,
556 					    &ctx->ah_tbl.pa, GFP_KERNEL);
557 	if (!ctx->ah_tbl.va) {
558 		kfree(ctx);
559 		return ERR_PTR(-ENOMEM);
560 	}
561 	memset(ctx->ah_tbl.va, 0, map_len);
562 	ctx->ah_tbl.len = map_len;
563 
564 	memset(&resp, 0, sizeof(resp));
565 	resp.ah_tbl_len = ctx->ah_tbl.len;
566 	resp.ah_tbl_page = virt_to_phys(ctx->ah_tbl.va);
567 
568 	status = ocrdma_add_mmap(ctx, resp.ah_tbl_page, resp.ah_tbl_len);
569 	if (status)
570 		goto map_err;
571 
572 	status = ocrdma_alloc_ucontext_pd(dev, ctx, udata);
573 	if (status)
574 		goto pd_err;
575 
576 	resp.dev_id = dev->id;
577 	resp.max_inline_data = dev->attr.max_inline_data;
578 	resp.wqe_size = dev->attr.wqe_size;
579 	resp.rqe_size = dev->attr.rqe_size;
580 	resp.dpp_wqe_size = dev->attr.wqe_size;
581 
582 	memcpy(resp.fw_ver, dev->attr.fw_ver, sizeof(resp.fw_ver));
583 	status = ib_copy_to_udata(udata, &resp, sizeof(resp));
584 	if (status)
585 		goto cpy_err;
586 	return &ctx->ibucontext;
587 
588 cpy_err:
589 pd_err:
590 	ocrdma_del_mmap(ctx, ctx->ah_tbl.pa, ctx->ah_tbl.len);
591 map_err:
592 	dma_free_coherent(&pdev->dev, ctx->ah_tbl.len, ctx->ah_tbl.va,
593 			  ctx->ah_tbl.pa);
594 	kfree(ctx);
595 	return ERR_PTR(status);
596 }
597 
598 int ocrdma_dealloc_ucontext(struct ib_ucontext *ibctx)
599 {
600 	int status;
601 	struct ocrdma_mm *mm, *tmp;
602 	struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ibctx);
603 	struct ocrdma_dev *dev = get_ocrdma_dev(ibctx->device);
604 	struct pci_dev *pdev = dev->nic_info.pdev;
605 
606 	status = ocrdma_dealloc_ucontext_pd(uctx);
607 
608 	ocrdma_del_mmap(uctx, uctx->ah_tbl.pa, uctx->ah_tbl.len);
609 	dma_free_coherent(&pdev->dev, uctx->ah_tbl.len, uctx->ah_tbl.va,
610 			  uctx->ah_tbl.pa);
611 
612 	list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
613 		list_del(&mm->entry);
614 		kfree(mm);
615 	}
616 	kfree(uctx);
617 	return status;
618 }
619 
620 int ocrdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
621 {
622 	struct ocrdma_ucontext *ucontext = get_ocrdma_ucontext(context);
623 	struct ocrdma_dev *dev = get_ocrdma_dev(context->device);
624 	unsigned long vm_page = vma->vm_pgoff << PAGE_SHIFT;
625 	u64 unmapped_db = (u64) dev->nic_info.unmapped_db;
626 	unsigned long len = (vma->vm_end - vma->vm_start);
627 	int status;
628 	bool found;
629 
630 	if (vma->vm_start & (PAGE_SIZE - 1))
631 		return -EINVAL;
632 	found = ocrdma_search_mmap(ucontext, vma->vm_pgoff << PAGE_SHIFT, len);
633 	if (!found)
634 		return -EINVAL;
635 
636 	if ((vm_page >= unmapped_db) && (vm_page <= (unmapped_db +
637 		dev->nic_info.db_total_size)) &&
638 		(len <=	dev->nic_info.db_page_size)) {
639 		if (vma->vm_flags & VM_READ)
640 			return -EPERM;
641 
642 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
643 		status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
644 					    len, vma->vm_page_prot);
645 	} else if (dev->nic_info.dpp_unmapped_len &&
646 		(vm_page >= (u64) dev->nic_info.dpp_unmapped_addr) &&
647 		(vm_page <= (u64) (dev->nic_info.dpp_unmapped_addr +
648 			dev->nic_info.dpp_unmapped_len)) &&
649 		(len <= dev->nic_info.dpp_unmapped_len)) {
650 		if (vma->vm_flags & VM_READ)
651 			return -EPERM;
652 
653 		vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
654 		status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
655 					    len, vma->vm_page_prot);
656 	} else {
657 		status = remap_pfn_range(vma, vma->vm_start,
658 					 vma->vm_pgoff, len, vma->vm_page_prot);
659 	}
660 	return status;
661 }
662 
663 static int ocrdma_copy_pd_uresp(struct ocrdma_dev *dev, struct ocrdma_pd *pd,
664 				struct ib_ucontext *ib_ctx,
665 				struct ib_udata *udata)
666 {
667 	int status;
668 	u64 db_page_addr;
669 	u64 dpp_page_addr = 0;
670 	u32 db_page_size;
671 	struct ocrdma_alloc_pd_uresp rsp;
672 	struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ib_ctx);
673 
674 	memset(&rsp, 0, sizeof(rsp));
675 	rsp.id = pd->id;
676 	rsp.dpp_enabled = pd->dpp_enabled;
677 	db_page_addr = ocrdma_get_db_addr(dev, pd->id);
678 	db_page_size = dev->nic_info.db_page_size;
679 
680 	status = ocrdma_add_mmap(uctx, db_page_addr, db_page_size);
681 	if (status)
682 		return status;
683 
684 	if (pd->dpp_enabled) {
685 		dpp_page_addr = dev->nic_info.dpp_unmapped_addr +
686 				(pd->id * PAGE_SIZE);
687 		status = ocrdma_add_mmap(uctx, dpp_page_addr,
688 				 PAGE_SIZE);
689 		if (status)
690 			goto dpp_map_err;
691 		rsp.dpp_page_addr_hi = upper_32_bits(dpp_page_addr);
692 		rsp.dpp_page_addr_lo = dpp_page_addr;
693 	}
694 
695 	status = ib_copy_to_udata(udata, &rsp, sizeof(rsp));
696 	if (status)
697 		goto ucopy_err;
698 
699 	pd->uctx = uctx;
700 	return 0;
701 
702 ucopy_err:
703 	if (pd->dpp_enabled)
704 		ocrdma_del_mmap(pd->uctx, dpp_page_addr, PAGE_SIZE);
705 dpp_map_err:
706 	ocrdma_del_mmap(pd->uctx, db_page_addr, db_page_size);
707 	return status;
708 }
709 
710 struct ib_pd *ocrdma_alloc_pd(struct ib_device *ibdev,
711 			      struct ib_ucontext *context,
712 			      struct ib_udata *udata)
713 {
714 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
715 	struct ocrdma_pd *pd;
716 	struct ocrdma_ucontext *uctx = NULL;
717 	int status;
718 	u8 is_uctx_pd = false;
719 
720 	if (udata && context) {
721 		uctx = get_ocrdma_ucontext(context);
722 		pd = ocrdma_get_ucontext_pd(uctx);
723 		if (pd) {
724 			is_uctx_pd = true;
725 			goto pd_mapping;
726 		}
727 	}
728 
729 	pd = _ocrdma_alloc_pd(dev, uctx, udata);
730 	if (IS_ERR(pd)) {
731 		status = PTR_ERR(pd);
732 		goto exit;
733 	}
734 
735 pd_mapping:
736 	if (udata && context) {
737 		status = ocrdma_copy_pd_uresp(dev, pd, context, udata);
738 		if (status)
739 			goto err;
740 	}
741 	return &pd->ibpd;
742 
743 err:
744 	if (is_uctx_pd) {
745 		ocrdma_release_ucontext_pd(uctx);
746 	} else {
747 		status = _ocrdma_dealloc_pd(dev, pd);
748 	}
749 exit:
750 	return ERR_PTR(status);
751 }
752 
753 int ocrdma_dealloc_pd(struct ib_pd *ibpd)
754 {
755 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
756 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
757 	struct ocrdma_ucontext *uctx = NULL;
758 	int status = 0;
759 	u64 usr_db;
760 
761 	uctx = pd->uctx;
762 	if (uctx) {
763 		u64 dpp_db = dev->nic_info.dpp_unmapped_addr +
764 			(pd->id * PAGE_SIZE);
765 		if (pd->dpp_enabled)
766 			ocrdma_del_mmap(pd->uctx, dpp_db, PAGE_SIZE);
767 		usr_db = ocrdma_get_db_addr(dev, pd->id);
768 		ocrdma_del_mmap(pd->uctx, usr_db, dev->nic_info.db_page_size);
769 
770 		if (is_ucontext_pd(uctx, pd)) {
771 			ocrdma_release_ucontext_pd(uctx);
772 			return status;
773 		}
774 	}
775 	status = _ocrdma_dealloc_pd(dev, pd);
776 	return status;
777 }
778 
779 static int ocrdma_alloc_lkey(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
780 			    u32 pdid, int acc, u32 num_pbls, u32 addr_check)
781 {
782 	int status;
783 
784 	mr->hwmr.fr_mr = 0;
785 	mr->hwmr.local_rd = 1;
786 	mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
787 	mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
788 	mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
789 	mr->hwmr.mw_bind = (acc & IB_ACCESS_MW_BIND) ? 1 : 0;
790 	mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
791 	mr->hwmr.num_pbls = num_pbls;
792 
793 	status = ocrdma_mbx_alloc_lkey(dev, &mr->hwmr, pdid, addr_check);
794 	if (status)
795 		return status;
796 
797 	mr->ibmr.lkey = mr->hwmr.lkey;
798 	if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
799 		mr->ibmr.rkey = mr->hwmr.lkey;
800 	return 0;
801 }
802 
803 struct ib_mr *ocrdma_get_dma_mr(struct ib_pd *ibpd, int acc)
804 {
805 	int status;
806 	struct ocrdma_mr *mr;
807 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
808 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
809 
810 	if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE)) {
811 		pr_err("%s err, invalid access rights\n", __func__);
812 		return ERR_PTR(-EINVAL);
813 	}
814 
815 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
816 	if (!mr)
817 		return ERR_PTR(-ENOMEM);
818 
819 	status = ocrdma_alloc_lkey(dev, mr, pd->id, acc, 0,
820 				   OCRDMA_ADDR_CHECK_DISABLE);
821 	if (status) {
822 		kfree(mr);
823 		return ERR_PTR(status);
824 	}
825 
826 	return &mr->ibmr;
827 }
828 
829 static void ocrdma_free_mr_pbl_tbl(struct ocrdma_dev *dev,
830 				   struct ocrdma_hw_mr *mr)
831 {
832 	struct pci_dev *pdev = dev->nic_info.pdev;
833 	int i = 0;
834 
835 	if (mr->pbl_table) {
836 		for (i = 0; i < mr->num_pbls; i++) {
837 			if (!mr->pbl_table[i].va)
838 				continue;
839 			dma_free_coherent(&pdev->dev, mr->pbl_size,
840 					  mr->pbl_table[i].va,
841 					  mr->pbl_table[i].pa);
842 		}
843 		kfree(mr->pbl_table);
844 		mr->pbl_table = NULL;
845 	}
846 }
847 
848 static int ocrdma_get_pbl_info(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
849 			      u32 num_pbes)
850 {
851 	u32 num_pbls = 0;
852 	u32 idx = 0;
853 	int status = 0;
854 	u32 pbl_size;
855 
856 	do {
857 		pbl_size = OCRDMA_MIN_HPAGE_SIZE * (1 << idx);
858 		if (pbl_size > MAX_OCRDMA_PBL_SIZE) {
859 			status = -EFAULT;
860 			break;
861 		}
862 		num_pbls = roundup(num_pbes, (pbl_size / sizeof(u64)));
863 		num_pbls = num_pbls / (pbl_size / sizeof(u64));
864 		idx++;
865 	} while (num_pbls >= dev->attr.max_num_mr_pbl);
866 
867 	mr->hwmr.num_pbes = num_pbes;
868 	mr->hwmr.num_pbls = num_pbls;
869 	mr->hwmr.pbl_size = pbl_size;
870 	return status;
871 }
872 
873 static int ocrdma_build_pbl_tbl(struct ocrdma_dev *dev, struct ocrdma_hw_mr *mr)
874 {
875 	int status = 0;
876 	int i;
877 	u32 dma_len = mr->pbl_size;
878 	struct pci_dev *pdev = dev->nic_info.pdev;
879 	void *va;
880 	dma_addr_t pa;
881 
882 	mr->pbl_table = kzalloc(sizeof(struct ocrdma_pbl) *
883 				mr->num_pbls, GFP_KERNEL);
884 
885 	if (!mr->pbl_table)
886 		return -ENOMEM;
887 
888 	for (i = 0; i < mr->num_pbls; i++) {
889 		va = dma_alloc_coherent(&pdev->dev, dma_len, &pa, GFP_KERNEL);
890 		if (!va) {
891 			ocrdma_free_mr_pbl_tbl(dev, mr);
892 			status = -ENOMEM;
893 			break;
894 		}
895 		memset(va, 0, dma_len);
896 		mr->pbl_table[i].va = va;
897 		mr->pbl_table[i].pa = pa;
898 	}
899 	return status;
900 }
901 
902 static void build_user_pbes(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
903 			    u32 num_pbes)
904 {
905 	struct ocrdma_pbe *pbe;
906 	struct scatterlist *sg;
907 	struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
908 	struct ib_umem *umem = mr->umem;
909 	int shift, pg_cnt, pages, pbe_cnt, entry, total_num_pbes = 0;
910 
911 	if (!mr->hwmr.num_pbes)
912 		return;
913 
914 	pbe = (struct ocrdma_pbe *)pbl_tbl->va;
915 	pbe_cnt = 0;
916 
917 	shift = ilog2(umem->page_size);
918 
919 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
920 		pages = sg_dma_len(sg) >> shift;
921 		for (pg_cnt = 0; pg_cnt < pages; pg_cnt++) {
922 			/* store the page address in pbe */
923 			pbe->pa_lo =
924 			    cpu_to_le32(sg_dma_address
925 					(sg) +
926 					(umem->page_size * pg_cnt));
927 			pbe->pa_hi =
928 			    cpu_to_le32(upper_32_bits
929 					((sg_dma_address
930 					  (sg) +
931 					  umem->page_size * pg_cnt)));
932 			pbe_cnt += 1;
933 			total_num_pbes += 1;
934 			pbe++;
935 
936 			/* if done building pbes, issue the mbx cmd. */
937 			if (total_num_pbes == num_pbes)
938 				return;
939 
940 			/* if the given pbl is full storing the pbes,
941 			 * move to next pbl.
942 			 */
943 			if (pbe_cnt ==
944 				(mr->hwmr.pbl_size / sizeof(u64))) {
945 				pbl_tbl++;
946 				pbe = (struct ocrdma_pbe *)pbl_tbl->va;
947 				pbe_cnt = 0;
948 			}
949 
950 		}
951 	}
952 }
953 
954 struct ib_mr *ocrdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
955 				 u64 usr_addr, int acc, struct ib_udata *udata)
956 {
957 	int status = -ENOMEM;
958 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
959 	struct ocrdma_mr *mr;
960 	struct ocrdma_pd *pd;
961 	u32 num_pbes;
962 
963 	pd = get_ocrdma_pd(ibpd);
964 
965 	if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE))
966 		return ERR_PTR(-EINVAL);
967 
968 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
969 	if (!mr)
970 		return ERR_PTR(status);
971 	mr->umem = ib_umem_get(ibpd->uobject->context, start, len, acc, 0);
972 	if (IS_ERR(mr->umem)) {
973 		status = -EFAULT;
974 		goto umem_err;
975 	}
976 	num_pbes = ib_umem_page_count(mr->umem);
977 	status = ocrdma_get_pbl_info(dev, mr, num_pbes);
978 	if (status)
979 		goto umem_err;
980 
981 	mr->hwmr.pbe_size = mr->umem->page_size;
982 	mr->hwmr.fbo = ib_umem_offset(mr->umem);
983 	mr->hwmr.va = usr_addr;
984 	mr->hwmr.len = len;
985 	mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
986 	mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
987 	mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
988 	mr->hwmr.local_rd = 1;
989 	mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
990 	status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
991 	if (status)
992 		goto umem_err;
993 	build_user_pbes(dev, mr, num_pbes);
994 	status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
995 	if (status)
996 		goto mbx_err;
997 	mr->ibmr.lkey = mr->hwmr.lkey;
998 	if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
999 		mr->ibmr.rkey = mr->hwmr.lkey;
1000 
1001 	return &mr->ibmr;
1002 
1003 mbx_err:
1004 	ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
1005 umem_err:
1006 	kfree(mr);
1007 	return ERR_PTR(status);
1008 }
1009 
1010 int ocrdma_dereg_mr(struct ib_mr *ib_mr)
1011 {
1012 	struct ocrdma_mr *mr = get_ocrdma_mr(ib_mr);
1013 	struct ocrdma_dev *dev = get_ocrdma_dev(ib_mr->device);
1014 
1015 	(void) ocrdma_mbx_dealloc_lkey(dev, mr->hwmr.fr_mr, mr->hwmr.lkey);
1016 
1017 	kfree(mr->pages);
1018 	ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
1019 
1020 	/* it could be user registered memory. */
1021 	if (mr->umem)
1022 		ib_umem_release(mr->umem);
1023 	kfree(mr);
1024 
1025 	/* Don't stop cleanup, in case FW is unresponsive */
1026 	if (dev->mqe_ctx.fw_error_state) {
1027 		pr_err("%s(%d) fw not responding.\n",
1028 		       __func__, dev->id);
1029 	}
1030 	return 0;
1031 }
1032 
1033 static int ocrdma_copy_cq_uresp(struct ocrdma_dev *dev, struct ocrdma_cq *cq,
1034 				struct ib_udata *udata,
1035 				struct ib_ucontext *ib_ctx)
1036 {
1037 	int status;
1038 	struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ib_ctx);
1039 	struct ocrdma_create_cq_uresp uresp;
1040 
1041 	memset(&uresp, 0, sizeof(uresp));
1042 	uresp.cq_id = cq->id;
1043 	uresp.page_size = PAGE_ALIGN(cq->len);
1044 	uresp.num_pages = 1;
1045 	uresp.max_hw_cqe = cq->max_hw_cqe;
1046 	uresp.page_addr[0] = virt_to_phys(cq->va);
1047 	uresp.db_page_addr =  ocrdma_get_db_addr(dev, uctx->cntxt_pd->id);
1048 	uresp.db_page_size = dev->nic_info.db_page_size;
1049 	uresp.phase_change = cq->phase_change ? 1 : 0;
1050 	status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1051 	if (status) {
1052 		pr_err("%s(%d) copy error cqid=0x%x.\n",
1053 		       __func__, dev->id, cq->id);
1054 		goto err;
1055 	}
1056 	status = ocrdma_add_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
1057 	if (status)
1058 		goto err;
1059 	status = ocrdma_add_mmap(uctx, uresp.page_addr[0], uresp.page_size);
1060 	if (status) {
1061 		ocrdma_del_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
1062 		goto err;
1063 	}
1064 	cq->ucontext = uctx;
1065 err:
1066 	return status;
1067 }
1068 
1069 struct ib_cq *ocrdma_create_cq(struct ib_device *ibdev,
1070 			       const struct ib_cq_init_attr *attr,
1071 			       struct ib_ucontext *ib_ctx,
1072 			       struct ib_udata *udata)
1073 {
1074 	int entries = attr->cqe;
1075 	struct ocrdma_cq *cq;
1076 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
1077 	struct ocrdma_ucontext *uctx = NULL;
1078 	u16 pd_id = 0;
1079 	int status;
1080 	struct ocrdma_create_cq_ureq ureq;
1081 
1082 	if (attr->flags)
1083 		return ERR_PTR(-EINVAL);
1084 
1085 	if (udata) {
1086 		if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1087 			return ERR_PTR(-EFAULT);
1088 	} else
1089 		ureq.dpp_cq = 0;
1090 	cq = kzalloc(sizeof(*cq), GFP_KERNEL);
1091 	if (!cq)
1092 		return ERR_PTR(-ENOMEM);
1093 
1094 	spin_lock_init(&cq->cq_lock);
1095 	spin_lock_init(&cq->comp_handler_lock);
1096 	INIT_LIST_HEAD(&cq->sq_head);
1097 	INIT_LIST_HEAD(&cq->rq_head);
1098 
1099 	if (ib_ctx) {
1100 		uctx = get_ocrdma_ucontext(ib_ctx);
1101 		pd_id = uctx->cntxt_pd->id;
1102 	}
1103 
1104 	status = ocrdma_mbx_create_cq(dev, cq, entries, ureq.dpp_cq, pd_id);
1105 	if (status) {
1106 		kfree(cq);
1107 		return ERR_PTR(status);
1108 	}
1109 	if (ib_ctx) {
1110 		status = ocrdma_copy_cq_uresp(dev, cq, udata, ib_ctx);
1111 		if (status)
1112 			goto ctx_err;
1113 	}
1114 	cq->phase = OCRDMA_CQE_VALID;
1115 	dev->cq_tbl[cq->id] = cq;
1116 	return &cq->ibcq;
1117 
1118 ctx_err:
1119 	ocrdma_mbx_destroy_cq(dev, cq);
1120 	kfree(cq);
1121 	return ERR_PTR(status);
1122 }
1123 
1124 int ocrdma_resize_cq(struct ib_cq *ibcq, int new_cnt,
1125 		     struct ib_udata *udata)
1126 {
1127 	int status = 0;
1128 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
1129 
1130 	if (new_cnt < 1 || new_cnt > cq->max_hw_cqe) {
1131 		status = -EINVAL;
1132 		return status;
1133 	}
1134 	ibcq->cqe = new_cnt;
1135 	return status;
1136 }
1137 
1138 static void ocrdma_flush_cq(struct ocrdma_cq *cq)
1139 {
1140 	int cqe_cnt;
1141 	int valid_count = 0;
1142 	unsigned long flags;
1143 
1144 	struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
1145 	struct ocrdma_cqe *cqe = NULL;
1146 
1147 	cqe = cq->va;
1148 	cqe_cnt = cq->cqe_cnt;
1149 
1150 	/* Last irq might have scheduled a polling thread
1151 	 * sync-up with it before hard flushing.
1152 	 */
1153 	spin_lock_irqsave(&cq->cq_lock, flags);
1154 	while (cqe_cnt) {
1155 		if (is_cqe_valid(cq, cqe))
1156 			valid_count++;
1157 		cqe++;
1158 		cqe_cnt--;
1159 	}
1160 	ocrdma_ring_cq_db(dev, cq->id, false, false, valid_count);
1161 	spin_unlock_irqrestore(&cq->cq_lock, flags);
1162 }
1163 
1164 int ocrdma_destroy_cq(struct ib_cq *ibcq)
1165 {
1166 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
1167 	struct ocrdma_eq *eq = NULL;
1168 	struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
1169 	int pdid = 0;
1170 	u32 irq, indx;
1171 
1172 	dev->cq_tbl[cq->id] = NULL;
1173 	indx = ocrdma_get_eq_table_index(dev, cq->eqn);
1174 	BUG_ON(indx == -EINVAL);
1175 
1176 	eq = &dev->eq_tbl[indx];
1177 	irq = ocrdma_get_irq(dev, eq);
1178 	synchronize_irq(irq);
1179 	ocrdma_flush_cq(cq);
1180 
1181 	(void)ocrdma_mbx_destroy_cq(dev, cq);
1182 	if (cq->ucontext) {
1183 		pdid = cq->ucontext->cntxt_pd->id;
1184 		ocrdma_del_mmap(cq->ucontext, (u64) cq->pa,
1185 				PAGE_ALIGN(cq->len));
1186 		ocrdma_del_mmap(cq->ucontext,
1187 				ocrdma_get_db_addr(dev, pdid),
1188 				dev->nic_info.db_page_size);
1189 	}
1190 
1191 	kfree(cq);
1192 	return 0;
1193 }
1194 
1195 static int ocrdma_add_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1196 {
1197 	int status = -EINVAL;
1198 
1199 	if (qp->id < OCRDMA_MAX_QP && dev->qp_tbl[qp->id] == NULL) {
1200 		dev->qp_tbl[qp->id] = qp;
1201 		status = 0;
1202 	}
1203 	return status;
1204 }
1205 
1206 static void ocrdma_del_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1207 {
1208 	dev->qp_tbl[qp->id] = NULL;
1209 }
1210 
1211 static int ocrdma_check_qp_params(struct ib_pd *ibpd, struct ocrdma_dev *dev,
1212 				  struct ib_qp_init_attr *attrs)
1213 {
1214 	if ((attrs->qp_type != IB_QPT_GSI) &&
1215 	    (attrs->qp_type != IB_QPT_RC) &&
1216 	    (attrs->qp_type != IB_QPT_UC) &&
1217 	    (attrs->qp_type != IB_QPT_UD)) {
1218 		pr_err("%s(%d) unsupported qp type=0x%x requested\n",
1219 		       __func__, dev->id, attrs->qp_type);
1220 		return -EINVAL;
1221 	}
1222 	/* Skip the check for QP1 to support CM size of 128 */
1223 	if ((attrs->qp_type != IB_QPT_GSI) &&
1224 	    (attrs->cap.max_send_wr > dev->attr.max_wqe)) {
1225 		pr_err("%s(%d) unsupported send_wr=0x%x requested\n",
1226 		       __func__, dev->id, attrs->cap.max_send_wr);
1227 		pr_err("%s(%d) supported send_wr=0x%x\n",
1228 		       __func__, dev->id, dev->attr.max_wqe);
1229 		return -EINVAL;
1230 	}
1231 	if (!attrs->srq && (attrs->cap.max_recv_wr > dev->attr.max_rqe)) {
1232 		pr_err("%s(%d) unsupported recv_wr=0x%x requested\n",
1233 		       __func__, dev->id, attrs->cap.max_recv_wr);
1234 		pr_err("%s(%d) supported recv_wr=0x%x\n",
1235 		       __func__, dev->id, dev->attr.max_rqe);
1236 		return -EINVAL;
1237 	}
1238 	if (attrs->cap.max_inline_data > dev->attr.max_inline_data) {
1239 		pr_err("%s(%d) unsupported inline data size=0x%x requested\n",
1240 		       __func__, dev->id, attrs->cap.max_inline_data);
1241 		pr_err("%s(%d) supported inline data size=0x%x\n",
1242 		       __func__, dev->id, dev->attr.max_inline_data);
1243 		return -EINVAL;
1244 	}
1245 	if (attrs->cap.max_send_sge > dev->attr.max_send_sge) {
1246 		pr_err("%s(%d) unsupported send_sge=0x%x requested\n",
1247 		       __func__, dev->id, attrs->cap.max_send_sge);
1248 		pr_err("%s(%d) supported send_sge=0x%x\n",
1249 		       __func__, dev->id, dev->attr.max_send_sge);
1250 		return -EINVAL;
1251 	}
1252 	if (attrs->cap.max_recv_sge > dev->attr.max_recv_sge) {
1253 		pr_err("%s(%d) unsupported recv_sge=0x%x requested\n",
1254 		       __func__, dev->id, attrs->cap.max_recv_sge);
1255 		pr_err("%s(%d) supported recv_sge=0x%x\n",
1256 		       __func__, dev->id, dev->attr.max_recv_sge);
1257 		return -EINVAL;
1258 	}
1259 	/* unprivileged user space cannot create special QP */
1260 	if (ibpd->uobject && attrs->qp_type == IB_QPT_GSI) {
1261 		pr_err
1262 		    ("%s(%d) Userspace can't create special QPs of type=0x%x\n",
1263 		     __func__, dev->id, attrs->qp_type);
1264 		return -EINVAL;
1265 	}
1266 	/* allow creating only one GSI type of QP */
1267 	if (attrs->qp_type == IB_QPT_GSI && dev->gsi_qp_created) {
1268 		pr_err("%s(%d) GSI special QPs already created.\n",
1269 		       __func__, dev->id);
1270 		return -EINVAL;
1271 	}
1272 	/* verify consumer QPs are not trying to use GSI QP's CQ */
1273 	if ((attrs->qp_type != IB_QPT_GSI) && (dev->gsi_qp_created)) {
1274 		if ((dev->gsi_sqcq == get_ocrdma_cq(attrs->send_cq)) ||
1275 			(dev->gsi_rqcq == get_ocrdma_cq(attrs->recv_cq))) {
1276 			pr_err("%s(%d) Consumer QP cannot use GSI CQs.\n",
1277 				__func__, dev->id);
1278 			return -EINVAL;
1279 		}
1280 	}
1281 	return 0;
1282 }
1283 
1284 static int ocrdma_copy_qp_uresp(struct ocrdma_qp *qp,
1285 				struct ib_udata *udata, int dpp_offset,
1286 				int dpp_credit_lmt, int srq)
1287 {
1288 	int status;
1289 	u64 usr_db;
1290 	struct ocrdma_create_qp_uresp uresp;
1291 	struct ocrdma_pd *pd = qp->pd;
1292 	struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
1293 
1294 	memset(&uresp, 0, sizeof(uresp));
1295 	usr_db = dev->nic_info.unmapped_db +
1296 			(pd->id * dev->nic_info.db_page_size);
1297 	uresp.qp_id = qp->id;
1298 	uresp.sq_dbid = qp->sq.dbid;
1299 	uresp.num_sq_pages = 1;
1300 	uresp.sq_page_size = PAGE_ALIGN(qp->sq.len);
1301 	uresp.sq_page_addr[0] = virt_to_phys(qp->sq.va);
1302 	uresp.num_wqe_allocated = qp->sq.max_cnt;
1303 	if (!srq) {
1304 		uresp.rq_dbid = qp->rq.dbid;
1305 		uresp.num_rq_pages = 1;
1306 		uresp.rq_page_size = PAGE_ALIGN(qp->rq.len);
1307 		uresp.rq_page_addr[0] = virt_to_phys(qp->rq.va);
1308 		uresp.num_rqe_allocated = qp->rq.max_cnt;
1309 	}
1310 	uresp.db_page_addr = usr_db;
1311 	uresp.db_page_size = dev->nic_info.db_page_size;
1312 	uresp.db_sq_offset = OCRDMA_DB_GEN2_SQ_OFFSET;
1313 	uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1314 	uresp.db_shift = OCRDMA_DB_RQ_SHIFT;
1315 
1316 	if (qp->dpp_enabled) {
1317 		uresp.dpp_credit = dpp_credit_lmt;
1318 		uresp.dpp_offset = dpp_offset;
1319 	}
1320 	status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1321 	if (status) {
1322 		pr_err("%s(%d) user copy error.\n", __func__, dev->id);
1323 		goto err;
1324 	}
1325 	status = ocrdma_add_mmap(pd->uctx, uresp.sq_page_addr[0],
1326 				 uresp.sq_page_size);
1327 	if (status)
1328 		goto err;
1329 
1330 	if (!srq) {
1331 		status = ocrdma_add_mmap(pd->uctx, uresp.rq_page_addr[0],
1332 					 uresp.rq_page_size);
1333 		if (status)
1334 			goto rq_map_err;
1335 	}
1336 	return status;
1337 rq_map_err:
1338 	ocrdma_del_mmap(pd->uctx, uresp.sq_page_addr[0], uresp.sq_page_size);
1339 err:
1340 	return status;
1341 }
1342 
1343 static void ocrdma_set_qp_db(struct ocrdma_dev *dev, struct ocrdma_qp *qp,
1344 			     struct ocrdma_pd *pd)
1345 {
1346 	if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1347 		qp->sq_db = dev->nic_info.db +
1348 			(pd->id * dev->nic_info.db_page_size) +
1349 			OCRDMA_DB_GEN2_SQ_OFFSET;
1350 		qp->rq_db = dev->nic_info.db +
1351 			(pd->id * dev->nic_info.db_page_size) +
1352 			OCRDMA_DB_GEN2_RQ_OFFSET;
1353 	} else {
1354 		qp->sq_db = dev->nic_info.db +
1355 			(pd->id * dev->nic_info.db_page_size) +
1356 			OCRDMA_DB_SQ_OFFSET;
1357 		qp->rq_db = dev->nic_info.db +
1358 			(pd->id * dev->nic_info.db_page_size) +
1359 			OCRDMA_DB_RQ_OFFSET;
1360 	}
1361 }
1362 
1363 static int ocrdma_alloc_wr_id_tbl(struct ocrdma_qp *qp)
1364 {
1365 	qp->wqe_wr_id_tbl =
1366 	    kzalloc(sizeof(*(qp->wqe_wr_id_tbl)) * qp->sq.max_cnt,
1367 		    GFP_KERNEL);
1368 	if (qp->wqe_wr_id_tbl == NULL)
1369 		return -ENOMEM;
1370 	qp->rqe_wr_id_tbl =
1371 	    kzalloc(sizeof(u64) * qp->rq.max_cnt, GFP_KERNEL);
1372 	if (qp->rqe_wr_id_tbl == NULL)
1373 		return -ENOMEM;
1374 
1375 	return 0;
1376 }
1377 
1378 static void ocrdma_set_qp_init_params(struct ocrdma_qp *qp,
1379 				      struct ocrdma_pd *pd,
1380 				      struct ib_qp_init_attr *attrs)
1381 {
1382 	qp->pd = pd;
1383 	spin_lock_init(&qp->q_lock);
1384 	INIT_LIST_HEAD(&qp->sq_entry);
1385 	INIT_LIST_HEAD(&qp->rq_entry);
1386 
1387 	qp->qp_type = attrs->qp_type;
1388 	qp->cap_flags = OCRDMA_QP_INB_RD | OCRDMA_QP_INB_WR;
1389 	qp->max_inline_data = attrs->cap.max_inline_data;
1390 	qp->sq.max_sges = attrs->cap.max_send_sge;
1391 	qp->rq.max_sges = attrs->cap.max_recv_sge;
1392 	qp->state = OCRDMA_QPS_RST;
1393 	qp->signaled = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR) ? true : false;
1394 }
1395 
1396 static void ocrdma_store_gsi_qp_cq(struct ocrdma_dev *dev,
1397 				   struct ib_qp_init_attr *attrs)
1398 {
1399 	if (attrs->qp_type == IB_QPT_GSI) {
1400 		dev->gsi_qp_created = 1;
1401 		dev->gsi_sqcq = get_ocrdma_cq(attrs->send_cq);
1402 		dev->gsi_rqcq = get_ocrdma_cq(attrs->recv_cq);
1403 	}
1404 }
1405 
1406 struct ib_qp *ocrdma_create_qp(struct ib_pd *ibpd,
1407 			       struct ib_qp_init_attr *attrs,
1408 			       struct ib_udata *udata)
1409 {
1410 	int status;
1411 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1412 	struct ocrdma_qp *qp;
1413 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1414 	struct ocrdma_create_qp_ureq ureq;
1415 	u16 dpp_credit_lmt, dpp_offset;
1416 
1417 	status = ocrdma_check_qp_params(ibpd, dev, attrs);
1418 	if (status)
1419 		goto gen_err;
1420 
1421 	memset(&ureq, 0, sizeof(ureq));
1422 	if (udata) {
1423 		if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1424 			return ERR_PTR(-EFAULT);
1425 	}
1426 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1427 	if (!qp) {
1428 		status = -ENOMEM;
1429 		goto gen_err;
1430 	}
1431 	ocrdma_set_qp_init_params(qp, pd, attrs);
1432 	if (udata == NULL)
1433 		qp->cap_flags |= (OCRDMA_QP_MW_BIND | OCRDMA_QP_LKEY0 |
1434 					OCRDMA_QP_FAST_REG);
1435 
1436 	mutex_lock(&dev->dev_lock);
1437 	status = ocrdma_mbx_create_qp(qp, attrs, ureq.enable_dpp_cq,
1438 					ureq.dpp_cq_id,
1439 					&dpp_offset, &dpp_credit_lmt);
1440 	if (status)
1441 		goto mbx_err;
1442 
1443 	/* user space QP's wr_id table are managed in library */
1444 	if (udata == NULL) {
1445 		status = ocrdma_alloc_wr_id_tbl(qp);
1446 		if (status)
1447 			goto map_err;
1448 	}
1449 
1450 	status = ocrdma_add_qpn_map(dev, qp);
1451 	if (status)
1452 		goto map_err;
1453 	ocrdma_set_qp_db(dev, qp, pd);
1454 	if (udata) {
1455 		status = ocrdma_copy_qp_uresp(qp, udata, dpp_offset,
1456 					      dpp_credit_lmt,
1457 					      (attrs->srq != NULL));
1458 		if (status)
1459 			goto cpy_err;
1460 	}
1461 	ocrdma_store_gsi_qp_cq(dev, attrs);
1462 	qp->ibqp.qp_num = qp->id;
1463 	mutex_unlock(&dev->dev_lock);
1464 	return &qp->ibqp;
1465 
1466 cpy_err:
1467 	ocrdma_del_qpn_map(dev, qp);
1468 map_err:
1469 	ocrdma_mbx_destroy_qp(dev, qp);
1470 mbx_err:
1471 	mutex_unlock(&dev->dev_lock);
1472 	kfree(qp->wqe_wr_id_tbl);
1473 	kfree(qp->rqe_wr_id_tbl);
1474 	kfree(qp);
1475 	pr_err("%s(%d) error=%d\n", __func__, dev->id, status);
1476 gen_err:
1477 	return ERR_PTR(status);
1478 }
1479 
1480 int _ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1481 		      int attr_mask)
1482 {
1483 	int status = 0;
1484 	struct ocrdma_qp *qp;
1485 	struct ocrdma_dev *dev;
1486 	enum ib_qp_state old_qps;
1487 
1488 	qp = get_ocrdma_qp(ibqp);
1489 	dev = get_ocrdma_dev(ibqp->device);
1490 	if (attr_mask & IB_QP_STATE)
1491 		status = ocrdma_qp_state_change(qp, attr->qp_state, &old_qps);
1492 	/* if new and previous states are same hw doesn't need to
1493 	 * know about it.
1494 	 */
1495 	if (status < 0)
1496 		return status;
1497 	return ocrdma_mbx_modify_qp(dev, qp, attr, attr_mask);
1498 }
1499 
1500 int ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1501 		     int attr_mask, struct ib_udata *udata)
1502 {
1503 	unsigned long flags;
1504 	int status = -EINVAL;
1505 	struct ocrdma_qp *qp;
1506 	struct ocrdma_dev *dev;
1507 	enum ib_qp_state old_qps, new_qps;
1508 
1509 	qp = get_ocrdma_qp(ibqp);
1510 	dev = get_ocrdma_dev(ibqp->device);
1511 
1512 	/* syncronize with multiple context trying to change, retrive qps */
1513 	mutex_lock(&dev->dev_lock);
1514 	/* syncronize with wqe, rqe posting and cqe processing contexts */
1515 	spin_lock_irqsave(&qp->q_lock, flags);
1516 	old_qps = get_ibqp_state(qp->state);
1517 	if (attr_mask & IB_QP_STATE)
1518 		new_qps = attr->qp_state;
1519 	else
1520 		new_qps = old_qps;
1521 	spin_unlock_irqrestore(&qp->q_lock, flags);
1522 
1523 	if (!ib_modify_qp_is_ok(old_qps, new_qps, ibqp->qp_type, attr_mask,
1524 				IB_LINK_LAYER_ETHERNET)) {
1525 		pr_err("%s(%d) invalid attribute mask=0x%x specified for\n"
1526 		       "qpn=0x%x of type=0x%x old_qps=0x%x, new_qps=0x%x\n",
1527 		       __func__, dev->id, attr_mask, qp->id, ibqp->qp_type,
1528 		       old_qps, new_qps);
1529 		goto param_err;
1530 	}
1531 
1532 	status = _ocrdma_modify_qp(ibqp, attr, attr_mask);
1533 	if (status > 0)
1534 		status = 0;
1535 param_err:
1536 	mutex_unlock(&dev->dev_lock);
1537 	return status;
1538 }
1539 
1540 static enum ib_mtu ocrdma_mtu_int_to_enum(u16 mtu)
1541 {
1542 	switch (mtu) {
1543 	case 256:
1544 		return IB_MTU_256;
1545 	case 512:
1546 		return IB_MTU_512;
1547 	case 1024:
1548 		return IB_MTU_1024;
1549 	case 2048:
1550 		return IB_MTU_2048;
1551 	case 4096:
1552 		return IB_MTU_4096;
1553 	default:
1554 		return IB_MTU_1024;
1555 	}
1556 }
1557 
1558 static int ocrdma_to_ib_qp_acc_flags(int qp_cap_flags)
1559 {
1560 	int ib_qp_acc_flags = 0;
1561 
1562 	if (qp_cap_flags & OCRDMA_QP_INB_WR)
1563 		ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE;
1564 	if (qp_cap_flags & OCRDMA_QP_INB_RD)
1565 		ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE;
1566 	return ib_qp_acc_flags;
1567 }
1568 
1569 int ocrdma_query_qp(struct ib_qp *ibqp,
1570 		    struct ib_qp_attr *qp_attr,
1571 		    int attr_mask, struct ib_qp_init_attr *qp_init_attr)
1572 {
1573 	int status;
1574 	u32 qp_state;
1575 	struct ocrdma_qp_params params;
1576 	struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
1577 	struct ocrdma_dev *dev = get_ocrdma_dev(ibqp->device);
1578 
1579 	memset(&params, 0, sizeof(params));
1580 	mutex_lock(&dev->dev_lock);
1581 	status = ocrdma_mbx_query_qp(dev, qp, &params);
1582 	mutex_unlock(&dev->dev_lock);
1583 	if (status)
1584 		goto mbx_err;
1585 	if (qp->qp_type == IB_QPT_UD)
1586 		qp_attr->qkey = params.qkey;
1587 	qp_attr->path_mtu =
1588 		ocrdma_mtu_int_to_enum(params.path_mtu_pkey_indx &
1589 				OCRDMA_QP_PARAMS_PATH_MTU_MASK) >>
1590 				OCRDMA_QP_PARAMS_PATH_MTU_SHIFT;
1591 	qp_attr->path_mig_state = IB_MIG_MIGRATED;
1592 	qp_attr->rq_psn = params.hop_lmt_rq_psn & OCRDMA_QP_PARAMS_RQ_PSN_MASK;
1593 	qp_attr->sq_psn = params.tclass_sq_psn & OCRDMA_QP_PARAMS_SQ_PSN_MASK;
1594 	qp_attr->dest_qp_num =
1595 	    params.ack_to_rnr_rtc_dest_qpn & OCRDMA_QP_PARAMS_DEST_QPN_MASK;
1596 
1597 	qp_attr->qp_access_flags = ocrdma_to_ib_qp_acc_flags(qp->cap_flags);
1598 	qp_attr->cap.max_send_wr = qp->sq.max_cnt - 1;
1599 	qp_attr->cap.max_recv_wr = qp->rq.max_cnt - 1;
1600 	qp_attr->cap.max_send_sge = qp->sq.max_sges;
1601 	qp_attr->cap.max_recv_sge = qp->rq.max_sges;
1602 	qp_attr->cap.max_inline_data = qp->max_inline_data;
1603 	qp_init_attr->cap = qp_attr->cap;
1604 	memcpy(&qp_attr->ah_attr.grh.dgid, &params.dgid[0],
1605 	       sizeof(params.dgid));
1606 	qp_attr->ah_attr.grh.flow_label = params.rnt_rc_sl_fl &
1607 	    OCRDMA_QP_PARAMS_FLOW_LABEL_MASK;
1608 	qp_attr->ah_attr.grh.sgid_index = qp->sgid_idx;
1609 	qp_attr->ah_attr.grh.hop_limit = (params.hop_lmt_rq_psn &
1610 					  OCRDMA_QP_PARAMS_HOP_LMT_MASK) >>
1611 						OCRDMA_QP_PARAMS_HOP_LMT_SHIFT;
1612 	qp_attr->ah_attr.grh.traffic_class = (params.tclass_sq_psn &
1613 					      OCRDMA_QP_PARAMS_TCLASS_MASK) >>
1614 						OCRDMA_QP_PARAMS_TCLASS_SHIFT;
1615 
1616 	qp_attr->ah_attr.ah_flags = IB_AH_GRH;
1617 	qp_attr->ah_attr.port_num = 1;
1618 	qp_attr->ah_attr.sl = (params.rnt_rc_sl_fl &
1619 			       OCRDMA_QP_PARAMS_SL_MASK) >>
1620 				OCRDMA_QP_PARAMS_SL_SHIFT;
1621 	qp_attr->timeout = (params.ack_to_rnr_rtc_dest_qpn &
1622 			    OCRDMA_QP_PARAMS_ACK_TIMEOUT_MASK) >>
1623 				OCRDMA_QP_PARAMS_ACK_TIMEOUT_SHIFT;
1624 	qp_attr->rnr_retry = (params.ack_to_rnr_rtc_dest_qpn &
1625 			      OCRDMA_QP_PARAMS_RNR_RETRY_CNT_MASK) >>
1626 				OCRDMA_QP_PARAMS_RNR_RETRY_CNT_SHIFT;
1627 	qp_attr->retry_cnt =
1628 	    (params.rnt_rc_sl_fl & OCRDMA_QP_PARAMS_RETRY_CNT_MASK) >>
1629 		OCRDMA_QP_PARAMS_RETRY_CNT_SHIFT;
1630 	qp_attr->min_rnr_timer = 0;
1631 	qp_attr->pkey_index = 0;
1632 	qp_attr->port_num = 1;
1633 	qp_attr->ah_attr.src_path_bits = 0;
1634 	qp_attr->ah_attr.static_rate = 0;
1635 	qp_attr->alt_pkey_index = 0;
1636 	qp_attr->alt_port_num = 0;
1637 	qp_attr->alt_timeout = 0;
1638 	memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr));
1639 	qp_state = (params.max_sge_recv_flags & OCRDMA_QP_PARAMS_STATE_MASK) >>
1640 		    OCRDMA_QP_PARAMS_STATE_SHIFT;
1641 	qp_attr->qp_state = get_ibqp_state(qp_state);
1642 	qp_attr->cur_qp_state = qp_attr->qp_state;
1643 	qp_attr->sq_draining = (qp_state == OCRDMA_QPS_SQ_DRAINING) ? 1 : 0;
1644 	qp_attr->max_dest_rd_atomic =
1645 	    params.max_ord_ird >> OCRDMA_QP_PARAMS_MAX_ORD_SHIFT;
1646 	qp_attr->max_rd_atomic =
1647 	    params.max_ord_ird & OCRDMA_QP_PARAMS_MAX_IRD_MASK;
1648 	qp_attr->en_sqd_async_notify = (params.max_sge_recv_flags &
1649 				OCRDMA_QP_PARAMS_FLAGS_SQD_ASYNC) ? 1 : 0;
1650 	/* Sync driver QP state with FW */
1651 	ocrdma_qp_state_change(qp, qp_attr->qp_state, NULL);
1652 mbx_err:
1653 	return status;
1654 }
1655 
1656 static void ocrdma_srq_toggle_bit(struct ocrdma_srq *srq, unsigned int idx)
1657 {
1658 	unsigned int i = idx / 32;
1659 	u32 mask = (1U << (idx % 32));
1660 
1661 	srq->idx_bit_fields[i] ^= mask;
1662 }
1663 
1664 static int ocrdma_hwq_free_cnt(struct ocrdma_qp_hwq_info *q)
1665 {
1666 	return ((q->max_wqe_idx - q->head) + q->tail) % q->max_cnt;
1667 }
1668 
1669 static int is_hw_sq_empty(struct ocrdma_qp *qp)
1670 {
1671 	return (qp->sq.tail == qp->sq.head);
1672 }
1673 
1674 static int is_hw_rq_empty(struct ocrdma_qp *qp)
1675 {
1676 	return (qp->rq.tail == qp->rq.head);
1677 }
1678 
1679 static void *ocrdma_hwq_head(struct ocrdma_qp_hwq_info *q)
1680 {
1681 	return q->va + (q->head * q->entry_size);
1682 }
1683 
1684 static void *ocrdma_hwq_head_from_idx(struct ocrdma_qp_hwq_info *q,
1685 				      u32 idx)
1686 {
1687 	return q->va + (idx * q->entry_size);
1688 }
1689 
1690 static void ocrdma_hwq_inc_head(struct ocrdma_qp_hwq_info *q)
1691 {
1692 	q->head = (q->head + 1) & q->max_wqe_idx;
1693 }
1694 
1695 static void ocrdma_hwq_inc_tail(struct ocrdma_qp_hwq_info *q)
1696 {
1697 	q->tail = (q->tail + 1) & q->max_wqe_idx;
1698 }
1699 
1700 /* discard the cqe for a given QP */
1701 static void ocrdma_discard_cqes(struct ocrdma_qp *qp, struct ocrdma_cq *cq)
1702 {
1703 	unsigned long cq_flags;
1704 	unsigned long flags;
1705 	int discard_cnt = 0;
1706 	u32 cur_getp, stop_getp;
1707 	struct ocrdma_cqe *cqe;
1708 	u32 qpn = 0, wqe_idx = 0;
1709 
1710 	spin_lock_irqsave(&cq->cq_lock, cq_flags);
1711 
1712 	/* traverse through the CQEs in the hw CQ,
1713 	 * find the matching CQE for a given qp,
1714 	 * mark the matching one discarded by clearing qpn.
1715 	 * ring the doorbell in the poll_cq() as
1716 	 * we don't complete out of order cqe.
1717 	 */
1718 
1719 	cur_getp = cq->getp;
1720 	/* find upto when do we reap the cq. */
1721 	stop_getp = cur_getp;
1722 	do {
1723 		if (is_hw_sq_empty(qp) && (!qp->srq && is_hw_rq_empty(qp)))
1724 			break;
1725 
1726 		cqe = cq->va + cur_getp;
1727 		/* if (a) done reaping whole hw cq, or
1728 		 *    (b) qp_xq becomes empty.
1729 		 * then exit
1730 		 */
1731 		qpn = cqe->cmn.qpn & OCRDMA_CQE_QPN_MASK;
1732 		/* if previously discarded cqe found, skip that too. */
1733 		/* check for matching qp */
1734 		if (qpn == 0 || qpn != qp->id)
1735 			goto skip_cqe;
1736 
1737 		if (is_cqe_for_sq(cqe)) {
1738 			ocrdma_hwq_inc_tail(&qp->sq);
1739 		} else {
1740 			if (qp->srq) {
1741 				wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
1742 					OCRDMA_CQE_BUFTAG_SHIFT) &
1743 					qp->srq->rq.max_wqe_idx;
1744 				BUG_ON(wqe_idx < 1);
1745 				spin_lock_irqsave(&qp->srq->q_lock, flags);
1746 				ocrdma_hwq_inc_tail(&qp->srq->rq);
1747 				ocrdma_srq_toggle_bit(qp->srq, wqe_idx - 1);
1748 				spin_unlock_irqrestore(&qp->srq->q_lock, flags);
1749 
1750 			} else {
1751 				ocrdma_hwq_inc_tail(&qp->rq);
1752 			}
1753 		}
1754 		/* mark cqe discarded so that it is not picked up later
1755 		 * in the poll_cq().
1756 		 */
1757 		discard_cnt += 1;
1758 		cqe->cmn.qpn = 0;
1759 skip_cqe:
1760 		cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
1761 	} while (cur_getp != stop_getp);
1762 	spin_unlock_irqrestore(&cq->cq_lock, cq_flags);
1763 }
1764 
1765 void ocrdma_del_flush_qp(struct ocrdma_qp *qp)
1766 {
1767 	int found = false;
1768 	unsigned long flags;
1769 	struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
1770 	/* sync with any active CQ poll */
1771 
1772 	spin_lock_irqsave(&dev->flush_q_lock, flags);
1773 	found = ocrdma_is_qp_in_sq_flushlist(qp->sq_cq, qp);
1774 	if (found)
1775 		list_del(&qp->sq_entry);
1776 	if (!qp->srq) {
1777 		found = ocrdma_is_qp_in_rq_flushlist(qp->rq_cq, qp);
1778 		if (found)
1779 			list_del(&qp->rq_entry);
1780 	}
1781 	spin_unlock_irqrestore(&dev->flush_q_lock, flags);
1782 }
1783 
1784 int ocrdma_destroy_qp(struct ib_qp *ibqp)
1785 {
1786 	struct ocrdma_pd *pd;
1787 	struct ocrdma_qp *qp;
1788 	struct ocrdma_dev *dev;
1789 	struct ib_qp_attr attrs;
1790 	int attr_mask;
1791 	unsigned long flags;
1792 
1793 	qp = get_ocrdma_qp(ibqp);
1794 	dev = get_ocrdma_dev(ibqp->device);
1795 
1796 	pd = qp->pd;
1797 
1798 	/* change the QP state to ERROR */
1799 	if (qp->state != OCRDMA_QPS_RST) {
1800 		attrs.qp_state = IB_QPS_ERR;
1801 		attr_mask = IB_QP_STATE;
1802 		_ocrdma_modify_qp(ibqp, &attrs, attr_mask);
1803 	}
1804 	/* ensure that CQEs for newly created QP (whose id may be same with
1805 	 * one which just getting destroyed are same), dont get
1806 	 * discarded until the old CQEs are discarded.
1807 	 */
1808 	mutex_lock(&dev->dev_lock);
1809 	(void) ocrdma_mbx_destroy_qp(dev, qp);
1810 
1811 	/*
1812 	 * acquire CQ lock while destroy is in progress, in order to
1813 	 * protect against proessing in-flight CQEs for this QP.
1814 	 */
1815 	spin_lock_irqsave(&qp->sq_cq->cq_lock, flags);
1816 	if (qp->rq_cq && (qp->rq_cq != qp->sq_cq))
1817 		spin_lock(&qp->rq_cq->cq_lock);
1818 
1819 	ocrdma_del_qpn_map(dev, qp);
1820 
1821 	if (qp->rq_cq && (qp->rq_cq != qp->sq_cq))
1822 		spin_unlock(&qp->rq_cq->cq_lock);
1823 	spin_unlock_irqrestore(&qp->sq_cq->cq_lock, flags);
1824 
1825 	if (!pd->uctx) {
1826 		ocrdma_discard_cqes(qp, qp->sq_cq);
1827 		ocrdma_discard_cqes(qp, qp->rq_cq);
1828 	}
1829 	mutex_unlock(&dev->dev_lock);
1830 
1831 	if (pd->uctx) {
1832 		ocrdma_del_mmap(pd->uctx, (u64) qp->sq.pa,
1833 				PAGE_ALIGN(qp->sq.len));
1834 		if (!qp->srq)
1835 			ocrdma_del_mmap(pd->uctx, (u64) qp->rq.pa,
1836 					PAGE_ALIGN(qp->rq.len));
1837 	}
1838 
1839 	ocrdma_del_flush_qp(qp);
1840 
1841 	kfree(qp->wqe_wr_id_tbl);
1842 	kfree(qp->rqe_wr_id_tbl);
1843 	kfree(qp);
1844 	return 0;
1845 }
1846 
1847 static int ocrdma_copy_srq_uresp(struct ocrdma_dev *dev, struct ocrdma_srq *srq,
1848 				struct ib_udata *udata)
1849 {
1850 	int status;
1851 	struct ocrdma_create_srq_uresp uresp;
1852 
1853 	memset(&uresp, 0, sizeof(uresp));
1854 	uresp.rq_dbid = srq->rq.dbid;
1855 	uresp.num_rq_pages = 1;
1856 	uresp.rq_page_addr[0] = virt_to_phys(srq->rq.va);
1857 	uresp.rq_page_size = srq->rq.len;
1858 	uresp.db_page_addr = dev->nic_info.unmapped_db +
1859 	    (srq->pd->id * dev->nic_info.db_page_size);
1860 	uresp.db_page_size = dev->nic_info.db_page_size;
1861 	uresp.num_rqe_allocated = srq->rq.max_cnt;
1862 	if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1863 		uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1864 		uresp.db_shift = 24;
1865 	} else {
1866 		uresp.db_rq_offset = OCRDMA_DB_RQ_OFFSET;
1867 		uresp.db_shift = 16;
1868 	}
1869 
1870 	status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1871 	if (status)
1872 		return status;
1873 	status = ocrdma_add_mmap(srq->pd->uctx, uresp.rq_page_addr[0],
1874 				 uresp.rq_page_size);
1875 	if (status)
1876 		return status;
1877 	return status;
1878 }
1879 
1880 struct ib_srq *ocrdma_create_srq(struct ib_pd *ibpd,
1881 				 struct ib_srq_init_attr *init_attr,
1882 				 struct ib_udata *udata)
1883 {
1884 	int status = -ENOMEM;
1885 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1886 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1887 	struct ocrdma_srq *srq;
1888 
1889 	if (init_attr->attr.max_sge > dev->attr.max_recv_sge)
1890 		return ERR_PTR(-EINVAL);
1891 	if (init_attr->attr.max_wr > dev->attr.max_rqe)
1892 		return ERR_PTR(-EINVAL);
1893 
1894 	srq = kzalloc(sizeof(*srq), GFP_KERNEL);
1895 	if (!srq)
1896 		return ERR_PTR(status);
1897 
1898 	spin_lock_init(&srq->q_lock);
1899 	srq->pd = pd;
1900 	srq->db = dev->nic_info.db + (pd->id * dev->nic_info.db_page_size);
1901 	status = ocrdma_mbx_create_srq(dev, srq, init_attr, pd);
1902 	if (status)
1903 		goto err;
1904 
1905 	if (udata == NULL) {
1906 		srq->rqe_wr_id_tbl = kzalloc(sizeof(u64) * srq->rq.max_cnt,
1907 			    GFP_KERNEL);
1908 		if (srq->rqe_wr_id_tbl == NULL)
1909 			goto arm_err;
1910 
1911 		srq->bit_fields_len = (srq->rq.max_cnt / 32) +
1912 		    (srq->rq.max_cnt % 32 ? 1 : 0);
1913 		srq->idx_bit_fields =
1914 		    kmalloc(srq->bit_fields_len * sizeof(u32), GFP_KERNEL);
1915 		if (srq->idx_bit_fields == NULL)
1916 			goto arm_err;
1917 		memset(srq->idx_bit_fields, 0xff,
1918 		       srq->bit_fields_len * sizeof(u32));
1919 	}
1920 
1921 	if (init_attr->attr.srq_limit) {
1922 		status = ocrdma_mbx_modify_srq(srq, &init_attr->attr);
1923 		if (status)
1924 			goto arm_err;
1925 	}
1926 
1927 	if (udata) {
1928 		status = ocrdma_copy_srq_uresp(dev, srq, udata);
1929 		if (status)
1930 			goto arm_err;
1931 	}
1932 
1933 	return &srq->ibsrq;
1934 
1935 arm_err:
1936 	ocrdma_mbx_destroy_srq(dev, srq);
1937 err:
1938 	kfree(srq->rqe_wr_id_tbl);
1939 	kfree(srq->idx_bit_fields);
1940 	kfree(srq);
1941 	return ERR_PTR(status);
1942 }
1943 
1944 int ocrdma_modify_srq(struct ib_srq *ibsrq,
1945 		      struct ib_srq_attr *srq_attr,
1946 		      enum ib_srq_attr_mask srq_attr_mask,
1947 		      struct ib_udata *udata)
1948 {
1949 	int status;
1950 	struct ocrdma_srq *srq;
1951 
1952 	srq = get_ocrdma_srq(ibsrq);
1953 	if (srq_attr_mask & IB_SRQ_MAX_WR)
1954 		status = -EINVAL;
1955 	else
1956 		status = ocrdma_mbx_modify_srq(srq, srq_attr);
1957 	return status;
1958 }
1959 
1960 int ocrdma_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
1961 {
1962 	int status;
1963 	struct ocrdma_srq *srq;
1964 
1965 	srq = get_ocrdma_srq(ibsrq);
1966 	status = ocrdma_mbx_query_srq(srq, srq_attr);
1967 	return status;
1968 }
1969 
1970 int ocrdma_destroy_srq(struct ib_srq *ibsrq)
1971 {
1972 	int status;
1973 	struct ocrdma_srq *srq;
1974 	struct ocrdma_dev *dev = get_ocrdma_dev(ibsrq->device);
1975 
1976 	srq = get_ocrdma_srq(ibsrq);
1977 
1978 	status = ocrdma_mbx_destroy_srq(dev, srq);
1979 
1980 	if (srq->pd->uctx)
1981 		ocrdma_del_mmap(srq->pd->uctx, (u64) srq->rq.pa,
1982 				PAGE_ALIGN(srq->rq.len));
1983 
1984 	kfree(srq->idx_bit_fields);
1985 	kfree(srq->rqe_wr_id_tbl);
1986 	kfree(srq);
1987 	return status;
1988 }
1989 
1990 /* unprivileged verbs and their support functions. */
1991 static void ocrdma_build_ud_hdr(struct ocrdma_qp *qp,
1992 				struct ocrdma_hdr_wqe *hdr,
1993 				struct ib_send_wr *wr)
1994 {
1995 	struct ocrdma_ewqe_ud_hdr *ud_hdr =
1996 		(struct ocrdma_ewqe_ud_hdr *)(hdr + 1);
1997 	struct ocrdma_ah *ah = get_ocrdma_ah(ud_wr(wr)->ah);
1998 
1999 	ud_hdr->rsvd_dest_qpn = ud_wr(wr)->remote_qpn;
2000 	if (qp->qp_type == IB_QPT_GSI)
2001 		ud_hdr->qkey = qp->qkey;
2002 	else
2003 		ud_hdr->qkey = ud_wr(wr)->remote_qkey;
2004 	ud_hdr->rsvd_ahid = ah->id;
2005 	ud_hdr->hdr_type = ah->hdr_type;
2006 	if (ah->av->valid & OCRDMA_AV_VLAN_VALID)
2007 		hdr->cw |= (OCRDMA_FLAG_AH_VLAN_PR << OCRDMA_WQE_FLAGS_SHIFT);
2008 }
2009 
2010 static void ocrdma_build_sges(struct ocrdma_hdr_wqe *hdr,
2011 			      struct ocrdma_sge *sge, int num_sge,
2012 			      struct ib_sge *sg_list)
2013 {
2014 	int i;
2015 
2016 	for (i = 0; i < num_sge; i++) {
2017 		sge[i].lrkey = sg_list[i].lkey;
2018 		sge[i].addr_lo = sg_list[i].addr;
2019 		sge[i].addr_hi = upper_32_bits(sg_list[i].addr);
2020 		sge[i].len = sg_list[i].length;
2021 		hdr->total_len += sg_list[i].length;
2022 	}
2023 	if (num_sge == 0)
2024 		memset(sge, 0, sizeof(*sge));
2025 }
2026 
2027 static inline uint32_t ocrdma_sglist_len(struct ib_sge *sg_list, int num_sge)
2028 {
2029 	uint32_t total_len = 0, i;
2030 
2031 	for (i = 0; i < num_sge; i++)
2032 		total_len += sg_list[i].length;
2033 	return total_len;
2034 }
2035 
2036 
2037 static int ocrdma_build_inline_sges(struct ocrdma_qp *qp,
2038 				    struct ocrdma_hdr_wqe *hdr,
2039 				    struct ocrdma_sge *sge,
2040 				    struct ib_send_wr *wr, u32 wqe_size)
2041 {
2042 	int i;
2043 	char *dpp_addr;
2044 
2045 	if (wr->send_flags & IB_SEND_INLINE && qp->qp_type != IB_QPT_UD) {
2046 		hdr->total_len = ocrdma_sglist_len(wr->sg_list, wr->num_sge);
2047 		if (unlikely(hdr->total_len > qp->max_inline_data)) {
2048 			pr_err("%s() supported_len=0x%x,\n"
2049 			       " unsupported len req=0x%x\n", __func__,
2050 				qp->max_inline_data, hdr->total_len);
2051 			return -EINVAL;
2052 		}
2053 		dpp_addr = (char *)sge;
2054 		for (i = 0; i < wr->num_sge; i++) {
2055 			memcpy(dpp_addr,
2056 			       (void *)(unsigned long)wr->sg_list[i].addr,
2057 			       wr->sg_list[i].length);
2058 			dpp_addr += wr->sg_list[i].length;
2059 		}
2060 
2061 		wqe_size += roundup(hdr->total_len, OCRDMA_WQE_ALIGN_BYTES);
2062 		if (0 == hdr->total_len)
2063 			wqe_size += sizeof(struct ocrdma_sge);
2064 		hdr->cw |= (OCRDMA_TYPE_INLINE << OCRDMA_WQE_TYPE_SHIFT);
2065 	} else {
2066 		ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
2067 		if (wr->num_sge)
2068 			wqe_size += (wr->num_sge * sizeof(struct ocrdma_sge));
2069 		else
2070 			wqe_size += sizeof(struct ocrdma_sge);
2071 		hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2072 	}
2073 	hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2074 	return 0;
2075 }
2076 
2077 static int ocrdma_build_send(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2078 			     struct ib_send_wr *wr)
2079 {
2080 	int status;
2081 	struct ocrdma_sge *sge;
2082 	u32 wqe_size = sizeof(*hdr);
2083 
2084 	if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2085 		ocrdma_build_ud_hdr(qp, hdr, wr);
2086 		sge = (struct ocrdma_sge *)(hdr + 2);
2087 		wqe_size += sizeof(struct ocrdma_ewqe_ud_hdr);
2088 	} else {
2089 		sge = (struct ocrdma_sge *)(hdr + 1);
2090 	}
2091 
2092 	status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
2093 	return status;
2094 }
2095 
2096 static int ocrdma_build_write(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2097 			      struct ib_send_wr *wr)
2098 {
2099 	int status;
2100 	struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
2101 	struct ocrdma_sge *sge = ext_rw + 1;
2102 	u32 wqe_size = sizeof(*hdr) + sizeof(*ext_rw);
2103 
2104 	status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
2105 	if (status)
2106 		return status;
2107 	ext_rw->addr_lo = rdma_wr(wr)->remote_addr;
2108 	ext_rw->addr_hi = upper_32_bits(rdma_wr(wr)->remote_addr);
2109 	ext_rw->lrkey = rdma_wr(wr)->rkey;
2110 	ext_rw->len = hdr->total_len;
2111 	return 0;
2112 }
2113 
2114 static void ocrdma_build_read(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2115 			      struct ib_send_wr *wr)
2116 {
2117 	struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
2118 	struct ocrdma_sge *sge = ext_rw + 1;
2119 	u32 wqe_size = ((wr->num_sge + 1) * sizeof(struct ocrdma_sge)) +
2120 	    sizeof(struct ocrdma_hdr_wqe);
2121 
2122 	ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
2123 	hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2124 	hdr->cw |= (OCRDMA_READ << OCRDMA_WQE_OPCODE_SHIFT);
2125 	hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2126 
2127 	ext_rw->addr_lo = rdma_wr(wr)->remote_addr;
2128 	ext_rw->addr_hi = upper_32_bits(rdma_wr(wr)->remote_addr);
2129 	ext_rw->lrkey = rdma_wr(wr)->rkey;
2130 	ext_rw->len = hdr->total_len;
2131 }
2132 
2133 static int get_encoded_page_size(int pg_sz)
2134 {
2135 	/* Max size is 256M 4096 << 16 */
2136 	int i = 0;
2137 	for (; i < 17; i++)
2138 		if (pg_sz == (4096 << i))
2139 			break;
2140 	return i;
2141 }
2142 
2143 static int ocrdma_build_reg(struct ocrdma_qp *qp,
2144 			    struct ocrdma_hdr_wqe *hdr,
2145 			    struct ib_reg_wr *wr)
2146 {
2147 	u64 fbo;
2148 	struct ocrdma_ewqe_fr *fast_reg = (struct ocrdma_ewqe_fr *)(hdr + 1);
2149 	struct ocrdma_mr *mr = get_ocrdma_mr(wr->mr);
2150 	struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
2151 	struct ocrdma_pbe *pbe;
2152 	u32 wqe_size = sizeof(*fast_reg) + sizeof(*hdr);
2153 	int num_pbes = 0, i;
2154 
2155 	wqe_size = roundup(wqe_size, OCRDMA_WQE_ALIGN_BYTES);
2156 
2157 	hdr->cw |= (OCRDMA_FR_MR << OCRDMA_WQE_OPCODE_SHIFT);
2158 	hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2159 
2160 	if (wr->access & IB_ACCESS_LOCAL_WRITE)
2161 		hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_LOCAL_WR;
2162 	if (wr->access & IB_ACCESS_REMOTE_WRITE)
2163 		hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_WR;
2164 	if (wr->access & IB_ACCESS_REMOTE_READ)
2165 		hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_RD;
2166 	hdr->lkey = wr->key;
2167 	hdr->total_len = mr->ibmr.length;
2168 
2169 	fbo = mr->ibmr.iova - mr->pages[0];
2170 
2171 	fast_reg->va_hi = upper_32_bits(mr->ibmr.iova);
2172 	fast_reg->va_lo = (u32) (mr->ibmr.iova & 0xffffffff);
2173 	fast_reg->fbo_hi = upper_32_bits(fbo);
2174 	fast_reg->fbo_lo = (u32) fbo & 0xffffffff;
2175 	fast_reg->num_sges = mr->npages;
2176 	fast_reg->size_sge = get_encoded_page_size(mr->ibmr.page_size);
2177 
2178 	pbe = pbl_tbl->va;
2179 	for (i = 0; i < mr->npages; i++) {
2180 		u64 buf_addr = mr->pages[i];
2181 
2182 		pbe->pa_lo = cpu_to_le32((u32) (buf_addr & PAGE_MASK));
2183 		pbe->pa_hi = cpu_to_le32((u32) upper_32_bits(buf_addr));
2184 		num_pbes += 1;
2185 		pbe++;
2186 
2187 		/* if the pbl is full storing the pbes,
2188 		 * move to next pbl.
2189 		*/
2190 		if (num_pbes == (mr->hwmr.pbl_size/sizeof(u64))) {
2191 			pbl_tbl++;
2192 			pbe = (struct ocrdma_pbe *)pbl_tbl->va;
2193 		}
2194 	}
2195 
2196 	return 0;
2197 }
2198 
2199 static void ocrdma_ring_sq_db(struct ocrdma_qp *qp)
2200 {
2201 	u32 val = qp->sq.dbid | (1 << OCRDMA_DB_SQ_SHIFT);
2202 
2203 	iowrite32(val, qp->sq_db);
2204 }
2205 
2206 int ocrdma_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2207 		     struct ib_send_wr **bad_wr)
2208 {
2209 	int status = 0;
2210 	struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2211 	struct ocrdma_hdr_wqe *hdr;
2212 	unsigned long flags;
2213 
2214 	spin_lock_irqsave(&qp->q_lock, flags);
2215 	if (qp->state != OCRDMA_QPS_RTS && qp->state != OCRDMA_QPS_SQD) {
2216 		spin_unlock_irqrestore(&qp->q_lock, flags);
2217 		*bad_wr = wr;
2218 		return -EINVAL;
2219 	}
2220 
2221 	while (wr) {
2222 		if (qp->qp_type == IB_QPT_UD &&
2223 		    (wr->opcode != IB_WR_SEND &&
2224 		     wr->opcode != IB_WR_SEND_WITH_IMM)) {
2225 			*bad_wr = wr;
2226 			status = -EINVAL;
2227 			break;
2228 		}
2229 		if (ocrdma_hwq_free_cnt(&qp->sq) == 0 ||
2230 		    wr->num_sge > qp->sq.max_sges) {
2231 			*bad_wr = wr;
2232 			status = -ENOMEM;
2233 			break;
2234 		}
2235 		hdr = ocrdma_hwq_head(&qp->sq);
2236 		hdr->cw = 0;
2237 		if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2238 			hdr->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2239 		if (wr->send_flags & IB_SEND_FENCE)
2240 			hdr->cw |=
2241 			    (OCRDMA_FLAG_FENCE_L << OCRDMA_WQE_FLAGS_SHIFT);
2242 		if (wr->send_flags & IB_SEND_SOLICITED)
2243 			hdr->cw |=
2244 			    (OCRDMA_FLAG_SOLICIT << OCRDMA_WQE_FLAGS_SHIFT);
2245 		hdr->total_len = 0;
2246 		switch (wr->opcode) {
2247 		case IB_WR_SEND_WITH_IMM:
2248 			hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2249 			hdr->immdt = ntohl(wr->ex.imm_data);
2250 		case IB_WR_SEND:
2251 			hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2252 			ocrdma_build_send(qp, hdr, wr);
2253 			break;
2254 		case IB_WR_SEND_WITH_INV:
2255 			hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
2256 			hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2257 			hdr->lkey = wr->ex.invalidate_rkey;
2258 			status = ocrdma_build_send(qp, hdr, wr);
2259 			break;
2260 		case IB_WR_RDMA_WRITE_WITH_IMM:
2261 			hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2262 			hdr->immdt = ntohl(wr->ex.imm_data);
2263 		case IB_WR_RDMA_WRITE:
2264 			hdr->cw |= (OCRDMA_WRITE << OCRDMA_WQE_OPCODE_SHIFT);
2265 			status = ocrdma_build_write(qp, hdr, wr);
2266 			break;
2267 		case IB_WR_RDMA_READ:
2268 			ocrdma_build_read(qp, hdr, wr);
2269 			break;
2270 		case IB_WR_LOCAL_INV:
2271 			hdr->cw |=
2272 			    (OCRDMA_LKEY_INV << OCRDMA_WQE_OPCODE_SHIFT);
2273 			hdr->cw |= ((sizeof(struct ocrdma_hdr_wqe) +
2274 					sizeof(struct ocrdma_sge)) /
2275 				OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT;
2276 			hdr->lkey = wr->ex.invalidate_rkey;
2277 			break;
2278 		case IB_WR_REG_MR:
2279 			status = ocrdma_build_reg(qp, hdr, reg_wr(wr));
2280 			break;
2281 		default:
2282 			status = -EINVAL;
2283 			break;
2284 		}
2285 		if (status) {
2286 			*bad_wr = wr;
2287 			break;
2288 		}
2289 		if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2290 			qp->wqe_wr_id_tbl[qp->sq.head].signaled = 1;
2291 		else
2292 			qp->wqe_wr_id_tbl[qp->sq.head].signaled = 0;
2293 		qp->wqe_wr_id_tbl[qp->sq.head].wrid = wr->wr_id;
2294 		ocrdma_cpu_to_le32(hdr, ((hdr->cw >> OCRDMA_WQE_SIZE_SHIFT) &
2295 				   OCRDMA_WQE_SIZE_MASK) * OCRDMA_WQE_STRIDE);
2296 		/* make sure wqe is written before adapter can access it */
2297 		wmb();
2298 		/* inform hw to start processing it */
2299 		ocrdma_ring_sq_db(qp);
2300 
2301 		/* update pointer, counter for next wr */
2302 		ocrdma_hwq_inc_head(&qp->sq);
2303 		wr = wr->next;
2304 	}
2305 	spin_unlock_irqrestore(&qp->q_lock, flags);
2306 	return status;
2307 }
2308 
2309 static void ocrdma_ring_rq_db(struct ocrdma_qp *qp)
2310 {
2311 	u32 val = qp->rq.dbid | (1 << OCRDMA_DB_RQ_SHIFT);
2312 
2313 	iowrite32(val, qp->rq_db);
2314 }
2315 
2316 static void ocrdma_build_rqe(struct ocrdma_hdr_wqe *rqe, struct ib_recv_wr *wr,
2317 			     u16 tag)
2318 {
2319 	u32 wqe_size = 0;
2320 	struct ocrdma_sge *sge;
2321 	if (wr->num_sge)
2322 		wqe_size = (wr->num_sge * sizeof(*sge)) + sizeof(*rqe);
2323 	else
2324 		wqe_size = sizeof(*sge) + sizeof(*rqe);
2325 
2326 	rqe->cw = ((wqe_size / OCRDMA_WQE_STRIDE) <<
2327 				OCRDMA_WQE_SIZE_SHIFT);
2328 	rqe->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2329 	rqe->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2330 	rqe->total_len = 0;
2331 	rqe->rsvd_tag = tag;
2332 	sge = (struct ocrdma_sge *)(rqe + 1);
2333 	ocrdma_build_sges(rqe, sge, wr->num_sge, wr->sg_list);
2334 	ocrdma_cpu_to_le32(rqe, wqe_size);
2335 }
2336 
2337 int ocrdma_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2338 		     struct ib_recv_wr **bad_wr)
2339 {
2340 	int status = 0;
2341 	unsigned long flags;
2342 	struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2343 	struct ocrdma_hdr_wqe *rqe;
2344 
2345 	spin_lock_irqsave(&qp->q_lock, flags);
2346 	if (qp->state == OCRDMA_QPS_RST || qp->state == OCRDMA_QPS_ERR) {
2347 		spin_unlock_irqrestore(&qp->q_lock, flags);
2348 		*bad_wr = wr;
2349 		return -EINVAL;
2350 	}
2351 	while (wr) {
2352 		if (ocrdma_hwq_free_cnt(&qp->rq) == 0 ||
2353 		    wr->num_sge > qp->rq.max_sges) {
2354 			*bad_wr = wr;
2355 			status = -ENOMEM;
2356 			break;
2357 		}
2358 		rqe = ocrdma_hwq_head(&qp->rq);
2359 		ocrdma_build_rqe(rqe, wr, 0);
2360 
2361 		qp->rqe_wr_id_tbl[qp->rq.head] = wr->wr_id;
2362 		/* make sure rqe is written before adapter can access it */
2363 		wmb();
2364 
2365 		/* inform hw to start processing it */
2366 		ocrdma_ring_rq_db(qp);
2367 
2368 		/* update pointer, counter for next wr */
2369 		ocrdma_hwq_inc_head(&qp->rq);
2370 		wr = wr->next;
2371 	}
2372 	spin_unlock_irqrestore(&qp->q_lock, flags);
2373 	return status;
2374 }
2375 
2376 /* cqe for srq's rqe can potentially arrive out of order.
2377  * index gives the entry in the shadow table where to store
2378  * the wr_id. tag/index is returned in cqe to reference back
2379  * for a given rqe.
2380  */
2381 static int ocrdma_srq_get_idx(struct ocrdma_srq *srq)
2382 {
2383 	int row = 0;
2384 	int indx = 0;
2385 
2386 	for (row = 0; row < srq->bit_fields_len; row++) {
2387 		if (srq->idx_bit_fields[row]) {
2388 			indx = ffs(srq->idx_bit_fields[row]);
2389 			indx = (row * 32) + (indx - 1);
2390 			BUG_ON(indx >= srq->rq.max_cnt);
2391 			ocrdma_srq_toggle_bit(srq, indx);
2392 			break;
2393 		}
2394 	}
2395 
2396 	BUG_ON(row == srq->bit_fields_len);
2397 	return indx + 1; /* Use from index 1 */
2398 }
2399 
2400 static void ocrdma_ring_srq_db(struct ocrdma_srq *srq)
2401 {
2402 	u32 val = srq->rq.dbid | (1 << 16);
2403 
2404 	iowrite32(val, srq->db + OCRDMA_DB_GEN2_SRQ_OFFSET);
2405 }
2406 
2407 int ocrdma_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
2408 			 struct ib_recv_wr **bad_wr)
2409 {
2410 	int status = 0;
2411 	unsigned long flags;
2412 	struct ocrdma_srq *srq;
2413 	struct ocrdma_hdr_wqe *rqe;
2414 	u16 tag;
2415 
2416 	srq = get_ocrdma_srq(ibsrq);
2417 
2418 	spin_lock_irqsave(&srq->q_lock, flags);
2419 	while (wr) {
2420 		if (ocrdma_hwq_free_cnt(&srq->rq) == 0 ||
2421 		    wr->num_sge > srq->rq.max_sges) {
2422 			status = -ENOMEM;
2423 			*bad_wr = wr;
2424 			break;
2425 		}
2426 		tag = ocrdma_srq_get_idx(srq);
2427 		rqe = ocrdma_hwq_head(&srq->rq);
2428 		ocrdma_build_rqe(rqe, wr, tag);
2429 
2430 		srq->rqe_wr_id_tbl[tag] = wr->wr_id;
2431 		/* make sure rqe is written before adapter can perform DMA */
2432 		wmb();
2433 		/* inform hw to start processing it */
2434 		ocrdma_ring_srq_db(srq);
2435 		/* update pointer, counter for next wr */
2436 		ocrdma_hwq_inc_head(&srq->rq);
2437 		wr = wr->next;
2438 	}
2439 	spin_unlock_irqrestore(&srq->q_lock, flags);
2440 	return status;
2441 }
2442 
2443 static enum ib_wc_status ocrdma_to_ibwc_err(u16 status)
2444 {
2445 	enum ib_wc_status ibwc_status;
2446 
2447 	switch (status) {
2448 	case OCRDMA_CQE_GENERAL_ERR:
2449 		ibwc_status = IB_WC_GENERAL_ERR;
2450 		break;
2451 	case OCRDMA_CQE_LOC_LEN_ERR:
2452 		ibwc_status = IB_WC_LOC_LEN_ERR;
2453 		break;
2454 	case OCRDMA_CQE_LOC_QP_OP_ERR:
2455 		ibwc_status = IB_WC_LOC_QP_OP_ERR;
2456 		break;
2457 	case OCRDMA_CQE_LOC_EEC_OP_ERR:
2458 		ibwc_status = IB_WC_LOC_EEC_OP_ERR;
2459 		break;
2460 	case OCRDMA_CQE_LOC_PROT_ERR:
2461 		ibwc_status = IB_WC_LOC_PROT_ERR;
2462 		break;
2463 	case OCRDMA_CQE_WR_FLUSH_ERR:
2464 		ibwc_status = IB_WC_WR_FLUSH_ERR;
2465 		break;
2466 	case OCRDMA_CQE_MW_BIND_ERR:
2467 		ibwc_status = IB_WC_MW_BIND_ERR;
2468 		break;
2469 	case OCRDMA_CQE_BAD_RESP_ERR:
2470 		ibwc_status = IB_WC_BAD_RESP_ERR;
2471 		break;
2472 	case OCRDMA_CQE_LOC_ACCESS_ERR:
2473 		ibwc_status = IB_WC_LOC_ACCESS_ERR;
2474 		break;
2475 	case OCRDMA_CQE_REM_INV_REQ_ERR:
2476 		ibwc_status = IB_WC_REM_INV_REQ_ERR;
2477 		break;
2478 	case OCRDMA_CQE_REM_ACCESS_ERR:
2479 		ibwc_status = IB_WC_REM_ACCESS_ERR;
2480 		break;
2481 	case OCRDMA_CQE_REM_OP_ERR:
2482 		ibwc_status = IB_WC_REM_OP_ERR;
2483 		break;
2484 	case OCRDMA_CQE_RETRY_EXC_ERR:
2485 		ibwc_status = IB_WC_RETRY_EXC_ERR;
2486 		break;
2487 	case OCRDMA_CQE_RNR_RETRY_EXC_ERR:
2488 		ibwc_status = IB_WC_RNR_RETRY_EXC_ERR;
2489 		break;
2490 	case OCRDMA_CQE_LOC_RDD_VIOL_ERR:
2491 		ibwc_status = IB_WC_LOC_RDD_VIOL_ERR;
2492 		break;
2493 	case OCRDMA_CQE_REM_INV_RD_REQ_ERR:
2494 		ibwc_status = IB_WC_REM_INV_RD_REQ_ERR;
2495 		break;
2496 	case OCRDMA_CQE_REM_ABORT_ERR:
2497 		ibwc_status = IB_WC_REM_ABORT_ERR;
2498 		break;
2499 	case OCRDMA_CQE_INV_EECN_ERR:
2500 		ibwc_status = IB_WC_INV_EECN_ERR;
2501 		break;
2502 	case OCRDMA_CQE_INV_EEC_STATE_ERR:
2503 		ibwc_status = IB_WC_INV_EEC_STATE_ERR;
2504 		break;
2505 	case OCRDMA_CQE_FATAL_ERR:
2506 		ibwc_status = IB_WC_FATAL_ERR;
2507 		break;
2508 	case OCRDMA_CQE_RESP_TIMEOUT_ERR:
2509 		ibwc_status = IB_WC_RESP_TIMEOUT_ERR;
2510 		break;
2511 	default:
2512 		ibwc_status = IB_WC_GENERAL_ERR;
2513 		break;
2514 	}
2515 	return ibwc_status;
2516 }
2517 
2518 static void ocrdma_update_wc(struct ocrdma_qp *qp, struct ib_wc *ibwc,
2519 		      u32 wqe_idx)
2520 {
2521 	struct ocrdma_hdr_wqe *hdr;
2522 	struct ocrdma_sge *rw;
2523 	int opcode;
2524 
2525 	hdr = ocrdma_hwq_head_from_idx(&qp->sq, wqe_idx);
2526 
2527 	ibwc->wr_id = qp->wqe_wr_id_tbl[wqe_idx].wrid;
2528 	/* Undo the hdr->cw swap */
2529 	opcode = le32_to_cpu(hdr->cw) & OCRDMA_WQE_OPCODE_MASK;
2530 	switch (opcode) {
2531 	case OCRDMA_WRITE:
2532 		ibwc->opcode = IB_WC_RDMA_WRITE;
2533 		break;
2534 	case OCRDMA_READ:
2535 		rw = (struct ocrdma_sge *)(hdr + 1);
2536 		ibwc->opcode = IB_WC_RDMA_READ;
2537 		ibwc->byte_len = rw->len;
2538 		break;
2539 	case OCRDMA_SEND:
2540 		ibwc->opcode = IB_WC_SEND;
2541 		break;
2542 	case OCRDMA_FR_MR:
2543 		ibwc->opcode = IB_WC_REG_MR;
2544 		break;
2545 	case OCRDMA_LKEY_INV:
2546 		ibwc->opcode = IB_WC_LOCAL_INV;
2547 		break;
2548 	default:
2549 		ibwc->status = IB_WC_GENERAL_ERR;
2550 		pr_err("%s() invalid opcode received = 0x%x\n",
2551 		       __func__, hdr->cw & OCRDMA_WQE_OPCODE_MASK);
2552 		break;
2553 	}
2554 }
2555 
2556 static void ocrdma_set_cqe_status_flushed(struct ocrdma_qp *qp,
2557 						struct ocrdma_cqe *cqe)
2558 {
2559 	if (is_cqe_for_sq(cqe)) {
2560 		cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2561 				cqe->flags_status_srcqpn) &
2562 					~OCRDMA_CQE_STATUS_MASK);
2563 		cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2564 				cqe->flags_status_srcqpn) |
2565 				(OCRDMA_CQE_WR_FLUSH_ERR <<
2566 					OCRDMA_CQE_STATUS_SHIFT));
2567 	} else {
2568 		if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2569 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2570 					cqe->flags_status_srcqpn) &
2571 						~OCRDMA_CQE_UD_STATUS_MASK);
2572 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2573 					cqe->flags_status_srcqpn) |
2574 					(OCRDMA_CQE_WR_FLUSH_ERR <<
2575 						OCRDMA_CQE_UD_STATUS_SHIFT));
2576 		} else {
2577 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2578 					cqe->flags_status_srcqpn) &
2579 						~OCRDMA_CQE_STATUS_MASK);
2580 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2581 					cqe->flags_status_srcqpn) |
2582 					(OCRDMA_CQE_WR_FLUSH_ERR <<
2583 						OCRDMA_CQE_STATUS_SHIFT));
2584 		}
2585 	}
2586 }
2587 
2588 static bool ocrdma_update_err_cqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2589 				  struct ocrdma_qp *qp, int status)
2590 {
2591 	bool expand = false;
2592 
2593 	ibwc->byte_len = 0;
2594 	ibwc->qp = &qp->ibqp;
2595 	ibwc->status = ocrdma_to_ibwc_err(status);
2596 
2597 	ocrdma_flush_qp(qp);
2598 	ocrdma_qp_state_change(qp, IB_QPS_ERR, NULL);
2599 
2600 	/* if wqe/rqe pending for which cqe needs to be returned,
2601 	 * trigger inflating it.
2602 	 */
2603 	if (!is_hw_rq_empty(qp) || !is_hw_sq_empty(qp)) {
2604 		expand = true;
2605 		ocrdma_set_cqe_status_flushed(qp, cqe);
2606 	}
2607 	return expand;
2608 }
2609 
2610 static int ocrdma_update_err_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2611 				  struct ocrdma_qp *qp, int status)
2612 {
2613 	ibwc->opcode = IB_WC_RECV;
2614 	ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2615 	ocrdma_hwq_inc_tail(&qp->rq);
2616 
2617 	return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2618 }
2619 
2620 static int ocrdma_update_err_scqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2621 				  struct ocrdma_qp *qp, int status)
2622 {
2623 	ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2624 	ocrdma_hwq_inc_tail(&qp->sq);
2625 
2626 	return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2627 }
2628 
2629 
2630 static bool ocrdma_poll_err_scqe(struct ocrdma_qp *qp,
2631 				 struct ocrdma_cqe *cqe, struct ib_wc *ibwc,
2632 				 bool *polled, bool *stop)
2633 {
2634 	bool expand;
2635 	struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2636 	int status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2637 		OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2638 	if (status < OCRDMA_MAX_CQE_ERR)
2639 		atomic_inc(&dev->cqe_err_stats[status]);
2640 
2641 	/* when hw sq is empty, but rq is not empty, so we continue
2642 	 * to keep the cqe in order to get the cq event again.
2643 	 */
2644 	if (is_hw_sq_empty(qp) && !is_hw_rq_empty(qp)) {
2645 		/* when cq for rq and sq is same, it is safe to return
2646 		 * flush cqe for RQEs.
2647 		 */
2648 		if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2649 			*polled = true;
2650 			status = OCRDMA_CQE_WR_FLUSH_ERR;
2651 			expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2652 		} else {
2653 			/* stop processing further cqe as this cqe is used for
2654 			 * triggering cq event on buddy cq of RQ.
2655 			 * When QP is destroyed, this cqe will be removed
2656 			 * from the cq's hardware q.
2657 			 */
2658 			*polled = false;
2659 			*stop = true;
2660 			expand = false;
2661 		}
2662 	} else if (is_hw_sq_empty(qp)) {
2663 		/* Do nothing */
2664 		expand = false;
2665 		*polled = false;
2666 		*stop = false;
2667 	} else {
2668 		*polled = true;
2669 		expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2670 	}
2671 	return expand;
2672 }
2673 
2674 static bool ocrdma_poll_success_scqe(struct ocrdma_qp *qp,
2675 				     struct ocrdma_cqe *cqe,
2676 				     struct ib_wc *ibwc, bool *polled)
2677 {
2678 	bool expand = false;
2679 	int tail = qp->sq.tail;
2680 	u32 wqe_idx;
2681 
2682 	if (!qp->wqe_wr_id_tbl[tail].signaled) {
2683 		*polled = false;    /* WC cannot be consumed yet */
2684 	} else {
2685 		ibwc->status = IB_WC_SUCCESS;
2686 		ibwc->wc_flags = 0;
2687 		ibwc->qp = &qp->ibqp;
2688 		ocrdma_update_wc(qp, ibwc, tail);
2689 		*polled = true;
2690 	}
2691 	wqe_idx = (le32_to_cpu(cqe->wq.wqeidx) &
2692 			OCRDMA_CQE_WQEIDX_MASK) & qp->sq.max_wqe_idx;
2693 	if (tail != wqe_idx)
2694 		expand = true; /* Coalesced CQE can't be consumed yet */
2695 
2696 	ocrdma_hwq_inc_tail(&qp->sq);
2697 	return expand;
2698 }
2699 
2700 static bool ocrdma_poll_scqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2701 			     struct ib_wc *ibwc, bool *polled, bool *stop)
2702 {
2703 	int status;
2704 	bool expand;
2705 
2706 	status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2707 		OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2708 
2709 	if (status == OCRDMA_CQE_SUCCESS)
2710 		expand = ocrdma_poll_success_scqe(qp, cqe, ibwc, polled);
2711 	else
2712 		expand = ocrdma_poll_err_scqe(qp, cqe, ibwc, polled, stop);
2713 	return expand;
2714 }
2715 
2716 static int ocrdma_update_ud_rcqe(struct ocrdma_dev *dev, struct ib_wc *ibwc,
2717 				 struct ocrdma_cqe *cqe)
2718 {
2719 	int status;
2720 	u16 hdr_type = 0;
2721 
2722 	status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2723 		OCRDMA_CQE_UD_STATUS_MASK) >> OCRDMA_CQE_UD_STATUS_SHIFT;
2724 	ibwc->src_qp = le32_to_cpu(cqe->flags_status_srcqpn) &
2725 						OCRDMA_CQE_SRCQP_MASK;
2726 	ibwc->pkey_index = 0;
2727 	ibwc->wc_flags = IB_WC_GRH;
2728 	ibwc->byte_len = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2729 			  OCRDMA_CQE_UD_XFER_LEN_SHIFT) &
2730 			  OCRDMA_CQE_UD_XFER_LEN_MASK;
2731 
2732 	if (ocrdma_is_udp_encap_supported(dev)) {
2733 		hdr_type = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2734 			    OCRDMA_CQE_UD_L3TYPE_SHIFT) &
2735 			    OCRDMA_CQE_UD_L3TYPE_MASK;
2736 		ibwc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
2737 		ibwc->network_hdr_type = hdr_type;
2738 	}
2739 
2740 	return status;
2741 }
2742 
2743 static void ocrdma_update_free_srq_cqe(struct ib_wc *ibwc,
2744 				       struct ocrdma_cqe *cqe,
2745 				       struct ocrdma_qp *qp)
2746 {
2747 	unsigned long flags;
2748 	struct ocrdma_srq *srq;
2749 	u32 wqe_idx;
2750 
2751 	srq = get_ocrdma_srq(qp->ibqp.srq);
2752 	wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
2753 		OCRDMA_CQE_BUFTAG_SHIFT) & srq->rq.max_wqe_idx;
2754 	BUG_ON(wqe_idx < 1);
2755 
2756 	ibwc->wr_id = srq->rqe_wr_id_tbl[wqe_idx];
2757 	spin_lock_irqsave(&srq->q_lock, flags);
2758 	ocrdma_srq_toggle_bit(srq, wqe_idx - 1);
2759 	spin_unlock_irqrestore(&srq->q_lock, flags);
2760 	ocrdma_hwq_inc_tail(&srq->rq);
2761 }
2762 
2763 static bool ocrdma_poll_err_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2764 				struct ib_wc *ibwc, bool *polled, bool *stop,
2765 				int status)
2766 {
2767 	bool expand;
2768 	struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2769 
2770 	if (status < OCRDMA_MAX_CQE_ERR)
2771 		atomic_inc(&dev->cqe_err_stats[status]);
2772 
2773 	/* when hw_rq is empty, but wq is not empty, so continue
2774 	 * to keep the cqe to get the cq event again.
2775 	 */
2776 	if (is_hw_rq_empty(qp) && !is_hw_sq_empty(qp)) {
2777 		if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2778 			*polled = true;
2779 			status = OCRDMA_CQE_WR_FLUSH_ERR;
2780 			expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2781 		} else {
2782 			*polled = false;
2783 			*stop = true;
2784 			expand = false;
2785 		}
2786 	} else if (is_hw_rq_empty(qp)) {
2787 		/* Do nothing */
2788 		expand = false;
2789 		*polled = false;
2790 		*stop = false;
2791 	} else {
2792 		*polled = true;
2793 		expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2794 	}
2795 	return expand;
2796 }
2797 
2798 static void ocrdma_poll_success_rcqe(struct ocrdma_qp *qp,
2799 				     struct ocrdma_cqe *cqe, struct ib_wc *ibwc)
2800 {
2801 	struct ocrdma_dev *dev;
2802 
2803 	dev = get_ocrdma_dev(qp->ibqp.device);
2804 	ibwc->opcode = IB_WC_RECV;
2805 	ibwc->qp = &qp->ibqp;
2806 	ibwc->status = IB_WC_SUCCESS;
2807 
2808 	if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI)
2809 		ocrdma_update_ud_rcqe(dev, ibwc, cqe);
2810 	else
2811 		ibwc->byte_len = le32_to_cpu(cqe->rq.rxlen);
2812 
2813 	if (is_cqe_imm(cqe)) {
2814 		ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2815 		ibwc->wc_flags |= IB_WC_WITH_IMM;
2816 	} else if (is_cqe_wr_imm(cqe)) {
2817 		ibwc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
2818 		ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2819 		ibwc->wc_flags |= IB_WC_WITH_IMM;
2820 	} else if (is_cqe_invalidated(cqe)) {
2821 		ibwc->ex.invalidate_rkey = le32_to_cpu(cqe->rq.lkey_immdt);
2822 		ibwc->wc_flags |= IB_WC_WITH_INVALIDATE;
2823 	}
2824 	if (qp->ibqp.srq) {
2825 		ocrdma_update_free_srq_cqe(ibwc, cqe, qp);
2826 	} else {
2827 		ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2828 		ocrdma_hwq_inc_tail(&qp->rq);
2829 	}
2830 }
2831 
2832 static bool ocrdma_poll_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2833 			     struct ib_wc *ibwc, bool *polled, bool *stop)
2834 {
2835 	int status;
2836 	bool expand = false;
2837 
2838 	ibwc->wc_flags = 0;
2839 	if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2840 		status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2841 					OCRDMA_CQE_UD_STATUS_MASK) >>
2842 					OCRDMA_CQE_UD_STATUS_SHIFT;
2843 	} else {
2844 		status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2845 			     OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2846 	}
2847 
2848 	if (status == OCRDMA_CQE_SUCCESS) {
2849 		*polled = true;
2850 		ocrdma_poll_success_rcqe(qp, cqe, ibwc);
2851 	} else {
2852 		expand = ocrdma_poll_err_rcqe(qp, cqe, ibwc, polled, stop,
2853 					      status);
2854 	}
2855 	return expand;
2856 }
2857 
2858 static void ocrdma_change_cq_phase(struct ocrdma_cq *cq, struct ocrdma_cqe *cqe,
2859 				   u16 cur_getp)
2860 {
2861 	if (cq->phase_change) {
2862 		if (cur_getp == 0)
2863 			cq->phase = (~cq->phase & OCRDMA_CQE_VALID);
2864 	} else {
2865 		/* clear valid bit */
2866 		cqe->flags_status_srcqpn = 0;
2867 	}
2868 }
2869 
2870 static int ocrdma_poll_hwcq(struct ocrdma_cq *cq, int num_entries,
2871 			    struct ib_wc *ibwc)
2872 {
2873 	u16 qpn = 0;
2874 	int i = 0;
2875 	bool expand = false;
2876 	int polled_hw_cqes = 0;
2877 	struct ocrdma_qp *qp = NULL;
2878 	struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
2879 	struct ocrdma_cqe *cqe;
2880 	u16 cur_getp; bool polled = false; bool stop = false;
2881 
2882 	cur_getp = cq->getp;
2883 	while (num_entries) {
2884 		cqe = cq->va + cur_getp;
2885 		/* check whether valid cqe or not */
2886 		if (!is_cqe_valid(cq, cqe))
2887 			break;
2888 		qpn = (le32_to_cpu(cqe->cmn.qpn) & OCRDMA_CQE_QPN_MASK);
2889 		/* ignore discarded cqe */
2890 		if (qpn == 0)
2891 			goto skip_cqe;
2892 		qp = dev->qp_tbl[qpn];
2893 		BUG_ON(qp == NULL);
2894 
2895 		if (is_cqe_for_sq(cqe)) {
2896 			expand = ocrdma_poll_scqe(qp, cqe, ibwc, &polled,
2897 						  &stop);
2898 		} else {
2899 			expand = ocrdma_poll_rcqe(qp, cqe, ibwc, &polled,
2900 						  &stop);
2901 		}
2902 		if (expand)
2903 			goto expand_cqe;
2904 		if (stop)
2905 			goto stop_cqe;
2906 		/* clear qpn to avoid duplicate processing by discard_cqe() */
2907 		cqe->cmn.qpn = 0;
2908 skip_cqe:
2909 		polled_hw_cqes += 1;
2910 		cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
2911 		ocrdma_change_cq_phase(cq, cqe, cur_getp);
2912 expand_cqe:
2913 		if (polled) {
2914 			num_entries -= 1;
2915 			i += 1;
2916 			ibwc = ibwc + 1;
2917 			polled = false;
2918 		}
2919 	}
2920 stop_cqe:
2921 	cq->getp = cur_getp;
2922 
2923 	if (polled_hw_cqes)
2924 		ocrdma_ring_cq_db(dev, cq->id, false, false, polled_hw_cqes);
2925 
2926 	return i;
2927 }
2928 
2929 /* insert error cqe if the QP's SQ or RQ's CQ matches the CQ under poll. */
2930 static int ocrdma_add_err_cqe(struct ocrdma_cq *cq, int num_entries,
2931 			      struct ocrdma_qp *qp, struct ib_wc *ibwc)
2932 {
2933 	int err_cqes = 0;
2934 
2935 	while (num_entries) {
2936 		if (is_hw_sq_empty(qp) && is_hw_rq_empty(qp))
2937 			break;
2938 		if (!is_hw_sq_empty(qp) && qp->sq_cq == cq) {
2939 			ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2940 			ocrdma_hwq_inc_tail(&qp->sq);
2941 		} else if (!is_hw_rq_empty(qp) && qp->rq_cq == cq) {
2942 			ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2943 			ocrdma_hwq_inc_tail(&qp->rq);
2944 		} else {
2945 			return err_cqes;
2946 		}
2947 		ibwc->byte_len = 0;
2948 		ibwc->status = IB_WC_WR_FLUSH_ERR;
2949 		ibwc = ibwc + 1;
2950 		err_cqes += 1;
2951 		num_entries -= 1;
2952 	}
2953 	return err_cqes;
2954 }
2955 
2956 int ocrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
2957 {
2958 	int cqes_to_poll = num_entries;
2959 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2960 	struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2961 	int num_os_cqe = 0, err_cqes = 0;
2962 	struct ocrdma_qp *qp;
2963 	unsigned long flags;
2964 
2965 	/* poll cqes from adapter CQ */
2966 	spin_lock_irqsave(&cq->cq_lock, flags);
2967 	num_os_cqe = ocrdma_poll_hwcq(cq, cqes_to_poll, wc);
2968 	spin_unlock_irqrestore(&cq->cq_lock, flags);
2969 	cqes_to_poll -= num_os_cqe;
2970 
2971 	if (cqes_to_poll) {
2972 		wc = wc + num_os_cqe;
2973 		/* adapter returns single error cqe when qp moves to
2974 		 * error state. So insert error cqes with wc_status as
2975 		 * FLUSHED for pending WQEs and RQEs of QP's SQ and RQ
2976 		 * respectively which uses this CQ.
2977 		 */
2978 		spin_lock_irqsave(&dev->flush_q_lock, flags);
2979 		list_for_each_entry(qp, &cq->sq_head, sq_entry) {
2980 			if (cqes_to_poll == 0)
2981 				break;
2982 			err_cqes = ocrdma_add_err_cqe(cq, cqes_to_poll, qp, wc);
2983 			cqes_to_poll -= err_cqes;
2984 			num_os_cqe += err_cqes;
2985 			wc = wc + err_cqes;
2986 		}
2987 		spin_unlock_irqrestore(&dev->flush_q_lock, flags);
2988 	}
2989 	return num_os_cqe;
2990 }
2991 
2992 int ocrdma_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags cq_flags)
2993 {
2994 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2995 	struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2996 	u16 cq_id;
2997 	unsigned long flags;
2998 	bool arm_needed = false, sol_needed = false;
2999 
3000 	cq_id = cq->id;
3001 
3002 	spin_lock_irqsave(&cq->cq_lock, flags);
3003 	if (cq_flags & IB_CQ_NEXT_COMP || cq_flags & IB_CQ_SOLICITED)
3004 		arm_needed = true;
3005 	if (cq_flags & IB_CQ_SOLICITED)
3006 		sol_needed = true;
3007 
3008 	ocrdma_ring_cq_db(dev, cq_id, arm_needed, sol_needed, 0);
3009 	spin_unlock_irqrestore(&cq->cq_lock, flags);
3010 
3011 	return 0;
3012 }
3013 
3014 struct ib_mr *ocrdma_alloc_mr(struct ib_pd *ibpd,
3015 			      enum ib_mr_type mr_type,
3016 			      u32 max_num_sg)
3017 {
3018 	int status;
3019 	struct ocrdma_mr *mr;
3020 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
3021 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
3022 
3023 	if (mr_type != IB_MR_TYPE_MEM_REG)
3024 		return ERR_PTR(-EINVAL);
3025 
3026 	if (max_num_sg > dev->attr.max_pages_per_frmr)
3027 		return ERR_PTR(-EINVAL);
3028 
3029 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3030 	if (!mr)
3031 		return ERR_PTR(-ENOMEM);
3032 
3033 	mr->pages = kcalloc(max_num_sg, sizeof(u64), GFP_KERNEL);
3034 	if (!mr->pages) {
3035 		status = -ENOMEM;
3036 		goto pl_err;
3037 	}
3038 
3039 	status = ocrdma_get_pbl_info(dev, mr, max_num_sg);
3040 	if (status)
3041 		goto pbl_err;
3042 	mr->hwmr.fr_mr = 1;
3043 	mr->hwmr.remote_rd = 0;
3044 	mr->hwmr.remote_wr = 0;
3045 	mr->hwmr.local_rd = 0;
3046 	mr->hwmr.local_wr = 0;
3047 	mr->hwmr.mw_bind = 0;
3048 	status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
3049 	if (status)
3050 		goto pbl_err;
3051 	status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, 0);
3052 	if (status)
3053 		goto mbx_err;
3054 	mr->ibmr.rkey = mr->hwmr.lkey;
3055 	mr->ibmr.lkey = mr->hwmr.lkey;
3056 	dev->stag_arr[(mr->hwmr.lkey >> 8) & (OCRDMA_MAX_STAG - 1)] =
3057 		(unsigned long) mr;
3058 	return &mr->ibmr;
3059 mbx_err:
3060 	ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
3061 pbl_err:
3062 	kfree(mr->pages);
3063 pl_err:
3064 	kfree(mr);
3065 	return ERR_PTR(-ENOMEM);
3066 }
3067 
3068 static int ocrdma_set_page(struct ib_mr *ibmr, u64 addr)
3069 {
3070 	struct ocrdma_mr *mr = get_ocrdma_mr(ibmr);
3071 
3072 	if (unlikely(mr->npages == mr->hwmr.num_pbes))
3073 		return -ENOMEM;
3074 
3075 	mr->pages[mr->npages++] = addr;
3076 
3077 	return 0;
3078 }
3079 
3080 int ocrdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
3081 		     unsigned int *sg_offset)
3082 {
3083 	struct ocrdma_mr *mr = get_ocrdma_mr(ibmr);
3084 
3085 	mr->npages = 0;
3086 
3087 	return ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, ocrdma_set_page);
3088 }
3089