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