xref: /openbmc/linux/drivers/infiniband/hw/cxgb4/qp.c (revision 2985bed6)
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 		goto free_dma;
299 	}
300 
301 	wq->rdev = rdev;
302 	wq->rq.msn = 1;
303 
304 	/* build fw_ri_res_wr */
305 	wr_len = sizeof(*res_wr) + 2 * sizeof(*res);
306 	if (need_rq)
307 		wr_len += sizeof(*res);
308 	skb = alloc_skb(wr_len, GFP_KERNEL);
309 	if (!skb) {
310 		ret = -ENOMEM;
311 		goto free_dma;
312 	}
313 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
314 
315 	res_wr = __skb_put_zero(skb, wr_len);
316 	res_wr->op_nres = cpu_to_be32(
317 			FW_WR_OP_V(FW_RI_RES_WR) |
318 			FW_RI_RES_WR_NRES_V(need_rq ? 2 : 1) |
319 			FW_WR_COMPL_F);
320 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
321 	res_wr->cookie = (uintptr_t)wr_waitp;
322 	res = res_wr->res;
323 	res->u.sqrq.restype = FW_RI_RES_TYPE_SQ;
324 	res->u.sqrq.op = FW_RI_RES_OP_WRITE;
325 
326 	/*
327 	 * eqsize is the number of 64B entries plus the status page size.
328 	 */
329 	eqsize = wq->sq.size * T4_SQ_NUM_SLOTS +
330 		rdev->hw_queue.t4_eq_status_entries;
331 
332 	res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
333 		FW_RI_RES_WR_HOSTFCMODE_V(0) |	/* no host cidx updates */
334 		FW_RI_RES_WR_CPRIO_V(0) |	/* don't keep in chip cache */
335 		FW_RI_RES_WR_PCIECHN_V(0) |	/* set by uP at ri_init time */
336 		(t4_sq_onchip(&wq->sq) ? FW_RI_RES_WR_ONCHIP_F : 0) |
337 		FW_RI_RES_WR_IQID_V(scq->cqid));
338 	res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
339 		FW_RI_RES_WR_DCAEN_V(0) |
340 		FW_RI_RES_WR_DCACPU_V(0) |
341 		FW_RI_RES_WR_FBMIN_V(2) |
342 		(t4_sq_onchip(&wq->sq) ? FW_RI_RES_WR_FBMAX_V(2) :
343 					 FW_RI_RES_WR_FBMAX_V(3)) |
344 		FW_RI_RES_WR_CIDXFTHRESHO_V(0) |
345 		FW_RI_RES_WR_CIDXFTHRESH_V(0) |
346 		FW_RI_RES_WR_EQSIZE_V(eqsize));
347 	res->u.sqrq.eqid = cpu_to_be32(wq->sq.qid);
348 	res->u.sqrq.eqaddr = cpu_to_be64(wq->sq.dma_addr);
349 
350 	if (need_rq) {
351 		res++;
352 		res->u.sqrq.restype = FW_RI_RES_TYPE_RQ;
353 		res->u.sqrq.op = FW_RI_RES_OP_WRITE;
354 
355 		/*
356 		 * eqsize is the number of 64B entries plus the status page size
357 		 */
358 		eqsize = wq->rq.size * T4_RQ_NUM_SLOTS +
359 			rdev->hw_queue.t4_eq_status_entries;
360 		res->u.sqrq.fetchszm_to_iqid =
361 			/* no host cidx updates */
362 			cpu_to_be32(FW_RI_RES_WR_HOSTFCMODE_V(0) |
363 			/* don't keep in chip cache */
364 			FW_RI_RES_WR_CPRIO_V(0) |
365 			/* set by uP at ri_init time */
366 			FW_RI_RES_WR_PCIECHN_V(0) |
367 			FW_RI_RES_WR_IQID_V(rcq->cqid));
368 		res->u.sqrq.dcaen_to_eqsize =
369 			cpu_to_be32(FW_RI_RES_WR_DCAEN_V(0) |
370 			FW_RI_RES_WR_DCACPU_V(0) |
371 			FW_RI_RES_WR_FBMIN_V(2) |
372 			FW_RI_RES_WR_FBMAX_V(3) |
373 			FW_RI_RES_WR_CIDXFTHRESHO_V(0) |
374 			FW_RI_RES_WR_CIDXFTHRESH_V(0) |
375 			FW_RI_RES_WR_EQSIZE_V(eqsize));
376 		res->u.sqrq.eqid = cpu_to_be32(wq->rq.qid);
377 		res->u.sqrq.eqaddr = cpu_to_be64(wq->rq.dma_addr);
378 	}
379 
380 	c4iw_init_wr_wait(wr_waitp);
381 	ret = c4iw_ref_send_wait(rdev, skb, wr_waitp, 0, wq->sq.qid, __func__);
382 	if (ret)
383 		goto free_dma;
384 
385 	pr_debug("sqid 0x%x rqid 0x%x kdb 0x%p sq_bar2_addr %p rq_bar2_addr %p\n",
386 		 wq->sq.qid, wq->rq.qid, wq->db,
387 		 wq->sq.bar2_va, wq->rq.bar2_va);
388 
389 	return 0;
390 free_dma:
391 	if (need_rq)
392 		dma_free_coherent(&rdev->lldi.pdev->dev,
393 				  wq->rq.memsize, wq->rq.queue,
394 				  dma_unmap_addr(&wq->rq, mapping));
395 free_sq:
396 	dealloc_sq(rdev, &wq->sq);
397 free_hwaddr:
398 	if (need_rq)
399 		c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
400 free_sw_rq:
401 	if (need_rq)
402 		kfree(wq->rq.sw_rq);
403 free_sw_sq:
404 	kfree(wq->sq.sw_sq);
405 free_rq_qid:
406 	if (need_rq)
407 		c4iw_put_qpid(rdev, wq->rq.qid, uctx);
408 free_sq_qid:
409 	c4iw_put_qpid(rdev, wq->sq.qid, uctx);
410 	return ret;
411 }
412 
413 static int build_immd(struct t4_sq *sq, struct fw_ri_immd *immdp,
414 		      const struct ib_send_wr *wr, int max, u32 *plenp)
415 {
416 	u8 *dstp, *srcp;
417 	u32 plen = 0;
418 	int i;
419 	int rem, len;
420 
421 	dstp = (u8 *)immdp->data;
422 	for (i = 0; i < wr->num_sge; i++) {
423 		if ((plen + wr->sg_list[i].length) > max)
424 			return -EMSGSIZE;
425 		srcp = (u8 *)(unsigned long)wr->sg_list[i].addr;
426 		plen += wr->sg_list[i].length;
427 		rem = wr->sg_list[i].length;
428 		while (rem) {
429 			if (dstp == (u8 *)&sq->queue[sq->size])
430 				dstp = (u8 *)sq->queue;
431 			if (rem <= (u8 *)&sq->queue[sq->size] - dstp)
432 				len = rem;
433 			else
434 				len = (u8 *)&sq->queue[sq->size] - dstp;
435 			memcpy(dstp, srcp, len);
436 			dstp += len;
437 			srcp += len;
438 			rem -= len;
439 		}
440 	}
441 	len = roundup(plen + sizeof(*immdp), 16) - (plen + sizeof(*immdp));
442 	if (len)
443 		memset(dstp, 0, len);
444 	immdp->op = FW_RI_DATA_IMMD;
445 	immdp->r1 = 0;
446 	immdp->r2 = 0;
447 	immdp->immdlen = cpu_to_be32(plen);
448 	*plenp = plen;
449 	return 0;
450 }
451 
452 static int build_isgl(__be64 *queue_start, __be64 *queue_end,
453 		      struct fw_ri_isgl *isglp, struct ib_sge *sg_list,
454 		      int num_sge, u32 *plenp)
455 
456 {
457 	int i;
458 	u32 plen = 0;
459 	__be64 *flitp;
460 
461 	if ((__be64 *)isglp == queue_end)
462 		isglp = (struct fw_ri_isgl *)queue_start;
463 
464 	flitp = (__be64 *)isglp->sge;
465 
466 	for (i = 0; i < num_sge; i++) {
467 		if ((plen + sg_list[i].length) < plen)
468 			return -EMSGSIZE;
469 		plen += sg_list[i].length;
470 		*flitp = cpu_to_be64(((u64)sg_list[i].lkey << 32) |
471 				     sg_list[i].length);
472 		if (++flitp == queue_end)
473 			flitp = queue_start;
474 		*flitp = cpu_to_be64(sg_list[i].addr);
475 		if (++flitp == queue_end)
476 			flitp = queue_start;
477 	}
478 	*flitp = (__force __be64)0;
479 	isglp->op = FW_RI_DATA_ISGL;
480 	isglp->r1 = 0;
481 	isglp->nsge = cpu_to_be16(num_sge);
482 	isglp->r2 = 0;
483 	if (plenp)
484 		*plenp = plen;
485 	return 0;
486 }
487 
488 static int build_rdma_send(struct t4_sq *sq, union t4_wr *wqe,
489 			   const struct ib_send_wr *wr, u8 *len16)
490 {
491 	u32 plen;
492 	int size;
493 	int ret;
494 
495 	if (wr->num_sge > T4_MAX_SEND_SGE)
496 		return -EINVAL;
497 	switch (wr->opcode) {
498 	case IB_WR_SEND:
499 		if (wr->send_flags & IB_SEND_SOLICITED)
500 			wqe->send.sendop_pkd = cpu_to_be32(
501 				FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_SE));
502 		else
503 			wqe->send.sendop_pkd = cpu_to_be32(
504 				FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND));
505 		wqe->send.stag_inv = 0;
506 		break;
507 	case IB_WR_SEND_WITH_INV:
508 		if (wr->send_flags & IB_SEND_SOLICITED)
509 			wqe->send.sendop_pkd = cpu_to_be32(
510 				FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_SE_INV));
511 		else
512 			wqe->send.sendop_pkd = cpu_to_be32(
513 				FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_INV));
514 		wqe->send.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
515 		break;
516 
517 	default:
518 		return -EINVAL;
519 	}
520 	wqe->send.r3 = 0;
521 	wqe->send.r4 = 0;
522 
523 	plen = 0;
524 	if (wr->num_sge) {
525 		if (wr->send_flags & IB_SEND_INLINE) {
526 			ret = build_immd(sq, wqe->send.u.immd_src, wr,
527 					 T4_MAX_SEND_INLINE, &plen);
528 			if (ret)
529 				return ret;
530 			size = sizeof(wqe->send) + sizeof(struct fw_ri_immd) +
531 			       plen;
532 		} else {
533 			ret = build_isgl((__be64 *)sq->queue,
534 					 (__be64 *)&sq->queue[sq->size],
535 					 wqe->send.u.isgl_src,
536 					 wr->sg_list, wr->num_sge, &plen);
537 			if (ret)
538 				return ret;
539 			size = sizeof(wqe->send) + sizeof(struct fw_ri_isgl) +
540 			       wr->num_sge * sizeof(struct fw_ri_sge);
541 		}
542 	} else {
543 		wqe->send.u.immd_src[0].op = FW_RI_DATA_IMMD;
544 		wqe->send.u.immd_src[0].r1 = 0;
545 		wqe->send.u.immd_src[0].r2 = 0;
546 		wqe->send.u.immd_src[0].immdlen = 0;
547 		size = sizeof(wqe->send) + sizeof(struct fw_ri_immd);
548 		plen = 0;
549 	}
550 	*len16 = DIV_ROUND_UP(size, 16);
551 	wqe->send.plen = cpu_to_be32(plen);
552 	return 0;
553 }
554 
555 static int build_rdma_write(struct t4_sq *sq, union t4_wr *wqe,
556 			    const struct ib_send_wr *wr, u8 *len16)
557 {
558 	u32 plen;
559 	int size;
560 	int ret;
561 
562 	if (wr->num_sge > T4_MAX_SEND_SGE)
563 		return -EINVAL;
564 
565 	/*
566 	 * iWARP protocol supports 64 bit immediate data but rdma api
567 	 * limits it to 32bit.
568 	 */
569 	if (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
570 		wqe->write.iw_imm_data.ib_imm_data.imm_data32 = wr->ex.imm_data;
571 	else
572 		wqe->write.iw_imm_data.ib_imm_data.imm_data32 = 0;
573 	wqe->write.stag_sink = cpu_to_be32(rdma_wr(wr)->rkey);
574 	wqe->write.to_sink = cpu_to_be64(rdma_wr(wr)->remote_addr);
575 	if (wr->num_sge) {
576 		if (wr->send_flags & IB_SEND_INLINE) {
577 			ret = build_immd(sq, wqe->write.u.immd_src, wr,
578 					 T4_MAX_WRITE_INLINE, &plen);
579 			if (ret)
580 				return ret;
581 			size = sizeof(wqe->write) + sizeof(struct fw_ri_immd) +
582 			       plen;
583 		} else {
584 			ret = build_isgl((__be64 *)sq->queue,
585 					 (__be64 *)&sq->queue[sq->size],
586 					 wqe->write.u.isgl_src,
587 					 wr->sg_list, wr->num_sge, &plen);
588 			if (ret)
589 				return ret;
590 			size = sizeof(wqe->write) + sizeof(struct fw_ri_isgl) +
591 			       wr->num_sge * sizeof(struct fw_ri_sge);
592 		}
593 	} else {
594 		wqe->write.u.immd_src[0].op = FW_RI_DATA_IMMD;
595 		wqe->write.u.immd_src[0].r1 = 0;
596 		wqe->write.u.immd_src[0].r2 = 0;
597 		wqe->write.u.immd_src[0].immdlen = 0;
598 		size = sizeof(wqe->write) + sizeof(struct fw_ri_immd);
599 		plen = 0;
600 	}
601 	*len16 = DIV_ROUND_UP(size, 16);
602 	wqe->write.plen = cpu_to_be32(plen);
603 	return 0;
604 }
605 
606 static void build_immd_cmpl(struct t4_sq *sq, struct fw_ri_immd_cmpl *immdp,
607 			    struct ib_send_wr *wr)
608 {
609 	memcpy((u8 *)immdp->data, (u8 *)(uintptr_t)wr->sg_list->addr, 16);
610 	memset(immdp->r1, 0, 6);
611 	immdp->op = FW_RI_DATA_IMMD;
612 	immdp->immdlen = 16;
613 }
614 
615 static void build_rdma_write_cmpl(struct t4_sq *sq,
616 				  struct fw_ri_rdma_write_cmpl_wr *wcwr,
617 				  const struct ib_send_wr *wr, u8 *len16)
618 {
619 	u32 plen;
620 	int size;
621 
622 	/*
623 	 * This code assumes the struct fields preceding the write isgl
624 	 * fit in one 64B WR slot.  This is because the WQE is built
625 	 * directly in the dma queue, and wrapping is only handled
626 	 * by the code buildling sgls.  IE the "fixed part" of the wr
627 	 * structs must all fit in 64B.  The WQE build code should probably be
628 	 * redesigned to avoid this restriction, but for now just add
629 	 * the BUILD_BUG_ON() to catch if this WQE struct gets too big.
630 	 */
631 	BUILD_BUG_ON(offsetof(struct fw_ri_rdma_write_cmpl_wr, u) > 64);
632 
633 	wcwr->stag_sink = cpu_to_be32(rdma_wr(wr)->rkey);
634 	wcwr->to_sink = cpu_to_be64(rdma_wr(wr)->remote_addr);
635 	if (wr->next->opcode == IB_WR_SEND)
636 		wcwr->stag_inv = 0;
637 	else
638 		wcwr->stag_inv = cpu_to_be32(wr->next->ex.invalidate_rkey);
639 	wcwr->r2 = 0;
640 	wcwr->r3 = 0;
641 
642 	/* SEND_INV SGL */
643 	if (wr->next->send_flags & IB_SEND_INLINE)
644 		build_immd_cmpl(sq, &wcwr->u_cmpl.immd_src, wr->next);
645 	else
646 		build_isgl((__be64 *)sq->queue, (__be64 *)&sq->queue[sq->size],
647 			   &wcwr->u_cmpl.isgl_src, wr->next->sg_list, 1, NULL);
648 
649 	/* WRITE SGL */
650 	build_isgl((__be64 *)sq->queue, (__be64 *)&sq->queue[sq->size],
651 		   wcwr->u.isgl_src, wr->sg_list, wr->num_sge, &plen);
652 
653 	size = sizeof(*wcwr) + sizeof(struct fw_ri_isgl) +
654 		wr->num_sge * sizeof(struct fw_ri_sge);
655 	wcwr->plen = cpu_to_be32(plen);
656 	*len16 = DIV_ROUND_UP(size, 16);
657 }
658 
659 static int build_rdma_read(union t4_wr *wqe, const struct ib_send_wr *wr,
660 			   u8 *len16)
661 {
662 	if (wr->num_sge > 1)
663 		return -EINVAL;
664 	if (wr->num_sge && wr->sg_list[0].length) {
665 		wqe->read.stag_src = cpu_to_be32(rdma_wr(wr)->rkey);
666 		wqe->read.to_src_hi = cpu_to_be32((u32)(rdma_wr(wr)->remote_addr
667 							>> 32));
668 		wqe->read.to_src_lo = cpu_to_be32((u32)rdma_wr(wr)->remote_addr);
669 		wqe->read.stag_sink = cpu_to_be32(wr->sg_list[0].lkey);
670 		wqe->read.plen = cpu_to_be32(wr->sg_list[0].length);
671 		wqe->read.to_sink_hi = cpu_to_be32((u32)(wr->sg_list[0].addr
672 							 >> 32));
673 		wqe->read.to_sink_lo = cpu_to_be32((u32)(wr->sg_list[0].addr));
674 	} else {
675 		wqe->read.stag_src = cpu_to_be32(2);
676 		wqe->read.to_src_hi = 0;
677 		wqe->read.to_src_lo = 0;
678 		wqe->read.stag_sink = cpu_to_be32(2);
679 		wqe->read.plen = 0;
680 		wqe->read.to_sink_hi = 0;
681 		wqe->read.to_sink_lo = 0;
682 	}
683 	wqe->read.r2 = 0;
684 	wqe->read.r5 = 0;
685 	*len16 = DIV_ROUND_UP(sizeof(wqe->read), 16);
686 	return 0;
687 }
688 
689 static void post_write_cmpl(struct c4iw_qp *qhp, const struct ib_send_wr *wr)
690 {
691 	bool send_signaled = (wr->next->send_flags & IB_SEND_SIGNALED) ||
692 			     qhp->sq_sig_all;
693 	bool write_signaled = (wr->send_flags & IB_SEND_SIGNALED) ||
694 			      qhp->sq_sig_all;
695 	struct t4_swsqe *swsqe;
696 	union t4_wr *wqe;
697 	u16 write_wrid;
698 	u8 len16;
699 	u16 idx;
700 
701 	/*
702 	 * The sw_sq entries still look like a WRITE and a SEND and consume
703 	 * 2 slots. The FW WR, however, will be a single uber-WR.
704 	 */
705 	wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
706 	       qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
707 	build_rdma_write_cmpl(&qhp->wq.sq, &wqe->write_cmpl, wr, &len16);
708 
709 	/* WRITE swsqe */
710 	swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
711 	swsqe->opcode = FW_RI_RDMA_WRITE;
712 	swsqe->idx = qhp->wq.sq.pidx;
713 	swsqe->complete = 0;
714 	swsqe->signaled = write_signaled;
715 	swsqe->flushed = 0;
716 	swsqe->wr_id = wr->wr_id;
717 	if (c4iw_wr_log) {
718 		swsqe->sge_ts =
719 			cxgb4_read_sge_timestamp(qhp->rhp->rdev.lldi.ports[0]);
720 		swsqe->host_time = ktime_get();
721 	}
722 
723 	write_wrid = qhp->wq.sq.pidx;
724 
725 	/* just bump the sw_sq */
726 	qhp->wq.sq.in_use++;
727 	if (++qhp->wq.sq.pidx == qhp->wq.sq.size)
728 		qhp->wq.sq.pidx = 0;
729 
730 	/* SEND_WITH_INV swsqe */
731 	swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
732 	if (wr->next->opcode == IB_WR_SEND)
733 		swsqe->opcode = FW_RI_SEND;
734 	else
735 		swsqe->opcode = FW_RI_SEND_WITH_INV;
736 	swsqe->idx = qhp->wq.sq.pidx;
737 	swsqe->complete = 0;
738 	swsqe->signaled = send_signaled;
739 	swsqe->flushed = 0;
740 	swsqe->wr_id = wr->next->wr_id;
741 	if (c4iw_wr_log) {
742 		swsqe->sge_ts =
743 			cxgb4_read_sge_timestamp(qhp->rhp->rdev.lldi.ports[0]);
744 		swsqe->host_time = ktime_get();
745 	}
746 
747 	wqe->write_cmpl.flags_send = send_signaled ? FW_RI_COMPLETION_FLAG : 0;
748 	wqe->write_cmpl.wrid_send = qhp->wq.sq.pidx;
749 
750 	init_wr_hdr(wqe, write_wrid, FW_RI_RDMA_WRITE_CMPL_WR,
751 		    write_signaled ? FW_RI_COMPLETION_FLAG : 0, len16);
752 	t4_sq_produce(&qhp->wq, len16);
753 	idx = DIV_ROUND_UP(len16 * 16, T4_EQ_ENTRY_SIZE);
754 
755 	t4_ring_sq_db(&qhp->wq, idx, wqe);
756 }
757 
758 static int build_rdma_recv(struct c4iw_qp *qhp, union t4_recv_wr *wqe,
759 			   const struct ib_recv_wr *wr, u8 *len16)
760 {
761 	int ret;
762 
763 	ret = build_isgl((__be64 *)qhp->wq.rq.queue,
764 			 (__be64 *)&qhp->wq.rq.queue[qhp->wq.rq.size],
765 			 &wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL);
766 	if (ret)
767 		return ret;
768 	*len16 = DIV_ROUND_UP(
769 		sizeof(wqe->recv) + wr->num_sge * sizeof(struct fw_ri_sge), 16);
770 	return 0;
771 }
772 
773 static int build_srq_recv(union t4_recv_wr *wqe, const struct ib_recv_wr *wr,
774 			  u8 *len16)
775 {
776 	int ret;
777 
778 	ret = build_isgl((__be64 *)wqe, (__be64 *)(wqe + 1),
779 			 &wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL);
780 	if (ret)
781 		return ret;
782 	*len16 = DIV_ROUND_UP(sizeof(wqe->recv) +
783 			      wr->num_sge * sizeof(struct fw_ri_sge), 16);
784 	return 0;
785 }
786 
787 static void build_tpte_memreg(struct fw_ri_fr_nsmr_tpte_wr *fr,
788 			      const struct ib_reg_wr *wr, struct c4iw_mr *mhp,
789 			      u8 *len16)
790 {
791 	__be64 *p = (__be64 *)fr->pbl;
792 
793 	fr->r2 = cpu_to_be32(0);
794 	fr->stag = cpu_to_be32(mhp->ibmr.rkey);
795 
796 	fr->tpte.valid_to_pdid = cpu_to_be32(FW_RI_TPTE_VALID_F |
797 		FW_RI_TPTE_STAGKEY_V((mhp->ibmr.rkey & FW_RI_TPTE_STAGKEY_M)) |
798 		FW_RI_TPTE_STAGSTATE_V(1) |
799 		FW_RI_TPTE_STAGTYPE_V(FW_RI_STAG_NSMR) |
800 		FW_RI_TPTE_PDID_V(mhp->attr.pdid));
801 	fr->tpte.locread_to_qpid = cpu_to_be32(
802 		FW_RI_TPTE_PERM_V(c4iw_ib_to_tpt_access(wr->access)) |
803 		FW_RI_TPTE_ADDRTYPE_V(FW_RI_VA_BASED_TO) |
804 		FW_RI_TPTE_PS_V(ilog2(wr->mr->page_size) - 12));
805 	fr->tpte.nosnoop_pbladdr = cpu_to_be32(FW_RI_TPTE_PBLADDR_V(
806 		PBL_OFF(&mhp->rhp->rdev, mhp->attr.pbl_addr)>>3));
807 	fr->tpte.dca_mwbcnt_pstag = cpu_to_be32(0);
808 	fr->tpte.len_hi = cpu_to_be32(0);
809 	fr->tpte.len_lo = cpu_to_be32(mhp->ibmr.length);
810 	fr->tpte.va_hi = cpu_to_be32(mhp->ibmr.iova >> 32);
811 	fr->tpte.va_lo_fbo = cpu_to_be32(mhp->ibmr.iova & 0xffffffff);
812 
813 	p[0] = cpu_to_be64((u64)mhp->mpl[0]);
814 	p[1] = cpu_to_be64((u64)mhp->mpl[1]);
815 
816 	*len16 = DIV_ROUND_UP(sizeof(*fr), 16);
817 }
818 
819 static int build_memreg(struct t4_sq *sq, union t4_wr *wqe,
820 			const struct ib_reg_wr *wr, struct c4iw_mr *mhp,
821 			u8 *len16, bool dsgl_supported)
822 {
823 	struct fw_ri_immd *imdp;
824 	__be64 *p;
825 	int i;
826 	int pbllen = roundup(mhp->mpl_len * sizeof(u64), 32);
827 	int rem;
828 
829 	if (mhp->mpl_len > t4_max_fr_depth(dsgl_supported && use_dsgl))
830 		return -EINVAL;
831 
832 	wqe->fr.qpbinde_to_dcacpu = 0;
833 	wqe->fr.pgsz_shift = ilog2(wr->mr->page_size) - 12;
834 	wqe->fr.addr_type = FW_RI_VA_BASED_TO;
835 	wqe->fr.mem_perms = c4iw_ib_to_tpt_access(wr->access);
836 	wqe->fr.len_hi = 0;
837 	wqe->fr.len_lo = cpu_to_be32(mhp->ibmr.length);
838 	wqe->fr.stag = cpu_to_be32(wr->key);
839 	wqe->fr.va_hi = cpu_to_be32(mhp->ibmr.iova >> 32);
840 	wqe->fr.va_lo_fbo = cpu_to_be32(mhp->ibmr.iova &
841 					0xffffffff);
842 
843 	if (dsgl_supported && use_dsgl && (pbllen > max_fr_immd)) {
844 		struct fw_ri_dsgl *sglp;
845 
846 		for (i = 0; i < mhp->mpl_len; i++)
847 			mhp->mpl[i] = (__force u64)cpu_to_be64((u64)mhp->mpl[i]);
848 
849 		sglp = (struct fw_ri_dsgl *)(&wqe->fr + 1);
850 		sglp->op = FW_RI_DATA_DSGL;
851 		sglp->r1 = 0;
852 		sglp->nsge = cpu_to_be16(1);
853 		sglp->addr0 = cpu_to_be64(mhp->mpl_addr);
854 		sglp->len0 = cpu_to_be32(pbllen);
855 
856 		*len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*sglp), 16);
857 	} else {
858 		imdp = (struct fw_ri_immd *)(&wqe->fr + 1);
859 		imdp->op = FW_RI_DATA_IMMD;
860 		imdp->r1 = 0;
861 		imdp->r2 = 0;
862 		imdp->immdlen = cpu_to_be32(pbllen);
863 		p = (__be64 *)(imdp + 1);
864 		rem = pbllen;
865 		for (i = 0; i < mhp->mpl_len; i++) {
866 			*p = cpu_to_be64((u64)mhp->mpl[i]);
867 			rem -= sizeof(*p);
868 			if (++p == (__be64 *)&sq->queue[sq->size])
869 				p = (__be64 *)sq->queue;
870 		}
871 		while (rem) {
872 			*p = 0;
873 			rem -= sizeof(*p);
874 			if (++p == (__be64 *)&sq->queue[sq->size])
875 				p = (__be64 *)sq->queue;
876 		}
877 		*len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*imdp)
878 				      + pbllen, 16);
879 	}
880 	return 0;
881 }
882 
883 static int build_inv_stag(union t4_wr *wqe, const struct ib_send_wr *wr,
884 			  u8 *len16)
885 {
886 	wqe->inv.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
887 	wqe->inv.r2 = 0;
888 	*len16 = DIV_ROUND_UP(sizeof(wqe->inv), 16);
889 	return 0;
890 }
891 
892 void c4iw_qp_add_ref(struct ib_qp *qp)
893 {
894 	pr_debug("ib_qp %p\n", qp);
895 	refcount_inc(&to_c4iw_qp(qp)->qp_refcnt);
896 }
897 
898 void c4iw_qp_rem_ref(struct ib_qp *qp)
899 {
900 	pr_debug("ib_qp %p\n", qp);
901 	if (refcount_dec_and_test(&to_c4iw_qp(qp)->qp_refcnt))
902 		complete(&to_c4iw_qp(qp)->qp_rel_comp);
903 }
904 
905 static void add_to_fc_list(struct list_head *head, struct list_head *entry)
906 {
907 	if (list_empty(entry))
908 		list_add_tail(entry, head);
909 }
910 
911 static int ring_kernel_sq_db(struct c4iw_qp *qhp, u16 inc)
912 {
913 	unsigned long flags;
914 
915 	xa_lock_irqsave(&qhp->rhp->qps, flags);
916 	spin_lock(&qhp->lock);
917 	if (qhp->rhp->db_state == NORMAL)
918 		t4_ring_sq_db(&qhp->wq, inc, NULL);
919 	else {
920 		add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
921 		qhp->wq.sq.wq_pidx_inc += inc;
922 	}
923 	spin_unlock(&qhp->lock);
924 	xa_unlock_irqrestore(&qhp->rhp->qps, flags);
925 	return 0;
926 }
927 
928 static int ring_kernel_rq_db(struct c4iw_qp *qhp, u16 inc)
929 {
930 	unsigned long flags;
931 
932 	xa_lock_irqsave(&qhp->rhp->qps, flags);
933 	spin_lock(&qhp->lock);
934 	if (qhp->rhp->db_state == NORMAL)
935 		t4_ring_rq_db(&qhp->wq, inc, NULL);
936 	else {
937 		add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
938 		qhp->wq.rq.wq_pidx_inc += inc;
939 	}
940 	spin_unlock(&qhp->lock);
941 	xa_unlock_irqrestore(&qhp->rhp->qps, flags);
942 	return 0;
943 }
944 
945 static int ib_to_fw_opcode(int ib_opcode)
946 {
947 	int opcode;
948 
949 	switch (ib_opcode) {
950 	case IB_WR_SEND_WITH_INV:
951 		opcode = FW_RI_SEND_WITH_INV;
952 		break;
953 	case IB_WR_SEND:
954 		opcode = FW_RI_SEND;
955 		break;
956 	case IB_WR_RDMA_WRITE:
957 		opcode = FW_RI_RDMA_WRITE;
958 		break;
959 	case IB_WR_RDMA_WRITE_WITH_IMM:
960 		opcode = FW_RI_WRITE_IMMEDIATE;
961 		break;
962 	case IB_WR_RDMA_READ:
963 	case IB_WR_RDMA_READ_WITH_INV:
964 		opcode = FW_RI_READ_REQ;
965 		break;
966 	case IB_WR_REG_MR:
967 		opcode = FW_RI_FAST_REGISTER;
968 		break;
969 	case IB_WR_LOCAL_INV:
970 		opcode = FW_RI_LOCAL_INV;
971 		break;
972 	default:
973 		opcode = -EINVAL;
974 	}
975 	return opcode;
976 }
977 
978 static int complete_sq_drain_wr(struct c4iw_qp *qhp,
979 				const struct ib_send_wr *wr)
980 {
981 	struct t4_cqe cqe = {};
982 	struct c4iw_cq *schp;
983 	unsigned long flag;
984 	struct t4_cq *cq;
985 	int opcode;
986 
987 	schp = to_c4iw_cq(qhp->ibqp.send_cq);
988 	cq = &schp->cq;
989 
990 	opcode = ib_to_fw_opcode(wr->opcode);
991 	if (opcode < 0)
992 		return opcode;
993 
994 	cqe.u.drain_cookie = wr->wr_id;
995 	cqe.header = cpu_to_be32(CQE_STATUS_V(T4_ERR_SWFLUSH) |
996 				 CQE_OPCODE_V(opcode) |
997 				 CQE_TYPE_V(1) |
998 				 CQE_SWCQE_V(1) |
999 				 CQE_DRAIN_V(1) |
1000 				 CQE_QPID_V(qhp->wq.sq.qid));
1001 
1002 	spin_lock_irqsave(&schp->lock, flag);
1003 	cqe.bits_type_ts = cpu_to_be64(CQE_GENBIT_V((u64)cq->gen));
1004 	cq->sw_queue[cq->sw_pidx] = cqe;
1005 	t4_swcq_produce(cq);
1006 	spin_unlock_irqrestore(&schp->lock, flag);
1007 
1008 	if (t4_clear_cq_armed(&schp->cq)) {
1009 		spin_lock_irqsave(&schp->comp_handler_lock, flag);
1010 		(*schp->ibcq.comp_handler)(&schp->ibcq,
1011 					   schp->ibcq.cq_context);
1012 		spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1013 	}
1014 	return 0;
1015 }
1016 
1017 static int complete_sq_drain_wrs(struct c4iw_qp *qhp,
1018 				 const struct ib_send_wr *wr,
1019 				 const struct ib_send_wr **bad_wr)
1020 {
1021 	int ret = 0;
1022 
1023 	while (wr) {
1024 		ret = complete_sq_drain_wr(qhp, wr);
1025 		if (ret) {
1026 			*bad_wr = wr;
1027 			break;
1028 		}
1029 		wr = wr->next;
1030 	}
1031 	return ret;
1032 }
1033 
1034 static void complete_rq_drain_wr(struct c4iw_qp *qhp,
1035 				 const struct ib_recv_wr *wr)
1036 {
1037 	struct t4_cqe cqe = {};
1038 	struct c4iw_cq *rchp;
1039 	unsigned long flag;
1040 	struct t4_cq *cq;
1041 
1042 	rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
1043 	cq = &rchp->cq;
1044 
1045 	cqe.u.drain_cookie = wr->wr_id;
1046 	cqe.header = cpu_to_be32(CQE_STATUS_V(T4_ERR_SWFLUSH) |
1047 				 CQE_OPCODE_V(FW_RI_SEND) |
1048 				 CQE_TYPE_V(0) |
1049 				 CQE_SWCQE_V(1) |
1050 				 CQE_DRAIN_V(1) |
1051 				 CQE_QPID_V(qhp->wq.sq.qid));
1052 
1053 	spin_lock_irqsave(&rchp->lock, flag);
1054 	cqe.bits_type_ts = cpu_to_be64(CQE_GENBIT_V((u64)cq->gen));
1055 	cq->sw_queue[cq->sw_pidx] = cqe;
1056 	t4_swcq_produce(cq);
1057 	spin_unlock_irqrestore(&rchp->lock, flag);
1058 
1059 	if (t4_clear_cq_armed(&rchp->cq)) {
1060 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1061 		(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1062 					   rchp->ibcq.cq_context);
1063 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1064 	}
1065 }
1066 
1067 static void complete_rq_drain_wrs(struct c4iw_qp *qhp,
1068 				  const struct ib_recv_wr *wr)
1069 {
1070 	while (wr) {
1071 		complete_rq_drain_wr(qhp, wr);
1072 		wr = wr->next;
1073 	}
1074 }
1075 
1076 int c4iw_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
1077 		   const struct ib_send_wr **bad_wr)
1078 {
1079 	int err = 0;
1080 	u8 len16 = 0;
1081 	enum fw_wr_opcodes fw_opcode = 0;
1082 	enum fw_ri_wr_flags fw_flags;
1083 	struct c4iw_qp *qhp;
1084 	struct c4iw_dev *rhp;
1085 	union t4_wr *wqe = NULL;
1086 	u32 num_wrs;
1087 	struct t4_swsqe *swsqe;
1088 	unsigned long flag;
1089 	u16 idx = 0;
1090 
1091 	qhp = to_c4iw_qp(ibqp);
1092 	rhp = qhp->rhp;
1093 	spin_lock_irqsave(&qhp->lock, flag);
1094 
1095 	/*
1096 	 * If the qp has been flushed, then just insert a special
1097 	 * drain cqe.
1098 	 */
1099 	if (qhp->wq.flushed) {
1100 		spin_unlock_irqrestore(&qhp->lock, flag);
1101 		err = complete_sq_drain_wrs(qhp, wr, bad_wr);
1102 		return err;
1103 	}
1104 	num_wrs = t4_sq_avail(&qhp->wq);
1105 	if (num_wrs == 0) {
1106 		spin_unlock_irqrestore(&qhp->lock, flag);
1107 		*bad_wr = wr;
1108 		return -ENOMEM;
1109 	}
1110 
1111 	/*
1112 	 * Fastpath for NVMe-oF target WRITE + SEND_WITH_INV wr chain which is
1113 	 * the response for small NVMEe-oF READ requests.  If the chain is
1114 	 * exactly a WRITE->SEND_WITH_INV or a WRITE->SEND and the sgl depths
1115 	 * and lengths meet the requirements of the fw_ri_write_cmpl_wr work
1116 	 * request, then build and post the write_cmpl WR. If any of the tests
1117 	 * below are not true, then we continue on with the tradtional WRITE
1118 	 * and SEND WRs.
1119 	 */
1120 	if (qhp->rhp->rdev.lldi.write_cmpl_support &&
1121 	    CHELSIO_CHIP_VERSION(qhp->rhp->rdev.lldi.adapter_type) >=
1122 	    CHELSIO_T5 &&
1123 	    wr && wr->next && !wr->next->next &&
1124 	    wr->opcode == IB_WR_RDMA_WRITE &&
1125 	    wr->sg_list[0].length && wr->num_sge <= T4_WRITE_CMPL_MAX_SGL &&
1126 	    (wr->next->opcode == IB_WR_SEND ||
1127 	    wr->next->opcode == IB_WR_SEND_WITH_INV) &&
1128 	    wr->next->sg_list[0].length == T4_WRITE_CMPL_MAX_CQE &&
1129 	    wr->next->num_sge == 1 && num_wrs >= 2) {
1130 		post_write_cmpl(qhp, wr);
1131 		spin_unlock_irqrestore(&qhp->lock, flag);
1132 		return 0;
1133 	}
1134 
1135 	while (wr) {
1136 		if (num_wrs == 0) {
1137 			err = -ENOMEM;
1138 			*bad_wr = wr;
1139 			break;
1140 		}
1141 		wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
1142 		      qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
1143 
1144 		fw_flags = 0;
1145 		if (wr->send_flags & IB_SEND_SOLICITED)
1146 			fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
1147 		if (wr->send_flags & IB_SEND_SIGNALED || qhp->sq_sig_all)
1148 			fw_flags |= FW_RI_COMPLETION_FLAG;
1149 		swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
1150 		switch (wr->opcode) {
1151 		case IB_WR_SEND_WITH_INV:
1152 		case IB_WR_SEND:
1153 			if (wr->send_flags & IB_SEND_FENCE)
1154 				fw_flags |= FW_RI_READ_FENCE_FLAG;
1155 			fw_opcode = FW_RI_SEND_WR;
1156 			if (wr->opcode == IB_WR_SEND)
1157 				swsqe->opcode = FW_RI_SEND;
1158 			else
1159 				swsqe->opcode = FW_RI_SEND_WITH_INV;
1160 			err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
1161 			break;
1162 		case IB_WR_RDMA_WRITE_WITH_IMM:
1163 			if (unlikely(!rhp->rdev.lldi.write_w_imm_support)) {
1164 				err = -EINVAL;
1165 				break;
1166 			}
1167 			fw_flags |= FW_RI_RDMA_WRITE_WITH_IMMEDIATE;
1168 			/*FALLTHROUGH*/
1169 		case IB_WR_RDMA_WRITE:
1170 			fw_opcode = FW_RI_RDMA_WRITE_WR;
1171 			swsqe->opcode = FW_RI_RDMA_WRITE;
1172 			err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
1173 			break;
1174 		case IB_WR_RDMA_READ:
1175 		case IB_WR_RDMA_READ_WITH_INV:
1176 			fw_opcode = FW_RI_RDMA_READ_WR;
1177 			swsqe->opcode = FW_RI_READ_REQ;
1178 			if (wr->opcode == IB_WR_RDMA_READ_WITH_INV) {
1179 				c4iw_invalidate_mr(rhp, wr->sg_list[0].lkey);
1180 				fw_flags = FW_RI_RDMA_READ_INVALIDATE;
1181 			} else {
1182 				fw_flags = 0;
1183 			}
1184 			err = build_rdma_read(wqe, wr, &len16);
1185 			if (err)
1186 				break;
1187 			swsqe->read_len = wr->sg_list[0].length;
1188 			if (!qhp->wq.sq.oldest_read)
1189 				qhp->wq.sq.oldest_read = swsqe;
1190 			break;
1191 		case IB_WR_REG_MR: {
1192 			struct c4iw_mr *mhp = to_c4iw_mr(reg_wr(wr)->mr);
1193 
1194 			swsqe->opcode = FW_RI_FAST_REGISTER;
1195 			if (rhp->rdev.lldi.fr_nsmr_tpte_wr_support &&
1196 			    !mhp->attr.state && mhp->mpl_len <= 2) {
1197 				fw_opcode = FW_RI_FR_NSMR_TPTE_WR;
1198 				build_tpte_memreg(&wqe->fr_tpte, reg_wr(wr),
1199 						  mhp, &len16);
1200 			} else {
1201 				fw_opcode = FW_RI_FR_NSMR_WR;
1202 				err = build_memreg(&qhp->wq.sq, wqe, reg_wr(wr),
1203 				       mhp, &len16,
1204 				       rhp->rdev.lldi.ulptx_memwrite_dsgl);
1205 				if (err)
1206 					break;
1207 			}
1208 			mhp->attr.state = 1;
1209 			break;
1210 		}
1211 		case IB_WR_LOCAL_INV:
1212 			if (wr->send_flags & IB_SEND_FENCE)
1213 				fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
1214 			fw_opcode = FW_RI_INV_LSTAG_WR;
1215 			swsqe->opcode = FW_RI_LOCAL_INV;
1216 			err = build_inv_stag(wqe, wr, &len16);
1217 			c4iw_invalidate_mr(rhp, wr->ex.invalidate_rkey);
1218 			break;
1219 		default:
1220 			pr_warn("%s post of type=%d TBD!\n", __func__,
1221 				wr->opcode);
1222 			err = -EINVAL;
1223 		}
1224 		if (err) {
1225 			*bad_wr = wr;
1226 			break;
1227 		}
1228 		swsqe->idx = qhp->wq.sq.pidx;
1229 		swsqe->complete = 0;
1230 		swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED) ||
1231 				  qhp->sq_sig_all;
1232 		swsqe->flushed = 0;
1233 		swsqe->wr_id = wr->wr_id;
1234 		if (c4iw_wr_log) {
1235 			swsqe->sge_ts = cxgb4_read_sge_timestamp(
1236 					rhp->rdev.lldi.ports[0]);
1237 			swsqe->host_time = ktime_get();
1238 		}
1239 
1240 		init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
1241 
1242 		pr_debug("cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u\n",
1243 			 (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
1244 			 swsqe->opcode, swsqe->read_len);
1245 		wr = wr->next;
1246 		num_wrs--;
1247 		t4_sq_produce(&qhp->wq, len16);
1248 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
1249 	}
1250 	if (!rhp->rdev.status_page->db_off) {
1251 		t4_ring_sq_db(&qhp->wq, idx, wqe);
1252 		spin_unlock_irqrestore(&qhp->lock, flag);
1253 	} else {
1254 		spin_unlock_irqrestore(&qhp->lock, flag);
1255 		ring_kernel_sq_db(qhp, idx);
1256 	}
1257 	return err;
1258 }
1259 
1260 int c4iw_post_receive(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
1261 		      const struct ib_recv_wr **bad_wr)
1262 {
1263 	int err = 0;
1264 	struct c4iw_qp *qhp;
1265 	union t4_recv_wr *wqe = NULL;
1266 	u32 num_wrs;
1267 	u8 len16 = 0;
1268 	unsigned long flag;
1269 	u16 idx = 0;
1270 
1271 	qhp = to_c4iw_qp(ibqp);
1272 	spin_lock_irqsave(&qhp->lock, flag);
1273 
1274 	/*
1275 	 * If the qp has been flushed, then just insert a special
1276 	 * drain cqe.
1277 	 */
1278 	if (qhp->wq.flushed) {
1279 		spin_unlock_irqrestore(&qhp->lock, flag);
1280 		complete_rq_drain_wrs(qhp, wr);
1281 		return err;
1282 	}
1283 	num_wrs = t4_rq_avail(&qhp->wq);
1284 	if (num_wrs == 0) {
1285 		spin_unlock_irqrestore(&qhp->lock, flag);
1286 		*bad_wr = wr;
1287 		return -ENOMEM;
1288 	}
1289 	while (wr) {
1290 		if (wr->num_sge > T4_MAX_RECV_SGE) {
1291 			err = -EINVAL;
1292 			*bad_wr = wr;
1293 			break;
1294 		}
1295 		wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
1296 					   qhp->wq.rq.wq_pidx *
1297 					   T4_EQ_ENTRY_SIZE);
1298 		if (num_wrs)
1299 			err = build_rdma_recv(qhp, wqe, wr, &len16);
1300 		else
1301 			err = -ENOMEM;
1302 		if (err) {
1303 			*bad_wr = wr;
1304 			break;
1305 		}
1306 
1307 		qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
1308 		if (c4iw_wr_log) {
1309 			qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].sge_ts =
1310 				cxgb4_read_sge_timestamp(
1311 						qhp->rhp->rdev.lldi.ports[0]);
1312 			qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].host_time =
1313 				ktime_get();
1314 		}
1315 
1316 		wqe->recv.opcode = FW_RI_RECV_WR;
1317 		wqe->recv.r1 = 0;
1318 		wqe->recv.wrid = qhp->wq.rq.pidx;
1319 		wqe->recv.r2[0] = 0;
1320 		wqe->recv.r2[1] = 0;
1321 		wqe->recv.r2[2] = 0;
1322 		wqe->recv.len16 = len16;
1323 		pr_debug("cookie 0x%llx pidx %u\n",
1324 			 (unsigned long long)wr->wr_id, qhp->wq.rq.pidx);
1325 		t4_rq_produce(&qhp->wq, len16);
1326 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
1327 		wr = wr->next;
1328 		num_wrs--;
1329 	}
1330 	if (!qhp->rhp->rdev.status_page->db_off) {
1331 		t4_ring_rq_db(&qhp->wq, idx, wqe);
1332 		spin_unlock_irqrestore(&qhp->lock, flag);
1333 	} else {
1334 		spin_unlock_irqrestore(&qhp->lock, flag);
1335 		ring_kernel_rq_db(qhp, idx);
1336 	}
1337 	return err;
1338 }
1339 
1340 static void defer_srq_wr(struct t4_srq *srq, union t4_recv_wr *wqe,
1341 			 u64 wr_id, u8 len16)
1342 {
1343 	struct t4_srq_pending_wr *pwr = &srq->pending_wrs[srq->pending_pidx];
1344 
1345 	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",
1346 		 __func__, srq->cidx, srq->pidx, srq->wq_pidx,
1347 		 srq->in_use, srq->ooo_count,
1348 		 (unsigned long long)wr_id, srq->pending_cidx,
1349 		 srq->pending_pidx, srq->pending_in_use);
1350 	pwr->wr_id = wr_id;
1351 	pwr->len16 = len16;
1352 	memcpy(&pwr->wqe, wqe, len16 * 16);
1353 	t4_srq_produce_pending_wr(srq);
1354 }
1355 
1356 int c4iw_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
1357 		       const struct ib_recv_wr **bad_wr)
1358 {
1359 	union t4_recv_wr *wqe, lwqe;
1360 	struct c4iw_srq *srq;
1361 	unsigned long flag;
1362 	u8 len16 = 0;
1363 	u16 idx = 0;
1364 	int err = 0;
1365 	u32 num_wrs;
1366 
1367 	srq = to_c4iw_srq(ibsrq);
1368 	spin_lock_irqsave(&srq->lock, flag);
1369 	num_wrs = t4_srq_avail(&srq->wq);
1370 	if (num_wrs == 0) {
1371 		spin_unlock_irqrestore(&srq->lock, flag);
1372 		return -ENOMEM;
1373 	}
1374 	while (wr) {
1375 		if (wr->num_sge > T4_MAX_RECV_SGE) {
1376 			err = -EINVAL;
1377 			*bad_wr = wr;
1378 			break;
1379 		}
1380 		wqe = &lwqe;
1381 		if (num_wrs)
1382 			err = build_srq_recv(wqe, wr, &len16);
1383 		else
1384 			err = -ENOMEM;
1385 		if (err) {
1386 			*bad_wr = wr;
1387 			break;
1388 		}
1389 
1390 		wqe->recv.opcode = FW_RI_RECV_WR;
1391 		wqe->recv.r1 = 0;
1392 		wqe->recv.wrid = srq->wq.pidx;
1393 		wqe->recv.r2[0] = 0;
1394 		wqe->recv.r2[1] = 0;
1395 		wqe->recv.r2[2] = 0;
1396 		wqe->recv.len16 = len16;
1397 
1398 		if (srq->wq.ooo_count ||
1399 		    srq->wq.pending_in_use ||
1400 		    srq->wq.sw_rq[srq->wq.pidx].valid) {
1401 			defer_srq_wr(&srq->wq, wqe, wr->wr_id, len16);
1402 		} else {
1403 			srq->wq.sw_rq[srq->wq.pidx].wr_id = wr->wr_id;
1404 			srq->wq.sw_rq[srq->wq.pidx].valid = 1;
1405 			c4iw_copy_wr_to_srq(&srq->wq, wqe, len16);
1406 			pr_debug("%s cidx %u pidx %u wq_pidx %u in_use %u wr_id 0x%llx\n",
1407 				 __func__, srq->wq.cidx,
1408 				 srq->wq.pidx, srq->wq.wq_pidx,
1409 				 srq->wq.in_use,
1410 				 (unsigned long long)wr->wr_id);
1411 			t4_srq_produce(&srq->wq, len16);
1412 			idx += DIV_ROUND_UP(len16 * 16, T4_EQ_ENTRY_SIZE);
1413 		}
1414 		wr = wr->next;
1415 		num_wrs--;
1416 	}
1417 	if (idx)
1418 		t4_ring_srq_db(&srq->wq, idx, len16, wqe);
1419 	spin_unlock_irqrestore(&srq->lock, flag);
1420 	return err;
1421 }
1422 
1423 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
1424 				    u8 *ecode)
1425 {
1426 	int status;
1427 	int tagged;
1428 	int opcode;
1429 	int rqtype;
1430 	int send_inv;
1431 
1432 	if (!err_cqe) {
1433 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
1434 		*ecode = 0;
1435 		return;
1436 	}
1437 
1438 	status = CQE_STATUS(err_cqe);
1439 	opcode = CQE_OPCODE(err_cqe);
1440 	rqtype = RQ_TYPE(err_cqe);
1441 	send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
1442 		   (opcode == FW_RI_SEND_WITH_SE_INV);
1443 	tagged = (opcode == FW_RI_RDMA_WRITE) ||
1444 		 (rqtype && (opcode == FW_RI_READ_RESP));
1445 
1446 	switch (status) {
1447 	case T4_ERR_STAG:
1448 		if (send_inv) {
1449 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1450 			*ecode = RDMAP_CANT_INV_STAG;
1451 		} else {
1452 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1453 			*ecode = RDMAP_INV_STAG;
1454 		}
1455 		break;
1456 	case T4_ERR_PDID:
1457 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1458 		if ((opcode == FW_RI_SEND_WITH_INV) ||
1459 		    (opcode == FW_RI_SEND_WITH_SE_INV))
1460 			*ecode = RDMAP_CANT_INV_STAG;
1461 		else
1462 			*ecode = RDMAP_STAG_NOT_ASSOC;
1463 		break;
1464 	case T4_ERR_QPID:
1465 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1466 		*ecode = RDMAP_STAG_NOT_ASSOC;
1467 		break;
1468 	case T4_ERR_ACCESS:
1469 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1470 		*ecode = RDMAP_ACC_VIOL;
1471 		break;
1472 	case T4_ERR_WRAP:
1473 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1474 		*ecode = RDMAP_TO_WRAP;
1475 		break;
1476 	case T4_ERR_BOUND:
1477 		if (tagged) {
1478 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1479 			*ecode = DDPT_BASE_BOUNDS;
1480 		} else {
1481 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1482 			*ecode = RDMAP_BASE_BOUNDS;
1483 		}
1484 		break;
1485 	case T4_ERR_INVALIDATE_SHARED_MR:
1486 	case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
1487 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1488 		*ecode = RDMAP_CANT_INV_STAG;
1489 		break;
1490 	case T4_ERR_ECC:
1491 	case T4_ERR_ECC_PSTAG:
1492 	case T4_ERR_INTERNAL_ERR:
1493 		*layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
1494 		*ecode = 0;
1495 		break;
1496 	case T4_ERR_OUT_OF_RQE:
1497 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1498 		*ecode = DDPU_INV_MSN_NOBUF;
1499 		break;
1500 	case T4_ERR_PBL_ADDR_BOUND:
1501 		*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1502 		*ecode = DDPT_BASE_BOUNDS;
1503 		break;
1504 	case T4_ERR_CRC:
1505 		*layer_type = LAYER_MPA|DDP_LLP;
1506 		*ecode = MPA_CRC_ERR;
1507 		break;
1508 	case T4_ERR_MARKER:
1509 		*layer_type = LAYER_MPA|DDP_LLP;
1510 		*ecode = MPA_MARKER_ERR;
1511 		break;
1512 	case T4_ERR_PDU_LEN_ERR:
1513 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1514 		*ecode = DDPU_MSG_TOOBIG;
1515 		break;
1516 	case T4_ERR_DDP_VERSION:
1517 		if (tagged) {
1518 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1519 			*ecode = DDPT_INV_VERS;
1520 		} else {
1521 			*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1522 			*ecode = DDPU_INV_VERS;
1523 		}
1524 		break;
1525 	case T4_ERR_RDMA_VERSION:
1526 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1527 		*ecode = RDMAP_INV_VERS;
1528 		break;
1529 	case T4_ERR_OPCODE:
1530 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1531 		*ecode = RDMAP_INV_OPCODE;
1532 		break;
1533 	case T4_ERR_DDP_QUEUE_NUM:
1534 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1535 		*ecode = DDPU_INV_QN;
1536 		break;
1537 	case T4_ERR_MSN:
1538 	case T4_ERR_MSN_GAP:
1539 	case T4_ERR_MSN_RANGE:
1540 	case T4_ERR_IRD_OVERFLOW:
1541 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1542 		*ecode = DDPU_INV_MSN_RANGE;
1543 		break;
1544 	case T4_ERR_TBIT:
1545 		*layer_type = LAYER_DDP|DDP_LOCAL_CATA;
1546 		*ecode = 0;
1547 		break;
1548 	case T4_ERR_MO:
1549 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1550 		*ecode = DDPU_INV_MO;
1551 		break;
1552 	default:
1553 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
1554 		*ecode = 0;
1555 		break;
1556 	}
1557 }
1558 
1559 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
1560 			   gfp_t gfp)
1561 {
1562 	struct fw_ri_wr *wqe;
1563 	struct sk_buff *skb;
1564 	struct terminate_message *term;
1565 
1566 	pr_debug("qhp %p qid 0x%x tid %u\n", qhp, qhp->wq.sq.qid,
1567 		 qhp->ep->hwtid);
1568 
1569 	skb = skb_dequeue(&qhp->ep->com.ep_skb_list);
1570 	if (WARN_ON(!skb))
1571 		return;
1572 
1573 	set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1574 
1575 	wqe = __skb_put_zero(skb, sizeof(*wqe));
1576 	wqe->op_compl = cpu_to_be32(FW_WR_OP_V(FW_RI_INIT_WR));
1577 	wqe->flowid_len16 = cpu_to_be32(
1578 		FW_WR_FLOWID_V(qhp->ep->hwtid) |
1579 		FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1580 
1581 	wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
1582 	wqe->u.terminate.immdlen = cpu_to_be32(sizeof(*term));
1583 	term = (struct terminate_message *)wqe->u.terminate.termmsg;
1584 	if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
1585 		term->layer_etype = qhp->attr.layer_etype;
1586 		term->ecode = qhp->attr.ecode;
1587 	} else
1588 		build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
1589 	c4iw_ofld_send(&qhp->rhp->rdev, skb);
1590 }
1591 
1592 /*
1593  * Assumes qhp lock is held.
1594  */
1595 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
1596 		       struct c4iw_cq *schp)
1597 {
1598 	int count;
1599 	int rq_flushed = 0, sq_flushed;
1600 	unsigned long flag;
1601 
1602 	pr_debug("qhp %p rchp %p schp %p\n", qhp, rchp, schp);
1603 
1604 	/* locking hierarchy: cqs lock first, then qp lock. */
1605 	spin_lock_irqsave(&rchp->lock, flag);
1606 	if (schp != rchp)
1607 		spin_lock(&schp->lock);
1608 	spin_lock(&qhp->lock);
1609 
1610 	if (qhp->wq.flushed) {
1611 		spin_unlock(&qhp->lock);
1612 		if (schp != rchp)
1613 			spin_unlock(&schp->lock);
1614 		spin_unlock_irqrestore(&rchp->lock, flag);
1615 		return;
1616 	}
1617 	qhp->wq.flushed = 1;
1618 	t4_set_wq_in_error(&qhp->wq, 0);
1619 
1620 	c4iw_flush_hw_cq(rchp, qhp);
1621 	if (!qhp->srq) {
1622 		c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
1623 		rq_flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
1624 	}
1625 
1626 	if (schp != rchp)
1627 		c4iw_flush_hw_cq(schp, qhp);
1628 	sq_flushed = c4iw_flush_sq(qhp);
1629 
1630 	spin_unlock(&qhp->lock);
1631 	if (schp != rchp)
1632 		spin_unlock(&schp->lock);
1633 	spin_unlock_irqrestore(&rchp->lock, flag);
1634 
1635 	if (schp == rchp) {
1636 		if ((rq_flushed || sq_flushed) &&
1637 		    t4_clear_cq_armed(&rchp->cq)) {
1638 			spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1639 			(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1640 						   rchp->ibcq.cq_context);
1641 			spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1642 		}
1643 	} else {
1644 		if (rq_flushed && t4_clear_cq_armed(&rchp->cq)) {
1645 			spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1646 			(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1647 						   rchp->ibcq.cq_context);
1648 			spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1649 		}
1650 		if (sq_flushed && t4_clear_cq_armed(&schp->cq)) {
1651 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1652 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1653 						   schp->ibcq.cq_context);
1654 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1655 		}
1656 	}
1657 }
1658 
1659 static void flush_qp(struct c4iw_qp *qhp)
1660 {
1661 	struct c4iw_cq *rchp, *schp;
1662 	unsigned long flag;
1663 
1664 	rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
1665 	schp = to_c4iw_cq(qhp->ibqp.send_cq);
1666 
1667 	if (qhp->ibqp.uobject) {
1668 
1669 		/* for user qps, qhp->wq.flushed is protected by qhp->mutex */
1670 		if (qhp->wq.flushed)
1671 			return;
1672 
1673 		qhp->wq.flushed = 1;
1674 		t4_set_wq_in_error(&qhp->wq, 0);
1675 		t4_set_cq_in_error(&rchp->cq);
1676 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1677 		(*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1678 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1679 		if (schp != rchp) {
1680 			t4_set_cq_in_error(&schp->cq);
1681 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1682 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1683 					schp->ibcq.cq_context);
1684 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1685 		}
1686 		return;
1687 	}
1688 	__flush_qp(qhp, rchp, schp);
1689 }
1690 
1691 static int rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1692 		     struct c4iw_ep *ep)
1693 {
1694 	struct fw_ri_wr *wqe;
1695 	int ret;
1696 	struct sk_buff *skb;
1697 
1698 	pr_debug("qhp %p qid 0x%x tid %u\n", qhp, qhp->wq.sq.qid, ep->hwtid);
1699 
1700 	skb = skb_dequeue(&ep->com.ep_skb_list);
1701 	if (WARN_ON(!skb))
1702 		return -ENOMEM;
1703 
1704 	set_wr_txq(skb, CPL_PRIORITY_DATA, ep->txq_idx);
1705 
1706 	wqe = __skb_put_zero(skb, sizeof(*wqe));
1707 	wqe->op_compl = cpu_to_be32(
1708 		FW_WR_OP_V(FW_RI_INIT_WR) |
1709 		FW_WR_COMPL_F);
1710 	wqe->flowid_len16 = cpu_to_be32(
1711 		FW_WR_FLOWID_V(ep->hwtid) |
1712 		FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1713 	wqe->cookie = (uintptr_t)ep->com.wr_waitp;
1714 
1715 	wqe->u.fini.type = FW_RI_TYPE_FINI;
1716 
1717 	ret = c4iw_ref_send_wait(&rhp->rdev, skb, ep->com.wr_waitp,
1718 				 qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
1719 
1720 	pr_debug("ret %d\n", ret);
1721 	return ret;
1722 }
1723 
1724 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1725 {
1726 	pr_debug("p2p_type = %d\n", p2p_type);
1727 	memset(&init->u, 0, sizeof(init->u));
1728 	switch (p2p_type) {
1729 	case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1730 		init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1731 		init->u.write.stag_sink = cpu_to_be32(1);
1732 		init->u.write.to_sink = cpu_to_be64(1);
1733 		init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1734 		init->u.write.len16 = DIV_ROUND_UP(
1735 			sizeof(init->u.write) + sizeof(struct fw_ri_immd), 16);
1736 		break;
1737 	case FW_RI_INIT_P2PTYPE_READ_REQ:
1738 		init->u.write.opcode = FW_RI_RDMA_READ_WR;
1739 		init->u.read.stag_src = cpu_to_be32(1);
1740 		init->u.read.to_src_lo = cpu_to_be32(1);
1741 		init->u.read.stag_sink = cpu_to_be32(1);
1742 		init->u.read.to_sink_lo = cpu_to_be32(1);
1743 		init->u.read.len16 = DIV_ROUND_UP(sizeof(init->u.read), 16);
1744 		break;
1745 	}
1746 }
1747 
1748 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1749 {
1750 	struct fw_ri_wr *wqe;
1751 	int ret;
1752 	struct sk_buff *skb;
1753 
1754 	pr_debug("qhp %p qid 0x%x tid %u ird %u ord %u\n", qhp,
1755 		 qhp->wq.sq.qid, qhp->ep->hwtid, qhp->ep->ird, qhp->ep->ord);
1756 
1757 	skb = alloc_skb(sizeof(*wqe), GFP_KERNEL);
1758 	if (!skb) {
1759 		ret = -ENOMEM;
1760 		goto out;
1761 	}
1762 	ret = alloc_ird(rhp, qhp->attr.max_ird);
1763 	if (ret) {
1764 		qhp->attr.max_ird = 0;
1765 		kfree_skb(skb);
1766 		goto out;
1767 	}
1768 	set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1769 
1770 	wqe = __skb_put_zero(skb, sizeof(*wqe));
1771 	wqe->op_compl = cpu_to_be32(
1772 		FW_WR_OP_V(FW_RI_INIT_WR) |
1773 		FW_WR_COMPL_F);
1774 	wqe->flowid_len16 = cpu_to_be32(
1775 		FW_WR_FLOWID_V(qhp->ep->hwtid) |
1776 		FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1777 
1778 	wqe->cookie = (uintptr_t)qhp->ep->com.wr_waitp;
1779 
1780 	wqe->u.init.type = FW_RI_TYPE_INIT;
1781 	wqe->u.init.mpareqbit_p2ptype =
1782 		FW_RI_WR_MPAREQBIT_V(qhp->attr.mpa_attr.initiator) |
1783 		FW_RI_WR_P2PTYPE_V(qhp->attr.mpa_attr.p2p_type);
1784 	wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1785 	if (qhp->attr.mpa_attr.recv_marker_enabled)
1786 		wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1787 	if (qhp->attr.mpa_attr.xmit_marker_enabled)
1788 		wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1789 	if (qhp->attr.mpa_attr.crc_enabled)
1790 		wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1791 
1792 	wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1793 			    FW_RI_QP_RDMA_WRITE_ENABLE |
1794 			    FW_RI_QP_BIND_ENABLE;
1795 	if (!qhp->ibqp.uobject)
1796 		wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1797 				     FW_RI_QP_STAG0_ENABLE;
1798 	wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1799 	wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1800 	wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1801 	wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1802 	if (qhp->srq) {
1803 		wqe->u.init.rq_eqid = cpu_to_be32(FW_RI_INIT_RQEQID_SRQ |
1804 						  qhp->srq->idx);
1805 	} else {
1806 		wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1807 		wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1808 		wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1809 						   rhp->rdev.lldi.vr->rq.start);
1810 	}
1811 	wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1812 	wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1813 	wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1814 	wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1815 	wqe->u.init.iss = cpu_to_be32(qhp->ep->snd_seq);
1816 	wqe->u.init.irs = cpu_to_be32(qhp->ep->rcv_seq);
1817 	if (qhp->attr.mpa_attr.initiator)
1818 		build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1819 
1820 	ret = c4iw_ref_send_wait(&rhp->rdev, skb, qhp->ep->com.wr_waitp,
1821 				 qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
1822 	if (!ret)
1823 		goto out;
1824 
1825 	free_ird(rhp, qhp->attr.max_ird);
1826 out:
1827 	pr_debug("ret %d\n", ret);
1828 	return ret;
1829 }
1830 
1831 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1832 		   enum c4iw_qp_attr_mask mask,
1833 		   struct c4iw_qp_attributes *attrs,
1834 		   int internal)
1835 {
1836 	int ret = 0;
1837 	struct c4iw_qp_attributes newattr = qhp->attr;
1838 	int disconnect = 0;
1839 	int terminate = 0;
1840 	int abort = 0;
1841 	int free = 0;
1842 	struct c4iw_ep *ep = NULL;
1843 
1844 	pr_debug("qhp %p sqid 0x%x rqid 0x%x ep %p state %d -> %d\n",
1845 		 qhp, qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep, qhp->attr.state,
1846 		 (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1847 
1848 	mutex_lock(&qhp->mutex);
1849 
1850 	/* Process attr changes if in IDLE */
1851 	if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1852 		if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1853 			ret = -EIO;
1854 			goto out;
1855 		}
1856 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1857 			newattr.enable_rdma_read = attrs->enable_rdma_read;
1858 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1859 			newattr.enable_rdma_write = attrs->enable_rdma_write;
1860 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1861 			newattr.enable_bind = attrs->enable_bind;
1862 		if (mask & C4IW_QP_ATTR_MAX_ORD) {
1863 			if (attrs->max_ord > c4iw_max_read_depth) {
1864 				ret = -EINVAL;
1865 				goto out;
1866 			}
1867 			newattr.max_ord = attrs->max_ord;
1868 		}
1869 		if (mask & C4IW_QP_ATTR_MAX_IRD) {
1870 			if (attrs->max_ird > cur_max_read_depth(rhp)) {
1871 				ret = -EINVAL;
1872 				goto out;
1873 			}
1874 			newattr.max_ird = attrs->max_ird;
1875 		}
1876 		qhp->attr = newattr;
1877 	}
1878 
1879 	if (mask & C4IW_QP_ATTR_SQ_DB) {
1880 		ret = ring_kernel_sq_db(qhp, attrs->sq_db_inc);
1881 		goto out;
1882 	}
1883 	if (mask & C4IW_QP_ATTR_RQ_DB) {
1884 		ret = ring_kernel_rq_db(qhp, attrs->rq_db_inc);
1885 		goto out;
1886 	}
1887 
1888 	if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1889 		goto out;
1890 	if (qhp->attr.state == attrs->next_state)
1891 		goto out;
1892 
1893 	switch (qhp->attr.state) {
1894 	case C4IW_QP_STATE_IDLE:
1895 		switch (attrs->next_state) {
1896 		case C4IW_QP_STATE_RTS:
1897 			if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1898 				ret = -EINVAL;
1899 				goto out;
1900 			}
1901 			if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1902 				ret = -EINVAL;
1903 				goto out;
1904 			}
1905 			qhp->attr.mpa_attr = attrs->mpa_attr;
1906 			qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1907 			qhp->ep = qhp->attr.llp_stream_handle;
1908 			set_state(qhp, C4IW_QP_STATE_RTS);
1909 
1910 			/*
1911 			 * Ref the endpoint here and deref when we
1912 			 * disassociate the endpoint from the QP.  This
1913 			 * happens in CLOSING->IDLE transition or *->ERROR
1914 			 * transition.
1915 			 */
1916 			c4iw_get_ep(&qhp->ep->com);
1917 			ret = rdma_init(rhp, qhp);
1918 			if (ret)
1919 				goto err;
1920 			break;
1921 		case C4IW_QP_STATE_ERROR:
1922 			set_state(qhp, C4IW_QP_STATE_ERROR);
1923 			flush_qp(qhp);
1924 			break;
1925 		default:
1926 			ret = -EINVAL;
1927 			goto out;
1928 		}
1929 		break;
1930 	case C4IW_QP_STATE_RTS:
1931 		switch (attrs->next_state) {
1932 		case C4IW_QP_STATE_CLOSING:
1933 			t4_set_wq_in_error(&qhp->wq, 0);
1934 			set_state(qhp, C4IW_QP_STATE_CLOSING);
1935 			ep = qhp->ep;
1936 			if (!internal) {
1937 				abort = 0;
1938 				disconnect = 1;
1939 				c4iw_get_ep(&qhp->ep->com);
1940 			}
1941 			ret = rdma_fini(rhp, qhp, ep);
1942 			if (ret)
1943 				goto err;
1944 			break;
1945 		case C4IW_QP_STATE_TERMINATE:
1946 			t4_set_wq_in_error(&qhp->wq, 0);
1947 			set_state(qhp, C4IW_QP_STATE_TERMINATE);
1948 			qhp->attr.layer_etype = attrs->layer_etype;
1949 			qhp->attr.ecode = attrs->ecode;
1950 			ep = qhp->ep;
1951 			if (!internal) {
1952 				c4iw_get_ep(&ep->com);
1953 				terminate = 1;
1954 				disconnect = 1;
1955 			} else {
1956 				terminate = qhp->attr.send_term;
1957 				ret = rdma_fini(rhp, qhp, ep);
1958 				if (ret)
1959 					goto err;
1960 			}
1961 			break;
1962 		case C4IW_QP_STATE_ERROR:
1963 			t4_set_wq_in_error(&qhp->wq, 0);
1964 			set_state(qhp, C4IW_QP_STATE_ERROR);
1965 			if (!internal) {
1966 				abort = 1;
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 
2107 	kfree(qhp);
2108 	return 0;
2109 }
2110 
2111 struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
2112 			     struct ib_udata *udata)
2113 {
2114 	struct c4iw_dev *rhp;
2115 	struct c4iw_qp *qhp;
2116 	struct c4iw_pd *php;
2117 	struct c4iw_cq *schp;
2118 	struct c4iw_cq *rchp;
2119 	struct c4iw_create_qp_resp uresp;
2120 	unsigned int sqsize, rqsize = 0;
2121 	struct c4iw_ucontext *ucontext = rdma_udata_to_drv_context(
2122 		udata, struct c4iw_ucontext, ibucontext);
2123 	int ret;
2124 	struct c4iw_mm_entry *sq_key_mm, *rq_key_mm = NULL, *sq_db_key_mm;
2125 	struct c4iw_mm_entry *rq_db_key_mm = NULL, *ma_sync_key_mm = NULL;
2126 
2127 	pr_debug("ib_pd %p\n", pd);
2128 
2129 	if (attrs->qp_type != IB_QPT_RC)
2130 		return ERR_PTR(-EINVAL);
2131 
2132 	php = to_c4iw_pd(pd);
2133 	rhp = php->rhp;
2134 	schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
2135 	rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
2136 	if (!schp || !rchp)
2137 		return ERR_PTR(-EINVAL);
2138 
2139 	if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
2140 		return ERR_PTR(-EINVAL);
2141 
2142 	if (!attrs->srq) {
2143 		if (attrs->cap.max_recv_wr > rhp->rdev.hw_queue.t4_max_rq_size)
2144 			return ERR_PTR(-E2BIG);
2145 		rqsize = attrs->cap.max_recv_wr + 1;
2146 		if (rqsize < 8)
2147 			rqsize = 8;
2148 	}
2149 
2150 	if (attrs->cap.max_send_wr > rhp->rdev.hw_queue.t4_max_sq_size)
2151 		return ERR_PTR(-E2BIG);
2152 	sqsize = attrs->cap.max_send_wr + 1;
2153 	if (sqsize < 8)
2154 		sqsize = 8;
2155 
2156 	qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
2157 	if (!qhp)
2158 		return ERR_PTR(-ENOMEM);
2159 
2160 	qhp->wr_waitp = c4iw_alloc_wr_wait(GFP_KERNEL);
2161 	if (!qhp->wr_waitp) {
2162 		ret = -ENOMEM;
2163 		goto err_free_qhp;
2164 	}
2165 
2166 	qhp->wq.sq.size = sqsize;
2167 	qhp->wq.sq.memsize =
2168 		(sqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
2169 		sizeof(*qhp->wq.sq.queue) + 16 * sizeof(__be64);
2170 	qhp->wq.sq.flush_cidx = -1;
2171 	if (!attrs->srq) {
2172 		qhp->wq.rq.size = rqsize;
2173 		qhp->wq.rq.memsize =
2174 			(rqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
2175 			sizeof(*qhp->wq.rq.queue);
2176 	}
2177 
2178 	if (ucontext) {
2179 		qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
2180 		if (!attrs->srq)
2181 			qhp->wq.rq.memsize =
2182 				roundup(qhp->wq.rq.memsize, PAGE_SIZE);
2183 	}
2184 
2185 	ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
2186 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx,
2187 			qhp->wr_waitp, !attrs->srq);
2188 	if (ret)
2189 		goto err_free_wr_wait;
2190 
2191 	attrs->cap.max_recv_wr = rqsize - 1;
2192 	attrs->cap.max_send_wr = sqsize - 1;
2193 	attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
2194 
2195 	qhp->rhp = rhp;
2196 	qhp->attr.pd = php->pdid;
2197 	qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
2198 	qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
2199 	qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
2200 	qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
2201 	qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
2202 	if (!attrs->srq) {
2203 		qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
2204 		qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
2205 	}
2206 	qhp->attr.state = C4IW_QP_STATE_IDLE;
2207 	qhp->attr.next_state = C4IW_QP_STATE_IDLE;
2208 	qhp->attr.enable_rdma_read = 1;
2209 	qhp->attr.enable_rdma_write = 1;
2210 	qhp->attr.enable_bind = 1;
2211 	qhp->attr.max_ord = 0;
2212 	qhp->attr.max_ird = 0;
2213 	qhp->sq_sig_all = attrs->sq_sig_type == IB_SIGNAL_ALL_WR;
2214 	spin_lock_init(&qhp->lock);
2215 	mutex_init(&qhp->mutex);
2216 	init_waitqueue_head(&qhp->wait);
2217 	init_completion(&qhp->qp_rel_comp);
2218 	refcount_set(&qhp->qp_refcnt, 1);
2219 
2220 	ret = xa_insert_irq(&rhp->qps, qhp->wq.sq.qid, qhp, GFP_KERNEL);
2221 	if (ret)
2222 		goto err_destroy_qp;
2223 
2224 	if (udata && ucontext) {
2225 		sq_key_mm = kmalloc(sizeof(*sq_key_mm), GFP_KERNEL);
2226 		if (!sq_key_mm) {
2227 			ret = -ENOMEM;
2228 			goto err_remove_handle;
2229 		}
2230 		if (!attrs->srq) {
2231 			rq_key_mm = kmalloc(sizeof(*rq_key_mm), GFP_KERNEL);
2232 			if (!rq_key_mm) {
2233 				ret = -ENOMEM;
2234 				goto err_free_sq_key;
2235 			}
2236 		}
2237 		sq_db_key_mm = kmalloc(sizeof(*sq_db_key_mm), GFP_KERNEL);
2238 		if (!sq_db_key_mm) {
2239 			ret = -ENOMEM;
2240 			goto err_free_rq_key;
2241 		}
2242 		if (!attrs->srq) {
2243 			rq_db_key_mm =
2244 				kmalloc(sizeof(*rq_db_key_mm), GFP_KERNEL);
2245 			if (!rq_db_key_mm) {
2246 				ret = -ENOMEM;
2247 				goto err_free_sq_db_key;
2248 			}
2249 		}
2250 		memset(&uresp, 0, sizeof(uresp));
2251 		if (t4_sq_onchip(&qhp->wq.sq)) {
2252 			ma_sync_key_mm = kmalloc(sizeof(*ma_sync_key_mm),
2253 						 GFP_KERNEL);
2254 			if (!ma_sync_key_mm) {
2255 				ret = -ENOMEM;
2256 				goto err_free_rq_db_key;
2257 			}
2258 			uresp.flags = C4IW_QPF_ONCHIP;
2259 		}
2260 		if (rhp->rdev.lldi.write_w_imm_support)
2261 			uresp.flags |= C4IW_QPF_WRITE_W_IMM;
2262 		uresp.qid_mask = rhp->rdev.qpmask;
2263 		uresp.sqid = qhp->wq.sq.qid;
2264 		uresp.sq_size = qhp->wq.sq.size;
2265 		uresp.sq_memsize = qhp->wq.sq.memsize;
2266 		if (!attrs->srq) {
2267 			uresp.rqid = qhp->wq.rq.qid;
2268 			uresp.rq_size = qhp->wq.rq.size;
2269 			uresp.rq_memsize = qhp->wq.rq.memsize;
2270 		}
2271 		spin_lock(&ucontext->mmap_lock);
2272 		if (ma_sync_key_mm) {
2273 			uresp.ma_sync_key = ucontext->key;
2274 			ucontext->key += PAGE_SIZE;
2275 		}
2276 		uresp.sq_key = ucontext->key;
2277 		ucontext->key += PAGE_SIZE;
2278 		if (!attrs->srq) {
2279 			uresp.rq_key = ucontext->key;
2280 			ucontext->key += PAGE_SIZE;
2281 		}
2282 		uresp.sq_db_gts_key = ucontext->key;
2283 		ucontext->key += PAGE_SIZE;
2284 		if (!attrs->srq) {
2285 			uresp.rq_db_gts_key = ucontext->key;
2286 			ucontext->key += PAGE_SIZE;
2287 		}
2288 		spin_unlock(&ucontext->mmap_lock);
2289 		ret = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
2290 		if (ret)
2291 			goto err_free_ma_sync_key;
2292 		sq_key_mm->key = uresp.sq_key;
2293 		sq_key_mm->addr = qhp->wq.sq.phys_addr;
2294 		sq_key_mm->len = PAGE_ALIGN(qhp->wq.sq.memsize);
2295 		insert_mmap(ucontext, sq_key_mm);
2296 		if (!attrs->srq) {
2297 			rq_key_mm->key = uresp.rq_key;
2298 			rq_key_mm->addr = virt_to_phys(qhp->wq.rq.queue);
2299 			rq_key_mm->len = PAGE_ALIGN(qhp->wq.rq.memsize);
2300 			insert_mmap(ucontext, rq_key_mm);
2301 		}
2302 		sq_db_key_mm->key = uresp.sq_db_gts_key;
2303 		sq_db_key_mm->addr = (u64)(unsigned long)qhp->wq.sq.bar2_pa;
2304 		sq_db_key_mm->len = PAGE_SIZE;
2305 		insert_mmap(ucontext, sq_db_key_mm);
2306 		if (!attrs->srq) {
2307 			rq_db_key_mm->key = uresp.rq_db_gts_key;
2308 			rq_db_key_mm->addr =
2309 				(u64)(unsigned long)qhp->wq.rq.bar2_pa;
2310 			rq_db_key_mm->len = PAGE_SIZE;
2311 			insert_mmap(ucontext, rq_db_key_mm);
2312 		}
2313 		if (ma_sync_key_mm) {
2314 			ma_sync_key_mm->key = uresp.ma_sync_key;
2315 			ma_sync_key_mm->addr =
2316 				(pci_resource_start(rhp->rdev.lldi.pdev, 0) +
2317 				PCIE_MA_SYNC_A) & PAGE_MASK;
2318 			ma_sync_key_mm->len = PAGE_SIZE;
2319 			insert_mmap(ucontext, ma_sync_key_mm);
2320 		}
2321 
2322 		qhp->ucontext = ucontext;
2323 	}
2324 	if (!attrs->srq) {
2325 		qhp->wq.qp_errp =
2326 			&qhp->wq.rq.queue[qhp->wq.rq.size].status.qp_err;
2327 	} else {
2328 		qhp->wq.qp_errp =
2329 			&qhp->wq.sq.queue[qhp->wq.sq.size].status.qp_err;
2330 		qhp->wq.srqidxp =
2331 			&qhp->wq.sq.queue[qhp->wq.sq.size].status.srqidx;
2332 	}
2333 
2334 	qhp->ibqp.qp_num = qhp->wq.sq.qid;
2335 	if (attrs->srq)
2336 		qhp->srq = to_c4iw_srq(attrs->srq);
2337 	INIT_LIST_HEAD(&qhp->db_fc_entry);
2338 	pr_debug("sq id %u size %u memsize %zu num_entries %u rq id %u size %u memsize %zu num_entries %u\n",
2339 		 qhp->wq.sq.qid, qhp->wq.sq.size, qhp->wq.sq.memsize,
2340 		 attrs->cap.max_send_wr, qhp->wq.rq.qid, qhp->wq.rq.size,
2341 		 qhp->wq.rq.memsize, attrs->cap.max_recv_wr);
2342 	return &qhp->ibqp;
2343 err_free_ma_sync_key:
2344 	kfree(ma_sync_key_mm);
2345 err_free_rq_db_key:
2346 	if (!attrs->srq)
2347 		kfree(rq_db_key_mm);
2348 err_free_sq_db_key:
2349 	kfree(sq_db_key_mm);
2350 err_free_rq_key:
2351 	if (!attrs->srq)
2352 		kfree(rq_key_mm);
2353 err_free_sq_key:
2354 	kfree(sq_key_mm);
2355 err_remove_handle:
2356 	xa_erase_irq(&rhp->qps, qhp->wq.sq.qid);
2357 err_destroy_qp:
2358 	destroy_qp(&rhp->rdev, &qhp->wq,
2359 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx, !attrs->srq);
2360 err_free_wr_wait:
2361 	c4iw_put_wr_wait(qhp->wr_waitp);
2362 err_free_qhp:
2363 	kfree(qhp);
2364 	return ERR_PTR(ret);
2365 }
2366 
2367 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2368 		      int attr_mask, struct ib_udata *udata)
2369 {
2370 	struct c4iw_dev *rhp;
2371 	struct c4iw_qp *qhp;
2372 	enum c4iw_qp_attr_mask mask = 0;
2373 	struct c4iw_qp_attributes attrs = {};
2374 
2375 	pr_debug("ib_qp %p\n", ibqp);
2376 
2377 	/* iwarp does not support the RTR state */
2378 	if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
2379 		attr_mask &= ~IB_QP_STATE;
2380 
2381 	/* Make sure we still have something left to do */
2382 	if (!attr_mask)
2383 		return 0;
2384 
2385 	qhp = to_c4iw_qp(ibqp);
2386 	rhp = qhp->rhp;
2387 
2388 	attrs.next_state = c4iw_convert_state(attr->qp_state);
2389 	attrs.enable_rdma_read = (attr->qp_access_flags &
2390 			       IB_ACCESS_REMOTE_READ) ?  1 : 0;
2391 	attrs.enable_rdma_write = (attr->qp_access_flags &
2392 				IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
2393 	attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
2394 
2395 
2396 	mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
2397 	mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
2398 			(C4IW_QP_ATTR_ENABLE_RDMA_READ |
2399 			 C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
2400 			 C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
2401 
2402 	/*
2403 	 * Use SQ_PSN and RQ_PSN to pass in IDX_INC values for
2404 	 * ringing the queue db when we're in DB_FULL mode.
2405 	 * Only allow this on T4 devices.
2406 	 */
2407 	attrs.sq_db_inc = attr->sq_psn;
2408 	attrs.rq_db_inc = attr->rq_psn;
2409 	mask |= (attr_mask & IB_QP_SQ_PSN) ? C4IW_QP_ATTR_SQ_DB : 0;
2410 	mask |= (attr_mask & IB_QP_RQ_PSN) ? C4IW_QP_ATTR_RQ_DB : 0;
2411 	if (!is_t4(to_c4iw_qp(ibqp)->rhp->rdev.lldi.adapter_type) &&
2412 	    (mask & (C4IW_QP_ATTR_SQ_DB|C4IW_QP_ATTR_RQ_DB)))
2413 		return -EINVAL;
2414 
2415 	return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
2416 }
2417 
2418 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
2419 {
2420 	pr_debug("ib_dev %p qpn 0x%x\n", dev, qpn);
2421 	return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
2422 }
2423 
2424 void c4iw_dispatch_srq_limit_reached_event(struct c4iw_srq *srq)
2425 {
2426 	struct ib_event event = {};
2427 
2428 	event.device = &srq->rhp->ibdev;
2429 	event.element.srq = &srq->ibsrq;
2430 	event.event = IB_EVENT_SRQ_LIMIT_REACHED;
2431 	ib_dispatch_event(&event);
2432 }
2433 
2434 int c4iw_modify_srq(struct ib_srq *ib_srq, struct ib_srq_attr *attr,
2435 		    enum ib_srq_attr_mask srq_attr_mask,
2436 		    struct ib_udata *udata)
2437 {
2438 	struct c4iw_srq *srq = to_c4iw_srq(ib_srq);
2439 	int ret = 0;
2440 
2441 	/*
2442 	 * XXX 0 mask == a SW interrupt for srq_limit reached...
2443 	 */
2444 	if (udata && !srq_attr_mask) {
2445 		c4iw_dispatch_srq_limit_reached_event(srq);
2446 		goto out;
2447 	}
2448 
2449 	/* no support for this yet */
2450 	if (srq_attr_mask & IB_SRQ_MAX_WR) {
2451 		ret = -EINVAL;
2452 		goto out;
2453 	}
2454 
2455 	if (!udata && (srq_attr_mask & IB_SRQ_LIMIT)) {
2456 		srq->armed = true;
2457 		srq->srq_limit = attr->srq_limit;
2458 	}
2459 out:
2460 	return ret;
2461 }
2462 
2463 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2464 		     int attr_mask, struct ib_qp_init_attr *init_attr)
2465 {
2466 	struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
2467 
2468 	memset(attr, 0, sizeof(*attr));
2469 	memset(init_attr, 0, sizeof(*init_attr));
2470 	attr->qp_state = to_ib_qp_state(qhp->attr.state);
2471 	init_attr->cap.max_send_wr = qhp->attr.sq_num_entries;
2472 	init_attr->cap.max_recv_wr = qhp->attr.rq_num_entries;
2473 	init_attr->cap.max_send_sge = qhp->attr.sq_max_sges;
2474 	init_attr->cap.max_recv_sge = qhp->attr.sq_max_sges;
2475 	init_attr->cap.max_inline_data = T4_MAX_SEND_INLINE;
2476 	init_attr->sq_sig_type = qhp->sq_sig_all ? IB_SIGNAL_ALL_WR : 0;
2477 	return 0;
2478 }
2479 
2480 static void free_srq_queue(struct c4iw_srq *srq, struct c4iw_dev_ucontext *uctx,
2481 			   struct c4iw_wr_wait *wr_waitp)
2482 {
2483 	struct c4iw_rdev *rdev = &srq->rhp->rdev;
2484 	struct sk_buff *skb = srq->destroy_skb;
2485 	struct t4_srq *wq = &srq->wq;
2486 	struct fw_ri_res_wr *res_wr;
2487 	struct fw_ri_res *res;
2488 	int wr_len;
2489 
2490 	wr_len = sizeof(*res_wr) + sizeof(*res);
2491 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
2492 
2493 	res_wr = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
2494 	memset(res_wr, 0, wr_len);
2495 	res_wr->op_nres = cpu_to_be32(FW_WR_OP_V(FW_RI_RES_WR) |
2496 			FW_RI_RES_WR_NRES_V(1) |
2497 			FW_WR_COMPL_F);
2498 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
2499 	res_wr->cookie = (uintptr_t)wr_waitp;
2500 	res = res_wr->res;
2501 	res->u.srq.restype = FW_RI_RES_TYPE_SRQ;
2502 	res->u.srq.op = FW_RI_RES_OP_RESET;
2503 	res->u.srq.srqid = cpu_to_be32(srq->idx);
2504 	res->u.srq.eqid = cpu_to_be32(wq->qid);
2505 
2506 	c4iw_init_wr_wait(wr_waitp);
2507 	c4iw_ref_send_wait(rdev, skb, wr_waitp, 0, 0, __func__);
2508 
2509 	dma_free_coherent(&rdev->lldi.pdev->dev,
2510 			  wq->memsize, wq->queue,
2511 			dma_unmap_addr(wq, mapping));
2512 	c4iw_rqtpool_free(rdev, wq->rqt_hwaddr, wq->rqt_size);
2513 	kfree(wq->sw_rq);
2514 	c4iw_put_qpid(rdev, wq->qid, uctx);
2515 }
2516 
2517 static int alloc_srq_queue(struct c4iw_srq *srq, struct c4iw_dev_ucontext *uctx,
2518 			   struct c4iw_wr_wait *wr_waitp)
2519 {
2520 	struct c4iw_rdev *rdev = &srq->rhp->rdev;
2521 	int user = (uctx != &rdev->uctx);
2522 	struct t4_srq *wq = &srq->wq;
2523 	struct fw_ri_res_wr *res_wr;
2524 	struct fw_ri_res *res;
2525 	struct sk_buff *skb;
2526 	int wr_len;
2527 	int eqsize;
2528 	int ret = -ENOMEM;
2529 
2530 	wq->qid = c4iw_get_qpid(rdev, uctx);
2531 	if (!wq->qid)
2532 		goto err;
2533 
2534 	if (!user) {
2535 		wq->sw_rq = kcalloc(wq->size, sizeof(*wq->sw_rq),
2536 				    GFP_KERNEL);
2537 		if (!wq->sw_rq)
2538 			goto err_put_qpid;
2539 		wq->pending_wrs = kcalloc(srq->wq.size,
2540 					  sizeof(*srq->wq.pending_wrs),
2541 					  GFP_KERNEL);
2542 		if (!wq->pending_wrs)
2543 			goto err_free_sw_rq;
2544 	}
2545 
2546 	wq->rqt_size = wq->size;
2547 	wq->rqt_hwaddr = c4iw_rqtpool_alloc(rdev, wq->rqt_size);
2548 	if (!wq->rqt_hwaddr)
2549 		goto err_free_pending_wrs;
2550 	wq->rqt_abs_idx = (wq->rqt_hwaddr - rdev->lldi.vr->rq.start) >>
2551 		T4_RQT_ENTRY_SHIFT;
2552 
2553 	wq->queue = dma_alloc_coherent(&rdev->lldi.pdev->dev, wq->memsize,
2554 				       &wq->dma_addr, GFP_KERNEL);
2555 	if (!wq->queue)
2556 		goto err_free_rqtpool;
2557 
2558 	dma_unmap_addr_set(wq, mapping, wq->dma_addr);
2559 
2560 	wq->bar2_va = c4iw_bar2_addrs(rdev, wq->qid, CXGB4_BAR2_QTYPE_EGRESS,
2561 				      &wq->bar2_qid,
2562 			user ? &wq->bar2_pa : NULL);
2563 
2564 	/*
2565 	 * User mode must have bar2 access.
2566 	 */
2567 
2568 	if (user && !wq->bar2_va) {
2569 		pr_warn(MOD "%s: srqid %u not in BAR2 range.\n",
2570 			pci_name(rdev->lldi.pdev), wq->qid);
2571 		ret = -EINVAL;
2572 		goto err_free_queue;
2573 	}
2574 
2575 	/* build fw_ri_res_wr */
2576 	wr_len = sizeof(*res_wr) + sizeof(*res);
2577 
2578 	skb = alloc_skb(wr_len, GFP_KERNEL);
2579 	if (!skb)
2580 		goto err_free_queue;
2581 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
2582 
2583 	res_wr = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
2584 	memset(res_wr, 0, wr_len);
2585 	res_wr->op_nres = cpu_to_be32(FW_WR_OP_V(FW_RI_RES_WR) |
2586 			FW_RI_RES_WR_NRES_V(1) |
2587 			FW_WR_COMPL_F);
2588 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
2589 	res_wr->cookie = (uintptr_t)wr_waitp;
2590 	res = res_wr->res;
2591 	res->u.srq.restype = FW_RI_RES_TYPE_SRQ;
2592 	res->u.srq.op = FW_RI_RES_OP_WRITE;
2593 
2594 	/*
2595 	 * eqsize is the number of 64B entries plus the status page size.
2596 	 */
2597 	eqsize = wq->size * T4_RQ_NUM_SLOTS +
2598 		rdev->hw_queue.t4_eq_status_entries;
2599 	res->u.srq.eqid = cpu_to_be32(wq->qid);
2600 	res->u.srq.fetchszm_to_iqid =
2601 						/* no host cidx updates */
2602 		cpu_to_be32(FW_RI_RES_WR_HOSTFCMODE_V(0) |
2603 		FW_RI_RES_WR_CPRIO_V(0) |       /* don't keep in chip cache */
2604 		FW_RI_RES_WR_PCIECHN_V(0) |     /* set by uP at ri_init time */
2605 		FW_RI_RES_WR_FETCHRO_V(0));     /* relaxed_ordering */
2606 	res->u.srq.dcaen_to_eqsize =
2607 		cpu_to_be32(FW_RI_RES_WR_DCAEN_V(0) |
2608 		FW_RI_RES_WR_DCACPU_V(0) |
2609 		FW_RI_RES_WR_FBMIN_V(2) |
2610 		FW_RI_RES_WR_FBMAX_V(3) |
2611 		FW_RI_RES_WR_CIDXFTHRESHO_V(0) |
2612 		FW_RI_RES_WR_CIDXFTHRESH_V(0) |
2613 		FW_RI_RES_WR_EQSIZE_V(eqsize));
2614 	res->u.srq.eqaddr = cpu_to_be64(wq->dma_addr);
2615 	res->u.srq.srqid = cpu_to_be32(srq->idx);
2616 	res->u.srq.pdid = cpu_to_be32(srq->pdid);
2617 	res->u.srq.hwsrqsize = cpu_to_be32(wq->rqt_size);
2618 	res->u.srq.hwsrqaddr = cpu_to_be32(wq->rqt_hwaddr -
2619 			rdev->lldi.vr->rq.start);
2620 
2621 	c4iw_init_wr_wait(wr_waitp);
2622 
2623 	ret = c4iw_ref_send_wait(rdev, skb, wr_waitp, 0, wq->qid, __func__);
2624 	if (ret)
2625 		goto err_free_queue;
2626 
2627 	pr_debug("%s srq %u eqid %u pdid %u queue va %p pa 0x%llx\n"
2628 			" bar2_addr %p rqt addr 0x%x size %d\n",
2629 			__func__, srq->idx, wq->qid, srq->pdid, wq->queue,
2630 			(u64)virt_to_phys(wq->queue), wq->bar2_va,
2631 			wq->rqt_hwaddr, wq->rqt_size);
2632 
2633 	return 0;
2634 err_free_queue:
2635 	dma_free_coherent(&rdev->lldi.pdev->dev,
2636 			  wq->memsize, wq->queue,
2637 			dma_unmap_addr(wq, mapping));
2638 err_free_rqtpool:
2639 	c4iw_rqtpool_free(rdev, wq->rqt_hwaddr, wq->rqt_size);
2640 err_free_pending_wrs:
2641 	if (!user)
2642 		kfree(wq->pending_wrs);
2643 err_free_sw_rq:
2644 	if (!user)
2645 		kfree(wq->sw_rq);
2646 err_put_qpid:
2647 	c4iw_put_qpid(rdev, wq->qid, uctx);
2648 err:
2649 	return ret;
2650 }
2651 
2652 void c4iw_copy_wr_to_srq(struct t4_srq *srq, union t4_recv_wr *wqe, u8 len16)
2653 {
2654 	u64 *src, *dst;
2655 
2656 	src = (u64 *)wqe;
2657 	dst = (u64 *)((u8 *)srq->queue + srq->wq_pidx * T4_EQ_ENTRY_SIZE);
2658 	while (len16) {
2659 		*dst++ = *src++;
2660 		if (dst >= (u64 *)&srq->queue[srq->size])
2661 			dst = (u64 *)srq->queue;
2662 		*dst++ = *src++;
2663 		if (dst >= (u64 *)&srq->queue[srq->size])
2664 			dst = (u64 *)srq->queue;
2665 		len16--;
2666 	}
2667 }
2668 
2669 int c4iw_create_srq(struct ib_srq *ib_srq, struct ib_srq_init_attr *attrs,
2670 			       struct ib_udata *udata)
2671 {
2672 	struct ib_pd *pd = ib_srq->pd;
2673 	struct c4iw_dev *rhp;
2674 	struct c4iw_srq *srq = to_c4iw_srq(ib_srq);
2675 	struct c4iw_pd *php;
2676 	struct c4iw_create_srq_resp uresp;
2677 	struct c4iw_ucontext *ucontext;
2678 	struct c4iw_mm_entry *srq_key_mm, *srq_db_key_mm;
2679 	int rqsize;
2680 	int ret;
2681 	int wr_len;
2682 
2683 	pr_debug("%s ib_pd %p\n", __func__, pd);
2684 
2685 	php = to_c4iw_pd(pd);
2686 	rhp = php->rhp;
2687 
2688 	if (!rhp->rdev.lldi.vr->srq.size)
2689 		return -EINVAL;
2690 	if (attrs->attr.max_wr > rhp->rdev.hw_queue.t4_max_rq_size)
2691 		return -E2BIG;
2692 	if (attrs->attr.max_sge > T4_MAX_RECV_SGE)
2693 		return -E2BIG;
2694 
2695 	/*
2696 	 * SRQ RQT and RQ must be a power of 2 and at least 16 deep.
2697 	 */
2698 	rqsize = attrs->attr.max_wr + 1;
2699 	rqsize = roundup_pow_of_two(max_t(u16, rqsize, 16));
2700 
2701 	ucontext = rdma_udata_to_drv_context(udata, struct c4iw_ucontext,
2702 					     ibucontext);
2703 
2704 	srq->wr_waitp = c4iw_alloc_wr_wait(GFP_KERNEL);
2705 	if (!srq->wr_waitp)
2706 		return -ENOMEM;
2707 
2708 	srq->idx = c4iw_alloc_srq_idx(&rhp->rdev);
2709 	if (srq->idx < 0) {
2710 		ret = -ENOMEM;
2711 		goto err_free_wr_wait;
2712 	}
2713 
2714 	wr_len = sizeof(struct fw_ri_res_wr) + sizeof(struct fw_ri_res);
2715 	srq->destroy_skb = alloc_skb(wr_len, GFP_KERNEL);
2716 	if (!srq->destroy_skb) {
2717 		ret = -ENOMEM;
2718 		goto err_free_srq_idx;
2719 	}
2720 
2721 	srq->rhp = rhp;
2722 	srq->pdid = php->pdid;
2723 
2724 	srq->wq.size = rqsize;
2725 	srq->wq.memsize =
2726 		(rqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
2727 		sizeof(*srq->wq.queue);
2728 	if (ucontext)
2729 		srq->wq.memsize = roundup(srq->wq.memsize, PAGE_SIZE);
2730 
2731 	ret = alloc_srq_queue(srq, ucontext ? &ucontext->uctx :
2732 			&rhp->rdev.uctx, srq->wr_waitp);
2733 	if (ret)
2734 		goto err_free_skb;
2735 	attrs->attr.max_wr = rqsize - 1;
2736 
2737 	if (CHELSIO_CHIP_VERSION(rhp->rdev.lldi.adapter_type) > CHELSIO_T6)
2738 		srq->flags = T4_SRQ_LIMIT_SUPPORT;
2739 
2740 	if (udata) {
2741 		srq_key_mm = kmalloc(sizeof(*srq_key_mm), GFP_KERNEL);
2742 		if (!srq_key_mm) {
2743 			ret = -ENOMEM;
2744 			goto err_free_queue;
2745 		}
2746 		srq_db_key_mm = kmalloc(sizeof(*srq_db_key_mm), GFP_KERNEL);
2747 		if (!srq_db_key_mm) {
2748 			ret = -ENOMEM;
2749 			goto err_free_srq_key_mm;
2750 		}
2751 		memset(&uresp, 0, sizeof(uresp));
2752 		uresp.flags = srq->flags;
2753 		uresp.qid_mask = rhp->rdev.qpmask;
2754 		uresp.srqid = srq->wq.qid;
2755 		uresp.srq_size = srq->wq.size;
2756 		uresp.srq_memsize = srq->wq.memsize;
2757 		uresp.rqt_abs_idx = srq->wq.rqt_abs_idx;
2758 		spin_lock(&ucontext->mmap_lock);
2759 		uresp.srq_key = ucontext->key;
2760 		ucontext->key += PAGE_SIZE;
2761 		uresp.srq_db_gts_key = ucontext->key;
2762 		ucontext->key += PAGE_SIZE;
2763 		spin_unlock(&ucontext->mmap_lock);
2764 		ret = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
2765 		if (ret)
2766 			goto err_free_srq_db_key_mm;
2767 		srq_key_mm->key = uresp.srq_key;
2768 		srq_key_mm->addr = virt_to_phys(srq->wq.queue);
2769 		srq_key_mm->len = PAGE_ALIGN(srq->wq.memsize);
2770 		insert_mmap(ucontext, srq_key_mm);
2771 		srq_db_key_mm->key = uresp.srq_db_gts_key;
2772 		srq_db_key_mm->addr = (u64)(unsigned long)srq->wq.bar2_pa;
2773 		srq_db_key_mm->len = PAGE_SIZE;
2774 		insert_mmap(ucontext, srq_db_key_mm);
2775 	}
2776 
2777 	pr_debug("%s srq qid %u idx %u size %u memsize %lu num_entries %u\n",
2778 		 __func__, srq->wq.qid, srq->idx, srq->wq.size,
2779 			(unsigned long)srq->wq.memsize, attrs->attr.max_wr);
2780 
2781 	spin_lock_init(&srq->lock);
2782 	return 0;
2783 
2784 err_free_srq_db_key_mm:
2785 	kfree(srq_db_key_mm);
2786 err_free_srq_key_mm:
2787 	kfree(srq_key_mm);
2788 err_free_queue:
2789 	free_srq_queue(srq, ucontext ? &ucontext->uctx : &rhp->rdev.uctx,
2790 		       srq->wr_waitp);
2791 err_free_skb:
2792 	kfree_skb(srq->destroy_skb);
2793 err_free_srq_idx:
2794 	c4iw_free_srq_idx(&rhp->rdev, srq->idx);
2795 err_free_wr_wait:
2796 	c4iw_put_wr_wait(srq->wr_waitp);
2797 	return ret;
2798 }
2799 
2800 void c4iw_destroy_srq(struct ib_srq *ibsrq, struct ib_udata *udata)
2801 {
2802 	struct c4iw_dev *rhp;
2803 	struct c4iw_srq *srq;
2804 	struct c4iw_ucontext *ucontext;
2805 
2806 	srq = to_c4iw_srq(ibsrq);
2807 	rhp = srq->rhp;
2808 
2809 	pr_debug("%s id %d\n", __func__, srq->wq.qid);
2810 	ucontext = rdma_udata_to_drv_context(udata, struct c4iw_ucontext,
2811 					     ibucontext);
2812 	free_srq_queue(srq, ucontext ? &ucontext->uctx : &rhp->rdev.uctx,
2813 		       srq->wr_waitp);
2814 	c4iw_free_srq_idx(&rhp->rdev, srq->idx);
2815 	c4iw_put_wr_wait(srq->wr_waitp);
2816 }
2817