xref: /linux/drivers/infiniband/hw/cxgb4/qp.c (revision c7546e2c3cb739a3c1a2f5acaf9bb629d401afe5)
1 /*
2  * Copyright (c) 2009-2010 Chelsio, Inc. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <linux/module.h>
34 #include <rdma/uverbs_ioctl.h>
35 
36 #include "iw_cxgb4.h"
37 
38 static int db_delay_usecs = 1;
39 module_param(db_delay_usecs, int, 0644);
40 MODULE_PARM_DESC(db_delay_usecs, "Usecs to delay awaiting db fifo to drain");
41 
42 static int ocqp_support = 1;
43 module_param(ocqp_support, int, 0644);
44 MODULE_PARM_DESC(ocqp_support, "Support on-chip SQs (default=1)");
45 
46 int db_fc_threshold = 1000;
47 module_param(db_fc_threshold, int, 0644);
48 MODULE_PARM_DESC(db_fc_threshold,
49 		 "QP count/threshold that triggers"
50 		 " automatic db flow control mode (default = 1000)");
51 
52 int db_coalescing_threshold;
53 module_param(db_coalescing_threshold, int, 0644);
54 MODULE_PARM_DESC(db_coalescing_threshold,
55 		 "QP count/threshold that triggers"
56 		 " disabling db coalescing (default = 0)");
57 
58 static int max_fr_immd = T4_MAX_FR_IMMD;
59 module_param(max_fr_immd, int, 0644);
60 MODULE_PARM_DESC(max_fr_immd, "fastreg threshold for using DSGL instead of immediate");
61 
62 static int alloc_ird(struct c4iw_dev *dev, u32 ird)
63 {
64 	int ret = 0;
65 
66 	xa_lock_irq(&dev->qps);
67 	if (ird <= dev->avail_ird)
68 		dev->avail_ird -= ird;
69 	else
70 		ret = -ENOMEM;
71 	xa_unlock_irq(&dev->qps);
72 
73 	if (ret)
74 		dev_warn(&dev->rdev.lldi.pdev->dev,
75 			 "device IRD resources exhausted\n");
76 
77 	return ret;
78 }
79 
80 static void free_ird(struct c4iw_dev *dev, int ird)
81 {
82 	xa_lock_irq(&dev->qps);
83 	dev->avail_ird += ird;
84 	xa_unlock_irq(&dev->qps);
85 }
86 
87 static void set_state(struct c4iw_qp *qhp, enum c4iw_qp_state state)
88 {
89 	unsigned long flag;
90 	spin_lock_irqsave(&qhp->lock, flag);
91 	qhp->attr.state = state;
92 	spin_unlock_irqrestore(&qhp->lock, flag);
93 }
94 
95 static void dealloc_oc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
96 {
97 	c4iw_ocqp_pool_free(rdev, sq->dma_addr, sq->memsize);
98 }
99 
100 static void dealloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
101 {
102 	dma_free_coherent(&(rdev->lldi.pdev->dev), sq->memsize, sq->queue,
103 			  dma_unmap_addr(sq, mapping));
104 }
105 
106 static void dealloc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
107 {
108 	if (t4_sq_onchip(sq))
109 		dealloc_oc_sq(rdev, sq);
110 	else
111 		dealloc_host_sq(rdev, sq);
112 }
113 
114 static int alloc_oc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
115 {
116 	if (!ocqp_support || !ocqp_supported(&rdev->lldi))
117 		return -ENOSYS;
118 	sq->dma_addr = c4iw_ocqp_pool_alloc(rdev, sq->memsize);
119 	if (!sq->dma_addr)
120 		return -ENOMEM;
121 	sq->phys_addr = rdev->oc_mw_pa + sq->dma_addr -
122 			rdev->lldi.vr->ocq.start;
123 	sq->queue = (__force union t4_wr *)(rdev->oc_mw_kva + sq->dma_addr -
124 					    rdev->lldi.vr->ocq.start);
125 	sq->flags |= T4_SQ_ONCHIP;
126 	return 0;
127 }
128 
129 static int alloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
130 {
131 	sq->queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev), sq->memsize,
132 				       &(sq->dma_addr), GFP_KERNEL);
133 	if (!sq->queue)
134 		return -ENOMEM;
135 	sq->phys_addr = virt_to_phys(sq->queue);
136 	dma_unmap_addr_set(sq, mapping, sq->dma_addr);
137 	return 0;
138 }
139 
140 static int alloc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq, int user)
141 {
142 	int ret = -ENOSYS;
143 	if (user)
144 		ret = alloc_oc_sq(rdev, sq);
145 	if (ret)
146 		ret = alloc_host_sq(rdev, sq);
147 	return ret;
148 }
149 
150 static int destroy_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
151 		      struct c4iw_dev_ucontext *uctx, int has_rq)
152 {
153 	/*
154 	 * uP clears EQ contexts when the connection exits rdma mode,
155 	 * so no need to post a RESET WR for these EQs.
156 	 */
157 	dealloc_sq(rdev, &wq->sq);
158 	kfree(wq->sq.sw_sq);
159 	c4iw_put_qpid(rdev, wq->sq.qid, uctx);
160 
161 	if (has_rq) {
162 		dma_free_coherent(&rdev->lldi.pdev->dev,
163 				  wq->rq.memsize, wq->rq.queue,
164 				  dma_unmap_addr(&wq->rq, mapping));
165 		c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
166 		kfree(wq->rq.sw_rq);
167 		c4iw_put_qpid(rdev, wq->rq.qid, uctx);
168 	}
169 	return 0;
170 }
171 
172 /*
173  * Determine the BAR2 virtual address and qid. If pbar2_pa is not NULL,
174  * then this is a user mapping so compute the page-aligned physical address
175  * for mapping.
176  */
177 void __iomem *c4iw_bar2_addrs(struct c4iw_rdev *rdev, unsigned int qid,
178 			      enum cxgb4_bar2_qtype qtype,
179 			      unsigned int *pbar2_qid, u64 *pbar2_pa)
180 {
181 	u64 bar2_qoffset;
182 	int ret;
183 
184 	ret = cxgb4_bar2_sge_qregs(rdev->lldi.ports[0], qid, qtype,
185 				   pbar2_pa ? 1 : 0,
186 				   &bar2_qoffset, pbar2_qid);
187 	if (ret)
188 		return NULL;
189 
190 	if (pbar2_pa)
191 		*pbar2_pa = (rdev->bar2_pa + bar2_qoffset) & PAGE_MASK;
192 
193 	if (is_t4(rdev->lldi.adapter_type))
194 		return NULL;
195 
196 	return rdev->bar2_kva + bar2_qoffset;
197 }
198 
199 static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
200 		     struct t4_cq *rcq, struct t4_cq *scq,
201 		     struct c4iw_dev_ucontext *uctx,
202 		     struct c4iw_wr_wait *wr_waitp,
203 		     int need_rq)
204 {
205 	int user = (uctx != &rdev->uctx);
206 	struct fw_ri_res_wr *res_wr;
207 	struct fw_ri_res *res;
208 	int wr_len;
209 	struct sk_buff *skb;
210 	int ret = 0;
211 	int eqsize;
212 
213 	wq->sq.qid = c4iw_get_qpid(rdev, uctx);
214 	if (!wq->sq.qid)
215 		return -ENOMEM;
216 
217 	if (need_rq) {
218 		wq->rq.qid = c4iw_get_qpid(rdev, uctx);
219 		if (!wq->rq.qid) {
220 			ret = -ENOMEM;
221 			goto free_sq_qid;
222 		}
223 	}
224 
225 	if (!user) {
226 		wq->sq.sw_sq = kcalloc(wq->sq.size, sizeof(*wq->sq.sw_sq),
227 				       GFP_KERNEL);
228 		if (!wq->sq.sw_sq) {
229 			ret = -ENOMEM;
230 			goto free_rq_qid;//FIXME
231 		}
232 
233 		if (need_rq) {
234 			wq->rq.sw_rq = kcalloc(wq->rq.size,
235 					       sizeof(*wq->rq.sw_rq),
236 					       GFP_KERNEL);
237 			if (!wq->rq.sw_rq) {
238 				ret = -ENOMEM;
239 				goto free_sw_sq;
240 			}
241 		}
242 	}
243 
244 	if (need_rq) {
245 		/*
246 		 * RQT must be a power of 2 and at least 16 deep.
247 		 */
248 		wq->rq.rqt_size =
249 			roundup_pow_of_two(max_t(u16, wq->rq.size, 16));
250 		wq->rq.rqt_hwaddr = c4iw_rqtpool_alloc(rdev, wq->rq.rqt_size);
251 		if (!wq->rq.rqt_hwaddr) {
252 			ret = -ENOMEM;
253 			goto free_sw_rq;
254 		}
255 	}
256 
257 	ret = alloc_sq(rdev, &wq->sq, user);
258 	if (ret)
259 		goto free_hwaddr;
260 	memset(wq->sq.queue, 0, wq->sq.memsize);
261 	dma_unmap_addr_set(&wq->sq, mapping, wq->sq.dma_addr);
262 
263 	if (need_rq) {
264 		wq->rq.queue = dma_alloc_coherent(&rdev->lldi.pdev->dev,
265 						  wq->rq.memsize,
266 						  &wq->rq.dma_addr,
267 						  GFP_KERNEL);
268 		if (!wq->rq.queue) {
269 			ret = -ENOMEM;
270 			goto free_sq;
271 		}
272 		pr_debug("sq base va 0x%p pa 0x%llx rq base va 0x%p pa 0x%llx\n",
273 			 wq->sq.queue,
274 			 (unsigned long long)virt_to_phys(wq->sq.queue),
275 			 wq->rq.queue,
276 			 (unsigned long long)virt_to_phys(wq->rq.queue));
277 		dma_unmap_addr_set(&wq->rq, mapping, wq->rq.dma_addr);
278 	}
279 
280 	wq->db = rdev->lldi.db_reg;
281 
282 	wq->sq.bar2_va = c4iw_bar2_addrs(rdev, wq->sq.qid,
283 					 CXGB4_BAR2_QTYPE_EGRESS,
284 					 &wq->sq.bar2_qid,
285 					 user ? &wq->sq.bar2_pa : NULL);
286 	if (need_rq)
287 		wq->rq.bar2_va = c4iw_bar2_addrs(rdev, wq->rq.qid,
288 						 CXGB4_BAR2_QTYPE_EGRESS,
289 						 &wq->rq.bar2_qid,
290 						 user ? &wq->rq.bar2_pa : NULL);
291 
292 	/*
293 	 * User mode must have bar2 access.
294 	 */
295 	if (user && (!wq->sq.bar2_pa || (need_rq && !wq->rq.bar2_pa))) {
296 		pr_warn("%s: sqid %u or rqid %u not in BAR2 range\n",
297 			pci_name(rdev->lldi.pdev), wq->sq.qid, wq->rq.qid);
298 		ret = -EINVAL;
299 		goto free_dma;
300 	}
301 
302 	wq->rdev = rdev;
303 	wq->rq.msn = 1;
304 
305 	/* build fw_ri_res_wr */
306 	wr_len = sizeof(*res_wr) + 2 * sizeof(*res);
307 	if (need_rq)
308 		wr_len += sizeof(*res);
309 	skb = alloc_skb(wr_len, GFP_KERNEL);
310 	if (!skb) {
311 		ret = -ENOMEM;
312 		goto free_dma;
313 	}
314 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
315 
316 	res_wr = __skb_put_zero(skb, wr_len);
317 	res_wr->op_nres = cpu_to_be32(
318 			FW_WR_OP_V(FW_RI_RES_WR) |
319 			FW_RI_RES_WR_NRES_V(need_rq ? 2 : 1) |
320 			FW_WR_COMPL_F);
321 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
322 	res_wr->cookie = (uintptr_t)wr_waitp;
323 	res = res_wr->res;
324 	res->u.sqrq.restype = FW_RI_RES_TYPE_SQ;
325 	res->u.sqrq.op = FW_RI_RES_OP_WRITE;
326 
327 	/*
328 	 * eqsize is the number of 64B entries plus the status page size.
329 	 */
330 	eqsize = wq->sq.size * T4_SQ_NUM_SLOTS +
331 		rdev->hw_queue.t4_eq_status_entries;
332 
333 	res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
334 		FW_RI_RES_WR_HOSTFCMODE_V(0) |	/* no host cidx updates */
335 		FW_RI_RES_WR_CPRIO_V(0) |	/* don't keep in chip cache */
336 		FW_RI_RES_WR_PCIECHN_V(0) |	/* set by uP at ri_init time */
337 		(t4_sq_onchip(&wq->sq) ? FW_RI_RES_WR_ONCHIP_F : 0) |
338 		FW_RI_RES_WR_IQID_V(scq->cqid));
339 	res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
340 		FW_RI_RES_WR_DCAEN_V(0) |
341 		FW_RI_RES_WR_DCACPU_V(0) |
342 		FW_RI_RES_WR_FBMIN_V(2) |
343 		(t4_sq_onchip(&wq->sq) ? FW_RI_RES_WR_FBMAX_V(2) :
344 					 FW_RI_RES_WR_FBMAX_V(3)) |
345 		FW_RI_RES_WR_CIDXFTHRESHO_V(0) |
346 		FW_RI_RES_WR_CIDXFTHRESH_V(0) |
347 		FW_RI_RES_WR_EQSIZE_V(eqsize));
348 	res->u.sqrq.eqid = cpu_to_be32(wq->sq.qid);
349 	res->u.sqrq.eqaddr = cpu_to_be64(wq->sq.dma_addr);
350 
351 	if (need_rq) {
352 		res++;
353 		res->u.sqrq.restype = FW_RI_RES_TYPE_RQ;
354 		res->u.sqrq.op = FW_RI_RES_OP_WRITE;
355 
356 		/*
357 		 * eqsize is the number of 64B entries plus the status page size
358 		 */
359 		eqsize = wq->rq.size * T4_RQ_NUM_SLOTS +
360 			rdev->hw_queue.t4_eq_status_entries;
361 		res->u.sqrq.fetchszm_to_iqid =
362 			/* no host cidx updates */
363 			cpu_to_be32(FW_RI_RES_WR_HOSTFCMODE_V(0) |
364 			/* don't keep in chip cache */
365 			FW_RI_RES_WR_CPRIO_V(0) |
366 			/* set by uP at ri_init time */
367 			FW_RI_RES_WR_PCIECHN_V(0) |
368 			FW_RI_RES_WR_IQID_V(rcq->cqid));
369 		res->u.sqrq.dcaen_to_eqsize =
370 			cpu_to_be32(FW_RI_RES_WR_DCAEN_V(0) |
371 			FW_RI_RES_WR_DCACPU_V(0) |
372 			FW_RI_RES_WR_FBMIN_V(2) |
373 			FW_RI_RES_WR_FBMAX_V(3) |
374 			FW_RI_RES_WR_CIDXFTHRESHO_V(0) |
375 			FW_RI_RES_WR_CIDXFTHRESH_V(0) |
376 			FW_RI_RES_WR_EQSIZE_V(eqsize));
377 		res->u.sqrq.eqid = cpu_to_be32(wq->rq.qid);
378 		res->u.sqrq.eqaddr = cpu_to_be64(wq->rq.dma_addr);
379 	}
380 
381 	c4iw_init_wr_wait(wr_waitp);
382 	ret = c4iw_ref_send_wait(rdev, skb, wr_waitp, 0, wq->sq.qid, __func__);
383 	if (ret)
384 		goto free_dma;
385 
386 	pr_debug("sqid 0x%x rqid 0x%x kdb 0x%p sq_bar2_addr %p rq_bar2_addr %p\n",
387 		 wq->sq.qid, wq->rq.qid, wq->db,
388 		 wq->sq.bar2_va, wq->rq.bar2_va);
389 
390 	return 0;
391 free_dma:
392 	if (need_rq)
393 		dma_free_coherent(&rdev->lldi.pdev->dev,
394 				  wq->rq.memsize, wq->rq.queue,
395 				  dma_unmap_addr(&wq->rq, mapping));
396 free_sq:
397 	dealloc_sq(rdev, &wq->sq);
398 free_hwaddr:
399 	if (need_rq)
400 		c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
401 free_sw_rq:
402 	if (need_rq)
403 		kfree(wq->rq.sw_rq);
404 free_sw_sq:
405 	kfree(wq->sq.sw_sq);
406 free_rq_qid:
407 	if (need_rq)
408 		c4iw_put_qpid(rdev, wq->rq.qid, uctx);
409 free_sq_qid:
410 	c4iw_put_qpid(rdev, wq->sq.qid, uctx);
411 	return ret;
412 }
413 
414 static int build_immd(struct t4_sq *sq, struct fw_ri_immd *immdp,
415 		      const struct ib_send_wr *wr, int max, u32 *plenp)
416 {
417 	u8 *dstp, *srcp;
418 	u32 plen = 0;
419 	int i;
420 	int rem, len;
421 
422 	dstp = (u8 *)immdp->data;
423 	for (i = 0; i < wr->num_sge; i++) {
424 		if ((plen + wr->sg_list[i].length) > max)
425 			return -EMSGSIZE;
426 		srcp = (u8 *)(unsigned long)wr->sg_list[i].addr;
427 		plen += wr->sg_list[i].length;
428 		rem = wr->sg_list[i].length;
429 		while (rem) {
430 			if (dstp == (u8 *)&sq->queue[sq->size])
431 				dstp = (u8 *)sq->queue;
432 			if (rem <= (u8 *)&sq->queue[sq->size] - dstp)
433 				len = rem;
434 			else
435 				len = (u8 *)&sq->queue[sq->size] - dstp;
436 			memcpy(dstp, srcp, len);
437 			dstp += len;
438 			srcp += len;
439 			rem -= len;
440 		}
441 	}
442 	len = roundup(plen + sizeof(*immdp), 16) - (plen + sizeof(*immdp));
443 	if (len)
444 		memset(dstp, 0, len);
445 	immdp->op = FW_RI_DATA_IMMD;
446 	immdp->r1 = 0;
447 	immdp->r2 = 0;
448 	immdp->immdlen = cpu_to_be32(plen);
449 	*plenp = plen;
450 	return 0;
451 }
452 
453 static int build_isgl(__be64 *queue_start, __be64 *queue_end,
454 		      struct fw_ri_isgl *isglp, struct ib_sge *sg_list,
455 		      int num_sge, u32 *plenp)
456 
457 {
458 	int i;
459 	u32 plen = 0;
460 	__be64 *flitp;
461 
462 	if ((__be64 *)isglp == queue_end)
463 		isglp = (struct fw_ri_isgl *)queue_start;
464 
465 	flitp = (__be64 *)isglp->sge;
466 
467 	for (i = 0; i < num_sge; i++) {
468 		if ((plen + sg_list[i].length) < plen)
469 			return -EMSGSIZE;
470 		plen += sg_list[i].length;
471 		*flitp = cpu_to_be64(((u64)sg_list[i].lkey << 32) |
472 				     sg_list[i].length);
473 		if (++flitp == queue_end)
474 			flitp = queue_start;
475 		*flitp = cpu_to_be64(sg_list[i].addr);
476 		if (++flitp == queue_end)
477 			flitp = queue_start;
478 	}
479 	*flitp = (__force __be64)0;
480 	isglp->op = FW_RI_DATA_ISGL;
481 	isglp->r1 = 0;
482 	isglp->nsge = cpu_to_be16(num_sge);
483 	isglp->r2 = 0;
484 	if (plenp)
485 		*plenp = plen;
486 	return 0;
487 }
488 
489 static int build_rdma_send(struct t4_sq *sq, union t4_wr *wqe,
490 			   const struct ib_send_wr *wr, u8 *len16)
491 {
492 	u32 plen;
493 	int size;
494 	int ret;
495 
496 	if (wr->num_sge > T4_MAX_SEND_SGE)
497 		return -EINVAL;
498 	switch (wr->opcode) {
499 	case IB_WR_SEND:
500 		if (wr->send_flags & IB_SEND_SOLICITED)
501 			wqe->send.sendop_pkd = cpu_to_be32(
502 				FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_SE));
503 		else
504 			wqe->send.sendop_pkd = cpu_to_be32(
505 				FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND));
506 		wqe->send.stag_inv = 0;
507 		break;
508 	case IB_WR_SEND_WITH_INV:
509 		if (wr->send_flags & IB_SEND_SOLICITED)
510 			wqe->send.sendop_pkd = cpu_to_be32(
511 				FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_SE_INV));
512 		else
513 			wqe->send.sendop_pkd = cpu_to_be32(
514 				FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_INV));
515 		wqe->send.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
516 		break;
517 
518 	default:
519 		return -EINVAL;
520 	}
521 	wqe->send.r3 = 0;
522 	wqe->send.r4 = 0;
523 
524 	plen = 0;
525 	if (wr->num_sge) {
526 		if (wr->send_flags & IB_SEND_INLINE) {
527 			ret = build_immd(sq, wqe->send.u.immd_src, wr,
528 					 T4_MAX_SEND_INLINE, &plen);
529 			if (ret)
530 				return ret;
531 			size = sizeof(wqe->send) + sizeof(struct fw_ri_immd) +
532 			       plen;
533 		} else {
534 			ret = build_isgl((__be64 *)sq->queue,
535 					 (__be64 *)&sq->queue[sq->size],
536 					 wqe->send.u.isgl_src,
537 					 wr->sg_list, wr->num_sge, &plen);
538 			if (ret)
539 				return ret;
540 			size = sizeof(wqe->send) + sizeof(struct fw_ri_isgl) +
541 			       wr->num_sge * sizeof(struct fw_ri_sge);
542 		}
543 	} else {
544 		wqe->send.u.immd_src[0].op = FW_RI_DATA_IMMD;
545 		wqe->send.u.immd_src[0].r1 = 0;
546 		wqe->send.u.immd_src[0].r2 = 0;
547 		wqe->send.u.immd_src[0].immdlen = 0;
548 		size = sizeof(wqe->send) + sizeof(struct fw_ri_immd);
549 		plen = 0;
550 	}
551 	*len16 = DIV_ROUND_UP(size, 16);
552 	wqe->send.plen = cpu_to_be32(plen);
553 	return 0;
554 }
555 
556 static int build_rdma_write(struct t4_sq *sq, union t4_wr *wqe,
557 			    const struct ib_send_wr *wr, u8 *len16)
558 {
559 	u32 plen;
560 	int size;
561 	int ret;
562 
563 	if (wr->num_sge > T4_MAX_SEND_SGE)
564 		return -EINVAL;
565 
566 	/*
567 	 * iWARP protocol supports 64 bit immediate data but rdma api
568 	 * limits it to 32bit.
569 	 */
570 	if (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
571 		wqe->write.iw_imm_data.ib_imm_data.imm_data32 = wr->ex.imm_data;
572 	else
573 		wqe->write.iw_imm_data.ib_imm_data.imm_data32 = 0;
574 	wqe->write.stag_sink = cpu_to_be32(rdma_wr(wr)->rkey);
575 	wqe->write.to_sink = cpu_to_be64(rdma_wr(wr)->remote_addr);
576 	if (wr->num_sge) {
577 		if (wr->send_flags & IB_SEND_INLINE) {
578 			ret = build_immd(sq, wqe->write.u.immd_src, wr,
579 					 T4_MAX_WRITE_INLINE, &plen);
580 			if (ret)
581 				return ret;
582 			size = sizeof(wqe->write) + sizeof(struct fw_ri_immd) +
583 			       plen;
584 		} else {
585 			ret = build_isgl((__be64 *)sq->queue,
586 					 (__be64 *)&sq->queue[sq->size],
587 					 wqe->write.u.isgl_src,
588 					 wr->sg_list, wr->num_sge, &plen);
589 			if (ret)
590 				return ret;
591 			size = sizeof(wqe->write) + sizeof(struct fw_ri_isgl) +
592 			       wr->num_sge * sizeof(struct fw_ri_sge);
593 		}
594 	} else {
595 		wqe->write.u.immd_src[0].op = FW_RI_DATA_IMMD;
596 		wqe->write.u.immd_src[0].r1 = 0;
597 		wqe->write.u.immd_src[0].r2 = 0;
598 		wqe->write.u.immd_src[0].immdlen = 0;
599 		size = sizeof(wqe->write) + sizeof(struct fw_ri_immd);
600 		plen = 0;
601 	}
602 	*len16 = DIV_ROUND_UP(size, 16);
603 	wqe->write.plen = cpu_to_be32(plen);
604 	return 0;
605 }
606 
607 static void build_immd_cmpl(struct t4_sq *sq, struct fw_ri_immd_cmpl *immdp,
608 			    struct ib_send_wr *wr)
609 {
610 	memcpy((u8 *)immdp->data, (u8 *)(uintptr_t)wr->sg_list->addr, 16);
611 	memset(immdp->r1, 0, 6);
612 	immdp->op = FW_RI_DATA_IMMD;
613 	immdp->immdlen = 16;
614 }
615 
616 static void build_rdma_write_cmpl(struct t4_sq *sq,
617 				  struct fw_ri_rdma_write_cmpl_wr *wcwr,
618 				  const struct ib_send_wr *wr, u8 *len16)
619 {
620 	u32 plen;
621 	int size;
622 
623 	/*
624 	 * This code assumes the struct fields preceding the write isgl
625 	 * fit in one 64B WR slot.  This is because the WQE is built
626 	 * directly in the dma queue, and wrapping is only handled
627 	 * by the code buildling sgls.  IE the "fixed part" of the wr
628 	 * structs must all fit in 64B.  The WQE build code should probably be
629 	 * redesigned to avoid this restriction, but for now just add
630 	 * the BUILD_BUG_ON() to catch if this WQE struct gets too big.
631 	 */
632 	BUILD_BUG_ON(offsetof(struct fw_ri_rdma_write_cmpl_wr, u) > 64);
633 
634 	wcwr->stag_sink = cpu_to_be32(rdma_wr(wr)->rkey);
635 	wcwr->to_sink = cpu_to_be64(rdma_wr(wr)->remote_addr);
636 	if (wr->next->opcode == IB_WR_SEND)
637 		wcwr->stag_inv = 0;
638 	else
639 		wcwr->stag_inv = cpu_to_be32(wr->next->ex.invalidate_rkey);
640 	wcwr->r2 = 0;
641 	wcwr->r3 = 0;
642 
643 	/* SEND_INV SGL */
644 	if (wr->next->send_flags & IB_SEND_INLINE)
645 		build_immd_cmpl(sq, &wcwr->u_cmpl.immd_src, wr->next);
646 	else
647 		build_isgl((__be64 *)sq->queue, (__be64 *)&sq->queue[sq->size],
648 			   &wcwr->u_cmpl.isgl_src, wr->next->sg_list, 1, NULL);
649 
650 	/* WRITE SGL */
651 	build_isgl((__be64 *)sq->queue, (__be64 *)&sq->queue[sq->size],
652 		   wcwr->u.isgl_src, wr->sg_list, wr->num_sge, &plen);
653 
654 	size = sizeof(*wcwr) + sizeof(struct fw_ri_isgl) +
655 		wr->num_sge * sizeof(struct fw_ri_sge);
656 	wcwr->plen = cpu_to_be32(plen);
657 	*len16 = DIV_ROUND_UP(size, 16);
658 }
659 
660 static int build_rdma_read(union t4_wr *wqe, const struct ib_send_wr *wr,
661 			   u8 *len16)
662 {
663 	if (wr->num_sge > 1)
664 		return -EINVAL;
665 	if (wr->num_sge && wr->sg_list[0].length) {
666 		wqe->read.stag_src = cpu_to_be32(rdma_wr(wr)->rkey);
667 		wqe->read.to_src_hi = cpu_to_be32((u32)(rdma_wr(wr)->remote_addr
668 							>> 32));
669 		wqe->read.to_src_lo = cpu_to_be32((u32)rdma_wr(wr)->remote_addr);
670 		wqe->read.stag_sink = cpu_to_be32(wr->sg_list[0].lkey);
671 		wqe->read.plen = cpu_to_be32(wr->sg_list[0].length);
672 		wqe->read.to_sink_hi = cpu_to_be32((u32)(wr->sg_list[0].addr
673 							 >> 32));
674 		wqe->read.to_sink_lo = cpu_to_be32((u32)(wr->sg_list[0].addr));
675 	} else {
676 		wqe->read.stag_src = cpu_to_be32(2);
677 		wqe->read.to_src_hi = 0;
678 		wqe->read.to_src_lo = 0;
679 		wqe->read.stag_sink = cpu_to_be32(2);
680 		wqe->read.plen = 0;
681 		wqe->read.to_sink_hi = 0;
682 		wqe->read.to_sink_lo = 0;
683 	}
684 	wqe->read.r2 = 0;
685 	wqe->read.r5 = 0;
686 	*len16 = DIV_ROUND_UP(sizeof(wqe->read), 16);
687 	return 0;
688 }
689 
690 static void post_write_cmpl(struct c4iw_qp *qhp, const struct ib_send_wr *wr)
691 {
692 	bool send_signaled = (wr->next->send_flags & IB_SEND_SIGNALED) ||
693 			     qhp->sq_sig_all;
694 	bool write_signaled = (wr->send_flags & IB_SEND_SIGNALED) ||
695 			      qhp->sq_sig_all;
696 	struct t4_swsqe *swsqe;
697 	union t4_wr *wqe;
698 	u16 write_wrid;
699 	u8 len16;
700 	u16 idx;
701 
702 	/*
703 	 * The sw_sq entries still look like a WRITE and a SEND and consume
704 	 * 2 slots. The FW WR, however, will be a single uber-WR.
705 	 */
706 	wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
707 	       qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
708 	build_rdma_write_cmpl(&qhp->wq.sq, &wqe->write_cmpl, wr, &len16);
709 
710 	/* WRITE swsqe */
711 	swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
712 	swsqe->opcode = FW_RI_RDMA_WRITE;
713 	swsqe->idx = qhp->wq.sq.pidx;
714 	swsqe->complete = 0;
715 	swsqe->signaled = write_signaled;
716 	swsqe->flushed = 0;
717 	swsqe->wr_id = wr->wr_id;
718 	if (c4iw_wr_log) {
719 		swsqe->sge_ts =
720 			cxgb4_read_sge_timestamp(qhp->rhp->rdev.lldi.ports[0]);
721 		swsqe->host_time = ktime_get();
722 	}
723 
724 	write_wrid = qhp->wq.sq.pidx;
725 
726 	/* just bump the sw_sq */
727 	qhp->wq.sq.in_use++;
728 	if (++qhp->wq.sq.pidx == qhp->wq.sq.size)
729 		qhp->wq.sq.pidx = 0;
730 
731 	/* SEND_WITH_INV swsqe */
732 	swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
733 	if (wr->next->opcode == IB_WR_SEND)
734 		swsqe->opcode = FW_RI_SEND;
735 	else
736 		swsqe->opcode = FW_RI_SEND_WITH_INV;
737 	swsqe->idx = qhp->wq.sq.pidx;
738 	swsqe->complete = 0;
739 	swsqe->signaled = send_signaled;
740 	swsqe->flushed = 0;
741 	swsqe->wr_id = wr->next->wr_id;
742 	if (c4iw_wr_log) {
743 		swsqe->sge_ts =
744 			cxgb4_read_sge_timestamp(qhp->rhp->rdev.lldi.ports[0]);
745 		swsqe->host_time = ktime_get();
746 	}
747 
748 	wqe->write_cmpl.flags_send = send_signaled ? FW_RI_COMPLETION_FLAG : 0;
749 	wqe->write_cmpl.wrid_send = qhp->wq.sq.pidx;
750 
751 	init_wr_hdr(wqe, write_wrid, FW_RI_RDMA_WRITE_CMPL_WR,
752 		    write_signaled ? FW_RI_COMPLETION_FLAG : 0, len16);
753 	t4_sq_produce(&qhp->wq, len16);
754 	idx = DIV_ROUND_UP(len16 * 16, T4_EQ_ENTRY_SIZE);
755 
756 	t4_ring_sq_db(&qhp->wq, idx, wqe);
757 }
758 
759 static int build_rdma_recv(struct c4iw_qp *qhp, union t4_recv_wr *wqe,
760 			   const struct ib_recv_wr *wr, u8 *len16)
761 {
762 	int ret;
763 
764 	ret = build_isgl((__be64 *)qhp->wq.rq.queue,
765 			 (__be64 *)&qhp->wq.rq.queue[qhp->wq.rq.size],
766 			 &wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL);
767 	if (ret)
768 		return ret;
769 	*len16 = DIV_ROUND_UP(
770 		sizeof(wqe->recv) + wr->num_sge * sizeof(struct fw_ri_sge), 16);
771 	return 0;
772 }
773 
774 static int build_srq_recv(union t4_recv_wr *wqe, const struct ib_recv_wr *wr,
775 			  u8 *len16)
776 {
777 	int ret;
778 
779 	ret = build_isgl((__be64 *)wqe, (__be64 *)(wqe + 1),
780 			 &wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL);
781 	if (ret)
782 		return ret;
783 	*len16 = DIV_ROUND_UP(sizeof(wqe->recv) +
784 			      wr->num_sge * sizeof(struct fw_ri_sge), 16);
785 	return 0;
786 }
787 
788 static void build_tpte_memreg(struct fw_ri_fr_nsmr_tpte_wr *fr,
789 			      const struct ib_reg_wr *wr, struct c4iw_mr *mhp,
790 			      u8 *len16)
791 {
792 	__be64 *p = (__be64 *)fr->pbl;
793 
794 	fr->r2 = cpu_to_be32(0);
795 	fr->stag = cpu_to_be32(mhp->ibmr.rkey);
796 
797 	fr->tpte.valid_to_pdid = cpu_to_be32(FW_RI_TPTE_VALID_F |
798 		FW_RI_TPTE_STAGKEY_V((mhp->ibmr.rkey & FW_RI_TPTE_STAGKEY_M)) |
799 		FW_RI_TPTE_STAGSTATE_V(1) |
800 		FW_RI_TPTE_STAGTYPE_V(FW_RI_STAG_NSMR) |
801 		FW_RI_TPTE_PDID_V(mhp->attr.pdid));
802 	fr->tpte.locread_to_qpid = cpu_to_be32(
803 		FW_RI_TPTE_PERM_V(c4iw_ib_to_tpt_access(wr->access)) |
804 		FW_RI_TPTE_ADDRTYPE_V(FW_RI_VA_BASED_TO) |
805 		FW_RI_TPTE_PS_V(ilog2(wr->mr->page_size) - 12));
806 	fr->tpte.nosnoop_pbladdr = cpu_to_be32(FW_RI_TPTE_PBLADDR_V(
807 		PBL_OFF(&mhp->rhp->rdev, mhp->attr.pbl_addr)>>3));
808 	fr->tpte.dca_mwbcnt_pstag = cpu_to_be32(0);
809 	fr->tpte.len_hi = cpu_to_be32(0);
810 	fr->tpte.len_lo = cpu_to_be32(mhp->ibmr.length);
811 	fr->tpte.va_hi = cpu_to_be32(mhp->ibmr.iova >> 32);
812 	fr->tpte.va_lo_fbo = cpu_to_be32(mhp->ibmr.iova & 0xffffffff);
813 
814 	p[0] = cpu_to_be64((u64)mhp->mpl[0]);
815 	p[1] = cpu_to_be64((u64)mhp->mpl[1]);
816 
817 	*len16 = DIV_ROUND_UP(sizeof(*fr), 16);
818 }
819 
820 static int build_memreg(struct t4_sq *sq, union t4_wr *wqe,
821 			const struct ib_reg_wr *wr, struct c4iw_mr *mhp,
822 			u8 *len16, bool dsgl_supported)
823 {
824 	struct fw_ri_immd *imdp;
825 	__be64 *p;
826 	int i;
827 	int pbllen = roundup(mhp->mpl_len * sizeof(u64), 32);
828 	int rem;
829 
830 	if (mhp->mpl_len > t4_max_fr_depth(dsgl_supported && use_dsgl))
831 		return -EINVAL;
832 
833 	wqe->fr.qpbinde_to_dcacpu = 0;
834 	wqe->fr.pgsz_shift = ilog2(wr->mr->page_size) - 12;
835 	wqe->fr.addr_type = FW_RI_VA_BASED_TO;
836 	wqe->fr.mem_perms = c4iw_ib_to_tpt_access(wr->access);
837 	wqe->fr.len_hi = 0;
838 	wqe->fr.len_lo = cpu_to_be32(mhp->ibmr.length);
839 	wqe->fr.stag = cpu_to_be32(wr->key);
840 	wqe->fr.va_hi = cpu_to_be32(mhp->ibmr.iova >> 32);
841 	wqe->fr.va_lo_fbo = cpu_to_be32(mhp->ibmr.iova &
842 					0xffffffff);
843 
844 	if (dsgl_supported && use_dsgl && (pbllen > max_fr_immd)) {
845 		struct fw_ri_dsgl *sglp;
846 
847 		for (i = 0; i < mhp->mpl_len; i++)
848 			mhp->mpl[i] = (__force u64)cpu_to_be64((u64)mhp->mpl[i]);
849 
850 		sglp = (struct fw_ri_dsgl *)(&wqe->fr + 1);
851 		sglp->op = FW_RI_DATA_DSGL;
852 		sglp->r1 = 0;
853 		sglp->nsge = cpu_to_be16(1);
854 		sglp->addr0 = cpu_to_be64(mhp->mpl_addr);
855 		sglp->len0 = cpu_to_be32(pbllen);
856 
857 		*len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*sglp), 16);
858 	} else {
859 		imdp = (struct fw_ri_immd *)(&wqe->fr + 1);
860 		imdp->op = FW_RI_DATA_IMMD;
861 		imdp->r1 = 0;
862 		imdp->r2 = 0;
863 		imdp->immdlen = cpu_to_be32(pbllen);
864 		p = (__be64 *)(imdp + 1);
865 		rem = pbllen;
866 		for (i = 0; i < mhp->mpl_len; i++) {
867 			*p = cpu_to_be64((u64)mhp->mpl[i]);
868 			rem -= sizeof(*p);
869 			if (++p == (__be64 *)&sq->queue[sq->size])
870 				p = (__be64 *)sq->queue;
871 		}
872 		while (rem) {
873 			*p = 0;
874 			rem -= sizeof(*p);
875 			if (++p == (__be64 *)&sq->queue[sq->size])
876 				p = (__be64 *)sq->queue;
877 		}
878 		*len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*imdp)
879 				      + pbllen, 16);
880 	}
881 	return 0;
882 }
883 
884 static int build_inv_stag(union t4_wr *wqe, const struct ib_send_wr *wr,
885 			  u8 *len16)
886 {
887 	wqe->inv.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
888 	wqe->inv.r2 = 0;
889 	*len16 = DIV_ROUND_UP(sizeof(wqe->inv), 16);
890 	return 0;
891 }
892 
893 void c4iw_qp_add_ref(struct ib_qp *qp)
894 {
895 	pr_debug("ib_qp %p\n", qp);
896 	refcount_inc(&to_c4iw_qp(qp)->qp_refcnt);
897 }
898 
899 void c4iw_qp_rem_ref(struct ib_qp *qp)
900 {
901 	pr_debug("ib_qp %p\n", qp);
902 	if (refcount_dec_and_test(&to_c4iw_qp(qp)->qp_refcnt))
903 		complete(&to_c4iw_qp(qp)->qp_rel_comp);
904 }
905 
906 static void add_to_fc_list(struct list_head *head, struct list_head *entry)
907 {
908 	if (list_empty(entry))
909 		list_add_tail(entry, head);
910 }
911 
912 static int ring_kernel_sq_db(struct c4iw_qp *qhp, u16 inc)
913 {
914 	unsigned long flags;
915 
916 	xa_lock_irqsave(&qhp->rhp->qps, flags);
917 	spin_lock(&qhp->lock);
918 	if (qhp->rhp->db_state == NORMAL)
919 		t4_ring_sq_db(&qhp->wq, inc, NULL);
920 	else {
921 		add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
922 		qhp->wq.sq.wq_pidx_inc += inc;
923 	}
924 	spin_unlock(&qhp->lock);
925 	xa_unlock_irqrestore(&qhp->rhp->qps, flags);
926 	return 0;
927 }
928 
929 static int ring_kernel_rq_db(struct c4iw_qp *qhp, u16 inc)
930 {
931 	unsigned long flags;
932 
933 	xa_lock_irqsave(&qhp->rhp->qps, flags);
934 	spin_lock(&qhp->lock);
935 	if (qhp->rhp->db_state == NORMAL)
936 		t4_ring_rq_db(&qhp->wq, inc, NULL);
937 	else {
938 		add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
939 		qhp->wq.rq.wq_pidx_inc += inc;
940 	}
941 	spin_unlock(&qhp->lock);
942 	xa_unlock_irqrestore(&qhp->rhp->qps, flags);
943 	return 0;
944 }
945 
946 static int ib_to_fw_opcode(int ib_opcode)
947 {
948 	int opcode;
949 
950 	switch (ib_opcode) {
951 	case IB_WR_SEND_WITH_INV:
952 		opcode = FW_RI_SEND_WITH_INV;
953 		break;
954 	case IB_WR_SEND:
955 		opcode = FW_RI_SEND;
956 		break;
957 	case IB_WR_RDMA_WRITE:
958 		opcode = FW_RI_RDMA_WRITE;
959 		break;
960 	case IB_WR_RDMA_WRITE_WITH_IMM:
961 		opcode = FW_RI_WRITE_IMMEDIATE;
962 		break;
963 	case IB_WR_RDMA_READ:
964 	case IB_WR_RDMA_READ_WITH_INV:
965 		opcode = FW_RI_READ_REQ;
966 		break;
967 	case IB_WR_REG_MR:
968 		opcode = FW_RI_FAST_REGISTER;
969 		break;
970 	case IB_WR_LOCAL_INV:
971 		opcode = FW_RI_LOCAL_INV;
972 		break;
973 	default:
974 		opcode = -EINVAL;
975 	}
976 	return opcode;
977 }
978 
979 static int complete_sq_drain_wr(struct c4iw_qp *qhp,
980 				const struct ib_send_wr *wr)
981 {
982 	struct t4_cqe cqe = {};
983 	struct c4iw_cq *schp;
984 	unsigned long flag;
985 	struct t4_cq *cq;
986 	int opcode;
987 
988 	schp = to_c4iw_cq(qhp->ibqp.send_cq);
989 	cq = &schp->cq;
990 
991 	opcode = ib_to_fw_opcode(wr->opcode);
992 	if (opcode < 0)
993 		return opcode;
994 
995 	cqe.u.drain_cookie = wr->wr_id;
996 	cqe.header = cpu_to_be32(CQE_STATUS_V(T4_ERR_SWFLUSH) |
997 				 CQE_OPCODE_V(opcode) |
998 				 CQE_TYPE_V(1) |
999 				 CQE_SWCQE_V(1) |
1000 				 CQE_DRAIN_V(1) |
1001 				 CQE_QPID_V(qhp->wq.sq.qid));
1002 
1003 	spin_lock_irqsave(&schp->lock, flag);
1004 	cqe.bits_type_ts = cpu_to_be64(CQE_GENBIT_V((u64)cq->gen));
1005 	cq->sw_queue[cq->sw_pidx] = cqe;
1006 	t4_swcq_produce(cq);
1007 	spin_unlock_irqrestore(&schp->lock, flag);
1008 
1009 	if (t4_clear_cq_armed(&schp->cq)) {
1010 		spin_lock_irqsave(&schp->comp_handler_lock, flag);
1011 		(*schp->ibcq.comp_handler)(&schp->ibcq,
1012 					   schp->ibcq.cq_context);
1013 		spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1014 	}
1015 	return 0;
1016 }
1017 
1018 static int complete_sq_drain_wrs(struct c4iw_qp *qhp,
1019 				 const struct ib_send_wr *wr,
1020 				 const struct ib_send_wr **bad_wr)
1021 {
1022 	int ret = 0;
1023 
1024 	while (wr) {
1025 		ret = complete_sq_drain_wr(qhp, wr);
1026 		if (ret) {
1027 			*bad_wr = wr;
1028 			break;
1029 		}
1030 		wr = wr->next;
1031 	}
1032 	return ret;
1033 }
1034 
1035 static void complete_rq_drain_wr(struct c4iw_qp *qhp,
1036 				 const struct ib_recv_wr *wr)
1037 {
1038 	struct t4_cqe cqe = {};
1039 	struct c4iw_cq *rchp;
1040 	unsigned long flag;
1041 	struct t4_cq *cq;
1042 
1043 	rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
1044 	cq = &rchp->cq;
1045 
1046 	cqe.u.drain_cookie = wr->wr_id;
1047 	cqe.header = cpu_to_be32(CQE_STATUS_V(T4_ERR_SWFLUSH) |
1048 				 CQE_OPCODE_V(FW_RI_SEND) |
1049 				 CQE_TYPE_V(0) |
1050 				 CQE_SWCQE_V(1) |
1051 				 CQE_DRAIN_V(1) |
1052 				 CQE_QPID_V(qhp->wq.sq.qid));
1053 
1054 	spin_lock_irqsave(&rchp->lock, flag);
1055 	cqe.bits_type_ts = cpu_to_be64(CQE_GENBIT_V((u64)cq->gen));
1056 	cq->sw_queue[cq->sw_pidx] = cqe;
1057 	t4_swcq_produce(cq);
1058 	spin_unlock_irqrestore(&rchp->lock, flag);
1059 
1060 	if (t4_clear_cq_armed(&rchp->cq)) {
1061 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1062 		(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1063 					   rchp->ibcq.cq_context);
1064 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1065 	}
1066 }
1067 
1068 static void complete_rq_drain_wrs(struct c4iw_qp *qhp,
1069 				  const struct ib_recv_wr *wr)
1070 {
1071 	while (wr) {
1072 		complete_rq_drain_wr(qhp, wr);
1073 		wr = wr->next;
1074 	}
1075 }
1076 
1077 int c4iw_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
1078 		   const struct ib_send_wr **bad_wr)
1079 {
1080 	int err = 0;
1081 	u8 len16 = 0;
1082 	enum fw_wr_opcodes fw_opcode = 0;
1083 	enum fw_ri_wr_flags fw_flags;
1084 	struct c4iw_qp *qhp;
1085 	struct c4iw_dev *rhp;
1086 	union t4_wr *wqe = NULL;
1087 	u32 num_wrs;
1088 	struct t4_swsqe *swsqe;
1089 	unsigned long flag;
1090 	u16 idx = 0;
1091 
1092 	qhp = to_c4iw_qp(ibqp);
1093 	rhp = qhp->rhp;
1094 	spin_lock_irqsave(&qhp->lock, flag);
1095 
1096 	/*
1097 	 * If the qp has been flushed, then just insert a special
1098 	 * drain cqe.
1099 	 */
1100 	if (qhp->wq.flushed) {
1101 		spin_unlock_irqrestore(&qhp->lock, flag);
1102 		err = complete_sq_drain_wrs(qhp, wr, bad_wr);
1103 		return err;
1104 	}
1105 	num_wrs = t4_sq_avail(&qhp->wq);
1106 	if (num_wrs == 0) {
1107 		spin_unlock_irqrestore(&qhp->lock, flag);
1108 		*bad_wr = wr;
1109 		return -ENOMEM;
1110 	}
1111 
1112 	/*
1113 	 * Fastpath for NVMe-oF target WRITE + SEND_WITH_INV wr chain which is
1114 	 * the response for small NVMEe-oF READ requests.  If the chain is
1115 	 * exactly a WRITE->SEND_WITH_INV or a WRITE->SEND and the sgl depths
1116 	 * and lengths meet the requirements of the fw_ri_write_cmpl_wr work
1117 	 * request, then build and post the write_cmpl WR. If any of the tests
1118 	 * below are not true, then we continue on with the tradtional WRITE
1119 	 * and SEND WRs.
1120 	 */
1121 	if (qhp->rhp->rdev.lldi.write_cmpl_support &&
1122 	    CHELSIO_CHIP_VERSION(qhp->rhp->rdev.lldi.adapter_type) >=
1123 	    CHELSIO_T5 &&
1124 	    wr && wr->next && !wr->next->next &&
1125 	    wr->opcode == IB_WR_RDMA_WRITE &&
1126 	    wr->sg_list[0].length && wr->num_sge <= T4_WRITE_CMPL_MAX_SGL &&
1127 	    (wr->next->opcode == IB_WR_SEND ||
1128 	    wr->next->opcode == IB_WR_SEND_WITH_INV) &&
1129 	    wr->next->sg_list[0].length == T4_WRITE_CMPL_MAX_CQE &&
1130 	    wr->next->num_sge == 1 && num_wrs >= 2) {
1131 		post_write_cmpl(qhp, wr);
1132 		spin_unlock_irqrestore(&qhp->lock, flag);
1133 		return 0;
1134 	}
1135 
1136 	while (wr) {
1137 		if (num_wrs == 0) {
1138 			err = -ENOMEM;
1139 			*bad_wr = wr;
1140 			break;
1141 		}
1142 		wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
1143 		      qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
1144 
1145 		fw_flags = 0;
1146 		if (wr->send_flags & IB_SEND_SOLICITED)
1147 			fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
1148 		if (wr->send_flags & IB_SEND_SIGNALED || qhp->sq_sig_all)
1149 			fw_flags |= FW_RI_COMPLETION_FLAG;
1150 		swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
1151 		switch (wr->opcode) {
1152 		case IB_WR_SEND_WITH_INV:
1153 		case IB_WR_SEND:
1154 			if (wr->send_flags & IB_SEND_FENCE)
1155 				fw_flags |= FW_RI_READ_FENCE_FLAG;
1156 			fw_opcode = FW_RI_SEND_WR;
1157 			if (wr->opcode == IB_WR_SEND)
1158 				swsqe->opcode = FW_RI_SEND;
1159 			else
1160 				swsqe->opcode = FW_RI_SEND_WITH_INV;
1161 			err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
1162 			break;
1163 		case IB_WR_RDMA_WRITE_WITH_IMM:
1164 			if (unlikely(!rhp->rdev.lldi.write_w_imm_support)) {
1165 				err = -EINVAL;
1166 				break;
1167 			}
1168 			fw_flags |= FW_RI_RDMA_WRITE_WITH_IMMEDIATE;
1169 			fallthrough;
1170 		case IB_WR_RDMA_WRITE:
1171 			fw_opcode = FW_RI_RDMA_WRITE_WR;
1172 			swsqe->opcode = FW_RI_RDMA_WRITE;
1173 			err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
1174 			break;
1175 		case IB_WR_RDMA_READ:
1176 		case IB_WR_RDMA_READ_WITH_INV:
1177 			fw_opcode = FW_RI_RDMA_READ_WR;
1178 			swsqe->opcode = FW_RI_READ_REQ;
1179 			if (wr->opcode == IB_WR_RDMA_READ_WITH_INV) {
1180 				c4iw_invalidate_mr(rhp, wr->sg_list[0].lkey);
1181 				fw_flags = FW_RI_RDMA_READ_INVALIDATE;
1182 			} else {
1183 				fw_flags = 0;
1184 			}
1185 			err = build_rdma_read(wqe, wr, &len16);
1186 			if (err)
1187 				break;
1188 			swsqe->read_len = wr->sg_list[0].length;
1189 			if (!qhp->wq.sq.oldest_read)
1190 				qhp->wq.sq.oldest_read = swsqe;
1191 			break;
1192 		case IB_WR_REG_MR: {
1193 			struct c4iw_mr *mhp = to_c4iw_mr(reg_wr(wr)->mr);
1194 
1195 			swsqe->opcode = FW_RI_FAST_REGISTER;
1196 			if (rhp->rdev.lldi.fr_nsmr_tpte_wr_support &&
1197 			    !mhp->attr.state && mhp->mpl_len <= 2) {
1198 				fw_opcode = FW_RI_FR_NSMR_TPTE_WR;
1199 				build_tpte_memreg(&wqe->fr_tpte, reg_wr(wr),
1200 						  mhp, &len16);
1201 			} else {
1202 				fw_opcode = FW_RI_FR_NSMR_WR;
1203 				err = build_memreg(&qhp->wq.sq, wqe, reg_wr(wr),
1204 				       mhp, &len16,
1205 				       rhp->rdev.lldi.ulptx_memwrite_dsgl);
1206 				if (err)
1207 					break;
1208 			}
1209 			mhp->attr.state = 1;
1210 			break;
1211 		}
1212 		case IB_WR_LOCAL_INV:
1213 			if (wr->send_flags & IB_SEND_FENCE)
1214 				fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
1215 			fw_opcode = FW_RI_INV_LSTAG_WR;
1216 			swsqe->opcode = FW_RI_LOCAL_INV;
1217 			err = build_inv_stag(wqe, wr, &len16);
1218 			c4iw_invalidate_mr(rhp, wr->ex.invalidate_rkey);
1219 			break;
1220 		default:
1221 			pr_warn("%s post of type=%d TBD!\n", __func__,
1222 				wr->opcode);
1223 			err = -EINVAL;
1224 		}
1225 		if (err) {
1226 			*bad_wr = wr;
1227 			break;
1228 		}
1229 		swsqe->idx = qhp->wq.sq.pidx;
1230 		swsqe->complete = 0;
1231 		swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED) ||
1232 				  qhp->sq_sig_all;
1233 		swsqe->flushed = 0;
1234 		swsqe->wr_id = wr->wr_id;
1235 		if (c4iw_wr_log) {
1236 			swsqe->sge_ts = cxgb4_read_sge_timestamp(
1237 					rhp->rdev.lldi.ports[0]);
1238 			swsqe->host_time = ktime_get();
1239 		}
1240 
1241 		init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
1242 
1243 		pr_debug("cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u\n",
1244 			 (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
1245 			 swsqe->opcode, swsqe->read_len);
1246 		wr = wr->next;
1247 		num_wrs--;
1248 		t4_sq_produce(&qhp->wq, len16);
1249 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
1250 	}
1251 	if (!rhp->rdev.status_page->db_off) {
1252 		t4_ring_sq_db(&qhp->wq, idx, wqe);
1253 		spin_unlock_irqrestore(&qhp->lock, flag);
1254 	} else {
1255 		spin_unlock_irqrestore(&qhp->lock, flag);
1256 		ring_kernel_sq_db(qhp, idx);
1257 	}
1258 	return err;
1259 }
1260 
1261 int c4iw_post_receive(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
1262 		      const struct ib_recv_wr **bad_wr)
1263 {
1264 	int err = 0;
1265 	struct c4iw_qp *qhp;
1266 	union t4_recv_wr *wqe = NULL;
1267 	u32 num_wrs;
1268 	u8 len16 = 0;
1269 	unsigned long flag;
1270 	u16 idx = 0;
1271 
1272 	qhp = to_c4iw_qp(ibqp);
1273 	spin_lock_irqsave(&qhp->lock, flag);
1274 
1275 	/*
1276 	 * If the qp has been flushed, then just insert a special
1277 	 * drain cqe.
1278 	 */
1279 	if (qhp->wq.flushed) {
1280 		spin_unlock_irqrestore(&qhp->lock, flag);
1281 		complete_rq_drain_wrs(qhp, wr);
1282 		return err;
1283 	}
1284 	num_wrs = t4_rq_avail(&qhp->wq);
1285 	if (num_wrs == 0) {
1286 		spin_unlock_irqrestore(&qhp->lock, flag);
1287 		*bad_wr = wr;
1288 		return -ENOMEM;
1289 	}
1290 	while (wr) {
1291 		if (wr->num_sge > T4_MAX_RECV_SGE) {
1292 			err = -EINVAL;
1293 			*bad_wr = wr;
1294 			break;
1295 		}
1296 		wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
1297 					   qhp->wq.rq.wq_pidx *
1298 					   T4_EQ_ENTRY_SIZE);
1299 		if (num_wrs)
1300 			err = build_rdma_recv(qhp, wqe, wr, &len16);
1301 		else
1302 			err = -ENOMEM;
1303 		if (err) {
1304 			*bad_wr = wr;
1305 			break;
1306 		}
1307 
1308 		qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
1309 		if (c4iw_wr_log) {
1310 			qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].sge_ts =
1311 				cxgb4_read_sge_timestamp(
1312 						qhp->rhp->rdev.lldi.ports[0]);
1313 			qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].host_time =
1314 				ktime_get();
1315 		}
1316 
1317 		wqe->recv.opcode = FW_RI_RECV_WR;
1318 		wqe->recv.r1 = 0;
1319 		wqe->recv.wrid = qhp->wq.rq.pidx;
1320 		wqe->recv.r2[0] = 0;
1321 		wqe->recv.r2[1] = 0;
1322 		wqe->recv.r2[2] = 0;
1323 		wqe->recv.len16 = len16;
1324 		pr_debug("cookie 0x%llx pidx %u\n",
1325 			 (unsigned long long)wr->wr_id, qhp->wq.rq.pidx);
1326 		t4_rq_produce(&qhp->wq, len16);
1327 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
1328 		wr = wr->next;
1329 		num_wrs--;
1330 	}
1331 	if (!qhp->rhp->rdev.status_page->db_off) {
1332 		t4_ring_rq_db(&qhp->wq, idx, wqe);
1333 		spin_unlock_irqrestore(&qhp->lock, flag);
1334 	} else {
1335 		spin_unlock_irqrestore(&qhp->lock, flag);
1336 		ring_kernel_rq_db(qhp, idx);
1337 	}
1338 	return err;
1339 }
1340 
1341 static void defer_srq_wr(struct t4_srq *srq, union t4_recv_wr *wqe,
1342 			 u64 wr_id, u8 len16)
1343 {
1344 	struct t4_srq_pending_wr *pwr = &srq->pending_wrs[srq->pending_pidx];
1345 
1346 	pr_debug("%s cidx %u pidx %u wq_pidx %u in_use %u ooo_count %u wr_id 0x%llx pending_cidx %u pending_pidx %u pending_in_use %u\n",
1347 		 __func__, srq->cidx, srq->pidx, srq->wq_pidx,
1348 		 srq->in_use, srq->ooo_count,
1349 		 (unsigned long long)wr_id, srq->pending_cidx,
1350 		 srq->pending_pidx, srq->pending_in_use);
1351 	pwr->wr_id = wr_id;
1352 	pwr->len16 = len16;
1353 	memcpy(&pwr->wqe, wqe, len16 * 16);
1354 	t4_srq_produce_pending_wr(srq);
1355 }
1356 
1357 int c4iw_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
1358 		       const struct ib_recv_wr **bad_wr)
1359 {
1360 	union t4_recv_wr *wqe, lwqe;
1361 	struct c4iw_srq *srq;
1362 	unsigned long flag;
1363 	u8 len16 = 0;
1364 	u16 idx = 0;
1365 	int err = 0;
1366 	u32 num_wrs;
1367 
1368 	srq = to_c4iw_srq(ibsrq);
1369 	spin_lock_irqsave(&srq->lock, flag);
1370 	num_wrs = t4_srq_avail(&srq->wq);
1371 	if (num_wrs == 0) {
1372 		spin_unlock_irqrestore(&srq->lock, flag);
1373 		return -ENOMEM;
1374 	}
1375 	while (wr) {
1376 		if (wr->num_sge > T4_MAX_RECV_SGE) {
1377 			err = -EINVAL;
1378 			*bad_wr = wr;
1379 			break;
1380 		}
1381 		wqe = &lwqe;
1382 		if (num_wrs)
1383 			err = build_srq_recv(wqe, wr, &len16);
1384 		else
1385 			err = -ENOMEM;
1386 		if (err) {
1387 			*bad_wr = wr;
1388 			break;
1389 		}
1390 
1391 		wqe->recv.opcode = FW_RI_RECV_WR;
1392 		wqe->recv.r1 = 0;
1393 		wqe->recv.wrid = srq->wq.pidx;
1394 		wqe->recv.r2[0] = 0;
1395 		wqe->recv.r2[1] = 0;
1396 		wqe->recv.r2[2] = 0;
1397 		wqe->recv.len16 = len16;
1398 
1399 		if (srq->wq.ooo_count ||
1400 		    srq->wq.pending_in_use ||
1401 		    srq->wq.sw_rq[srq->wq.pidx].valid) {
1402 			defer_srq_wr(&srq->wq, wqe, wr->wr_id, len16);
1403 		} else {
1404 			srq->wq.sw_rq[srq->wq.pidx].wr_id = wr->wr_id;
1405 			srq->wq.sw_rq[srq->wq.pidx].valid = 1;
1406 			c4iw_copy_wr_to_srq(&srq->wq, wqe, len16);
1407 			pr_debug("%s cidx %u pidx %u wq_pidx %u in_use %u wr_id 0x%llx\n",
1408 				 __func__, srq->wq.cidx,
1409 				 srq->wq.pidx, srq->wq.wq_pidx,
1410 				 srq->wq.in_use,
1411 				 (unsigned long long)wr->wr_id);
1412 			t4_srq_produce(&srq->wq, len16);
1413 			idx += DIV_ROUND_UP(len16 * 16, T4_EQ_ENTRY_SIZE);
1414 		}
1415 		wr = wr->next;
1416 		num_wrs--;
1417 	}
1418 	if (idx)
1419 		t4_ring_srq_db(&srq->wq, idx, len16, wqe);
1420 	spin_unlock_irqrestore(&srq->lock, flag);
1421 	return err;
1422 }
1423 
1424 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
1425 				    u8 *ecode)
1426 {
1427 	int status;
1428 	int tagged;
1429 	int opcode;
1430 	int rqtype;
1431 	int send_inv;
1432 
1433 	if (!err_cqe) {
1434 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
1435 		*ecode = 0;
1436 		return;
1437 	}
1438 
1439 	status = CQE_STATUS(err_cqe);
1440 	opcode = CQE_OPCODE(err_cqe);
1441 	rqtype = RQ_TYPE(err_cqe);
1442 	send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
1443 		   (opcode == FW_RI_SEND_WITH_SE_INV);
1444 	tagged = (opcode == FW_RI_RDMA_WRITE) ||
1445 		 (rqtype && (opcode == FW_RI_READ_RESP));
1446 
1447 	switch (status) {
1448 	case T4_ERR_STAG:
1449 		if (send_inv) {
1450 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1451 			*ecode = RDMAP_CANT_INV_STAG;
1452 		} else {
1453 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1454 			*ecode = RDMAP_INV_STAG;
1455 		}
1456 		break;
1457 	case T4_ERR_PDID:
1458 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1459 		if ((opcode == FW_RI_SEND_WITH_INV) ||
1460 		    (opcode == FW_RI_SEND_WITH_SE_INV))
1461 			*ecode = RDMAP_CANT_INV_STAG;
1462 		else
1463 			*ecode = RDMAP_STAG_NOT_ASSOC;
1464 		break;
1465 	case T4_ERR_QPID:
1466 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1467 		*ecode = RDMAP_STAG_NOT_ASSOC;
1468 		break;
1469 	case T4_ERR_ACCESS:
1470 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1471 		*ecode = RDMAP_ACC_VIOL;
1472 		break;
1473 	case T4_ERR_WRAP:
1474 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1475 		*ecode = RDMAP_TO_WRAP;
1476 		break;
1477 	case T4_ERR_BOUND:
1478 		if (tagged) {
1479 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1480 			*ecode = DDPT_BASE_BOUNDS;
1481 		} else {
1482 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1483 			*ecode = RDMAP_BASE_BOUNDS;
1484 		}
1485 		break;
1486 	case T4_ERR_INVALIDATE_SHARED_MR:
1487 	case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
1488 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1489 		*ecode = RDMAP_CANT_INV_STAG;
1490 		break;
1491 	case T4_ERR_ECC:
1492 	case T4_ERR_ECC_PSTAG:
1493 	case T4_ERR_INTERNAL_ERR:
1494 		*layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
1495 		*ecode = 0;
1496 		break;
1497 	case T4_ERR_OUT_OF_RQE:
1498 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1499 		*ecode = DDPU_INV_MSN_NOBUF;
1500 		break;
1501 	case T4_ERR_PBL_ADDR_BOUND:
1502 		*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1503 		*ecode = DDPT_BASE_BOUNDS;
1504 		break;
1505 	case T4_ERR_CRC:
1506 		*layer_type = LAYER_MPA|DDP_LLP;
1507 		*ecode = MPA_CRC_ERR;
1508 		break;
1509 	case T4_ERR_MARKER:
1510 		*layer_type = LAYER_MPA|DDP_LLP;
1511 		*ecode = MPA_MARKER_ERR;
1512 		break;
1513 	case T4_ERR_PDU_LEN_ERR:
1514 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1515 		*ecode = DDPU_MSG_TOOBIG;
1516 		break;
1517 	case T4_ERR_DDP_VERSION:
1518 		if (tagged) {
1519 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1520 			*ecode = DDPT_INV_VERS;
1521 		} else {
1522 			*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1523 			*ecode = DDPU_INV_VERS;
1524 		}
1525 		break;
1526 	case T4_ERR_RDMA_VERSION:
1527 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1528 		*ecode = RDMAP_INV_VERS;
1529 		break;
1530 	case T4_ERR_OPCODE:
1531 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1532 		*ecode = RDMAP_INV_OPCODE;
1533 		break;
1534 	case T4_ERR_DDP_QUEUE_NUM:
1535 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1536 		*ecode = DDPU_INV_QN;
1537 		break;
1538 	case T4_ERR_MSN:
1539 	case T4_ERR_MSN_GAP:
1540 	case T4_ERR_MSN_RANGE:
1541 	case T4_ERR_IRD_OVERFLOW:
1542 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1543 		*ecode = DDPU_INV_MSN_RANGE;
1544 		break;
1545 	case T4_ERR_TBIT:
1546 		*layer_type = LAYER_DDP|DDP_LOCAL_CATA;
1547 		*ecode = 0;
1548 		break;
1549 	case T4_ERR_MO:
1550 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1551 		*ecode = DDPU_INV_MO;
1552 		break;
1553 	default:
1554 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
1555 		*ecode = 0;
1556 		break;
1557 	}
1558 }
1559 
1560 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
1561 			   gfp_t gfp)
1562 {
1563 	struct fw_ri_wr *wqe;
1564 	struct sk_buff *skb;
1565 	struct terminate_message *term;
1566 
1567 	pr_debug("qhp %p qid 0x%x tid %u\n", qhp, qhp->wq.sq.qid,
1568 		 qhp->ep->hwtid);
1569 
1570 	skb = skb_dequeue(&qhp->ep->com.ep_skb_list);
1571 	if (WARN_ON(!skb))
1572 		return;
1573 
1574 	set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1575 
1576 	wqe = __skb_put_zero(skb, sizeof(*wqe));
1577 	wqe->op_compl = cpu_to_be32(FW_WR_OP_V(FW_RI_INIT_WR));
1578 	wqe->flowid_len16 = cpu_to_be32(
1579 		FW_WR_FLOWID_V(qhp->ep->hwtid) |
1580 		FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1581 
1582 	wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
1583 	wqe->u.terminate.immdlen = cpu_to_be32(sizeof(*term));
1584 	term = (struct terminate_message *)wqe->u.terminate.termmsg;
1585 	if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
1586 		term->layer_etype = qhp->attr.layer_etype;
1587 		term->ecode = qhp->attr.ecode;
1588 	} else
1589 		build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
1590 	c4iw_ofld_send(&qhp->rhp->rdev, skb);
1591 }
1592 
1593 /*
1594  * Assumes qhp lock is held.
1595  */
1596 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
1597 		       struct c4iw_cq *schp)
1598 {
1599 	int count;
1600 	int rq_flushed = 0, sq_flushed;
1601 	unsigned long flag;
1602 
1603 	pr_debug("qhp %p rchp %p schp %p\n", qhp, rchp, schp);
1604 
1605 	/* locking hierarchy: cqs lock first, then qp lock. */
1606 	spin_lock_irqsave(&rchp->lock, flag);
1607 	if (schp != rchp)
1608 		spin_lock(&schp->lock);
1609 	spin_lock(&qhp->lock);
1610 
1611 	if (qhp->wq.flushed) {
1612 		spin_unlock(&qhp->lock);
1613 		if (schp != rchp)
1614 			spin_unlock(&schp->lock);
1615 		spin_unlock_irqrestore(&rchp->lock, flag);
1616 		return;
1617 	}
1618 	qhp->wq.flushed = 1;
1619 	t4_set_wq_in_error(&qhp->wq, 0);
1620 
1621 	c4iw_flush_hw_cq(rchp, qhp);
1622 	if (!qhp->srq) {
1623 		c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
1624 		rq_flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
1625 	}
1626 
1627 	if (schp != rchp)
1628 		c4iw_flush_hw_cq(schp, qhp);
1629 	sq_flushed = c4iw_flush_sq(qhp);
1630 
1631 	spin_unlock(&qhp->lock);
1632 	if (schp != rchp)
1633 		spin_unlock(&schp->lock);
1634 	spin_unlock_irqrestore(&rchp->lock, flag);
1635 
1636 	if (schp == rchp) {
1637 		if ((rq_flushed || sq_flushed) &&
1638 		    t4_clear_cq_armed(&rchp->cq)) {
1639 			spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1640 			(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1641 						   rchp->ibcq.cq_context);
1642 			spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1643 		}
1644 	} else {
1645 		if (rq_flushed && t4_clear_cq_armed(&rchp->cq)) {
1646 			spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1647 			(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1648 						   rchp->ibcq.cq_context);
1649 			spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1650 		}
1651 		if (sq_flushed && t4_clear_cq_armed(&schp->cq)) {
1652 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1653 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1654 						   schp->ibcq.cq_context);
1655 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1656 		}
1657 	}
1658 }
1659 
1660 static void flush_qp(struct c4iw_qp *qhp)
1661 {
1662 	struct c4iw_cq *rchp, *schp;
1663 	unsigned long flag;
1664 
1665 	rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
1666 	schp = to_c4iw_cq(qhp->ibqp.send_cq);
1667 
1668 	if (qhp->ibqp.uobject) {
1669 
1670 		/* for user qps, qhp->wq.flushed is protected by qhp->mutex */
1671 		if (qhp->wq.flushed)
1672 			return;
1673 
1674 		qhp->wq.flushed = 1;
1675 		t4_set_wq_in_error(&qhp->wq, 0);
1676 		t4_set_cq_in_error(&rchp->cq);
1677 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1678 		(*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1679 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1680 		if (schp != rchp) {
1681 			t4_set_cq_in_error(&schp->cq);
1682 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1683 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1684 					schp->ibcq.cq_context);
1685 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1686 		}
1687 		return;
1688 	}
1689 	__flush_qp(qhp, rchp, schp);
1690 }
1691 
1692 static int rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1693 		     struct c4iw_ep *ep)
1694 {
1695 	struct fw_ri_wr *wqe;
1696 	int ret;
1697 	struct sk_buff *skb;
1698 
1699 	pr_debug("qhp %p qid 0x%x tid %u\n", qhp, qhp->wq.sq.qid, ep->hwtid);
1700 
1701 	skb = skb_dequeue(&ep->com.ep_skb_list);
1702 	if (WARN_ON(!skb))
1703 		return -ENOMEM;
1704 
1705 	set_wr_txq(skb, CPL_PRIORITY_DATA, ep->txq_idx);
1706 
1707 	wqe = __skb_put_zero(skb, sizeof(*wqe));
1708 	wqe->op_compl = cpu_to_be32(
1709 		FW_WR_OP_V(FW_RI_INIT_WR) |
1710 		FW_WR_COMPL_F);
1711 	wqe->flowid_len16 = cpu_to_be32(
1712 		FW_WR_FLOWID_V(ep->hwtid) |
1713 		FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1714 	wqe->cookie = (uintptr_t)ep->com.wr_waitp;
1715 
1716 	wqe->u.fini.type = FW_RI_TYPE_FINI;
1717 
1718 	ret = c4iw_ref_send_wait(&rhp->rdev, skb, ep->com.wr_waitp,
1719 				 qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
1720 
1721 	pr_debug("ret %d\n", ret);
1722 	return ret;
1723 }
1724 
1725 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1726 {
1727 	pr_debug("p2p_type = %d\n", p2p_type);
1728 	memset(&init->u, 0, sizeof(init->u));
1729 	switch (p2p_type) {
1730 	case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1731 		init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1732 		init->u.write.stag_sink = cpu_to_be32(1);
1733 		init->u.write.to_sink = cpu_to_be64(1);
1734 		init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1735 		init->u.write.len16 = DIV_ROUND_UP(
1736 			sizeof(init->u.write) + sizeof(struct fw_ri_immd), 16);
1737 		break;
1738 	case FW_RI_INIT_P2PTYPE_READ_REQ:
1739 		init->u.write.opcode = FW_RI_RDMA_READ_WR;
1740 		init->u.read.stag_src = cpu_to_be32(1);
1741 		init->u.read.to_src_lo = cpu_to_be32(1);
1742 		init->u.read.stag_sink = cpu_to_be32(1);
1743 		init->u.read.to_sink_lo = cpu_to_be32(1);
1744 		init->u.read.len16 = DIV_ROUND_UP(sizeof(init->u.read), 16);
1745 		break;
1746 	}
1747 }
1748 
1749 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1750 {
1751 	struct fw_ri_wr *wqe;
1752 	int ret;
1753 	struct sk_buff *skb;
1754 
1755 	pr_debug("qhp %p qid 0x%x tid %u ird %u ord %u\n", qhp,
1756 		 qhp->wq.sq.qid, qhp->ep->hwtid, qhp->ep->ird, qhp->ep->ord);
1757 
1758 	skb = alloc_skb(sizeof(*wqe), GFP_KERNEL);
1759 	if (!skb) {
1760 		ret = -ENOMEM;
1761 		goto out;
1762 	}
1763 	ret = alloc_ird(rhp, qhp->attr.max_ird);
1764 	if (ret) {
1765 		qhp->attr.max_ird = 0;
1766 		kfree_skb(skb);
1767 		goto out;
1768 	}
1769 	set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1770 
1771 	wqe = __skb_put_zero(skb, sizeof(*wqe));
1772 	wqe->op_compl = cpu_to_be32(
1773 		FW_WR_OP_V(FW_RI_INIT_WR) |
1774 		FW_WR_COMPL_F);
1775 	wqe->flowid_len16 = cpu_to_be32(
1776 		FW_WR_FLOWID_V(qhp->ep->hwtid) |
1777 		FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1778 
1779 	wqe->cookie = (uintptr_t)qhp->ep->com.wr_waitp;
1780 
1781 	wqe->u.init.type = FW_RI_TYPE_INIT;
1782 	wqe->u.init.mpareqbit_p2ptype =
1783 		FW_RI_WR_MPAREQBIT_V(qhp->attr.mpa_attr.initiator) |
1784 		FW_RI_WR_P2PTYPE_V(qhp->attr.mpa_attr.p2p_type);
1785 	wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1786 	if (qhp->attr.mpa_attr.recv_marker_enabled)
1787 		wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1788 	if (qhp->attr.mpa_attr.xmit_marker_enabled)
1789 		wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1790 	if (qhp->attr.mpa_attr.crc_enabled)
1791 		wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1792 
1793 	wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1794 			    FW_RI_QP_RDMA_WRITE_ENABLE |
1795 			    FW_RI_QP_BIND_ENABLE;
1796 	if (!qhp->ibqp.uobject)
1797 		wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1798 				     FW_RI_QP_STAG0_ENABLE;
1799 	wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1800 	wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1801 	wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1802 	wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1803 	if (qhp->srq) {
1804 		wqe->u.init.rq_eqid = cpu_to_be32(FW_RI_INIT_RQEQID_SRQ |
1805 						  qhp->srq->idx);
1806 	} else {
1807 		wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1808 		wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1809 		wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1810 						   rhp->rdev.lldi.vr->rq.start);
1811 	}
1812 	wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1813 	wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1814 	wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1815 	wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1816 	wqe->u.init.iss = cpu_to_be32(qhp->ep->snd_seq);
1817 	wqe->u.init.irs = cpu_to_be32(qhp->ep->rcv_seq);
1818 	if (qhp->attr.mpa_attr.initiator)
1819 		build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1820 
1821 	ret = c4iw_ref_send_wait(&rhp->rdev, skb, qhp->ep->com.wr_waitp,
1822 				 qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
1823 	if (!ret)
1824 		goto out;
1825 
1826 	free_ird(rhp, qhp->attr.max_ird);
1827 out:
1828 	pr_debug("ret %d\n", ret);
1829 	return ret;
1830 }
1831 
1832 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1833 		   enum c4iw_qp_attr_mask mask,
1834 		   struct c4iw_qp_attributes *attrs,
1835 		   int internal)
1836 {
1837 	int ret = 0;
1838 	struct c4iw_qp_attributes newattr = qhp->attr;
1839 	int disconnect = 0;
1840 	int terminate = 0;
1841 	int abort = 0;
1842 	int free = 0;
1843 	struct c4iw_ep *ep = NULL;
1844 
1845 	pr_debug("qhp %p sqid 0x%x rqid 0x%x ep %p state %d -> %d\n",
1846 		 qhp, qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep, qhp->attr.state,
1847 		 (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1848 
1849 	mutex_lock(&qhp->mutex);
1850 
1851 	/* Process attr changes if in IDLE */
1852 	if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1853 		if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1854 			ret = -EIO;
1855 			goto out;
1856 		}
1857 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1858 			newattr.enable_rdma_read = attrs->enable_rdma_read;
1859 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1860 			newattr.enable_rdma_write = attrs->enable_rdma_write;
1861 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1862 			newattr.enable_bind = attrs->enable_bind;
1863 		if (mask & C4IW_QP_ATTR_MAX_ORD) {
1864 			if (attrs->max_ord > c4iw_max_read_depth) {
1865 				ret = -EINVAL;
1866 				goto out;
1867 			}
1868 			newattr.max_ord = attrs->max_ord;
1869 		}
1870 		if (mask & C4IW_QP_ATTR_MAX_IRD) {
1871 			if (attrs->max_ird > cur_max_read_depth(rhp)) {
1872 				ret = -EINVAL;
1873 				goto out;
1874 			}
1875 			newattr.max_ird = attrs->max_ird;
1876 		}
1877 		qhp->attr = newattr;
1878 	}
1879 
1880 	if (mask & C4IW_QP_ATTR_SQ_DB) {
1881 		ret = ring_kernel_sq_db(qhp, attrs->sq_db_inc);
1882 		goto out;
1883 	}
1884 	if (mask & C4IW_QP_ATTR_RQ_DB) {
1885 		ret = ring_kernel_rq_db(qhp, attrs->rq_db_inc);
1886 		goto out;
1887 	}
1888 
1889 	if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1890 		goto out;
1891 	if (qhp->attr.state == attrs->next_state)
1892 		goto out;
1893 
1894 	switch (qhp->attr.state) {
1895 	case C4IW_QP_STATE_IDLE:
1896 		switch (attrs->next_state) {
1897 		case C4IW_QP_STATE_RTS:
1898 			if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1899 				ret = -EINVAL;
1900 				goto out;
1901 			}
1902 			if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1903 				ret = -EINVAL;
1904 				goto out;
1905 			}
1906 			qhp->attr.mpa_attr = attrs->mpa_attr;
1907 			qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1908 			qhp->ep = qhp->attr.llp_stream_handle;
1909 			set_state(qhp, C4IW_QP_STATE_RTS);
1910 
1911 			/*
1912 			 * Ref the endpoint here and deref when we
1913 			 * disassociate the endpoint from the QP.  This
1914 			 * happens in CLOSING->IDLE transition or *->ERROR
1915 			 * transition.
1916 			 */
1917 			c4iw_get_ep(&qhp->ep->com);
1918 			ret = rdma_init(rhp, qhp);
1919 			if (ret)
1920 				goto err;
1921 			break;
1922 		case C4IW_QP_STATE_ERROR:
1923 			set_state(qhp, C4IW_QP_STATE_ERROR);
1924 			flush_qp(qhp);
1925 			break;
1926 		default:
1927 			ret = -EINVAL;
1928 			goto out;
1929 		}
1930 		break;
1931 	case C4IW_QP_STATE_RTS:
1932 		switch (attrs->next_state) {
1933 		case C4IW_QP_STATE_CLOSING:
1934 			t4_set_wq_in_error(&qhp->wq, 0);
1935 			set_state(qhp, C4IW_QP_STATE_CLOSING);
1936 			ep = qhp->ep;
1937 			if (!internal) {
1938 				abort = 0;
1939 				disconnect = 1;
1940 				c4iw_get_ep(&qhp->ep->com);
1941 			}
1942 			ret = rdma_fini(rhp, qhp, ep);
1943 			if (ret)
1944 				goto err;
1945 			break;
1946 		case C4IW_QP_STATE_TERMINATE:
1947 			t4_set_wq_in_error(&qhp->wq, 0);
1948 			set_state(qhp, C4IW_QP_STATE_TERMINATE);
1949 			qhp->attr.layer_etype = attrs->layer_etype;
1950 			qhp->attr.ecode = attrs->ecode;
1951 			ep = qhp->ep;
1952 			if (!internal) {
1953 				c4iw_get_ep(&ep->com);
1954 				terminate = 1;
1955 				disconnect = 1;
1956 			} else {
1957 				terminate = qhp->attr.send_term;
1958 				ret = rdma_fini(rhp, qhp, ep);
1959 				if (ret)
1960 					goto err;
1961 			}
1962 			break;
1963 		case C4IW_QP_STATE_ERROR:
1964 			t4_set_wq_in_error(&qhp->wq, 0);
1965 			set_state(qhp, C4IW_QP_STATE_ERROR);
1966 			if (!internal) {
1967 				disconnect = 1;
1968 				ep = qhp->ep;
1969 				c4iw_get_ep(&qhp->ep->com);
1970 			}
1971 			goto err;
1972 			break;
1973 		default:
1974 			ret = -EINVAL;
1975 			goto out;
1976 		}
1977 		break;
1978 	case C4IW_QP_STATE_CLOSING:
1979 
1980 		/*
1981 		 * Allow kernel users to move to ERROR for qp draining.
1982 		 */
1983 		if (!internal && (qhp->ibqp.uobject || attrs->next_state !=
1984 				  C4IW_QP_STATE_ERROR)) {
1985 			ret = -EINVAL;
1986 			goto out;
1987 		}
1988 		switch (attrs->next_state) {
1989 		case C4IW_QP_STATE_IDLE:
1990 			flush_qp(qhp);
1991 			set_state(qhp, C4IW_QP_STATE_IDLE);
1992 			qhp->attr.llp_stream_handle = NULL;
1993 			c4iw_put_ep(&qhp->ep->com);
1994 			qhp->ep = NULL;
1995 			wake_up(&qhp->wait);
1996 			break;
1997 		case C4IW_QP_STATE_ERROR:
1998 			goto err;
1999 		default:
2000 			ret = -EINVAL;
2001 			goto err;
2002 		}
2003 		break;
2004 	case C4IW_QP_STATE_ERROR:
2005 		if (attrs->next_state != C4IW_QP_STATE_IDLE) {
2006 			ret = -EINVAL;
2007 			goto out;
2008 		}
2009 		if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) {
2010 			ret = -EINVAL;
2011 			goto out;
2012 		}
2013 		set_state(qhp, C4IW_QP_STATE_IDLE);
2014 		break;
2015 	case C4IW_QP_STATE_TERMINATE:
2016 		if (!internal) {
2017 			ret = -EINVAL;
2018 			goto out;
2019 		}
2020 		goto err;
2021 		break;
2022 	default:
2023 		pr_err("%s in a bad state %d\n", __func__, qhp->attr.state);
2024 		ret = -EINVAL;
2025 		goto err;
2026 		break;
2027 	}
2028 	goto out;
2029 err:
2030 	pr_debug("disassociating ep %p qpid 0x%x\n", qhp->ep,
2031 		 qhp->wq.sq.qid);
2032 
2033 	/* disassociate the LLP connection */
2034 	qhp->attr.llp_stream_handle = NULL;
2035 	if (!ep)
2036 		ep = qhp->ep;
2037 	qhp->ep = NULL;
2038 	set_state(qhp, C4IW_QP_STATE_ERROR);
2039 	free = 1;
2040 	abort = 1;
2041 	flush_qp(qhp);
2042 	wake_up(&qhp->wait);
2043 out:
2044 	mutex_unlock(&qhp->mutex);
2045 
2046 	if (terminate)
2047 		post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
2048 
2049 	/*
2050 	 * If disconnect is 1, then we need to initiate a disconnect
2051 	 * on the EP.  This can be a normal close (RTS->CLOSING) or
2052 	 * an abnormal close (RTS/CLOSING->ERROR).
2053 	 */
2054 	if (disconnect) {
2055 		c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC :
2056 							 GFP_KERNEL);
2057 		c4iw_put_ep(&ep->com);
2058 	}
2059 
2060 	/*
2061 	 * If free is 1, then we've disassociated the EP from the QP
2062 	 * and we need to dereference the EP.
2063 	 */
2064 	if (free)
2065 		c4iw_put_ep(&ep->com);
2066 	pr_debug("exit state %d\n", qhp->attr.state);
2067 	return ret;
2068 }
2069 
2070 int c4iw_destroy_qp(struct ib_qp *ib_qp, struct ib_udata *udata)
2071 {
2072 	struct c4iw_dev *rhp;
2073 	struct c4iw_qp *qhp;
2074 	struct c4iw_ucontext *ucontext;
2075 	struct c4iw_qp_attributes attrs;
2076 
2077 	qhp = to_c4iw_qp(ib_qp);
2078 	rhp = qhp->rhp;
2079 	ucontext = qhp->ucontext;
2080 
2081 	attrs.next_state = C4IW_QP_STATE_ERROR;
2082 	if (qhp->attr.state == C4IW_QP_STATE_TERMINATE)
2083 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
2084 	else
2085 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
2086 	wait_event(qhp->wait, !qhp->ep);
2087 
2088 	xa_lock_irq(&rhp->qps);
2089 	__xa_erase(&rhp->qps, qhp->wq.sq.qid);
2090 	if (!list_empty(&qhp->db_fc_entry))
2091 		list_del_init(&qhp->db_fc_entry);
2092 	xa_unlock_irq(&rhp->qps);
2093 	free_ird(rhp, qhp->attr.max_ird);
2094 
2095 	c4iw_qp_rem_ref(ib_qp);
2096 
2097 	wait_for_completion(&qhp->qp_rel_comp);
2098 
2099 	pr_debug("ib_qp %p qpid 0x%0x\n", ib_qp, qhp->wq.sq.qid);
2100 	pr_debug("qhp %p ucontext %p\n", qhp, ucontext);
2101 
2102 	destroy_qp(&rhp->rdev, &qhp->wq,
2103 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx, !qhp->srq);
2104 
2105 	c4iw_put_wr_wait(qhp->wr_waitp);
2106 	return 0;
2107 }
2108 
2109 int c4iw_create_qp(struct ib_qp *qp, struct ib_qp_init_attr *attrs,
2110 		   struct ib_udata *udata)
2111 {
2112 	struct ib_pd *pd = qp->pd;
2113 	struct c4iw_dev *rhp;
2114 	struct c4iw_qp *qhp = to_c4iw_qp(qp);
2115 	struct c4iw_pd *php;
2116 	struct c4iw_cq *schp;
2117 	struct c4iw_cq *rchp;
2118 	struct c4iw_create_qp_resp uresp;
2119 	unsigned int sqsize, rqsize = 0;
2120 	struct c4iw_ucontext *ucontext = rdma_udata_to_drv_context(
2121 		udata, struct c4iw_ucontext, ibucontext);
2122 	int ret;
2123 	struct c4iw_mm_entry *sq_key_mm, *rq_key_mm = NULL, *sq_db_key_mm;
2124 	struct c4iw_mm_entry *rq_db_key_mm = NULL, *ma_sync_key_mm = NULL;
2125 
2126 	if (attrs->qp_type != IB_QPT_RC || attrs->create_flags)
2127 		return -EOPNOTSUPP;
2128 
2129 	php = to_c4iw_pd(pd);
2130 	rhp = php->rhp;
2131 	schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
2132 	rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
2133 	if (!schp || !rchp)
2134 		return -EINVAL;
2135 
2136 	if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
2137 		return -EINVAL;
2138 
2139 	if (!attrs->srq) {
2140 		if (attrs->cap.max_recv_wr > rhp->rdev.hw_queue.t4_max_rq_size)
2141 			return -E2BIG;
2142 		rqsize = attrs->cap.max_recv_wr + 1;
2143 		if (rqsize < 8)
2144 			rqsize = 8;
2145 	}
2146 
2147 	if (attrs->cap.max_send_wr > rhp->rdev.hw_queue.t4_max_sq_size)
2148 		return -E2BIG;
2149 	sqsize = attrs->cap.max_send_wr + 1;
2150 	if (sqsize < 8)
2151 		sqsize = 8;
2152 
2153 	qhp->wr_waitp = c4iw_alloc_wr_wait(GFP_KERNEL);
2154 	if (!qhp->wr_waitp)
2155 		return -ENOMEM;
2156 
2157 	qhp->wq.sq.size = sqsize;
2158 	qhp->wq.sq.memsize =
2159 		(sqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
2160 		sizeof(*qhp->wq.sq.queue) + 16 * sizeof(__be64);
2161 	qhp->wq.sq.flush_cidx = -1;
2162 	if (!attrs->srq) {
2163 		qhp->wq.rq.size = rqsize;
2164 		qhp->wq.rq.memsize =
2165 			(rqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
2166 			sizeof(*qhp->wq.rq.queue);
2167 	}
2168 
2169 	if (ucontext) {
2170 		qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
2171 		if (!attrs->srq)
2172 			qhp->wq.rq.memsize =
2173 				roundup(qhp->wq.rq.memsize, PAGE_SIZE);
2174 	}
2175 
2176 	ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
2177 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx,
2178 			qhp->wr_waitp, !attrs->srq);
2179 	if (ret)
2180 		goto err_free_wr_wait;
2181 
2182 	attrs->cap.max_recv_wr = rqsize - 1;
2183 	attrs->cap.max_send_wr = sqsize - 1;
2184 	attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
2185 
2186 	qhp->rhp = rhp;
2187 	qhp->attr.pd = php->pdid;
2188 	qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
2189 	qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
2190 	qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
2191 	qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
2192 	qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
2193 	if (!attrs->srq) {
2194 		qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
2195 		qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
2196 	}
2197 	qhp->attr.state = C4IW_QP_STATE_IDLE;
2198 	qhp->attr.next_state = C4IW_QP_STATE_IDLE;
2199 	qhp->attr.enable_rdma_read = 1;
2200 	qhp->attr.enable_rdma_write = 1;
2201 	qhp->attr.enable_bind = 1;
2202 	qhp->attr.max_ord = 0;
2203 	qhp->attr.max_ird = 0;
2204 	qhp->sq_sig_all = attrs->sq_sig_type == IB_SIGNAL_ALL_WR;
2205 	spin_lock_init(&qhp->lock);
2206 	mutex_init(&qhp->mutex);
2207 	init_waitqueue_head(&qhp->wait);
2208 	init_completion(&qhp->qp_rel_comp);
2209 	refcount_set(&qhp->qp_refcnt, 1);
2210 
2211 	ret = xa_insert_irq(&rhp->qps, qhp->wq.sq.qid, qhp, GFP_KERNEL);
2212 	if (ret)
2213 		goto err_destroy_qp;
2214 
2215 	if (udata && ucontext) {
2216 		sq_key_mm = kmalloc(sizeof(*sq_key_mm), GFP_KERNEL);
2217 		if (!sq_key_mm) {
2218 			ret = -ENOMEM;
2219 			goto err_remove_handle;
2220 		}
2221 		if (!attrs->srq) {
2222 			rq_key_mm = kmalloc(sizeof(*rq_key_mm), GFP_KERNEL);
2223 			if (!rq_key_mm) {
2224 				ret = -ENOMEM;
2225 				goto err_free_sq_key;
2226 			}
2227 		}
2228 		sq_db_key_mm = kmalloc(sizeof(*sq_db_key_mm), GFP_KERNEL);
2229 		if (!sq_db_key_mm) {
2230 			ret = -ENOMEM;
2231 			goto err_free_rq_key;
2232 		}
2233 		if (!attrs->srq) {
2234 			rq_db_key_mm =
2235 				kmalloc(sizeof(*rq_db_key_mm), GFP_KERNEL);
2236 			if (!rq_db_key_mm) {
2237 				ret = -ENOMEM;
2238 				goto err_free_sq_db_key;
2239 			}
2240 		}
2241 		memset(&uresp, 0, sizeof(uresp));
2242 		if (t4_sq_onchip(&qhp->wq.sq)) {
2243 			ma_sync_key_mm = kmalloc(sizeof(*ma_sync_key_mm),
2244 						 GFP_KERNEL);
2245 			if (!ma_sync_key_mm) {
2246 				ret = -ENOMEM;
2247 				goto err_free_rq_db_key;
2248 			}
2249 			uresp.flags = C4IW_QPF_ONCHIP;
2250 		}
2251 		if (rhp->rdev.lldi.write_w_imm_support)
2252 			uresp.flags |= C4IW_QPF_WRITE_W_IMM;
2253 		uresp.qid_mask = rhp->rdev.qpmask;
2254 		uresp.sqid = qhp->wq.sq.qid;
2255 		uresp.sq_size = qhp->wq.sq.size;
2256 		uresp.sq_memsize = qhp->wq.sq.memsize;
2257 		if (!attrs->srq) {
2258 			uresp.rqid = qhp->wq.rq.qid;
2259 			uresp.rq_size = qhp->wq.rq.size;
2260 			uresp.rq_memsize = qhp->wq.rq.memsize;
2261 		}
2262 		spin_lock(&ucontext->mmap_lock);
2263 		if (ma_sync_key_mm) {
2264 			uresp.ma_sync_key = ucontext->key;
2265 			ucontext->key += PAGE_SIZE;
2266 		}
2267 		uresp.sq_key = ucontext->key;
2268 		ucontext->key += PAGE_SIZE;
2269 		if (!attrs->srq) {
2270 			uresp.rq_key = ucontext->key;
2271 			ucontext->key += PAGE_SIZE;
2272 		}
2273 		uresp.sq_db_gts_key = ucontext->key;
2274 		ucontext->key += PAGE_SIZE;
2275 		if (!attrs->srq) {
2276 			uresp.rq_db_gts_key = ucontext->key;
2277 			ucontext->key += PAGE_SIZE;
2278 		}
2279 		spin_unlock(&ucontext->mmap_lock);
2280 		ret = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
2281 		if (ret)
2282 			goto err_free_ma_sync_key;
2283 		sq_key_mm->key = uresp.sq_key;
2284 		sq_key_mm->addr = 0;
2285 		sq_key_mm->vaddr = qhp->wq.sq.queue;
2286 		sq_key_mm->dma_addr = qhp->wq.sq.dma_addr;
2287 		sq_key_mm->len = PAGE_ALIGN(qhp->wq.sq.memsize);
2288 		insert_flag_to_mmap(&rhp->rdev, sq_key_mm, sq_key_mm->addr);
2289 		insert_mmap(ucontext, sq_key_mm);
2290 		if (!attrs->srq) {
2291 			rq_key_mm->key = uresp.rq_key;
2292 			rq_key_mm->addr = 0;
2293 			rq_key_mm->vaddr = qhp->wq.rq.queue;
2294 			rq_key_mm->dma_addr = qhp->wq.rq.dma_addr;
2295 			rq_key_mm->len = PAGE_ALIGN(qhp->wq.rq.memsize);
2296 			insert_flag_to_mmap(&rhp->rdev, rq_key_mm,
2297 					    rq_key_mm->addr);
2298 			insert_mmap(ucontext, rq_key_mm);
2299 		}
2300 		sq_db_key_mm->key = uresp.sq_db_gts_key;
2301 		sq_db_key_mm->addr = (u64)(unsigned long)qhp->wq.sq.bar2_pa;
2302 		sq_db_key_mm->vaddr = NULL;
2303 		sq_db_key_mm->dma_addr = 0;
2304 		sq_db_key_mm->len = PAGE_SIZE;
2305 		insert_flag_to_mmap(&rhp->rdev, sq_db_key_mm,
2306 				    sq_db_key_mm->addr);
2307 		insert_mmap(ucontext, sq_db_key_mm);
2308 		if (!attrs->srq) {
2309 			rq_db_key_mm->key = uresp.rq_db_gts_key;
2310 			rq_db_key_mm->addr =
2311 				(u64)(unsigned long)qhp->wq.rq.bar2_pa;
2312 			rq_db_key_mm->len = PAGE_SIZE;
2313 			rq_db_key_mm->vaddr = NULL;
2314 			rq_db_key_mm->dma_addr = 0;
2315 			insert_flag_to_mmap(&rhp->rdev, rq_db_key_mm,
2316 					    rq_db_key_mm->addr);
2317 			insert_mmap(ucontext, rq_db_key_mm);
2318 		}
2319 		if (ma_sync_key_mm) {
2320 			ma_sync_key_mm->key = uresp.ma_sync_key;
2321 			ma_sync_key_mm->addr =
2322 				(pci_resource_start(rhp->rdev.lldi.pdev, 0) +
2323 				PCIE_MA_SYNC_A) & PAGE_MASK;
2324 			ma_sync_key_mm->len = PAGE_SIZE;
2325 			ma_sync_key_mm->vaddr = NULL;
2326 			ma_sync_key_mm->dma_addr = 0;
2327 			insert_flag_to_mmap(&rhp->rdev, ma_sync_key_mm,
2328 					    ma_sync_key_mm->addr);
2329 			insert_mmap(ucontext, ma_sync_key_mm);
2330 		}
2331 
2332 		qhp->ucontext = ucontext;
2333 	}
2334 	if (!attrs->srq) {
2335 		qhp->wq.qp_errp =
2336 			&qhp->wq.rq.queue[qhp->wq.rq.size].status.qp_err;
2337 	} else {
2338 		qhp->wq.qp_errp =
2339 			&qhp->wq.sq.queue[qhp->wq.sq.size].status.qp_err;
2340 		qhp->wq.srqidxp =
2341 			&qhp->wq.sq.queue[qhp->wq.sq.size].status.srqidx;
2342 	}
2343 
2344 	qhp->ibqp.qp_num = qhp->wq.sq.qid;
2345 	if (attrs->srq)
2346 		qhp->srq = to_c4iw_srq(attrs->srq);
2347 	INIT_LIST_HEAD(&qhp->db_fc_entry);
2348 	pr_debug("sq id %u size %u memsize %zu num_entries %u rq id %u size %u memsize %zu num_entries %u\n",
2349 		 qhp->wq.sq.qid, qhp->wq.sq.size, qhp->wq.sq.memsize,
2350 		 attrs->cap.max_send_wr, qhp->wq.rq.qid, qhp->wq.rq.size,
2351 		 qhp->wq.rq.memsize, attrs->cap.max_recv_wr);
2352 	return 0;
2353 err_free_ma_sync_key:
2354 	kfree(ma_sync_key_mm);
2355 err_free_rq_db_key:
2356 	if (!attrs->srq)
2357 		kfree(rq_db_key_mm);
2358 err_free_sq_db_key:
2359 	kfree(sq_db_key_mm);
2360 err_free_rq_key:
2361 	if (!attrs->srq)
2362 		kfree(rq_key_mm);
2363 err_free_sq_key:
2364 	kfree(sq_key_mm);
2365 err_remove_handle:
2366 	xa_erase_irq(&rhp->qps, qhp->wq.sq.qid);
2367 err_destroy_qp:
2368 	destroy_qp(&rhp->rdev, &qhp->wq,
2369 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx, !attrs->srq);
2370 err_free_wr_wait:
2371 	c4iw_put_wr_wait(qhp->wr_waitp);
2372 	return ret;
2373 }
2374 
2375 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2376 		      int attr_mask, struct ib_udata *udata)
2377 {
2378 	struct c4iw_dev *rhp;
2379 	struct c4iw_qp *qhp;
2380 	enum c4iw_qp_attr_mask mask = 0;
2381 	struct c4iw_qp_attributes attrs = {};
2382 
2383 	pr_debug("ib_qp %p\n", ibqp);
2384 
2385 	if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
2386 		return -EOPNOTSUPP;
2387 
2388 	/* iwarp does not support the RTR state */
2389 	if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
2390 		attr_mask &= ~IB_QP_STATE;
2391 
2392 	/* Make sure we still have something left to do */
2393 	if (!attr_mask)
2394 		return 0;
2395 
2396 	qhp = to_c4iw_qp(ibqp);
2397 	rhp = qhp->rhp;
2398 
2399 	attrs.next_state = c4iw_convert_state(attr->qp_state);
2400 	attrs.enable_rdma_read = (attr->qp_access_flags &
2401 			       IB_ACCESS_REMOTE_READ) ?  1 : 0;
2402 	attrs.enable_rdma_write = (attr->qp_access_flags &
2403 				IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
2404 	attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
2405 
2406 
2407 	mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
2408 	mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
2409 			(C4IW_QP_ATTR_ENABLE_RDMA_READ |
2410 			 C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
2411 			 C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
2412 
2413 	/*
2414 	 * Use SQ_PSN and RQ_PSN to pass in IDX_INC values for
2415 	 * ringing the queue db when we're in DB_FULL mode.
2416 	 * Only allow this on T4 devices.
2417 	 */
2418 	attrs.sq_db_inc = attr->sq_psn;
2419 	attrs.rq_db_inc = attr->rq_psn;
2420 	mask |= (attr_mask & IB_QP_SQ_PSN) ? C4IW_QP_ATTR_SQ_DB : 0;
2421 	mask |= (attr_mask & IB_QP_RQ_PSN) ? C4IW_QP_ATTR_RQ_DB : 0;
2422 	if (!is_t4(to_c4iw_qp(ibqp)->rhp->rdev.lldi.adapter_type) &&
2423 	    (mask & (C4IW_QP_ATTR_SQ_DB|C4IW_QP_ATTR_RQ_DB)))
2424 		return -EINVAL;
2425 
2426 	return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
2427 }
2428 
2429 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
2430 {
2431 	pr_debug("ib_dev %p qpn 0x%x\n", dev, qpn);
2432 	return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
2433 }
2434 
2435 void c4iw_dispatch_srq_limit_reached_event(struct c4iw_srq *srq)
2436 {
2437 	struct ib_event event = {};
2438 
2439 	event.device = &srq->rhp->ibdev;
2440 	event.element.srq = &srq->ibsrq;
2441 	event.event = IB_EVENT_SRQ_LIMIT_REACHED;
2442 	ib_dispatch_event(&event);
2443 }
2444 
2445 int c4iw_modify_srq(struct ib_srq *ib_srq, struct ib_srq_attr *attr,
2446 		    enum ib_srq_attr_mask srq_attr_mask,
2447 		    struct ib_udata *udata)
2448 {
2449 	struct c4iw_srq *srq = to_c4iw_srq(ib_srq);
2450 	int ret = 0;
2451 
2452 	/*
2453 	 * XXX 0 mask == a SW interrupt for srq_limit reached...
2454 	 */
2455 	if (udata && !srq_attr_mask) {
2456 		c4iw_dispatch_srq_limit_reached_event(srq);
2457 		goto out;
2458 	}
2459 
2460 	/* no support for this yet */
2461 	if (srq_attr_mask & IB_SRQ_MAX_WR) {
2462 		ret = -EINVAL;
2463 		goto out;
2464 	}
2465 
2466 	if (!udata && (srq_attr_mask & IB_SRQ_LIMIT)) {
2467 		srq->armed = true;
2468 		srq->srq_limit = attr->srq_limit;
2469 	}
2470 out:
2471 	return ret;
2472 }
2473 
2474 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2475 		     int attr_mask, struct ib_qp_init_attr *init_attr)
2476 {
2477 	struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
2478 
2479 	memset(attr, 0, sizeof(*attr));
2480 	memset(init_attr, 0, sizeof(*init_attr));
2481 	attr->qp_state = to_ib_qp_state(qhp->attr.state);
2482 	attr->cur_qp_state = to_ib_qp_state(qhp->attr.state);
2483 	init_attr->cap.max_send_wr = qhp->attr.sq_num_entries;
2484 	init_attr->cap.max_recv_wr = qhp->attr.rq_num_entries;
2485 	init_attr->cap.max_send_sge = qhp->attr.sq_max_sges;
2486 	init_attr->cap.max_recv_sge = qhp->attr.rq_max_sges;
2487 	init_attr->cap.max_inline_data = T4_MAX_SEND_INLINE;
2488 	init_attr->sq_sig_type = qhp->sq_sig_all ? IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
2489 	return 0;
2490 }
2491 
2492 static void free_srq_queue(struct c4iw_srq *srq, struct c4iw_dev_ucontext *uctx,
2493 			   struct c4iw_wr_wait *wr_waitp)
2494 {
2495 	struct c4iw_rdev *rdev = &srq->rhp->rdev;
2496 	struct sk_buff *skb = srq->destroy_skb;
2497 	struct t4_srq *wq = &srq->wq;
2498 	struct fw_ri_res_wr *res_wr;
2499 	struct fw_ri_res *res;
2500 	int wr_len;
2501 
2502 	wr_len = sizeof(*res_wr) + sizeof(*res);
2503 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
2504 
2505 	res_wr = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
2506 	memset(res_wr, 0, wr_len);
2507 	res_wr->op_nres = cpu_to_be32(FW_WR_OP_V(FW_RI_RES_WR) |
2508 			FW_RI_RES_WR_NRES_V(1) |
2509 			FW_WR_COMPL_F);
2510 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
2511 	res_wr->cookie = (uintptr_t)wr_waitp;
2512 	res = res_wr->res;
2513 	res->u.srq.restype = FW_RI_RES_TYPE_SRQ;
2514 	res->u.srq.op = FW_RI_RES_OP_RESET;
2515 	res->u.srq.srqid = cpu_to_be32(srq->idx);
2516 	res->u.srq.eqid = cpu_to_be32(wq->qid);
2517 
2518 	c4iw_init_wr_wait(wr_waitp);
2519 	c4iw_ref_send_wait(rdev, skb, wr_waitp, 0, 0, __func__);
2520 
2521 	dma_free_coherent(&rdev->lldi.pdev->dev,
2522 			  wq->memsize, wq->queue,
2523 			dma_unmap_addr(wq, mapping));
2524 	c4iw_rqtpool_free(rdev, wq->rqt_hwaddr, wq->rqt_size);
2525 	kfree(wq->sw_rq);
2526 	c4iw_put_qpid(rdev, wq->qid, uctx);
2527 }
2528 
2529 static int alloc_srq_queue(struct c4iw_srq *srq, struct c4iw_dev_ucontext *uctx,
2530 			   struct c4iw_wr_wait *wr_waitp)
2531 {
2532 	struct c4iw_rdev *rdev = &srq->rhp->rdev;
2533 	int user = (uctx != &rdev->uctx);
2534 	struct t4_srq *wq = &srq->wq;
2535 	struct fw_ri_res_wr *res_wr;
2536 	struct fw_ri_res *res;
2537 	struct sk_buff *skb;
2538 	int wr_len;
2539 	int eqsize;
2540 	int ret = -ENOMEM;
2541 
2542 	wq->qid = c4iw_get_qpid(rdev, uctx);
2543 	if (!wq->qid)
2544 		goto err;
2545 
2546 	if (!user) {
2547 		wq->sw_rq = kcalloc(wq->size, sizeof(*wq->sw_rq),
2548 				    GFP_KERNEL);
2549 		if (!wq->sw_rq)
2550 			goto err_put_qpid;
2551 		wq->pending_wrs = kcalloc(srq->wq.size,
2552 					  sizeof(*srq->wq.pending_wrs),
2553 					  GFP_KERNEL);
2554 		if (!wq->pending_wrs)
2555 			goto err_free_sw_rq;
2556 	}
2557 
2558 	wq->rqt_size = wq->size;
2559 	wq->rqt_hwaddr = c4iw_rqtpool_alloc(rdev, wq->rqt_size);
2560 	if (!wq->rqt_hwaddr)
2561 		goto err_free_pending_wrs;
2562 	wq->rqt_abs_idx = (wq->rqt_hwaddr - rdev->lldi.vr->rq.start) >>
2563 		T4_RQT_ENTRY_SHIFT;
2564 
2565 	wq->queue = dma_alloc_coherent(&rdev->lldi.pdev->dev, wq->memsize,
2566 				       &wq->dma_addr, GFP_KERNEL);
2567 	if (!wq->queue)
2568 		goto err_free_rqtpool;
2569 
2570 	dma_unmap_addr_set(wq, mapping, wq->dma_addr);
2571 
2572 	wq->bar2_va = c4iw_bar2_addrs(rdev, wq->qid, CXGB4_BAR2_QTYPE_EGRESS,
2573 				      &wq->bar2_qid,
2574 			user ? &wq->bar2_pa : NULL);
2575 
2576 	/*
2577 	 * User mode must have bar2 access.
2578 	 */
2579 
2580 	if (user && !wq->bar2_va) {
2581 		pr_warn(MOD "%s: srqid %u not in BAR2 range.\n",
2582 			pci_name(rdev->lldi.pdev), wq->qid);
2583 		ret = -EINVAL;
2584 		goto err_free_queue;
2585 	}
2586 
2587 	/* build fw_ri_res_wr */
2588 	wr_len = sizeof(*res_wr) + sizeof(*res);
2589 
2590 	skb = alloc_skb(wr_len, GFP_KERNEL);
2591 	if (!skb)
2592 		goto err_free_queue;
2593 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
2594 
2595 	res_wr = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
2596 	memset(res_wr, 0, wr_len);
2597 	res_wr->op_nres = cpu_to_be32(FW_WR_OP_V(FW_RI_RES_WR) |
2598 			FW_RI_RES_WR_NRES_V(1) |
2599 			FW_WR_COMPL_F);
2600 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
2601 	res_wr->cookie = (uintptr_t)wr_waitp;
2602 	res = res_wr->res;
2603 	res->u.srq.restype = FW_RI_RES_TYPE_SRQ;
2604 	res->u.srq.op = FW_RI_RES_OP_WRITE;
2605 
2606 	/*
2607 	 * eqsize is the number of 64B entries plus the status page size.
2608 	 */
2609 	eqsize = wq->size * T4_RQ_NUM_SLOTS +
2610 		rdev->hw_queue.t4_eq_status_entries;
2611 	res->u.srq.eqid = cpu_to_be32(wq->qid);
2612 	res->u.srq.fetchszm_to_iqid =
2613 						/* no host cidx updates */
2614 		cpu_to_be32(FW_RI_RES_WR_HOSTFCMODE_V(0) |
2615 		FW_RI_RES_WR_CPRIO_V(0) |       /* don't keep in chip cache */
2616 		FW_RI_RES_WR_PCIECHN_V(0) |     /* set by uP at ri_init time */
2617 		FW_RI_RES_WR_FETCHRO_V(0));     /* relaxed_ordering */
2618 	res->u.srq.dcaen_to_eqsize =
2619 		cpu_to_be32(FW_RI_RES_WR_DCAEN_V(0) |
2620 		FW_RI_RES_WR_DCACPU_V(0) |
2621 		FW_RI_RES_WR_FBMIN_V(2) |
2622 		FW_RI_RES_WR_FBMAX_V(3) |
2623 		FW_RI_RES_WR_CIDXFTHRESHO_V(0) |
2624 		FW_RI_RES_WR_CIDXFTHRESH_V(0) |
2625 		FW_RI_RES_WR_EQSIZE_V(eqsize));
2626 	res->u.srq.eqaddr = cpu_to_be64(wq->dma_addr);
2627 	res->u.srq.srqid = cpu_to_be32(srq->idx);
2628 	res->u.srq.pdid = cpu_to_be32(srq->pdid);
2629 	res->u.srq.hwsrqsize = cpu_to_be32(wq->rqt_size);
2630 	res->u.srq.hwsrqaddr = cpu_to_be32(wq->rqt_hwaddr -
2631 			rdev->lldi.vr->rq.start);
2632 
2633 	c4iw_init_wr_wait(wr_waitp);
2634 
2635 	ret = c4iw_ref_send_wait(rdev, skb, wr_waitp, 0, wq->qid, __func__);
2636 	if (ret)
2637 		goto err_free_queue;
2638 
2639 	pr_debug("%s srq %u eqid %u pdid %u queue va %p pa 0x%llx\n"
2640 			" bar2_addr %p rqt addr 0x%x size %d\n",
2641 			__func__, srq->idx, wq->qid, srq->pdid, wq->queue,
2642 			(u64)virt_to_phys(wq->queue), wq->bar2_va,
2643 			wq->rqt_hwaddr, wq->rqt_size);
2644 
2645 	return 0;
2646 err_free_queue:
2647 	dma_free_coherent(&rdev->lldi.pdev->dev,
2648 			  wq->memsize, wq->queue,
2649 			dma_unmap_addr(wq, mapping));
2650 err_free_rqtpool:
2651 	c4iw_rqtpool_free(rdev, wq->rqt_hwaddr, wq->rqt_size);
2652 err_free_pending_wrs:
2653 	if (!user)
2654 		kfree(wq->pending_wrs);
2655 err_free_sw_rq:
2656 	if (!user)
2657 		kfree(wq->sw_rq);
2658 err_put_qpid:
2659 	c4iw_put_qpid(rdev, wq->qid, uctx);
2660 err:
2661 	return ret;
2662 }
2663 
2664 void c4iw_copy_wr_to_srq(struct t4_srq *srq, union t4_recv_wr *wqe, u8 len16)
2665 {
2666 	u64 *src, *dst;
2667 
2668 	src = (u64 *)wqe;
2669 	dst = (u64 *)((u8 *)srq->queue + srq->wq_pidx * T4_EQ_ENTRY_SIZE);
2670 	while (len16) {
2671 		*dst++ = *src++;
2672 		if (dst >= (u64 *)&srq->queue[srq->size])
2673 			dst = (u64 *)srq->queue;
2674 		*dst++ = *src++;
2675 		if (dst >= (u64 *)&srq->queue[srq->size])
2676 			dst = (u64 *)srq->queue;
2677 		len16--;
2678 	}
2679 }
2680 
2681 int c4iw_create_srq(struct ib_srq *ib_srq, struct ib_srq_init_attr *attrs,
2682 			       struct ib_udata *udata)
2683 {
2684 	struct ib_pd *pd = ib_srq->pd;
2685 	struct c4iw_dev *rhp;
2686 	struct c4iw_srq *srq = to_c4iw_srq(ib_srq);
2687 	struct c4iw_pd *php;
2688 	struct c4iw_create_srq_resp uresp;
2689 	struct c4iw_ucontext *ucontext;
2690 	struct c4iw_mm_entry *srq_key_mm, *srq_db_key_mm;
2691 	int rqsize;
2692 	int ret;
2693 	int wr_len;
2694 
2695 	if (attrs->srq_type != IB_SRQT_BASIC)
2696 		return -EOPNOTSUPP;
2697 
2698 	pr_debug("%s ib_pd %p\n", __func__, pd);
2699 
2700 	php = to_c4iw_pd(pd);
2701 	rhp = php->rhp;
2702 
2703 	if (!rhp->rdev.lldi.vr->srq.size)
2704 		return -EINVAL;
2705 	if (attrs->attr.max_wr > rhp->rdev.hw_queue.t4_max_rq_size)
2706 		return -E2BIG;
2707 	if (attrs->attr.max_sge > T4_MAX_RECV_SGE)
2708 		return -E2BIG;
2709 
2710 	/*
2711 	 * SRQ RQT and RQ must be a power of 2 and at least 16 deep.
2712 	 */
2713 	rqsize = attrs->attr.max_wr + 1;
2714 	rqsize = roundup_pow_of_two(max_t(u16, rqsize, 16));
2715 
2716 	ucontext = rdma_udata_to_drv_context(udata, struct c4iw_ucontext,
2717 					     ibucontext);
2718 
2719 	srq->wr_waitp = c4iw_alloc_wr_wait(GFP_KERNEL);
2720 	if (!srq->wr_waitp)
2721 		return -ENOMEM;
2722 
2723 	srq->idx = c4iw_alloc_srq_idx(&rhp->rdev);
2724 	if (srq->idx < 0) {
2725 		ret = -ENOMEM;
2726 		goto err_free_wr_wait;
2727 	}
2728 
2729 	wr_len = sizeof(struct fw_ri_res_wr) + sizeof(struct fw_ri_res);
2730 	srq->destroy_skb = alloc_skb(wr_len, GFP_KERNEL);
2731 	if (!srq->destroy_skb) {
2732 		ret = -ENOMEM;
2733 		goto err_free_srq_idx;
2734 	}
2735 
2736 	srq->rhp = rhp;
2737 	srq->pdid = php->pdid;
2738 
2739 	srq->wq.size = rqsize;
2740 	srq->wq.memsize =
2741 		(rqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
2742 		sizeof(*srq->wq.queue);
2743 	if (ucontext)
2744 		srq->wq.memsize = roundup(srq->wq.memsize, PAGE_SIZE);
2745 
2746 	ret = alloc_srq_queue(srq, ucontext ? &ucontext->uctx :
2747 			&rhp->rdev.uctx, srq->wr_waitp);
2748 	if (ret)
2749 		goto err_free_skb;
2750 	attrs->attr.max_wr = rqsize - 1;
2751 
2752 	if (CHELSIO_CHIP_VERSION(rhp->rdev.lldi.adapter_type) > CHELSIO_T6)
2753 		srq->flags = T4_SRQ_LIMIT_SUPPORT;
2754 
2755 	if (udata) {
2756 		srq_key_mm = kmalloc(sizeof(*srq_key_mm), GFP_KERNEL);
2757 		if (!srq_key_mm) {
2758 			ret = -ENOMEM;
2759 			goto err_free_queue;
2760 		}
2761 		srq_db_key_mm = kmalloc(sizeof(*srq_db_key_mm), GFP_KERNEL);
2762 		if (!srq_db_key_mm) {
2763 			ret = -ENOMEM;
2764 			goto err_free_srq_key_mm;
2765 		}
2766 		memset(&uresp, 0, sizeof(uresp));
2767 		uresp.flags = srq->flags;
2768 		uresp.qid_mask = rhp->rdev.qpmask;
2769 		uresp.srqid = srq->wq.qid;
2770 		uresp.srq_size = srq->wq.size;
2771 		uresp.srq_memsize = srq->wq.memsize;
2772 		uresp.rqt_abs_idx = srq->wq.rqt_abs_idx;
2773 		spin_lock(&ucontext->mmap_lock);
2774 		uresp.srq_key = ucontext->key;
2775 		ucontext->key += PAGE_SIZE;
2776 		uresp.srq_db_gts_key = ucontext->key;
2777 		ucontext->key += PAGE_SIZE;
2778 		spin_unlock(&ucontext->mmap_lock);
2779 		ret = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
2780 		if (ret)
2781 			goto err_free_srq_db_key_mm;
2782 		srq_key_mm->key = uresp.srq_key;
2783 		srq_key_mm->addr = 0;
2784 		srq_key_mm->len = PAGE_ALIGN(srq->wq.memsize);
2785 		srq_key_mm->vaddr = srq->wq.queue;
2786 		srq_key_mm->dma_addr = srq->wq.dma_addr;
2787 		insert_flag_to_mmap(&rhp->rdev, srq_key_mm, srq_key_mm->addr);
2788 		insert_mmap(ucontext, srq_key_mm);
2789 		srq_db_key_mm->key = uresp.srq_db_gts_key;
2790 		srq_db_key_mm->addr = (u64)(unsigned long)srq->wq.bar2_pa;
2791 		srq_db_key_mm->len = PAGE_SIZE;
2792 		srq_db_key_mm->vaddr = NULL;
2793 		srq_db_key_mm->dma_addr = 0;
2794 		insert_flag_to_mmap(&rhp->rdev, srq_db_key_mm,
2795 				    srq_db_key_mm->addr);
2796 		insert_mmap(ucontext, srq_db_key_mm);
2797 	}
2798 
2799 	pr_debug("%s srq qid %u idx %u size %u memsize %lu num_entries %u\n",
2800 		 __func__, srq->wq.qid, srq->idx, srq->wq.size,
2801 			(unsigned long)srq->wq.memsize, attrs->attr.max_wr);
2802 
2803 	spin_lock_init(&srq->lock);
2804 	return 0;
2805 
2806 err_free_srq_db_key_mm:
2807 	kfree(srq_db_key_mm);
2808 err_free_srq_key_mm:
2809 	kfree(srq_key_mm);
2810 err_free_queue:
2811 	free_srq_queue(srq, ucontext ? &ucontext->uctx : &rhp->rdev.uctx,
2812 		       srq->wr_waitp);
2813 err_free_skb:
2814 	kfree_skb(srq->destroy_skb);
2815 err_free_srq_idx:
2816 	c4iw_free_srq_idx(&rhp->rdev, srq->idx);
2817 err_free_wr_wait:
2818 	c4iw_put_wr_wait(srq->wr_waitp);
2819 	return ret;
2820 }
2821 
2822 int c4iw_destroy_srq(struct ib_srq *ibsrq, struct ib_udata *udata)
2823 {
2824 	struct c4iw_dev *rhp;
2825 	struct c4iw_srq *srq;
2826 	struct c4iw_ucontext *ucontext;
2827 
2828 	srq = to_c4iw_srq(ibsrq);
2829 	rhp = srq->rhp;
2830 
2831 	pr_debug("%s id %d\n", __func__, srq->wq.qid);
2832 	ucontext = rdma_udata_to_drv_context(udata, struct c4iw_ucontext,
2833 					     ibucontext);
2834 	free_srq_queue(srq, ucontext ? &ucontext->uctx : &rhp->rdev.uctx,
2835 		       srq->wr_waitp);
2836 	c4iw_free_srq_idx(&rhp->rdev, srq->idx);
2837 	c4iw_put_wr_wait(srq->wr_waitp);
2838 	return 0;
2839 }
2840