1 /* This file is part of the Emulex RoCE Device Driver for
2  * RoCE (RDMA over Converged Ethernet) adapters.
3  * Copyright (C) 2012-2015 Emulex. All rights reserved.
4  * EMULEX and SLI are trademarks of Emulex.
5  * www.emulex.com
6  *
7  * This software is available to you under a choice of one of two licenses.
8  * You may choose to be licensed under the terms of the GNU General Public
9  * License (GPL) Version 2, available from the file COPYING in the main
10  * directory of this source tree, or the BSD license below:
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  *
16  * - Redistributions of source code must retain the above copyright notice,
17  *   this list of conditions and the following disclaimer.
18  *
19  * - Redistributions in binary form must reproduce the above copyright
20  *   notice, this list of conditions and the following disclaimer in
21  *   the documentation and/or other materials provided with the distribution.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
24  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
27  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
28  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
29  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
30  * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
31  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
32  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
33  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
34  *
35  * Contact Information:
36  * linux-drivers@emulex.com
37  *
38  * Emulex
39  * 3333 Susan Street
40  * Costa Mesa, CA 92626
41  */
42 
43 #include <linux/dma-mapping.h>
44 #include <rdma/ib_verbs.h>
45 #include <rdma/ib_user_verbs.h>
46 #include <rdma/iw_cm.h>
47 #include <rdma/ib_umem.h>
48 #include <rdma/ib_addr.h>
49 #include <rdma/ib_cache.h>
50 #include <rdma/uverbs_ioctl.h>
51 
52 #include "ocrdma.h"
53 #include "ocrdma_hw.h"
54 #include "ocrdma_verbs.h"
55 #include <rdma/ocrdma-abi.h>
56 
57 int ocrdma_query_pkey(struct ib_device *ibdev, u8 port, u16 index, u16 *pkey)
58 {
59 	if (index > 0)
60 		return -EINVAL;
61 
62 	*pkey = 0xffff;
63 	return 0;
64 }
65 
66 int ocrdma_query_device(struct ib_device *ibdev, struct ib_device_attr *attr,
67 			struct ib_udata *uhw)
68 {
69 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
70 
71 	if (uhw->inlen || uhw->outlen)
72 		return -EINVAL;
73 
74 	memset(attr, 0, sizeof *attr);
75 	memcpy(&attr->fw_ver, &dev->attr.fw_ver[0],
76 	       min(sizeof(dev->attr.fw_ver), sizeof(attr->fw_ver)));
77 	ocrdma_get_guid(dev, (u8 *)&attr->sys_image_guid);
78 	attr->max_mr_size = dev->attr.max_mr_size;
79 	attr->page_size_cap = 0xffff000;
80 	attr->vendor_id = dev->nic_info.pdev->vendor;
81 	attr->vendor_part_id = dev->nic_info.pdev->device;
82 	attr->hw_ver = dev->asic_id;
83 	attr->max_qp = dev->attr.max_qp;
84 	attr->max_ah = OCRDMA_MAX_AH;
85 	attr->max_qp_wr = dev->attr.max_wqe;
86 
87 	attr->device_cap_flags = IB_DEVICE_CURR_QP_STATE_MOD |
88 					IB_DEVICE_RC_RNR_NAK_GEN |
89 					IB_DEVICE_SHUTDOWN_PORT |
90 					IB_DEVICE_SYS_IMAGE_GUID |
91 					IB_DEVICE_LOCAL_DMA_LKEY |
92 					IB_DEVICE_MEM_MGT_EXTENSIONS;
93 	attr->max_send_sge = dev->attr.max_send_sge;
94 	attr->max_recv_sge = dev->attr.max_recv_sge;
95 	attr->max_sge_rd = dev->attr.max_rdma_sge;
96 	attr->max_cq = dev->attr.max_cq;
97 	attr->max_cqe = dev->attr.max_cqe;
98 	attr->max_mr = dev->attr.max_mr;
99 	attr->max_mw = dev->attr.max_mw;
100 	attr->max_pd = dev->attr.max_pd;
101 	attr->atomic_cap = 0;
102 	attr->max_fmr = 0;
103 	attr->max_map_per_fmr = 0;
104 	attr->max_qp_rd_atom =
105 	    min(dev->attr.max_ord_per_qp, dev->attr.max_ird_per_qp);
106 	attr->max_qp_init_rd_atom = dev->attr.max_ord_per_qp;
107 	attr->max_srq = dev->attr.max_srq;
108 	attr->max_srq_sge = dev->attr.max_srq_sge;
109 	attr->max_srq_wr = dev->attr.max_rqe;
110 	attr->local_ca_ack_delay = dev->attr.local_ca_ack_delay;
111 	attr->max_fast_reg_page_list_len = dev->attr.max_pages_per_frmr;
112 	attr->max_pkeys = 1;
113 	return 0;
114 }
115 
116 static inline void get_link_speed_and_width(struct ocrdma_dev *dev,
117 					    u8 *ib_speed, u8 *ib_width)
118 {
119 	int status;
120 	u8 speed;
121 
122 	status = ocrdma_mbx_get_link_speed(dev, &speed, NULL);
123 	if (status)
124 		speed = OCRDMA_PHYS_LINK_SPEED_ZERO;
125 
126 	switch (speed) {
127 	case OCRDMA_PHYS_LINK_SPEED_1GBPS:
128 		*ib_speed = IB_SPEED_SDR;
129 		*ib_width = IB_WIDTH_1X;
130 		break;
131 
132 	case OCRDMA_PHYS_LINK_SPEED_10GBPS:
133 		*ib_speed = IB_SPEED_QDR;
134 		*ib_width = IB_WIDTH_1X;
135 		break;
136 
137 	case OCRDMA_PHYS_LINK_SPEED_20GBPS:
138 		*ib_speed = IB_SPEED_DDR;
139 		*ib_width = IB_WIDTH_4X;
140 		break;
141 
142 	case OCRDMA_PHYS_LINK_SPEED_40GBPS:
143 		*ib_speed = IB_SPEED_QDR;
144 		*ib_width = IB_WIDTH_4X;
145 		break;
146 
147 	default:
148 		/* Unsupported */
149 		*ib_speed = IB_SPEED_SDR;
150 		*ib_width = IB_WIDTH_1X;
151 	}
152 }
153 
154 int ocrdma_query_port(struct ib_device *ibdev,
155 		      u8 port, struct ib_port_attr *props)
156 {
157 	enum ib_port_state port_state;
158 	struct ocrdma_dev *dev;
159 	struct net_device *netdev;
160 
161 	/* props being zeroed by the caller, avoid zeroing it here */
162 	dev = get_ocrdma_dev(ibdev);
163 	netdev = dev->nic_info.netdev;
164 	if (netif_running(netdev) && netif_oper_up(netdev)) {
165 		port_state = IB_PORT_ACTIVE;
166 		props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
167 	} else {
168 		port_state = IB_PORT_DOWN;
169 		props->phys_state = IB_PORT_PHYS_STATE_DISABLED;
170 	}
171 	props->max_mtu = IB_MTU_4096;
172 	props->active_mtu = iboe_get_mtu(netdev->mtu);
173 	props->lid = 0;
174 	props->lmc = 0;
175 	props->sm_lid = 0;
176 	props->sm_sl = 0;
177 	props->state = port_state;
178 	props->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_REINIT_SUP |
179 				IB_PORT_DEVICE_MGMT_SUP |
180 				IB_PORT_VENDOR_CLASS_SUP;
181 	props->ip_gids = true;
182 	props->gid_tbl_len = OCRDMA_MAX_SGID;
183 	props->pkey_tbl_len = 1;
184 	props->bad_pkey_cntr = 0;
185 	props->qkey_viol_cntr = 0;
186 	get_link_speed_and_width(dev, &props->active_speed,
187 				 &props->active_width);
188 	props->max_msg_sz = 0x80000000;
189 	props->max_vl_num = 4;
190 	return 0;
191 }
192 
193 static int ocrdma_add_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
194 			   unsigned long len)
195 {
196 	struct ocrdma_mm *mm;
197 
198 	mm = kzalloc(sizeof(*mm), GFP_KERNEL);
199 	if (mm == NULL)
200 		return -ENOMEM;
201 	mm->key.phy_addr = phy_addr;
202 	mm->key.len = len;
203 	INIT_LIST_HEAD(&mm->entry);
204 
205 	mutex_lock(&uctx->mm_list_lock);
206 	list_add_tail(&mm->entry, &uctx->mm_head);
207 	mutex_unlock(&uctx->mm_list_lock);
208 	return 0;
209 }
210 
211 static void ocrdma_del_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
212 			    unsigned long len)
213 {
214 	struct ocrdma_mm *mm, *tmp;
215 
216 	mutex_lock(&uctx->mm_list_lock);
217 	list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
218 		if (len != mm->key.len && phy_addr != mm->key.phy_addr)
219 			continue;
220 
221 		list_del(&mm->entry);
222 		kfree(mm);
223 		break;
224 	}
225 	mutex_unlock(&uctx->mm_list_lock);
226 }
227 
228 static bool ocrdma_search_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
229 			      unsigned long len)
230 {
231 	bool found = false;
232 	struct ocrdma_mm *mm;
233 
234 	mutex_lock(&uctx->mm_list_lock);
235 	list_for_each_entry(mm, &uctx->mm_head, entry) {
236 		if (len != mm->key.len && phy_addr != mm->key.phy_addr)
237 			continue;
238 
239 		found = true;
240 		break;
241 	}
242 	mutex_unlock(&uctx->mm_list_lock);
243 	return found;
244 }
245 
246 
247 static u16 _ocrdma_pd_mgr_get_bitmap(struct ocrdma_dev *dev, bool dpp_pool)
248 {
249 	u16 pd_bitmap_idx = 0;
250 	const unsigned long *pd_bitmap;
251 
252 	if (dpp_pool) {
253 		pd_bitmap = dev->pd_mgr->pd_dpp_bitmap;
254 		pd_bitmap_idx = find_first_zero_bit(pd_bitmap,
255 						    dev->pd_mgr->max_dpp_pd);
256 		__set_bit(pd_bitmap_idx, dev->pd_mgr->pd_dpp_bitmap);
257 		dev->pd_mgr->pd_dpp_count++;
258 		if (dev->pd_mgr->pd_dpp_count > dev->pd_mgr->pd_dpp_thrsh)
259 			dev->pd_mgr->pd_dpp_thrsh = dev->pd_mgr->pd_dpp_count;
260 	} else {
261 		pd_bitmap = dev->pd_mgr->pd_norm_bitmap;
262 		pd_bitmap_idx = find_first_zero_bit(pd_bitmap,
263 						    dev->pd_mgr->max_normal_pd);
264 		__set_bit(pd_bitmap_idx, dev->pd_mgr->pd_norm_bitmap);
265 		dev->pd_mgr->pd_norm_count++;
266 		if (dev->pd_mgr->pd_norm_count > dev->pd_mgr->pd_norm_thrsh)
267 			dev->pd_mgr->pd_norm_thrsh = dev->pd_mgr->pd_norm_count;
268 	}
269 	return pd_bitmap_idx;
270 }
271 
272 static int _ocrdma_pd_mgr_put_bitmap(struct ocrdma_dev *dev, u16 pd_id,
273 					bool dpp_pool)
274 {
275 	u16 pd_count;
276 	u16 pd_bit_index;
277 
278 	pd_count = dpp_pool ? dev->pd_mgr->pd_dpp_count :
279 			      dev->pd_mgr->pd_norm_count;
280 	if (pd_count == 0)
281 		return -EINVAL;
282 
283 	if (dpp_pool) {
284 		pd_bit_index = pd_id - dev->pd_mgr->pd_dpp_start;
285 		if (pd_bit_index >= dev->pd_mgr->max_dpp_pd) {
286 			return -EINVAL;
287 		} else {
288 			__clear_bit(pd_bit_index, dev->pd_mgr->pd_dpp_bitmap);
289 			dev->pd_mgr->pd_dpp_count--;
290 		}
291 	} else {
292 		pd_bit_index = pd_id - dev->pd_mgr->pd_norm_start;
293 		if (pd_bit_index >= dev->pd_mgr->max_normal_pd) {
294 			return -EINVAL;
295 		} else {
296 			__clear_bit(pd_bit_index, dev->pd_mgr->pd_norm_bitmap);
297 			dev->pd_mgr->pd_norm_count--;
298 		}
299 	}
300 
301 	return 0;
302 }
303 
304 static int ocrdma_put_pd_num(struct ocrdma_dev *dev, u16 pd_id,
305 				   bool dpp_pool)
306 {
307 	int status;
308 
309 	mutex_lock(&dev->dev_lock);
310 	status = _ocrdma_pd_mgr_put_bitmap(dev, pd_id, dpp_pool);
311 	mutex_unlock(&dev->dev_lock);
312 	return status;
313 }
314 
315 static int ocrdma_get_pd_num(struct ocrdma_dev *dev, struct ocrdma_pd *pd)
316 {
317 	u16 pd_idx = 0;
318 	int status = 0;
319 
320 	mutex_lock(&dev->dev_lock);
321 	if (pd->dpp_enabled) {
322 		/* try allocating DPP PD, if not available then normal PD */
323 		if (dev->pd_mgr->pd_dpp_count < dev->pd_mgr->max_dpp_pd) {
324 			pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, true);
325 			pd->id = dev->pd_mgr->pd_dpp_start + pd_idx;
326 			pd->dpp_page = dev->pd_mgr->dpp_page_index + pd_idx;
327 		} else if (dev->pd_mgr->pd_norm_count <
328 			   dev->pd_mgr->max_normal_pd) {
329 			pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, false);
330 			pd->id = dev->pd_mgr->pd_norm_start + pd_idx;
331 			pd->dpp_enabled = false;
332 		} else {
333 			status = -EINVAL;
334 		}
335 	} else {
336 		if (dev->pd_mgr->pd_norm_count < dev->pd_mgr->max_normal_pd) {
337 			pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, false);
338 			pd->id = dev->pd_mgr->pd_norm_start + pd_idx;
339 		} else {
340 			status = -EINVAL;
341 		}
342 	}
343 	mutex_unlock(&dev->dev_lock);
344 	return status;
345 }
346 
347 /*
348  * NOTE:
349  *
350  * ocrdma_ucontext must be used here because this function is also
351  * called from ocrdma_alloc_ucontext where ib_udata does not have
352  * valid ib_ucontext pointer. ib_uverbs_get_context does not call
353  * uobj_{alloc|get_xxx} helpers which are used to store the
354  * ib_ucontext in uverbs_attr_bundle wrapping the ib_udata. so
355  * ib_udata does NOT imply valid ib_ucontext here!
356  */
357 static int _ocrdma_alloc_pd(struct ocrdma_dev *dev, struct ocrdma_pd *pd,
358 			    struct ocrdma_ucontext *uctx,
359 			    struct ib_udata *udata)
360 {
361 	int status;
362 
363 	if (udata && uctx && dev->attr.max_dpp_pds) {
364 		pd->dpp_enabled =
365 			ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R;
366 		pd->num_dpp_qp =
367 			pd->dpp_enabled ? (dev->nic_info.db_page_size /
368 					   dev->attr.wqe_size) : 0;
369 	}
370 
371 	if (dev->pd_mgr->pd_prealloc_valid)
372 		return ocrdma_get_pd_num(dev, pd);
373 
374 retry:
375 	status = ocrdma_mbx_alloc_pd(dev, pd);
376 	if (status) {
377 		if (pd->dpp_enabled) {
378 			pd->dpp_enabled = false;
379 			pd->num_dpp_qp = 0;
380 			goto retry;
381 		}
382 		return status;
383 	}
384 
385 	return 0;
386 }
387 
388 static inline int is_ucontext_pd(struct ocrdma_ucontext *uctx,
389 				 struct ocrdma_pd *pd)
390 {
391 	return (uctx->cntxt_pd == pd);
392 }
393 
394 static void _ocrdma_dealloc_pd(struct ocrdma_dev *dev,
395 			      struct ocrdma_pd *pd)
396 {
397 	if (dev->pd_mgr->pd_prealloc_valid)
398 		ocrdma_put_pd_num(dev, pd->id, pd->dpp_enabled);
399 	else
400 		ocrdma_mbx_dealloc_pd(dev, pd);
401 }
402 
403 static int ocrdma_alloc_ucontext_pd(struct ocrdma_dev *dev,
404 				    struct ocrdma_ucontext *uctx,
405 				    struct ib_udata *udata)
406 {
407 	struct ib_device *ibdev = &dev->ibdev;
408 	struct ib_pd *pd;
409 	int status;
410 
411 	pd = rdma_zalloc_drv_obj(ibdev, ib_pd);
412 	if (!pd)
413 		return -ENOMEM;
414 
415 	pd->device  = ibdev;
416 	uctx->cntxt_pd = get_ocrdma_pd(pd);
417 
418 	status = _ocrdma_alloc_pd(dev, uctx->cntxt_pd, uctx, udata);
419 	if (status) {
420 		kfree(uctx->cntxt_pd);
421 		goto err;
422 	}
423 
424 	uctx->cntxt_pd->uctx = uctx;
425 	uctx->cntxt_pd->ibpd.device = &dev->ibdev;
426 err:
427 	return status;
428 }
429 
430 static void ocrdma_dealloc_ucontext_pd(struct ocrdma_ucontext *uctx)
431 {
432 	struct ocrdma_pd *pd = uctx->cntxt_pd;
433 	struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
434 
435 	if (uctx->pd_in_use) {
436 		pr_err("%s(%d) Freeing in use pdid=0x%x.\n",
437 		       __func__, dev->id, pd->id);
438 	}
439 	kfree(uctx->cntxt_pd);
440 	uctx->cntxt_pd = NULL;
441 	_ocrdma_dealloc_pd(dev, pd);
442 }
443 
444 static struct ocrdma_pd *ocrdma_get_ucontext_pd(struct ocrdma_ucontext *uctx)
445 {
446 	struct ocrdma_pd *pd = NULL;
447 
448 	mutex_lock(&uctx->mm_list_lock);
449 	if (!uctx->pd_in_use) {
450 		uctx->pd_in_use = true;
451 		pd = uctx->cntxt_pd;
452 	}
453 	mutex_unlock(&uctx->mm_list_lock);
454 
455 	return pd;
456 }
457 
458 static void ocrdma_release_ucontext_pd(struct ocrdma_ucontext *uctx)
459 {
460 	mutex_lock(&uctx->mm_list_lock);
461 	uctx->pd_in_use = false;
462 	mutex_unlock(&uctx->mm_list_lock);
463 }
464 
465 int ocrdma_alloc_ucontext(struct ib_ucontext *uctx, struct ib_udata *udata)
466 {
467 	struct ib_device *ibdev = uctx->device;
468 	int status;
469 	struct ocrdma_ucontext *ctx = get_ocrdma_ucontext(uctx);
470 	struct ocrdma_alloc_ucontext_resp resp = {};
471 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
472 	struct pci_dev *pdev = dev->nic_info.pdev;
473 	u32 map_len = roundup(sizeof(u32) * 2048, PAGE_SIZE);
474 
475 	if (!udata)
476 		return -EFAULT;
477 	INIT_LIST_HEAD(&ctx->mm_head);
478 	mutex_init(&ctx->mm_list_lock);
479 
480 	ctx->ah_tbl.va = dma_alloc_coherent(&pdev->dev, map_len,
481 					    &ctx->ah_tbl.pa, GFP_KERNEL);
482 	if (!ctx->ah_tbl.va)
483 		return -ENOMEM;
484 
485 	ctx->ah_tbl.len = map_len;
486 
487 	resp.ah_tbl_len = ctx->ah_tbl.len;
488 	resp.ah_tbl_page = virt_to_phys(ctx->ah_tbl.va);
489 
490 	status = ocrdma_add_mmap(ctx, resp.ah_tbl_page, resp.ah_tbl_len);
491 	if (status)
492 		goto map_err;
493 
494 	status = ocrdma_alloc_ucontext_pd(dev, ctx, udata);
495 	if (status)
496 		goto pd_err;
497 
498 	resp.dev_id = dev->id;
499 	resp.max_inline_data = dev->attr.max_inline_data;
500 	resp.wqe_size = dev->attr.wqe_size;
501 	resp.rqe_size = dev->attr.rqe_size;
502 	resp.dpp_wqe_size = dev->attr.wqe_size;
503 
504 	memcpy(resp.fw_ver, dev->attr.fw_ver, sizeof(resp.fw_ver));
505 	status = ib_copy_to_udata(udata, &resp, sizeof(resp));
506 	if (status)
507 		goto cpy_err;
508 	return 0;
509 
510 cpy_err:
511 	ocrdma_dealloc_ucontext_pd(ctx);
512 pd_err:
513 	ocrdma_del_mmap(ctx, ctx->ah_tbl.pa, ctx->ah_tbl.len);
514 map_err:
515 	dma_free_coherent(&pdev->dev, ctx->ah_tbl.len, ctx->ah_tbl.va,
516 			  ctx->ah_tbl.pa);
517 	return status;
518 }
519 
520 void ocrdma_dealloc_ucontext(struct ib_ucontext *ibctx)
521 {
522 	struct ocrdma_mm *mm, *tmp;
523 	struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ibctx);
524 	struct ocrdma_dev *dev = get_ocrdma_dev(ibctx->device);
525 	struct pci_dev *pdev = dev->nic_info.pdev;
526 
527 	ocrdma_dealloc_ucontext_pd(uctx);
528 
529 	ocrdma_del_mmap(uctx, uctx->ah_tbl.pa, uctx->ah_tbl.len);
530 	dma_free_coherent(&pdev->dev, uctx->ah_tbl.len, uctx->ah_tbl.va,
531 			  uctx->ah_tbl.pa);
532 
533 	list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
534 		list_del(&mm->entry);
535 		kfree(mm);
536 	}
537 }
538 
539 int ocrdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
540 {
541 	struct ocrdma_ucontext *ucontext = get_ocrdma_ucontext(context);
542 	struct ocrdma_dev *dev = get_ocrdma_dev(context->device);
543 	unsigned long vm_page = vma->vm_pgoff << PAGE_SHIFT;
544 	u64 unmapped_db = (u64) dev->nic_info.unmapped_db;
545 	unsigned long len = (vma->vm_end - vma->vm_start);
546 	int status;
547 	bool found;
548 
549 	if (vma->vm_start & (PAGE_SIZE - 1))
550 		return -EINVAL;
551 	found = ocrdma_search_mmap(ucontext, vma->vm_pgoff << PAGE_SHIFT, len);
552 	if (!found)
553 		return -EINVAL;
554 
555 	if ((vm_page >= unmapped_db) && (vm_page <= (unmapped_db +
556 		dev->nic_info.db_total_size)) &&
557 		(len <=	dev->nic_info.db_page_size)) {
558 		if (vma->vm_flags & VM_READ)
559 			return -EPERM;
560 
561 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
562 		status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
563 					    len, vma->vm_page_prot);
564 	} else if (dev->nic_info.dpp_unmapped_len &&
565 		(vm_page >= (u64) dev->nic_info.dpp_unmapped_addr) &&
566 		(vm_page <= (u64) (dev->nic_info.dpp_unmapped_addr +
567 			dev->nic_info.dpp_unmapped_len)) &&
568 		(len <= dev->nic_info.dpp_unmapped_len)) {
569 		if (vma->vm_flags & VM_READ)
570 			return -EPERM;
571 
572 		vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
573 		status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
574 					    len, vma->vm_page_prot);
575 	} else {
576 		status = remap_pfn_range(vma, vma->vm_start,
577 					 vma->vm_pgoff, len, vma->vm_page_prot);
578 	}
579 	return status;
580 }
581 
582 static int ocrdma_copy_pd_uresp(struct ocrdma_dev *dev, struct ocrdma_pd *pd,
583 				struct ib_udata *udata)
584 {
585 	int status;
586 	u64 db_page_addr;
587 	u64 dpp_page_addr = 0;
588 	u32 db_page_size;
589 	struct ocrdma_alloc_pd_uresp rsp;
590 	struct ocrdma_ucontext *uctx = rdma_udata_to_drv_context(
591 		udata, struct ocrdma_ucontext, ibucontext);
592 
593 	memset(&rsp, 0, sizeof(rsp));
594 	rsp.id = pd->id;
595 	rsp.dpp_enabled = pd->dpp_enabled;
596 	db_page_addr = ocrdma_get_db_addr(dev, pd->id);
597 	db_page_size = dev->nic_info.db_page_size;
598 
599 	status = ocrdma_add_mmap(uctx, db_page_addr, db_page_size);
600 	if (status)
601 		return status;
602 
603 	if (pd->dpp_enabled) {
604 		dpp_page_addr = dev->nic_info.dpp_unmapped_addr +
605 				(pd->id * PAGE_SIZE);
606 		status = ocrdma_add_mmap(uctx, dpp_page_addr,
607 				 PAGE_SIZE);
608 		if (status)
609 			goto dpp_map_err;
610 		rsp.dpp_page_addr_hi = upper_32_bits(dpp_page_addr);
611 		rsp.dpp_page_addr_lo = dpp_page_addr;
612 	}
613 
614 	status = ib_copy_to_udata(udata, &rsp, sizeof(rsp));
615 	if (status)
616 		goto ucopy_err;
617 
618 	pd->uctx = uctx;
619 	return 0;
620 
621 ucopy_err:
622 	if (pd->dpp_enabled)
623 		ocrdma_del_mmap(pd->uctx, dpp_page_addr, PAGE_SIZE);
624 dpp_map_err:
625 	ocrdma_del_mmap(pd->uctx, db_page_addr, db_page_size);
626 	return status;
627 }
628 
629 int ocrdma_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
630 {
631 	struct ib_device *ibdev = ibpd->device;
632 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
633 	struct ocrdma_pd *pd;
634 	int status;
635 	u8 is_uctx_pd = false;
636 	struct ocrdma_ucontext *uctx = rdma_udata_to_drv_context(
637 		udata, struct ocrdma_ucontext, ibucontext);
638 
639 	if (udata) {
640 		pd = ocrdma_get_ucontext_pd(uctx);
641 		if (pd) {
642 			is_uctx_pd = true;
643 			goto pd_mapping;
644 		}
645 	}
646 
647 	pd = get_ocrdma_pd(ibpd);
648 	status = _ocrdma_alloc_pd(dev, pd, uctx, udata);
649 	if (status)
650 		goto exit;
651 
652 pd_mapping:
653 	if (udata) {
654 		status = ocrdma_copy_pd_uresp(dev, pd, udata);
655 		if (status)
656 			goto err;
657 	}
658 	return 0;
659 
660 err:
661 	if (is_uctx_pd)
662 		ocrdma_release_ucontext_pd(uctx);
663 	else
664 		_ocrdma_dealloc_pd(dev, pd);
665 exit:
666 	return status;
667 }
668 
669 void ocrdma_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
670 {
671 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
672 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
673 	struct ocrdma_ucontext *uctx = NULL;
674 	u64 usr_db;
675 
676 	uctx = pd->uctx;
677 	if (uctx) {
678 		u64 dpp_db = dev->nic_info.dpp_unmapped_addr +
679 			(pd->id * PAGE_SIZE);
680 		if (pd->dpp_enabled)
681 			ocrdma_del_mmap(pd->uctx, dpp_db, PAGE_SIZE);
682 		usr_db = ocrdma_get_db_addr(dev, pd->id);
683 		ocrdma_del_mmap(pd->uctx, usr_db, dev->nic_info.db_page_size);
684 
685 		if (is_ucontext_pd(uctx, pd)) {
686 			ocrdma_release_ucontext_pd(uctx);
687 			return;
688 		}
689 	}
690 	_ocrdma_dealloc_pd(dev, pd);
691 }
692 
693 static int ocrdma_alloc_lkey(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
694 			    u32 pdid, int acc, u32 num_pbls, u32 addr_check)
695 {
696 	int status;
697 
698 	mr->hwmr.fr_mr = 0;
699 	mr->hwmr.local_rd = 1;
700 	mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
701 	mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
702 	mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
703 	mr->hwmr.mw_bind = (acc & IB_ACCESS_MW_BIND) ? 1 : 0;
704 	mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
705 	mr->hwmr.num_pbls = num_pbls;
706 
707 	status = ocrdma_mbx_alloc_lkey(dev, &mr->hwmr, pdid, addr_check);
708 	if (status)
709 		return status;
710 
711 	mr->ibmr.lkey = mr->hwmr.lkey;
712 	if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
713 		mr->ibmr.rkey = mr->hwmr.lkey;
714 	return 0;
715 }
716 
717 struct ib_mr *ocrdma_get_dma_mr(struct ib_pd *ibpd, int acc)
718 {
719 	int status;
720 	struct ocrdma_mr *mr;
721 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
722 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
723 
724 	if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE)) {
725 		pr_err("%s err, invalid access rights\n", __func__);
726 		return ERR_PTR(-EINVAL);
727 	}
728 
729 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
730 	if (!mr)
731 		return ERR_PTR(-ENOMEM);
732 
733 	status = ocrdma_alloc_lkey(dev, mr, pd->id, acc, 0,
734 				   OCRDMA_ADDR_CHECK_DISABLE);
735 	if (status) {
736 		kfree(mr);
737 		return ERR_PTR(status);
738 	}
739 
740 	return &mr->ibmr;
741 }
742 
743 static void ocrdma_free_mr_pbl_tbl(struct ocrdma_dev *dev,
744 				   struct ocrdma_hw_mr *mr)
745 {
746 	struct pci_dev *pdev = dev->nic_info.pdev;
747 	int i = 0;
748 
749 	if (mr->pbl_table) {
750 		for (i = 0; i < mr->num_pbls; i++) {
751 			if (!mr->pbl_table[i].va)
752 				continue;
753 			dma_free_coherent(&pdev->dev, mr->pbl_size,
754 					  mr->pbl_table[i].va,
755 					  mr->pbl_table[i].pa);
756 		}
757 		kfree(mr->pbl_table);
758 		mr->pbl_table = NULL;
759 	}
760 }
761 
762 static int ocrdma_get_pbl_info(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
763 			      u32 num_pbes)
764 {
765 	u32 num_pbls = 0;
766 	u32 idx = 0;
767 	int status = 0;
768 	u32 pbl_size;
769 
770 	do {
771 		pbl_size = OCRDMA_MIN_HPAGE_SIZE * (1 << idx);
772 		if (pbl_size > MAX_OCRDMA_PBL_SIZE) {
773 			status = -EFAULT;
774 			break;
775 		}
776 		num_pbls = roundup(num_pbes, (pbl_size / sizeof(u64)));
777 		num_pbls = num_pbls / (pbl_size / sizeof(u64));
778 		idx++;
779 	} while (num_pbls >= dev->attr.max_num_mr_pbl);
780 
781 	mr->hwmr.num_pbes = num_pbes;
782 	mr->hwmr.num_pbls = num_pbls;
783 	mr->hwmr.pbl_size = pbl_size;
784 	return status;
785 }
786 
787 static int ocrdma_build_pbl_tbl(struct ocrdma_dev *dev, struct ocrdma_hw_mr *mr)
788 {
789 	int status = 0;
790 	int i;
791 	u32 dma_len = mr->pbl_size;
792 	struct pci_dev *pdev = dev->nic_info.pdev;
793 	void *va;
794 	dma_addr_t pa;
795 
796 	mr->pbl_table = kcalloc(mr->num_pbls, sizeof(struct ocrdma_pbl),
797 				GFP_KERNEL);
798 
799 	if (!mr->pbl_table)
800 		return -ENOMEM;
801 
802 	for (i = 0; i < mr->num_pbls; i++) {
803 		va = dma_alloc_coherent(&pdev->dev, dma_len, &pa, GFP_KERNEL);
804 		if (!va) {
805 			ocrdma_free_mr_pbl_tbl(dev, mr);
806 			status = -ENOMEM;
807 			break;
808 		}
809 		mr->pbl_table[i].va = va;
810 		mr->pbl_table[i].pa = pa;
811 	}
812 	return status;
813 }
814 
815 static void build_user_pbes(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
816 			    u32 num_pbes)
817 {
818 	struct ocrdma_pbe *pbe;
819 	struct sg_dma_page_iter sg_iter;
820 	struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
821 	struct ib_umem *umem = mr->umem;
822 	int pbe_cnt, total_num_pbes = 0;
823 	u64 pg_addr;
824 
825 	if (!mr->hwmr.num_pbes)
826 		return;
827 
828 	pbe = (struct ocrdma_pbe *)pbl_tbl->va;
829 	pbe_cnt = 0;
830 
831 	for_each_sg_dma_page (umem->sg_head.sgl, &sg_iter, umem->nmap, 0) {
832 		/* store the page address in pbe */
833 		pg_addr = sg_page_iter_dma_address(&sg_iter);
834 		pbe->pa_lo = cpu_to_le32(pg_addr);
835 		pbe->pa_hi = cpu_to_le32(upper_32_bits(pg_addr));
836 		pbe_cnt += 1;
837 		total_num_pbes += 1;
838 		pbe++;
839 
840 		/* if done building pbes, issue the mbx cmd. */
841 		if (total_num_pbes == num_pbes)
842 			return;
843 
844 		/* if the given pbl is full storing the pbes,
845 		 * move to next pbl.
846 		 */
847 		if (pbe_cnt == (mr->hwmr.pbl_size / sizeof(u64))) {
848 			pbl_tbl++;
849 			pbe = (struct ocrdma_pbe *)pbl_tbl->va;
850 			pbe_cnt = 0;
851 		}
852 	}
853 }
854 
855 struct ib_mr *ocrdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
856 				 u64 usr_addr, int acc, struct ib_udata *udata)
857 {
858 	int status = -ENOMEM;
859 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
860 	struct ocrdma_mr *mr;
861 	struct ocrdma_pd *pd;
862 	u32 num_pbes;
863 
864 	pd = get_ocrdma_pd(ibpd);
865 
866 	if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE))
867 		return ERR_PTR(-EINVAL);
868 
869 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
870 	if (!mr)
871 		return ERR_PTR(status);
872 	mr->umem = ib_umem_get(ibpd->device, start, len, acc);
873 	if (IS_ERR(mr->umem)) {
874 		status = -EFAULT;
875 		goto umem_err;
876 	}
877 	num_pbes = ib_umem_page_count(mr->umem);
878 	status = ocrdma_get_pbl_info(dev, mr, num_pbes);
879 	if (status)
880 		goto umem_err;
881 
882 	mr->hwmr.pbe_size = PAGE_SIZE;
883 	mr->hwmr.fbo = ib_umem_offset(mr->umem);
884 	mr->hwmr.va = usr_addr;
885 	mr->hwmr.len = len;
886 	mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
887 	mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
888 	mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
889 	mr->hwmr.local_rd = 1;
890 	mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
891 	status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
892 	if (status)
893 		goto umem_err;
894 	build_user_pbes(dev, mr, num_pbes);
895 	status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
896 	if (status)
897 		goto mbx_err;
898 	mr->ibmr.lkey = mr->hwmr.lkey;
899 	if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
900 		mr->ibmr.rkey = mr->hwmr.lkey;
901 
902 	return &mr->ibmr;
903 
904 mbx_err:
905 	ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
906 umem_err:
907 	kfree(mr);
908 	return ERR_PTR(status);
909 }
910 
911 int ocrdma_dereg_mr(struct ib_mr *ib_mr, struct ib_udata *udata)
912 {
913 	struct ocrdma_mr *mr = get_ocrdma_mr(ib_mr);
914 	struct ocrdma_dev *dev = get_ocrdma_dev(ib_mr->device);
915 
916 	(void) ocrdma_mbx_dealloc_lkey(dev, mr->hwmr.fr_mr, mr->hwmr.lkey);
917 
918 	kfree(mr->pages);
919 	ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
920 
921 	/* it could be user registered memory. */
922 	ib_umem_release(mr->umem);
923 	kfree(mr);
924 
925 	/* Don't stop cleanup, in case FW is unresponsive */
926 	if (dev->mqe_ctx.fw_error_state) {
927 		pr_err("%s(%d) fw not responding.\n",
928 		       __func__, dev->id);
929 	}
930 	return 0;
931 }
932 
933 static int ocrdma_copy_cq_uresp(struct ocrdma_dev *dev, struct ocrdma_cq *cq,
934 				struct ib_udata *udata)
935 {
936 	int status;
937 	struct ocrdma_ucontext *uctx = rdma_udata_to_drv_context(
938 		udata, struct ocrdma_ucontext, ibucontext);
939 	struct ocrdma_create_cq_uresp uresp;
940 
941 	/* this must be user flow! */
942 	if (!udata)
943 		return -EINVAL;
944 
945 	memset(&uresp, 0, sizeof(uresp));
946 	uresp.cq_id = cq->id;
947 	uresp.page_size = PAGE_ALIGN(cq->len);
948 	uresp.num_pages = 1;
949 	uresp.max_hw_cqe = cq->max_hw_cqe;
950 	uresp.page_addr[0] = virt_to_phys(cq->va);
951 	uresp.db_page_addr =  ocrdma_get_db_addr(dev, uctx->cntxt_pd->id);
952 	uresp.db_page_size = dev->nic_info.db_page_size;
953 	uresp.phase_change = cq->phase_change ? 1 : 0;
954 	status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
955 	if (status) {
956 		pr_err("%s(%d) copy error cqid=0x%x.\n",
957 		       __func__, dev->id, cq->id);
958 		goto err;
959 	}
960 	status = ocrdma_add_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
961 	if (status)
962 		goto err;
963 	status = ocrdma_add_mmap(uctx, uresp.page_addr[0], uresp.page_size);
964 	if (status) {
965 		ocrdma_del_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
966 		goto err;
967 	}
968 	cq->ucontext = uctx;
969 err:
970 	return status;
971 }
972 
973 int ocrdma_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
974 		     struct ib_udata *udata)
975 {
976 	struct ib_device *ibdev = ibcq->device;
977 	int entries = attr->cqe;
978 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
979 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
980 	struct ocrdma_ucontext *uctx = rdma_udata_to_drv_context(
981 		udata, struct ocrdma_ucontext, ibucontext);
982 	u16 pd_id = 0;
983 	int status;
984 	struct ocrdma_create_cq_ureq ureq;
985 
986 	if (attr->flags)
987 		return -EINVAL;
988 
989 	if (udata) {
990 		if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
991 			return -EFAULT;
992 	} else
993 		ureq.dpp_cq = 0;
994 
995 	spin_lock_init(&cq->cq_lock);
996 	spin_lock_init(&cq->comp_handler_lock);
997 	INIT_LIST_HEAD(&cq->sq_head);
998 	INIT_LIST_HEAD(&cq->rq_head);
999 
1000 	if (udata)
1001 		pd_id = uctx->cntxt_pd->id;
1002 
1003 	status = ocrdma_mbx_create_cq(dev, cq, entries, ureq.dpp_cq, pd_id);
1004 	if (status)
1005 		return status;
1006 
1007 	if (udata) {
1008 		status = ocrdma_copy_cq_uresp(dev, cq, udata);
1009 		if (status)
1010 			goto ctx_err;
1011 	}
1012 	cq->phase = OCRDMA_CQE_VALID;
1013 	dev->cq_tbl[cq->id] = cq;
1014 	return 0;
1015 
1016 ctx_err:
1017 	ocrdma_mbx_destroy_cq(dev, cq);
1018 	return status;
1019 }
1020 
1021 int ocrdma_resize_cq(struct ib_cq *ibcq, int new_cnt,
1022 		     struct ib_udata *udata)
1023 {
1024 	int status = 0;
1025 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
1026 
1027 	if (new_cnt < 1 || new_cnt > cq->max_hw_cqe) {
1028 		status = -EINVAL;
1029 		return status;
1030 	}
1031 	ibcq->cqe = new_cnt;
1032 	return status;
1033 }
1034 
1035 static void ocrdma_flush_cq(struct ocrdma_cq *cq)
1036 {
1037 	int cqe_cnt;
1038 	int valid_count = 0;
1039 	unsigned long flags;
1040 
1041 	struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
1042 	struct ocrdma_cqe *cqe = NULL;
1043 
1044 	cqe = cq->va;
1045 	cqe_cnt = cq->cqe_cnt;
1046 
1047 	/* Last irq might have scheduled a polling thread
1048 	 * sync-up with it before hard flushing.
1049 	 */
1050 	spin_lock_irqsave(&cq->cq_lock, flags);
1051 	while (cqe_cnt) {
1052 		if (is_cqe_valid(cq, cqe))
1053 			valid_count++;
1054 		cqe++;
1055 		cqe_cnt--;
1056 	}
1057 	ocrdma_ring_cq_db(dev, cq->id, false, false, valid_count);
1058 	spin_unlock_irqrestore(&cq->cq_lock, flags);
1059 }
1060 
1061 void ocrdma_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
1062 {
1063 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
1064 	struct ocrdma_eq *eq = NULL;
1065 	struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
1066 	int pdid = 0;
1067 	u32 irq, indx;
1068 
1069 	dev->cq_tbl[cq->id] = NULL;
1070 	indx = ocrdma_get_eq_table_index(dev, cq->eqn);
1071 
1072 	eq = &dev->eq_tbl[indx];
1073 	irq = ocrdma_get_irq(dev, eq);
1074 	synchronize_irq(irq);
1075 	ocrdma_flush_cq(cq);
1076 
1077 	ocrdma_mbx_destroy_cq(dev, cq);
1078 	if (cq->ucontext) {
1079 		pdid = cq->ucontext->cntxt_pd->id;
1080 		ocrdma_del_mmap(cq->ucontext, (u64) cq->pa,
1081 				PAGE_ALIGN(cq->len));
1082 		ocrdma_del_mmap(cq->ucontext,
1083 				ocrdma_get_db_addr(dev, pdid),
1084 				dev->nic_info.db_page_size);
1085 	}
1086 }
1087 
1088 static int ocrdma_add_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1089 {
1090 	int status = -EINVAL;
1091 
1092 	if (qp->id < OCRDMA_MAX_QP && dev->qp_tbl[qp->id] == NULL) {
1093 		dev->qp_tbl[qp->id] = qp;
1094 		status = 0;
1095 	}
1096 	return status;
1097 }
1098 
1099 static void ocrdma_del_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1100 {
1101 	dev->qp_tbl[qp->id] = NULL;
1102 }
1103 
1104 static int ocrdma_check_qp_params(struct ib_pd *ibpd, struct ocrdma_dev *dev,
1105 				  struct ib_qp_init_attr *attrs,
1106 				  struct ib_udata *udata)
1107 {
1108 	if ((attrs->qp_type != IB_QPT_GSI) &&
1109 	    (attrs->qp_type != IB_QPT_RC) &&
1110 	    (attrs->qp_type != IB_QPT_UC) &&
1111 	    (attrs->qp_type != IB_QPT_UD)) {
1112 		pr_err("%s(%d) unsupported qp type=0x%x requested\n",
1113 		       __func__, dev->id, attrs->qp_type);
1114 		return -EOPNOTSUPP;
1115 	}
1116 	/* Skip the check for QP1 to support CM size of 128 */
1117 	if ((attrs->qp_type != IB_QPT_GSI) &&
1118 	    (attrs->cap.max_send_wr > dev->attr.max_wqe)) {
1119 		pr_err("%s(%d) unsupported send_wr=0x%x requested\n",
1120 		       __func__, dev->id, attrs->cap.max_send_wr);
1121 		pr_err("%s(%d) supported send_wr=0x%x\n",
1122 		       __func__, dev->id, dev->attr.max_wqe);
1123 		return -EINVAL;
1124 	}
1125 	if (!attrs->srq && (attrs->cap.max_recv_wr > dev->attr.max_rqe)) {
1126 		pr_err("%s(%d) unsupported recv_wr=0x%x requested\n",
1127 		       __func__, dev->id, attrs->cap.max_recv_wr);
1128 		pr_err("%s(%d) supported recv_wr=0x%x\n",
1129 		       __func__, dev->id, dev->attr.max_rqe);
1130 		return -EINVAL;
1131 	}
1132 	if (attrs->cap.max_inline_data > dev->attr.max_inline_data) {
1133 		pr_err("%s(%d) unsupported inline data size=0x%x requested\n",
1134 		       __func__, dev->id, attrs->cap.max_inline_data);
1135 		pr_err("%s(%d) supported inline data size=0x%x\n",
1136 		       __func__, dev->id, dev->attr.max_inline_data);
1137 		return -EINVAL;
1138 	}
1139 	if (attrs->cap.max_send_sge > dev->attr.max_send_sge) {
1140 		pr_err("%s(%d) unsupported send_sge=0x%x requested\n",
1141 		       __func__, dev->id, attrs->cap.max_send_sge);
1142 		pr_err("%s(%d) supported send_sge=0x%x\n",
1143 		       __func__, dev->id, dev->attr.max_send_sge);
1144 		return -EINVAL;
1145 	}
1146 	if (attrs->cap.max_recv_sge > dev->attr.max_recv_sge) {
1147 		pr_err("%s(%d) unsupported recv_sge=0x%x requested\n",
1148 		       __func__, dev->id, attrs->cap.max_recv_sge);
1149 		pr_err("%s(%d) supported recv_sge=0x%x\n",
1150 		       __func__, dev->id, dev->attr.max_recv_sge);
1151 		return -EINVAL;
1152 	}
1153 	/* unprivileged user space cannot create special QP */
1154 	if (udata && attrs->qp_type == IB_QPT_GSI) {
1155 		pr_err
1156 		    ("%s(%d) Userspace can't create special QPs of type=0x%x\n",
1157 		     __func__, dev->id, attrs->qp_type);
1158 		return -EINVAL;
1159 	}
1160 	/* allow creating only one GSI type of QP */
1161 	if (attrs->qp_type == IB_QPT_GSI && dev->gsi_qp_created) {
1162 		pr_err("%s(%d) GSI special QPs already created.\n",
1163 		       __func__, dev->id);
1164 		return -EINVAL;
1165 	}
1166 	/* verify consumer QPs are not trying to use GSI QP's CQ */
1167 	if ((attrs->qp_type != IB_QPT_GSI) && (dev->gsi_qp_created)) {
1168 		if ((dev->gsi_sqcq == get_ocrdma_cq(attrs->send_cq)) ||
1169 			(dev->gsi_rqcq == get_ocrdma_cq(attrs->recv_cq))) {
1170 			pr_err("%s(%d) Consumer QP cannot use GSI CQs.\n",
1171 				__func__, dev->id);
1172 			return -EINVAL;
1173 		}
1174 	}
1175 	return 0;
1176 }
1177 
1178 static int ocrdma_copy_qp_uresp(struct ocrdma_qp *qp,
1179 				struct ib_udata *udata, int dpp_offset,
1180 				int dpp_credit_lmt, int srq)
1181 {
1182 	int status;
1183 	u64 usr_db;
1184 	struct ocrdma_create_qp_uresp uresp;
1185 	struct ocrdma_pd *pd = qp->pd;
1186 	struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
1187 
1188 	memset(&uresp, 0, sizeof(uresp));
1189 	usr_db = dev->nic_info.unmapped_db +
1190 			(pd->id * dev->nic_info.db_page_size);
1191 	uresp.qp_id = qp->id;
1192 	uresp.sq_dbid = qp->sq.dbid;
1193 	uresp.num_sq_pages = 1;
1194 	uresp.sq_page_size = PAGE_ALIGN(qp->sq.len);
1195 	uresp.sq_page_addr[0] = virt_to_phys(qp->sq.va);
1196 	uresp.num_wqe_allocated = qp->sq.max_cnt;
1197 	if (!srq) {
1198 		uresp.rq_dbid = qp->rq.dbid;
1199 		uresp.num_rq_pages = 1;
1200 		uresp.rq_page_size = PAGE_ALIGN(qp->rq.len);
1201 		uresp.rq_page_addr[0] = virt_to_phys(qp->rq.va);
1202 		uresp.num_rqe_allocated = qp->rq.max_cnt;
1203 	}
1204 	uresp.db_page_addr = usr_db;
1205 	uresp.db_page_size = dev->nic_info.db_page_size;
1206 	uresp.db_sq_offset = OCRDMA_DB_GEN2_SQ_OFFSET;
1207 	uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1208 	uresp.db_shift = OCRDMA_DB_RQ_SHIFT;
1209 
1210 	if (qp->dpp_enabled) {
1211 		uresp.dpp_credit = dpp_credit_lmt;
1212 		uresp.dpp_offset = dpp_offset;
1213 	}
1214 	status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1215 	if (status) {
1216 		pr_err("%s(%d) user copy error.\n", __func__, dev->id);
1217 		goto err;
1218 	}
1219 	status = ocrdma_add_mmap(pd->uctx, uresp.sq_page_addr[0],
1220 				 uresp.sq_page_size);
1221 	if (status)
1222 		goto err;
1223 
1224 	if (!srq) {
1225 		status = ocrdma_add_mmap(pd->uctx, uresp.rq_page_addr[0],
1226 					 uresp.rq_page_size);
1227 		if (status)
1228 			goto rq_map_err;
1229 	}
1230 	return status;
1231 rq_map_err:
1232 	ocrdma_del_mmap(pd->uctx, uresp.sq_page_addr[0], uresp.sq_page_size);
1233 err:
1234 	return status;
1235 }
1236 
1237 static void ocrdma_set_qp_db(struct ocrdma_dev *dev, struct ocrdma_qp *qp,
1238 			     struct ocrdma_pd *pd)
1239 {
1240 	if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1241 		qp->sq_db = dev->nic_info.db +
1242 			(pd->id * dev->nic_info.db_page_size) +
1243 			OCRDMA_DB_GEN2_SQ_OFFSET;
1244 		qp->rq_db = dev->nic_info.db +
1245 			(pd->id * dev->nic_info.db_page_size) +
1246 			OCRDMA_DB_GEN2_RQ_OFFSET;
1247 	} else {
1248 		qp->sq_db = dev->nic_info.db +
1249 			(pd->id * dev->nic_info.db_page_size) +
1250 			OCRDMA_DB_SQ_OFFSET;
1251 		qp->rq_db = dev->nic_info.db +
1252 			(pd->id * dev->nic_info.db_page_size) +
1253 			OCRDMA_DB_RQ_OFFSET;
1254 	}
1255 }
1256 
1257 static int ocrdma_alloc_wr_id_tbl(struct ocrdma_qp *qp)
1258 {
1259 	qp->wqe_wr_id_tbl =
1260 	    kcalloc(qp->sq.max_cnt, sizeof(*(qp->wqe_wr_id_tbl)),
1261 		    GFP_KERNEL);
1262 	if (qp->wqe_wr_id_tbl == NULL)
1263 		return -ENOMEM;
1264 	qp->rqe_wr_id_tbl =
1265 	    kcalloc(qp->rq.max_cnt, sizeof(u64), GFP_KERNEL);
1266 	if (qp->rqe_wr_id_tbl == NULL)
1267 		return -ENOMEM;
1268 
1269 	return 0;
1270 }
1271 
1272 static void ocrdma_set_qp_init_params(struct ocrdma_qp *qp,
1273 				      struct ocrdma_pd *pd,
1274 				      struct ib_qp_init_attr *attrs)
1275 {
1276 	qp->pd = pd;
1277 	spin_lock_init(&qp->q_lock);
1278 	INIT_LIST_HEAD(&qp->sq_entry);
1279 	INIT_LIST_HEAD(&qp->rq_entry);
1280 
1281 	qp->qp_type = attrs->qp_type;
1282 	qp->cap_flags = OCRDMA_QP_INB_RD | OCRDMA_QP_INB_WR;
1283 	qp->max_inline_data = attrs->cap.max_inline_data;
1284 	qp->sq.max_sges = attrs->cap.max_send_sge;
1285 	qp->rq.max_sges = attrs->cap.max_recv_sge;
1286 	qp->state = OCRDMA_QPS_RST;
1287 	qp->signaled = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR) ? true : false;
1288 }
1289 
1290 static void ocrdma_store_gsi_qp_cq(struct ocrdma_dev *dev,
1291 				   struct ib_qp_init_attr *attrs)
1292 {
1293 	if (attrs->qp_type == IB_QPT_GSI) {
1294 		dev->gsi_qp_created = 1;
1295 		dev->gsi_sqcq = get_ocrdma_cq(attrs->send_cq);
1296 		dev->gsi_rqcq = get_ocrdma_cq(attrs->recv_cq);
1297 	}
1298 }
1299 
1300 struct ib_qp *ocrdma_create_qp(struct ib_pd *ibpd,
1301 			       struct ib_qp_init_attr *attrs,
1302 			       struct ib_udata *udata)
1303 {
1304 	int status;
1305 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1306 	struct ocrdma_qp *qp;
1307 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1308 	struct ocrdma_create_qp_ureq ureq;
1309 	u16 dpp_credit_lmt, dpp_offset;
1310 
1311 	status = ocrdma_check_qp_params(ibpd, dev, attrs, udata);
1312 	if (status)
1313 		goto gen_err;
1314 
1315 	memset(&ureq, 0, sizeof(ureq));
1316 	if (udata) {
1317 		if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1318 			return ERR_PTR(-EFAULT);
1319 	}
1320 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1321 	if (!qp) {
1322 		status = -ENOMEM;
1323 		goto gen_err;
1324 	}
1325 	ocrdma_set_qp_init_params(qp, pd, attrs);
1326 	if (udata == NULL)
1327 		qp->cap_flags |= (OCRDMA_QP_MW_BIND | OCRDMA_QP_LKEY0 |
1328 					OCRDMA_QP_FAST_REG);
1329 
1330 	mutex_lock(&dev->dev_lock);
1331 	status = ocrdma_mbx_create_qp(qp, attrs, ureq.enable_dpp_cq,
1332 					ureq.dpp_cq_id,
1333 					&dpp_offset, &dpp_credit_lmt);
1334 	if (status)
1335 		goto mbx_err;
1336 
1337 	/* user space QP's wr_id table are managed in library */
1338 	if (udata == NULL) {
1339 		status = ocrdma_alloc_wr_id_tbl(qp);
1340 		if (status)
1341 			goto map_err;
1342 	}
1343 
1344 	status = ocrdma_add_qpn_map(dev, qp);
1345 	if (status)
1346 		goto map_err;
1347 	ocrdma_set_qp_db(dev, qp, pd);
1348 	if (udata) {
1349 		status = ocrdma_copy_qp_uresp(qp, udata, dpp_offset,
1350 					      dpp_credit_lmt,
1351 					      (attrs->srq != NULL));
1352 		if (status)
1353 			goto cpy_err;
1354 	}
1355 	ocrdma_store_gsi_qp_cq(dev, attrs);
1356 	qp->ibqp.qp_num = qp->id;
1357 	mutex_unlock(&dev->dev_lock);
1358 	return &qp->ibqp;
1359 
1360 cpy_err:
1361 	ocrdma_del_qpn_map(dev, qp);
1362 map_err:
1363 	ocrdma_mbx_destroy_qp(dev, qp);
1364 mbx_err:
1365 	mutex_unlock(&dev->dev_lock);
1366 	kfree(qp->wqe_wr_id_tbl);
1367 	kfree(qp->rqe_wr_id_tbl);
1368 	kfree(qp);
1369 	pr_err("%s(%d) error=%d\n", __func__, dev->id, status);
1370 gen_err:
1371 	return ERR_PTR(status);
1372 }
1373 
1374 int _ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1375 		      int attr_mask)
1376 {
1377 	int status = 0;
1378 	struct ocrdma_qp *qp;
1379 	struct ocrdma_dev *dev;
1380 	enum ib_qp_state old_qps;
1381 
1382 	qp = get_ocrdma_qp(ibqp);
1383 	dev = get_ocrdma_dev(ibqp->device);
1384 	if (attr_mask & IB_QP_STATE)
1385 		status = ocrdma_qp_state_change(qp, attr->qp_state, &old_qps);
1386 	/* if new and previous states are same hw doesn't need to
1387 	 * know about it.
1388 	 */
1389 	if (status < 0)
1390 		return status;
1391 	return ocrdma_mbx_modify_qp(dev, qp, attr, attr_mask);
1392 }
1393 
1394 int ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1395 		     int attr_mask, struct ib_udata *udata)
1396 {
1397 	unsigned long flags;
1398 	int status = -EINVAL;
1399 	struct ocrdma_qp *qp;
1400 	struct ocrdma_dev *dev;
1401 	enum ib_qp_state old_qps, new_qps;
1402 
1403 	qp = get_ocrdma_qp(ibqp);
1404 	dev = get_ocrdma_dev(ibqp->device);
1405 
1406 	/* syncronize with multiple context trying to change, retrive qps */
1407 	mutex_lock(&dev->dev_lock);
1408 	/* syncronize with wqe, rqe posting and cqe processing contexts */
1409 	spin_lock_irqsave(&qp->q_lock, flags);
1410 	old_qps = get_ibqp_state(qp->state);
1411 	if (attr_mask & IB_QP_STATE)
1412 		new_qps = attr->qp_state;
1413 	else
1414 		new_qps = old_qps;
1415 	spin_unlock_irqrestore(&qp->q_lock, flags);
1416 
1417 	if (!ib_modify_qp_is_ok(old_qps, new_qps, ibqp->qp_type, attr_mask)) {
1418 		pr_err("%s(%d) invalid attribute mask=0x%x specified for\n"
1419 		       "qpn=0x%x of type=0x%x old_qps=0x%x, new_qps=0x%x\n",
1420 		       __func__, dev->id, attr_mask, qp->id, ibqp->qp_type,
1421 		       old_qps, new_qps);
1422 		goto param_err;
1423 	}
1424 
1425 	status = _ocrdma_modify_qp(ibqp, attr, attr_mask);
1426 	if (status > 0)
1427 		status = 0;
1428 param_err:
1429 	mutex_unlock(&dev->dev_lock);
1430 	return status;
1431 }
1432 
1433 static enum ib_mtu ocrdma_mtu_int_to_enum(u16 mtu)
1434 {
1435 	switch (mtu) {
1436 	case 256:
1437 		return IB_MTU_256;
1438 	case 512:
1439 		return IB_MTU_512;
1440 	case 1024:
1441 		return IB_MTU_1024;
1442 	case 2048:
1443 		return IB_MTU_2048;
1444 	case 4096:
1445 		return IB_MTU_4096;
1446 	default:
1447 		return IB_MTU_1024;
1448 	}
1449 }
1450 
1451 static int ocrdma_to_ib_qp_acc_flags(int qp_cap_flags)
1452 {
1453 	int ib_qp_acc_flags = 0;
1454 
1455 	if (qp_cap_flags & OCRDMA_QP_INB_WR)
1456 		ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE;
1457 	if (qp_cap_flags & OCRDMA_QP_INB_RD)
1458 		ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE;
1459 	return ib_qp_acc_flags;
1460 }
1461 
1462 int ocrdma_query_qp(struct ib_qp *ibqp,
1463 		    struct ib_qp_attr *qp_attr,
1464 		    int attr_mask, struct ib_qp_init_attr *qp_init_attr)
1465 {
1466 	int status;
1467 	u32 qp_state;
1468 	struct ocrdma_qp_params params;
1469 	struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
1470 	struct ocrdma_dev *dev = get_ocrdma_dev(ibqp->device);
1471 
1472 	memset(&params, 0, sizeof(params));
1473 	mutex_lock(&dev->dev_lock);
1474 	status = ocrdma_mbx_query_qp(dev, qp, &params);
1475 	mutex_unlock(&dev->dev_lock);
1476 	if (status)
1477 		goto mbx_err;
1478 	if (qp->qp_type == IB_QPT_UD)
1479 		qp_attr->qkey = params.qkey;
1480 	qp_attr->path_mtu =
1481 		ocrdma_mtu_int_to_enum(params.path_mtu_pkey_indx &
1482 				OCRDMA_QP_PARAMS_PATH_MTU_MASK) >>
1483 				OCRDMA_QP_PARAMS_PATH_MTU_SHIFT;
1484 	qp_attr->path_mig_state = IB_MIG_MIGRATED;
1485 	qp_attr->rq_psn = params.hop_lmt_rq_psn & OCRDMA_QP_PARAMS_RQ_PSN_MASK;
1486 	qp_attr->sq_psn = params.tclass_sq_psn & OCRDMA_QP_PARAMS_SQ_PSN_MASK;
1487 	qp_attr->dest_qp_num =
1488 	    params.ack_to_rnr_rtc_dest_qpn & OCRDMA_QP_PARAMS_DEST_QPN_MASK;
1489 
1490 	qp_attr->qp_access_flags = ocrdma_to_ib_qp_acc_flags(qp->cap_flags);
1491 	qp_attr->cap.max_send_wr = qp->sq.max_cnt - 1;
1492 	qp_attr->cap.max_recv_wr = qp->rq.max_cnt - 1;
1493 	qp_attr->cap.max_send_sge = qp->sq.max_sges;
1494 	qp_attr->cap.max_recv_sge = qp->rq.max_sges;
1495 	qp_attr->cap.max_inline_data = qp->max_inline_data;
1496 	qp_init_attr->cap = qp_attr->cap;
1497 	qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
1498 
1499 	rdma_ah_set_grh(&qp_attr->ah_attr, NULL,
1500 			params.rnt_rc_sl_fl &
1501 			  OCRDMA_QP_PARAMS_FLOW_LABEL_MASK,
1502 			qp->sgid_idx,
1503 			(params.hop_lmt_rq_psn &
1504 			 OCRDMA_QP_PARAMS_HOP_LMT_MASK) >>
1505 			 OCRDMA_QP_PARAMS_HOP_LMT_SHIFT,
1506 			(params.tclass_sq_psn &
1507 			 OCRDMA_QP_PARAMS_TCLASS_MASK) >>
1508 			 OCRDMA_QP_PARAMS_TCLASS_SHIFT);
1509 	rdma_ah_set_dgid_raw(&qp_attr->ah_attr, &params.dgid[0]);
1510 
1511 	rdma_ah_set_port_num(&qp_attr->ah_attr, 1);
1512 	rdma_ah_set_sl(&qp_attr->ah_attr, (params.rnt_rc_sl_fl &
1513 					   OCRDMA_QP_PARAMS_SL_MASK) >>
1514 					   OCRDMA_QP_PARAMS_SL_SHIFT);
1515 	qp_attr->timeout = (params.ack_to_rnr_rtc_dest_qpn &
1516 			    OCRDMA_QP_PARAMS_ACK_TIMEOUT_MASK) >>
1517 				OCRDMA_QP_PARAMS_ACK_TIMEOUT_SHIFT;
1518 	qp_attr->rnr_retry = (params.ack_to_rnr_rtc_dest_qpn &
1519 			      OCRDMA_QP_PARAMS_RNR_RETRY_CNT_MASK) >>
1520 				OCRDMA_QP_PARAMS_RNR_RETRY_CNT_SHIFT;
1521 	qp_attr->retry_cnt =
1522 	    (params.rnt_rc_sl_fl & OCRDMA_QP_PARAMS_RETRY_CNT_MASK) >>
1523 		OCRDMA_QP_PARAMS_RETRY_CNT_SHIFT;
1524 	qp_attr->min_rnr_timer = 0;
1525 	qp_attr->pkey_index = 0;
1526 	qp_attr->port_num = 1;
1527 	rdma_ah_set_path_bits(&qp_attr->ah_attr, 0);
1528 	rdma_ah_set_static_rate(&qp_attr->ah_attr, 0);
1529 	qp_attr->alt_pkey_index = 0;
1530 	qp_attr->alt_port_num = 0;
1531 	qp_attr->alt_timeout = 0;
1532 	memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr));
1533 	qp_state = (params.max_sge_recv_flags & OCRDMA_QP_PARAMS_STATE_MASK) >>
1534 		    OCRDMA_QP_PARAMS_STATE_SHIFT;
1535 	qp_attr->qp_state = get_ibqp_state(qp_state);
1536 	qp_attr->cur_qp_state = qp_attr->qp_state;
1537 	qp_attr->sq_draining = (qp_state == OCRDMA_QPS_SQ_DRAINING) ? 1 : 0;
1538 	qp_attr->max_dest_rd_atomic =
1539 	    params.max_ord_ird >> OCRDMA_QP_PARAMS_MAX_ORD_SHIFT;
1540 	qp_attr->max_rd_atomic =
1541 	    params.max_ord_ird & OCRDMA_QP_PARAMS_MAX_IRD_MASK;
1542 	qp_attr->en_sqd_async_notify = (params.max_sge_recv_flags &
1543 				OCRDMA_QP_PARAMS_FLAGS_SQD_ASYNC) ? 1 : 0;
1544 	/* Sync driver QP state with FW */
1545 	ocrdma_qp_state_change(qp, qp_attr->qp_state, NULL);
1546 mbx_err:
1547 	return status;
1548 }
1549 
1550 static void ocrdma_srq_toggle_bit(struct ocrdma_srq *srq, unsigned int idx)
1551 {
1552 	unsigned int i = idx / 32;
1553 	u32 mask = (1U << (idx % 32));
1554 
1555 	srq->idx_bit_fields[i] ^= mask;
1556 }
1557 
1558 static int ocrdma_hwq_free_cnt(struct ocrdma_qp_hwq_info *q)
1559 {
1560 	return ((q->max_wqe_idx - q->head) + q->tail) % q->max_cnt;
1561 }
1562 
1563 static int is_hw_sq_empty(struct ocrdma_qp *qp)
1564 {
1565 	return (qp->sq.tail == qp->sq.head);
1566 }
1567 
1568 static int is_hw_rq_empty(struct ocrdma_qp *qp)
1569 {
1570 	return (qp->rq.tail == qp->rq.head);
1571 }
1572 
1573 static void *ocrdma_hwq_head(struct ocrdma_qp_hwq_info *q)
1574 {
1575 	return q->va + (q->head * q->entry_size);
1576 }
1577 
1578 static void *ocrdma_hwq_head_from_idx(struct ocrdma_qp_hwq_info *q,
1579 				      u32 idx)
1580 {
1581 	return q->va + (idx * q->entry_size);
1582 }
1583 
1584 static void ocrdma_hwq_inc_head(struct ocrdma_qp_hwq_info *q)
1585 {
1586 	q->head = (q->head + 1) & q->max_wqe_idx;
1587 }
1588 
1589 static void ocrdma_hwq_inc_tail(struct ocrdma_qp_hwq_info *q)
1590 {
1591 	q->tail = (q->tail + 1) & q->max_wqe_idx;
1592 }
1593 
1594 /* discard the cqe for a given QP */
1595 static void ocrdma_discard_cqes(struct ocrdma_qp *qp, struct ocrdma_cq *cq)
1596 {
1597 	unsigned long cq_flags;
1598 	unsigned long flags;
1599 	int discard_cnt = 0;
1600 	u32 cur_getp, stop_getp;
1601 	struct ocrdma_cqe *cqe;
1602 	u32 qpn = 0, wqe_idx = 0;
1603 
1604 	spin_lock_irqsave(&cq->cq_lock, cq_flags);
1605 
1606 	/* traverse through the CQEs in the hw CQ,
1607 	 * find the matching CQE for a given qp,
1608 	 * mark the matching one discarded by clearing qpn.
1609 	 * ring the doorbell in the poll_cq() as
1610 	 * we don't complete out of order cqe.
1611 	 */
1612 
1613 	cur_getp = cq->getp;
1614 	/* find upto when do we reap the cq. */
1615 	stop_getp = cur_getp;
1616 	do {
1617 		if (is_hw_sq_empty(qp) && (!qp->srq && is_hw_rq_empty(qp)))
1618 			break;
1619 
1620 		cqe = cq->va + cur_getp;
1621 		/* if (a) done reaping whole hw cq, or
1622 		 *    (b) qp_xq becomes empty.
1623 		 * then exit
1624 		 */
1625 		qpn = cqe->cmn.qpn & OCRDMA_CQE_QPN_MASK;
1626 		/* if previously discarded cqe found, skip that too. */
1627 		/* check for matching qp */
1628 		if (qpn == 0 || qpn != qp->id)
1629 			goto skip_cqe;
1630 
1631 		if (is_cqe_for_sq(cqe)) {
1632 			ocrdma_hwq_inc_tail(&qp->sq);
1633 		} else {
1634 			if (qp->srq) {
1635 				wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
1636 					OCRDMA_CQE_BUFTAG_SHIFT) &
1637 					qp->srq->rq.max_wqe_idx;
1638 				BUG_ON(wqe_idx < 1);
1639 				spin_lock_irqsave(&qp->srq->q_lock, flags);
1640 				ocrdma_hwq_inc_tail(&qp->srq->rq);
1641 				ocrdma_srq_toggle_bit(qp->srq, wqe_idx - 1);
1642 				spin_unlock_irqrestore(&qp->srq->q_lock, flags);
1643 
1644 			} else {
1645 				ocrdma_hwq_inc_tail(&qp->rq);
1646 			}
1647 		}
1648 		/* mark cqe discarded so that it is not picked up later
1649 		 * in the poll_cq().
1650 		 */
1651 		discard_cnt += 1;
1652 		cqe->cmn.qpn = 0;
1653 skip_cqe:
1654 		cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
1655 	} while (cur_getp != stop_getp);
1656 	spin_unlock_irqrestore(&cq->cq_lock, cq_flags);
1657 }
1658 
1659 void ocrdma_del_flush_qp(struct ocrdma_qp *qp)
1660 {
1661 	int found = false;
1662 	unsigned long flags;
1663 	struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
1664 	/* sync with any active CQ poll */
1665 
1666 	spin_lock_irqsave(&dev->flush_q_lock, flags);
1667 	found = ocrdma_is_qp_in_sq_flushlist(qp->sq_cq, qp);
1668 	if (found)
1669 		list_del(&qp->sq_entry);
1670 	if (!qp->srq) {
1671 		found = ocrdma_is_qp_in_rq_flushlist(qp->rq_cq, qp);
1672 		if (found)
1673 			list_del(&qp->rq_entry);
1674 	}
1675 	spin_unlock_irqrestore(&dev->flush_q_lock, flags);
1676 }
1677 
1678 int ocrdma_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
1679 {
1680 	struct ocrdma_pd *pd;
1681 	struct ocrdma_qp *qp;
1682 	struct ocrdma_dev *dev;
1683 	struct ib_qp_attr attrs;
1684 	int attr_mask;
1685 	unsigned long flags;
1686 
1687 	qp = get_ocrdma_qp(ibqp);
1688 	dev = get_ocrdma_dev(ibqp->device);
1689 
1690 	pd = qp->pd;
1691 
1692 	/* change the QP state to ERROR */
1693 	if (qp->state != OCRDMA_QPS_RST) {
1694 		attrs.qp_state = IB_QPS_ERR;
1695 		attr_mask = IB_QP_STATE;
1696 		_ocrdma_modify_qp(ibqp, &attrs, attr_mask);
1697 	}
1698 	/* ensure that CQEs for newly created QP (whose id may be same with
1699 	 * one which just getting destroyed are same), dont get
1700 	 * discarded until the old CQEs are discarded.
1701 	 */
1702 	mutex_lock(&dev->dev_lock);
1703 	(void) ocrdma_mbx_destroy_qp(dev, qp);
1704 
1705 	/*
1706 	 * acquire CQ lock while destroy is in progress, in order to
1707 	 * protect against proessing in-flight CQEs for this QP.
1708 	 */
1709 	spin_lock_irqsave(&qp->sq_cq->cq_lock, flags);
1710 	if (qp->rq_cq && (qp->rq_cq != qp->sq_cq)) {
1711 		spin_lock(&qp->rq_cq->cq_lock);
1712 		ocrdma_del_qpn_map(dev, qp);
1713 		spin_unlock(&qp->rq_cq->cq_lock);
1714 	} else {
1715 		ocrdma_del_qpn_map(dev, qp);
1716 	}
1717 	spin_unlock_irqrestore(&qp->sq_cq->cq_lock, flags);
1718 
1719 	if (!pd->uctx) {
1720 		ocrdma_discard_cqes(qp, qp->sq_cq);
1721 		ocrdma_discard_cqes(qp, qp->rq_cq);
1722 	}
1723 	mutex_unlock(&dev->dev_lock);
1724 
1725 	if (pd->uctx) {
1726 		ocrdma_del_mmap(pd->uctx, (u64) qp->sq.pa,
1727 				PAGE_ALIGN(qp->sq.len));
1728 		if (!qp->srq)
1729 			ocrdma_del_mmap(pd->uctx, (u64) qp->rq.pa,
1730 					PAGE_ALIGN(qp->rq.len));
1731 	}
1732 
1733 	ocrdma_del_flush_qp(qp);
1734 
1735 	kfree(qp->wqe_wr_id_tbl);
1736 	kfree(qp->rqe_wr_id_tbl);
1737 	kfree(qp);
1738 	return 0;
1739 }
1740 
1741 static int ocrdma_copy_srq_uresp(struct ocrdma_dev *dev, struct ocrdma_srq *srq,
1742 				struct ib_udata *udata)
1743 {
1744 	int status;
1745 	struct ocrdma_create_srq_uresp uresp;
1746 
1747 	memset(&uresp, 0, sizeof(uresp));
1748 	uresp.rq_dbid = srq->rq.dbid;
1749 	uresp.num_rq_pages = 1;
1750 	uresp.rq_page_addr[0] = virt_to_phys(srq->rq.va);
1751 	uresp.rq_page_size = srq->rq.len;
1752 	uresp.db_page_addr = dev->nic_info.unmapped_db +
1753 	    (srq->pd->id * dev->nic_info.db_page_size);
1754 	uresp.db_page_size = dev->nic_info.db_page_size;
1755 	uresp.num_rqe_allocated = srq->rq.max_cnt;
1756 	if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1757 		uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1758 		uresp.db_shift = 24;
1759 	} else {
1760 		uresp.db_rq_offset = OCRDMA_DB_RQ_OFFSET;
1761 		uresp.db_shift = 16;
1762 	}
1763 
1764 	status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1765 	if (status)
1766 		return status;
1767 	status = ocrdma_add_mmap(srq->pd->uctx, uresp.rq_page_addr[0],
1768 				 uresp.rq_page_size);
1769 	if (status)
1770 		return status;
1771 	return status;
1772 }
1773 
1774 int ocrdma_create_srq(struct ib_srq *ibsrq, struct ib_srq_init_attr *init_attr,
1775 		      struct ib_udata *udata)
1776 {
1777 	int status;
1778 	struct ocrdma_pd *pd = get_ocrdma_pd(ibsrq->pd);
1779 	struct ocrdma_dev *dev = get_ocrdma_dev(ibsrq->device);
1780 	struct ocrdma_srq *srq = get_ocrdma_srq(ibsrq);
1781 
1782 	if (init_attr->attr.max_sge > dev->attr.max_recv_sge)
1783 		return -EINVAL;
1784 	if (init_attr->attr.max_wr > dev->attr.max_rqe)
1785 		return -EINVAL;
1786 
1787 	spin_lock_init(&srq->q_lock);
1788 	srq->pd = pd;
1789 	srq->db = dev->nic_info.db + (pd->id * dev->nic_info.db_page_size);
1790 	status = ocrdma_mbx_create_srq(dev, srq, init_attr, pd);
1791 	if (status)
1792 		return status;
1793 
1794 	if (!udata) {
1795 		srq->rqe_wr_id_tbl = kcalloc(srq->rq.max_cnt, sizeof(u64),
1796 					     GFP_KERNEL);
1797 		if (!srq->rqe_wr_id_tbl) {
1798 			status = -ENOMEM;
1799 			goto arm_err;
1800 		}
1801 
1802 		srq->bit_fields_len = (srq->rq.max_cnt / 32) +
1803 		    (srq->rq.max_cnt % 32 ? 1 : 0);
1804 		srq->idx_bit_fields =
1805 		    kmalloc_array(srq->bit_fields_len, sizeof(u32),
1806 				  GFP_KERNEL);
1807 		if (!srq->idx_bit_fields) {
1808 			status = -ENOMEM;
1809 			goto arm_err;
1810 		}
1811 		memset(srq->idx_bit_fields, 0xff,
1812 		       srq->bit_fields_len * sizeof(u32));
1813 	}
1814 
1815 	if (init_attr->attr.srq_limit) {
1816 		status = ocrdma_mbx_modify_srq(srq, &init_attr->attr);
1817 		if (status)
1818 			goto arm_err;
1819 	}
1820 
1821 	if (udata) {
1822 		status = ocrdma_copy_srq_uresp(dev, srq, udata);
1823 		if (status)
1824 			goto arm_err;
1825 	}
1826 
1827 	return 0;
1828 
1829 arm_err:
1830 	ocrdma_mbx_destroy_srq(dev, srq);
1831 	kfree(srq->rqe_wr_id_tbl);
1832 	kfree(srq->idx_bit_fields);
1833 	return status;
1834 }
1835 
1836 int ocrdma_modify_srq(struct ib_srq *ibsrq,
1837 		      struct ib_srq_attr *srq_attr,
1838 		      enum ib_srq_attr_mask srq_attr_mask,
1839 		      struct ib_udata *udata)
1840 {
1841 	int status;
1842 	struct ocrdma_srq *srq;
1843 
1844 	srq = get_ocrdma_srq(ibsrq);
1845 	if (srq_attr_mask & IB_SRQ_MAX_WR)
1846 		status = -EINVAL;
1847 	else
1848 		status = ocrdma_mbx_modify_srq(srq, srq_attr);
1849 	return status;
1850 }
1851 
1852 int ocrdma_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
1853 {
1854 	int status;
1855 	struct ocrdma_srq *srq;
1856 
1857 	srq = get_ocrdma_srq(ibsrq);
1858 	status = ocrdma_mbx_query_srq(srq, srq_attr);
1859 	return status;
1860 }
1861 
1862 void ocrdma_destroy_srq(struct ib_srq *ibsrq, struct ib_udata *udata)
1863 {
1864 	struct ocrdma_srq *srq;
1865 	struct ocrdma_dev *dev = get_ocrdma_dev(ibsrq->device);
1866 
1867 	srq = get_ocrdma_srq(ibsrq);
1868 
1869 	ocrdma_mbx_destroy_srq(dev, srq);
1870 
1871 	if (srq->pd->uctx)
1872 		ocrdma_del_mmap(srq->pd->uctx, (u64) srq->rq.pa,
1873 				PAGE_ALIGN(srq->rq.len));
1874 
1875 	kfree(srq->idx_bit_fields);
1876 	kfree(srq->rqe_wr_id_tbl);
1877 }
1878 
1879 /* unprivileged verbs and their support functions. */
1880 static void ocrdma_build_ud_hdr(struct ocrdma_qp *qp,
1881 				struct ocrdma_hdr_wqe *hdr,
1882 				const struct ib_send_wr *wr)
1883 {
1884 	struct ocrdma_ewqe_ud_hdr *ud_hdr =
1885 		(struct ocrdma_ewqe_ud_hdr *)(hdr + 1);
1886 	struct ocrdma_ah *ah = get_ocrdma_ah(ud_wr(wr)->ah);
1887 
1888 	ud_hdr->rsvd_dest_qpn = ud_wr(wr)->remote_qpn;
1889 	if (qp->qp_type == IB_QPT_GSI)
1890 		ud_hdr->qkey = qp->qkey;
1891 	else
1892 		ud_hdr->qkey = ud_wr(wr)->remote_qkey;
1893 	ud_hdr->rsvd_ahid = ah->id;
1894 	ud_hdr->hdr_type = ah->hdr_type;
1895 	if (ah->av->valid & OCRDMA_AV_VLAN_VALID)
1896 		hdr->cw |= (OCRDMA_FLAG_AH_VLAN_PR << OCRDMA_WQE_FLAGS_SHIFT);
1897 }
1898 
1899 static void ocrdma_build_sges(struct ocrdma_hdr_wqe *hdr,
1900 			      struct ocrdma_sge *sge, int num_sge,
1901 			      struct ib_sge *sg_list)
1902 {
1903 	int i;
1904 
1905 	for (i = 0; i < num_sge; i++) {
1906 		sge[i].lrkey = sg_list[i].lkey;
1907 		sge[i].addr_lo = sg_list[i].addr;
1908 		sge[i].addr_hi = upper_32_bits(sg_list[i].addr);
1909 		sge[i].len = sg_list[i].length;
1910 		hdr->total_len += sg_list[i].length;
1911 	}
1912 	if (num_sge == 0)
1913 		memset(sge, 0, sizeof(*sge));
1914 }
1915 
1916 static inline uint32_t ocrdma_sglist_len(struct ib_sge *sg_list, int num_sge)
1917 {
1918 	uint32_t total_len = 0, i;
1919 
1920 	for (i = 0; i < num_sge; i++)
1921 		total_len += sg_list[i].length;
1922 	return total_len;
1923 }
1924 
1925 
1926 static int ocrdma_build_inline_sges(struct ocrdma_qp *qp,
1927 				    struct ocrdma_hdr_wqe *hdr,
1928 				    struct ocrdma_sge *sge,
1929 				    const struct ib_send_wr *wr, u32 wqe_size)
1930 {
1931 	int i;
1932 	char *dpp_addr;
1933 
1934 	if (wr->send_flags & IB_SEND_INLINE && qp->qp_type != IB_QPT_UD) {
1935 		hdr->total_len = ocrdma_sglist_len(wr->sg_list, wr->num_sge);
1936 		if (unlikely(hdr->total_len > qp->max_inline_data)) {
1937 			pr_err("%s() supported_len=0x%x,\n"
1938 			       " unsupported len req=0x%x\n", __func__,
1939 				qp->max_inline_data, hdr->total_len);
1940 			return -EINVAL;
1941 		}
1942 		dpp_addr = (char *)sge;
1943 		for (i = 0; i < wr->num_sge; i++) {
1944 			memcpy(dpp_addr,
1945 			       (void *)(unsigned long)wr->sg_list[i].addr,
1946 			       wr->sg_list[i].length);
1947 			dpp_addr += wr->sg_list[i].length;
1948 		}
1949 
1950 		wqe_size += roundup(hdr->total_len, OCRDMA_WQE_ALIGN_BYTES);
1951 		if (0 == hdr->total_len)
1952 			wqe_size += sizeof(struct ocrdma_sge);
1953 		hdr->cw |= (OCRDMA_TYPE_INLINE << OCRDMA_WQE_TYPE_SHIFT);
1954 	} else {
1955 		ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
1956 		if (wr->num_sge)
1957 			wqe_size += (wr->num_sge * sizeof(struct ocrdma_sge));
1958 		else
1959 			wqe_size += sizeof(struct ocrdma_sge);
1960 		hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
1961 	}
1962 	hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
1963 	return 0;
1964 }
1965 
1966 static int ocrdma_build_send(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1967 			     const struct ib_send_wr *wr)
1968 {
1969 	int status;
1970 	struct ocrdma_sge *sge;
1971 	u32 wqe_size = sizeof(*hdr);
1972 
1973 	if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
1974 		ocrdma_build_ud_hdr(qp, hdr, wr);
1975 		sge = (struct ocrdma_sge *)(hdr + 2);
1976 		wqe_size += sizeof(struct ocrdma_ewqe_ud_hdr);
1977 	} else {
1978 		sge = (struct ocrdma_sge *)(hdr + 1);
1979 	}
1980 
1981 	status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
1982 	return status;
1983 }
1984 
1985 static int ocrdma_build_write(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1986 			      const struct ib_send_wr *wr)
1987 {
1988 	int status;
1989 	struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
1990 	struct ocrdma_sge *sge = ext_rw + 1;
1991 	u32 wqe_size = sizeof(*hdr) + sizeof(*ext_rw);
1992 
1993 	status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
1994 	if (status)
1995 		return status;
1996 	ext_rw->addr_lo = rdma_wr(wr)->remote_addr;
1997 	ext_rw->addr_hi = upper_32_bits(rdma_wr(wr)->remote_addr);
1998 	ext_rw->lrkey = rdma_wr(wr)->rkey;
1999 	ext_rw->len = hdr->total_len;
2000 	return 0;
2001 }
2002 
2003 static void ocrdma_build_read(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2004 			      const struct ib_send_wr *wr)
2005 {
2006 	struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
2007 	struct ocrdma_sge *sge = ext_rw + 1;
2008 	u32 wqe_size = ((wr->num_sge + 1) * sizeof(struct ocrdma_sge)) +
2009 	    sizeof(struct ocrdma_hdr_wqe);
2010 
2011 	ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
2012 	hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2013 	hdr->cw |= (OCRDMA_READ << OCRDMA_WQE_OPCODE_SHIFT);
2014 	hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2015 
2016 	ext_rw->addr_lo = rdma_wr(wr)->remote_addr;
2017 	ext_rw->addr_hi = upper_32_bits(rdma_wr(wr)->remote_addr);
2018 	ext_rw->lrkey = rdma_wr(wr)->rkey;
2019 	ext_rw->len = hdr->total_len;
2020 }
2021 
2022 static int get_encoded_page_size(int pg_sz)
2023 {
2024 	/* Max size is 256M 4096 << 16 */
2025 	int i = 0;
2026 	for (; i < 17; i++)
2027 		if (pg_sz == (4096 << i))
2028 			break;
2029 	return i;
2030 }
2031 
2032 static int ocrdma_build_reg(struct ocrdma_qp *qp,
2033 			    struct ocrdma_hdr_wqe *hdr,
2034 			    const struct ib_reg_wr *wr)
2035 {
2036 	u64 fbo;
2037 	struct ocrdma_ewqe_fr *fast_reg = (struct ocrdma_ewqe_fr *)(hdr + 1);
2038 	struct ocrdma_mr *mr = get_ocrdma_mr(wr->mr);
2039 	struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
2040 	struct ocrdma_pbe *pbe;
2041 	u32 wqe_size = sizeof(*fast_reg) + sizeof(*hdr);
2042 	int num_pbes = 0, i;
2043 
2044 	wqe_size = roundup(wqe_size, OCRDMA_WQE_ALIGN_BYTES);
2045 
2046 	hdr->cw |= (OCRDMA_FR_MR << OCRDMA_WQE_OPCODE_SHIFT);
2047 	hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2048 
2049 	if (wr->access & IB_ACCESS_LOCAL_WRITE)
2050 		hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_LOCAL_WR;
2051 	if (wr->access & IB_ACCESS_REMOTE_WRITE)
2052 		hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_WR;
2053 	if (wr->access & IB_ACCESS_REMOTE_READ)
2054 		hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_RD;
2055 	hdr->lkey = wr->key;
2056 	hdr->total_len = mr->ibmr.length;
2057 
2058 	fbo = mr->ibmr.iova - mr->pages[0];
2059 
2060 	fast_reg->va_hi = upper_32_bits(mr->ibmr.iova);
2061 	fast_reg->va_lo = (u32) (mr->ibmr.iova & 0xffffffff);
2062 	fast_reg->fbo_hi = upper_32_bits(fbo);
2063 	fast_reg->fbo_lo = (u32) fbo & 0xffffffff;
2064 	fast_reg->num_sges = mr->npages;
2065 	fast_reg->size_sge = get_encoded_page_size(mr->ibmr.page_size);
2066 
2067 	pbe = pbl_tbl->va;
2068 	for (i = 0; i < mr->npages; i++) {
2069 		u64 buf_addr = mr->pages[i];
2070 
2071 		pbe->pa_lo = cpu_to_le32((u32) (buf_addr & PAGE_MASK));
2072 		pbe->pa_hi = cpu_to_le32((u32) upper_32_bits(buf_addr));
2073 		num_pbes += 1;
2074 		pbe++;
2075 
2076 		/* if the pbl is full storing the pbes,
2077 		 * move to next pbl.
2078 		*/
2079 		if (num_pbes == (mr->hwmr.pbl_size/sizeof(u64))) {
2080 			pbl_tbl++;
2081 			pbe = (struct ocrdma_pbe *)pbl_tbl->va;
2082 		}
2083 	}
2084 
2085 	return 0;
2086 }
2087 
2088 static void ocrdma_ring_sq_db(struct ocrdma_qp *qp)
2089 {
2090 	u32 val = qp->sq.dbid | (1 << OCRDMA_DB_SQ_SHIFT);
2091 
2092 	iowrite32(val, qp->sq_db);
2093 }
2094 
2095 int ocrdma_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
2096 		     const struct ib_send_wr **bad_wr)
2097 {
2098 	int status = 0;
2099 	struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2100 	struct ocrdma_hdr_wqe *hdr;
2101 	unsigned long flags;
2102 
2103 	spin_lock_irqsave(&qp->q_lock, flags);
2104 	if (qp->state != OCRDMA_QPS_RTS && qp->state != OCRDMA_QPS_SQD) {
2105 		spin_unlock_irqrestore(&qp->q_lock, flags);
2106 		*bad_wr = wr;
2107 		return -EINVAL;
2108 	}
2109 
2110 	while (wr) {
2111 		if (qp->qp_type == IB_QPT_UD &&
2112 		    (wr->opcode != IB_WR_SEND &&
2113 		     wr->opcode != IB_WR_SEND_WITH_IMM)) {
2114 			*bad_wr = wr;
2115 			status = -EINVAL;
2116 			break;
2117 		}
2118 		if (ocrdma_hwq_free_cnt(&qp->sq) == 0 ||
2119 		    wr->num_sge > qp->sq.max_sges) {
2120 			*bad_wr = wr;
2121 			status = -ENOMEM;
2122 			break;
2123 		}
2124 		hdr = ocrdma_hwq_head(&qp->sq);
2125 		hdr->cw = 0;
2126 		if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2127 			hdr->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2128 		if (wr->send_flags & IB_SEND_FENCE)
2129 			hdr->cw |=
2130 			    (OCRDMA_FLAG_FENCE_L << OCRDMA_WQE_FLAGS_SHIFT);
2131 		if (wr->send_flags & IB_SEND_SOLICITED)
2132 			hdr->cw |=
2133 			    (OCRDMA_FLAG_SOLICIT << OCRDMA_WQE_FLAGS_SHIFT);
2134 		hdr->total_len = 0;
2135 		switch (wr->opcode) {
2136 		case IB_WR_SEND_WITH_IMM:
2137 			hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2138 			hdr->immdt = ntohl(wr->ex.imm_data);
2139 			/* fall through */
2140 		case IB_WR_SEND:
2141 			hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2142 			ocrdma_build_send(qp, hdr, wr);
2143 			break;
2144 		case IB_WR_SEND_WITH_INV:
2145 			hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
2146 			hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2147 			hdr->lkey = wr->ex.invalidate_rkey;
2148 			status = ocrdma_build_send(qp, hdr, wr);
2149 			break;
2150 		case IB_WR_RDMA_WRITE_WITH_IMM:
2151 			hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2152 			hdr->immdt = ntohl(wr->ex.imm_data);
2153 			/* fall through */
2154 		case IB_WR_RDMA_WRITE:
2155 			hdr->cw |= (OCRDMA_WRITE << OCRDMA_WQE_OPCODE_SHIFT);
2156 			status = ocrdma_build_write(qp, hdr, wr);
2157 			break;
2158 		case IB_WR_RDMA_READ:
2159 			ocrdma_build_read(qp, hdr, wr);
2160 			break;
2161 		case IB_WR_LOCAL_INV:
2162 			hdr->cw |=
2163 			    (OCRDMA_LKEY_INV << OCRDMA_WQE_OPCODE_SHIFT);
2164 			hdr->cw |= ((sizeof(struct ocrdma_hdr_wqe) +
2165 					sizeof(struct ocrdma_sge)) /
2166 				OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT;
2167 			hdr->lkey = wr->ex.invalidate_rkey;
2168 			break;
2169 		case IB_WR_REG_MR:
2170 			status = ocrdma_build_reg(qp, hdr, reg_wr(wr));
2171 			break;
2172 		default:
2173 			status = -EINVAL;
2174 			break;
2175 		}
2176 		if (status) {
2177 			*bad_wr = wr;
2178 			break;
2179 		}
2180 		if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2181 			qp->wqe_wr_id_tbl[qp->sq.head].signaled = 1;
2182 		else
2183 			qp->wqe_wr_id_tbl[qp->sq.head].signaled = 0;
2184 		qp->wqe_wr_id_tbl[qp->sq.head].wrid = wr->wr_id;
2185 		ocrdma_cpu_to_le32(hdr, ((hdr->cw >> OCRDMA_WQE_SIZE_SHIFT) &
2186 				   OCRDMA_WQE_SIZE_MASK) * OCRDMA_WQE_STRIDE);
2187 		/* make sure wqe is written before adapter can access it */
2188 		wmb();
2189 		/* inform hw to start processing it */
2190 		ocrdma_ring_sq_db(qp);
2191 
2192 		/* update pointer, counter for next wr */
2193 		ocrdma_hwq_inc_head(&qp->sq);
2194 		wr = wr->next;
2195 	}
2196 	spin_unlock_irqrestore(&qp->q_lock, flags);
2197 	return status;
2198 }
2199 
2200 static void ocrdma_ring_rq_db(struct ocrdma_qp *qp)
2201 {
2202 	u32 val = qp->rq.dbid | (1 << OCRDMA_DB_RQ_SHIFT);
2203 
2204 	iowrite32(val, qp->rq_db);
2205 }
2206 
2207 static void ocrdma_build_rqe(struct ocrdma_hdr_wqe *rqe,
2208 			     const struct ib_recv_wr *wr, u16 tag)
2209 {
2210 	u32 wqe_size = 0;
2211 	struct ocrdma_sge *sge;
2212 	if (wr->num_sge)
2213 		wqe_size = (wr->num_sge * sizeof(*sge)) + sizeof(*rqe);
2214 	else
2215 		wqe_size = sizeof(*sge) + sizeof(*rqe);
2216 
2217 	rqe->cw = ((wqe_size / OCRDMA_WQE_STRIDE) <<
2218 				OCRDMA_WQE_SIZE_SHIFT);
2219 	rqe->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2220 	rqe->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2221 	rqe->total_len = 0;
2222 	rqe->rsvd_tag = tag;
2223 	sge = (struct ocrdma_sge *)(rqe + 1);
2224 	ocrdma_build_sges(rqe, sge, wr->num_sge, wr->sg_list);
2225 	ocrdma_cpu_to_le32(rqe, wqe_size);
2226 }
2227 
2228 int ocrdma_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
2229 		     const struct ib_recv_wr **bad_wr)
2230 {
2231 	int status = 0;
2232 	unsigned long flags;
2233 	struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2234 	struct ocrdma_hdr_wqe *rqe;
2235 
2236 	spin_lock_irqsave(&qp->q_lock, flags);
2237 	if (qp->state == OCRDMA_QPS_RST || qp->state == OCRDMA_QPS_ERR) {
2238 		spin_unlock_irqrestore(&qp->q_lock, flags);
2239 		*bad_wr = wr;
2240 		return -EINVAL;
2241 	}
2242 	while (wr) {
2243 		if (ocrdma_hwq_free_cnt(&qp->rq) == 0 ||
2244 		    wr->num_sge > qp->rq.max_sges) {
2245 			*bad_wr = wr;
2246 			status = -ENOMEM;
2247 			break;
2248 		}
2249 		rqe = ocrdma_hwq_head(&qp->rq);
2250 		ocrdma_build_rqe(rqe, wr, 0);
2251 
2252 		qp->rqe_wr_id_tbl[qp->rq.head] = wr->wr_id;
2253 		/* make sure rqe is written before adapter can access it */
2254 		wmb();
2255 
2256 		/* inform hw to start processing it */
2257 		ocrdma_ring_rq_db(qp);
2258 
2259 		/* update pointer, counter for next wr */
2260 		ocrdma_hwq_inc_head(&qp->rq);
2261 		wr = wr->next;
2262 	}
2263 	spin_unlock_irqrestore(&qp->q_lock, flags);
2264 	return status;
2265 }
2266 
2267 /* cqe for srq's rqe can potentially arrive out of order.
2268  * index gives the entry in the shadow table where to store
2269  * the wr_id. tag/index is returned in cqe to reference back
2270  * for a given rqe.
2271  */
2272 static int ocrdma_srq_get_idx(struct ocrdma_srq *srq)
2273 {
2274 	int row = 0;
2275 	int indx = 0;
2276 
2277 	for (row = 0; row < srq->bit_fields_len; row++) {
2278 		if (srq->idx_bit_fields[row]) {
2279 			indx = ffs(srq->idx_bit_fields[row]);
2280 			indx = (row * 32) + (indx - 1);
2281 			BUG_ON(indx >= srq->rq.max_cnt);
2282 			ocrdma_srq_toggle_bit(srq, indx);
2283 			break;
2284 		}
2285 	}
2286 
2287 	BUG_ON(row == srq->bit_fields_len);
2288 	return indx + 1; /* Use from index 1 */
2289 }
2290 
2291 static void ocrdma_ring_srq_db(struct ocrdma_srq *srq)
2292 {
2293 	u32 val = srq->rq.dbid | (1 << 16);
2294 
2295 	iowrite32(val, srq->db + OCRDMA_DB_GEN2_SRQ_OFFSET);
2296 }
2297 
2298 int ocrdma_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
2299 			 const struct ib_recv_wr **bad_wr)
2300 {
2301 	int status = 0;
2302 	unsigned long flags;
2303 	struct ocrdma_srq *srq;
2304 	struct ocrdma_hdr_wqe *rqe;
2305 	u16 tag;
2306 
2307 	srq = get_ocrdma_srq(ibsrq);
2308 
2309 	spin_lock_irqsave(&srq->q_lock, flags);
2310 	while (wr) {
2311 		if (ocrdma_hwq_free_cnt(&srq->rq) == 0 ||
2312 		    wr->num_sge > srq->rq.max_sges) {
2313 			status = -ENOMEM;
2314 			*bad_wr = wr;
2315 			break;
2316 		}
2317 		tag = ocrdma_srq_get_idx(srq);
2318 		rqe = ocrdma_hwq_head(&srq->rq);
2319 		ocrdma_build_rqe(rqe, wr, tag);
2320 
2321 		srq->rqe_wr_id_tbl[tag] = wr->wr_id;
2322 		/* make sure rqe is written before adapter can perform DMA */
2323 		wmb();
2324 		/* inform hw to start processing it */
2325 		ocrdma_ring_srq_db(srq);
2326 		/* update pointer, counter for next wr */
2327 		ocrdma_hwq_inc_head(&srq->rq);
2328 		wr = wr->next;
2329 	}
2330 	spin_unlock_irqrestore(&srq->q_lock, flags);
2331 	return status;
2332 }
2333 
2334 static enum ib_wc_status ocrdma_to_ibwc_err(u16 status)
2335 {
2336 	enum ib_wc_status ibwc_status;
2337 
2338 	switch (status) {
2339 	case OCRDMA_CQE_GENERAL_ERR:
2340 		ibwc_status = IB_WC_GENERAL_ERR;
2341 		break;
2342 	case OCRDMA_CQE_LOC_LEN_ERR:
2343 		ibwc_status = IB_WC_LOC_LEN_ERR;
2344 		break;
2345 	case OCRDMA_CQE_LOC_QP_OP_ERR:
2346 		ibwc_status = IB_WC_LOC_QP_OP_ERR;
2347 		break;
2348 	case OCRDMA_CQE_LOC_EEC_OP_ERR:
2349 		ibwc_status = IB_WC_LOC_EEC_OP_ERR;
2350 		break;
2351 	case OCRDMA_CQE_LOC_PROT_ERR:
2352 		ibwc_status = IB_WC_LOC_PROT_ERR;
2353 		break;
2354 	case OCRDMA_CQE_WR_FLUSH_ERR:
2355 		ibwc_status = IB_WC_WR_FLUSH_ERR;
2356 		break;
2357 	case OCRDMA_CQE_MW_BIND_ERR:
2358 		ibwc_status = IB_WC_MW_BIND_ERR;
2359 		break;
2360 	case OCRDMA_CQE_BAD_RESP_ERR:
2361 		ibwc_status = IB_WC_BAD_RESP_ERR;
2362 		break;
2363 	case OCRDMA_CQE_LOC_ACCESS_ERR:
2364 		ibwc_status = IB_WC_LOC_ACCESS_ERR;
2365 		break;
2366 	case OCRDMA_CQE_REM_INV_REQ_ERR:
2367 		ibwc_status = IB_WC_REM_INV_REQ_ERR;
2368 		break;
2369 	case OCRDMA_CQE_REM_ACCESS_ERR:
2370 		ibwc_status = IB_WC_REM_ACCESS_ERR;
2371 		break;
2372 	case OCRDMA_CQE_REM_OP_ERR:
2373 		ibwc_status = IB_WC_REM_OP_ERR;
2374 		break;
2375 	case OCRDMA_CQE_RETRY_EXC_ERR:
2376 		ibwc_status = IB_WC_RETRY_EXC_ERR;
2377 		break;
2378 	case OCRDMA_CQE_RNR_RETRY_EXC_ERR:
2379 		ibwc_status = IB_WC_RNR_RETRY_EXC_ERR;
2380 		break;
2381 	case OCRDMA_CQE_LOC_RDD_VIOL_ERR:
2382 		ibwc_status = IB_WC_LOC_RDD_VIOL_ERR;
2383 		break;
2384 	case OCRDMA_CQE_REM_INV_RD_REQ_ERR:
2385 		ibwc_status = IB_WC_REM_INV_RD_REQ_ERR;
2386 		break;
2387 	case OCRDMA_CQE_REM_ABORT_ERR:
2388 		ibwc_status = IB_WC_REM_ABORT_ERR;
2389 		break;
2390 	case OCRDMA_CQE_INV_EECN_ERR:
2391 		ibwc_status = IB_WC_INV_EECN_ERR;
2392 		break;
2393 	case OCRDMA_CQE_INV_EEC_STATE_ERR:
2394 		ibwc_status = IB_WC_INV_EEC_STATE_ERR;
2395 		break;
2396 	case OCRDMA_CQE_FATAL_ERR:
2397 		ibwc_status = IB_WC_FATAL_ERR;
2398 		break;
2399 	case OCRDMA_CQE_RESP_TIMEOUT_ERR:
2400 		ibwc_status = IB_WC_RESP_TIMEOUT_ERR;
2401 		break;
2402 	default:
2403 		ibwc_status = IB_WC_GENERAL_ERR;
2404 		break;
2405 	}
2406 	return ibwc_status;
2407 }
2408 
2409 static void ocrdma_update_wc(struct ocrdma_qp *qp, struct ib_wc *ibwc,
2410 		      u32 wqe_idx)
2411 {
2412 	struct ocrdma_hdr_wqe *hdr;
2413 	struct ocrdma_sge *rw;
2414 	int opcode;
2415 
2416 	hdr = ocrdma_hwq_head_from_idx(&qp->sq, wqe_idx);
2417 
2418 	ibwc->wr_id = qp->wqe_wr_id_tbl[wqe_idx].wrid;
2419 	/* Undo the hdr->cw swap */
2420 	opcode = le32_to_cpu(hdr->cw) & OCRDMA_WQE_OPCODE_MASK;
2421 	switch (opcode) {
2422 	case OCRDMA_WRITE:
2423 		ibwc->opcode = IB_WC_RDMA_WRITE;
2424 		break;
2425 	case OCRDMA_READ:
2426 		rw = (struct ocrdma_sge *)(hdr + 1);
2427 		ibwc->opcode = IB_WC_RDMA_READ;
2428 		ibwc->byte_len = rw->len;
2429 		break;
2430 	case OCRDMA_SEND:
2431 		ibwc->opcode = IB_WC_SEND;
2432 		break;
2433 	case OCRDMA_FR_MR:
2434 		ibwc->opcode = IB_WC_REG_MR;
2435 		break;
2436 	case OCRDMA_LKEY_INV:
2437 		ibwc->opcode = IB_WC_LOCAL_INV;
2438 		break;
2439 	default:
2440 		ibwc->status = IB_WC_GENERAL_ERR;
2441 		pr_err("%s() invalid opcode received = 0x%x\n",
2442 		       __func__, hdr->cw & OCRDMA_WQE_OPCODE_MASK);
2443 		break;
2444 	}
2445 }
2446 
2447 static void ocrdma_set_cqe_status_flushed(struct ocrdma_qp *qp,
2448 						struct ocrdma_cqe *cqe)
2449 {
2450 	if (is_cqe_for_sq(cqe)) {
2451 		cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2452 				cqe->flags_status_srcqpn) &
2453 					~OCRDMA_CQE_STATUS_MASK);
2454 		cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2455 				cqe->flags_status_srcqpn) |
2456 				(OCRDMA_CQE_WR_FLUSH_ERR <<
2457 					OCRDMA_CQE_STATUS_SHIFT));
2458 	} else {
2459 		if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2460 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2461 					cqe->flags_status_srcqpn) &
2462 						~OCRDMA_CQE_UD_STATUS_MASK);
2463 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2464 					cqe->flags_status_srcqpn) |
2465 					(OCRDMA_CQE_WR_FLUSH_ERR <<
2466 						OCRDMA_CQE_UD_STATUS_SHIFT));
2467 		} else {
2468 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2469 					cqe->flags_status_srcqpn) &
2470 						~OCRDMA_CQE_STATUS_MASK);
2471 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2472 					cqe->flags_status_srcqpn) |
2473 					(OCRDMA_CQE_WR_FLUSH_ERR <<
2474 						OCRDMA_CQE_STATUS_SHIFT));
2475 		}
2476 	}
2477 }
2478 
2479 static bool ocrdma_update_err_cqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2480 				  struct ocrdma_qp *qp, int status)
2481 {
2482 	bool expand = false;
2483 
2484 	ibwc->byte_len = 0;
2485 	ibwc->qp = &qp->ibqp;
2486 	ibwc->status = ocrdma_to_ibwc_err(status);
2487 
2488 	ocrdma_flush_qp(qp);
2489 	ocrdma_qp_state_change(qp, IB_QPS_ERR, NULL);
2490 
2491 	/* if wqe/rqe pending for which cqe needs to be returned,
2492 	 * trigger inflating it.
2493 	 */
2494 	if (!is_hw_rq_empty(qp) || !is_hw_sq_empty(qp)) {
2495 		expand = true;
2496 		ocrdma_set_cqe_status_flushed(qp, cqe);
2497 	}
2498 	return expand;
2499 }
2500 
2501 static int ocrdma_update_err_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2502 				  struct ocrdma_qp *qp, int status)
2503 {
2504 	ibwc->opcode = IB_WC_RECV;
2505 	ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2506 	ocrdma_hwq_inc_tail(&qp->rq);
2507 
2508 	return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2509 }
2510 
2511 static int ocrdma_update_err_scqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2512 				  struct ocrdma_qp *qp, int status)
2513 {
2514 	ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2515 	ocrdma_hwq_inc_tail(&qp->sq);
2516 
2517 	return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2518 }
2519 
2520 
2521 static bool ocrdma_poll_err_scqe(struct ocrdma_qp *qp,
2522 				 struct ocrdma_cqe *cqe, struct ib_wc *ibwc,
2523 				 bool *polled, bool *stop)
2524 {
2525 	bool expand;
2526 	struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2527 	int status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2528 		OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2529 	if (status < OCRDMA_MAX_CQE_ERR)
2530 		atomic_inc(&dev->cqe_err_stats[status]);
2531 
2532 	/* when hw sq is empty, but rq is not empty, so we continue
2533 	 * to keep the cqe in order to get the cq event again.
2534 	 */
2535 	if (is_hw_sq_empty(qp) && !is_hw_rq_empty(qp)) {
2536 		/* when cq for rq and sq is same, it is safe to return
2537 		 * flush cqe for RQEs.
2538 		 */
2539 		if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2540 			*polled = true;
2541 			status = OCRDMA_CQE_WR_FLUSH_ERR;
2542 			expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2543 		} else {
2544 			/* stop processing further cqe as this cqe is used for
2545 			 * triggering cq event on buddy cq of RQ.
2546 			 * When QP is destroyed, this cqe will be removed
2547 			 * from the cq's hardware q.
2548 			 */
2549 			*polled = false;
2550 			*stop = true;
2551 			expand = false;
2552 		}
2553 	} else if (is_hw_sq_empty(qp)) {
2554 		/* Do nothing */
2555 		expand = false;
2556 		*polled = false;
2557 		*stop = false;
2558 	} else {
2559 		*polled = true;
2560 		expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2561 	}
2562 	return expand;
2563 }
2564 
2565 static bool ocrdma_poll_success_scqe(struct ocrdma_qp *qp,
2566 				     struct ocrdma_cqe *cqe,
2567 				     struct ib_wc *ibwc, bool *polled)
2568 {
2569 	bool expand = false;
2570 	int tail = qp->sq.tail;
2571 	u32 wqe_idx;
2572 
2573 	if (!qp->wqe_wr_id_tbl[tail].signaled) {
2574 		*polled = false;    /* WC cannot be consumed yet */
2575 	} else {
2576 		ibwc->status = IB_WC_SUCCESS;
2577 		ibwc->wc_flags = 0;
2578 		ibwc->qp = &qp->ibqp;
2579 		ocrdma_update_wc(qp, ibwc, tail);
2580 		*polled = true;
2581 	}
2582 	wqe_idx = (le32_to_cpu(cqe->wq.wqeidx) &
2583 			OCRDMA_CQE_WQEIDX_MASK) & qp->sq.max_wqe_idx;
2584 	if (tail != wqe_idx)
2585 		expand = true; /* Coalesced CQE can't be consumed yet */
2586 
2587 	ocrdma_hwq_inc_tail(&qp->sq);
2588 	return expand;
2589 }
2590 
2591 static bool ocrdma_poll_scqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2592 			     struct ib_wc *ibwc, bool *polled, bool *stop)
2593 {
2594 	int status;
2595 	bool expand;
2596 
2597 	status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2598 		OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2599 
2600 	if (status == OCRDMA_CQE_SUCCESS)
2601 		expand = ocrdma_poll_success_scqe(qp, cqe, ibwc, polled);
2602 	else
2603 		expand = ocrdma_poll_err_scqe(qp, cqe, ibwc, polled, stop);
2604 	return expand;
2605 }
2606 
2607 static int ocrdma_update_ud_rcqe(struct ocrdma_dev *dev, struct ib_wc *ibwc,
2608 				 struct ocrdma_cqe *cqe)
2609 {
2610 	int status;
2611 	u16 hdr_type = 0;
2612 
2613 	status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2614 		OCRDMA_CQE_UD_STATUS_MASK) >> OCRDMA_CQE_UD_STATUS_SHIFT;
2615 	ibwc->src_qp = le32_to_cpu(cqe->flags_status_srcqpn) &
2616 						OCRDMA_CQE_SRCQP_MASK;
2617 	ibwc->pkey_index = 0;
2618 	ibwc->wc_flags = IB_WC_GRH;
2619 	ibwc->byte_len = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2620 			  OCRDMA_CQE_UD_XFER_LEN_SHIFT) &
2621 			  OCRDMA_CQE_UD_XFER_LEN_MASK;
2622 
2623 	if (ocrdma_is_udp_encap_supported(dev)) {
2624 		hdr_type = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2625 			    OCRDMA_CQE_UD_L3TYPE_SHIFT) &
2626 			    OCRDMA_CQE_UD_L3TYPE_MASK;
2627 		ibwc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
2628 		ibwc->network_hdr_type = hdr_type;
2629 	}
2630 
2631 	return status;
2632 }
2633 
2634 static void ocrdma_update_free_srq_cqe(struct ib_wc *ibwc,
2635 				       struct ocrdma_cqe *cqe,
2636 				       struct ocrdma_qp *qp)
2637 {
2638 	unsigned long flags;
2639 	struct ocrdma_srq *srq;
2640 	u32 wqe_idx;
2641 
2642 	srq = get_ocrdma_srq(qp->ibqp.srq);
2643 	wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
2644 		OCRDMA_CQE_BUFTAG_SHIFT) & srq->rq.max_wqe_idx;
2645 	BUG_ON(wqe_idx < 1);
2646 
2647 	ibwc->wr_id = srq->rqe_wr_id_tbl[wqe_idx];
2648 	spin_lock_irqsave(&srq->q_lock, flags);
2649 	ocrdma_srq_toggle_bit(srq, wqe_idx - 1);
2650 	spin_unlock_irqrestore(&srq->q_lock, flags);
2651 	ocrdma_hwq_inc_tail(&srq->rq);
2652 }
2653 
2654 static bool ocrdma_poll_err_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2655 				struct ib_wc *ibwc, bool *polled, bool *stop,
2656 				int status)
2657 {
2658 	bool expand;
2659 	struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2660 
2661 	if (status < OCRDMA_MAX_CQE_ERR)
2662 		atomic_inc(&dev->cqe_err_stats[status]);
2663 
2664 	/* when hw_rq is empty, but wq is not empty, so continue
2665 	 * to keep the cqe to get the cq event again.
2666 	 */
2667 	if (is_hw_rq_empty(qp) && !is_hw_sq_empty(qp)) {
2668 		if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2669 			*polled = true;
2670 			status = OCRDMA_CQE_WR_FLUSH_ERR;
2671 			expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2672 		} else {
2673 			*polled = false;
2674 			*stop = true;
2675 			expand = false;
2676 		}
2677 	} else if (is_hw_rq_empty(qp)) {
2678 		/* Do nothing */
2679 		expand = false;
2680 		*polled = false;
2681 		*stop = false;
2682 	} else {
2683 		*polled = true;
2684 		expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2685 	}
2686 	return expand;
2687 }
2688 
2689 static void ocrdma_poll_success_rcqe(struct ocrdma_qp *qp,
2690 				     struct ocrdma_cqe *cqe, struct ib_wc *ibwc)
2691 {
2692 	struct ocrdma_dev *dev;
2693 
2694 	dev = get_ocrdma_dev(qp->ibqp.device);
2695 	ibwc->opcode = IB_WC_RECV;
2696 	ibwc->qp = &qp->ibqp;
2697 	ibwc->status = IB_WC_SUCCESS;
2698 
2699 	if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI)
2700 		ocrdma_update_ud_rcqe(dev, ibwc, cqe);
2701 	else
2702 		ibwc->byte_len = le32_to_cpu(cqe->rq.rxlen);
2703 
2704 	if (is_cqe_imm(cqe)) {
2705 		ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2706 		ibwc->wc_flags |= IB_WC_WITH_IMM;
2707 	} else if (is_cqe_wr_imm(cqe)) {
2708 		ibwc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
2709 		ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2710 		ibwc->wc_flags |= IB_WC_WITH_IMM;
2711 	} else if (is_cqe_invalidated(cqe)) {
2712 		ibwc->ex.invalidate_rkey = le32_to_cpu(cqe->rq.lkey_immdt);
2713 		ibwc->wc_flags |= IB_WC_WITH_INVALIDATE;
2714 	}
2715 	if (qp->ibqp.srq) {
2716 		ocrdma_update_free_srq_cqe(ibwc, cqe, qp);
2717 	} else {
2718 		ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2719 		ocrdma_hwq_inc_tail(&qp->rq);
2720 	}
2721 }
2722 
2723 static bool ocrdma_poll_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2724 			     struct ib_wc *ibwc, bool *polled, bool *stop)
2725 {
2726 	int status;
2727 	bool expand = false;
2728 
2729 	ibwc->wc_flags = 0;
2730 	if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2731 		status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2732 					OCRDMA_CQE_UD_STATUS_MASK) >>
2733 					OCRDMA_CQE_UD_STATUS_SHIFT;
2734 	} else {
2735 		status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2736 			     OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2737 	}
2738 
2739 	if (status == OCRDMA_CQE_SUCCESS) {
2740 		*polled = true;
2741 		ocrdma_poll_success_rcqe(qp, cqe, ibwc);
2742 	} else {
2743 		expand = ocrdma_poll_err_rcqe(qp, cqe, ibwc, polled, stop,
2744 					      status);
2745 	}
2746 	return expand;
2747 }
2748 
2749 static void ocrdma_change_cq_phase(struct ocrdma_cq *cq, struct ocrdma_cqe *cqe,
2750 				   u16 cur_getp)
2751 {
2752 	if (cq->phase_change) {
2753 		if (cur_getp == 0)
2754 			cq->phase = (~cq->phase & OCRDMA_CQE_VALID);
2755 	} else {
2756 		/* clear valid bit */
2757 		cqe->flags_status_srcqpn = 0;
2758 	}
2759 }
2760 
2761 static int ocrdma_poll_hwcq(struct ocrdma_cq *cq, int num_entries,
2762 			    struct ib_wc *ibwc)
2763 {
2764 	u16 qpn = 0;
2765 	int i = 0;
2766 	bool expand = false;
2767 	int polled_hw_cqes = 0;
2768 	struct ocrdma_qp *qp = NULL;
2769 	struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
2770 	struct ocrdma_cqe *cqe;
2771 	u16 cur_getp; bool polled = false; bool stop = false;
2772 
2773 	cur_getp = cq->getp;
2774 	while (num_entries) {
2775 		cqe = cq->va + cur_getp;
2776 		/* check whether valid cqe or not */
2777 		if (!is_cqe_valid(cq, cqe))
2778 			break;
2779 		qpn = (le32_to_cpu(cqe->cmn.qpn) & OCRDMA_CQE_QPN_MASK);
2780 		/* ignore discarded cqe */
2781 		if (qpn == 0)
2782 			goto skip_cqe;
2783 		qp = dev->qp_tbl[qpn];
2784 		BUG_ON(qp == NULL);
2785 
2786 		if (is_cqe_for_sq(cqe)) {
2787 			expand = ocrdma_poll_scqe(qp, cqe, ibwc, &polled,
2788 						  &stop);
2789 		} else {
2790 			expand = ocrdma_poll_rcqe(qp, cqe, ibwc, &polled,
2791 						  &stop);
2792 		}
2793 		if (expand)
2794 			goto expand_cqe;
2795 		if (stop)
2796 			goto stop_cqe;
2797 		/* clear qpn to avoid duplicate processing by discard_cqe() */
2798 		cqe->cmn.qpn = 0;
2799 skip_cqe:
2800 		polled_hw_cqes += 1;
2801 		cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
2802 		ocrdma_change_cq_phase(cq, cqe, cur_getp);
2803 expand_cqe:
2804 		if (polled) {
2805 			num_entries -= 1;
2806 			i += 1;
2807 			ibwc = ibwc + 1;
2808 			polled = false;
2809 		}
2810 	}
2811 stop_cqe:
2812 	cq->getp = cur_getp;
2813 
2814 	if (polled_hw_cqes)
2815 		ocrdma_ring_cq_db(dev, cq->id, false, false, polled_hw_cqes);
2816 
2817 	return i;
2818 }
2819 
2820 /* insert error cqe if the QP's SQ or RQ's CQ matches the CQ under poll. */
2821 static int ocrdma_add_err_cqe(struct ocrdma_cq *cq, int num_entries,
2822 			      struct ocrdma_qp *qp, struct ib_wc *ibwc)
2823 {
2824 	int err_cqes = 0;
2825 
2826 	while (num_entries) {
2827 		if (is_hw_sq_empty(qp) && is_hw_rq_empty(qp))
2828 			break;
2829 		if (!is_hw_sq_empty(qp) && qp->sq_cq == cq) {
2830 			ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2831 			ocrdma_hwq_inc_tail(&qp->sq);
2832 		} else if (!is_hw_rq_empty(qp) && qp->rq_cq == cq) {
2833 			ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2834 			ocrdma_hwq_inc_tail(&qp->rq);
2835 		} else {
2836 			return err_cqes;
2837 		}
2838 		ibwc->byte_len = 0;
2839 		ibwc->status = IB_WC_WR_FLUSH_ERR;
2840 		ibwc = ibwc + 1;
2841 		err_cqes += 1;
2842 		num_entries -= 1;
2843 	}
2844 	return err_cqes;
2845 }
2846 
2847 int ocrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
2848 {
2849 	int cqes_to_poll = num_entries;
2850 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2851 	struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2852 	int num_os_cqe = 0, err_cqes = 0;
2853 	struct ocrdma_qp *qp;
2854 	unsigned long flags;
2855 
2856 	/* poll cqes from adapter CQ */
2857 	spin_lock_irqsave(&cq->cq_lock, flags);
2858 	num_os_cqe = ocrdma_poll_hwcq(cq, cqes_to_poll, wc);
2859 	spin_unlock_irqrestore(&cq->cq_lock, flags);
2860 	cqes_to_poll -= num_os_cqe;
2861 
2862 	if (cqes_to_poll) {
2863 		wc = wc + num_os_cqe;
2864 		/* adapter returns single error cqe when qp moves to
2865 		 * error state. So insert error cqes with wc_status as
2866 		 * FLUSHED for pending WQEs and RQEs of QP's SQ and RQ
2867 		 * respectively which uses this CQ.
2868 		 */
2869 		spin_lock_irqsave(&dev->flush_q_lock, flags);
2870 		list_for_each_entry(qp, &cq->sq_head, sq_entry) {
2871 			if (cqes_to_poll == 0)
2872 				break;
2873 			err_cqes = ocrdma_add_err_cqe(cq, cqes_to_poll, qp, wc);
2874 			cqes_to_poll -= err_cqes;
2875 			num_os_cqe += err_cqes;
2876 			wc = wc + err_cqes;
2877 		}
2878 		spin_unlock_irqrestore(&dev->flush_q_lock, flags);
2879 	}
2880 	return num_os_cqe;
2881 }
2882 
2883 int ocrdma_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags cq_flags)
2884 {
2885 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2886 	struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2887 	u16 cq_id;
2888 	unsigned long flags;
2889 	bool arm_needed = false, sol_needed = false;
2890 
2891 	cq_id = cq->id;
2892 
2893 	spin_lock_irqsave(&cq->cq_lock, flags);
2894 	if (cq_flags & IB_CQ_NEXT_COMP || cq_flags & IB_CQ_SOLICITED)
2895 		arm_needed = true;
2896 	if (cq_flags & IB_CQ_SOLICITED)
2897 		sol_needed = true;
2898 
2899 	ocrdma_ring_cq_db(dev, cq_id, arm_needed, sol_needed, 0);
2900 	spin_unlock_irqrestore(&cq->cq_lock, flags);
2901 
2902 	return 0;
2903 }
2904 
2905 struct ib_mr *ocrdma_alloc_mr(struct ib_pd *ibpd, enum ib_mr_type mr_type,
2906 			      u32 max_num_sg, struct ib_udata *udata)
2907 {
2908 	int status;
2909 	struct ocrdma_mr *mr;
2910 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
2911 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
2912 
2913 	if (mr_type != IB_MR_TYPE_MEM_REG)
2914 		return ERR_PTR(-EINVAL);
2915 
2916 	if (max_num_sg > dev->attr.max_pages_per_frmr)
2917 		return ERR_PTR(-EINVAL);
2918 
2919 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
2920 	if (!mr)
2921 		return ERR_PTR(-ENOMEM);
2922 
2923 	mr->pages = kcalloc(max_num_sg, sizeof(u64), GFP_KERNEL);
2924 	if (!mr->pages) {
2925 		status = -ENOMEM;
2926 		goto pl_err;
2927 	}
2928 
2929 	status = ocrdma_get_pbl_info(dev, mr, max_num_sg);
2930 	if (status)
2931 		goto pbl_err;
2932 	mr->hwmr.fr_mr = 1;
2933 	mr->hwmr.remote_rd = 0;
2934 	mr->hwmr.remote_wr = 0;
2935 	mr->hwmr.local_rd = 0;
2936 	mr->hwmr.local_wr = 0;
2937 	mr->hwmr.mw_bind = 0;
2938 	status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
2939 	if (status)
2940 		goto pbl_err;
2941 	status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, 0);
2942 	if (status)
2943 		goto mbx_err;
2944 	mr->ibmr.rkey = mr->hwmr.lkey;
2945 	mr->ibmr.lkey = mr->hwmr.lkey;
2946 	dev->stag_arr[(mr->hwmr.lkey >> 8) & (OCRDMA_MAX_STAG - 1)] =
2947 		(unsigned long) mr;
2948 	return &mr->ibmr;
2949 mbx_err:
2950 	ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
2951 pbl_err:
2952 	kfree(mr->pages);
2953 pl_err:
2954 	kfree(mr);
2955 	return ERR_PTR(-ENOMEM);
2956 }
2957 
2958 static int ocrdma_set_page(struct ib_mr *ibmr, u64 addr)
2959 {
2960 	struct ocrdma_mr *mr = get_ocrdma_mr(ibmr);
2961 
2962 	if (unlikely(mr->npages == mr->hwmr.num_pbes))
2963 		return -ENOMEM;
2964 
2965 	mr->pages[mr->npages++] = addr;
2966 
2967 	return 0;
2968 }
2969 
2970 int ocrdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
2971 		     unsigned int *sg_offset)
2972 {
2973 	struct ocrdma_mr *mr = get_ocrdma_mr(ibmr);
2974 
2975 	mr->npages = 0;
2976 
2977 	return ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, ocrdma_set_page);
2978 }
2979