xref: /linux/drivers/infiniband/hw/ocrdma/ocrdma_verbs.c (revision 55d0969c451159cff86949b38c39171cab962069)
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 <net/addrconf.h>
45 #include <rdma/ib_verbs.h>
46 #include <rdma/ib_user_verbs.h>
47 #include <rdma/iw_cm.h>
48 #include <rdma/ib_umem.h>
49 #include <rdma/ib_addr.h>
50 #include <rdma/ib_cache.h>
51 #include <rdma/uverbs_ioctl.h>
52 
53 #include "ocrdma.h"
54 #include "ocrdma_hw.h"
55 #include "ocrdma_verbs.h"
56 #include <rdma/ocrdma-abi.h>
57 
58 int ocrdma_query_pkey(struct ib_device *ibdev, u32 port, u16 index, u16 *pkey)
59 {
60 	if (index > 0)
61 		return -EINVAL;
62 
63 	*pkey = 0xffff;
64 	return 0;
65 }
66 
67 int ocrdma_query_device(struct ib_device *ibdev, struct ib_device_attr *attr,
68 			struct ib_udata *uhw)
69 {
70 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
71 
72 	if (uhw->inlen || uhw->outlen)
73 		return -EINVAL;
74 
75 	memset(attr, 0, sizeof *attr);
76 	memcpy(&attr->fw_ver, &dev->attr.fw_ver[0],
77 	       min(sizeof(dev->attr.fw_ver), sizeof(attr->fw_ver)));
78 	addrconf_addr_eui48((u8 *)&attr->sys_image_guid,
79 			    dev->nic_info.mac_addr);
80 	attr->max_mr_size = dev->attr.max_mr_size;
81 	attr->page_size_cap = 0xffff000;
82 	attr->vendor_id = dev->nic_info.pdev->vendor;
83 	attr->vendor_part_id = dev->nic_info.pdev->device;
84 	attr->hw_ver = dev->asic_id;
85 	attr->max_qp = dev->attr.max_qp;
86 	attr->max_ah = OCRDMA_MAX_AH;
87 	attr->max_qp_wr = dev->attr.max_wqe;
88 
89 	attr->device_cap_flags = IB_DEVICE_CURR_QP_STATE_MOD |
90 					IB_DEVICE_RC_RNR_NAK_GEN |
91 					IB_DEVICE_SHUTDOWN_PORT |
92 					IB_DEVICE_SYS_IMAGE_GUID |
93 					IB_DEVICE_MEM_MGT_EXTENSIONS;
94 	attr->kernel_cap_flags = IBK_LOCAL_DMA_LKEY;
95 	attr->max_send_sge = dev->attr.max_send_sge;
96 	attr->max_recv_sge = dev->attr.max_recv_sge;
97 	attr->max_sge_rd = dev->attr.max_rdma_sge;
98 	attr->max_cq = dev->attr.max_cq;
99 	attr->max_cqe = dev->attr.max_cqe;
100 	attr->max_mr = dev->attr.max_mr;
101 	attr->max_mw = dev->attr.max_mw;
102 	attr->max_pd = dev->attr.max_pd;
103 	attr->atomic_cap = 0;
104 	attr->max_qp_rd_atom =
105 	    min(dev->attr.max_ord_per_qp, dev->attr.max_ird_per_qp);
106 	attr->max_qp_init_rd_atom = dev->attr.max_ord_per_qp;
107 	attr->max_srq = dev->attr.max_srq;
108 	attr->max_srq_sge = dev->attr.max_srq_sge;
109 	attr->max_srq_wr = dev->attr.max_rqe;
110 	attr->local_ca_ack_delay = dev->attr.local_ca_ack_delay;
111 	attr->max_fast_reg_page_list_len = dev->attr.max_pages_per_frmr;
112 	attr->max_pkeys = 1;
113 	return 0;
114 }
115 
116 static inline void get_link_speed_and_width(struct ocrdma_dev *dev,
117 					    u16 *ib_speed, u8 *ib_width)
118 {
119 	int status;
120 	u8 speed;
121 
122 	status = ocrdma_mbx_get_link_speed(dev, &speed, NULL);
123 	if (status)
124 		speed = OCRDMA_PHYS_LINK_SPEED_ZERO;
125 
126 	switch (speed) {
127 	case OCRDMA_PHYS_LINK_SPEED_1GBPS:
128 		*ib_speed = IB_SPEED_SDR;
129 		*ib_width = IB_WIDTH_1X;
130 		break;
131 
132 	case OCRDMA_PHYS_LINK_SPEED_10GBPS:
133 		*ib_speed = IB_SPEED_QDR;
134 		*ib_width = IB_WIDTH_1X;
135 		break;
136 
137 	case OCRDMA_PHYS_LINK_SPEED_20GBPS:
138 		*ib_speed = IB_SPEED_DDR;
139 		*ib_width = IB_WIDTH_4X;
140 		break;
141 
142 	case OCRDMA_PHYS_LINK_SPEED_40GBPS:
143 		*ib_speed = IB_SPEED_QDR;
144 		*ib_width = IB_WIDTH_4X;
145 		break;
146 
147 	default:
148 		/* Unsupported */
149 		*ib_speed = IB_SPEED_SDR;
150 		*ib_width = IB_WIDTH_1X;
151 	}
152 }
153 
154 int ocrdma_query_port(struct ib_device *ibdev,
155 		      u32 port, struct ib_port_attr *props)
156 {
157 	enum ib_port_state port_state;
158 	struct ocrdma_dev *dev;
159 	struct net_device *netdev;
160 
161 	/* props being zeroed by the caller, avoid zeroing it here */
162 	dev = get_ocrdma_dev(ibdev);
163 	netdev = dev->nic_info.netdev;
164 	if (netif_running(netdev) && netif_oper_up(netdev)) {
165 		port_state = IB_PORT_ACTIVE;
166 		props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
167 	} else {
168 		port_state = IB_PORT_DOWN;
169 		props->phys_state = IB_PORT_PHYS_STATE_DISABLED;
170 	}
171 	props->max_mtu = IB_MTU_4096;
172 	props->active_mtu = iboe_get_mtu(netdev->mtu);
173 	props->lid = 0;
174 	props->lmc = 0;
175 	props->sm_lid = 0;
176 	props->sm_sl = 0;
177 	props->state = port_state;
178 	props->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_REINIT_SUP |
179 				IB_PORT_DEVICE_MGMT_SUP |
180 				IB_PORT_VENDOR_CLASS_SUP;
181 	props->ip_gids = true;
182 	props->gid_tbl_len = OCRDMA_MAX_SGID;
183 	props->pkey_tbl_len = 1;
184 	props->bad_pkey_cntr = 0;
185 	props->qkey_viol_cntr = 0;
186 	get_link_speed_and_width(dev, &props->active_speed,
187 				 &props->active_width);
188 	props->max_msg_sz = 0x80000000;
189 	props->max_vl_num = 4;
190 	return 0;
191 }
192 
193 static int ocrdma_add_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
194 			   unsigned long len)
195 {
196 	struct ocrdma_mm *mm;
197 
198 	mm = kzalloc(sizeof(*mm), GFP_KERNEL);
199 	if (mm == NULL)
200 		return -ENOMEM;
201 	mm->key.phy_addr = phy_addr;
202 	mm->key.len = len;
203 	INIT_LIST_HEAD(&mm->entry);
204 
205 	mutex_lock(&uctx->mm_list_lock);
206 	list_add_tail(&mm->entry, &uctx->mm_head);
207 	mutex_unlock(&uctx->mm_list_lock);
208 	return 0;
209 }
210 
211 static void ocrdma_del_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
212 			    unsigned long len)
213 {
214 	struct ocrdma_mm *mm, *tmp;
215 
216 	mutex_lock(&uctx->mm_list_lock);
217 	list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
218 		if (len != mm->key.len && phy_addr != mm->key.phy_addr)
219 			continue;
220 
221 		list_del(&mm->entry);
222 		kfree(mm);
223 		break;
224 	}
225 	mutex_unlock(&uctx->mm_list_lock);
226 }
227 
228 static bool ocrdma_search_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
229 			      unsigned long len)
230 {
231 	bool found = false;
232 	struct ocrdma_mm *mm;
233 
234 	mutex_lock(&uctx->mm_list_lock);
235 	list_for_each_entry(mm, &uctx->mm_head, entry) {
236 		if (len != mm->key.len && phy_addr != mm->key.phy_addr)
237 			continue;
238 
239 		found = true;
240 		break;
241 	}
242 	mutex_unlock(&uctx->mm_list_lock);
243 	return found;
244 }
245 
246 
247 static u16 _ocrdma_pd_mgr_get_bitmap(struct ocrdma_dev *dev, bool dpp_pool)
248 {
249 	u16 pd_bitmap_idx = 0;
250 	unsigned long *pd_bitmap;
251 
252 	if (dpp_pool) {
253 		pd_bitmap = dev->pd_mgr->pd_dpp_bitmap;
254 		pd_bitmap_idx = find_first_zero_bit(pd_bitmap,
255 						    dev->pd_mgr->max_dpp_pd);
256 		__set_bit(pd_bitmap_idx, pd_bitmap);
257 		dev->pd_mgr->pd_dpp_count++;
258 		if (dev->pd_mgr->pd_dpp_count > dev->pd_mgr->pd_dpp_thrsh)
259 			dev->pd_mgr->pd_dpp_thrsh = dev->pd_mgr->pd_dpp_count;
260 	} else {
261 		pd_bitmap = dev->pd_mgr->pd_norm_bitmap;
262 		pd_bitmap_idx = find_first_zero_bit(pd_bitmap,
263 						    dev->pd_mgr->max_normal_pd);
264 		__set_bit(pd_bitmap_idx, pd_bitmap);
265 		dev->pd_mgr->pd_norm_count++;
266 		if (dev->pd_mgr->pd_norm_count > dev->pd_mgr->pd_norm_thrsh)
267 			dev->pd_mgr->pd_norm_thrsh = dev->pd_mgr->pd_norm_count;
268 	}
269 	return pd_bitmap_idx;
270 }
271 
272 static int _ocrdma_pd_mgr_put_bitmap(struct ocrdma_dev *dev, u16 pd_id,
273 					bool dpp_pool)
274 {
275 	u16 pd_count;
276 	u16 pd_bit_index;
277 
278 	pd_count = dpp_pool ? dev->pd_mgr->pd_dpp_count :
279 			      dev->pd_mgr->pd_norm_count;
280 	if (pd_count == 0)
281 		return -EINVAL;
282 
283 	if (dpp_pool) {
284 		pd_bit_index = pd_id - dev->pd_mgr->pd_dpp_start;
285 		if (pd_bit_index >= dev->pd_mgr->max_dpp_pd) {
286 			return -EINVAL;
287 		} else {
288 			__clear_bit(pd_bit_index, dev->pd_mgr->pd_dpp_bitmap);
289 			dev->pd_mgr->pd_dpp_count--;
290 		}
291 	} else {
292 		pd_bit_index = pd_id - dev->pd_mgr->pd_norm_start;
293 		if (pd_bit_index >= dev->pd_mgr->max_normal_pd) {
294 			return -EINVAL;
295 		} else {
296 			__clear_bit(pd_bit_index, dev->pd_mgr->pd_norm_bitmap);
297 			dev->pd_mgr->pd_norm_count--;
298 		}
299 	}
300 
301 	return 0;
302 }
303 
304 static int ocrdma_put_pd_num(struct ocrdma_dev *dev, u16 pd_id,
305 				   bool dpp_pool)
306 {
307 	int status;
308 
309 	mutex_lock(&dev->dev_lock);
310 	status = _ocrdma_pd_mgr_put_bitmap(dev, pd_id, dpp_pool);
311 	mutex_unlock(&dev->dev_lock);
312 	return status;
313 }
314 
315 static int ocrdma_get_pd_num(struct ocrdma_dev *dev, struct ocrdma_pd *pd)
316 {
317 	u16 pd_idx = 0;
318 	int status = 0;
319 
320 	mutex_lock(&dev->dev_lock);
321 	if (pd->dpp_enabled) {
322 		/* try allocating DPP PD, if not available then normal PD */
323 		if (dev->pd_mgr->pd_dpp_count < dev->pd_mgr->max_dpp_pd) {
324 			pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, true);
325 			pd->id = dev->pd_mgr->pd_dpp_start + pd_idx;
326 			pd->dpp_page = dev->pd_mgr->dpp_page_index + pd_idx;
327 		} else if (dev->pd_mgr->pd_norm_count <
328 			   dev->pd_mgr->max_normal_pd) {
329 			pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, false);
330 			pd->id = dev->pd_mgr->pd_norm_start + pd_idx;
331 			pd->dpp_enabled = false;
332 		} else {
333 			status = -EINVAL;
334 		}
335 	} else {
336 		if (dev->pd_mgr->pd_norm_count < dev->pd_mgr->max_normal_pd) {
337 			pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, false);
338 			pd->id = dev->pd_mgr->pd_norm_start + pd_idx;
339 		} else {
340 			status = -EINVAL;
341 		}
342 	}
343 	mutex_unlock(&dev->dev_lock);
344 	return status;
345 }
346 
347 /*
348  * NOTE:
349  *
350  * ocrdma_ucontext must be used here because this function is also
351  * called from ocrdma_alloc_ucontext where ib_udata does not have
352  * valid ib_ucontext pointer. ib_uverbs_get_context does not call
353  * uobj_{alloc|get_xxx} helpers which are used to store the
354  * ib_ucontext in uverbs_attr_bundle wrapping the ib_udata. so
355  * ib_udata does NOT imply valid ib_ucontext here!
356  */
357 static int _ocrdma_alloc_pd(struct ocrdma_dev *dev, struct ocrdma_pd *pd,
358 			    struct ocrdma_ucontext *uctx,
359 			    struct ib_udata *udata)
360 {
361 	int status;
362 
363 	if (udata && uctx && dev->attr.max_dpp_pds) {
364 		pd->dpp_enabled =
365 			ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R;
366 		pd->num_dpp_qp =
367 			pd->dpp_enabled ? (dev->nic_info.db_page_size /
368 					   dev->attr.wqe_size) : 0;
369 	}
370 
371 	if (dev->pd_mgr->pd_prealloc_valid)
372 		return ocrdma_get_pd_num(dev, pd);
373 
374 retry:
375 	status = ocrdma_mbx_alloc_pd(dev, pd);
376 	if (status) {
377 		if (pd->dpp_enabled) {
378 			pd->dpp_enabled = false;
379 			pd->num_dpp_qp = 0;
380 			goto retry;
381 		}
382 		return status;
383 	}
384 
385 	return 0;
386 }
387 
388 static inline int is_ucontext_pd(struct ocrdma_ucontext *uctx,
389 				 struct ocrdma_pd *pd)
390 {
391 	return (uctx->cntxt_pd == pd);
392 }
393 
394 static void _ocrdma_dealloc_pd(struct ocrdma_dev *dev,
395 			      struct ocrdma_pd *pd)
396 {
397 	if (dev->pd_mgr->pd_prealloc_valid)
398 		ocrdma_put_pd_num(dev, pd->id, pd->dpp_enabled);
399 	else
400 		ocrdma_mbx_dealloc_pd(dev, pd);
401 }
402 
403 static int ocrdma_alloc_ucontext_pd(struct ocrdma_dev *dev,
404 				    struct ocrdma_ucontext *uctx,
405 				    struct ib_udata *udata)
406 {
407 	struct ib_device *ibdev = &dev->ibdev;
408 	struct ib_pd *pd;
409 	int status;
410 
411 	pd = rdma_zalloc_drv_obj(ibdev, ib_pd);
412 	if (!pd)
413 		return -ENOMEM;
414 
415 	pd->device  = ibdev;
416 	uctx->cntxt_pd = get_ocrdma_pd(pd);
417 
418 	status = _ocrdma_alloc_pd(dev, uctx->cntxt_pd, uctx, udata);
419 	if (status) {
420 		kfree(uctx->cntxt_pd);
421 		goto err;
422 	}
423 
424 	uctx->cntxt_pd->uctx = uctx;
425 	uctx->cntxt_pd->ibpd.device = &dev->ibdev;
426 err:
427 	return status;
428 }
429 
430 static void ocrdma_dealloc_ucontext_pd(struct ocrdma_ucontext *uctx)
431 {
432 	struct ocrdma_pd *pd = uctx->cntxt_pd;
433 	struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
434 
435 	if (uctx->pd_in_use) {
436 		pr_err("%s(%d) Freeing in use pdid=0x%x.\n",
437 		       __func__, dev->id, pd->id);
438 	}
439 	uctx->cntxt_pd = NULL;
440 	_ocrdma_dealloc_pd(dev, pd);
441 	kfree(pd);
442 }
443 
444 static struct ocrdma_pd *ocrdma_get_ucontext_pd(struct ocrdma_ucontext *uctx)
445 {
446 	struct ocrdma_pd *pd = NULL;
447 
448 	mutex_lock(&uctx->mm_list_lock);
449 	if (!uctx->pd_in_use) {
450 		uctx->pd_in_use = true;
451 		pd = uctx->cntxt_pd;
452 	}
453 	mutex_unlock(&uctx->mm_list_lock);
454 
455 	return pd;
456 }
457 
458 static void ocrdma_release_ucontext_pd(struct ocrdma_ucontext *uctx)
459 {
460 	mutex_lock(&uctx->mm_list_lock);
461 	uctx->pd_in_use = false;
462 	mutex_unlock(&uctx->mm_list_lock);
463 }
464 
465 int ocrdma_alloc_ucontext(struct ib_ucontext *uctx, struct ib_udata *udata)
466 {
467 	struct ib_device *ibdev = uctx->device;
468 	int status;
469 	struct ocrdma_ucontext *ctx = get_ocrdma_ucontext(uctx);
470 	struct ocrdma_alloc_ucontext_resp resp = {};
471 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
472 	struct pci_dev *pdev = dev->nic_info.pdev;
473 	u32 map_len = roundup(sizeof(u32) * 2048, PAGE_SIZE);
474 
475 	if (!udata)
476 		return -EFAULT;
477 	INIT_LIST_HEAD(&ctx->mm_head);
478 	mutex_init(&ctx->mm_list_lock);
479 
480 	ctx->ah_tbl.va = dma_alloc_coherent(&pdev->dev, map_len,
481 					    &ctx->ah_tbl.pa, GFP_KERNEL);
482 	if (!ctx->ah_tbl.va)
483 		return -ENOMEM;
484 
485 	ctx->ah_tbl.len = map_len;
486 
487 	resp.ah_tbl_len = ctx->ah_tbl.len;
488 	resp.ah_tbl_page = virt_to_phys(ctx->ah_tbl.va);
489 
490 	status = ocrdma_add_mmap(ctx, resp.ah_tbl_page, resp.ah_tbl_len);
491 	if (status)
492 		goto map_err;
493 
494 	status = ocrdma_alloc_ucontext_pd(dev, ctx, udata);
495 	if (status)
496 		goto pd_err;
497 
498 	resp.dev_id = dev->id;
499 	resp.max_inline_data = dev->attr.max_inline_data;
500 	resp.wqe_size = dev->attr.wqe_size;
501 	resp.rqe_size = dev->attr.rqe_size;
502 	resp.dpp_wqe_size = dev->attr.wqe_size;
503 
504 	memcpy(resp.fw_ver, dev->attr.fw_ver, sizeof(resp.fw_ver));
505 	status = ib_copy_to_udata(udata, &resp, sizeof(resp));
506 	if (status)
507 		goto cpy_err;
508 	return 0;
509 
510 cpy_err:
511 	ocrdma_dealloc_ucontext_pd(ctx);
512 pd_err:
513 	ocrdma_del_mmap(ctx, ctx->ah_tbl.pa, ctx->ah_tbl.len);
514 map_err:
515 	dma_free_coherent(&pdev->dev, ctx->ah_tbl.len, ctx->ah_tbl.va,
516 			  ctx->ah_tbl.pa);
517 	return status;
518 }
519 
520 void ocrdma_dealloc_ucontext(struct ib_ucontext *ibctx)
521 {
522 	struct ocrdma_mm *mm, *tmp;
523 	struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ibctx);
524 	struct ocrdma_dev *dev = get_ocrdma_dev(ibctx->device);
525 	struct pci_dev *pdev = dev->nic_info.pdev;
526 
527 	ocrdma_dealloc_ucontext_pd(uctx);
528 
529 	ocrdma_del_mmap(uctx, uctx->ah_tbl.pa, uctx->ah_tbl.len);
530 	dma_free_coherent(&pdev->dev, uctx->ah_tbl.len, uctx->ah_tbl.va,
531 			  uctx->ah_tbl.pa);
532 
533 	list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
534 		list_del(&mm->entry);
535 		kfree(mm);
536 	}
537 }
538 
539 int ocrdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
540 {
541 	struct ocrdma_ucontext *ucontext = get_ocrdma_ucontext(context);
542 	struct ocrdma_dev *dev = get_ocrdma_dev(context->device);
543 	unsigned long vm_page = vma->vm_pgoff << PAGE_SHIFT;
544 	u64 unmapped_db = (u64) dev->nic_info.unmapped_db;
545 	unsigned long len = (vma->vm_end - vma->vm_start);
546 	int status;
547 	bool found;
548 
549 	if (vma->vm_start & (PAGE_SIZE - 1))
550 		return -EINVAL;
551 	found = ocrdma_search_mmap(ucontext, vma->vm_pgoff << PAGE_SHIFT, len);
552 	if (!found)
553 		return -EINVAL;
554 
555 	if ((vm_page >= unmapped_db) && (vm_page <= (unmapped_db +
556 		dev->nic_info.db_total_size)) &&
557 		(len <=	dev->nic_info.db_page_size)) {
558 		if (vma->vm_flags & VM_READ)
559 			return -EPERM;
560 
561 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
562 		status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
563 					    len, vma->vm_page_prot);
564 	} else if (dev->nic_info.dpp_unmapped_len &&
565 		(vm_page >= (u64) dev->nic_info.dpp_unmapped_addr) &&
566 		(vm_page <= (u64) (dev->nic_info.dpp_unmapped_addr +
567 			dev->nic_info.dpp_unmapped_len)) &&
568 		(len <= dev->nic_info.dpp_unmapped_len)) {
569 		if (vma->vm_flags & VM_READ)
570 			return -EPERM;
571 
572 		vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
573 		status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
574 					    len, vma->vm_page_prot);
575 	} else {
576 		status = remap_pfn_range(vma, vma->vm_start,
577 					 vma->vm_pgoff, len, vma->vm_page_prot);
578 	}
579 	return status;
580 }
581 
582 static int ocrdma_copy_pd_uresp(struct ocrdma_dev *dev, struct ocrdma_pd *pd,
583 				struct ib_udata *udata)
584 {
585 	int status;
586 	u64 db_page_addr;
587 	u64 dpp_page_addr = 0;
588 	u32 db_page_size;
589 	struct ocrdma_alloc_pd_uresp rsp;
590 	struct ocrdma_ucontext *uctx = rdma_udata_to_drv_context(
591 		udata, struct ocrdma_ucontext, ibucontext);
592 
593 	memset(&rsp, 0, sizeof(rsp));
594 	rsp.id = pd->id;
595 	rsp.dpp_enabled = pd->dpp_enabled;
596 	db_page_addr = ocrdma_get_db_addr(dev, pd->id);
597 	db_page_size = dev->nic_info.db_page_size;
598 
599 	status = ocrdma_add_mmap(uctx, db_page_addr, db_page_size);
600 	if (status)
601 		return status;
602 
603 	if (pd->dpp_enabled) {
604 		dpp_page_addr = dev->nic_info.dpp_unmapped_addr +
605 				(pd->id * PAGE_SIZE);
606 		status = ocrdma_add_mmap(uctx, dpp_page_addr,
607 				 PAGE_SIZE);
608 		if (status)
609 			goto dpp_map_err;
610 		rsp.dpp_page_addr_hi = upper_32_bits(dpp_page_addr);
611 		rsp.dpp_page_addr_lo = dpp_page_addr;
612 	}
613 
614 	status = ib_copy_to_udata(udata, &rsp, sizeof(rsp));
615 	if (status)
616 		goto ucopy_err;
617 
618 	pd->uctx = uctx;
619 	return 0;
620 
621 ucopy_err:
622 	if (pd->dpp_enabled)
623 		ocrdma_del_mmap(pd->uctx, dpp_page_addr, PAGE_SIZE);
624 dpp_map_err:
625 	ocrdma_del_mmap(pd->uctx, db_page_addr, db_page_size);
626 	return status;
627 }
628 
629 int ocrdma_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
630 {
631 	struct ib_device *ibdev = ibpd->device;
632 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
633 	struct ocrdma_pd *pd;
634 	int status;
635 	u8 is_uctx_pd = false;
636 	struct ocrdma_ucontext *uctx = rdma_udata_to_drv_context(
637 		udata, struct ocrdma_ucontext, ibucontext);
638 
639 	if (udata) {
640 		pd = ocrdma_get_ucontext_pd(uctx);
641 		if (pd) {
642 			is_uctx_pd = true;
643 			goto pd_mapping;
644 		}
645 	}
646 
647 	pd = get_ocrdma_pd(ibpd);
648 	status = _ocrdma_alloc_pd(dev, pd, uctx, udata);
649 	if (status)
650 		goto exit;
651 
652 pd_mapping:
653 	if (udata) {
654 		status = ocrdma_copy_pd_uresp(dev, pd, udata);
655 		if (status)
656 			goto err;
657 	}
658 	return 0;
659 
660 err:
661 	if (is_uctx_pd)
662 		ocrdma_release_ucontext_pd(uctx);
663 	else
664 		_ocrdma_dealloc_pd(dev, pd);
665 exit:
666 	return status;
667 }
668 
669 int ocrdma_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
670 {
671 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
672 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
673 	struct ocrdma_ucontext *uctx = NULL;
674 	u64 usr_db;
675 
676 	uctx = pd->uctx;
677 	if (uctx) {
678 		u64 dpp_db = dev->nic_info.dpp_unmapped_addr +
679 			(pd->id * PAGE_SIZE);
680 		if (pd->dpp_enabled)
681 			ocrdma_del_mmap(pd->uctx, dpp_db, PAGE_SIZE);
682 		usr_db = ocrdma_get_db_addr(dev, pd->id);
683 		ocrdma_del_mmap(pd->uctx, usr_db, dev->nic_info.db_page_size);
684 
685 		if (is_ucontext_pd(uctx, pd)) {
686 			ocrdma_release_ucontext_pd(uctx);
687 			return 0;
688 		}
689 	}
690 	_ocrdma_dealloc_pd(dev, pd);
691 	return 0;
692 }
693 
694 static int ocrdma_alloc_lkey(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
695 			    u32 pdid, int acc, u32 num_pbls, u32 addr_check)
696 {
697 	int status;
698 
699 	mr->hwmr.fr_mr = 0;
700 	mr->hwmr.local_rd = 1;
701 	mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
702 	mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
703 	mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
704 	mr->hwmr.mw_bind = (acc & IB_ACCESS_MW_BIND) ? 1 : 0;
705 	mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
706 	mr->hwmr.num_pbls = num_pbls;
707 
708 	status = ocrdma_mbx_alloc_lkey(dev, &mr->hwmr, pdid, addr_check);
709 	if (status)
710 		return status;
711 
712 	mr->ibmr.lkey = mr->hwmr.lkey;
713 	if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
714 		mr->ibmr.rkey = mr->hwmr.lkey;
715 	return 0;
716 }
717 
718 struct ib_mr *ocrdma_get_dma_mr(struct ib_pd *ibpd, int acc)
719 {
720 	int status;
721 	struct ocrdma_mr *mr;
722 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
723 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
724 
725 	if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE)) {
726 		pr_err("%s err, invalid access rights\n", __func__);
727 		return ERR_PTR(-EINVAL);
728 	}
729 
730 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
731 	if (!mr)
732 		return ERR_PTR(-ENOMEM);
733 
734 	status = ocrdma_alloc_lkey(dev, mr, pd->id, acc, 0,
735 				   OCRDMA_ADDR_CHECK_DISABLE);
736 	if (status) {
737 		kfree(mr);
738 		return ERR_PTR(status);
739 	}
740 
741 	return &mr->ibmr;
742 }
743 
744 static void ocrdma_free_mr_pbl_tbl(struct ocrdma_dev *dev,
745 				   struct ocrdma_hw_mr *mr)
746 {
747 	struct pci_dev *pdev = dev->nic_info.pdev;
748 	int i = 0;
749 
750 	if (mr->pbl_table) {
751 		for (i = 0; i < mr->num_pbls; i++) {
752 			if (!mr->pbl_table[i].va)
753 				continue;
754 			dma_free_coherent(&pdev->dev, mr->pbl_size,
755 					  mr->pbl_table[i].va,
756 					  mr->pbl_table[i].pa);
757 		}
758 		kfree(mr->pbl_table);
759 		mr->pbl_table = NULL;
760 	}
761 }
762 
763 static int ocrdma_get_pbl_info(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
764 			      u32 num_pbes)
765 {
766 	u32 num_pbls = 0;
767 	u32 idx = 0;
768 	int status = 0;
769 	u32 pbl_size;
770 
771 	do {
772 		pbl_size = OCRDMA_MIN_HPAGE_SIZE * (1 << idx);
773 		if (pbl_size > MAX_OCRDMA_PBL_SIZE) {
774 			status = -EFAULT;
775 			break;
776 		}
777 		num_pbls = roundup(num_pbes, (pbl_size / sizeof(u64)));
778 		num_pbls = num_pbls / (pbl_size / sizeof(u64));
779 		idx++;
780 	} while (num_pbls >= dev->attr.max_num_mr_pbl);
781 
782 	mr->hwmr.num_pbes = num_pbes;
783 	mr->hwmr.num_pbls = num_pbls;
784 	mr->hwmr.pbl_size = pbl_size;
785 	return status;
786 }
787 
788 static int ocrdma_build_pbl_tbl(struct ocrdma_dev *dev, struct ocrdma_hw_mr *mr)
789 {
790 	int status = 0;
791 	int i;
792 	u32 dma_len = mr->pbl_size;
793 	struct pci_dev *pdev = dev->nic_info.pdev;
794 	void *va;
795 	dma_addr_t pa;
796 
797 	mr->pbl_table = kcalloc(mr->num_pbls, sizeof(struct ocrdma_pbl),
798 				GFP_KERNEL);
799 
800 	if (!mr->pbl_table)
801 		return -ENOMEM;
802 
803 	for (i = 0; i < mr->num_pbls; i++) {
804 		va = dma_alloc_coherent(&pdev->dev, dma_len, &pa, GFP_KERNEL);
805 		if (!va) {
806 			ocrdma_free_mr_pbl_tbl(dev, mr);
807 			status = -ENOMEM;
808 			break;
809 		}
810 		mr->pbl_table[i].va = va;
811 		mr->pbl_table[i].pa = pa;
812 	}
813 	return status;
814 }
815 
816 static void build_user_pbes(struct ocrdma_dev *dev, struct ocrdma_mr *mr)
817 {
818 	struct ocrdma_pbe *pbe;
819 	struct ib_block_iter biter;
820 	struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
821 	int pbe_cnt;
822 	u64 pg_addr;
823 
824 	if (!mr->hwmr.num_pbes)
825 		return;
826 
827 	pbe = (struct ocrdma_pbe *)pbl_tbl->va;
828 	pbe_cnt = 0;
829 
830 	rdma_umem_for_each_dma_block (mr->umem, &biter, PAGE_SIZE) {
831 		/* store the page address in pbe */
832 		pg_addr = rdma_block_iter_dma_address(&biter);
833 		pbe->pa_lo = cpu_to_le32(pg_addr);
834 		pbe->pa_hi = cpu_to_le32(upper_32_bits(pg_addr));
835 		pbe_cnt += 1;
836 		pbe++;
837 
838 		/* if the given pbl is full storing the pbes,
839 		 * move to next pbl.
840 		 */
841 		if (pbe_cnt == (mr->hwmr.pbl_size / sizeof(u64))) {
842 			pbl_tbl++;
843 			pbe = (struct ocrdma_pbe *)pbl_tbl->va;
844 			pbe_cnt = 0;
845 		}
846 	}
847 }
848 
849 struct ib_mr *ocrdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
850 				 u64 usr_addr, int acc, struct ib_udata *udata)
851 {
852 	int status = -ENOMEM;
853 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
854 	struct ocrdma_mr *mr;
855 	struct ocrdma_pd *pd;
856 
857 	pd = get_ocrdma_pd(ibpd);
858 
859 	if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE))
860 		return ERR_PTR(-EINVAL);
861 
862 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
863 	if (!mr)
864 		return ERR_PTR(status);
865 	mr->umem = ib_umem_get(ibpd->device, start, len, acc);
866 	if (IS_ERR(mr->umem)) {
867 		status = -EFAULT;
868 		goto umem_err;
869 	}
870 	status = ocrdma_get_pbl_info(
871 		dev, mr, ib_umem_num_dma_blocks(mr->umem, PAGE_SIZE));
872 	if (status)
873 		goto umem_err;
874 
875 	mr->hwmr.pbe_size = PAGE_SIZE;
876 	mr->hwmr.va = usr_addr;
877 	mr->hwmr.len = len;
878 	mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
879 	mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
880 	mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
881 	mr->hwmr.local_rd = 1;
882 	mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
883 	status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
884 	if (status)
885 		goto umem_err;
886 	build_user_pbes(dev, mr);
887 	status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
888 	if (status)
889 		goto mbx_err;
890 	mr->ibmr.lkey = mr->hwmr.lkey;
891 	if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
892 		mr->ibmr.rkey = mr->hwmr.lkey;
893 
894 	return &mr->ibmr;
895 
896 mbx_err:
897 	ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
898 umem_err:
899 	kfree(mr);
900 	return ERR_PTR(status);
901 }
902 
903 int ocrdma_dereg_mr(struct ib_mr *ib_mr, struct ib_udata *udata)
904 {
905 	struct ocrdma_mr *mr = get_ocrdma_mr(ib_mr);
906 	struct ocrdma_dev *dev = get_ocrdma_dev(ib_mr->device);
907 
908 	(void) ocrdma_mbx_dealloc_lkey(dev, mr->hwmr.fr_mr, mr->hwmr.lkey);
909 
910 	kfree(mr->pages);
911 	ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
912 
913 	/* it could be user registered memory. */
914 	ib_umem_release(mr->umem);
915 	kfree(mr);
916 
917 	/* Don't stop cleanup, in case FW is unresponsive */
918 	if (dev->mqe_ctx.fw_error_state) {
919 		pr_err("%s(%d) fw not responding.\n",
920 		       __func__, dev->id);
921 	}
922 	return 0;
923 }
924 
925 static int ocrdma_copy_cq_uresp(struct ocrdma_dev *dev, struct ocrdma_cq *cq,
926 				struct ib_udata *udata)
927 {
928 	int status;
929 	struct ocrdma_ucontext *uctx = rdma_udata_to_drv_context(
930 		udata, struct ocrdma_ucontext, ibucontext);
931 	struct ocrdma_create_cq_uresp uresp;
932 
933 	/* this must be user flow! */
934 	if (!udata)
935 		return -EINVAL;
936 
937 	memset(&uresp, 0, sizeof(uresp));
938 	uresp.cq_id = cq->id;
939 	uresp.page_size = PAGE_ALIGN(cq->len);
940 	uresp.num_pages = 1;
941 	uresp.max_hw_cqe = cq->max_hw_cqe;
942 	uresp.page_addr[0] = virt_to_phys(cq->va);
943 	uresp.db_page_addr =  ocrdma_get_db_addr(dev, uctx->cntxt_pd->id);
944 	uresp.db_page_size = dev->nic_info.db_page_size;
945 	uresp.phase_change = cq->phase_change ? 1 : 0;
946 	status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
947 	if (status) {
948 		pr_err("%s(%d) copy error cqid=0x%x.\n",
949 		       __func__, dev->id, cq->id);
950 		goto err;
951 	}
952 	status = ocrdma_add_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
953 	if (status)
954 		goto err;
955 	status = ocrdma_add_mmap(uctx, uresp.page_addr[0], uresp.page_size);
956 	if (status) {
957 		ocrdma_del_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
958 		goto err;
959 	}
960 	cq->ucontext = uctx;
961 err:
962 	return status;
963 }
964 
965 int ocrdma_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
966 		     struct uverbs_attr_bundle *attrs)
967 {
968 	struct ib_udata *udata = &attrs->driver_udata;
969 	struct ib_device *ibdev = ibcq->device;
970 	int entries = attr->cqe;
971 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
972 	struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
973 	struct ocrdma_ucontext *uctx = rdma_udata_to_drv_context(
974 		udata, struct ocrdma_ucontext, ibucontext);
975 	u16 pd_id = 0;
976 	int status;
977 	struct ocrdma_create_cq_ureq ureq;
978 
979 	if (attr->flags)
980 		return -EOPNOTSUPP;
981 
982 	if (udata) {
983 		if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
984 			return -EFAULT;
985 	} else
986 		ureq.dpp_cq = 0;
987 
988 	spin_lock_init(&cq->cq_lock);
989 	spin_lock_init(&cq->comp_handler_lock);
990 	INIT_LIST_HEAD(&cq->sq_head);
991 	INIT_LIST_HEAD(&cq->rq_head);
992 
993 	if (udata)
994 		pd_id = uctx->cntxt_pd->id;
995 
996 	status = ocrdma_mbx_create_cq(dev, cq, entries, ureq.dpp_cq, pd_id);
997 	if (status)
998 		return status;
999 
1000 	if (udata) {
1001 		status = ocrdma_copy_cq_uresp(dev, cq, udata);
1002 		if (status)
1003 			goto ctx_err;
1004 	}
1005 	cq->phase = OCRDMA_CQE_VALID;
1006 	dev->cq_tbl[cq->id] = cq;
1007 	return 0;
1008 
1009 ctx_err:
1010 	ocrdma_mbx_destroy_cq(dev, cq);
1011 	return status;
1012 }
1013 
1014 int ocrdma_resize_cq(struct ib_cq *ibcq, int new_cnt,
1015 		     struct ib_udata *udata)
1016 {
1017 	int status = 0;
1018 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
1019 
1020 	if (new_cnt < 1 || new_cnt > cq->max_hw_cqe) {
1021 		status = -EINVAL;
1022 		return status;
1023 	}
1024 	ibcq->cqe = new_cnt;
1025 	return status;
1026 }
1027 
1028 static void ocrdma_flush_cq(struct ocrdma_cq *cq)
1029 {
1030 	int cqe_cnt;
1031 	int valid_count = 0;
1032 	unsigned long flags;
1033 
1034 	struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
1035 	struct ocrdma_cqe *cqe = NULL;
1036 
1037 	cqe = cq->va;
1038 	cqe_cnt = cq->cqe_cnt;
1039 
1040 	/* Last irq might have scheduled a polling thread
1041 	 * sync-up with it before hard flushing.
1042 	 */
1043 	spin_lock_irqsave(&cq->cq_lock, flags);
1044 	while (cqe_cnt) {
1045 		if (is_cqe_valid(cq, cqe))
1046 			valid_count++;
1047 		cqe++;
1048 		cqe_cnt--;
1049 	}
1050 	ocrdma_ring_cq_db(dev, cq->id, false, false, valid_count);
1051 	spin_unlock_irqrestore(&cq->cq_lock, flags);
1052 }
1053 
1054 int ocrdma_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
1055 {
1056 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
1057 	struct ocrdma_eq *eq = NULL;
1058 	struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
1059 	int pdid = 0;
1060 	u32 irq, indx;
1061 
1062 	dev->cq_tbl[cq->id] = NULL;
1063 	indx = ocrdma_get_eq_table_index(dev, cq->eqn);
1064 
1065 	eq = &dev->eq_tbl[indx];
1066 	irq = ocrdma_get_irq(dev, eq);
1067 	synchronize_irq(irq);
1068 	ocrdma_flush_cq(cq);
1069 
1070 	ocrdma_mbx_destroy_cq(dev, cq);
1071 	if (cq->ucontext) {
1072 		pdid = cq->ucontext->cntxt_pd->id;
1073 		ocrdma_del_mmap(cq->ucontext, (u64) cq->pa,
1074 				PAGE_ALIGN(cq->len));
1075 		ocrdma_del_mmap(cq->ucontext,
1076 				ocrdma_get_db_addr(dev, pdid),
1077 				dev->nic_info.db_page_size);
1078 	}
1079 	return 0;
1080 }
1081 
1082 static int ocrdma_add_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1083 {
1084 	int status = -EINVAL;
1085 
1086 	if (qp->id < OCRDMA_MAX_QP && dev->qp_tbl[qp->id] == NULL) {
1087 		dev->qp_tbl[qp->id] = qp;
1088 		status = 0;
1089 	}
1090 	return status;
1091 }
1092 
1093 static void ocrdma_del_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1094 {
1095 	dev->qp_tbl[qp->id] = NULL;
1096 }
1097 
1098 static int ocrdma_check_qp_params(struct ib_pd *ibpd, struct ocrdma_dev *dev,
1099 				  struct ib_qp_init_attr *attrs,
1100 				  struct ib_udata *udata)
1101 {
1102 	if ((attrs->qp_type != IB_QPT_GSI) &&
1103 	    (attrs->qp_type != IB_QPT_RC) &&
1104 	    (attrs->qp_type != IB_QPT_UC) &&
1105 	    (attrs->qp_type != IB_QPT_UD)) {
1106 		pr_err("%s(%d) unsupported qp type=0x%x requested\n",
1107 		       __func__, dev->id, attrs->qp_type);
1108 		return -EOPNOTSUPP;
1109 	}
1110 	/* Skip the check for QP1 to support CM size of 128 */
1111 	if ((attrs->qp_type != IB_QPT_GSI) &&
1112 	    (attrs->cap.max_send_wr > dev->attr.max_wqe)) {
1113 		pr_err("%s(%d) unsupported send_wr=0x%x requested\n",
1114 		       __func__, dev->id, attrs->cap.max_send_wr);
1115 		pr_err("%s(%d) supported send_wr=0x%x\n",
1116 		       __func__, dev->id, dev->attr.max_wqe);
1117 		return -EINVAL;
1118 	}
1119 	if (!attrs->srq && (attrs->cap.max_recv_wr > dev->attr.max_rqe)) {
1120 		pr_err("%s(%d) unsupported recv_wr=0x%x requested\n",
1121 		       __func__, dev->id, attrs->cap.max_recv_wr);
1122 		pr_err("%s(%d) supported recv_wr=0x%x\n",
1123 		       __func__, dev->id, dev->attr.max_rqe);
1124 		return -EINVAL;
1125 	}
1126 	if (attrs->cap.max_inline_data > dev->attr.max_inline_data) {
1127 		pr_err("%s(%d) unsupported inline data size=0x%x requested\n",
1128 		       __func__, dev->id, attrs->cap.max_inline_data);
1129 		pr_err("%s(%d) supported inline data size=0x%x\n",
1130 		       __func__, dev->id, dev->attr.max_inline_data);
1131 		return -EINVAL;
1132 	}
1133 	if (attrs->cap.max_send_sge > dev->attr.max_send_sge) {
1134 		pr_err("%s(%d) unsupported send_sge=0x%x requested\n",
1135 		       __func__, dev->id, attrs->cap.max_send_sge);
1136 		pr_err("%s(%d) supported send_sge=0x%x\n",
1137 		       __func__, dev->id, dev->attr.max_send_sge);
1138 		return -EINVAL;
1139 	}
1140 	if (attrs->cap.max_recv_sge > dev->attr.max_recv_sge) {
1141 		pr_err("%s(%d) unsupported recv_sge=0x%x requested\n",
1142 		       __func__, dev->id, attrs->cap.max_recv_sge);
1143 		pr_err("%s(%d) supported recv_sge=0x%x\n",
1144 		       __func__, dev->id, dev->attr.max_recv_sge);
1145 		return -EINVAL;
1146 	}
1147 	/* unprivileged user space cannot create special QP */
1148 	if (udata && attrs->qp_type == IB_QPT_GSI) {
1149 		pr_err
1150 		    ("%s(%d) Userspace can't create special QPs of type=0x%x\n",
1151 		     __func__, dev->id, attrs->qp_type);
1152 		return -EINVAL;
1153 	}
1154 	/* allow creating only one GSI type of QP */
1155 	if (attrs->qp_type == IB_QPT_GSI && dev->gsi_qp_created) {
1156 		pr_err("%s(%d) GSI special QPs already created.\n",
1157 		       __func__, dev->id);
1158 		return -EINVAL;
1159 	}
1160 	/* verify consumer QPs are not trying to use GSI QP's CQ */
1161 	if ((attrs->qp_type != IB_QPT_GSI) && (dev->gsi_qp_created)) {
1162 		if ((dev->gsi_sqcq == get_ocrdma_cq(attrs->send_cq)) ||
1163 			(dev->gsi_rqcq == get_ocrdma_cq(attrs->recv_cq))) {
1164 			pr_err("%s(%d) Consumer QP cannot use GSI CQs.\n",
1165 				__func__, dev->id);
1166 			return -EINVAL;
1167 		}
1168 	}
1169 	return 0;
1170 }
1171 
1172 static int ocrdma_copy_qp_uresp(struct ocrdma_qp *qp,
1173 				struct ib_udata *udata, int dpp_offset,
1174 				int dpp_credit_lmt, int srq)
1175 {
1176 	int status;
1177 	u64 usr_db;
1178 	struct ocrdma_create_qp_uresp uresp;
1179 	struct ocrdma_pd *pd = qp->pd;
1180 	struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
1181 
1182 	memset(&uresp, 0, sizeof(uresp));
1183 	usr_db = dev->nic_info.unmapped_db +
1184 			(pd->id * dev->nic_info.db_page_size);
1185 	uresp.qp_id = qp->id;
1186 	uresp.sq_dbid = qp->sq.dbid;
1187 	uresp.num_sq_pages = 1;
1188 	uresp.sq_page_size = PAGE_ALIGN(qp->sq.len);
1189 	uresp.sq_page_addr[0] = virt_to_phys(qp->sq.va);
1190 	uresp.num_wqe_allocated = qp->sq.max_cnt;
1191 	if (!srq) {
1192 		uresp.rq_dbid = qp->rq.dbid;
1193 		uresp.num_rq_pages = 1;
1194 		uresp.rq_page_size = PAGE_ALIGN(qp->rq.len);
1195 		uresp.rq_page_addr[0] = virt_to_phys(qp->rq.va);
1196 		uresp.num_rqe_allocated = qp->rq.max_cnt;
1197 	}
1198 	uresp.db_page_addr = usr_db;
1199 	uresp.db_page_size = dev->nic_info.db_page_size;
1200 	uresp.db_sq_offset = OCRDMA_DB_GEN2_SQ_OFFSET;
1201 	uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1202 	uresp.db_shift = OCRDMA_DB_RQ_SHIFT;
1203 
1204 	if (qp->dpp_enabled) {
1205 		uresp.dpp_credit = dpp_credit_lmt;
1206 		uresp.dpp_offset = dpp_offset;
1207 	}
1208 	status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1209 	if (status) {
1210 		pr_err("%s(%d) user copy error.\n", __func__, dev->id);
1211 		goto err;
1212 	}
1213 	status = ocrdma_add_mmap(pd->uctx, uresp.sq_page_addr[0],
1214 				 uresp.sq_page_size);
1215 	if (status)
1216 		goto err;
1217 
1218 	if (!srq) {
1219 		status = ocrdma_add_mmap(pd->uctx, uresp.rq_page_addr[0],
1220 					 uresp.rq_page_size);
1221 		if (status)
1222 			goto rq_map_err;
1223 	}
1224 	return status;
1225 rq_map_err:
1226 	ocrdma_del_mmap(pd->uctx, uresp.sq_page_addr[0], uresp.sq_page_size);
1227 err:
1228 	return status;
1229 }
1230 
1231 static void ocrdma_set_qp_db(struct ocrdma_dev *dev, struct ocrdma_qp *qp,
1232 			     struct ocrdma_pd *pd)
1233 {
1234 	if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1235 		qp->sq_db = dev->nic_info.db +
1236 			(pd->id * dev->nic_info.db_page_size) +
1237 			OCRDMA_DB_GEN2_SQ_OFFSET;
1238 		qp->rq_db = dev->nic_info.db +
1239 			(pd->id * dev->nic_info.db_page_size) +
1240 			OCRDMA_DB_GEN2_RQ_OFFSET;
1241 	} else {
1242 		qp->sq_db = dev->nic_info.db +
1243 			(pd->id * dev->nic_info.db_page_size) +
1244 			OCRDMA_DB_SQ_OFFSET;
1245 		qp->rq_db = dev->nic_info.db +
1246 			(pd->id * dev->nic_info.db_page_size) +
1247 			OCRDMA_DB_RQ_OFFSET;
1248 	}
1249 }
1250 
1251 static int ocrdma_alloc_wr_id_tbl(struct ocrdma_qp *qp)
1252 {
1253 	qp->wqe_wr_id_tbl =
1254 	    kcalloc(qp->sq.max_cnt, sizeof(*(qp->wqe_wr_id_tbl)),
1255 		    GFP_KERNEL);
1256 	if (qp->wqe_wr_id_tbl == NULL)
1257 		return -ENOMEM;
1258 	qp->rqe_wr_id_tbl =
1259 	    kcalloc(qp->rq.max_cnt, sizeof(u64), GFP_KERNEL);
1260 	if (qp->rqe_wr_id_tbl == NULL)
1261 		return -ENOMEM;
1262 
1263 	return 0;
1264 }
1265 
1266 static void ocrdma_set_qp_init_params(struct ocrdma_qp *qp,
1267 				      struct ocrdma_pd *pd,
1268 				      struct ib_qp_init_attr *attrs)
1269 {
1270 	qp->pd = pd;
1271 	spin_lock_init(&qp->q_lock);
1272 	INIT_LIST_HEAD(&qp->sq_entry);
1273 	INIT_LIST_HEAD(&qp->rq_entry);
1274 
1275 	qp->qp_type = attrs->qp_type;
1276 	qp->cap_flags = OCRDMA_QP_INB_RD | OCRDMA_QP_INB_WR;
1277 	qp->max_inline_data = attrs->cap.max_inline_data;
1278 	qp->sq.max_sges = attrs->cap.max_send_sge;
1279 	qp->rq.max_sges = attrs->cap.max_recv_sge;
1280 	qp->state = OCRDMA_QPS_RST;
1281 	qp->signaled = attrs->sq_sig_type == IB_SIGNAL_ALL_WR;
1282 }
1283 
1284 static void ocrdma_store_gsi_qp_cq(struct ocrdma_dev *dev,
1285 				   struct ib_qp_init_attr *attrs)
1286 {
1287 	if (attrs->qp_type == IB_QPT_GSI) {
1288 		dev->gsi_qp_created = 1;
1289 		dev->gsi_sqcq = get_ocrdma_cq(attrs->send_cq);
1290 		dev->gsi_rqcq = get_ocrdma_cq(attrs->recv_cq);
1291 	}
1292 }
1293 
1294 int ocrdma_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *attrs,
1295 		     struct ib_udata *udata)
1296 {
1297 	int status;
1298 	struct ib_pd *ibpd = ibqp->pd;
1299 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1300 	struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
1301 	struct ocrdma_dev *dev = get_ocrdma_dev(ibqp->device);
1302 	struct ocrdma_create_qp_ureq ureq;
1303 	u16 dpp_credit_lmt, dpp_offset;
1304 
1305 	if (attrs->create_flags)
1306 		return -EOPNOTSUPP;
1307 
1308 	status = ocrdma_check_qp_params(ibpd, dev, attrs, udata);
1309 	if (status)
1310 		goto gen_err;
1311 
1312 	memset(&ureq, 0, sizeof(ureq));
1313 	if (udata) {
1314 		if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1315 			return -EFAULT;
1316 	}
1317 	ocrdma_set_qp_init_params(qp, pd, attrs);
1318 	if (udata == NULL)
1319 		qp->cap_flags |= (OCRDMA_QP_MW_BIND | OCRDMA_QP_LKEY0 |
1320 					OCRDMA_QP_FAST_REG);
1321 
1322 	mutex_lock(&dev->dev_lock);
1323 	status = ocrdma_mbx_create_qp(qp, attrs, ureq.enable_dpp_cq,
1324 					ureq.dpp_cq_id,
1325 					&dpp_offset, &dpp_credit_lmt);
1326 	if (status)
1327 		goto mbx_err;
1328 
1329 	/* user space QP's wr_id table are managed in library */
1330 	if (udata == NULL) {
1331 		status = ocrdma_alloc_wr_id_tbl(qp);
1332 		if (status)
1333 			goto map_err;
1334 	}
1335 
1336 	status = ocrdma_add_qpn_map(dev, qp);
1337 	if (status)
1338 		goto map_err;
1339 	ocrdma_set_qp_db(dev, qp, pd);
1340 	if (udata) {
1341 		status = ocrdma_copy_qp_uresp(qp, udata, dpp_offset,
1342 					      dpp_credit_lmt,
1343 					      (attrs->srq != NULL));
1344 		if (status)
1345 			goto cpy_err;
1346 	}
1347 	ocrdma_store_gsi_qp_cq(dev, attrs);
1348 	qp->ibqp.qp_num = qp->id;
1349 	mutex_unlock(&dev->dev_lock);
1350 	return 0;
1351 
1352 cpy_err:
1353 	ocrdma_del_qpn_map(dev, qp);
1354 map_err:
1355 	ocrdma_mbx_destroy_qp(dev, qp);
1356 mbx_err:
1357 	mutex_unlock(&dev->dev_lock);
1358 	kfree(qp->wqe_wr_id_tbl);
1359 	kfree(qp->rqe_wr_id_tbl);
1360 	pr_err("%s(%d) error=%d\n", __func__, dev->id, status);
1361 gen_err:
1362 	return status;
1363 }
1364 
1365 int _ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1366 		      int attr_mask)
1367 {
1368 	int status = 0;
1369 	struct ocrdma_qp *qp;
1370 	struct ocrdma_dev *dev;
1371 	enum ib_qp_state old_qps;
1372 
1373 	qp = get_ocrdma_qp(ibqp);
1374 	dev = get_ocrdma_dev(ibqp->device);
1375 	if (attr_mask & IB_QP_STATE)
1376 		status = ocrdma_qp_state_change(qp, attr->qp_state, &old_qps);
1377 	/* if new and previous states are same hw doesn't need to
1378 	 * know about it.
1379 	 */
1380 	if (status < 0)
1381 		return status;
1382 	return ocrdma_mbx_modify_qp(dev, qp, attr, attr_mask);
1383 }
1384 
1385 int ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1386 		     int attr_mask, struct ib_udata *udata)
1387 {
1388 	unsigned long flags;
1389 	int status = -EINVAL;
1390 	struct ocrdma_qp *qp;
1391 	struct ocrdma_dev *dev;
1392 	enum ib_qp_state old_qps, new_qps;
1393 
1394 	if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1395 		return -EOPNOTSUPP;
1396 
1397 	qp = get_ocrdma_qp(ibqp);
1398 	dev = get_ocrdma_dev(ibqp->device);
1399 
1400 	/* syncronize with multiple context trying to change, retrive qps */
1401 	mutex_lock(&dev->dev_lock);
1402 	/* syncronize with wqe, rqe posting and cqe processing contexts */
1403 	spin_lock_irqsave(&qp->q_lock, flags);
1404 	old_qps = get_ibqp_state(qp->state);
1405 	if (attr_mask & IB_QP_STATE)
1406 		new_qps = attr->qp_state;
1407 	else
1408 		new_qps = old_qps;
1409 	spin_unlock_irqrestore(&qp->q_lock, flags);
1410 
1411 	if (!ib_modify_qp_is_ok(old_qps, new_qps, ibqp->qp_type, attr_mask)) {
1412 		pr_err("%s(%d) invalid attribute mask=0x%x specified for\n"
1413 		       "qpn=0x%x of type=0x%x old_qps=0x%x, new_qps=0x%x\n",
1414 		       __func__, dev->id, attr_mask, qp->id, ibqp->qp_type,
1415 		       old_qps, new_qps);
1416 		goto param_err;
1417 	}
1418 
1419 	status = _ocrdma_modify_qp(ibqp, attr, attr_mask);
1420 	if (status > 0)
1421 		status = 0;
1422 param_err:
1423 	mutex_unlock(&dev->dev_lock);
1424 	return status;
1425 }
1426 
1427 static enum ib_mtu ocrdma_mtu_int_to_enum(u16 mtu)
1428 {
1429 	switch (mtu) {
1430 	case 256:
1431 		return IB_MTU_256;
1432 	case 512:
1433 		return IB_MTU_512;
1434 	case 1024:
1435 		return IB_MTU_1024;
1436 	case 2048:
1437 		return IB_MTU_2048;
1438 	case 4096:
1439 		return IB_MTU_4096;
1440 	default:
1441 		return IB_MTU_1024;
1442 	}
1443 }
1444 
1445 static int ocrdma_to_ib_qp_acc_flags(int qp_cap_flags)
1446 {
1447 	int ib_qp_acc_flags = 0;
1448 
1449 	if (qp_cap_flags & OCRDMA_QP_INB_WR)
1450 		ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE;
1451 	if (qp_cap_flags & OCRDMA_QP_INB_RD)
1452 		ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE;
1453 	return ib_qp_acc_flags;
1454 }
1455 
1456 int ocrdma_query_qp(struct ib_qp *ibqp,
1457 		    struct ib_qp_attr *qp_attr,
1458 		    int attr_mask, struct ib_qp_init_attr *qp_init_attr)
1459 {
1460 	int status;
1461 	u32 qp_state;
1462 	struct ocrdma_qp_params params;
1463 	struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
1464 	struct ocrdma_dev *dev = get_ocrdma_dev(ibqp->device);
1465 
1466 	memset(&params, 0, sizeof(params));
1467 	mutex_lock(&dev->dev_lock);
1468 	status = ocrdma_mbx_query_qp(dev, qp, &params);
1469 	mutex_unlock(&dev->dev_lock);
1470 	if (status)
1471 		goto mbx_err;
1472 	if (qp->qp_type == IB_QPT_UD)
1473 		qp_attr->qkey = params.qkey;
1474 	qp_attr->path_mtu =
1475 		ocrdma_mtu_int_to_enum(params.path_mtu_pkey_indx &
1476 				OCRDMA_QP_PARAMS_PATH_MTU_MASK) >>
1477 				OCRDMA_QP_PARAMS_PATH_MTU_SHIFT;
1478 	qp_attr->path_mig_state = IB_MIG_MIGRATED;
1479 	qp_attr->rq_psn = params.hop_lmt_rq_psn & OCRDMA_QP_PARAMS_RQ_PSN_MASK;
1480 	qp_attr->sq_psn = params.tclass_sq_psn & OCRDMA_QP_PARAMS_SQ_PSN_MASK;
1481 	qp_attr->dest_qp_num =
1482 	    params.ack_to_rnr_rtc_dest_qpn & OCRDMA_QP_PARAMS_DEST_QPN_MASK;
1483 
1484 	qp_attr->qp_access_flags = ocrdma_to_ib_qp_acc_flags(qp->cap_flags);
1485 	qp_attr->cap.max_send_wr = qp->sq.max_cnt - 1;
1486 	qp_attr->cap.max_recv_wr = qp->rq.max_cnt - 1;
1487 	qp_attr->cap.max_send_sge = qp->sq.max_sges;
1488 	qp_attr->cap.max_recv_sge = qp->rq.max_sges;
1489 	qp_attr->cap.max_inline_data = qp->max_inline_data;
1490 	qp_init_attr->cap = qp_attr->cap;
1491 	qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
1492 
1493 	rdma_ah_set_grh(&qp_attr->ah_attr, NULL,
1494 			params.rnt_rc_sl_fl &
1495 			  OCRDMA_QP_PARAMS_FLOW_LABEL_MASK,
1496 			qp->sgid_idx,
1497 			(params.hop_lmt_rq_psn &
1498 			 OCRDMA_QP_PARAMS_HOP_LMT_MASK) >>
1499 			 OCRDMA_QP_PARAMS_HOP_LMT_SHIFT,
1500 			(params.tclass_sq_psn &
1501 			 OCRDMA_QP_PARAMS_TCLASS_MASK) >>
1502 			 OCRDMA_QP_PARAMS_TCLASS_SHIFT);
1503 	rdma_ah_set_dgid_raw(&qp_attr->ah_attr, &params.dgid[0]);
1504 
1505 	rdma_ah_set_port_num(&qp_attr->ah_attr, 1);
1506 	rdma_ah_set_sl(&qp_attr->ah_attr, (params.rnt_rc_sl_fl &
1507 					   OCRDMA_QP_PARAMS_SL_MASK) >>
1508 					   OCRDMA_QP_PARAMS_SL_SHIFT);
1509 	qp_attr->timeout = (params.ack_to_rnr_rtc_dest_qpn &
1510 			    OCRDMA_QP_PARAMS_ACK_TIMEOUT_MASK) >>
1511 				OCRDMA_QP_PARAMS_ACK_TIMEOUT_SHIFT;
1512 	qp_attr->rnr_retry = (params.ack_to_rnr_rtc_dest_qpn &
1513 			      OCRDMA_QP_PARAMS_RNR_RETRY_CNT_MASK) >>
1514 				OCRDMA_QP_PARAMS_RNR_RETRY_CNT_SHIFT;
1515 	qp_attr->retry_cnt =
1516 	    (params.rnt_rc_sl_fl & OCRDMA_QP_PARAMS_RETRY_CNT_MASK) >>
1517 		OCRDMA_QP_PARAMS_RETRY_CNT_SHIFT;
1518 	qp_attr->min_rnr_timer = 0;
1519 	qp_attr->pkey_index = 0;
1520 	qp_attr->port_num = 1;
1521 	rdma_ah_set_path_bits(&qp_attr->ah_attr, 0);
1522 	rdma_ah_set_static_rate(&qp_attr->ah_attr, 0);
1523 	qp_attr->alt_pkey_index = 0;
1524 	qp_attr->alt_port_num = 0;
1525 	qp_attr->alt_timeout = 0;
1526 	memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr));
1527 	qp_state = (params.max_sge_recv_flags & OCRDMA_QP_PARAMS_STATE_MASK) >>
1528 		    OCRDMA_QP_PARAMS_STATE_SHIFT;
1529 	qp_attr->qp_state = get_ibqp_state(qp_state);
1530 	qp_attr->cur_qp_state = qp_attr->qp_state;
1531 	qp_attr->sq_draining = (qp_state == OCRDMA_QPS_SQ_DRAINING) ? 1 : 0;
1532 	qp_attr->max_dest_rd_atomic =
1533 	    params.max_ord_ird >> OCRDMA_QP_PARAMS_MAX_ORD_SHIFT;
1534 	qp_attr->max_rd_atomic =
1535 	    params.max_ord_ird & OCRDMA_QP_PARAMS_MAX_IRD_MASK;
1536 	qp_attr->en_sqd_async_notify = (params.max_sge_recv_flags &
1537 				OCRDMA_QP_PARAMS_FLAGS_SQD_ASYNC) ? 1 : 0;
1538 	/* Sync driver QP state with FW */
1539 	ocrdma_qp_state_change(qp, qp_attr->qp_state, NULL);
1540 mbx_err:
1541 	return status;
1542 }
1543 
1544 static void ocrdma_srq_toggle_bit(struct ocrdma_srq *srq, unsigned int idx)
1545 {
1546 	unsigned int i = idx / 32;
1547 	u32 mask = (1U << (idx % 32));
1548 
1549 	srq->idx_bit_fields[i] ^= mask;
1550 }
1551 
1552 static int ocrdma_hwq_free_cnt(struct ocrdma_qp_hwq_info *q)
1553 {
1554 	return ((q->max_wqe_idx - q->head) + q->tail) % q->max_cnt;
1555 }
1556 
1557 static int is_hw_sq_empty(struct ocrdma_qp *qp)
1558 {
1559 	return (qp->sq.tail == qp->sq.head);
1560 }
1561 
1562 static int is_hw_rq_empty(struct ocrdma_qp *qp)
1563 {
1564 	return (qp->rq.tail == qp->rq.head);
1565 }
1566 
1567 static void *ocrdma_hwq_head(struct ocrdma_qp_hwq_info *q)
1568 {
1569 	return q->va + (q->head * q->entry_size);
1570 }
1571 
1572 static void *ocrdma_hwq_head_from_idx(struct ocrdma_qp_hwq_info *q,
1573 				      u32 idx)
1574 {
1575 	return q->va + (idx * q->entry_size);
1576 }
1577 
1578 static void ocrdma_hwq_inc_head(struct ocrdma_qp_hwq_info *q)
1579 {
1580 	q->head = (q->head + 1) & q->max_wqe_idx;
1581 }
1582 
1583 static void ocrdma_hwq_inc_tail(struct ocrdma_qp_hwq_info *q)
1584 {
1585 	q->tail = (q->tail + 1) & q->max_wqe_idx;
1586 }
1587 
1588 /* discard the cqe for a given QP */
1589 static void ocrdma_discard_cqes(struct ocrdma_qp *qp, struct ocrdma_cq *cq)
1590 {
1591 	unsigned long cq_flags;
1592 	unsigned long flags;
1593 	u32 cur_getp, stop_getp;
1594 	struct ocrdma_cqe *cqe;
1595 	u32 qpn = 0, wqe_idx = 0;
1596 
1597 	spin_lock_irqsave(&cq->cq_lock, cq_flags);
1598 
1599 	/* traverse through the CQEs in the hw CQ,
1600 	 * find the matching CQE for a given qp,
1601 	 * mark the matching one discarded by clearing qpn.
1602 	 * ring the doorbell in the poll_cq() as
1603 	 * we don't complete out of order cqe.
1604 	 */
1605 
1606 	cur_getp = cq->getp;
1607 	/* find upto when do we reap the cq. */
1608 	stop_getp = cur_getp;
1609 	do {
1610 		if (is_hw_sq_empty(qp) && (!qp->srq && is_hw_rq_empty(qp)))
1611 			break;
1612 
1613 		cqe = cq->va + cur_getp;
1614 		/* if (a) done reaping whole hw cq, or
1615 		 *    (b) qp_xq becomes empty.
1616 		 * then exit
1617 		 */
1618 		qpn = cqe->cmn.qpn & OCRDMA_CQE_QPN_MASK;
1619 		/* if previously discarded cqe found, skip that too. */
1620 		/* check for matching qp */
1621 		if (qpn == 0 || qpn != qp->id)
1622 			goto skip_cqe;
1623 
1624 		if (is_cqe_for_sq(cqe)) {
1625 			ocrdma_hwq_inc_tail(&qp->sq);
1626 		} else {
1627 			if (qp->srq) {
1628 				wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
1629 					OCRDMA_CQE_BUFTAG_SHIFT) &
1630 					qp->srq->rq.max_wqe_idx;
1631 				BUG_ON(wqe_idx < 1);
1632 				spin_lock_irqsave(&qp->srq->q_lock, flags);
1633 				ocrdma_hwq_inc_tail(&qp->srq->rq);
1634 				ocrdma_srq_toggle_bit(qp->srq, wqe_idx - 1);
1635 				spin_unlock_irqrestore(&qp->srq->q_lock, flags);
1636 
1637 			} else {
1638 				ocrdma_hwq_inc_tail(&qp->rq);
1639 			}
1640 		}
1641 		/* mark cqe discarded so that it is not picked up later
1642 		 * in the poll_cq().
1643 		 */
1644 		cqe->cmn.qpn = 0;
1645 skip_cqe:
1646 		cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
1647 	} while (cur_getp != stop_getp);
1648 	spin_unlock_irqrestore(&cq->cq_lock, cq_flags);
1649 }
1650 
1651 void ocrdma_del_flush_qp(struct ocrdma_qp *qp)
1652 {
1653 	int found = false;
1654 	unsigned long flags;
1655 	struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
1656 	/* sync with any active CQ poll */
1657 
1658 	spin_lock_irqsave(&dev->flush_q_lock, flags);
1659 	found = ocrdma_is_qp_in_sq_flushlist(qp->sq_cq, qp);
1660 	if (found)
1661 		list_del(&qp->sq_entry);
1662 	if (!qp->srq) {
1663 		found = ocrdma_is_qp_in_rq_flushlist(qp->rq_cq, qp);
1664 		if (found)
1665 			list_del(&qp->rq_entry);
1666 	}
1667 	spin_unlock_irqrestore(&dev->flush_q_lock, flags);
1668 }
1669 
1670 int ocrdma_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
1671 {
1672 	struct ocrdma_pd *pd;
1673 	struct ocrdma_qp *qp;
1674 	struct ocrdma_dev *dev;
1675 	struct ib_qp_attr attrs;
1676 	int attr_mask;
1677 	unsigned long flags;
1678 
1679 	qp = get_ocrdma_qp(ibqp);
1680 	dev = get_ocrdma_dev(ibqp->device);
1681 
1682 	pd = qp->pd;
1683 
1684 	/* change the QP state to ERROR */
1685 	if (qp->state != OCRDMA_QPS_RST) {
1686 		attrs.qp_state = IB_QPS_ERR;
1687 		attr_mask = IB_QP_STATE;
1688 		_ocrdma_modify_qp(ibqp, &attrs, attr_mask);
1689 	}
1690 	/* ensure that CQEs for newly created QP (whose id may be same with
1691 	 * one which just getting destroyed are same), dont get
1692 	 * discarded until the old CQEs are discarded.
1693 	 */
1694 	mutex_lock(&dev->dev_lock);
1695 	(void) ocrdma_mbx_destroy_qp(dev, qp);
1696 
1697 	/*
1698 	 * acquire CQ lock while destroy is in progress, in order to
1699 	 * protect against proessing in-flight CQEs for this QP.
1700 	 */
1701 	spin_lock_irqsave(&qp->sq_cq->cq_lock, flags);
1702 	if (qp->rq_cq && (qp->rq_cq != qp->sq_cq)) {
1703 		spin_lock(&qp->rq_cq->cq_lock);
1704 		ocrdma_del_qpn_map(dev, qp);
1705 		spin_unlock(&qp->rq_cq->cq_lock);
1706 	} else {
1707 		ocrdma_del_qpn_map(dev, qp);
1708 	}
1709 	spin_unlock_irqrestore(&qp->sq_cq->cq_lock, flags);
1710 
1711 	if (!pd->uctx) {
1712 		ocrdma_discard_cqes(qp, qp->sq_cq);
1713 		ocrdma_discard_cqes(qp, qp->rq_cq);
1714 	}
1715 	mutex_unlock(&dev->dev_lock);
1716 
1717 	if (pd->uctx) {
1718 		ocrdma_del_mmap(pd->uctx, (u64) qp->sq.pa,
1719 				PAGE_ALIGN(qp->sq.len));
1720 		if (!qp->srq)
1721 			ocrdma_del_mmap(pd->uctx, (u64) qp->rq.pa,
1722 					PAGE_ALIGN(qp->rq.len));
1723 	}
1724 
1725 	ocrdma_del_flush_qp(qp);
1726 
1727 	kfree(qp->wqe_wr_id_tbl);
1728 	kfree(qp->rqe_wr_id_tbl);
1729 	return 0;
1730 }
1731 
1732 static int ocrdma_copy_srq_uresp(struct ocrdma_dev *dev, struct ocrdma_srq *srq,
1733 				struct ib_udata *udata)
1734 {
1735 	int status;
1736 	struct ocrdma_create_srq_uresp uresp;
1737 
1738 	memset(&uresp, 0, sizeof(uresp));
1739 	uresp.rq_dbid = srq->rq.dbid;
1740 	uresp.num_rq_pages = 1;
1741 	uresp.rq_page_addr[0] = virt_to_phys(srq->rq.va);
1742 	uresp.rq_page_size = srq->rq.len;
1743 	uresp.db_page_addr = dev->nic_info.unmapped_db +
1744 	    (srq->pd->id * dev->nic_info.db_page_size);
1745 	uresp.db_page_size = dev->nic_info.db_page_size;
1746 	uresp.num_rqe_allocated = srq->rq.max_cnt;
1747 	if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1748 		uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1749 		uresp.db_shift = 24;
1750 	} else {
1751 		uresp.db_rq_offset = OCRDMA_DB_RQ_OFFSET;
1752 		uresp.db_shift = 16;
1753 	}
1754 
1755 	status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1756 	if (status)
1757 		return status;
1758 	status = ocrdma_add_mmap(srq->pd->uctx, uresp.rq_page_addr[0],
1759 				 uresp.rq_page_size);
1760 	if (status)
1761 		return status;
1762 	return status;
1763 }
1764 
1765 int ocrdma_create_srq(struct ib_srq *ibsrq, struct ib_srq_init_attr *init_attr,
1766 		      struct ib_udata *udata)
1767 {
1768 	int status;
1769 	struct ocrdma_pd *pd = get_ocrdma_pd(ibsrq->pd);
1770 	struct ocrdma_dev *dev = get_ocrdma_dev(ibsrq->device);
1771 	struct ocrdma_srq *srq = get_ocrdma_srq(ibsrq);
1772 
1773 	if (init_attr->srq_type != IB_SRQT_BASIC)
1774 		return -EOPNOTSUPP;
1775 
1776 	if (init_attr->attr.max_sge > dev->attr.max_recv_sge)
1777 		return -EINVAL;
1778 	if (init_attr->attr.max_wr > dev->attr.max_rqe)
1779 		return -EINVAL;
1780 
1781 	spin_lock_init(&srq->q_lock);
1782 	srq->pd = pd;
1783 	srq->db = dev->nic_info.db + (pd->id * dev->nic_info.db_page_size);
1784 	status = ocrdma_mbx_create_srq(dev, srq, init_attr, pd);
1785 	if (status)
1786 		return status;
1787 
1788 	if (!udata) {
1789 		srq->rqe_wr_id_tbl = kcalloc(srq->rq.max_cnt, sizeof(u64),
1790 					     GFP_KERNEL);
1791 		if (!srq->rqe_wr_id_tbl) {
1792 			status = -ENOMEM;
1793 			goto arm_err;
1794 		}
1795 
1796 		srq->bit_fields_len = (srq->rq.max_cnt / 32) +
1797 		    (srq->rq.max_cnt % 32 ? 1 : 0);
1798 		srq->idx_bit_fields =
1799 		    kmalloc_array(srq->bit_fields_len, sizeof(u32),
1800 				  GFP_KERNEL);
1801 		if (!srq->idx_bit_fields) {
1802 			status = -ENOMEM;
1803 			goto arm_err;
1804 		}
1805 		memset(srq->idx_bit_fields, 0xff,
1806 		       srq->bit_fields_len * sizeof(u32));
1807 	}
1808 
1809 	if (init_attr->attr.srq_limit) {
1810 		status = ocrdma_mbx_modify_srq(srq, &init_attr->attr);
1811 		if (status)
1812 			goto arm_err;
1813 	}
1814 
1815 	if (udata) {
1816 		status = ocrdma_copy_srq_uresp(dev, srq, udata);
1817 		if (status)
1818 			goto arm_err;
1819 	}
1820 
1821 	return 0;
1822 
1823 arm_err:
1824 	ocrdma_mbx_destroy_srq(dev, srq);
1825 	kfree(srq->rqe_wr_id_tbl);
1826 	kfree(srq->idx_bit_fields);
1827 	return status;
1828 }
1829 
1830 int ocrdma_modify_srq(struct ib_srq *ibsrq,
1831 		      struct ib_srq_attr *srq_attr,
1832 		      enum ib_srq_attr_mask srq_attr_mask,
1833 		      struct ib_udata *udata)
1834 {
1835 	int status;
1836 	struct ocrdma_srq *srq;
1837 
1838 	srq = get_ocrdma_srq(ibsrq);
1839 	if (srq_attr_mask & IB_SRQ_MAX_WR)
1840 		status = -EINVAL;
1841 	else
1842 		status = ocrdma_mbx_modify_srq(srq, srq_attr);
1843 	return status;
1844 }
1845 
1846 int ocrdma_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
1847 {
1848 	struct ocrdma_srq *srq;
1849 
1850 	srq = get_ocrdma_srq(ibsrq);
1851 	return ocrdma_mbx_query_srq(srq, srq_attr);
1852 }
1853 
1854 int ocrdma_destroy_srq(struct ib_srq *ibsrq, struct ib_udata *udata)
1855 {
1856 	struct ocrdma_srq *srq;
1857 	struct ocrdma_dev *dev = get_ocrdma_dev(ibsrq->device);
1858 
1859 	srq = get_ocrdma_srq(ibsrq);
1860 
1861 	ocrdma_mbx_destroy_srq(dev, srq);
1862 
1863 	if (srq->pd->uctx)
1864 		ocrdma_del_mmap(srq->pd->uctx, (u64) srq->rq.pa,
1865 				PAGE_ALIGN(srq->rq.len));
1866 
1867 	kfree(srq->idx_bit_fields);
1868 	kfree(srq->rqe_wr_id_tbl);
1869 	return 0;
1870 }
1871 
1872 /* unprivileged verbs and their support functions. */
1873 static void ocrdma_build_ud_hdr(struct ocrdma_qp *qp,
1874 				struct ocrdma_hdr_wqe *hdr,
1875 				const struct ib_send_wr *wr)
1876 {
1877 	struct ocrdma_ewqe_ud_hdr *ud_hdr =
1878 		(struct ocrdma_ewqe_ud_hdr *)(hdr + 1);
1879 	struct ocrdma_ah *ah = get_ocrdma_ah(ud_wr(wr)->ah);
1880 
1881 	ud_hdr->rsvd_dest_qpn = ud_wr(wr)->remote_qpn;
1882 	if (qp->qp_type == IB_QPT_GSI)
1883 		ud_hdr->qkey = qp->qkey;
1884 	else
1885 		ud_hdr->qkey = ud_wr(wr)->remote_qkey;
1886 	ud_hdr->rsvd_ahid = ah->id;
1887 	ud_hdr->hdr_type = ah->hdr_type;
1888 	if (ah->av->valid & OCRDMA_AV_VLAN_VALID)
1889 		hdr->cw |= (OCRDMA_FLAG_AH_VLAN_PR << OCRDMA_WQE_FLAGS_SHIFT);
1890 }
1891 
1892 static void ocrdma_build_sges(struct ocrdma_hdr_wqe *hdr,
1893 			      struct ocrdma_sge *sge, int num_sge,
1894 			      struct ib_sge *sg_list)
1895 {
1896 	int i;
1897 
1898 	for (i = 0; i < num_sge; i++) {
1899 		sge[i].lrkey = sg_list[i].lkey;
1900 		sge[i].addr_lo = sg_list[i].addr;
1901 		sge[i].addr_hi = upper_32_bits(sg_list[i].addr);
1902 		sge[i].len = sg_list[i].length;
1903 		hdr->total_len += sg_list[i].length;
1904 	}
1905 	if (num_sge == 0)
1906 		memset(sge, 0, sizeof(*sge));
1907 }
1908 
1909 static inline uint32_t ocrdma_sglist_len(struct ib_sge *sg_list, int num_sge)
1910 {
1911 	uint32_t total_len = 0, i;
1912 
1913 	for (i = 0; i < num_sge; i++)
1914 		total_len += sg_list[i].length;
1915 	return total_len;
1916 }
1917 
1918 
1919 static int ocrdma_build_inline_sges(struct ocrdma_qp *qp,
1920 				    struct ocrdma_hdr_wqe *hdr,
1921 				    struct ocrdma_sge *sge,
1922 				    const struct ib_send_wr *wr, u32 wqe_size)
1923 {
1924 	int i;
1925 	char *dpp_addr;
1926 
1927 	if (wr->send_flags & IB_SEND_INLINE && qp->qp_type != IB_QPT_UD) {
1928 		hdr->total_len = ocrdma_sglist_len(wr->sg_list, wr->num_sge);
1929 		if (unlikely(hdr->total_len > qp->max_inline_data)) {
1930 			pr_err("%s() supported_len=0x%x,\n"
1931 			       " unsupported len req=0x%x\n", __func__,
1932 				qp->max_inline_data, hdr->total_len);
1933 			return -EINVAL;
1934 		}
1935 		dpp_addr = (char *)sge;
1936 		for (i = 0; i < wr->num_sge; i++) {
1937 			memcpy(dpp_addr,
1938 			       (void *)(unsigned long)wr->sg_list[i].addr,
1939 			       wr->sg_list[i].length);
1940 			dpp_addr += wr->sg_list[i].length;
1941 		}
1942 
1943 		wqe_size += roundup(hdr->total_len, OCRDMA_WQE_ALIGN_BYTES);
1944 		if (0 == hdr->total_len)
1945 			wqe_size += sizeof(struct ocrdma_sge);
1946 		hdr->cw |= (OCRDMA_TYPE_INLINE << OCRDMA_WQE_TYPE_SHIFT);
1947 	} else {
1948 		ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
1949 		if (wr->num_sge)
1950 			wqe_size += (wr->num_sge * sizeof(struct ocrdma_sge));
1951 		else
1952 			wqe_size += sizeof(struct ocrdma_sge);
1953 		hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
1954 	}
1955 	hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
1956 	return 0;
1957 }
1958 
1959 static int ocrdma_build_send(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1960 			     const struct ib_send_wr *wr)
1961 {
1962 	struct ocrdma_sge *sge;
1963 	u32 wqe_size = sizeof(*hdr);
1964 
1965 	if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
1966 		ocrdma_build_ud_hdr(qp, hdr, wr);
1967 		sge = (struct ocrdma_sge *)(hdr + 2);
1968 		wqe_size += sizeof(struct ocrdma_ewqe_ud_hdr);
1969 	} else {
1970 		sge = (struct ocrdma_sge *)(hdr + 1);
1971 	}
1972 
1973 	return ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
1974 }
1975 
1976 static int ocrdma_build_write(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1977 			      const struct ib_send_wr *wr)
1978 {
1979 	int status;
1980 	struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
1981 	struct ocrdma_sge *sge = ext_rw + 1;
1982 	u32 wqe_size = sizeof(*hdr) + sizeof(*ext_rw);
1983 
1984 	status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
1985 	if (status)
1986 		return status;
1987 	ext_rw->addr_lo = rdma_wr(wr)->remote_addr;
1988 	ext_rw->addr_hi = upper_32_bits(rdma_wr(wr)->remote_addr);
1989 	ext_rw->lrkey = rdma_wr(wr)->rkey;
1990 	ext_rw->len = hdr->total_len;
1991 	return 0;
1992 }
1993 
1994 static void ocrdma_build_read(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1995 			      const struct ib_send_wr *wr)
1996 {
1997 	struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
1998 	struct ocrdma_sge *sge = ext_rw + 1;
1999 	u32 wqe_size = ((wr->num_sge + 1) * sizeof(struct ocrdma_sge)) +
2000 	    sizeof(struct ocrdma_hdr_wqe);
2001 
2002 	ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
2003 	hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2004 	hdr->cw |= (OCRDMA_READ << OCRDMA_WQE_OPCODE_SHIFT);
2005 	hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2006 
2007 	ext_rw->addr_lo = rdma_wr(wr)->remote_addr;
2008 	ext_rw->addr_hi = upper_32_bits(rdma_wr(wr)->remote_addr);
2009 	ext_rw->lrkey = rdma_wr(wr)->rkey;
2010 	ext_rw->len = hdr->total_len;
2011 }
2012 
2013 static int get_encoded_page_size(int pg_sz)
2014 {
2015 	/* Max size is 256M 4096 << 16 */
2016 	int i = 0;
2017 	for (; i < 17; i++)
2018 		if (pg_sz == (4096 << i))
2019 			break;
2020 	return i;
2021 }
2022 
2023 static int ocrdma_build_reg(struct ocrdma_qp *qp,
2024 			    struct ocrdma_hdr_wqe *hdr,
2025 			    const struct ib_reg_wr *wr)
2026 {
2027 	u64 fbo;
2028 	struct ocrdma_ewqe_fr *fast_reg = (struct ocrdma_ewqe_fr *)(hdr + 1);
2029 	struct ocrdma_mr *mr = get_ocrdma_mr(wr->mr);
2030 	struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
2031 	struct ocrdma_pbe *pbe;
2032 	u32 wqe_size = sizeof(*fast_reg) + sizeof(*hdr);
2033 	int num_pbes = 0, i;
2034 
2035 	wqe_size = roundup(wqe_size, OCRDMA_WQE_ALIGN_BYTES);
2036 
2037 	hdr->cw |= (OCRDMA_FR_MR << OCRDMA_WQE_OPCODE_SHIFT);
2038 	hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2039 
2040 	if (wr->access & IB_ACCESS_LOCAL_WRITE)
2041 		hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_LOCAL_WR;
2042 	if (wr->access & IB_ACCESS_REMOTE_WRITE)
2043 		hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_WR;
2044 	if (wr->access & IB_ACCESS_REMOTE_READ)
2045 		hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_RD;
2046 	hdr->lkey = wr->key;
2047 	hdr->total_len = mr->ibmr.length;
2048 
2049 	fbo = mr->ibmr.iova - mr->pages[0];
2050 
2051 	fast_reg->va_hi = upper_32_bits(mr->ibmr.iova);
2052 	fast_reg->va_lo = (u32) (mr->ibmr.iova & 0xffffffff);
2053 	fast_reg->fbo_hi = upper_32_bits(fbo);
2054 	fast_reg->fbo_lo = (u32) fbo & 0xffffffff;
2055 	fast_reg->num_sges = mr->npages;
2056 	fast_reg->size_sge = get_encoded_page_size(mr->ibmr.page_size);
2057 
2058 	pbe = pbl_tbl->va;
2059 	for (i = 0; i < mr->npages; i++) {
2060 		u64 buf_addr = mr->pages[i];
2061 
2062 		pbe->pa_lo = cpu_to_le32((u32) (buf_addr & PAGE_MASK));
2063 		pbe->pa_hi = cpu_to_le32((u32) upper_32_bits(buf_addr));
2064 		num_pbes += 1;
2065 		pbe++;
2066 
2067 		/* if the pbl is full storing the pbes,
2068 		 * move to next pbl.
2069 		*/
2070 		if (num_pbes == (mr->hwmr.pbl_size/sizeof(u64))) {
2071 			pbl_tbl++;
2072 			pbe = (struct ocrdma_pbe *)pbl_tbl->va;
2073 		}
2074 	}
2075 
2076 	return 0;
2077 }
2078 
2079 static void ocrdma_ring_sq_db(struct ocrdma_qp *qp)
2080 {
2081 	u32 val = qp->sq.dbid | (1 << OCRDMA_DB_SQ_SHIFT);
2082 
2083 	iowrite32(val, qp->sq_db);
2084 }
2085 
2086 int ocrdma_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
2087 		     const struct ib_send_wr **bad_wr)
2088 {
2089 	int status = 0;
2090 	struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2091 	struct ocrdma_hdr_wqe *hdr;
2092 	unsigned long flags;
2093 
2094 	spin_lock_irqsave(&qp->q_lock, flags);
2095 	if (qp->state != OCRDMA_QPS_RTS && qp->state != OCRDMA_QPS_SQD) {
2096 		spin_unlock_irqrestore(&qp->q_lock, flags);
2097 		*bad_wr = wr;
2098 		return -EINVAL;
2099 	}
2100 
2101 	while (wr) {
2102 		if (qp->qp_type == IB_QPT_UD &&
2103 		    (wr->opcode != IB_WR_SEND &&
2104 		     wr->opcode != IB_WR_SEND_WITH_IMM)) {
2105 			*bad_wr = wr;
2106 			status = -EINVAL;
2107 			break;
2108 		}
2109 		if (ocrdma_hwq_free_cnt(&qp->sq) == 0 ||
2110 		    wr->num_sge > qp->sq.max_sges) {
2111 			*bad_wr = wr;
2112 			status = -ENOMEM;
2113 			break;
2114 		}
2115 		hdr = ocrdma_hwq_head(&qp->sq);
2116 		hdr->cw = 0;
2117 		if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2118 			hdr->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2119 		if (wr->send_flags & IB_SEND_FENCE)
2120 			hdr->cw |=
2121 			    (OCRDMA_FLAG_FENCE_L << OCRDMA_WQE_FLAGS_SHIFT);
2122 		if (wr->send_flags & IB_SEND_SOLICITED)
2123 			hdr->cw |=
2124 			    (OCRDMA_FLAG_SOLICIT << OCRDMA_WQE_FLAGS_SHIFT);
2125 		hdr->total_len = 0;
2126 		switch (wr->opcode) {
2127 		case IB_WR_SEND_WITH_IMM:
2128 			hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2129 			hdr->immdt = ntohl(wr->ex.imm_data);
2130 			fallthrough;
2131 		case IB_WR_SEND:
2132 			hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2133 			ocrdma_build_send(qp, hdr, wr);
2134 			break;
2135 		case IB_WR_SEND_WITH_INV:
2136 			hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
2137 			hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2138 			hdr->lkey = wr->ex.invalidate_rkey;
2139 			status = ocrdma_build_send(qp, hdr, wr);
2140 			break;
2141 		case IB_WR_RDMA_WRITE_WITH_IMM:
2142 			hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2143 			hdr->immdt = ntohl(wr->ex.imm_data);
2144 			fallthrough;
2145 		case IB_WR_RDMA_WRITE:
2146 			hdr->cw |= (OCRDMA_WRITE << OCRDMA_WQE_OPCODE_SHIFT);
2147 			status = ocrdma_build_write(qp, hdr, wr);
2148 			break;
2149 		case IB_WR_RDMA_READ:
2150 			ocrdma_build_read(qp, hdr, wr);
2151 			break;
2152 		case IB_WR_LOCAL_INV:
2153 			hdr->cw |=
2154 			    (OCRDMA_LKEY_INV << OCRDMA_WQE_OPCODE_SHIFT);
2155 			hdr->cw |= ((sizeof(struct ocrdma_hdr_wqe) +
2156 					sizeof(struct ocrdma_sge)) /
2157 				OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT;
2158 			hdr->lkey = wr->ex.invalidate_rkey;
2159 			break;
2160 		case IB_WR_REG_MR:
2161 			status = ocrdma_build_reg(qp, hdr, reg_wr(wr));
2162 			break;
2163 		default:
2164 			status = -EINVAL;
2165 			break;
2166 		}
2167 		if (status) {
2168 			*bad_wr = wr;
2169 			break;
2170 		}
2171 		if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2172 			qp->wqe_wr_id_tbl[qp->sq.head].signaled = 1;
2173 		else
2174 			qp->wqe_wr_id_tbl[qp->sq.head].signaled = 0;
2175 		qp->wqe_wr_id_tbl[qp->sq.head].wrid = wr->wr_id;
2176 		ocrdma_cpu_to_le32(hdr, ((hdr->cw >> OCRDMA_WQE_SIZE_SHIFT) &
2177 				   OCRDMA_WQE_SIZE_MASK) * OCRDMA_WQE_STRIDE);
2178 		/* make sure wqe is written before adapter can access it */
2179 		wmb();
2180 		/* inform hw to start processing it */
2181 		ocrdma_ring_sq_db(qp);
2182 
2183 		/* update pointer, counter for next wr */
2184 		ocrdma_hwq_inc_head(&qp->sq);
2185 		wr = wr->next;
2186 	}
2187 	spin_unlock_irqrestore(&qp->q_lock, flags);
2188 	return status;
2189 }
2190 
2191 static void ocrdma_ring_rq_db(struct ocrdma_qp *qp)
2192 {
2193 	u32 val = qp->rq.dbid | (1 << OCRDMA_DB_RQ_SHIFT);
2194 
2195 	iowrite32(val, qp->rq_db);
2196 }
2197 
2198 static void ocrdma_build_rqe(struct ocrdma_hdr_wqe *rqe,
2199 			     const struct ib_recv_wr *wr, u16 tag)
2200 {
2201 	u32 wqe_size = 0;
2202 	struct ocrdma_sge *sge;
2203 	if (wr->num_sge)
2204 		wqe_size = (wr->num_sge * sizeof(*sge)) + sizeof(*rqe);
2205 	else
2206 		wqe_size = sizeof(*sge) + sizeof(*rqe);
2207 
2208 	rqe->cw = ((wqe_size / OCRDMA_WQE_STRIDE) <<
2209 				OCRDMA_WQE_SIZE_SHIFT);
2210 	rqe->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2211 	rqe->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2212 	rqe->total_len = 0;
2213 	rqe->rsvd_tag = tag;
2214 	sge = (struct ocrdma_sge *)(rqe + 1);
2215 	ocrdma_build_sges(rqe, sge, wr->num_sge, wr->sg_list);
2216 	ocrdma_cpu_to_le32(rqe, wqe_size);
2217 }
2218 
2219 int ocrdma_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
2220 		     const struct ib_recv_wr **bad_wr)
2221 {
2222 	int status = 0;
2223 	unsigned long flags;
2224 	struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2225 	struct ocrdma_hdr_wqe *rqe;
2226 
2227 	spin_lock_irqsave(&qp->q_lock, flags);
2228 	if (qp->state == OCRDMA_QPS_RST || qp->state == OCRDMA_QPS_ERR) {
2229 		spin_unlock_irqrestore(&qp->q_lock, flags);
2230 		*bad_wr = wr;
2231 		return -EINVAL;
2232 	}
2233 	while (wr) {
2234 		if (ocrdma_hwq_free_cnt(&qp->rq) == 0 ||
2235 		    wr->num_sge > qp->rq.max_sges) {
2236 			*bad_wr = wr;
2237 			status = -ENOMEM;
2238 			break;
2239 		}
2240 		rqe = ocrdma_hwq_head(&qp->rq);
2241 		ocrdma_build_rqe(rqe, wr, 0);
2242 
2243 		qp->rqe_wr_id_tbl[qp->rq.head] = wr->wr_id;
2244 		/* make sure rqe is written before adapter can access it */
2245 		wmb();
2246 
2247 		/* inform hw to start processing it */
2248 		ocrdma_ring_rq_db(qp);
2249 
2250 		/* update pointer, counter for next wr */
2251 		ocrdma_hwq_inc_head(&qp->rq);
2252 		wr = wr->next;
2253 	}
2254 	spin_unlock_irqrestore(&qp->q_lock, flags);
2255 	return status;
2256 }
2257 
2258 /* cqe for srq's rqe can potentially arrive out of order.
2259  * index gives the entry in the shadow table where to store
2260  * the wr_id. tag/index is returned in cqe to reference back
2261  * for a given rqe.
2262  */
2263 static int ocrdma_srq_get_idx(struct ocrdma_srq *srq)
2264 {
2265 	int row = 0;
2266 	int indx = 0;
2267 
2268 	for (row = 0; row < srq->bit_fields_len; row++) {
2269 		if (srq->idx_bit_fields[row]) {
2270 			indx = ffs(srq->idx_bit_fields[row]);
2271 			indx = (row * 32) + (indx - 1);
2272 			BUG_ON(indx >= srq->rq.max_cnt);
2273 			ocrdma_srq_toggle_bit(srq, indx);
2274 			break;
2275 		}
2276 	}
2277 
2278 	BUG_ON(row == srq->bit_fields_len);
2279 	return indx + 1; /* Use from index 1 */
2280 }
2281 
2282 static void ocrdma_ring_srq_db(struct ocrdma_srq *srq)
2283 {
2284 	u32 val = srq->rq.dbid | (1 << 16);
2285 
2286 	iowrite32(val, srq->db + OCRDMA_DB_GEN2_SRQ_OFFSET);
2287 }
2288 
2289 int ocrdma_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
2290 			 const struct ib_recv_wr **bad_wr)
2291 {
2292 	int status = 0;
2293 	unsigned long flags;
2294 	struct ocrdma_srq *srq;
2295 	struct ocrdma_hdr_wqe *rqe;
2296 	u16 tag;
2297 
2298 	srq = get_ocrdma_srq(ibsrq);
2299 
2300 	spin_lock_irqsave(&srq->q_lock, flags);
2301 	while (wr) {
2302 		if (ocrdma_hwq_free_cnt(&srq->rq) == 0 ||
2303 		    wr->num_sge > srq->rq.max_sges) {
2304 			status = -ENOMEM;
2305 			*bad_wr = wr;
2306 			break;
2307 		}
2308 		tag = ocrdma_srq_get_idx(srq);
2309 		rqe = ocrdma_hwq_head(&srq->rq);
2310 		ocrdma_build_rqe(rqe, wr, tag);
2311 
2312 		srq->rqe_wr_id_tbl[tag] = wr->wr_id;
2313 		/* make sure rqe is written before adapter can perform DMA */
2314 		wmb();
2315 		/* inform hw to start processing it */
2316 		ocrdma_ring_srq_db(srq);
2317 		/* update pointer, counter for next wr */
2318 		ocrdma_hwq_inc_head(&srq->rq);
2319 		wr = wr->next;
2320 	}
2321 	spin_unlock_irqrestore(&srq->q_lock, flags);
2322 	return status;
2323 }
2324 
2325 static enum ib_wc_status ocrdma_to_ibwc_err(u16 status)
2326 {
2327 	enum ib_wc_status ibwc_status;
2328 
2329 	switch (status) {
2330 	case OCRDMA_CQE_GENERAL_ERR:
2331 		ibwc_status = IB_WC_GENERAL_ERR;
2332 		break;
2333 	case OCRDMA_CQE_LOC_LEN_ERR:
2334 		ibwc_status = IB_WC_LOC_LEN_ERR;
2335 		break;
2336 	case OCRDMA_CQE_LOC_QP_OP_ERR:
2337 		ibwc_status = IB_WC_LOC_QP_OP_ERR;
2338 		break;
2339 	case OCRDMA_CQE_LOC_EEC_OP_ERR:
2340 		ibwc_status = IB_WC_LOC_EEC_OP_ERR;
2341 		break;
2342 	case OCRDMA_CQE_LOC_PROT_ERR:
2343 		ibwc_status = IB_WC_LOC_PROT_ERR;
2344 		break;
2345 	case OCRDMA_CQE_WR_FLUSH_ERR:
2346 		ibwc_status = IB_WC_WR_FLUSH_ERR;
2347 		break;
2348 	case OCRDMA_CQE_MW_BIND_ERR:
2349 		ibwc_status = IB_WC_MW_BIND_ERR;
2350 		break;
2351 	case OCRDMA_CQE_BAD_RESP_ERR:
2352 		ibwc_status = IB_WC_BAD_RESP_ERR;
2353 		break;
2354 	case OCRDMA_CQE_LOC_ACCESS_ERR:
2355 		ibwc_status = IB_WC_LOC_ACCESS_ERR;
2356 		break;
2357 	case OCRDMA_CQE_REM_INV_REQ_ERR:
2358 		ibwc_status = IB_WC_REM_INV_REQ_ERR;
2359 		break;
2360 	case OCRDMA_CQE_REM_ACCESS_ERR:
2361 		ibwc_status = IB_WC_REM_ACCESS_ERR;
2362 		break;
2363 	case OCRDMA_CQE_REM_OP_ERR:
2364 		ibwc_status = IB_WC_REM_OP_ERR;
2365 		break;
2366 	case OCRDMA_CQE_RETRY_EXC_ERR:
2367 		ibwc_status = IB_WC_RETRY_EXC_ERR;
2368 		break;
2369 	case OCRDMA_CQE_RNR_RETRY_EXC_ERR:
2370 		ibwc_status = IB_WC_RNR_RETRY_EXC_ERR;
2371 		break;
2372 	case OCRDMA_CQE_LOC_RDD_VIOL_ERR:
2373 		ibwc_status = IB_WC_LOC_RDD_VIOL_ERR;
2374 		break;
2375 	case OCRDMA_CQE_REM_INV_RD_REQ_ERR:
2376 		ibwc_status = IB_WC_REM_INV_RD_REQ_ERR;
2377 		break;
2378 	case OCRDMA_CQE_REM_ABORT_ERR:
2379 		ibwc_status = IB_WC_REM_ABORT_ERR;
2380 		break;
2381 	case OCRDMA_CQE_INV_EECN_ERR:
2382 		ibwc_status = IB_WC_INV_EECN_ERR;
2383 		break;
2384 	case OCRDMA_CQE_INV_EEC_STATE_ERR:
2385 		ibwc_status = IB_WC_INV_EEC_STATE_ERR;
2386 		break;
2387 	case OCRDMA_CQE_FATAL_ERR:
2388 		ibwc_status = IB_WC_FATAL_ERR;
2389 		break;
2390 	case OCRDMA_CQE_RESP_TIMEOUT_ERR:
2391 		ibwc_status = IB_WC_RESP_TIMEOUT_ERR;
2392 		break;
2393 	default:
2394 		ibwc_status = IB_WC_GENERAL_ERR;
2395 		break;
2396 	}
2397 	return ibwc_status;
2398 }
2399 
2400 static void ocrdma_update_wc(struct ocrdma_qp *qp, struct ib_wc *ibwc,
2401 		      u32 wqe_idx)
2402 {
2403 	struct ocrdma_hdr_wqe *hdr;
2404 	struct ocrdma_sge *rw;
2405 	int opcode;
2406 
2407 	hdr = ocrdma_hwq_head_from_idx(&qp->sq, wqe_idx);
2408 
2409 	ibwc->wr_id = qp->wqe_wr_id_tbl[wqe_idx].wrid;
2410 	/* Undo the hdr->cw swap */
2411 	opcode = le32_to_cpu(hdr->cw) & OCRDMA_WQE_OPCODE_MASK;
2412 	switch (opcode) {
2413 	case OCRDMA_WRITE:
2414 		ibwc->opcode = IB_WC_RDMA_WRITE;
2415 		break;
2416 	case OCRDMA_READ:
2417 		rw = (struct ocrdma_sge *)(hdr + 1);
2418 		ibwc->opcode = IB_WC_RDMA_READ;
2419 		ibwc->byte_len = rw->len;
2420 		break;
2421 	case OCRDMA_SEND:
2422 		ibwc->opcode = IB_WC_SEND;
2423 		break;
2424 	case OCRDMA_FR_MR:
2425 		ibwc->opcode = IB_WC_REG_MR;
2426 		break;
2427 	case OCRDMA_LKEY_INV:
2428 		ibwc->opcode = IB_WC_LOCAL_INV;
2429 		break;
2430 	default:
2431 		ibwc->status = IB_WC_GENERAL_ERR;
2432 		pr_err("%s() invalid opcode received = 0x%x\n",
2433 		       __func__, hdr->cw & OCRDMA_WQE_OPCODE_MASK);
2434 		break;
2435 	}
2436 }
2437 
2438 static void ocrdma_set_cqe_status_flushed(struct ocrdma_qp *qp,
2439 						struct ocrdma_cqe *cqe)
2440 {
2441 	if (is_cqe_for_sq(cqe)) {
2442 		cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2443 				cqe->flags_status_srcqpn) &
2444 					~OCRDMA_CQE_STATUS_MASK);
2445 		cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2446 				cqe->flags_status_srcqpn) |
2447 				(OCRDMA_CQE_WR_FLUSH_ERR <<
2448 					OCRDMA_CQE_STATUS_SHIFT));
2449 	} else {
2450 		if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2451 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2452 					cqe->flags_status_srcqpn) &
2453 						~OCRDMA_CQE_UD_STATUS_MASK);
2454 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2455 					cqe->flags_status_srcqpn) |
2456 					(OCRDMA_CQE_WR_FLUSH_ERR <<
2457 						OCRDMA_CQE_UD_STATUS_SHIFT));
2458 		} else {
2459 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2460 					cqe->flags_status_srcqpn) &
2461 						~OCRDMA_CQE_STATUS_MASK);
2462 			cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2463 					cqe->flags_status_srcqpn) |
2464 					(OCRDMA_CQE_WR_FLUSH_ERR <<
2465 						OCRDMA_CQE_STATUS_SHIFT));
2466 		}
2467 	}
2468 }
2469 
2470 static bool ocrdma_update_err_cqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2471 				  struct ocrdma_qp *qp, int status)
2472 {
2473 	bool expand = false;
2474 
2475 	ibwc->byte_len = 0;
2476 	ibwc->qp = &qp->ibqp;
2477 	ibwc->status = ocrdma_to_ibwc_err(status);
2478 
2479 	ocrdma_flush_qp(qp);
2480 	ocrdma_qp_state_change(qp, IB_QPS_ERR, NULL);
2481 
2482 	/* if wqe/rqe pending for which cqe needs to be returned,
2483 	 * trigger inflating it.
2484 	 */
2485 	if (!is_hw_rq_empty(qp) || !is_hw_sq_empty(qp)) {
2486 		expand = true;
2487 		ocrdma_set_cqe_status_flushed(qp, cqe);
2488 	}
2489 	return expand;
2490 }
2491 
2492 static int ocrdma_update_err_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2493 				  struct ocrdma_qp *qp, int status)
2494 {
2495 	ibwc->opcode = IB_WC_RECV;
2496 	ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2497 	ocrdma_hwq_inc_tail(&qp->rq);
2498 
2499 	return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2500 }
2501 
2502 static int ocrdma_update_err_scqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2503 				  struct ocrdma_qp *qp, int status)
2504 {
2505 	ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2506 	ocrdma_hwq_inc_tail(&qp->sq);
2507 
2508 	return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2509 }
2510 
2511 
2512 static bool ocrdma_poll_err_scqe(struct ocrdma_qp *qp,
2513 				 struct ocrdma_cqe *cqe, struct ib_wc *ibwc,
2514 				 bool *polled, bool *stop)
2515 {
2516 	bool expand;
2517 	struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2518 	int status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2519 		OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2520 	if (status < OCRDMA_MAX_CQE_ERR)
2521 		atomic_inc(&dev->cqe_err_stats[status]);
2522 
2523 	/* when hw sq is empty, but rq is not empty, so we continue
2524 	 * to keep the cqe in order to get the cq event again.
2525 	 */
2526 	if (is_hw_sq_empty(qp) && !is_hw_rq_empty(qp)) {
2527 		/* when cq for rq and sq is same, it is safe to return
2528 		 * flush cqe for RQEs.
2529 		 */
2530 		if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2531 			*polled = true;
2532 			status = OCRDMA_CQE_WR_FLUSH_ERR;
2533 			expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2534 		} else {
2535 			/* stop processing further cqe as this cqe is used for
2536 			 * triggering cq event on buddy cq of RQ.
2537 			 * When QP is destroyed, this cqe will be removed
2538 			 * from the cq's hardware q.
2539 			 */
2540 			*polled = false;
2541 			*stop = true;
2542 			expand = false;
2543 		}
2544 	} else if (is_hw_sq_empty(qp)) {
2545 		/* Do nothing */
2546 		expand = false;
2547 		*polled = false;
2548 		*stop = false;
2549 	} else {
2550 		*polled = true;
2551 		expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2552 	}
2553 	return expand;
2554 }
2555 
2556 static bool ocrdma_poll_success_scqe(struct ocrdma_qp *qp,
2557 				     struct ocrdma_cqe *cqe,
2558 				     struct ib_wc *ibwc, bool *polled)
2559 {
2560 	bool expand = false;
2561 	int tail = qp->sq.tail;
2562 	u32 wqe_idx;
2563 
2564 	if (!qp->wqe_wr_id_tbl[tail].signaled) {
2565 		*polled = false;    /* WC cannot be consumed yet */
2566 	} else {
2567 		ibwc->status = IB_WC_SUCCESS;
2568 		ibwc->wc_flags = 0;
2569 		ibwc->qp = &qp->ibqp;
2570 		ocrdma_update_wc(qp, ibwc, tail);
2571 		*polled = true;
2572 	}
2573 	wqe_idx = (le32_to_cpu(cqe->wq.wqeidx) &
2574 			OCRDMA_CQE_WQEIDX_MASK) & qp->sq.max_wqe_idx;
2575 	if (tail != wqe_idx)
2576 		expand = true; /* Coalesced CQE can't be consumed yet */
2577 
2578 	ocrdma_hwq_inc_tail(&qp->sq);
2579 	return expand;
2580 }
2581 
2582 static bool ocrdma_poll_scqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2583 			     struct ib_wc *ibwc, bool *polled, bool *stop)
2584 {
2585 	int status;
2586 	bool expand;
2587 
2588 	status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2589 		OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2590 
2591 	if (status == OCRDMA_CQE_SUCCESS)
2592 		expand = ocrdma_poll_success_scqe(qp, cqe, ibwc, polled);
2593 	else
2594 		expand = ocrdma_poll_err_scqe(qp, cqe, ibwc, polled, stop);
2595 	return expand;
2596 }
2597 
2598 static int ocrdma_update_ud_rcqe(struct ocrdma_dev *dev, struct ib_wc *ibwc,
2599 				 struct ocrdma_cqe *cqe)
2600 {
2601 	int status;
2602 	u16 hdr_type = 0;
2603 
2604 	status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2605 		OCRDMA_CQE_UD_STATUS_MASK) >> OCRDMA_CQE_UD_STATUS_SHIFT;
2606 	ibwc->src_qp = le32_to_cpu(cqe->flags_status_srcqpn) &
2607 						OCRDMA_CQE_SRCQP_MASK;
2608 	ibwc->pkey_index = 0;
2609 	ibwc->wc_flags = IB_WC_GRH;
2610 	ibwc->byte_len = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2611 			  OCRDMA_CQE_UD_XFER_LEN_SHIFT) &
2612 			  OCRDMA_CQE_UD_XFER_LEN_MASK;
2613 
2614 	if (ocrdma_is_udp_encap_supported(dev)) {
2615 		hdr_type = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2616 			    OCRDMA_CQE_UD_L3TYPE_SHIFT) &
2617 			    OCRDMA_CQE_UD_L3TYPE_MASK;
2618 		ibwc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
2619 		ibwc->network_hdr_type = hdr_type;
2620 	}
2621 
2622 	return status;
2623 }
2624 
2625 static void ocrdma_update_free_srq_cqe(struct ib_wc *ibwc,
2626 				       struct ocrdma_cqe *cqe,
2627 				       struct ocrdma_qp *qp)
2628 {
2629 	unsigned long flags;
2630 	struct ocrdma_srq *srq;
2631 	u32 wqe_idx;
2632 
2633 	srq = get_ocrdma_srq(qp->ibqp.srq);
2634 	wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
2635 		OCRDMA_CQE_BUFTAG_SHIFT) & srq->rq.max_wqe_idx;
2636 	BUG_ON(wqe_idx < 1);
2637 
2638 	ibwc->wr_id = srq->rqe_wr_id_tbl[wqe_idx];
2639 	spin_lock_irqsave(&srq->q_lock, flags);
2640 	ocrdma_srq_toggle_bit(srq, wqe_idx - 1);
2641 	spin_unlock_irqrestore(&srq->q_lock, flags);
2642 	ocrdma_hwq_inc_tail(&srq->rq);
2643 }
2644 
2645 static bool ocrdma_poll_err_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2646 				struct ib_wc *ibwc, bool *polled, bool *stop,
2647 				int status)
2648 {
2649 	bool expand;
2650 	struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2651 
2652 	if (status < OCRDMA_MAX_CQE_ERR)
2653 		atomic_inc(&dev->cqe_err_stats[status]);
2654 
2655 	/* when hw_rq is empty, but wq is not empty, so continue
2656 	 * to keep the cqe to get the cq event again.
2657 	 */
2658 	if (is_hw_rq_empty(qp) && !is_hw_sq_empty(qp)) {
2659 		if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2660 			*polled = true;
2661 			status = OCRDMA_CQE_WR_FLUSH_ERR;
2662 			expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2663 		} else {
2664 			*polled = false;
2665 			*stop = true;
2666 			expand = false;
2667 		}
2668 	} else if (is_hw_rq_empty(qp)) {
2669 		/* Do nothing */
2670 		expand = false;
2671 		*polled = false;
2672 		*stop = false;
2673 	} else {
2674 		*polled = true;
2675 		expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2676 	}
2677 	return expand;
2678 }
2679 
2680 static void ocrdma_poll_success_rcqe(struct ocrdma_qp *qp,
2681 				     struct ocrdma_cqe *cqe, struct ib_wc *ibwc)
2682 {
2683 	struct ocrdma_dev *dev;
2684 
2685 	dev = get_ocrdma_dev(qp->ibqp.device);
2686 	ibwc->opcode = IB_WC_RECV;
2687 	ibwc->qp = &qp->ibqp;
2688 	ibwc->status = IB_WC_SUCCESS;
2689 
2690 	if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI)
2691 		ocrdma_update_ud_rcqe(dev, ibwc, cqe);
2692 	else
2693 		ibwc->byte_len = le32_to_cpu(cqe->rq.rxlen);
2694 
2695 	if (is_cqe_imm(cqe)) {
2696 		ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2697 		ibwc->wc_flags |= IB_WC_WITH_IMM;
2698 	} else if (is_cqe_wr_imm(cqe)) {
2699 		ibwc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
2700 		ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2701 		ibwc->wc_flags |= IB_WC_WITH_IMM;
2702 	} else if (is_cqe_invalidated(cqe)) {
2703 		ibwc->ex.invalidate_rkey = le32_to_cpu(cqe->rq.lkey_immdt);
2704 		ibwc->wc_flags |= IB_WC_WITH_INVALIDATE;
2705 	}
2706 	if (qp->ibqp.srq) {
2707 		ocrdma_update_free_srq_cqe(ibwc, cqe, qp);
2708 	} else {
2709 		ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2710 		ocrdma_hwq_inc_tail(&qp->rq);
2711 	}
2712 }
2713 
2714 static bool ocrdma_poll_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2715 			     struct ib_wc *ibwc, bool *polled, bool *stop)
2716 {
2717 	int status;
2718 	bool expand = false;
2719 
2720 	ibwc->wc_flags = 0;
2721 	if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2722 		status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2723 					OCRDMA_CQE_UD_STATUS_MASK) >>
2724 					OCRDMA_CQE_UD_STATUS_SHIFT;
2725 	} else {
2726 		status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2727 			     OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2728 	}
2729 
2730 	if (status == OCRDMA_CQE_SUCCESS) {
2731 		*polled = true;
2732 		ocrdma_poll_success_rcqe(qp, cqe, ibwc);
2733 	} else {
2734 		expand = ocrdma_poll_err_rcqe(qp, cqe, ibwc, polled, stop,
2735 					      status);
2736 	}
2737 	return expand;
2738 }
2739 
2740 static void ocrdma_change_cq_phase(struct ocrdma_cq *cq, struct ocrdma_cqe *cqe,
2741 				   u16 cur_getp)
2742 {
2743 	if (cq->phase_change) {
2744 		if (cur_getp == 0)
2745 			cq->phase = (~cq->phase & OCRDMA_CQE_VALID);
2746 	} else {
2747 		/* clear valid bit */
2748 		cqe->flags_status_srcqpn = 0;
2749 	}
2750 }
2751 
2752 static int ocrdma_poll_hwcq(struct ocrdma_cq *cq, int num_entries,
2753 			    struct ib_wc *ibwc)
2754 {
2755 	u16 qpn = 0;
2756 	int i = 0;
2757 	bool expand = false;
2758 	int polled_hw_cqes = 0;
2759 	struct ocrdma_qp *qp = NULL;
2760 	struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
2761 	struct ocrdma_cqe *cqe;
2762 	u16 cur_getp; bool polled = false; bool stop = false;
2763 
2764 	cur_getp = cq->getp;
2765 	while (num_entries) {
2766 		cqe = cq->va + cur_getp;
2767 		/* check whether valid cqe or not */
2768 		if (!is_cqe_valid(cq, cqe))
2769 			break;
2770 		qpn = (le32_to_cpu(cqe->cmn.qpn) & OCRDMA_CQE_QPN_MASK);
2771 		/* ignore discarded cqe */
2772 		if (qpn == 0)
2773 			goto skip_cqe;
2774 		qp = dev->qp_tbl[qpn];
2775 		BUG_ON(qp == NULL);
2776 
2777 		if (is_cqe_for_sq(cqe)) {
2778 			expand = ocrdma_poll_scqe(qp, cqe, ibwc, &polled,
2779 						  &stop);
2780 		} else {
2781 			expand = ocrdma_poll_rcqe(qp, cqe, ibwc, &polled,
2782 						  &stop);
2783 		}
2784 		if (expand)
2785 			goto expand_cqe;
2786 		if (stop)
2787 			goto stop_cqe;
2788 		/* clear qpn to avoid duplicate processing by discard_cqe() */
2789 		cqe->cmn.qpn = 0;
2790 skip_cqe:
2791 		polled_hw_cqes += 1;
2792 		cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
2793 		ocrdma_change_cq_phase(cq, cqe, cur_getp);
2794 expand_cqe:
2795 		if (polled) {
2796 			num_entries -= 1;
2797 			i += 1;
2798 			ibwc = ibwc + 1;
2799 			polled = false;
2800 		}
2801 	}
2802 stop_cqe:
2803 	cq->getp = cur_getp;
2804 
2805 	if (polled_hw_cqes)
2806 		ocrdma_ring_cq_db(dev, cq->id, false, false, polled_hw_cqes);
2807 
2808 	return i;
2809 }
2810 
2811 /* insert error cqe if the QP's SQ or RQ's CQ matches the CQ under poll. */
2812 static int ocrdma_add_err_cqe(struct ocrdma_cq *cq, int num_entries,
2813 			      struct ocrdma_qp *qp, struct ib_wc *ibwc)
2814 {
2815 	int err_cqes = 0;
2816 
2817 	while (num_entries) {
2818 		if (is_hw_sq_empty(qp) && is_hw_rq_empty(qp))
2819 			break;
2820 		if (!is_hw_sq_empty(qp) && qp->sq_cq == cq) {
2821 			ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2822 			ocrdma_hwq_inc_tail(&qp->sq);
2823 		} else if (!is_hw_rq_empty(qp) && qp->rq_cq == cq) {
2824 			ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2825 			ocrdma_hwq_inc_tail(&qp->rq);
2826 		} else {
2827 			return err_cqes;
2828 		}
2829 		ibwc->byte_len = 0;
2830 		ibwc->status = IB_WC_WR_FLUSH_ERR;
2831 		ibwc = ibwc + 1;
2832 		err_cqes += 1;
2833 		num_entries -= 1;
2834 	}
2835 	return err_cqes;
2836 }
2837 
2838 int ocrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
2839 {
2840 	int cqes_to_poll = num_entries;
2841 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2842 	struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2843 	int num_os_cqe = 0, err_cqes = 0;
2844 	struct ocrdma_qp *qp;
2845 	unsigned long flags;
2846 
2847 	/* poll cqes from adapter CQ */
2848 	spin_lock_irqsave(&cq->cq_lock, flags);
2849 	num_os_cqe = ocrdma_poll_hwcq(cq, cqes_to_poll, wc);
2850 	spin_unlock_irqrestore(&cq->cq_lock, flags);
2851 	cqes_to_poll -= num_os_cqe;
2852 
2853 	if (cqes_to_poll) {
2854 		wc = wc + num_os_cqe;
2855 		/* adapter returns single error cqe when qp moves to
2856 		 * error state. So insert error cqes with wc_status as
2857 		 * FLUSHED for pending WQEs and RQEs of QP's SQ and RQ
2858 		 * respectively which uses this CQ.
2859 		 */
2860 		spin_lock_irqsave(&dev->flush_q_lock, flags);
2861 		list_for_each_entry(qp, &cq->sq_head, sq_entry) {
2862 			if (cqes_to_poll == 0)
2863 				break;
2864 			err_cqes = ocrdma_add_err_cqe(cq, cqes_to_poll, qp, wc);
2865 			cqes_to_poll -= err_cqes;
2866 			num_os_cqe += err_cqes;
2867 			wc = wc + err_cqes;
2868 		}
2869 		spin_unlock_irqrestore(&dev->flush_q_lock, flags);
2870 	}
2871 	return num_os_cqe;
2872 }
2873 
2874 int ocrdma_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags cq_flags)
2875 {
2876 	struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2877 	struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2878 	u16 cq_id;
2879 	unsigned long flags;
2880 	bool arm_needed = false, sol_needed = false;
2881 
2882 	cq_id = cq->id;
2883 
2884 	spin_lock_irqsave(&cq->cq_lock, flags);
2885 	if (cq_flags & IB_CQ_NEXT_COMP || cq_flags & IB_CQ_SOLICITED)
2886 		arm_needed = true;
2887 	if (cq_flags & IB_CQ_SOLICITED)
2888 		sol_needed = true;
2889 
2890 	ocrdma_ring_cq_db(dev, cq_id, arm_needed, sol_needed, 0);
2891 	spin_unlock_irqrestore(&cq->cq_lock, flags);
2892 
2893 	return 0;
2894 }
2895 
2896 struct ib_mr *ocrdma_alloc_mr(struct ib_pd *ibpd, enum ib_mr_type mr_type,
2897 			      u32 max_num_sg)
2898 {
2899 	int status;
2900 	struct ocrdma_mr *mr;
2901 	struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
2902 	struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
2903 
2904 	if (mr_type != IB_MR_TYPE_MEM_REG)
2905 		return ERR_PTR(-EINVAL);
2906 
2907 	if (max_num_sg > dev->attr.max_pages_per_frmr)
2908 		return ERR_PTR(-EINVAL);
2909 
2910 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
2911 	if (!mr)
2912 		return ERR_PTR(-ENOMEM);
2913 
2914 	mr->pages = kcalloc(max_num_sg, sizeof(u64), GFP_KERNEL);
2915 	if (!mr->pages) {
2916 		status = -ENOMEM;
2917 		goto pl_err;
2918 	}
2919 
2920 	status = ocrdma_get_pbl_info(dev, mr, max_num_sg);
2921 	if (status)
2922 		goto pbl_err;
2923 	mr->hwmr.fr_mr = 1;
2924 	mr->hwmr.remote_rd = 0;
2925 	mr->hwmr.remote_wr = 0;
2926 	mr->hwmr.local_rd = 0;
2927 	mr->hwmr.local_wr = 0;
2928 	mr->hwmr.mw_bind = 0;
2929 	status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
2930 	if (status)
2931 		goto pbl_err;
2932 	status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, 0);
2933 	if (status)
2934 		goto mbx_err;
2935 	mr->ibmr.rkey = mr->hwmr.lkey;
2936 	mr->ibmr.lkey = mr->hwmr.lkey;
2937 	dev->stag_arr[(mr->hwmr.lkey >> 8) & (OCRDMA_MAX_STAG - 1)] =
2938 		(unsigned long) mr;
2939 	return &mr->ibmr;
2940 mbx_err:
2941 	ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
2942 pbl_err:
2943 	kfree(mr->pages);
2944 pl_err:
2945 	kfree(mr);
2946 	return ERR_PTR(-ENOMEM);
2947 }
2948 
2949 static int ocrdma_set_page(struct ib_mr *ibmr, u64 addr)
2950 {
2951 	struct ocrdma_mr *mr = get_ocrdma_mr(ibmr);
2952 
2953 	if (unlikely(mr->npages == mr->hwmr.num_pbes))
2954 		return -ENOMEM;
2955 
2956 	mr->pages[mr->npages++] = addr;
2957 
2958 	return 0;
2959 }
2960 
2961 int ocrdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
2962 		     unsigned int *sg_offset)
2963 {
2964 	struct ocrdma_mr *mr = get_ocrdma_mr(ibmr);
2965 
2966 	mr->npages = 0;
2967 
2968 	return ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, ocrdma_set_page);
2969 }
2970