1 /* This file is part of the Emulex RoCE Device Driver for
2  * RoCE (RDMA over Converged Ethernet) adapters.
3  * Copyright (C) 2012-2015 Emulex. All rights reserved.
4  * EMULEX and SLI are trademarks of Emulex.
5  * www.emulex.com
6  *
7  * This software is available to you under a choice of one of two licenses.
8  * You may choose to be licensed under the terms of the GNU General Public
9  * License (GPL) Version 2, available from the file COPYING in the main
10  * directory of this source tree, or the BSD license below:
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  *
16  * - Redistributions of source code must retain the above copyright notice,
17  *   this list of conditions and the following disclaimer.
18  *
19  * - Redistributions in binary form must reproduce the above copyright
20  *   notice, this list of conditions and the following disclaimer in
21  *   the documentation and/or other materials provided with the distribution.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
24  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
27  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
28  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
29  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
30  * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
31  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
32  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
33  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
34  *
35  * Contact Information:
36  * linux-drivers@emulex.com
37  *
38  * Emulex
39  * 3333 Susan Street
40  * Costa Mesa, CA 92626
41  */
42 
43 #include <linux/dma-mapping.h>
44 #include <rdma/ib_verbs.h>
45 #include <rdma/ib_user_verbs.h>
46 #include <rdma/iw_cm.h>
47 #include <rdma/ib_umem.h>
48 #include <rdma/ib_addr.h>
49 #include <rdma/ib_cache.h>
50 #include <rdma/uverbs_ioctl.h>
51 
52 #include "ocrdma.h"
53 #include "ocrdma_hw.h"
54 #include "ocrdma_verbs.h"
55 #include <rdma/ocrdma-abi.h>
56 
57 int ocrdma_query_pkey(struct ib_device *ibdev, u8 port, u16 index, u16 *pkey)
58 {
59 	if (index > 0)
60 		return -EINVAL;
61 
62 	*pkey = 0xffff;
63 	return 0;
64 }
65 
66 int ocrdma_query_device(struct ib_device *ibdev, struct ib_device_attr *attr,
67 			struct ib_udata *uhw)
68 {
69 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
70 
71 	if (uhw->inlen || uhw->outlen)
72 		return -EINVAL;
73 
74 	memset(attr, 0, sizeof *attr);
75 	memcpy(&attr->fw_ver, &dev->attr.fw_ver[0],
76 	       min(sizeof(dev->attr.fw_ver), sizeof(attr->fw_ver)));
77 	ocrdma_get_guid(dev, (u8 *)&attr->sys_image_guid);
78 	attr->max_mr_size = dev->attr.max_mr_size;
79 	attr->page_size_cap = 0xffff000;
80 	attr->vendor_id = dev->nic_info.pdev->vendor;
81 	attr->vendor_part_id = dev->nic_info.pdev->device;
82 	attr->hw_ver = dev->asic_id;
83 	attr->max_qp = dev->attr.max_qp;
84 	attr->max_ah = OCRDMA_MAX_AH;
85 	attr->max_qp_wr = dev->attr.max_wqe;
86 
87 	attr->device_cap_flags = IB_DEVICE_CURR_QP_STATE_MOD |
88 					IB_DEVICE_RC_RNR_NAK_GEN |
89 					IB_DEVICE_SHUTDOWN_PORT |
90 					IB_DEVICE_SYS_IMAGE_GUID |
91 					IB_DEVICE_LOCAL_DMA_LKEY |
92 					IB_DEVICE_MEM_MGT_EXTENSIONS;
93 	attr->max_send_sge = dev->attr.max_send_sge;
94 	attr->max_recv_sge = dev->attr.max_recv_sge;
95 	attr->max_sge_rd = dev->attr.max_rdma_sge;
96 	attr->max_cq = dev->attr.max_cq;
97 	attr->max_cqe = dev->attr.max_cqe;
98 	attr->max_mr = dev->attr.max_mr;
99 	attr->max_mw = dev->attr.max_mw;
100 	attr->max_pd = dev->attr.max_pd;
101 	attr->atomic_cap = 0;
102 	attr->max_qp_rd_atom =
103 	    min(dev->attr.max_ord_per_qp, dev->attr.max_ird_per_qp);
104 	attr->max_qp_init_rd_atom = dev->attr.max_ord_per_qp;
105 	attr->max_srq = dev->attr.max_srq;
106 	attr->max_srq_sge = dev->attr.max_srq_sge;
107 	attr->max_srq_wr = dev->attr.max_rqe;
108 	attr->local_ca_ack_delay = dev->attr.local_ca_ack_delay;
109 	attr->max_fast_reg_page_list_len = dev->attr.max_pages_per_frmr;
110 	attr->max_pkeys = 1;
111 	return 0;
112 }
113 
114 static inline void get_link_speed_and_width(struct ocrdma_dev *dev,
115 					    u16 *ib_speed, u8 *ib_width)
116 {
117 	int status;
118 	u8 speed;
119 
120 	status = ocrdma_mbx_get_link_speed(dev, &speed, NULL);
121 	if (status)
122 		speed = OCRDMA_PHYS_LINK_SPEED_ZERO;
123 
124 	switch (speed) {
125 	case OCRDMA_PHYS_LINK_SPEED_1GBPS:
126 		*ib_speed = IB_SPEED_SDR;
127 		*ib_width = IB_WIDTH_1X;
128 		break;
129 
130 	case OCRDMA_PHYS_LINK_SPEED_10GBPS:
131 		*ib_speed = IB_SPEED_QDR;
132 		*ib_width = IB_WIDTH_1X;
133 		break;
134 
135 	case OCRDMA_PHYS_LINK_SPEED_20GBPS:
136 		*ib_speed = IB_SPEED_DDR;
137 		*ib_width = IB_WIDTH_4X;
138 		break;
139 
140 	case OCRDMA_PHYS_LINK_SPEED_40GBPS:
141 		*ib_speed = IB_SPEED_QDR;
142 		*ib_width = IB_WIDTH_4X;
143 		break;
144 
145 	default:
146 		/* Unsupported */
147 		*ib_speed = IB_SPEED_SDR;
148 		*ib_width = IB_WIDTH_1X;
149 	}
150 }
151 
152 int ocrdma_query_port(struct ib_device *ibdev,
153 		      u8 port, struct ib_port_attr *props)
154 {
155 	enum ib_port_state port_state;
156 	struct ocrdma_dev *dev;
157 	struct net_device *netdev;
158 
159 	/* props being zeroed by the caller, avoid zeroing it here */
160 	dev = get_ocrdma_dev(ibdev);
161 	netdev = dev->nic_info.netdev;
162 	if (netif_running(netdev) && netif_oper_up(netdev)) {
163 		port_state = IB_PORT_ACTIVE;
164 		props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
165 	} else {
166 		port_state = IB_PORT_DOWN;
167 		props->phys_state = IB_PORT_PHYS_STATE_DISABLED;
168 	}
169 	props->max_mtu = IB_MTU_4096;
170 	props->active_mtu = iboe_get_mtu(netdev->mtu);
171 	props->lid = 0;
172 	props->lmc = 0;
173 	props->sm_lid = 0;
174 	props->sm_sl = 0;
175 	props->state = port_state;
176 	props->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_REINIT_SUP |
177 				IB_PORT_DEVICE_MGMT_SUP |
178 				IB_PORT_VENDOR_CLASS_SUP;
179 	props->ip_gids = true;
180 	props->gid_tbl_len = OCRDMA_MAX_SGID;
181 	props->pkey_tbl_len = 1;
182 	props->bad_pkey_cntr = 0;
183 	props->qkey_viol_cntr = 0;
184 	get_link_speed_and_width(dev, &props->active_speed,
185 				 &props->active_width);
186 	props->max_msg_sz = 0x80000000;
187 	props->max_vl_num = 4;
188 	return 0;
189 }
190 
191 static int ocrdma_add_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
192 			   unsigned long len)
193 {
194 	struct ocrdma_mm *mm;
195 
196 	mm = kzalloc(sizeof(*mm), GFP_KERNEL);
197 	if (mm == NULL)
198 		return -ENOMEM;
199 	mm->key.phy_addr = phy_addr;
200 	mm->key.len = len;
201 	INIT_LIST_HEAD(&mm->entry);
202 
203 	mutex_lock(&uctx->mm_list_lock);
204 	list_add_tail(&mm->entry, &uctx->mm_head);
205 	mutex_unlock(&uctx->mm_list_lock);
206 	return 0;
207 }
208 
209 static void ocrdma_del_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
210 			    unsigned long len)
211 {
212 	struct ocrdma_mm *mm, *tmp;
213 
214 	mutex_lock(&uctx->mm_list_lock);
215 	list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
216 		if (len != mm->key.len && phy_addr != mm->key.phy_addr)
217 			continue;
218 
219 		list_del(&mm->entry);
220 		kfree(mm);
221 		break;
222 	}
223 	mutex_unlock(&uctx->mm_list_lock);
224 }
225 
226 static bool ocrdma_search_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
227 			      unsigned long len)
228 {
229 	bool found = false;
230 	struct ocrdma_mm *mm;
231 
232 	mutex_lock(&uctx->mm_list_lock);
233 	list_for_each_entry(mm, &uctx->mm_head, entry) {
234 		if (len != mm->key.len && phy_addr != mm->key.phy_addr)
235 			continue;
236 
237 		found = true;
238 		break;
239 	}
240 	mutex_unlock(&uctx->mm_list_lock);
241 	return found;
242 }
243 
244 
245 static u16 _ocrdma_pd_mgr_get_bitmap(struct ocrdma_dev *dev, bool dpp_pool)
246 {
247 	u16 pd_bitmap_idx = 0;
248 	const unsigned long *pd_bitmap;
249 
250 	if (dpp_pool) {
251 		pd_bitmap = dev->pd_mgr->pd_dpp_bitmap;
252 		pd_bitmap_idx = find_first_zero_bit(pd_bitmap,
253 						    dev->pd_mgr->max_dpp_pd);
254 		__set_bit(pd_bitmap_idx, dev->pd_mgr->pd_dpp_bitmap);
255 		dev->pd_mgr->pd_dpp_count++;
256 		if (dev->pd_mgr->pd_dpp_count > dev->pd_mgr->pd_dpp_thrsh)
257 			dev->pd_mgr->pd_dpp_thrsh = dev->pd_mgr->pd_dpp_count;
258 	} else {
259 		pd_bitmap = dev->pd_mgr->pd_norm_bitmap;
260 		pd_bitmap_idx = find_first_zero_bit(pd_bitmap,
261 						    dev->pd_mgr->max_normal_pd);
262 		__set_bit(pd_bitmap_idx, dev->pd_mgr->pd_norm_bitmap);
263 		dev->pd_mgr->pd_norm_count++;
264 		if (dev->pd_mgr->pd_norm_count > dev->pd_mgr->pd_norm_thrsh)
265 			dev->pd_mgr->pd_norm_thrsh = dev->pd_mgr->pd_norm_count;
266 	}
267 	return pd_bitmap_idx;
268 }
269 
270 static int _ocrdma_pd_mgr_put_bitmap(struct ocrdma_dev *dev, u16 pd_id,
271 					bool dpp_pool)
272 {
273 	u16 pd_count;
274 	u16 pd_bit_index;
275 
276 	pd_count = dpp_pool ? dev->pd_mgr->pd_dpp_count :
277 			      dev->pd_mgr->pd_norm_count;
278 	if (pd_count == 0)
279 		return -EINVAL;
280 
281 	if (dpp_pool) {
282 		pd_bit_index = pd_id - dev->pd_mgr->pd_dpp_start;
283 		if (pd_bit_index >= dev->pd_mgr->max_dpp_pd) {
284 			return -EINVAL;
285 		} else {
286 			__clear_bit(pd_bit_index, dev->pd_mgr->pd_dpp_bitmap);
287 			dev->pd_mgr->pd_dpp_count--;
288 		}
289 	} else {
290 		pd_bit_index = pd_id - dev->pd_mgr->pd_norm_start;
291 		if (pd_bit_index >= dev->pd_mgr->max_normal_pd) {
292 			return -EINVAL;
293 		} else {
294 			__clear_bit(pd_bit_index, dev->pd_mgr->pd_norm_bitmap);
295 			dev->pd_mgr->pd_norm_count--;
296 		}
297 	}
298 
299 	return 0;
300 }
301 
302 static int ocrdma_put_pd_num(struct ocrdma_dev *dev, u16 pd_id,
303 				   bool dpp_pool)
304 {
305 	int status;
306 
307 	mutex_lock(&dev->dev_lock);
308 	status = _ocrdma_pd_mgr_put_bitmap(dev, pd_id, dpp_pool);
309 	mutex_unlock(&dev->dev_lock);
310 	return status;
311 }
312 
313 static int ocrdma_get_pd_num(struct ocrdma_dev *dev, struct ocrdma_pd *pd)
314 {
315 	u16 pd_idx = 0;
316 	int status = 0;
317 
318 	mutex_lock(&dev->dev_lock);
319 	if (pd->dpp_enabled) {
320 		/* try allocating DPP PD, if not available then normal PD */
321 		if (dev->pd_mgr->pd_dpp_count < dev->pd_mgr->max_dpp_pd) {
322 			pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, true);
323 			pd->id = dev->pd_mgr->pd_dpp_start + pd_idx;
324 			pd->dpp_page = dev->pd_mgr->dpp_page_index + pd_idx;
325 		} else if (dev->pd_mgr->pd_norm_count <
326 			   dev->pd_mgr->max_normal_pd) {
327 			pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, false);
328 			pd->id = dev->pd_mgr->pd_norm_start + pd_idx;
329 			pd->dpp_enabled = false;
330 		} else {
331 			status = -EINVAL;
332 		}
333 	} else {
334 		if (dev->pd_mgr->pd_norm_count < dev->pd_mgr->max_normal_pd) {
335 			pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, false);
336 			pd->id = dev->pd_mgr->pd_norm_start + pd_idx;
337 		} else {
338 			status = -EINVAL;
339 		}
340 	}
341 	mutex_unlock(&dev->dev_lock);
342 	return status;
343 }
344 
345 /*
346  * NOTE:
347  *
348  * ocrdma_ucontext must be used here because this function is also
349  * called from ocrdma_alloc_ucontext where ib_udata does not have
350  * valid ib_ucontext pointer. ib_uverbs_get_context does not call
351  * uobj_{alloc|get_xxx} helpers which are used to store the
352  * ib_ucontext in uverbs_attr_bundle wrapping the ib_udata. so
353  * ib_udata does NOT imply valid ib_ucontext here!
354  */
355 static int _ocrdma_alloc_pd(struct ocrdma_dev *dev, struct ocrdma_pd *pd,
356 			    struct ocrdma_ucontext *uctx,
357 			    struct ib_udata *udata)
358 {
359 	int status;
360 
361 	if (udata && uctx && dev->attr.max_dpp_pds) {
362 		pd->dpp_enabled =
363 			ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R;
364 		pd->num_dpp_qp =
365 			pd->dpp_enabled ? (dev->nic_info.db_page_size /
366 					   dev->attr.wqe_size) : 0;
367 	}
368 
369 	if (dev->pd_mgr->pd_prealloc_valid)
370 		return ocrdma_get_pd_num(dev, pd);
371 
372 retry:
373 	status = ocrdma_mbx_alloc_pd(dev, pd);
374 	if (status) {
375 		if (pd->dpp_enabled) {
376 			pd->dpp_enabled = false;
377 			pd->num_dpp_qp = 0;
378 			goto retry;
379 		}
380 		return status;
381 	}
382 
383 	return 0;
384 }
385 
386 static inline int is_ucontext_pd(struct ocrdma_ucontext *uctx,
387 				 struct ocrdma_pd *pd)
388 {
389 	return (uctx->cntxt_pd == pd);
390 }
391 
392 static void _ocrdma_dealloc_pd(struct ocrdma_dev *dev,
393 			      struct ocrdma_pd *pd)
394 {
395 	if (dev->pd_mgr->pd_prealloc_valid)
396 		ocrdma_put_pd_num(dev, pd->id, pd->dpp_enabled);
397 	else
398 		ocrdma_mbx_dealloc_pd(dev, pd);
399 }
400 
401 static int ocrdma_alloc_ucontext_pd(struct ocrdma_dev *dev,
402 				    struct ocrdma_ucontext *uctx,
403 				    struct ib_udata *udata)
404 {
405 	struct ib_device *ibdev = &dev->ibdev;
406 	struct ib_pd *pd;
407 	int status;
408 
409 	pd = rdma_zalloc_drv_obj(ibdev, ib_pd);
410 	if (!pd)
411 		return -ENOMEM;
412 
413 	pd->device  = ibdev;
414 	uctx->cntxt_pd = get_ocrdma_pd(pd);
415 
416 	status = _ocrdma_alloc_pd(dev, uctx->cntxt_pd, uctx, udata);
417 	if (status) {
418 		kfree(uctx->cntxt_pd);
419 		goto err;
420 	}
421 
422 	uctx->cntxt_pd->uctx = uctx;
423 	uctx->cntxt_pd->ibpd.device = &dev->ibdev;
424 err:
425 	return status;
426 }
427 
428 static void ocrdma_dealloc_ucontext_pd(struct ocrdma_ucontext *uctx)
429 {
430 	struct ocrdma_pd *pd = uctx->cntxt_pd;
431 	struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
432 
433 	if (uctx->pd_in_use) {
434 		pr_err("%s(%d) Freeing in use pdid=0x%x.\n",
435 		       __func__, dev->id, pd->id);
436 	}
437 	uctx->cntxt_pd = NULL;
438 	_ocrdma_dealloc_pd(dev, pd);
439 	kfree(pd);
440 }
441 
442 static struct ocrdma_pd *ocrdma_get_ucontext_pd(struct ocrdma_ucontext *uctx)
443 {
444 	struct ocrdma_pd *pd = NULL;
445 
446 	mutex_lock(&uctx->mm_list_lock);
447 	if (!uctx->pd_in_use) {
448 		uctx->pd_in_use = true;
449 		pd = uctx->cntxt_pd;
450 	}
451 	mutex_unlock(&uctx->mm_list_lock);
452 
453 	return pd;
454 }
455 
456 static void ocrdma_release_ucontext_pd(struct ocrdma_ucontext *uctx)
457 {
458 	mutex_lock(&uctx->mm_list_lock);
459 	uctx->pd_in_use = false;
460 	mutex_unlock(&uctx->mm_list_lock);
461 }
462 
463 int ocrdma_alloc_ucontext(struct ib_ucontext *uctx, struct ib_udata *udata)
464 {
465 	struct ib_device *ibdev = uctx->device;
466 	int status;
467 	struct ocrdma_ucontext *ctx = get_ocrdma_ucontext(uctx);
468 	struct ocrdma_alloc_ucontext_resp resp = {};
469 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
470 	struct pci_dev *pdev = dev->nic_info.pdev;
471 	u32 map_len = roundup(sizeof(u32) * 2048, PAGE_SIZE);
472 
473 	if (!udata)
474 		return -EFAULT;
475 	INIT_LIST_HEAD(&ctx->mm_head);
476 	mutex_init(&ctx->mm_list_lock);
477 
478 	ctx->ah_tbl.va = dma_alloc_coherent(&pdev->dev, map_len,
479 					    &ctx->ah_tbl.pa, GFP_KERNEL);
480 	if (!ctx->ah_tbl.va)
481 		return -ENOMEM;
482 
483 	ctx->ah_tbl.len = map_len;
484 
485 	resp.ah_tbl_len = ctx->ah_tbl.len;
486 	resp.ah_tbl_page = virt_to_phys(ctx->ah_tbl.va);
487 
488 	status = ocrdma_add_mmap(ctx, resp.ah_tbl_page, resp.ah_tbl_len);
489 	if (status)
490 		goto map_err;
491 
492 	status = ocrdma_alloc_ucontext_pd(dev, ctx, udata);
493 	if (status)
494 		goto pd_err;
495 
496 	resp.dev_id = dev->id;
497 	resp.max_inline_data = dev->attr.max_inline_data;
498 	resp.wqe_size = dev->attr.wqe_size;
499 	resp.rqe_size = dev->attr.rqe_size;
500 	resp.dpp_wqe_size = dev->attr.wqe_size;
501 
502 	memcpy(resp.fw_ver, dev->attr.fw_ver, sizeof(resp.fw_ver));
503 	status = ib_copy_to_udata(udata, &resp, sizeof(resp));
504 	if (status)
505 		goto cpy_err;
506 	return 0;
507 
508 cpy_err:
509 	ocrdma_dealloc_ucontext_pd(ctx);
510 pd_err:
511 	ocrdma_del_mmap(ctx, ctx->ah_tbl.pa, ctx->ah_tbl.len);
512 map_err:
513 	dma_free_coherent(&pdev->dev, ctx->ah_tbl.len, ctx->ah_tbl.va,
514 			  ctx->ah_tbl.pa);
515 	return status;
516 }
517 
518 void ocrdma_dealloc_ucontext(struct ib_ucontext *ibctx)
519 {
520 	struct ocrdma_mm *mm, *tmp;
521 	struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ibctx);
522 	struct ocrdma_dev *dev = get_ocrdma_dev(ibctx->device);
523 	struct pci_dev *pdev = dev->nic_info.pdev;
524 
525 	ocrdma_dealloc_ucontext_pd(uctx);
526 
527 	ocrdma_del_mmap(uctx, uctx->ah_tbl.pa, uctx->ah_tbl.len);
528 	dma_free_coherent(&pdev->dev, uctx->ah_tbl.len, uctx->ah_tbl.va,
529 			  uctx->ah_tbl.pa);
530 
531 	list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
532 		list_del(&mm->entry);
533 		kfree(mm);
534 	}
535 }
536 
537 int ocrdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
538 {
539 	struct ocrdma_ucontext *ucontext = get_ocrdma_ucontext(context);
540 	struct ocrdma_dev *dev = get_ocrdma_dev(context->device);
541 	unsigned long vm_page = vma->vm_pgoff << PAGE_SHIFT;
542 	u64 unmapped_db = (u64) dev->nic_info.unmapped_db;
543 	unsigned long len = (vma->vm_end - vma->vm_start);
544 	int status;
545 	bool found;
546 
547 	if (vma->vm_start & (PAGE_SIZE - 1))
548 		return -EINVAL;
549 	found = ocrdma_search_mmap(ucontext, vma->vm_pgoff << PAGE_SHIFT, len);
550 	if (!found)
551 		return -EINVAL;
552 
553 	if ((vm_page >= unmapped_db) && (vm_page <= (unmapped_db +
554 		dev->nic_info.db_total_size)) &&
555 		(len <=	dev->nic_info.db_page_size)) {
556 		if (vma->vm_flags & VM_READ)
557 			return -EPERM;
558 
559 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
560 		status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
561 					    len, vma->vm_page_prot);
562 	} else if (dev->nic_info.dpp_unmapped_len &&
563 		(vm_page >= (u64) dev->nic_info.dpp_unmapped_addr) &&
564 		(vm_page <= (u64) (dev->nic_info.dpp_unmapped_addr +
565 			dev->nic_info.dpp_unmapped_len)) &&
566 		(len <= dev->nic_info.dpp_unmapped_len)) {
567 		if (vma->vm_flags & VM_READ)
568 			return -EPERM;
569 
570 		vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
571 		status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
572 					    len, vma->vm_page_prot);
573 	} else {
574 		status = remap_pfn_range(vma, vma->vm_start,
575 					 vma->vm_pgoff, len, vma->vm_page_prot);
576 	}
577 	return status;
578 }
579 
580 static int ocrdma_copy_pd_uresp(struct ocrdma_dev *dev, struct ocrdma_pd *pd,
581 				struct ib_udata *udata)
582 {
583 	int status;
584 	u64 db_page_addr;
585 	u64 dpp_page_addr = 0;
586 	u32 db_page_size;
587 	struct ocrdma_alloc_pd_uresp rsp;
588 	struct ocrdma_ucontext *uctx = rdma_udata_to_drv_context(
589 		udata, struct ocrdma_ucontext, ibucontext);
590 
591 	memset(&rsp, 0, sizeof(rsp));
592 	rsp.id = pd->id;
593 	rsp.dpp_enabled = pd->dpp_enabled;
594 	db_page_addr = ocrdma_get_db_addr(dev, pd->id);
595 	db_page_size = dev->nic_info.db_page_size;
596 
597 	status = ocrdma_add_mmap(uctx, db_page_addr, db_page_size);
598 	if (status)
599 		return status;
600 
601 	if (pd->dpp_enabled) {
602 		dpp_page_addr = dev->nic_info.dpp_unmapped_addr +
603 				(pd->id * PAGE_SIZE);
604 		status = ocrdma_add_mmap(uctx, dpp_page_addr,
605 				 PAGE_SIZE);
606 		if (status)
607 			goto dpp_map_err;
608 		rsp.dpp_page_addr_hi = upper_32_bits(dpp_page_addr);
609 		rsp.dpp_page_addr_lo = dpp_page_addr;
610 	}
611 
612 	status = ib_copy_to_udata(udata, &rsp, sizeof(rsp));
613 	if (status)
614 		goto ucopy_err;
615 
616 	pd->uctx = uctx;
617 	return 0;
618 
619 ucopy_err:
620 	if (pd->dpp_enabled)
621 		ocrdma_del_mmap(pd->uctx, dpp_page_addr, PAGE_SIZE);
622 dpp_map_err:
623 	ocrdma_del_mmap(pd->uctx, db_page_addr, db_page_size);
624 	return status;
625 }
626 
627 int ocrdma_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
628 {
629 	struct ib_device *ibdev = ibpd->device;
630 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
631 	struct ocrdma_pd *pd;
632 	int status;
633 	u8 is_uctx_pd = false;
634 	struct ocrdma_ucontext *uctx = rdma_udata_to_drv_context(
635 		udata, struct ocrdma_ucontext, ibucontext);
636 
637 	if (udata) {
638 		pd = ocrdma_get_ucontext_pd(uctx);
639 		if (pd) {
640 			is_uctx_pd = true;
641 			goto pd_mapping;
642 		}
643 	}
644 
645 	pd = get_ocrdma_pd(ibpd);
646 	status = _ocrdma_alloc_pd(dev, pd, uctx, udata);
647 	if (status)
648 		goto exit;
649 
650 pd_mapping:
651 	if (udata) {
652 		status = ocrdma_copy_pd_uresp(dev, pd, udata);
653 		if (status)
654 			goto err;
655 	}
656 	return 0;
657 
658 err:
659 	if (is_uctx_pd)
660 		ocrdma_release_ucontext_pd(uctx);
661 	else
662 		_ocrdma_dealloc_pd(dev, pd);
663 exit:
664 	return status;
665 }
666 
667 int ocrdma_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
668 {
669 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
670 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
671 	struct ocrdma_ucontext *uctx = NULL;
672 	u64 usr_db;
673 
674 	uctx = pd->uctx;
675 	if (uctx) {
676 		u64 dpp_db = dev->nic_info.dpp_unmapped_addr +
677 			(pd->id * PAGE_SIZE);
678 		if (pd->dpp_enabled)
679 			ocrdma_del_mmap(pd->uctx, dpp_db, PAGE_SIZE);
680 		usr_db = ocrdma_get_db_addr(dev, pd->id);
681 		ocrdma_del_mmap(pd->uctx, usr_db, dev->nic_info.db_page_size);
682 
683 		if (is_ucontext_pd(uctx, pd)) {
684 			ocrdma_release_ucontext_pd(uctx);
685 			return 0;
686 		}
687 	}
688 	_ocrdma_dealloc_pd(dev, pd);
689 	return 0;
690 }
691 
692 static int ocrdma_alloc_lkey(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
693 			    u32 pdid, int acc, u32 num_pbls, u32 addr_check)
694 {
695 	int status;
696 
697 	mr->hwmr.fr_mr = 0;
698 	mr->hwmr.local_rd = 1;
699 	mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
700 	mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
701 	mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
702 	mr->hwmr.mw_bind = (acc & IB_ACCESS_MW_BIND) ? 1 : 0;
703 	mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
704 	mr->hwmr.num_pbls = num_pbls;
705 
706 	status = ocrdma_mbx_alloc_lkey(dev, &mr->hwmr, pdid, addr_check);
707 	if (status)
708 		return status;
709 
710 	mr->ibmr.lkey = mr->hwmr.lkey;
711 	if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
712 		mr->ibmr.rkey = mr->hwmr.lkey;
713 	return 0;
714 }
715 
716 struct ib_mr *ocrdma_get_dma_mr(struct ib_pd *ibpd, int acc)
717 {
718 	int status;
719 	struct ocrdma_mr *mr;
720 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
721 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
722 
723 	if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE)) {
724 		pr_err("%s err, invalid access rights\n", __func__);
725 		return ERR_PTR(-EINVAL);
726 	}
727 
728 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
729 	if (!mr)
730 		return ERR_PTR(-ENOMEM);
731 
732 	status = ocrdma_alloc_lkey(dev, mr, pd->id, acc, 0,
733 				   OCRDMA_ADDR_CHECK_DISABLE);
734 	if (status) {
735 		kfree(mr);
736 		return ERR_PTR(status);
737 	}
738 
739 	return &mr->ibmr;
740 }
741 
742 static void ocrdma_free_mr_pbl_tbl(struct ocrdma_dev *dev,
743 				   struct ocrdma_hw_mr *mr)
744 {
745 	struct pci_dev *pdev = dev->nic_info.pdev;
746 	int i = 0;
747 
748 	if (mr->pbl_table) {
749 		for (i = 0; i < mr->num_pbls; i++) {
750 			if (!mr->pbl_table[i].va)
751 				continue;
752 			dma_free_coherent(&pdev->dev, mr->pbl_size,
753 					  mr->pbl_table[i].va,
754 					  mr->pbl_table[i].pa);
755 		}
756 		kfree(mr->pbl_table);
757 		mr->pbl_table = NULL;
758 	}
759 }
760 
761 static int ocrdma_get_pbl_info(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
762 			      u32 num_pbes)
763 {
764 	u32 num_pbls = 0;
765 	u32 idx = 0;
766 	int status = 0;
767 	u32 pbl_size;
768 
769 	do {
770 		pbl_size = OCRDMA_MIN_HPAGE_SIZE * (1 << idx);
771 		if (pbl_size > MAX_OCRDMA_PBL_SIZE) {
772 			status = -EFAULT;
773 			break;
774 		}
775 		num_pbls = roundup(num_pbes, (pbl_size / sizeof(u64)));
776 		num_pbls = num_pbls / (pbl_size / sizeof(u64));
777 		idx++;
778 	} while (num_pbls >= dev->attr.max_num_mr_pbl);
779 
780 	mr->hwmr.num_pbes = num_pbes;
781 	mr->hwmr.num_pbls = num_pbls;
782 	mr->hwmr.pbl_size = pbl_size;
783 	return status;
784 }
785 
786 static int ocrdma_build_pbl_tbl(struct ocrdma_dev *dev, struct ocrdma_hw_mr *mr)
787 {
788 	int status = 0;
789 	int i;
790 	u32 dma_len = mr->pbl_size;
791 	struct pci_dev *pdev = dev->nic_info.pdev;
792 	void *va;
793 	dma_addr_t pa;
794 
795 	mr->pbl_table = kcalloc(mr->num_pbls, sizeof(struct ocrdma_pbl),
796 				GFP_KERNEL);
797 
798 	if (!mr->pbl_table)
799 		return -ENOMEM;
800 
801 	for (i = 0; i < mr->num_pbls; i++) {
802 		va = dma_alloc_coherent(&pdev->dev, dma_len, &pa, GFP_KERNEL);
803 		if (!va) {
804 			ocrdma_free_mr_pbl_tbl(dev, mr);
805 			status = -ENOMEM;
806 			break;
807 		}
808 		mr->pbl_table[i].va = va;
809 		mr->pbl_table[i].pa = pa;
810 	}
811 	return status;
812 }
813 
814 static void build_user_pbes(struct ocrdma_dev *dev, struct ocrdma_mr *mr)
815 {
816 	struct ocrdma_pbe *pbe;
817 	struct ib_block_iter biter;
818 	struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
819 	int pbe_cnt;
820 	u64 pg_addr;
821 
822 	if (!mr->hwmr.num_pbes)
823 		return;
824 
825 	pbe = (struct ocrdma_pbe *)pbl_tbl->va;
826 	pbe_cnt = 0;
827 
828 	rdma_umem_for_each_dma_block (mr->umem, &biter, PAGE_SIZE) {
829 		/* store the page address in pbe */
830 		pg_addr = rdma_block_iter_dma_address(&biter);
831 		pbe->pa_lo = cpu_to_le32(pg_addr);
832 		pbe->pa_hi = cpu_to_le32(upper_32_bits(pg_addr));
833 		pbe_cnt += 1;
834 		pbe++;
835 
836 		/* if the given pbl is full storing the pbes,
837 		 * move to next pbl.
838 		 */
839 		if (pbe_cnt == (mr->hwmr.pbl_size / sizeof(u64))) {
840 			pbl_tbl++;
841 			pbe = (struct ocrdma_pbe *)pbl_tbl->va;
842 			pbe_cnt = 0;
843 		}
844 	}
845 }
846 
847 struct ib_mr *ocrdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
848 				 u64 usr_addr, int acc, struct ib_udata *udata)
849 {
850 	int status = -ENOMEM;
851 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
852 	struct ocrdma_mr *mr;
853 	struct ocrdma_pd *pd;
854 
855 	pd = get_ocrdma_pd(ibpd);
856 
857 	if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE))
858 		return ERR_PTR(-EINVAL);
859 
860 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
861 	if (!mr)
862 		return ERR_PTR(status);
863 	mr->umem = ib_umem_get(ibpd->device, start, len, acc);
864 	if (IS_ERR(mr->umem)) {
865 		status = -EFAULT;
866 		goto umem_err;
867 	}
868 	status = ocrdma_get_pbl_info(
869 		dev, mr, ib_umem_num_dma_blocks(mr->umem, PAGE_SIZE));
870 	if (status)
871 		goto umem_err;
872 
873 	mr->hwmr.pbe_size = PAGE_SIZE;
874 	mr->hwmr.va = usr_addr;
875 	mr->hwmr.len = len;
876 	mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
877 	mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
878 	mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
879 	mr->hwmr.local_rd = 1;
880 	mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
881 	status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
882 	if (status)
883 		goto umem_err;
884 	build_user_pbes(dev, mr);
885 	status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
886 	if (status)
887 		goto mbx_err;
888 	mr->ibmr.lkey = mr->hwmr.lkey;
889 	if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
890 		mr->ibmr.rkey = mr->hwmr.lkey;
891 
892 	return &mr->ibmr;
893 
894 mbx_err:
895 	ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
896 umem_err:
897 	kfree(mr);
898 	return ERR_PTR(status);
899 }
900 
901 int ocrdma_dereg_mr(struct ib_mr *ib_mr, struct ib_udata *udata)
902 {
903 	struct ocrdma_mr *mr = get_ocrdma_mr(ib_mr);
904 	struct ocrdma_dev *dev = get_ocrdma_dev(ib_mr->device);
905 
906 	(void) ocrdma_mbx_dealloc_lkey(dev, mr->hwmr.fr_mr, mr->hwmr.lkey);
907 
908 	kfree(mr->pages);
909 	ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
910 
911 	/* it could be user registered memory. */
912 	ib_umem_release(mr->umem);
913 	kfree(mr);
914 
915 	/* Don't stop cleanup, in case FW is unresponsive */
916 	if (dev->mqe_ctx.fw_error_state) {
917 		pr_err("%s(%d) fw not responding.\n",
918 		       __func__, dev->id);
919 	}
920 	return 0;
921 }
922 
923 static int ocrdma_copy_cq_uresp(struct ocrdma_dev *dev, struct ocrdma_cq *cq,
924 				struct ib_udata *udata)
925 {
926 	int status;
927 	struct ocrdma_ucontext *uctx = rdma_udata_to_drv_context(
928 		udata, struct ocrdma_ucontext, ibucontext);
929 	struct ocrdma_create_cq_uresp uresp;
930 
931 	/* this must be user flow! */
932 	if (!udata)
933 		return -EINVAL;
934 
935 	memset(&uresp, 0, sizeof(uresp));
936 	uresp.cq_id = cq->id;
937 	uresp.page_size = PAGE_ALIGN(cq->len);
938 	uresp.num_pages = 1;
939 	uresp.max_hw_cqe = cq->max_hw_cqe;
940 	uresp.page_addr[0] = virt_to_phys(cq->va);
941 	uresp.db_page_addr =  ocrdma_get_db_addr(dev, uctx->cntxt_pd->id);
942 	uresp.db_page_size = dev->nic_info.db_page_size;
943 	uresp.phase_change = cq->phase_change ? 1 : 0;
944 	status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
945 	if (status) {
946 		pr_err("%s(%d) copy error cqid=0x%x.\n",
947 		       __func__, dev->id, cq->id);
948 		goto err;
949 	}
950 	status = ocrdma_add_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
951 	if (status)
952 		goto err;
953 	status = ocrdma_add_mmap(uctx, uresp.page_addr[0], uresp.page_size);
954 	if (status) {
955 		ocrdma_del_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
956 		goto err;
957 	}
958 	cq->ucontext = uctx;
959 err:
960 	return status;
961 }
962 
963 int ocrdma_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
964 		     struct ib_udata *udata)
965 {
966 	struct ib_device *ibdev = ibcq->device;
967 	int entries = attr->cqe;
968 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
969 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
970 	struct ocrdma_ucontext *uctx = rdma_udata_to_drv_context(
971 		udata, struct ocrdma_ucontext, ibucontext);
972 	u16 pd_id = 0;
973 	int status;
974 	struct ocrdma_create_cq_ureq ureq;
975 
976 	if (attr->flags)
977 		return -EOPNOTSUPP;
978 
979 	if (udata) {
980 		if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
981 			return -EFAULT;
982 	} else
983 		ureq.dpp_cq = 0;
984 
985 	spin_lock_init(&cq->cq_lock);
986 	spin_lock_init(&cq->comp_handler_lock);
987 	INIT_LIST_HEAD(&cq->sq_head);
988 	INIT_LIST_HEAD(&cq->rq_head);
989 
990 	if (udata)
991 		pd_id = uctx->cntxt_pd->id;
992 
993 	status = ocrdma_mbx_create_cq(dev, cq, entries, ureq.dpp_cq, pd_id);
994 	if (status)
995 		return status;
996 
997 	if (udata) {
998 		status = ocrdma_copy_cq_uresp(dev, cq, udata);
999 		if (status)
1000 			goto ctx_err;
1001 	}
1002 	cq->phase = OCRDMA_CQE_VALID;
1003 	dev->cq_tbl[cq->id] = cq;
1004 	return 0;
1005 
1006 ctx_err:
1007 	ocrdma_mbx_destroy_cq(dev, cq);
1008 	return status;
1009 }
1010 
1011 int ocrdma_resize_cq(struct ib_cq *ibcq, int new_cnt,
1012 		     struct ib_udata *udata)
1013 {
1014 	int status = 0;
1015 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
1016 
1017 	if (new_cnt < 1 || new_cnt > cq->max_hw_cqe) {
1018 		status = -EINVAL;
1019 		return status;
1020 	}
1021 	ibcq->cqe = new_cnt;
1022 	return status;
1023 }
1024 
1025 static void ocrdma_flush_cq(struct ocrdma_cq *cq)
1026 {
1027 	int cqe_cnt;
1028 	int valid_count = 0;
1029 	unsigned long flags;
1030 
1031 	struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
1032 	struct ocrdma_cqe *cqe = NULL;
1033 
1034 	cqe = cq->va;
1035 	cqe_cnt = cq->cqe_cnt;
1036 
1037 	/* Last irq might have scheduled a polling thread
1038 	 * sync-up with it before hard flushing.
1039 	 */
1040 	spin_lock_irqsave(&cq->cq_lock, flags);
1041 	while (cqe_cnt) {
1042 		if (is_cqe_valid(cq, cqe))
1043 			valid_count++;
1044 		cqe++;
1045 		cqe_cnt--;
1046 	}
1047 	ocrdma_ring_cq_db(dev, cq->id, false, false, valid_count);
1048 	spin_unlock_irqrestore(&cq->cq_lock, flags);
1049 }
1050 
1051 int ocrdma_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
1052 {
1053 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
1054 	struct ocrdma_eq *eq = NULL;
1055 	struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
1056 	int pdid = 0;
1057 	u32 irq, indx;
1058 
1059 	dev->cq_tbl[cq->id] = NULL;
1060 	indx = ocrdma_get_eq_table_index(dev, cq->eqn);
1061 
1062 	eq = &dev->eq_tbl[indx];
1063 	irq = ocrdma_get_irq(dev, eq);
1064 	synchronize_irq(irq);
1065 	ocrdma_flush_cq(cq);
1066 
1067 	ocrdma_mbx_destroy_cq(dev, cq);
1068 	if (cq->ucontext) {
1069 		pdid = cq->ucontext->cntxt_pd->id;
1070 		ocrdma_del_mmap(cq->ucontext, (u64) cq->pa,
1071 				PAGE_ALIGN(cq->len));
1072 		ocrdma_del_mmap(cq->ucontext,
1073 				ocrdma_get_db_addr(dev, pdid),
1074 				dev->nic_info.db_page_size);
1075 	}
1076 	return 0;
1077 }
1078 
1079 static int ocrdma_add_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1080 {
1081 	int status = -EINVAL;
1082 
1083 	if (qp->id < OCRDMA_MAX_QP && dev->qp_tbl[qp->id] == NULL) {
1084 		dev->qp_tbl[qp->id] = qp;
1085 		status = 0;
1086 	}
1087 	return status;
1088 }
1089 
1090 static void ocrdma_del_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1091 {
1092 	dev->qp_tbl[qp->id] = NULL;
1093 }
1094 
1095 static int ocrdma_check_qp_params(struct ib_pd *ibpd, struct ocrdma_dev *dev,
1096 				  struct ib_qp_init_attr *attrs,
1097 				  struct ib_udata *udata)
1098 {
1099 	if ((attrs->qp_type != IB_QPT_GSI) &&
1100 	    (attrs->qp_type != IB_QPT_RC) &&
1101 	    (attrs->qp_type != IB_QPT_UC) &&
1102 	    (attrs->qp_type != IB_QPT_UD)) {
1103 		pr_err("%s(%d) unsupported qp type=0x%x requested\n",
1104 		       __func__, dev->id, attrs->qp_type);
1105 		return -EOPNOTSUPP;
1106 	}
1107 	/* Skip the check for QP1 to support CM size of 128 */
1108 	if ((attrs->qp_type != IB_QPT_GSI) &&
1109 	    (attrs->cap.max_send_wr > dev->attr.max_wqe)) {
1110 		pr_err("%s(%d) unsupported send_wr=0x%x requested\n",
1111 		       __func__, dev->id, attrs->cap.max_send_wr);
1112 		pr_err("%s(%d) supported send_wr=0x%x\n",
1113 		       __func__, dev->id, dev->attr.max_wqe);
1114 		return -EINVAL;
1115 	}
1116 	if (!attrs->srq && (attrs->cap.max_recv_wr > dev->attr.max_rqe)) {
1117 		pr_err("%s(%d) unsupported recv_wr=0x%x requested\n",
1118 		       __func__, dev->id, attrs->cap.max_recv_wr);
1119 		pr_err("%s(%d) supported recv_wr=0x%x\n",
1120 		       __func__, dev->id, dev->attr.max_rqe);
1121 		return -EINVAL;
1122 	}
1123 	if (attrs->cap.max_inline_data > dev->attr.max_inline_data) {
1124 		pr_err("%s(%d) unsupported inline data size=0x%x requested\n",
1125 		       __func__, dev->id, attrs->cap.max_inline_data);
1126 		pr_err("%s(%d) supported inline data size=0x%x\n",
1127 		       __func__, dev->id, dev->attr.max_inline_data);
1128 		return -EINVAL;
1129 	}
1130 	if (attrs->cap.max_send_sge > dev->attr.max_send_sge) {
1131 		pr_err("%s(%d) unsupported send_sge=0x%x requested\n",
1132 		       __func__, dev->id, attrs->cap.max_send_sge);
1133 		pr_err("%s(%d) supported send_sge=0x%x\n",
1134 		       __func__, dev->id, dev->attr.max_send_sge);
1135 		return -EINVAL;
1136 	}
1137 	if (attrs->cap.max_recv_sge > dev->attr.max_recv_sge) {
1138 		pr_err("%s(%d) unsupported recv_sge=0x%x requested\n",
1139 		       __func__, dev->id, attrs->cap.max_recv_sge);
1140 		pr_err("%s(%d) supported recv_sge=0x%x\n",
1141 		       __func__, dev->id, dev->attr.max_recv_sge);
1142 		return -EINVAL;
1143 	}
1144 	/* unprivileged user space cannot create special QP */
1145 	if (udata && attrs->qp_type == IB_QPT_GSI) {
1146 		pr_err
1147 		    ("%s(%d) Userspace can't create special QPs of type=0x%x\n",
1148 		     __func__, dev->id, attrs->qp_type);
1149 		return -EINVAL;
1150 	}
1151 	/* allow creating only one GSI type of QP */
1152 	if (attrs->qp_type == IB_QPT_GSI && dev->gsi_qp_created) {
1153 		pr_err("%s(%d) GSI special QPs already created.\n",
1154 		       __func__, dev->id);
1155 		return -EINVAL;
1156 	}
1157 	/* verify consumer QPs are not trying to use GSI QP's CQ */
1158 	if ((attrs->qp_type != IB_QPT_GSI) && (dev->gsi_qp_created)) {
1159 		if ((dev->gsi_sqcq == get_ocrdma_cq(attrs->send_cq)) ||
1160 			(dev->gsi_rqcq == get_ocrdma_cq(attrs->recv_cq))) {
1161 			pr_err("%s(%d) Consumer QP cannot use GSI CQs.\n",
1162 				__func__, dev->id);
1163 			return -EINVAL;
1164 		}
1165 	}
1166 	return 0;
1167 }
1168 
1169 static int ocrdma_copy_qp_uresp(struct ocrdma_qp *qp,
1170 				struct ib_udata *udata, int dpp_offset,
1171 				int dpp_credit_lmt, int srq)
1172 {
1173 	int status;
1174 	u64 usr_db;
1175 	struct ocrdma_create_qp_uresp uresp;
1176 	struct ocrdma_pd *pd = qp->pd;
1177 	struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
1178 
1179 	memset(&uresp, 0, sizeof(uresp));
1180 	usr_db = dev->nic_info.unmapped_db +
1181 			(pd->id * dev->nic_info.db_page_size);
1182 	uresp.qp_id = qp->id;
1183 	uresp.sq_dbid = qp->sq.dbid;
1184 	uresp.num_sq_pages = 1;
1185 	uresp.sq_page_size = PAGE_ALIGN(qp->sq.len);
1186 	uresp.sq_page_addr[0] = virt_to_phys(qp->sq.va);
1187 	uresp.num_wqe_allocated = qp->sq.max_cnt;
1188 	if (!srq) {
1189 		uresp.rq_dbid = qp->rq.dbid;
1190 		uresp.num_rq_pages = 1;
1191 		uresp.rq_page_size = PAGE_ALIGN(qp->rq.len);
1192 		uresp.rq_page_addr[0] = virt_to_phys(qp->rq.va);
1193 		uresp.num_rqe_allocated = qp->rq.max_cnt;
1194 	}
1195 	uresp.db_page_addr = usr_db;
1196 	uresp.db_page_size = dev->nic_info.db_page_size;
1197 	uresp.db_sq_offset = OCRDMA_DB_GEN2_SQ_OFFSET;
1198 	uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1199 	uresp.db_shift = OCRDMA_DB_RQ_SHIFT;
1200 
1201 	if (qp->dpp_enabled) {
1202 		uresp.dpp_credit = dpp_credit_lmt;
1203 		uresp.dpp_offset = dpp_offset;
1204 	}
1205 	status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1206 	if (status) {
1207 		pr_err("%s(%d) user copy error.\n", __func__, dev->id);
1208 		goto err;
1209 	}
1210 	status = ocrdma_add_mmap(pd->uctx, uresp.sq_page_addr[0],
1211 				 uresp.sq_page_size);
1212 	if (status)
1213 		goto err;
1214 
1215 	if (!srq) {
1216 		status = ocrdma_add_mmap(pd->uctx, uresp.rq_page_addr[0],
1217 					 uresp.rq_page_size);
1218 		if (status)
1219 			goto rq_map_err;
1220 	}
1221 	return status;
1222 rq_map_err:
1223 	ocrdma_del_mmap(pd->uctx, uresp.sq_page_addr[0], uresp.sq_page_size);
1224 err:
1225 	return status;
1226 }
1227 
1228 static void ocrdma_set_qp_db(struct ocrdma_dev *dev, struct ocrdma_qp *qp,
1229 			     struct ocrdma_pd *pd)
1230 {
1231 	if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1232 		qp->sq_db = dev->nic_info.db +
1233 			(pd->id * dev->nic_info.db_page_size) +
1234 			OCRDMA_DB_GEN2_SQ_OFFSET;
1235 		qp->rq_db = dev->nic_info.db +
1236 			(pd->id * dev->nic_info.db_page_size) +
1237 			OCRDMA_DB_GEN2_RQ_OFFSET;
1238 	} else {
1239 		qp->sq_db = dev->nic_info.db +
1240 			(pd->id * dev->nic_info.db_page_size) +
1241 			OCRDMA_DB_SQ_OFFSET;
1242 		qp->rq_db = dev->nic_info.db +
1243 			(pd->id * dev->nic_info.db_page_size) +
1244 			OCRDMA_DB_RQ_OFFSET;
1245 	}
1246 }
1247 
1248 static int ocrdma_alloc_wr_id_tbl(struct ocrdma_qp *qp)
1249 {
1250 	qp->wqe_wr_id_tbl =
1251 	    kcalloc(qp->sq.max_cnt, sizeof(*(qp->wqe_wr_id_tbl)),
1252 		    GFP_KERNEL);
1253 	if (qp->wqe_wr_id_tbl == NULL)
1254 		return -ENOMEM;
1255 	qp->rqe_wr_id_tbl =
1256 	    kcalloc(qp->rq.max_cnt, sizeof(u64), GFP_KERNEL);
1257 	if (qp->rqe_wr_id_tbl == NULL)
1258 		return -ENOMEM;
1259 
1260 	return 0;
1261 }
1262 
1263 static void ocrdma_set_qp_init_params(struct ocrdma_qp *qp,
1264 				      struct ocrdma_pd *pd,
1265 				      struct ib_qp_init_attr *attrs)
1266 {
1267 	qp->pd = pd;
1268 	spin_lock_init(&qp->q_lock);
1269 	INIT_LIST_HEAD(&qp->sq_entry);
1270 	INIT_LIST_HEAD(&qp->rq_entry);
1271 
1272 	qp->qp_type = attrs->qp_type;
1273 	qp->cap_flags = OCRDMA_QP_INB_RD | OCRDMA_QP_INB_WR;
1274 	qp->max_inline_data = attrs->cap.max_inline_data;
1275 	qp->sq.max_sges = attrs->cap.max_send_sge;
1276 	qp->rq.max_sges = attrs->cap.max_recv_sge;
1277 	qp->state = OCRDMA_QPS_RST;
1278 	qp->signaled = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR) ? true : false;
1279 }
1280 
1281 static void ocrdma_store_gsi_qp_cq(struct ocrdma_dev *dev,
1282 				   struct ib_qp_init_attr *attrs)
1283 {
1284 	if (attrs->qp_type == IB_QPT_GSI) {
1285 		dev->gsi_qp_created = 1;
1286 		dev->gsi_sqcq = get_ocrdma_cq(attrs->send_cq);
1287 		dev->gsi_rqcq = get_ocrdma_cq(attrs->recv_cq);
1288 	}
1289 }
1290 
1291 struct ib_qp *ocrdma_create_qp(struct ib_pd *ibpd,
1292 			       struct ib_qp_init_attr *attrs,
1293 			       struct ib_udata *udata)
1294 {
1295 	int status;
1296 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1297 	struct ocrdma_qp *qp;
1298 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1299 	struct ocrdma_create_qp_ureq ureq;
1300 	u16 dpp_credit_lmt, dpp_offset;
1301 
1302 	if (attrs->create_flags)
1303 		return ERR_PTR(-EOPNOTSUPP);
1304 
1305 	status = ocrdma_check_qp_params(ibpd, dev, attrs, udata);
1306 	if (status)
1307 		goto gen_err;
1308 
1309 	memset(&ureq, 0, sizeof(ureq));
1310 	if (udata) {
1311 		if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1312 			return ERR_PTR(-EFAULT);
1313 	}
1314 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1315 	if (!qp) {
1316 		status = -ENOMEM;
1317 		goto gen_err;
1318 	}
1319 	ocrdma_set_qp_init_params(qp, pd, attrs);
1320 	if (udata == NULL)
1321 		qp->cap_flags |= (OCRDMA_QP_MW_BIND | OCRDMA_QP_LKEY0 |
1322 					OCRDMA_QP_FAST_REG);
1323 
1324 	mutex_lock(&dev->dev_lock);
1325 	status = ocrdma_mbx_create_qp(qp, attrs, ureq.enable_dpp_cq,
1326 					ureq.dpp_cq_id,
1327 					&dpp_offset, &dpp_credit_lmt);
1328 	if (status)
1329 		goto mbx_err;
1330 
1331 	/* user space QP's wr_id table are managed in library */
1332 	if (udata == NULL) {
1333 		status = ocrdma_alloc_wr_id_tbl(qp);
1334 		if (status)
1335 			goto map_err;
1336 	}
1337 
1338 	status = ocrdma_add_qpn_map(dev, qp);
1339 	if (status)
1340 		goto map_err;
1341 	ocrdma_set_qp_db(dev, qp, pd);
1342 	if (udata) {
1343 		status = ocrdma_copy_qp_uresp(qp, udata, dpp_offset,
1344 					      dpp_credit_lmt,
1345 					      (attrs->srq != NULL));
1346 		if (status)
1347 			goto cpy_err;
1348 	}
1349 	ocrdma_store_gsi_qp_cq(dev, attrs);
1350 	qp->ibqp.qp_num = qp->id;
1351 	mutex_unlock(&dev->dev_lock);
1352 	return &qp->ibqp;
1353 
1354 cpy_err:
1355 	ocrdma_del_qpn_map(dev, qp);
1356 map_err:
1357 	ocrdma_mbx_destroy_qp(dev, qp);
1358 mbx_err:
1359 	mutex_unlock(&dev->dev_lock);
1360 	kfree(qp->wqe_wr_id_tbl);
1361 	kfree(qp->rqe_wr_id_tbl);
1362 	kfree(qp);
1363 	pr_err("%s(%d) error=%d\n", __func__, dev->id, status);
1364 gen_err:
1365 	return ERR_PTR(status);
1366 }
1367 
1368 int _ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1369 		      int attr_mask)
1370 {
1371 	int status = 0;
1372 	struct ocrdma_qp *qp;
1373 	struct ocrdma_dev *dev;
1374 	enum ib_qp_state old_qps;
1375 
1376 	qp = get_ocrdma_qp(ibqp);
1377 	dev = get_ocrdma_dev(ibqp->device);
1378 	if (attr_mask & IB_QP_STATE)
1379 		status = ocrdma_qp_state_change(qp, attr->qp_state, &old_qps);
1380 	/* if new and previous states are same hw doesn't need to
1381 	 * know about it.
1382 	 */
1383 	if (status < 0)
1384 		return status;
1385 	return ocrdma_mbx_modify_qp(dev, qp, attr, attr_mask);
1386 }
1387 
1388 int ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1389 		     int attr_mask, struct ib_udata *udata)
1390 {
1391 	unsigned long flags;
1392 	int status = -EINVAL;
1393 	struct ocrdma_qp *qp;
1394 	struct ocrdma_dev *dev;
1395 	enum ib_qp_state old_qps, new_qps;
1396 
1397 	if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1398 		return -EOPNOTSUPP;
1399 
1400 	qp = get_ocrdma_qp(ibqp);
1401 	dev = get_ocrdma_dev(ibqp->device);
1402 
1403 	/* syncronize with multiple context trying to change, retrive qps */
1404 	mutex_lock(&dev->dev_lock);
1405 	/* syncronize with wqe, rqe posting and cqe processing contexts */
1406 	spin_lock_irqsave(&qp->q_lock, flags);
1407 	old_qps = get_ibqp_state(qp->state);
1408 	if (attr_mask & IB_QP_STATE)
1409 		new_qps = attr->qp_state;
1410 	else
1411 		new_qps = old_qps;
1412 	spin_unlock_irqrestore(&qp->q_lock, flags);
1413 
1414 	if (!ib_modify_qp_is_ok(old_qps, new_qps, ibqp->qp_type, attr_mask)) {
1415 		pr_err("%s(%d) invalid attribute mask=0x%x specified for\n"
1416 		       "qpn=0x%x of type=0x%x old_qps=0x%x, new_qps=0x%x\n",
1417 		       __func__, dev->id, attr_mask, qp->id, ibqp->qp_type,
1418 		       old_qps, new_qps);
1419 		goto param_err;
1420 	}
1421 
1422 	status = _ocrdma_modify_qp(ibqp, attr, attr_mask);
1423 	if (status > 0)
1424 		status = 0;
1425 param_err:
1426 	mutex_unlock(&dev->dev_lock);
1427 	return status;
1428 }
1429 
1430 static enum ib_mtu ocrdma_mtu_int_to_enum(u16 mtu)
1431 {
1432 	switch (mtu) {
1433 	case 256:
1434 		return IB_MTU_256;
1435 	case 512:
1436 		return IB_MTU_512;
1437 	case 1024:
1438 		return IB_MTU_1024;
1439 	case 2048:
1440 		return IB_MTU_2048;
1441 	case 4096:
1442 		return IB_MTU_4096;
1443 	default:
1444 		return IB_MTU_1024;
1445 	}
1446 }
1447 
1448 static int ocrdma_to_ib_qp_acc_flags(int qp_cap_flags)
1449 {
1450 	int ib_qp_acc_flags = 0;
1451 
1452 	if (qp_cap_flags & OCRDMA_QP_INB_WR)
1453 		ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE;
1454 	if (qp_cap_flags & OCRDMA_QP_INB_RD)
1455 		ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE;
1456 	return ib_qp_acc_flags;
1457 }
1458 
1459 int ocrdma_query_qp(struct ib_qp *ibqp,
1460 		    struct ib_qp_attr *qp_attr,
1461 		    int attr_mask, struct ib_qp_init_attr *qp_init_attr)
1462 {
1463 	int status;
1464 	u32 qp_state;
1465 	struct ocrdma_qp_params params;
1466 	struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
1467 	struct ocrdma_dev *dev = get_ocrdma_dev(ibqp->device);
1468 
1469 	memset(&params, 0, sizeof(params));
1470 	mutex_lock(&dev->dev_lock);
1471 	status = ocrdma_mbx_query_qp(dev, qp, &params);
1472 	mutex_unlock(&dev->dev_lock);
1473 	if (status)
1474 		goto mbx_err;
1475 	if (qp->qp_type == IB_QPT_UD)
1476 		qp_attr->qkey = params.qkey;
1477 	qp_attr->path_mtu =
1478 		ocrdma_mtu_int_to_enum(params.path_mtu_pkey_indx &
1479 				OCRDMA_QP_PARAMS_PATH_MTU_MASK) >>
1480 				OCRDMA_QP_PARAMS_PATH_MTU_SHIFT;
1481 	qp_attr->path_mig_state = IB_MIG_MIGRATED;
1482 	qp_attr->rq_psn = params.hop_lmt_rq_psn & OCRDMA_QP_PARAMS_RQ_PSN_MASK;
1483 	qp_attr->sq_psn = params.tclass_sq_psn & OCRDMA_QP_PARAMS_SQ_PSN_MASK;
1484 	qp_attr->dest_qp_num =
1485 	    params.ack_to_rnr_rtc_dest_qpn & OCRDMA_QP_PARAMS_DEST_QPN_MASK;
1486 
1487 	qp_attr->qp_access_flags = ocrdma_to_ib_qp_acc_flags(qp->cap_flags);
1488 	qp_attr->cap.max_send_wr = qp->sq.max_cnt - 1;
1489 	qp_attr->cap.max_recv_wr = qp->rq.max_cnt - 1;
1490 	qp_attr->cap.max_send_sge = qp->sq.max_sges;
1491 	qp_attr->cap.max_recv_sge = qp->rq.max_sges;
1492 	qp_attr->cap.max_inline_data = qp->max_inline_data;
1493 	qp_init_attr->cap = qp_attr->cap;
1494 	qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
1495 
1496 	rdma_ah_set_grh(&qp_attr->ah_attr, NULL,
1497 			params.rnt_rc_sl_fl &
1498 			  OCRDMA_QP_PARAMS_FLOW_LABEL_MASK,
1499 			qp->sgid_idx,
1500 			(params.hop_lmt_rq_psn &
1501 			 OCRDMA_QP_PARAMS_HOP_LMT_MASK) >>
1502 			 OCRDMA_QP_PARAMS_HOP_LMT_SHIFT,
1503 			(params.tclass_sq_psn &
1504 			 OCRDMA_QP_PARAMS_TCLASS_MASK) >>
1505 			 OCRDMA_QP_PARAMS_TCLASS_SHIFT);
1506 	rdma_ah_set_dgid_raw(&qp_attr->ah_attr, &params.dgid[0]);
1507 
1508 	rdma_ah_set_port_num(&qp_attr->ah_attr, 1);
1509 	rdma_ah_set_sl(&qp_attr->ah_attr, (params.rnt_rc_sl_fl &
1510 					   OCRDMA_QP_PARAMS_SL_MASK) >>
1511 					   OCRDMA_QP_PARAMS_SL_SHIFT);
1512 	qp_attr->timeout = (params.ack_to_rnr_rtc_dest_qpn &
1513 			    OCRDMA_QP_PARAMS_ACK_TIMEOUT_MASK) >>
1514 				OCRDMA_QP_PARAMS_ACK_TIMEOUT_SHIFT;
1515 	qp_attr->rnr_retry = (params.ack_to_rnr_rtc_dest_qpn &
1516 			      OCRDMA_QP_PARAMS_RNR_RETRY_CNT_MASK) >>
1517 				OCRDMA_QP_PARAMS_RNR_RETRY_CNT_SHIFT;
1518 	qp_attr->retry_cnt =
1519 	    (params.rnt_rc_sl_fl & OCRDMA_QP_PARAMS_RETRY_CNT_MASK) >>
1520 		OCRDMA_QP_PARAMS_RETRY_CNT_SHIFT;
1521 	qp_attr->min_rnr_timer = 0;
1522 	qp_attr->pkey_index = 0;
1523 	qp_attr->port_num = 1;
1524 	rdma_ah_set_path_bits(&qp_attr->ah_attr, 0);
1525 	rdma_ah_set_static_rate(&qp_attr->ah_attr, 0);
1526 	qp_attr->alt_pkey_index = 0;
1527 	qp_attr->alt_port_num = 0;
1528 	qp_attr->alt_timeout = 0;
1529 	memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr));
1530 	qp_state = (params.max_sge_recv_flags & OCRDMA_QP_PARAMS_STATE_MASK) >>
1531 		    OCRDMA_QP_PARAMS_STATE_SHIFT;
1532 	qp_attr->qp_state = get_ibqp_state(qp_state);
1533 	qp_attr->cur_qp_state = qp_attr->qp_state;
1534 	qp_attr->sq_draining = (qp_state == OCRDMA_QPS_SQ_DRAINING) ? 1 : 0;
1535 	qp_attr->max_dest_rd_atomic =
1536 	    params.max_ord_ird >> OCRDMA_QP_PARAMS_MAX_ORD_SHIFT;
1537 	qp_attr->max_rd_atomic =
1538 	    params.max_ord_ird & OCRDMA_QP_PARAMS_MAX_IRD_MASK;
1539 	qp_attr->en_sqd_async_notify = (params.max_sge_recv_flags &
1540 				OCRDMA_QP_PARAMS_FLAGS_SQD_ASYNC) ? 1 : 0;
1541 	/* Sync driver QP state with FW */
1542 	ocrdma_qp_state_change(qp, qp_attr->qp_state, NULL);
1543 mbx_err:
1544 	return status;
1545 }
1546 
1547 static void ocrdma_srq_toggle_bit(struct ocrdma_srq *srq, unsigned int idx)
1548 {
1549 	unsigned int i = idx / 32;
1550 	u32 mask = (1U << (idx % 32));
1551 
1552 	srq->idx_bit_fields[i] ^= mask;
1553 }
1554 
1555 static int ocrdma_hwq_free_cnt(struct ocrdma_qp_hwq_info *q)
1556 {
1557 	return ((q->max_wqe_idx - q->head) + q->tail) % q->max_cnt;
1558 }
1559 
1560 static int is_hw_sq_empty(struct ocrdma_qp *qp)
1561 {
1562 	return (qp->sq.tail == qp->sq.head);
1563 }
1564 
1565 static int is_hw_rq_empty(struct ocrdma_qp *qp)
1566 {
1567 	return (qp->rq.tail == qp->rq.head);
1568 }
1569 
1570 static void *ocrdma_hwq_head(struct ocrdma_qp_hwq_info *q)
1571 {
1572 	return q->va + (q->head * q->entry_size);
1573 }
1574 
1575 static void *ocrdma_hwq_head_from_idx(struct ocrdma_qp_hwq_info *q,
1576 				      u32 idx)
1577 {
1578 	return q->va + (idx * q->entry_size);
1579 }
1580 
1581 static void ocrdma_hwq_inc_head(struct ocrdma_qp_hwq_info *q)
1582 {
1583 	q->head = (q->head + 1) & q->max_wqe_idx;
1584 }
1585 
1586 static void ocrdma_hwq_inc_tail(struct ocrdma_qp_hwq_info *q)
1587 {
1588 	q->tail = (q->tail + 1) & q->max_wqe_idx;
1589 }
1590 
1591 /* discard the cqe for a given QP */
1592 static void ocrdma_discard_cqes(struct ocrdma_qp *qp, struct ocrdma_cq *cq)
1593 {
1594 	unsigned long cq_flags;
1595 	unsigned long flags;
1596 	int discard_cnt = 0;
1597 	u32 cur_getp, stop_getp;
1598 	struct ocrdma_cqe *cqe;
1599 	u32 qpn = 0, wqe_idx = 0;
1600 
1601 	spin_lock_irqsave(&cq->cq_lock, cq_flags);
1602 
1603 	/* traverse through the CQEs in the hw CQ,
1604 	 * find the matching CQE for a given qp,
1605 	 * mark the matching one discarded by clearing qpn.
1606 	 * ring the doorbell in the poll_cq() as
1607 	 * we don't complete out of order cqe.
1608 	 */
1609 
1610 	cur_getp = cq->getp;
1611 	/* find upto when do we reap the cq. */
1612 	stop_getp = cur_getp;
1613 	do {
1614 		if (is_hw_sq_empty(qp) && (!qp->srq && is_hw_rq_empty(qp)))
1615 			break;
1616 
1617 		cqe = cq->va + cur_getp;
1618 		/* if (a) done reaping whole hw cq, or
1619 		 *    (b) qp_xq becomes empty.
1620 		 * then exit
1621 		 */
1622 		qpn = cqe->cmn.qpn & OCRDMA_CQE_QPN_MASK;
1623 		/* if previously discarded cqe found, skip that too. */
1624 		/* check for matching qp */
1625 		if (qpn == 0 || qpn != qp->id)
1626 			goto skip_cqe;
1627 
1628 		if (is_cqe_for_sq(cqe)) {
1629 			ocrdma_hwq_inc_tail(&qp->sq);
1630 		} else {
1631 			if (qp->srq) {
1632 				wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
1633 					OCRDMA_CQE_BUFTAG_SHIFT) &
1634 					qp->srq->rq.max_wqe_idx;
1635 				BUG_ON(wqe_idx < 1);
1636 				spin_lock_irqsave(&qp->srq->q_lock, flags);
1637 				ocrdma_hwq_inc_tail(&qp->srq->rq);
1638 				ocrdma_srq_toggle_bit(qp->srq, wqe_idx - 1);
1639 				spin_unlock_irqrestore(&qp->srq->q_lock, flags);
1640 
1641 			} else {
1642 				ocrdma_hwq_inc_tail(&qp->rq);
1643 			}
1644 		}
1645 		/* mark cqe discarded so that it is not picked up later
1646 		 * in the poll_cq().
1647 		 */
1648 		discard_cnt += 1;
1649 		cqe->cmn.qpn = 0;
1650 skip_cqe:
1651 		cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
1652 	} while (cur_getp != stop_getp);
1653 	spin_unlock_irqrestore(&cq->cq_lock, cq_flags);
1654 }
1655 
1656 void ocrdma_del_flush_qp(struct ocrdma_qp *qp)
1657 {
1658 	int found = false;
1659 	unsigned long flags;
1660 	struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
1661 	/* sync with any active CQ poll */
1662 
1663 	spin_lock_irqsave(&dev->flush_q_lock, flags);
1664 	found = ocrdma_is_qp_in_sq_flushlist(qp->sq_cq, qp);
1665 	if (found)
1666 		list_del(&qp->sq_entry);
1667 	if (!qp->srq) {
1668 		found = ocrdma_is_qp_in_rq_flushlist(qp->rq_cq, qp);
1669 		if (found)
1670 			list_del(&qp->rq_entry);
1671 	}
1672 	spin_unlock_irqrestore(&dev->flush_q_lock, flags);
1673 }
1674 
1675 int ocrdma_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
1676 {
1677 	struct ocrdma_pd *pd;
1678 	struct ocrdma_qp *qp;
1679 	struct ocrdma_dev *dev;
1680 	struct ib_qp_attr attrs;
1681 	int attr_mask;
1682 	unsigned long flags;
1683 
1684 	qp = get_ocrdma_qp(ibqp);
1685 	dev = get_ocrdma_dev(ibqp->device);
1686 
1687 	pd = qp->pd;
1688 
1689 	/* change the QP state to ERROR */
1690 	if (qp->state != OCRDMA_QPS_RST) {
1691 		attrs.qp_state = IB_QPS_ERR;
1692 		attr_mask = IB_QP_STATE;
1693 		_ocrdma_modify_qp(ibqp, &attrs, attr_mask);
1694 	}
1695 	/* ensure that CQEs for newly created QP (whose id may be same with
1696 	 * one which just getting destroyed are same), dont get
1697 	 * discarded until the old CQEs are discarded.
1698 	 */
1699 	mutex_lock(&dev->dev_lock);
1700 	(void) ocrdma_mbx_destroy_qp(dev, qp);
1701 
1702 	/*
1703 	 * acquire CQ lock while destroy is in progress, in order to
1704 	 * protect against proessing in-flight CQEs for this QP.
1705 	 */
1706 	spin_lock_irqsave(&qp->sq_cq->cq_lock, flags);
1707 	if (qp->rq_cq && (qp->rq_cq != qp->sq_cq)) {
1708 		spin_lock(&qp->rq_cq->cq_lock);
1709 		ocrdma_del_qpn_map(dev, qp);
1710 		spin_unlock(&qp->rq_cq->cq_lock);
1711 	} else {
1712 		ocrdma_del_qpn_map(dev, qp);
1713 	}
1714 	spin_unlock_irqrestore(&qp->sq_cq->cq_lock, flags);
1715 
1716 	if (!pd->uctx) {
1717 		ocrdma_discard_cqes(qp, qp->sq_cq);
1718 		ocrdma_discard_cqes(qp, qp->rq_cq);
1719 	}
1720 	mutex_unlock(&dev->dev_lock);
1721 
1722 	if (pd->uctx) {
1723 		ocrdma_del_mmap(pd->uctx, (u64) qp->sq.pa,
1724 				PAGE_ALIGN(qp->sq.len));
1725 		if (!qp->srq)
1726 			ocrdma_del_mmap(pd->uctx, (u64) qp->rq.pa,
1727 					PAGE_ALIGN(qp->rq.len));
1728 	}
1729 
1730 	ocrdma_del_flush_qp(qp);
1731 
1732 	kfree(qp->wqe_wr_id_tbl);
1733 	kfree(qp->rqe_wr_id_tbl);
1734 	kfree(qp);
1735 	return 0;
1736 }
1737 
1738 static int ocrdma_copy_srq_uresp(struct ocrdma_dev *dev, struct ocrdma_srq *srq,
1739 				struct ib_udata *udata)
1740 {
1741 	int status;
1742 	struct ocrdma_create_srq_uresp uresp;
1743 
1744 	memset(&uresp, 0, sizeof(uresp));
1745 	uresp.rq_dbid = srq->rq.dbid;
1746 	uresp.num_rq_pages = 1;
1747 	uresp.rq_page_addr[0] = virt_to_phys(srq->rq.va);
1748 	uresp.rq_page_size = srq->rq.len;
1749 	uresp.db_page_addr = dev->nic_info.unmapped_db +
1750 	    (srq->pd->id * dev->nic_info.db_page_size);
1751 	uresp.db_page_size = dev->nic_info.db_page_size;
1752 	uresp.num_rqe_allocated = srq->rq.max_cnt;
1753 	if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1754 		uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1755 		uresp.db_shift = 24;
1756 	} else {
1757 		uresp.db_rq_offset = OCRDMA_DB_RQ_OFFSET;
1758 		uresp.db_shift = 16;
1759 	}
1760 
1761 	status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1762 	if (status)
1763 		return status;
1764 	status = ocrdma_add_mmap(srq->pd->uctx, uresp.rq_page_addr[0],
1765 				 uresp.rq_page_size);
1766 	if (status)
1767 		return status;
1768 	return status;
1769 }
1770 
1771 int ocrdma_create_srq(struct ib_srq *ibsrq, struct ib_srq_init_attr *init_attr,
1772 		      struct ib_udata *udata)
1773 {
1774 	int status;
1775 	struct ocrdma_pd *pd = get_ocrdma_pd(ibsrq->pd);
1776 	struct ocrdma_dev *dev = get_ocrdma_dev(ibsrq->device);
1777 	struct ocrdma_srq *srq = get_ocrdma_srq(ibsrq);
1778 
1779 	if (init_attr->srq_type != IB_SRQT_BASIC)
1780 		return -EOPNOTSUPP;
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 int 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 	return 0;
1878 }
1879 
1880 /* unprivileged verbs and their support functions. */
1881 static void ocrdma_build_ud_hdr(struct ocrdma_qp *qp,
1882 				struct ocrdma_hdr_wqe *hdr,
1883 				const struct ib_send_wr *wr)
1884 {
1885 	struct ocrdma_ewqe_ud_hdr *ud_hdr =
1886 		(struct ocrdma_ewqe_ud_hdr *)(hdr + 1);
1887 	struct ocrdma_ah *ah = get_ocrdma_ah(ud_wr(wr)->ah);
1888 
1889 	ud_hdr->rsvd_dest_qpn = ud_wr(wr)->remote_qpn;
1890 	if (qp->qp_type == IB_QPT_GSI)
1891 		ud_hdr->qkey = qp->qkey;
1892 	else
1893 		ud_hdr->qkey = ud_wr(wr)->remote_qkey;
1894 	ud_hdr->rsvd_ahid = ah->id;
1895 	ud_hdr->hdr_type = ah->hdr_type;
1896 	if (ah->av->valid & OCRDMA_AV_VLAN_VALID)
1897 		hdr->cw |= (OCRDMA_FLAG_AH_VLAN_PR << OCRDMA_WQE_FLAGS_SHIFT);
1898 }
1899 
1900 static void ocrdma_build_sges(struct ocrdma_hdr_wqe *hdr,
1901 			      struct ocrdma_sge *sge, int num_sge,
1902 			      struct ib_sge *sg_list)
1903 {
1904 	int i;
1905 
1906 	for (i = 0; i < num_sge; i++) {
1907 		sge[i].lrkey = sg_list[i].lkey;
1908 		sge[i].addr_lo = sg_list[i].addr;
1909 		sge[i].addr_hi = upper_32_bits(sg_list[i].addr);
1910 		sge[i].len = sg_list[i].length;
1911 		hdr->total_len += sg_list[i].length;
1912 	}
1913 	if (num_sge == 0)
1914 		memset(sge, 0, sizeof(*sge));
1915 }
1916 
1917 static inline uint32_t ocrdma_sglist_len(struct ib_sge *sg_list, int num_sge)
1918 {
1919 	uint32_t total_len = 0, i;
1920 
1921 	for (i = 0; i < num_sge; i++)
1922 		total_len += sg_list[i].length;
1923 	return total_len;
1924 }
1925 
1926 
1927 static int ocrdma_build_inline_sges(struct ocrdma_qp *qp,
1928 				    struct ocrdma_hdr_wqe *hdr,
1929 				    struct ocrdma_sge *sge,
1930 				    const struct ib_send_wr *wr, u32 wqe_size)
1931 {
1932 	int i;
1933 	char *dpp_addr;
1934 
1935 	if (wr->send_flags & IB_SEND_INLINE && qp->qp_type != IB_QPT_UD) {
1936 		hdr->total_len = ocrdma_sglist_len(wr->sg_list, wr->num_sge);
1937 		if (unlikely(hdr->total_len > qp->max_inline_data)) {
1938 			pr_err("%s() supported_len=0x%x,\n"
1939 			       " unsupported len req=0x%x\n", __func__,
1940 				qp->max_inline_data, hdr->total_len);
1941 			return -EINVAL;
1942 		}
1943 		dpp_addr = (char *)sge;
1944 		for (i = 0; i < wr->num_sge; i++) {
1945 			memcpy(dpp_addr,
1946 			       (void *)(unsigned long)wr->sg_list[i].addr,
1947 			       wr->sg_list[i].length);
1948 			dpp_addr += wr->sg_list[i].length;
1949 		}
1950 
1951 		wqe_size += roundup(hdr->total_len, OCRDMA_WQE_ALIGN_BYTES);
1952 		if (0 == hdr->total_len)
1953 			wqe_size += sizeof(struct ocrdma_sge);
1954 		hdr->cw |= (OCRDMA_TYPE_INLINE << OCRDMA_WQE_TYPE_SHIFT);
1955 	} else {
1956 		ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
1957 		if (wr->num_sge)
1958 			wqe_size += (wr->num_sge * sizeof(struct ocrdma_sge));
1959 		else
1960 			wqe_size += sizeof(struct ocrdma_sge);
1961 		hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
1962 	}
1963 	hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
1964 	return 0;
1965 }
1966 
1967 static int ocrdma_build_send(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1968 			     const struct ib_send_wr *wr)
1969 {
1970 	int status;
1971 	struct ocrdma_sge *sge;
1972 	u32 wqe_size = sizeof(*hdr);
1973 
1974 	if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
1975 		ocrdma_build_ud_hdr(qp, hdr, wr);
1976 		sge = (struct ocrdma_sge *)(hdr + 2);
1977 		wqe_size += sizeof(struct ocrdma_ewqe_ud_hdr);
1978 	} else {
1979 		sge = (struct ocrdma_sge *)(hdr + 1);
1980 	}
1981 
1982 	status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
1983 	return status;
1984 }
1985 
1986 static int ocrdma_build_write(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1987 			      const struct ib_send_wr *wr)
1988 {
1989 	int status;
1990 	struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
1991 	struct ocrdma_sge *sge = ext_rw + 1;
1992 	u32 wqe_size = sizeof(*hdr) + sizeof(*ext_rw);
1993 
1994 	status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
1995 	if (status)
1996 		return status;
1997 	ext_rw->addr_lo = rdma_wr(wr)->remote_addr;
1998 	ext_rw->addr_hi = upper_32_bits(rdma_wr(wr)->remote_addr);
1999 	ext_rw->lrkey = rdma_wr(wr)->rkey;
2000 	ext_rw->len = hdr->total_len;
2001 	return 0;
2002 }
2003 
2004 static void ocrdma_build_read(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2005 			      const struct ib_send_wr *wr)
2006 {
2007 	struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
2008 	struct ocrdma_sge *sge = ext_rw + 1;
2009 	u32 wqe_size = ((wr->num_sge + 1) * sizeof(struct ocrdma_sge)) +
2010 	    sizeof(struct ocrdma_hdr_wqe);
2011 
2012 	ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
2013 	hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2014 	hdr->cw |= (OCRDMA_READ << OCRDMA_WQE_OPCODE_SHIFT);
2015 	hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2016 
2017 	ext_rw->addr_lo = rdma_wr(wr)->remote_addr;
2018 	ext_rw->addr_hi = upper_32_bits(rdma_wr(wr)->remote_addr);
2019 	ext_rw->lrkey = rdma_wr(wr)->rkey;
2020 	ext_rw->len = hdr->total_len;
2021 }
2022 
2023 static int get_encoded_page_size(int pg_sz)
2024 {
2025 	/* Max size is 256M 4096 << 16 */
2026 	int i = 0;
2027 	for (; i < 17; i++)
2028 		if (pg_sz == (4096 << i))
2029 			break;
2030 	return i;
2031 }
2032 
2033 static int ocrdma_build_reg(struct ocrdma_qp *qp,
2034 			    struct ocrdma_hdr_wqe *hdr,
2035 			    const struct ib_reg_wr *wr)
2036 {
2037 	u64 fbo;
2038 	struct ocrdma_ewqe_fr *fast_reg = (struct ocrdma_ewqe_fr *)(hdr + 1);
2039 	struct ocrdma_mr *mr = get_ocrdma_mr(wr->mr);
2040 	struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
2041 	struct ocrdma_pbe *pbe;
2042 	u32 wqe_size = sizeof(*fast_reg) + sizeof(*hdr);
2043 	int num_pbes = 0, i;
2044 
2045 	wqe_size = roundup(wqe_size, OCRDMA_WQE_ALIGN_BYTES);
2046 
2047 	hdr->cw |= (OCRDMA_FR_MR << OCRDMA_WQE_OPCODE_SHIFT);
2048 	hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2049 
2050 	if (wr->access & IB_ACCESS_LOCAL_WRITE)
2051 		hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_LOCAL_WR;
2052 	if (wr->access & IB_ACCESS_REMOTE_WRITE)
2053 		hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_WR;
2054 	if (wr->access & IB_ACCESS_REMOTE_READ)
2055 		hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_RD;
2056 	hdr->lkey = wr->key;
2057 	hdr->total_len = mr->ibmr.length;
2058 
2059 	fbo = mr->ibmr.iova - mr->pages[0];
2060 
2061 	fast_reg->va_hi = upper_32_bits(mr->ibmr.iova);
2062 	fast_reg->va_lo = (u32) (mr->ibmr.iova & 0xffffffff);
2063 	fast_reg->fbo_hi = upper_32_bits(fbo);
2064 	fast_reg->fbo_lo = (u32) fbo & 0xffffffff;
2065 	fast_reg->num_sges = mr->npages;
2066 	fast_reg->size_sge = get_encoded_page_size(mr->ibmr.page_size);
2067 
2068 	pbe = pbl_tbl->va;
2069 	for (i = 0; i < mr->npages; i++) {
2070 		u64 buf_addr = mr->pages[i];
2071 
2072 		pbe->pa_lo = cpu_to_le32((u32) (buf_addr & PAGE_MASK));
2073 		pbe->pa_hi = cpu_to_le32((u32) upper_32_bits(buf_addr));
2074 		num_pbes += 1;
2075 		pbe++;
2076 
2077 		/* if the pbl is full storing the pbes,
2078 		 * move to next pbl.
2079 		*/
2080 		if (num_pbes == (mr->hwmr.pbl_size/sizeof(u64))) {
2081 			pbl_tbl++;
2082 			pbe = (struct ocrdma_pbe *)pbl_tbl->va;
2083 		}
2084 	}
2085 
2086 	return 0;
2087 }
2088 
2089 static void ocrdma_ring_sq_db(struct ocrdma_qp *qp)
2090 {
2091 	u32 val = qp->sq.dbid | (1 << OCRDMA_DB_SQ_SHIFT);
2092 
2093 	iowrite32(val, qp->sq_db);
2094 }
2095 
2096 int ocrdma_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
2097 		     const struct ib_send_wr **bad_wr)
2098 {
2099 	int status = 0;
2100 	struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2101 	struct ocrdma_hdr_wqe *hdr;
2102 	unsigned long flags;
2103 
2104 	spin_lock_irqsave(&qp->q_lock, flags);
2105 	if (qp->state != OCRDMA_QPS_RTS && qp->state != OCRDMA_QPS_SQD) {
2106 		spin_unlock_irqrestore(&qp->q_lock, flags);
2107 		*bad_wr = wr;
2108 		return -EINVAL;
2109 	}
2110 
2111 	while (wr) {
2112 		if (qp->qp_type == IB_QPT_UD &&
2113 		    (wr->opcode != IB_WR_SEND &&
2114 		     wr->opcode != IB_WR_SEND_WITH_IMM)) {
2115 			*bad_wr = wr;
2116 			status = -EINVAL;
2117 			break;
2118 		}
2119 		if (ocrdma_hwq_free_cnt(&qp->sq) == 0 ||
2120 		    wr->num_sge > qp->sq.max_sges) {
2121 			*bad_wr = wr;
2122 			status = -ENOMEM;
2123 			break;
2124 		}
2125 		hdr = ocrdma_hwq_head(&qp->sq);
2126 		hdr->cw = 0;
2127 		if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2128 			hdr->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2129 		if (wr->send_flags & IB_SEND_FENCE)
2130 			hdr->cw |=
2131 			    (OCRDMA_FLAG_FENCE_L << OCRDMA_WQE_FLAGS_SHIFT);
2132 		if (wr->send_flags & IB_SEND_SOLICITED)
2133 			hdr->cw |=
2134 			    (OCRDMA_FLAG_SOLICIT << OCRDMA_WQE_FLAGS_SHIFT);
2135 		hdr->total_len = 0;
2136 		switch (wr->opcode) {
2137 		case IB_WR_SEND_WITH_IMM:
2138 			hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2139 			hdr->immdt = ntohl(wr->ex.imm_data);
2140 			fallthrough;
2141 		case IB_WR_SEND:
2142 			hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2143 			ocrdma_build_send(qp, hdr, wr);
2144 			break;
2145 		case IB_WR_SEND_WITH_INV:
2146 			hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
2147 			hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2148 			hdr->lkey = wr->ex.invalidate_rkey;
2149 			status = ocrdma_build_send(qp, hdr, wr);
2150 			break;
2151 		case IB_WR_RDMA_WRITE_WITH_IMM:
2152 			hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2153 			hdr->immdt = ntohl(wr->ex.imm_data);
2154 			fallthrough;
2155 		case IB_WR_RDMA_WRITE:
2156 			hdr->cw |= (OCRDMA_WRITE << OCRDMA_WQE_OPCODE_SHIFT);
2157 			status = ocrdma_build_write(qp, hdr, wr);
2158 			break;
2159 		case IB_WR_RDMA_READ:
2160 			ocrdma_build_read(qp, hdr, wr);
2161 			break;
2162 		case IB_WR_LOCAL_INV:
2163 			hdr->cw |=
2164 			    (OCRDMA_LKEY_INV << OCRDMA_WQE_OPCODE_SHIFT);
2165 			hdr->cw |= ((sizeof(struct ocrdma_hdr_wqe) +
2166 					sizeof(struct ocrdma_sge)) /
2167 				OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT;
2168 			hdr->lkey = wr->ex.invalidate_rkey;
2169 			break;
2170 		case IB_WR_REG_MR:
2171 			status = ocrdma_build_reg(qp, hdr, reg_wr(wr));
2172 			break;
2173 		default:
2174 			status = -EINVAL;
2175 			break;
2176 		}
2177 		if (status) {
2178 			*bad_wr = wr;
2179 			break;
2180 		}
2181 		if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2182 			qp->wqe_wr_id_tbl[qp->sq.head].signaled = 1;
2183 		else
2184 			qp->wqe_wr_id_tbl[qp->sq.head].signaled = 0;
2185 		qp->wqe_wr_id_tbl[qp->sq.head].wrid = wr->wr_id;
2186 		ocrdma_cpu_to_le32(hdr, ((hdr->cw >> OCRDMA_WQE_SIZE_SHIFT) &
2187 				   OCRDMA_WQE_SIZE_MASK) * OCRDMA_WQE_STRIDE);
2188 		/* make sure wqe is written before adapter can access it */
2189 		wmb();
2190 		/* inform hw to start processing it */
2191 		ocrdma_ring_sq_db(qp);
2192 
2193 		/* update pointer, counter for next wr */
2194 		ocrdma_hwq_inc_head(&qp->sq);
2195 		wr = wr->next;
2196 	}
2197 	spin_unlock_irqrestore(&qp->q_lock, flags);
2198 	return status;
2199 }
2200 
2201 static void ocrdma_ring_rq_db(struct ocrdma_qp *qp)
2202 {
2203 	u32 val = qp->rq.dbid | (1 << OCRDMA_DB_RQ_SHIFT);
2204 
2205 	iowrite32(val, qp->rq_db);
2206 }
2207 
2208 static void ocrdma_build_rqe(struct ocrdma_hdr_wqe *rqe,
2209 			     const struct ib_recv_wr *wr, u16 tag)
2210 {
2211 	u32 wqe_size = 0;
2212 	struct ocrdma_sge *sge;
2213 	if (wr->num_sge)
2214 		wqe_size = (wr->num_sge * sizeof(*sge)) + sizeof(*rqe);
2215 	else
2216 		wqe_size = sizeof(*sge) + sizeof(*rqe);
2217 
2218 	rqe->cw = ((wqe_size / OCRDMA_WQE_STRIDE) <<
2219 				OCRDMA_WQE_SIZE_SHIFT);
2220 	rqe->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2221 	rqe->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2222 	rqe->total_len = 0;
2223 	rqe->rsvd_tag = tag;
2224 	sge = (struct ocrdma_sge *)(rqe + 1);
2225 	ocrdma_build_sges(rqe, sge, wr->num_sge, wr->sg_list);
2226 	ocrdma_cpu_to_le32(rqe, wqe_size);
2227 }
2228 
2229 int ocrdma_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
2230 		     const struct ib_recv_wr **bad_wr)
2231 {
2232 	int status = 0;
2233 	unsigned long flags;
2234 	struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2235 	struct ocrdma_hdr_wqe *rqe;
2236 
2237 	spin_lock_irqsave(&qp->q_lock, flags);
2238 	if (qp->state == OCRDMA_QPS_RST || qp->state == OCRDMA_QPS_ERR) {
2239 		spin_unlock_irqrestore(&qp->q_lock, flags);
2240 		*bad_wr = wr;
2241 		return -EINVAL;
2242 	}
2243 	while (wr) {
2244 		if (ocrdma_hwq_free_cnt(&qp->rq) == 0 ||
2245 		    wr->num_sge > qp->rq.max_sges) {
2246 			*bad_wr = wr;
2247 			status = -ENOMEM;
2248 			break;
2249 		}
2250 		rqe = ocrdma_hwq_head(&qp->rq);
2251 		ocrdma_build_rqe(rqe, wr, 0);
2252 
2253 		qp->rqe_wr_id_tbl[qp->rq.head] = wr->wr_id;
2254 		/* make sure rqe is written before adapter can access it */
2255 		wmb();
2256 
2257 		/* inform hw to start processing it */
2258 		ocrdma_ring_rq_db(qp);
2259 
2260 		/* update pointer, counter for next wr */
2261 		ocrdma_hwq_inc_head(&qp->rq);
2262 		wr = wr->next;
2263 	}
2264 	spin_unlock_irqrestore(&qp->q_lock, flags);
2265 	return status;
2266 }
2267 
2268 /* cqe for srq's rqe can potentially arrive out of order.
2269  * index gives the entry in the shadow table where to store
2270  * the wr_id. tag/index is returned in cqe to reference back
2271  * for a given rqe.
2272  */
2273 static int ocrdma_srq_get_idx(struct ocrdma_srq *srq)
2274 {
2275 	int row = 0;
2276 	int indx = 0;
2277 
2278 	for (row = 0; row < srq->bit_fields_len; row++) {
2279 		if (srq->idx_bit_fields[row]) {
2280 			indx = ffs(srq->idx_bit_fields[row]);
2281 			indx = (row * 32) + (indx - 1);
2282 			BUG_ON(indx >= srq->rq.max_cnt);
2283 			ocrdma_srq_toggle_bit(srq, indx);
2284 			break;
2285 		}
2286 	}
2287 
2288 	BUG_ON(row == srq->bit_fields_len);
2289 	return indx + 1; /* Use from index 1 */
2290 }
2291 
2292 static void ocrdma_ring_srq_db(struct ocrdma_srq *srq)
2293 {
2294 	u32 val = srq->rq.dbid | (1 << 16);
2295 
2296 	iowrite32(val, srq->db + OCRDMA_DB_GEN2_SRQ_OFFSET);
2297 }
2298 
2299 int ocrdma_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
2300 			 const struct ib_recv_wr **bad_wr)
2301 {
2302 	int status = 0;
2303 	unsigned long flags;
2304 	struct ocrdma_srq *srq;
2305 	struct ocrdma_hdr_wqe *rqe;
2306 	u16 tag;
2307 
2308 	srq = get_ocrdma_srq(ibsrq);
2309 
2310 	spin_lock_irqsave(&srq->q_lock, flags);
2311 	while (wr) {
2312 		if (ocrdma_hwq_free_cnt(&srq->rq) == 0 ||
2313 		    wr->num_sge > srq->rq.max_sges) {
2314 			status = -ENOMEM;
2315 			*bad_wr = wr;
2316 			break;
2317 		}
2318 		tag = ocrdma_srq_get_idx(srq);
2319 		rqe = ocrdma_hwq_head(&srq->rq);
2320 		ocrdma_build_rqe(rqe, wr, tag);
2321 
2322 		srq->rqe_wr_id_tbl[tag] = wr->wr_id;
2323 		/* make sure rqe is written before adapter can perform DMA */
2324 		wmb();
2325 		/* inform hw to start processing it */
2326 		ocrdma_ring_srq_db(srq);
2327 		/* update pointer, counter for next wr */
2328 		ocrdma_hwq_inc_head(&srq->rq);
2329 		wr = wr->next;
2330 	}
2331 	spin_unlock_irqrestore(&srq->q_lock, flags);
2332 	return status;
2333 }
2334 
2335 static enum ib_wc_status ocrdma_to_ibwc_err(u16 status)
2336 {
2337 	enum ib_wc_status ibwc_status;
2338 
2339 	switch (status) {
2340 	case OCRDMA_CQE_GENERAL_ERR:
2341 		ibwc_status = IB_WC_GENERAL_ERR;
2342 		break;
2343 	case OCRDMA_CQE_LOC_LEN_ERR:
2344 		ibwc_status = IB_WC_LOC_LEN_ERR;
2345 		break;
2346 	case OCRDMA_CQE_LOC_QP_OP_ERR:
2347 		ibwc_status = IB_WC_LOC_QP_OP_ERR;
2348 		break;
2349 	case OCRDMA_CQE_LOC_EEC_OP_ERR:
2350 		ibwc_status = IB_WC_LOC_EEC_OP_ERR;
2351 		break;
2352 	case OCRDMA_CQE_LOC_PROT_ERR:
2353 		ibwc_status = IB_WC_LOC_PROT_ERR;
2354 		break;
2355 	case OCRDMA_CQE_WR_FLUSH_ERR:
2356 		ibwc_status = IB_WC_WR_FLUSH_ERR;
2357 		break;
2358 	case OCRDMA_CQE_MW_BIND_ERR:
2359 		ibwc_status = IB_WC_MW_BIND_ERR;
2360 		break;
2361 	case OCRDMA_CQE_BAD_RESP_ERR:
2362 		ibwc_status = IB_WC_BAD_RESP_ERR;
2363 		break;
2364 	case OCRDMA_CQE_LOC_ACCESS_ERR:
2365 		ibwc_status = IB_WC_LOC_ACCESS_ERR;
2366 		break;
2367 	case OCRDMA_CQE_REM_INV_REQ_ERR:
2368 		ibwc_status = IB_WC_REM_INV_REQ_ERR;
2369 		break;
2370 	case OCRDMA_CQE_REM_ACCESS_ERR:
2371 		ibwc_status = IB_WC_REM_ACCESS_ERR;
2372 		break;
2373 	case OCRDMA_CQE_REM_OP_ERR:
2374 		ibwc_status = IB_WC_REM_OP_ERR;
2375 		break;
2376 	case OCRDMA_CQE_RETRY_EXC_ERR:
2377 		ibwc_status = IB_WC_RETRY_EXC_ERR;
2378 		break;
2379 	case OCRDMA_CQE_RNR_RETRY_EXC_ERR:
2380 		ibwc_status = IB_WC_RNR_RETRY_EXC_ERR;
2381 		break;
2382 	case OCRDMA_CQE_LOC_RDD_VIOL_ERR:
2383 		ibwc_status = IB_WC_LOC_RDD_VIOL_ERR;
2384 		break;
2385 	case OCRDMA_CQE_REM_INV_RD_REQ_ERR:
2386 		ibwc_status = IB_WC_REM_INV_RD_REQ_ERR;
2387 		break;
2388 	case OCRDMA_CQE_REM_ABORT_ERR:
2389 		ibwc_status = IB_WC_REM_ABORT_ERR;
2390 		break;
2391 	case OCRDMA_CQE_INV_EECN_ERR:
2392 		ibwc_status = IB_WC_INV_EECN_ERR;
2393 		break;
2394 	case OCRDMA_CQE_INV_EEC_STATE_ERR:
2395 		ibwc_status = IB_WC_INV_EEC_STATE_ERR;
2396 		break;
2397 	case OCRDMA_CQE_FATAL_ERR:
2398 		ibwc_status = IB_WC_FATAL_ERR;
2399 		break;
2400 	case OCRDMA_CQE_RESP_TIMEOUT_ERR:
2401 		ibwc_status = IB_WC_RESP_TIMEOUT_ERR;
2402 		break;
2403 	default:
2404 		ibwc_status = IB_WC_GENERAL_ERR;
2405 		break;
2406 	}
2407 	return ibwc_status;
2408 }
2409 
2410 static void ocrdma_update_wc(struct ocrdma_qp *qp, struct ib_wc *ibwc,
2411 		      u32 wqe_idx)
2412 {
2413 	struct ocrdma_hdr_wqe *hdr;
2414 	struct ocrdma_sge *rw;
2415 	int opcode;
2416 
2417 	hdr = ocrdma_hwq_head_from_idx(&qp->sq, wqe_idx);
2418 
2419 	ibwc->wr_id = qp->wqe_wr_id_tbl[wqe_idx].wrid;
2420 	/* Undo the hdr->cw swap */
2421 	opcode = le32_to_cpu(hdr->cw) & OCRDMA_WQE_OPCODE_MASK;
2422 	switch (opcode) {
2423 	case OCRDMA_WRITE:
2424 		ibwc->opcode = IB_WC_RDMA_WRITE;
2425 		break;
2426 	case OCRDMA_READ:
2427 		rw = (struct ocrdma_sge *)(hdr + 1);
2428 		ibwc->opcode = IB_WC_RDMA_READ;
2429 		ibwc->byte_len = rw->len;
2430 		break;
2431 	case OCRDMA_SEND:
2432 		ibwc->opcode = IB_WC_SEND;
2433 		break;
2434 	case OCRDMA_FR_MR:
2435 		ibwc->opcode = IB_WC_REG_MR;
2436 		break;
2437 	case OCRDMA_LKEY_INV:
2438 		ibwc->opcode = IB_WC_LOCAL_INV;
2439 		break;
2440 	default:
2441 		ibwc->status = IB_WC_GENERAL_ERR;
2442 		pr_err("%s() invalid opcode received = 0x%x\n",
2443 		       __func__, hdr->cw & OCRDMA_WQE_OPCODE_MASK);
2444 		break;
2445 	}
2446 }
2447 
2448 static void ocrdma_set_cqe_status_flushed(struct ocrdma_qp *qp,
2449 						struct ocrdma_cqe *cqe)
2450 {
2451 	if (is_cqe_for_sq(cqe)) {
2452 		cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2453 				cqe->flags_status_srcqpn) &
2454 					~OCRDMA_CQE_STATUS_MASK);
2455 		cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2456 				cqe->flags_status_srcqpn) |
2457 				(OCRDMA_CQE_WR_FLUSH_ERR <<
2458 					OCRDMA_CQE_STATUS_SHIFT));
2459 	} else {
2460 		if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2461 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2462 					cqe->flags_status_srcqpn) &
2463 						~OCRDMA_CQE_UD_STATUS_MASK);
2464 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2465 					cqe->flags_status_srcqpn) |
2466 					(OCRDMA_CQE_WR_FLUSH_ERR <<
2467 						OCRDMA_CQE_UD_STATUS_SHIFT));
2468 		} else {
2469 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2470 					cqe->flags_status_srcqpn) &
2471 						~OCRDMA_CQE_STATUS_MASK);
2472 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2473 					cqe->flags_status_srcqpn) |
2474 					(OCRDMA_CQE_WR_FLUSH_ERR <<
2475 						OCRDMA_CQE_STATUS_SHIFT));
2476 		}
2477 	}
2478 }
2479 
2480 static bool ocrdma_update_err_cqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2481 				  struct ocrdma_qp *qp, int status)
2482 {
2483 	bool expand = false;
2484 
2485 	ibwc->byte_len = 0;
2486 	ibwc->qp = &qp->ibqp;
2487 	ibwc->status = ocrdma_to_ibwc_err(status);
2488 
2489 	ocrdma_flush_qp(qp);
2490 	ocrdma_qp_state_change(qp, IB_QPS_ERR, NULL);
2491 
2492 	/* if wqe/rqe pending for which cqe needs to be returned,
2493 	 * trigger inflating it.
2494 	 */
2495 	if (!is_hw_rq_empty(qp) || !is_hw_sq_empty(qp)) {
2496 		expand = true;
2497 		ocrdma_set_cqe_status_flushed(qp, cqe);
2498 	}
2499 	return expand;
2500 }
2501 
2502 static int ocrdma_update_err_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2503 				  struct ocrdma_qp *qp, int status)
2504 {
2505 	ibwc->opcode = IB_WC_RECV;
2506 	ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2507 	ocrdma_hwq_inc_tail(&qp->rq);
2508 
2509 	return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2510 }
2511 
2512 static int ocrdma_update_err_scqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2513 				  struct ocrdma_qp *qp, int status)
2514 {
2515 	ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2516 	ocrdma_hwq_inc_tail(&qp->sq);
2517 
2518 	return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2519 }
2520 
2521 
2522 static bool ocrdma_poll_err_scqe(struct ocrdma_qp *qp,
2523 				 struct ocrdma_cqe *cqe, struct ib_wc *ibwc,
2524 				 bool *polled, bool *stop)
2525 {
2526 	bool expand;
2527 	struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2528 	int status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2529 		OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2530 	if (status < OCRDMA_MAX_CQE_ERR)
2531 		atomic_inc(&dev->cqe_err_stats[status]);
2532 
2533 	/* when hw sq is empty, but rq is not empty, so we continue
2534 	 * to keep the cqe in order to get the cq event again.
2535 	 */
2536 	if (is_hw_sq_empty(qp) && !is_hw_rq_empty(qp)) {
2537 		/* when cq for rq and sq is same, it is safe to return
2538 		 * flush cqe for RQEs.
2539 		 */
2540 		if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2541 			*polled = true;
2542 			status = OCRDMA_CQE_WR_FLUSH_ERR;
2543 			expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2544 		} else {
2545 			/* stop processing further cqe as this cqe is used for
2546 			 * triggering cq event on buddy cq of RQ.
2547 			 * When QP is destroyed, this cqe will be removed
2548 			 * from the cq's hardware q.
2549 			 */
2550 			*polled = false;
2551 			*stop = true;
2552 			expand = false;
2553 		}
2554 	} else if (is_hw_sq_empty(qp)) {
2555 		/* Do nothing */
2556 		expand = false;
2557 		*polled = false;
2558 		*stop = false;
2559 	} else {
2560 		*polled = true;
2561 		expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2562 	}
2563 	return expand;
2564 }
2565 
2566 static bool ocrdma_poll_success_scqe(struct ocrdma_qp *qp,
2567 				     struct ocrdma_cqe *cqe,
2568 				     struct ib_wc *ibwc, bool *polled)
2569 {
2570 	bool expand = false;
2571 	int tail = qp->sq.tail;
2572 	u32 wqe_idx;
2573 
2574 	if (!qp->wqe_wr_id_tbl[tail].signaled) {
2575 		*polled = false;    /* WC cannot be consumed yet */
2576 	} else {
2577 		ibwc->status = IB_WC_SUCCESS;
2578 		ibwc->wc_flags = 0;
2579 		ibwc->qp = &qp->ibqp;
2580 		ocrdma_update_wc(qp, ibwc, tail);
2581 		*polled = true;
2582 	}
2583 	wqe_idx = (le32_to_cpu(cqe->wq.wqeidx) &
2584 			OCRDMA_CQE_WQEIDX_MASK) & qp->sq.max_wqe_idx;
2585 	if (tail != wqe_idx)
2586 		expand = true; /* Coalesced CQE can't be consumed yet */
2587 
2588 	ocrdma_hwq_inc_tail(&qp->sq);
2589 	return expand;
2590 }
2591 
2592 static bool ocrdma_poll_scqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2593 			     struct ib_wc *ibwc, bool *polled, bool *stop)
2594 {
2595 	int status;
2596 	bool expand;
2597 
2598 	status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2599 		OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2600 
2601 	if (status == OCRDMA_CQE_SUCCESS)
2602 		expand = ocrdma_poll_success_scqe(qp, cqe, ibwc, polled);
2603 	else
2604 		expand = ocrdma_poll_err_scqe(qp, cqe, ibwc, polled, stop);
2605 	return expand;
2606 }
2607 
2608 static int ocrdma_update_ud_rcqe(struct ocrdma_dev *dev, struct ib_wc *ibwc,
2609 				 struct ocrdma_cqe *cqe)
2610 {
2611 	int status;
2612 	u16 hdr_type = 0;
2613 
2614 	status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2615 		OCRDMA_CQE_UD_STATUS_MASK) >> OCRDMA_CQE_UD_STATUS_SHIFT;
2616 	ibwc->src_qp = le32_to_cpu(cqe->flags_status_srcqpn) &
2617 						OCRDMA_CQE_SRCQP_MASK;
2618 	ibwc->pkey_index = 0;
2619 	ibwc->wc_flags = IB_WC_GRH;
2620 	ibwc->byte_len = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2621 			  OCRDMA_CQE_UD_XFER_LEN_SHIFT) &
2622 			  OCRDMA_CQE_UD_XFER_LEN_MASK;
2623 
2624 	if (ocrdma_is_udp_encap_supported(dev)) {
2625 		hdr_type = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2626 			    OCRDMA_CQE_UD_L3TYPE_SHIFT) &
2627 			    OCRDMA_CQE_UD_L3TYPE_MASK;
2628 		ibwc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
2629 		ibwc->network_hdr_type = hdr_type;
2630 	}
2631 
2632 	return status;
2633 }
2634 
2635 static void ocrdma_update_free_srq_cqe(struct ib_wc *ibwc,
2636 				       struct ocrdma_cqe *cqe,
2637 				       struct ocrdma_qp *qp)
2638 {
2639 	unsigned long flags;
2640 	struct ocrdma_srq *srq;
2641 	u32 wqe_idx;
2642 
2643 	srq = get_ocrdma_srq(qp->ibqp.srq);
2644 	wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
2645 		OCRDMA_CQE_BUFTAG_SHIFT) & srq->rq.max_wqe_idx;
2646 	BUG_ON(wqe_idx < 1);
2647 
2648 	ibwc->wr_id = srq->rqe_wr_id_tbl[wqe_idx];
2649 	spin_lock_irqsave(&srq->q_lock, flags);
2650 	ocrdma_srq_toggle_bit(srq, wqe_idx - 1);
2651 	spin_unlock_irqrestore(&srq->q_lock, flags);
2652 	ocrdma_hwq_inc_tail(&srq->rq);
2653 }
2654 
2655 static bool ocrdma_poll_err_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2656 				struct ib_wc *ibwc, bool *polled, bool *stop,
2657 				int status)
2658 {
2659 	bool expand;
2660 	struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2661 
2662 	if (status < OCRDMA_MAX_CQE_ERR)
2663 		atomic_inc(&dev->cqe_err_stats[status]);
2664 
2665 	/* when hw_rq is empty, but wq is not empty, so continue
2666 	 * to keep the cqe to get the cq event again.
2667 	 */
2668 	if (is_hw_rq_empty(qp) && !is_hw_sq_empty(qp)) {
2669 		if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2670 			*polled = true;
2671 			status = OCRDMA_CQE_WR_FLUSH_ERR;
2672 			expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2673 		} else {
2674 			*polled = false;
2675 			*stop = true;
2676 			expand = false;
2677 		}
2678 	} else if (is_hw_rq_empty(qp)) {
2679 		/* Do nothing */
2680 		expand = false;
2681 		*polled = false;
2682 		*stop = false;
2683 	} else {
2684 		*polled = true;
2685 		expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2686 	}
2687 	return expand;
2688 }
2689 
2690 static void ocrdma_poll_success_rcqe(struct ocrdma_qp *qp,
2691 				     struct ocrdma_cqe *cqe, struct ib_wc *ibwc)
2692 {
2693 	struct ocrdma_dev *dev;
2694 
2695 	dev = get_ocrdma_dev(qp->ibqp.device);
2696 	ibwc->opcode = IB_WC_RECV;
2697 	ibwc->qp = &qp->ibqp;
2698 	ibwc->status = IB_WC_SUCCESS;
2699 
2700 	if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI)
2701 		ocrdma_update_ud_rcqe(dev, ibwc, cqe);
2702 	else
2703 		ibwc->byte_len = le32_to_cpu(cqe->rq.rxlen);
2704 
2705 	if (is_cqe_imm(cqe)) {
2706 		ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2707 		ibwc->wc_flags |= IB_WC_WITH_IMM;
2708 	} else if (is_cqe_wr_imm(cqe)) {
2709 		ibwc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
2710 		ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2711 		ibwc->wc_flags |= IB_WC_WITH_IMM;
2712 	} else if (is_cqe_invalidated(cqe)) {
2713 		ibwc->ex.invalidate_rkey = le32_to_cpu(cqe->rq.lkey_immdt);
2714 		ibwc->wc_flags |= IB_WC_WITH_INVALIDATE;
2715 	}
2716 	if (qp->ibqp.srq) {
2717 		ocrdma_update_free_srq_cqe(ibwc, cqe, qp);
2718 	} else {
2719 		ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2720 		ocrdma_hwq_inc_tail(&qp->rq);
2721 	}
2722 }
2723 
2724 static bool ocrdma_poll_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2725 			     struct ib_wc *ibwc, bool *polled, bool *stop)
2726 {
2727 	int status;
2728 	bool expand = false;
2729 
2730 	ibwc->wc_flags = 0;
2731 	if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2732 		status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2733 					OCRDMA_CQE_UD_STATUS_MASK) >>
2734 					OCRDMA_CQE_UD_STATUS_SHIFT;
2735 	} else {
2736 		status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2737 			     OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2738 	}
2739 
2740 	if (status == OCRDMA_CQE_SUCCESS) {
2741 		*polled = true;
2742 		ocrdma_poll_success_rcqe(qp, cqe, ibwc);
2743 	} else {
2744 		expand = ocrdma_poll_err_rcqe(qp, cqe, ibwc, polled, stop,
2745 					      status);
2746 	}
2747 	return expand;
2748 }
2749 
2750 static void ocrdma_change_cq_phase(struct ocrdma_cq *cq, struct ocrdma_cqe *cqe,
2751 				   u16 cur_getp)
2752 {
2753 	if (cq->phase_change) {
2754 		if (cur_getp == 0)
2755 			cq->phase = (~cq->phase & OCRDMA_CQE_VALID);
2756 	} else {
2757 		/* clear valid bit */
2758 		cqe->flags_status_srcqpn = 0;
2759 	}
2760 }
2761 
2762 static int ocrdma_poll_hwcq(struct ocrdma_cq *cq, int num_entries,
2763 			    struct ib_wc *ibwc)
2764 {
2765 	u16 qpn = 0;
2766 	int i = 0;
2767 	bool expand = false;
2768 	int polled_hw_cqes = 0;
2769 	struct ocrdma_qp *qp = NULL;
2770 	struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
2771 	struct ocrdma_cqe *cqe;
2772 	u16 cur_getp; bool polled = false; bool stop = false;
2773 
2774 	cur_getp = cq->getp;
2775 	while (num_entries) {
2776 		cqe = cq->va + cur_getp;
2777 		/* check whether valid cqe or not */
2778 		if (!is_cqe_valid(cq, cqe))
2779 			break;
2780 		qpn = (le32_to_cpu(cqe->cmn.qpn) & OCRDMA_CQE_QPN_MASK);
2781 		/* ignore discarded cqe */
2782 		if (qpn == 0)
2783 			goto skip_cqe;
2784 		qp = dev->qp_tbl[qpn];
2785 		BUG_ON(qp == NULL);
2786 
2787 		if (is_cqe_for_sq(cqe)) {
2788 			expand = ocrdma_poll_scqe(qp, cqe, ibwc, &polled,
2789 						  &stop);
2790 		} else {
2791 			expand = ocrdma_poll_rcqe(qp, cqe, ibwc, &polled,
2792 						  &stop);
2793 		}
2794 		if (expand)
2795 			goto expand_cqe;
2796 		if (stop)
2797 			goto stop_cqe;
2798 		/* clear qpn to avoid duplicate processing by discard_cqe() */
2799 		cqe->cmn.qpn = 0;
2800 skip_cqe:
2801 		polled_hw_cqes += 1;
2802 		cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
2803 		ocrdma_change_cq_phase(cq, cqe, cur_getp);
2804 expand_cqe:
2805 		if (polled) {
2806 			num_entries -= 1;
2807 			i += 1;
2808 			ibwc = ibwc + 1;
2809 			polled = false;
2810 		}
2811 	}
2812 stop_cqe:
2813 	cq->getp = cur_getp;
2814 
2815 	if (polled_hw_cqes)
2816 		ocrdma_ring_cq_db(dev, cq->id, false, false, polled_hw_cqes);
2817 
2818 	return i;
2819 }
2820 
2821 /* insert error cqe if the QP's SQ or RQ's CQ matches the CQ under poll. */
2822 static int ocrdma_add_err_cqe(struct ocrdma_cq *cq, int num_entries,
2823 			      struct ocrdma_qp *qp, struct ib_wc *ibwc)
2824 {
2825 	int err_cqes = 0;
2826 
2827 	while (num_entries) {
2828 		if (is_hw_sq_empty(qp) && is_hw_rq_empty(qp))
2829 			break;
2830 		if (!is_hw_sq_empty(qp) && qp->sq_cq == cq) {
2831 			ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2832 			ocrdma_hwq_inc_tail(&qp->sq);
2833 		} else if (!is_hw_rq_empty(qp) && qp->rq_cq == cq) {
2834 			ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2835 			ocrdma_hwq_inc_tail(&qp->rq);
2836 		} else {
2837 			return err_cqes;
2838 		}
2839 		ibwc->byte_len = 0;
2840 		ibwc->status = IB_WC_WR_FLUSH_ERR;
2841 		ibwc = ibwc + 1;
2842 		err_cqes += 1;
2843 		num_entries -= 1;
2844 	}
2845 	return err_cqes;
2846 }
2847 
2848 int ocrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
2849 {
2850 	int cqes_to_poll = num_entries;
2851 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2852 	struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2853 	int num_os_cqe = 0, err_cqes = 0;
2854 	struct ocrdma_qp *qp;
2855 	unsigned long flags;
2856 
2857 	/* poll cqes from adapter CQ */
2858 	spin_lock_irqsave(&cq->cq_lock, flags);
2859 	num_os_cqe = ocrdma_poll_hwcq(cq, cqes_to_poll, wc);
2860 	spin_unlock_irqrestore(&cq->cq_lock, flags);
2861 	cqes_to_poll -= num_os_cqe;
2862 
2863 	if (cqes_to_poll) {
2864 		wc = wc + num_os_cqe;
2865 		/* adapter returns single error cqe when qp moves to
2866 		 * error state. So insert error cqes with wc_status as
2867 		 * FLUSHED for pending WQEs and RQEs of QP's SQ and RQ
2868 		 * respectively which uses this CQ.
2869 		 */
2870 		spin_lock_irqsave(&dev->flush_q_lock, flags);
2871 		list_for_each_entry(qp, &cq->sq_head, sq_entry) {
2872 			if (cqes_to_poll == 0)
2873 				break;
2874 			err_cqes = ocrdma_add_err_cqe(cq, cqes_to_poll, qp, wc);
2875 			cqes_to_poll -= err_cqes;
2876 			num_os_cqe += err_cqes;
2877 			wc = wc + err_cqes;
2878 		}
2879 		spin_unlock_irqrestore(&dev->flush_q_lock, flags);
2880 	}
2881 	return num_os_cqe;
2882 }
2883 
2884 int ocrdma_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags cq_flags)
2885 {
2886 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2887 	struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2888 	u16 cq_id;
2889 	unsigned long flags;
2890 	bool arm_needed = false, sol_needed = false;
2891 
2892 	cq_id = cq->id;
2893 
2894 	spin_lock_irqsave(&cq->cq_lock, flags);
2895 	if (cq_flags & IB_CQ_NEXT_COMP || cq_flags & IB_CQ_SOLICITED)
2896 		arm_needed = true;
2897 	if (cq_flags & IB_CQ_SOLICITED)
2898 		sol_needed = true;
2899 
2900 	ocrdma_ring_cq_db(dev, cq_id, arm_needed, sol_needed, 0);
2901 	spin_unlock_irqrestore(&cq->cq_lock, flags);
2902 
2903 	return 0;
2904 }
2905 
2906 struct ib_mr *ocrdma_alloc_mr(struct ib_pd *ibpd, enum ib_mr_type mr_type,
2907 			      u32 max_num_sg)
2908 {
2909 	int status;
2910 	struct ocrdma_mr *mr;
2911 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
2912 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
2913 
2914 	if (mr_type != IB_MR_TYPE_MEM_REG)
2915 		return ERR_PTR(-EINVAL);
2916 
2917 	if (max_num_sg > dev->attr.max_pages_per_frmr)
2918 		return ERR_PTR(-EINVAL);
2919 
2920 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
2921 	if (!mr)
2922 		return ERR_PTR(-ENOMEM);
2923 
2924 	mr->pages = kcalloc(max_num_sg, sizeof(u64), GFP_KERNEL);
2925 	if (!mr->pages) {
2926 		status = -ENOMEM;
2927 		goto pl_err;
2928 	}
2929 
2930 	status = ocrdma_get_pbl_info(dev, mr, max_num_sg);
2931 	if (status)
2932 		goto pbl_err;
2933 	mr->hwmr.fr_mr = 1;
2934 	mr->hwmr.remote_rd = 0;
2935 	mr->hwmr.remote_wr = 0;
2936 	mr->hwmr.local_rd = 0;
2937 	mr->hwmr.local_wr = 0;
2938 	mr->hwmr.mw_bind = 0;
2939 	status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
2940 	if (status)
2941 		goto pbl_err;
2942 	status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, 0);
2943 	if (status)
2944 		goto mbx_err;
2945 	mr->ibmr.rkey = mr->hwmr.lkey;
2946 	mr->ibmr.lkey = mr->hwmr.lkey;
2947 	dev->stag_arr[(mr->hwmr.lkey >> 8) & (OCRDMA_MAX_STAG - 1)] =
2948 		(unsigned long) mr;
2949 	return &mr->ibmr;
2950 mbx_err:
2951 	ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
2952 pbl_err:
2953 	kfree(mr->pages);
2954 pl_err:
2955 	kfree(mr);
2956 	return ERR_PTR(-ENOMEM);
2957 }
2958 
2959 static int ocrdma_set_page(struct ib_mr *ibmr, u64 addr)
2960 {
2961 	struct ocrdma_mr *mr = get_ocrdma_mr(ibmr);
2962 
2963 	if (unlikely(mr->npages == mr->hwmr.num_pbes))
2964 		return -ENOMEM;
2965 
2966 	mr->pages[mr->npages++] = addr;
2967 
2968 	return 0;
2969 }
2970 
2971 int ocrdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
2972 		     unsigned int *sg_offset)
2973 {
2974 	struct ocrdma_mr *mr = get_ocrdma_mr(ibmr);
2975 
2976 	mr->npages = 0;
2977 
2978 	return ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, ocrdma_set_page);
2979 }
2980