xref: /linux/drivers/infiniband/hw/cxgb4/qp.c (revision d2912cb15bdda8ba4a5dd73396ad62641af2f520)
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 		memset(wq->rq.queue, 0, wq->rq.memsize);
278 		dma_unmap_addr_set(&wq->rq, mapping, wq->rq.dma_addr);
279 	}
280 
281 	wq->db = rdev->lldi.db_reg;
282 
283 	wq->sq.bar2_va = c4iw_bar2_addrs(rdev, wq->sq.qid,
284 					 CXGB4_BAR2_QTYPE_EGRESS,
285 					 &wq->sq.bar2_qid,
286 					 user ? &wq->sq.bar2_pa : NULL);
287 	if (need_rq)
288 		wq->rq.bar2_va = c4iw_bar2_addrs(rdev, wq->rq.qid,
289 						 CXGB4_BAR2_QTYPE_EGRESS,
290 						 &wq->rq.bar2_qid,
291 						 user ? &wq->rq.bar2_pa : NULL);
292 
293 	/*
294 	 * User mode must have bar2 access.
295 	 */
296 	if (user && (!wq->sq.bar2_pa || (need_rq && !wq->rq.bar2_pa))) {
297 		pr_warn("%s: sqid %u or rqid %u not in BAR2 range\n",
298 			pci_name(rdev->lldi.pdev), wq->sq.qid, wq->rq.qid);
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(sizeof wqe->recv +
770 			      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 static void free_qp_work(struct work_struct *work)
894 {
895 	struct c4iw_ucontext *ucontext;
896 	struct c4iw_qp *qhp;
897 	struct c4iw_dev *rhp;
898 
899 	qhp = container_of(work, struct c4iw_qp, free_work);
900 	ucontext = qhp->ucontext;
901 	rhp = qhp->rhp;
902 
903 	pr_debug("qhp %p ucontext %p\n", qhp, ucontext);
904 	destroy_qp(&rhp->rdev, &qhp->wq,
905 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx, !qhp->srq);
906 
907 	c4iw_put_wr_wait(qhp->wr_waitp);
908 	kfree(qhp);
909 }
910 
911 static void queue_qp_free(struct kref *kref)
912 {
913 	struct c4iw_qp *qhp;
914 
915 	qhp = container_of(kref, struct c4iw_qp, kref);
916 	pr_debug("qhp %p\n", qhp);
917 	queue_work(qhp->rhp->rdev.free_workq, &qhp->free_work);
918 }
919 
920 void c4iw_qp_add_ref(struct ib_qp *qp)
921 {
922 	pr_debug("ib_qp %p\n", qp);
923 	kref_get(&to_c4iw_qp(qp)->kref);
924 }
925 
926 void c4iw_qp_rem_ref(struct ib_qp *qp)
927 {
928 	pr_debug("ib_qp %p\n", qp);
929 	kref_put(&to_c4iw_qp(qp)->kref, queue_qp_free);
930 }
931 
932 static void add_to_fc_list(struct list_head *head, struct list_head *entry)
933 {
934 	if (list_empty(entry))
935 		list_add_tail(entry, head);
936 }
937 
938 static int ring_kernel_sq_db(struct c4iw_qp *qhp, u16 inc)
939 {
940 	unsigned long flags;
941 
942 	xa_lock_irqsave(&qhp->rhp->qps, flags);
943 	spin_lock(&qhp->lock);
944 	if (qhp->rhp->db_state == NORMAL)
945 		t4_ring_sq_db(&qhp->wq, inc, NULL);
946 	else {
947 		add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
948 		qhp->wq.sq.wq_pidx_inc += inc;
949 	}
950 	spin_unlock(&qhp->lock);
951 	xa_unlock_irqrestore(&qhp->rhp->qps, flags);
952 	return 0;
953 }
954 
955 static int ring_kernel_rq_db(struct c4iw_qp *qhp, u16 inc)
956 {
957 	unsigned long flags;
958 
959 	xa_lock_irqsave(&qhp->rhp->qps, flags);
960 	spin_lock(&qhp->lock);
961 	if (qhp->rhp->db_state == NORMAL)
962 		t4_ring_rq_db(&qhp->wq, inc, NULL);
963 	else {
964 		add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
965 		qhp->wq.rq.wq_pidx_inc += inc;
966 	}
967 	spin_unlock(&qhp->lock);
968 	xa_unlock_irqrestore(&qhp->rhp->qps, flags);
969 	return 0;
970 }
971 
972 static int ib_to_fw_opcode(int ib_opcode)
973 {
974 	int opcode;
975 
976 	switch (ib_opcode) {
977 	case IB_WR_SEND_WITH_INV:
978 		opcode = FW_RI_SEND_WITH_INV;
979 		break;
980 	case IB_WR_SEND:
981 		opcode = FW_RI_SEND;
982 		break;
983 	case IB_WR_RDMA_WRITE:
984 		opcode = FW_RI_RDMA_WRITE;
985 		break;
986 	case IB_WR_RDMA_WRITE_WITH_IMM:
987 		opcode = FW_RI_WRITE_IMMEDIATE;
988 		break;
989 	case IB_WR_RDMA_READ:
990 	case IB_WR_RDMA_READ_WITH_INV:
991 		opcode = FW_RI_READ_REQ;
992 		break;
993 	case IB_WR_REG_MR:
994 		opcode = FW_RI_FAST_REGISTER;
995 		break;
996 	case IB_WR_LOCAL_INV:
997 		opcode = FW_RI_LOCAL_INV;
998 		break;
999 	default:
1000 		opcode = -EINVAL;
1001 	}
1002 	return opcode;
1003 }
1004 
1005 static int complete_sq_drain_wr(struct c4iw_qp *qhp,
1006 				const struct ib_send_wr *wr)
1007 {
1008 	struct t4_cqe cqe = {};
1009 	struct c4iw_cq *schp;
1010 	unsigned long flag;
1011 	struct t4_cq *cq;
1012 	int opcode;
1013 
1014 	schp = to_c4iw_cq(qhp->ibqp.send_cq);
1015 	cq = &schp->cq;
1016 
1017 	opcode = ib_to_fw_opcode(wr->opcode);
1018 	if (opcode < 0)
1019 		return opcode;
1020 
1021 	cqe.u.drain_cookie = wr->wr_id;
1022 	cqe.header = cpu_to_be32(CQE_STATUS_V(T4_ERR_SWFLUSH) |
1023 				 CQE_OPCODE_V(opcode) |
1024 				 CQE_TYPE_V(1) |
1025 				 CQE_SWCQE_V(1) |
1026 				 CQE_DRAIN_V(1) |
1027 				 CQE_QPID_V(qhp->wq.sq.qid));
1028 
1029 	spin_lock_irqsave(&schp->lock, flag);
1030 	cqe.bits_type_ts = cpu_to_be64(CQE_GENBIT_V((u64)cq->gen));
1031 	cq->sw_queue[cq->sw_pidx] = cqe;
1032 	t4_swcq_produce(cq);
1033 	spin_unlock_irqrestore(&schp->lock, flag);
1034 
1035 	if (t4_clear_cq_armed(&schp->cq)) {
1036 		spin_lock_irqsave(&schp->comp_handler_lock, flag);
1037 		(*schp->ibcq.comp_handler)(&schp->ibcq,
1038 					   schp->ibcq.cq_context);
1039 		spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1040 	}
1041 	return 0;
1042 }
1043 
1044 static int complete_sq_drain_wrs(struct c4iw_qp *qhp,
1045 				 const struct ib_send_wr *wr,
1046 				 const struct ib_send_wr **bad_wr)
1047 {
1048 	int ret = 0;
1049 
1050 	while (wr) {
1051 		ret = complete_sq_drain_wr(qhp, wr);
1052 		if (ret) {
1053 			*bad_wr = wr;
1054 			break;
1055 		}
1056 		wr = wr->next;
1057 	}
1058 	return ret;
1059 }
1060 
1061 static void complete_rq_drain_wr(struct c4iw_qp *qhp,
1062 				 const struct ib_recv_wr *wr)
1063 {
1064 	struct t4_cqe cqe = {};
1065 	struct c4iw_cq *rchp;
1066 	unsigned long flag;
1067 	struct t4_cq *cq;
1068 
1069 	rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
1070 	cq = &rchp->cq;
1071 
1072 	cqe.u.drain_cookie = wr->wr_id;
1073 	cqe.header = cpu_to_be32(CQE_STATUS_V(T4_ERR_SWFLUSH) |
1074 				 CQE_OPCODE_V(FW_RI_SEND) |
1075 				 CQE_TYPE_V(0) |
1076 				 CQE_SWCQE_V(1) |
1077 				 CQE_DRAIN_V(1) |
1078 				 CQE_QPID_V(qhp->wq.sq.qid));
1079 
1080 	spin_lock_irqsave(&rchp->lock, flag);
1081 	cqe.bits_type_ts = cpu_to_be64(CQE_GENBIT_V((u64)cq->gen));
1082 	cq->sw_queue[cq->sw_pidx] = cqe;
1083 	t4_swcq_produce(cq);
1084 	spin_unlock_irqrestore(&rchp->lock, flag);
1085 
1086 	if (t4_clear_cq_armed(&rchp->cq)) {
1087 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1088 		(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1089 					   rchp->ibcq.cq_context);
1090 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1091 	}
1092 }
1093 
1094 static void complete_rq_drain_wrs(struct c4iw_qp *qhp,
1095 				  const struct ib_recv_wr *wr)
1096 {
1097 	while (wr) {
1098 		complete_rq_drain_wr(qhp, wr);
1099 		wr = wr->next;
1100 	}
1101 }
1102 
1103 int c4iw_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
1104 		   const struct ib_send_wr **bad_wr)
1105 {
1106 	int err = 0;
1107 	u8 len16 = 0;
1108 	enum fw_wr_opcodes fw_opcode = 0;
1109 	enum fw_ri_wr_flags fw_flags;
1110 	struct c4iw_qp *qhp;
1111 	struct c4iw_dev *rhp;
1112 	union t4_wr *wqe = NULL;
1113 	u32 num_wrs;
1114 	struct t4_swsqe *swsqe;
1115 	unsigned long flag;
1116 	u16 idx = 0;
1117 
1118 	qhp = to_c4iw_qp(ibqp);
1119 	rhp = qhp->rhp;
1120 	spin_lock_irqsave(&qhp->lock, flag);
1121 
1122 	/*
1123 	 * If the qp has been flushed, then just insert a special
1124 	 * drain cqe.
1125 	 */
1126 	if (qhp->wq.flushed) {
1127 		spin_unlock_irqrestore(&qhp->lock, flag);
1128 		err = complete_sq_drain_wrs(qhp, wr, bad_wr);
1129 		return err;
1130 	}
1131 	num_wrs = t4_sq_avail(&qhp->wq);
1132 	if (num_wrs == 0) {
1133 		spin_unlock_irqrestore(&qhp->lock, flag);
1134 		*bad_wr = wr;
1135 		return -ENOMEM;
1136 	}
1137 
1138 	/*
1139 	 * Fastpath for NVMe-oF target WRITE + SEND_WITH_INV wr chain which is
1140 	 * the response for small NVMEe-oF READ requests.  If the chain is
1141 	 * exactly a WRITE->SEND_WITH_INV or a WRITE->SEND and the sgl depths
1142 	 * and lengths meet the requirements of the fw_ri_write_cmpl_wr work
1143 	 * request, then build and post the write_cmpl WR. If any of the tests
1144 	 * below are not true, then we continue on with the tradtional WRITE
1145 	 * and SEND WRs.
1146 	 */
1147 	if (qhp->rhp->rdev.lldi.write_cmpl_support &&
1148 	    CHELSIO_CHIP_VERSION(qhp->rhp->rdev.lldi.adapter_type) >=
1149 	    CHELSIO_T5 &&
1150 	    wr && wr->next && !wr->next->next &&
1151 	    wr->opcode == IB_WR_RDMA_WRITE &&
1152 	    wr->sg_list[0].length && wr->num_sge <= T4_WRITE_CMPL_MAX_SGL &&
1153 	    (wr->next->opcode == IB_WR_SEND ||
1154 	    wr->next->opcode == IB_WR_SEND_WITH_INV) &&
1155 	    wr->next->sg_list[0].length == T4_WRITE_CMPL_MAX_CQE &&
1156 	    wr->next->num_sge == 1 && num_wrs >= 2) {
1157 		post_write_cmpl(qhp, wr);
1158 		spin_unlock_irqrestore(&qhp->lock, flag);
1159 		return 0;
1160 	}
1161 
1162 	while (wr) {
1163 		if (num_wrs == 0) {
1164 			err = -ENOMEM;
1165 			*bad_wr = wr;
1166 			break;
1167 		}
1168 		wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
1169 		      qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
1170 
1171 		fw_flags = 0;
1172 		if (wr->send_flags & IB_SEND_SOLICITED)
1173 			fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
1174 		if (wr->send_flags & IB_SEND_SIGNALED || qhp->sq_sig_all)
1175 			fw_flags |= FW_RI_COMPLETION_FLAG;
1176 		swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
1177 		switch (wr->opcode) {
1178 		case IB_WR_SEND_WITH_INV:
1179 		case IB_WR_SEND:
1180 			if (wr->send_flags & IB_SEND_FENCE)
1181 				fw_flags |= FW_RI_READ_FENCE_FLAG;
1182 			fw_opcode = FW_RI_SEND_WR;
1183 			if (wr->opcode == IB_WR_SEND)
1184 				swsqe->opcode = FW_RI_SEND;
1185 			else
1186 				swsqe->opcode = FW_RI_SEND_WITH_INV;
1187 			err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
1188 			break;
1189 		case IB_WR_RDMA_WRITE_WITH_IMM:
1190 			if (unlikely(!rhp->rdev.lldi.write_w_imm_support)) {
1191 				err = -EINVAL;
1192 				break;
1193 			}
1194 			fw_flags |= FW_RI_RDMA_WRITE_WITH_IMMEDIATE;
1195 			/*FALLTHROUGH*/
1196 		case IB_WR_RDMA_WRITE:
1197 			fw_opcode = FW_RI_RDMA_WRITE_WR;
1198 			swsqe->opcode = FW_RI_RDMA_WRITE;
1199 			err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
1200 			break;
1201 		case IB_WR_RDMA_READ:
1202 		case IB_WR_RDMA_READ_WITH_INV:
1203 			fw_opcode = FW_RI_RDMA_READ_WR;
1204 			swsqe->opcode = FW_RI_READ_REQ;
1205 			if (wr->opcode == IB_WR_RDMA_READ_WITH_INV) {
1206 				c4iw_invalidate_mr(rhp, wr->sg_list[0].lkey);
1207 				fw_flags = FW_RI_RDMA_READ_INVALIDATE;
1208 			} else {
1209 				fw_flags = 0;
1210 			}
1211 			err = build_rdma_read(wqe, wr, &len16);
1212 			if (err)
1213 				break;
1214 			swsqe->read_len = wr->sg_list[0].length;
1215 			if (!qhp->wq.sq.oldest_read)
1216 				qhp->wq.sq.oldest_read = swsqe;
1217 			break;
1218 		case IB_WR_REG_MR: {
1219 			struct c4iw_mr *mhp = to_c4iw_mr(reg_wr(wr)->mr);
1220 
1221 			swsqe->opcode = FW_RI_FAST_REGISTER;
1222 			if (rhp->rdev.lldi.fr_nsmr_tpte_wr_support &&
1223 			    !mhp->attr.state && mhp->mpl_len <= 2) {
1224 				fw_opcode = FW_RI_FR_NSMR_TPTE_WR;
1225 				build_tpte_memreg(&wqe->fr_tpte, reg_wr(wr),
1226 						  mhp, &len16);
1227 			} else {
1228 				fw_opcode = FW_RI_FR_NSMR_WR;
1229 				err = build_memreg(&qhp->wq.sq, wqe, reg_wr(wr),
1230 				       mhp, &len16,
1231 				       rhp->rdev.lldi.ulptx_memwrite_dsgl);
1232 				if (err)
1233 					break;
1234 			}
1235 			mhp->attr.state = 1;
1236 			break;
1237 		}
1238 		case IB_WR_LOCAL_INV:
1239 			if (wr->send_flags & IB_SEND_FENCE)
1240 				fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
1241 			fw_opcode = FW_RI_INV_LSTAG_WR;
1242 			swsqe->opcode = FW_RI_LOCAL_INV;
1243 			err = build_inv_stag(wqe, wr, &len16);
1244 			c4iw_invalidate_mr(rhp, wr->ex.invalidate_rkey);
1245 			break;
1246 		default:
1247 			pr_warn("%s post of type=%d TBD!\n", __func__,
1248 				wr->opcode);
1249 			err = -EINVAL;
1250 		}
1251 		if (err) {
1252 			*bad_wr = wr;
1253 			break;
1254 		}
1255 		swsqe->idx = qhp->wq.sq.pidx;
1256 		swsqe->complete = 0;
1257 		swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED) ||
1258 				  qhp->sq_sig_all;
1259 		swsqe->flushed = 0;
1260 		swsqe->wr_id = wr->wr_id;
1261 		if (c4iw_wr_log) {
1262 			swsqe->sge_ts = cxgb4_read_sge_timestamp(
1263 					rhp->rdev.lldi.ports[0]);
1264 			swsqe->host_time = ktime_get();
1265 		}
1266 
1267 		init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
1268 
1269 		pr_debug("cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u\n",
1270 			 (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
1271 			 swsqe->opcode, swsqe->read_len);
1272 		wr = wr->next;
1273 		num_wrs--;
1274 		t4_sq_produce(&qhp->wq, len16);
1275 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
1276 	}
1277 	if (!rhp->rdev.status_page->db_off) {
1278 		t4_ring_sq_db(&qhp->wq, idx, wqe);
1279 		spin_unlock_irqrestore(&qhp->lock, flag);
1280 	} else {
1281 		spin_unlock_irqrestore(&qhp->lock, flag);
1282 		ring_kernel_sq_db(qhp, idx);
1283 	}
1284 	return err;
1285 }
1286 
1287 int c4iw_post_receive(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
1288 		      const struct ib_recv_wr **bad_wr)
1289 {
1290 	int err = 0;
1291 	struct c4iw_qp *qhp;
1292 	union t4_recv_wr *wqe = NULL;
1293 	u32 num_wrs;
1294 	u8 len16 = 0;
1295 	unsigned long flag;
1296 	u16 idx = 0;
1297 
1298 	qhp = to_c4iw_qp(ibqp);
1299 	spin_lock_irqsave(&qhp->lock, flag);
1300 
1301 	/*
1302 	 * If the qp has been flushed, then just insert a special
1303 	 * drain cqe.
1304 	 */
1305 	if (qhp->wq.flushed) {
1306 		spin_unlock_irqrestore(&qhp->lock, flag);
1307 		complete_rq_drain_wrs(qhp, wr);
1308 		return err;
1309 	}
1310 	num_wrs = t4_rq_avail(&qhp->wq);
1311 	if (num_wrs == 0) {
1312 		spin_unlock_irqrestore(&qhp->lock, flag);
1313 		*bad_wr = wr;
1314 		return -ENOMEM;
1315 	}
1316 	while (wr) {
1317 		if (wr->num_sge > T4_MAX_RECV_SGE) {
1318 			err = -EINVAL;
1319 			*bad_wr = wr;
1320 			break;
1321 		}
1322 		wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
1323 					   qhp->wq.rq.wq_pidx *
1324 					   T4_EQ_ENTRY_SIZE);
1325 		if (num_wrs)
1326 			err = build_rdma_recv(qhp, wqe, wr, &len16);
1327 		else
1328 			err = -ENOMEM;
1329 		if (err) {
1330 			*bad_wr = wr;
1331 			break;
1332 		}
1333 
1334 		qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
1335 		if (c4iw_wr_log) {
1336 			qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].sge_ts =
1337 				cxgb4_read_sge_timestamp(
1338 						qhp->rhp->rdev.lldi.ports[0]);
1339 			qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].host_time =
1340 				ktime_get();
1341 		}
1342 
1343 		wqe->recv.opcode = FW_RI_RECV_WR;
1344 		wqe->recv.r1 = 0;
1345 		wqe->recv.wrid = qhp->wq.rq.pidx;
1346 		wqe->recv.r2[0] = 0;
1347 		wqe->recv.r2[1] = 0;
1348 		wqe->recv.r2[2] = 0;
1349 		wqe->recv.len16 = len16;
1350 		pr_debug("cookie 0x%llx pidx %u\n",
1351 			 (unsigned long long)wr->wr_id, qhp->wq.rq.pidx);
1352 		t4_rq_produce(&qhp->wq, len16);
1353 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
1354 		wr = wr->next;
1355 		num_wrs--;
1356 	}
1357 	if (!qhp->rhp->rdev.status_page->db_off) {
1358 		t4_ring_rq_db(&qhp->wq, idx, wqe);
1359 		spin_unlock_irqrestore(&qhp->lock, flag);
1360 	} else {
1361 		spin_unlock_irqrestore(&qhp->lock, flag);
1362 		ring_kernel_rq_db(qhp, idx);
1363 	}
1364 	return err;
1365 }
1366 
1367 static void defer_srq_wr(struct t4_srq *srq, union t4_recv_wr *wqe,
1368 			 u64 wr_id, u8 len16)
1369 {
1370 	struct t4_srq_pending_wr *pwr = &srq->pending_wrs[srq->pending_pidx];
1371 
1372 	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",
1373 		 __func__, srq->cidx, srq->pidx, srq->wq_pidx,
1374 		 srq->in_use, srq->ooo_count,
1375 		 (unsigned long long)wr_id, srq->pending_cidx,
1376 		 srq->pending_pidx, srq->pending_in_use);
1377 	pwr->wr_id = wr_id;
1378 	pwr->len16 = len16;
1379 	memcpy(&pwr->wqe, wqe, len16 * 16);
1380 	t4_srq_produce_pending_wr(srq);
1381 }
1382 
1383 int c4iw_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
1384 		       const struct ib_recv_wr **bad_wr)
1385 {
1386 	union t4_recv_wr *wqe, lwqe;
1387 	struct c4iw_srq *srq;
1388 	unsigned long flag;
1389 	u8 len16 = 0;
1390 	u16 idx = 0;
1391 	int err = 0;
1392 	u32 num_wrs;
1393 
1394 	srq = to_c4iw_srq(ibsrq);
1395 	spin_lock_irqsave(&srq->lock, flag);
1396 	num_wrs = t4_srq_avail(&srq->wq);
1397 	if (num_wrs == 0) {
1398 		spin_unlock_irqrestore(&srq->lock, flag);
1399 		return -ENOMEM;
1400 	}
1401 	while (wr) {
1402 		if (wr->num_sge > T4_MAX_RECV_SGE) {
1403 			err = -EINVAL;
1404 			*bad_wr = wr;
1405 			break;
1406 		}
1407 		wqe = &lwqe;
1408 		if (num_wrs)
1409 			err = build_srq_recv(wqe, wr, &len16);
1410 		else
1411 			err = -ENOMEM;
1412 		if (err) {
1413 			*bad_wr = wr;
1414 			break;
1415 		}
1416 
1417 		wqe->recv.opcode = FW_RI_RECV_WR;
1418 		wqe->recv.r1 = 0;
1419 		wqe->recv.wrid = srq->wq.pidx;
1420 		wqe->recv.r2[0] = 0;
1421 		wqe->recv.r2[1] = 0;
1422 		wqe->recv.r2[2] = 0;
1423 		wqe->recv.len16 = len16;
1424 
1425 		if (srq->wq.ooo_count ||
1426 		    srq->wq.pending_in_use ||
1427 		    srq->wq.sw_rq[srq->wq.pidx].valid) {
1428 			defer_srq_wr(&srq->wq, wqe, wr->wr_id, len16);
1429 		} else {
1430 			srq->wq.sw_rq[srq->wq.pidx].wr_id = wr->wr_id;
1431 			srq->wq.sw_rq[srq->wq.pidx].valid = 1;
1432 			c4iw_copy_wr_to_srq(&srq->wq, wqe, len16);
1433 			pr_debug("%s cidx %u pidx %u wq_pidx %u in_use %u wr_id 0x%llx\n",
1434 				 __func__, srq->wq.cidx,
1435 				 srq->wq.pidx, srq->wq.wq_pidx,
1436 				 srq->wq.in_use,
1437 				 (unsigned long long)wr->wr_id);
1438 			t4_srq_produce(&srq->wq, len16);
1439 			idx += DIV_ROUND_UP(len16 * 16, T4_EQ_ENTRY_SIZE);
1440 		}
1441 		wr = wr->next;
1442 		num_wrs--;
1443 	}
1444 	if (idx)
1445 		t4_ring_srq_db(&srq->wq, idx, len16, wqe);
1446 	spin_unlock_irqrestore(&srq->lock, flag);
1447 	return err;
1448 }
1449 
1450 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
1451 				    u8 *ecode)
1452 {
1453 	int status;
1454 	int tagged;
1455 	int opcode;
1456 	int rqtype;
1457 	int send_inv;
1458 
1459 	if (!err_cqe) {
1460 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
1461 		*ecode = 0;
1462 		return;
1463 	}
1464 
1465 	status = CQE_STATUS(err_cqe);
1466 	opcode = CQE_OPCODE(err_cqe);
1467 	rqtype = RQ_TYPE(err_cqe);
1468 	send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
1469 		   (opcode == FW_RI_SEND_WITH_SE_INV);
1470 	tagged = (opcode == FW_RI_RDMA_WRITE) ||
1471 		 (rqtype && (opcode == FW_RI_READ_RESP));
1472 
1473 	switch (status) {
1474 	case T4_ERR_STAG:
1475 		if (send_inv) {
1476 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1477 			*ecode = RDMAP_CANT_INV_STAG;
1478 		} else {
1479 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1480 			*ecode = RDMAP_INV_STAG;
1481 		}
1482 		break;
1483 	case T4_ERR_PDID:
1484 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1485 		if ((opcode == FW_RI_SEND_WITH_INV) ||
1486 		    (opcode == FW_RI_SEND_WITH_SE_INV))
1487 			*ecode = RDMAP_CANT_INV_STAG;
1488 		else
1489 			*ecode = RDMAP_STAG_NOT_ASSOC;
1490 		break;
1491 	case T4_ERR_QPID:
1492 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1493 		*ecode = RDMAP_STAG_NOT_ASSOC;
1494 		break;
1495 	case T4_ERR_ACCESS:
1496 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1497 		*ecode = RDMAP_ACC_VIOL;
1498 		break;
1499 	case T4_ERR_WRAP:
1500 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1501 		*ecode = RDMAP_TO_WRAP;
1502 		break;
1503 	case T4_ERR_BOUND:
1504 		if (tagged) {
1505 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1506 			*ecode = DDPT_BASE_BOUNDS;
1507 		} else {
1508 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1509 			*ecode = RDMAP_BASE_BOUNDS;
1510 		}
1511 		break;
1512 	case T4_ERR_INVALIDATE_SHARED_MR:
1513 	case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
1514 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1515 		*ecode = RDMAP_CANT_INV_STAG;
1516 		break;
1517 	case T4_ERR_ECC:
1518 	case T4_ERR_ECC_PSTAG:
1519 	case T4_ERR_INTERNAL_ERR:
1520 		*layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
1521 		*ecode = 0;
1522 		break;
1523 	case T4_ERR_OUT_OF_RQE:
1524 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1525 		*ecode = DDPU_INV_MSN_NOBUF;
1526 		break;
1527 	case T4_ERR_PBL_ADDR_BOUND:
1528 		*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1529 		*ecode = DDPT_BASE_BOUNDS;
1530 		break;
1531 	case T4_ERR_CRC:
1532 		*layer_type = LAYER_MPA|DDP_LLP;
1533 		*ecode = MPA_CRC_ERR;
1534 		break;
1535 	case T4_ERR_MARKER:
1536 		*layer_type = LAYER_MPA|DDP_LLP;
1537 		*ecode = MPA_MARKER_ERR;
1538 		break;
1539 	case T4_ERR_PDU_LEN_ERR:
1540 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1541 		*ecode = DDPU_MSG_TOOBIG;
1542 		break;
1543 	case T4_ERR_DDP_VERSION:
1544 		if (tagged) {
1545 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1546 			*ecode = DDPT_INV_VERS;
1547 		} else {
1548 			*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1549 			*ecode = DDPU_INV_VERS;
1550 		}
1551 		break;
1552 	case T4_ERR_RDMA_VERSION:
1553 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1554 		*ecode = RDMAP_INV_VERS;
1555 		break;
1556 	case T4_ERR_OPCODE:
1557 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1558 		*ecode = RDMAP_INV_OPCODE;
1559 		break;
1560 	case T4_ERR_DDP_QUEUE_NUM:
1561 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1562 		*ecode = DDPU_INV_QN;
1563 		break;
1564 	case T4_ERR_MSN:
1565 	case T4_ERR_MSN_GAP:
1566 	case T4_ERR_MSN_RANGE:
1567 	case T4_ERR_IRD_OVERFLOW:
1568 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1569 		*ecode = DDPU_INV_MSN_RANGE;
1570 		break;
1571 	case T4_ERR_TBIT:
1572 		*layer_type = LAYER_DDP|DDP_LOCAL_CATA;
1573 		*ecode = 0;
1574 		break;
1575 	case T4_ERR_MO:
1576 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1577 		*ecode = DDPU_INV_MO;
1578 		break;
1579 	default:
1580 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
1581 		*ecode = 0;
1582 		break;
1583 	}
1584 }
1585 
1586 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
1587 			   gfp_t gfp)
1588 {
1589 	struct fw_ri_wr *wqe;
1590 	struct sk_buff *skb;
1591 	struct terminate_message *term;
1592 
1593 	pr_debug("qhp %p qid 0x%x tid %u\n", qhp, qhp->wq.sq.qid,
1594 		 qhp->ep->hwtid);
1595 
1596 	skb = skb_dequeue(&qhp->ep->com.ep_skb_list);
1597 	if (WARN_ON(!skb))
1598 		return;
1599 
1600 	set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1601 
1602 	wqe = __skb_put_zero(skb, sizeof(*wqe));
1603 	wqe->op_compl = cpu_to_be32(FW_WR_OP_V(FW_RI_INIT_WR));
1604 	wqe->flowid_len16 = cpu_to_be32(
1605 		FW_WR_FLOWID_V(qhp->ep->hwtid) |
1606 		FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1607 
1608 	wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
1609 	wqe->u.terminate.immdlen = cpu_to_be32(sizeof *term);
1610 	term = (struct terminate_message *)wqe->u.terminate.termmsg;
1611 	if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
1612 		term->layer_etype = qhp->attr.layer_etype;
1613 		term->ecode = qhp->attr.ecode;
1614 	} else
1615 		build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
1616 	c4iw_ofld_send(&qhp->rhp->rdev, skb);
1617 }
1618 
1619 /*
1620  * Assumes qhp lock is held.
1621  */
1622 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
1623 		       struct c4iw_cq *schp)
1624 {
1625 	int count;
1626 	int rq_flushed = 0, sq_flushed;
1627 	unsigned long flag;
1628 
1629 	pr_debug("qhp %p rchp %p schp %p\n", qhp, rchp, schp);
1630 
1631 	/* locking hierarchy: cqs lock first, then qp lock. */
1632 	spin_lock_irqsave(&rchp->lock, flag);
1633 	if (schp != rchp)
1634 		spin_lock(&schp->lock);
1635 	spin_lock(&qhp->lock);
1636 
1637 	if (qhp->wq.flushed) {
1638 		spin_unlock(&qhp->lock);
1639 		if (schp != rchp)
1640 			spin_unlock(&schp->lock);
1641 		spin_unlock_irqrestore(&rchp->lock, flag);
1642 		return;
1643 	}
1644 	qhp->wq.flushed = 1;
1645 	t4_set_wq_in_error(&qhp->wq, 0);
1646 
1647 	c4iw_flush_hw_cq(rchp, qhp);
1648 	if (!qhp->srq) {
1649 		c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
1650 		rq_flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
1651 	}
1652 
1653 	if (schp != rchp)
1654 		c4iw_flush_hw_cq(schp, qhp);
1655 	sq_flushed = c4iw_flush_sq(qhp);
1656 
1657 	spin_unlock(&qhp->lock);
1658 	if (schp != rchp)
1659 		spin_unlock(&schp->lock);
1660 	spin_unlock_irqrestore(&rchp->lock, flag);
1661 
1662 	if (schp == rchp) {
1663 		if ((rq_flushed || sq_flushed) &&
1664 		    t4_clear_cq_armed(&rchp->cq)) {
1665 			spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1666 			(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1667 						   rchp->ibcq.cq_context);
1668 			spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1669 		}
1670 	} else {
1671 		if (rq_flushed && t4_clear_cq_armed(&rchp->cq)) {
1672 			spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1673 			(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1674 						   rchp->ibcq.cq_context);
1675 			spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1676 		}
1677 		if (sq_flushed && t4_clear_cq_armed(&schp->cq)) {
1678 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1679 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1680 						   schp->ibcq.cq_context);
1681 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1682 		}
1683 	}
1684 }
1685 
1686 static void flush_qp(struct c4iw_qp *qhp)
1687 {
1688 	struct c4iw_cq *rchp, *schp;
1689 	unsigned long flag;
1690 
1691 	rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
1692 	schp = to_c4iw_cq(qhp->ibqp.send_cq);
1693 
1694 	if (qhp->ibqp.uobject) {
1695 
1696 		/* for user qps, qhp->wq.flushed is protected by qhp->mutex */
1697 		if (qhp->wq.flushed)
1698 			return;
1699 
1700 		qhp->wq.flushed = 1;
1701 		t4_set_wq_in_error(&qhp->wq, 0);
1702 		t4_set_cq_in_error(&rchp->cq);
1703 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1704 		(*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1705 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1706 		if (schp != rchp) {
1707 			t4_set_cq_in_error(&schp->cq);
1708 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1709 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1710 					schp->ibcq.cq_context);
1711 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1712 		}
1713 		return;
1714 	}
1715 	__flush_qp(qhp, rchp, schp);
1716 }
1717 
1718 static int rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1719 		     struct c4iw_ep *ep)
1720 {
1721 	struct fw_ri_wr *wqe;
1722 	int ret;
1723 	struct sk_buff *skb;
1724 
1725 	pr_debug("qhp %p qid 0x%x tid %u\n", qhp, qhp->wq.sq.qid, ep->hwtid);
1726 
1727 	skb = skb_dequeue(&ep->com.ep_skb_list);
1728 	if (WARN_ON(!skb))
1729 		return -ENOMEM;
1730 
1731 	set_wr_txq(skb, CPL_PRIORITY_DATA, ep->txq_idx);
1732 
1733 	wqe = __skb_put_zero(skb, sizeof(*wqe));
1734 	wqe->op_compl = cpu_to_be32(
1735 		FW_WR_OP_V(FW_RI_INIT_WR) |
1736 		FW_WR_COMPL_F);
1737 	wqe->flowid_len16 = cpu_to_be32(
1738 		FW_WR_FLOWID_V(ep->hwtid) |
1739 		FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1740 	wqe->cookie = (uintptr_t)ep->com.wr_waitp;
1741 
1742 	wqe->u.fini.type = FW_RI_TYPE_FINI;
1743 
1744 	ret = c4iw_ref_send_wait(&rhp->rdev, skb, ep->com.wr_waitp,
1745 				 qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
1746 
1747 	pr_debug("ret %d\n", ret);
1748 	return ret;
1749 }
1750 
1751 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1752 {
1753 	pr_debug("p2p_type = %d\n", p2p_type);
1754 	memset(&init->u, 0, sizeof init->u);
1755 	switch (p2p_type) {
1756 	case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1757 		init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1758 		init->u.write.stag_sink = cpu_to_be32(1);
1759 		init->u.write.to_sink = cpu_to_be64(1);
1760 		init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1761 		init->u.write.len16 = DIV_ROUND_UP(sizeof init->u.write +
1762 						   sizeof(struct fw_ri_immd),
1763 						   16);
1764 		break;
1765 	case FW_RI_INIT_P2PTYPE_READ_REQ:
1766 		init->u.write.opcode = FW_RI_RDMA_READ_WR;
1767 		init->u.read.stag_src = cpu_to_be32(1);
1768 		init->u.read.to_src_lo = cpu_to_be32(1);
1769 		init->u.read.stag_sink = cpu_to_be32(1);
1770 		init->u.read.to_sink_lo = cpu_to_be32(1);
1771 		init->u.read.len16 = DIV_ROUND_UP(sizeof init->u.read, 16);
1772 		break;
1773 	}
1774 }
1775 
1776 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1777 {
1778 	struct fw_ri_wr *wqe;
1779 	int ret;
1780 	struct sk_buff *skb;
1781 
1782 	pr_debug("qhp %p qid 0x%x tid %u ird %u ord %u\n", qhp,
1783 		 qhp->wq.sq.qid, qhp->ep->hwtid, qhp->ep->ird, qhp->ep->ord);
1784 
1785 	skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1786 	if (!skb) {
1787 		ret = -ENOMEM;
1788 		goto out;
1789 	}
1790 	ret = alloc_ird(rhp, qhp->attr.max_ird);
1791 	if (ret) {
1792 		qhp->attr.max_ird = 0;
1793 		kfree_skb(skb);
1794 		goto out;
1795 	}
1796 	set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1797 
1798 	wqe = __skb_put_zero(skb, sizeof(*wqe));
1799 	wqe->op_compl = cpu_to_be32(
1800 		FW_WR_OP_V(FW_RI_INIT_WR) |
1801 		FW_WR_COMPL_F);
1802 	wqe->flowid_len16 = cpu_to_be32(
1803 		FW_WR_FLOWID_V(qhp->ep->hwtid) |
1804 		FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1805 
1806 	wqe->cookie = (uintptr_t)qhp->ep->com.wr_waitp;
1807 
1808 	wqe->u.init.type = FW_RI_TYPE_INIT;
1809 	wqe->u.init.mpareqbit_p2ptype =
1810 		FW_RI_WR_MPAREQBIT_V(qhp->attr.mpa_attr.initiator) |
1811 		FW_RI_WR_P2PTYPE_V(qhp->attr.mpa_attr.p2p_type);
1812 	wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1813 	if (qhp->attr.mpa_attr.recv_marker_enabled)
1814 		wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1815 	if (qhp->attr.mpa_attr.xmit_marker_enabled)
1816 		wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1817 	if (qhp->attr.mpa_attr.crc_enabled)
1818 		wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1819 
1820 	wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1821 			    FW_RI_QP_RDMA_WRITE_ENABLE |
1822 			    FW_RI_QP_BIND_ENABLE;
1823 	if (!qhp->ibqp.uobject)
1824 		wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1825 				     FW_RI_QP_STAG0_ENABLE;
1826 	wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1827 	wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1828 	wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1829 	wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1830 	if (qhp->srq) {
1831 		wqe->u.init.rq_eqid = cpu_to_be32(FW_RI_INIT_RQEQID_SRQ |
1832 						  qhp->srq->idx);
1833 	} else {
1834 		wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1835 		wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1836 		wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1837 						   rhp->rdev.lldi.vr->rq.start);
1838 	}
1839 	wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1840 	wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1841 	wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1842 	wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1843 	wqe->u.init.iss = cpu_to_be32(qhp->ep->snd_seq);
1844 	wqe->u.init.irs = cpu_to_be32(qhp->ep->rcv_seq);
1845 	if (qhp->attr.mpa_attr.initiator)
1846 		build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1847 
1848 	ret = c4iw_ref_send_wait(&rhp->rdev, skb, qhp->ep->com.wr_waitp,
1849 				 qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
1850 	if (!ret)
1851 		goto out;
1852 
1853 	free_ird(rhp, qhp->attr.max_ird);
1854 out:
1855 	pr_debug("ret %d\n", ret);
1856 	return ret;
1857 }
1858 
1859 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1860 		   enum c4iw_qp_attr_mask mask,
1861 		   struct c4iw_qp_attributes *attrs,
1862 		   int internal)
1863 {
1864 	int ret = 0;
1865 	struct c4iw_qp_attributes newattr = qhp->attr;
1866 	int disconnect = 0;
1867 	int terminate = 0;
1868 	int abort = 0;
1869 	int free = 0;
1870 	struct c4iw_ep *ep = NULL;
1871 
1872 	pr_debug("qhp %p sqid 0x%x rqid 0x%x ep %p state %d -> %d\n",
1873 		 qhp, qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep, qhp->attr.state,
1874 		 (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1875 
1876 	mutex_lock(&qhp->mutex);
1877 
1878 	/* Process attr changes if in IDLE */
1879 	if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1880 		if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1881 			ret = -EIO;
1882 			goto out;
1883 		}
1884 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1885 			newattr.enable_rdma_read = attrs->enable_rdma_read;
1886 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1887 			newattr.enable_rdma_write = attrs->enable_rdma_write;
1888 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1889 			newattr.enable_bind = attrs->enable_bind;
1890 		if (mask & C4IW_QP_ATTR_MAX_ORD) {
1891 			if (attrs->max_ord > c4iw_max_read_depth) {
1892 				ret = -EINVAL;
1893 				goto out;
1894 			}
1895 			newattr.max_ord = attrs->max_ord;
1896 		}
1897 		if (mask & C4IW_QP_ATTR_MAX_IRD) {
1898 			if (attrs->max_ird > cur_max_read_depth(rhp)) {
1899 				ret = -EINVAL;
1900 				goto out;
1901 			}
1902 			newattr.max_ird = attrs->max_ird;
1903 		}
1904 		qhp->attr = newattr;
1905 	}
1906 
1907 	if (mask & C4IW_QP_ATTR_SQ_DB) {
1908 		ret = ring_kernel_sq_db(qhp, attrs->sq_db_inc);
1909 		goto out;
1910 	}
1911 	if (mask & C4IW_QP_ATTR_RQ_DB) {
1912 		ret = ring_kernel_rq_db(qhp, attrs->rq_db_inc);
1913 		goto out;
1914 	}
1915 
1916 	if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1917 		goto out;
1918 	if (qhp->attr.state == attrs->next_state)
1919 		goto out;
1920 
1921 	switch (qhp->attr.state) {
1922 	case C4IW_QP_STATE_IDLE:
1923 		switch (attrs->next_state) {
1924 		case C4IW_QP_STATE_RTS:
1925 			if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1926 				ret = -EINVAL;
1927 				goto out;
1928 			}
1929 			if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1930 				ret = -EINVAL;
1931 				goto out;
1932 			}
1933 			qhp->attr.mpa_attr = attrs->mpa_attr;
1934 			qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1935 			qhp->ep = qhp->attr.llp_stream_handle;
1936 			set_state(qhp, C4IW_QP_STATE_RTS);
1937 
1938 			/*
1939 			 * Ref the endpoint here and deref when we
1940 			 * disassociate the endpoint from the QP.  This
1941 			 * happens in CLOSING->IDLE transition or *->ERROR
1942 			 * transition.
1943 			 */
1944 			c4iw_get_ep(&qhp->ep->com);
1945 			ret = rdma_init(rhp, qhp);
1946 			if (ret)
1947 				goto err;
1948 			break;
1949 		case C4IW_QP_STATE_ERROR:
1950 			set_state(qhp, C4IW_QP_STATE_ERROR);
1951 			flush_qp(qhp);
1952 			break;
1953 		default:
1954 			ret = -EINVAL;
1955 			goto out;
1956 		}
1957 		break;
1958 	case C4IW_QP_STATE_RTS:
1959 		switch (attrs->next_state) {
1960 		case C4IW_QP_STATE_CLOSING:
1961 			t4_set_wq_in_error(&qhp->wq, 0);
1962 			set_state(qhp, C4IW_QP_STATE_CLOSING);
1963 			ep = qhp->ep;
1964 			if (!internal) {
1965 				abort = 0;
1966 				disconnect = 1;
1967 				c4iw_get_ep(&qhp->ep->com);
1968 			}
1969 			ret = rdma_fini(rhp, qhp, ep);
1970 			if (ret)
1971 				goto err;
1972 			break;
1973 		case C4IW_QP_STATE_TERMINATE:
1974 			t4_set_wq_in_error(&qhp->wq, 0);
1975 			set_state(qhp, C4IW_QP_STATE_TERMINATE);
1976 			qhp->attr.layer_etype = attrs->layer_etype;
1977 			qhp->attr.ecode = attrs->ecode;
1978 			ep = qhp->ep;
1979 			c4iw_get_ep(&ep->com);
1980 			disconnect = 1;
1981 			if (!internal) {
1982 				terminate = 1;
1983 			} else {
1984 				terminate = qhp->attr.send_term;
1985 				ret = rdma_fini(rhp, qhp, ep);
1986 				if (ret)
1987 					goto err;
1988 			}
1989 			break;
1990 		case C4IW_QP_STATE_ERROR:
1991 			t4_set_wq_in_error(&qhp->wq, 0);
1992 			set_state(qhp, C4IW_QP_STATE_ERROR);
1993 			if (!internal) {
1994 				abort = 1;
1995 				disconnect = 1;
1996 				ep = qhp->ep;
1997 				c4iw_get_ep(&qhp->ep->com);
1998 			}
1999 			goto err;
2000 			break;
2001 		default:
2002 			ret = -EINVAL;
2003 			goto out;
2004 		}
2005 		break;
2006 	case C4IW_QP_STATE_CLOSING:
2007 
2008 		/*
2009 		 * Allow kernel users to move to ERROR for qp draining.
2010 		 */
2011 		if (!internal && (qhp->ibqp.uobject || attrs->next_state !=
2012 				  C4IW_QP_STATE_ERROR)) {
2013 			ret = -EINVAL;
2014 			goto out;
2015 		}
2016 		switch (attrs->next_state) {
2017 		case C4IW_QP_STATE_IDLE:
2018 			flush_qp(qhp);
2019 			set_state(qhp, C4IW_QP_STATE_IDLE);
2020 			qhp->attr.llp_stream_handle = NULL;
2021 			c4iw_put_ep(&qhp->ep->com);
2022 			qhp->ep = NULL;
2023 			wake_up(&qhp->wait);
2024 			break;
2025 		case C4IW_QP_STATE_ERROR:
2026 			goto err;
2027 		default:
2028 			ret = -EINVAL;
2029 			goto err;
2030 		}
2031 		break;
2032 	case C4IW_QP_STATE_ERROR:
2033 		if (attrs->next_state != C4IW_QP_STATE_IDLE) {
2034 			ret = -EINVAL;
2035 			goto out;
2036 		}
2037 		if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) {
2038 			ret = -EINVAL;
2039 			goto out;
2040 		}
2041 		set_state(qhp, C4IW_QP_STATE_IDLE);
2042 		break;
2043 	case C4IW_QP_STATE_TERMINATE:
2044 		if (!internal) {
2045 			ret = -EINVAL;
2046 			goto out;
2047 		}
2048 		goto err;
2049 		break;
2050 	default:
2051 		pr_err("%s in a bad state %d\n", __func__, qhp->attr.state);
2052 		ret = -EINVAL;
2053 		goto err;
2054 		break;
2055 	}
2056 	goto out;
2057 err:
2058 	pr_debug("disassociating ep %p qpid 0x%x\n", qhp->ep,
2059 		 qhp->wq.sq.qid);
2060 
2061 	/* disassociate the LLP connection */
2062 	qhp->attr.llp_stream_handle = NULL;
2063 	if (!ep)
2064 		ep = qhp->ep;
2065 	qhp->ep = NULL;
2066 	set_state(qhp, C4IW_QP_STATE_ERROR);
2067 	free = 1;
2068 	abort = 1;
2069 	flush_qp(qhp);
2070 	wake_up(&qhp->wait);
2071 out:
2072 	mutex_unlock(&qhp->mutex);
2073 
2074 	if (terminate)
2075 		post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
2076 
2077 	/*
2078 	 * If disconnect is 1, then we need to initiate a disconnect
2079 	 * on the EP.  This can be a normal close (RTS->CLOSING) or
2080 	 * an abnormal close (RTS/CLOSING->ERROR).
2081 	 */
2082 	if (disconnect) {
2083 		c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC :
2084 							 GFP_KERNEL);
2085 		c4iw_put_ep(&ep->com);
2086 	}
2087 
2088 	/*
2089 	 * If free is 1, then we've disassociated the EP from the QP
2090 	 * and we need to dereference the EP.
2091 	 */
2092 	if (free)
2093 		c4iw_put_ep(&ep->com);
2094 	pr_debug("exit state %d\n", qhp->attr.state);
2095 	return ret;
2096 }
2097 
2098 int c4iw_destroy_qp(struct ib_qp *ib_qp, struct ib_udata *udata)
2099 {
2100 	struct c4iw_dev *rhp;
2101 	struct c4iw_qp *qhp;
2102 	struct c4iw_qp_attributes attrs;
2103 
2104 	qhp = to_c4iw_qp(ib_qp);
2105 	rhp = qhp->rhp;
2106 
2107 	attrs.next_state = C4IW_QP_STATE_ERROR;
2108 	if (qhp->attr.state == C4IW_QP_STATE_TERMINATE)
2109 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
2110 	else
2111 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
2112 	wait_event(qhp->wait, !qhp->ep);
2113 
2114 	xa_lock_irq(&rhp->qps);
2115 	__xa_erase(&rhp->qps, qhp->wq.sq.qid);
2116 	if (!list_empty(&qhp->db_fc_entry))
2117 		list_del_init(&qhp->db_fc_entry);
2118 	xa_unlock_irq(&rhp->qps);
2119 	free_ird(rhp, qhp->attr.max_ird);
2120 
2121 	c4iw_qp_rem_ref(ib_qp);
2122 
2123 	pr_debug("ib_qp %p qpid 0x%0x\n", ib_qp, qhp->wq.sq.qid);
2124 	return 0;
2125 }
2126 
2127 struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
2128 			     struct ib_udata *udata)
2129 {
2130 	struct c4iw_dev *rhp;
2131 	struct c4iw_qp *qhp;
2132 	struct c4iw_pd *php;
2133 	struct c4iw_cq *schp;
2134 	struct c4iw_cq *rchp;
2135 	struct c4iw_create_qp_resp uresp;
2136 	unsigned int sqsize, rqsize = 0;
2137 	struct c4iw_ucontext *ucontext = rdma_udata_to_drv_context(
2138 		udata, struct c4iw_ucontext, ibucontext);
2139 	int ret;
2140 	struct c4iw_mm_entry *sq_key_mm, *rq_key_mm = NULL, *sq_db_key_mm;
2141 	struct c4iw_mm_entry *rq_db_key_mm = NULL, *ma_sync_key_mm = NULL;
2142 
2143 	pr_debug("ib_pd %p\n", pd);
2144 
2145 	if (attrs->qp_type != IB_QPT_RC)
2146 		return ERR_PTR(-EINVAL);
2147 
2148 	php = to_c4iw_pd(pd);
2149 	rhp = php->rhp;
2150 	schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
2151 	rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
2152 	if (!schp || !rchp)
2153 		return ERR_PTR(-EINVAL);
2154 
2155 	if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
2156 		return ERR_PTR(-EINVAL);
2157 
2158 	if (!attrs->srq) {
2159 		if (attrs->cap.max_recv_wr > rhp->rdev.hw_queue.t4_max_rq_size)
2160 			return ERR_PTR(-E2BIG);
2161 		rqsize = attrs->cap.max_recv_wr + 1;
2162 		if (rqsize < 8)
2163 			rqsize = 8;
2164 	}
2165 
2166 	if (attrs->cap.max_send_wr > rhp->rdev.hw_queue.t4_max_sq_size)
2167 		return ERR_PTR(-E2BIG);
2168 	sqsize = attrs->cap.max_send_wr + 1;
2169 	if (sqsize < 8)
2170 		sqsize = 8;
2171 
2172 	qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
2173 	if (!qhp)
2174 		return ERR_PTR(-ENOMEM);
2175 
2176 	qhp->wr_waitp = c4iw_alloc_wr_wait(GFP_KERNEL);
2177 	if (!qhp->wr_waitp) {
2178 		ret = -ENOMEM;
2179 		goto err_free_qhp;
2180 	}
2181 
2182 	qhp->wq.sq.size = sqsize;
2183 	qhp->wq.sq.memsize =
2184 		(sqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
2185 		sizeof(*qhp->wq.sq.queue) + 16 * sizeof(__be64);
2186 	qhp->wq.sq.flush_cidx = -1;
2187 	if (!attrs->srq) {
2188 		qhp->wq.rq.size = rqsize;
2189 		qhp->wq.rq.memsize =
2190 			(rqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
2191 			sizeof(*qhp->wq.rq.queue);
2192 	}
2193 
2194 	if (ucontext) {
2195 		qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
2196 		if (!attrs->srq)
2197 			qhp->wq.rq.memsize =
2198 				roundup(qhp->wq.rq.memsize, PAGE_SIZE);
2199 	}
2200 
2201 	ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
2202 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx,
2203 			qhp->wr_waitp, !attrs->srq);
2204 	if (ret)
2205 		goto err_free_wr_wait;
2206 
2207 	attrs->cap.max_recv_wr = rqsize - 1;
2208 	attrs->cap.max_send_wr = sqsize - 1;
2209 	attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
2210 
2211 	qhp->rhp = rhp;
2212 	qhp->attr.pd = php->pdid;
2213 	qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
2214 	qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
2215 	qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
2216 	qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
2217 	qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
2218 	if (!attrs->srq) {
2219 		qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
2220 		qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
2221 	}
2222 	qhp->attr.state = C4IW_QP_STATE_IDLE;
2223 	qhp->attr.next_state = C4IW_QP_STATE_IDLE;
2224 	qhp->attr.enable_rdma_read = 1;
2225 	qhp->attr.enable_rdma_write = 1;
2226 	qhp->attr.enable_bind = 1;
2227 	qhp->attr.max_ord = 0;
2228 	qhp->attr.max_ird = 0;
2229 	qhp->sq_sig_all = attrs->sq_sig_type == IB_SIGNAL_ALL_WR;
2230 	spin_lock_init(&qhp->lock);
2231 	mutex_init(&qhp->mutex);
2232 	init_waitqueue_head(&qhp->wait);
2233 	kref_init(&qhp->kref);
2234 	INIT_WORK(&qhp->free_work, free_qp_work);
2235 
2236 	ret = xa_insert_irq(&rhp->qps, qhp->wq.sq.qid, qhp, GFP_KERNEL);
2237 	if (ret)
2238 		goto err_destroy_qp;
2239 
2240 	if (udata && ucontext) {
2241 		sq_key_mm = kmalloc(sizeof(*sq_key_mm), GFP_KERNEL);
2242 		if (!sq_key_mm) {
2243 			ret = -ENOMEM;
2244 			goto err_remove_handle;
2245 		}
2246 		if (!attrs->srq) {
2247 			rq_key_mm = kmalloc(sizeof(*rq_key_mm), GFP_KERNEL);
2248 			if (!rq_key_mm) {
2249 				ret = -ENOMEM;
2250 				goto err_free_sq_key;
2251 			}
2252 		}
2253 		sq_db_key_mm = kmalloc(sizeof(*sq_db_key_mm), GFP_KERNEL);
2254 		if (!sq_db_key_mm) {
2255 			ret = -ENOMEM;
2256 			goto err_free_rq_key;
2257 		}
2258 		if (!attrs->srq) {
2259 			rq_db_key_mm =
2260 				kmalloc(sizeof(*rq_db_key_mm), GFP_KERNEL);
2261 			if (!rq_db_key_mm) {
2262 				ret = -ENOMEM;
2263 				goto err_free_sq_db_key;
2264 			}
2265 		}
2266 		memset(&uresp, 0, sizeof(uresp));
2267 		if (t4_sq_onchip(&qhp->wq.sq)) {
2268 			ma_sync_key_mm = kmalloc(sizeof(*ma_sync_key_mm),
2269 						 GFP_KERNEL);
2270 			if (!ma_sync_key_mm) {
2271 				ret = -ENOMEM;
2272 				goto err_free_rq_db_key;
2273 			}
2274 			uresp.flags = C4IW_QPF_ONCHIP;
2275 		}
2276 		if (rhp->rdev.lldi.write_w_imm_support)
2277 			uresp.flags |= C4IW_QPF_WRITE_W_IMM;
2278 		uresp.qid_mask = rhp->rdev.qpmask;
2279 		uresp.sqid = qhp->wq.sq.qid;
2280 		uresp.sq_size = qhp->wq.sq.size;
2281 		uresp.sq_memsize = qhp->wq.sq.memsize;
2282 		if (!attrs->srq) {
2283 			uresp.rqid = qhp->wq.rq.qid;
2284 			uresp.rq_size = qhp->wq.rq.size;
2285 			uresp.rq_memsize = qhp->wq.rq.memsize;
2286 		}
2287 		spin_lock(&ucontext->mmap_lock);
2288 		if (ma_sync_key_mm) {
2289 			uresp.ma_sync_key = ucontext->key;
2290 			ucontext->key += PAGE_SIZE;
2291 		}
2292 		uresp.sq_key = ucontext->key;
2293 		ucontext->key += PAGE_SIZE;
2294 		if (!attrs->srq) {
2295 			uresp.rq_key = ucontext->key;
2296 			ucontext->key += PAGE_SIZE;
2297 		}
2298 		uresp.sq_db_gts_key = ucontext->key;
2299 		ucontext->key += PAGE_SIZE;
2300 		if (!attrs->srq) {
2301 			uresp.rq_db_gts_key = ucontext->key;
2302 			ucontext->key += PAGE_SIZE;
2303 		}
2304 		spin_unlock(&ucontext->mmap_lock);
2305 		ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
2306 		if (ret)
2307 			goto err_free_ma_sync_key;
2308 		sq_key_mm->key = uresp.sq_key;
2309 		sq_key_mm->addr = qhp->wq.sq.phys_addr;
2310 		sq_key_mm->len = PAGE_ALIGN(qhp->wq.sq.memsize);
2311 		insert_mmap(ucontext, sq_key_mm);
2312 		if (!attrs->srq) {
2313 			rq_key_mm->key = uresp.rq_key;
2314 			rq_key_mm->addr = virt_to_phys(qhp->wq.rq.queue);
2315 			rq_key_mm->len = PAGE_ALIGN(qhp->wq.rq.memsize);
2316 			insert_mmap(ucontext, rq_key_mm);
2317 		}
2318 		sq_db_key_mm->key = uresp.sq_db_gts_key;
2319 		sq_db_key_mm->addr = (u64)(unsigned long)qhp->wq.sq.bar2_pa;
2320 		sq_db_key_mm->len = PAGE_SIZE;
2321 		insert_mmap(ucontext, sq_db_key_mm);
2322 		if (!attrs->srq) {
2323 			rq_db_key_mm->key = uresp.rq_db_gts_key;
2324 			rq_db_key_mm->addr =
2325 				(u64)(unsigned long)qhp->wq.rq.bar2_pa;
2326 			rq_db_key_mm->len = PAGE_SIZE;
2327 			insert_mmap(ucontext, rq_db_key_mm);
2328 		}
2329 		if (ma_sync_key_mm) {
2330 			ma_sync_key_mm->key = uresp.ma_sync_key;
2331 			ma_sync_key_mm->addr =
2332 				(pci_resource_start(rhp->rdev.lldi.pdev, 0) +
2333 				PCIE_MA_SYNC_A) & PAGE_MASK;
2334 			ma_sync_key_mm->len = PAGE_SIZE;
2335 			insert_mmap(ucontext, ma_sync_key_mm);
2336 		}
2337 
2338 		qhp->ucontext = ucontext;
2339 	}
2340 	if (!attrs->srq) {
2341 		qhp->wq.qp_errp =
2342 			&qhp->wq.rq.queue[qhp->wq.rq.size].status.qp_err;
2343 	} else {
2344 		qhp->wq.qp_errp =
2345 			&qhp->wq.sq.queue[qhp->wq.sq.size].status.qp_err;
2346 		qhp->wq.srqidxp =
2347 			&qhp->wq.sq.queue[qhp->wq.sq.size].status.srqidx;
2348 	}
2349 
2350 	qhp->ibqp.qp_num = qhp->wq.sq.qid;
2351 	if (attrs->srq)
2352 		qhp->srq = to_c4iw_srq(attrs->srq);
2353 	INIT_LIST_HEAD(&qhp->db_fc_entry);
2354 	pr_debug("sq id %u size %u memsize %zu num_entries %u rq id %u size %u memsize %zu num_entries %u\n",
2355 		 qhp->wq.sq.qid, qhp->wq.sq.size, qhp->wq.sq.memsize,
2356 		 attrs->cap.max_send_wr, qhp->wq.rq.qid, qhp->wq.rq.size,
2357 		 qhp->wq.rq.memsize, attrs->cap.max_recv_wr);
2358 	return &qhp->ibqp;
2359 err_free_ma_sync_key:
2360 	kfree(ma_sync_key_mm);
2361 err_free_rq_db_key:
2362 	if (!attrs->srq)
2363 		kfree(rq_db_key_mm);
2364 err_free_sq_db_key:
2365 	kfree(sq_db_key_mm);
2366 err_free_rq_key:
2367 	if (!attrs->srq)
2368 		kfree(rq_key_mm);
2369 err_free_sq_key:
2370 	kfree(sq_key_mm);
2371 err_remove_handle:
2372 	xa_erase_irq(&rhp->qps, qhp->wq.sq.qid);
2373 err_destroy_qp:
2374 	destroy_qp(&rhp->rdev, &qhp->wq,
2375 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx, !attrs->srq);
2376 err_free_wr_wait:
2377 	c4iw_put_wr_wait(qhp->wr_waitp);
2378 err_free_qhp:
2379 	kfree(qhp);
2380 	return ERR_PTR(ret);
2381 }
2382 
2383 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2384 		      int attr_mask, struct ib_udata *udata)
2385 {
2386 	struct c4iw_dev *rhp;
2387 	struct c4iw_qp *qhp;
2388 	enum c4iw_qp_attr_mask mask = 0;
2389 	struct c4iw_qp_attributes attrs;
2390 
2391 	pr_debug("ib_qp %p\n", ibqp);
2392 
2393 	/* iwarp does not support the RTR state */
2394 	if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
2395 		attr_mask &= ~IB_QP_STATE;
2396 
2397 	/* Make sure we still have something left to do */
2398 	if (!attr_mask)
2399 		return 0;
2400 
2401 	memset(&attrs, 0, sizeof attrs);
2402 	qhp = to_c4iw_qp(ibqp);
2403 	rhp = qhp->rhp;
2404 
2405 	attrs.next_state = c4iw_convert_state(attr->qp_state);
2406 	attrs.enable_rdma_read = (attr->qp_access_flags &
2407 			       IB_ACCESS_REMOTE_READ) ?  1 : 0;
2408 	attrs.enable_rdma_write = (attr->qp_access_flags &
2409 				IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
2410 	attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
2411 
2412 
2413 	mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
2414 	mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
2415 			(C4IW_QP_ATTR_ENABLE_RDMA_READ |
2416 			 C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
2417 			 C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
2418 
2419 	/*
2420 	 * Use SQ_PSN and RQ_PSN to pass in IDX_INC values for
2421 	 * ringing the queue db when we're in DB_FULL mode.
2422 	 * Only allow this on T4 devices.
2423 	 */
2424 	attrs.sq_db_inc = attr->sq_psn;
2425 	attrs.rq_db_inc = attr->rq_psn;
2426 	mask |= (attr_mask & IB_QP_SQ_PSN) ? C4IW_QP_ATTR_SQ_DB : 0;
2427 	mask |= (attr_mask & IB_QP_RQ_PSN) ? C4IW_QP_ATTR_RQ_DB : 0;
2428 	if (!is_t4(to_c4iw_qp(ibqp)->rhp->rdev.lldi.adapter_type) &&
2429 	    (mask & (C4IW_QP_ATTR_SQ_DB|C4IW_QP_ATTR_RQ_DB)))
2430 		return -EINVAL;
2431 
2432 	return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
2433 }
2434 
2435 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
2436 {
2437 	pr_debug("ib_dev %p qpn 0x%x\n", dev, qpn);
2438 	return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
2439 }
2440 
2441 void c4iw_dispatch_srq_limit_reached_event(struct c4iw_srq *srq)
2442 {
2443 	struct ib_event event = {};
2444 
2445 	event.device = &srq->rhp->ibdev;
2446 	event.element.srq = &srq->ibsrq;
2447 	event.event = IB_EVENT_SRQ_LIMIT_REACHED;
2448 	ib_dispatch_event(&event);
2449 }
2450 
2451 int c4iw_modify_srq(struct ib_srq *ib_srq, struct ib_srq_attr *attr,
2452 		    enum ib_srq_attr_mask srq_attr_mask,
2453 		    struct ib_udata *udata)
2454 {
2455 	struct c4iw_srq *srq = to_c4iw_srq(ib_srq);
2456 	int ret = 0;
2457 
2458 	/*
2459 	 * XXX 0 mask == a SW interrupt for srq_limit reached...
2460 	 */
2461 	if (udata && !srq_attr_mask) {
2462 		c4iw_dispatch_srq_limit_reached_event(srq);
2463 		goto out;
2464 	}
2465 
2466 	/* no support for this yet */
2467 	if (srq_attr_mask & IB_SRQ_MAX_WR) {
2468 		ret = -EINVAL;
2469 		goto out;
2470 	}
2471 
2472 	if (!udata && (srq_attr_mask & IB_SRQ_LIMIT)) {
2473 		srq->armed = true;
2474 		srq->srq_limit = attr->srq_limit;
2475 	}
2476 out:
2477 	return ret;
2478 }
2479 
2480 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2481 		     int attr_mask, struct ib_qp_init_attr *init_attr)
2482 {
2483 	struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
2484 
2485 	memset(attr, 0, sizeof *attr);
2486 	memset(init_attr, 0, sizeof *init_attr);
2487 	attr->qp_state = to_ib_qp_state(qhp->attr.state);
2488 	init_attr->cap.max_send_wr = qhp->attr.sq_num_entries;
2489 	init_attr->cap.max_recv_wr = qhp->attr.rq_num_entries;
2490 	init_attr->cap.max_send_sge = qhp->attr.sq_max_sges;
2491 	init_attr->cap.max_recv_sge = qhp->attr.sq_max_sges;
2492 	init_attr->cap.max_inline_data = T4_MAX_SEND_INLINE;
2493 	init_attr->sq_sig_type = qhp->sq_sig_all ? IB_SIGNAL_ALL_WR : 0;
2494 	return 0;
2495 }
2496 
2497 static void free_srq_queue(struct c4iw_srq *srq, struct c4iw_dev_ucontext *uctx,
2498 			   struct c4iw_wr_wait *wr_waitp)
2499 {
2500 	struct c4iw_rdev *rdev = &srq->rhp->rdev;
2501 	struct sk_buff *skb = srq->destroy_skb;
2502 	struct t4_srq *wq = &srq->wq;
2503 	struct fw_ri_res_wr *res_wr;
2504 	struct fw_ri_res *res;
2505 	int wr_len;
2506 
2507 	wr_len = sizeof(*res_wr) + sizeof(*res);
2508 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
2509 
2510 	res_wr = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
2511 	memset(res_wr, 0, wr_len);
2512 	res_wr->op_nres = cpu_to_be32(FW_WR_OP_V(FW_RI_RES_WR) |
2513 			FW_RI_RES_WR_NRES_V(1) |
2514 			FW_WR_COMPL_F);
2515 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
2516 	res_wr->cookie = (uintptr_t)wr_waitp;
2517 	res = res_wr->res;
2518 	res->u.srq.restype = FW_RI_RES_TYPE_SRQ;
2519 	res->u.srq.op = FW_RI_RES_OP_RESET;
2520 	res->u.srq.srqid = cpu_to_be32(srq->idx);
2521 	res->u.srq.eqid = cpu_to_be32(wq->qid);
2522 
2523 	c4iw_init_wr_wait(wr_waitp);
2524 	c4iw_ref_send_wait(rdev, skb, wr_waitp, 0, 0, __func__);
2525 
2526 	dma_free_coherent(&rdev->lldi.pdev->dev,
2527 			  wq->memsize, wq->queue,
2528 			dma_unmap_addr(wq, mapping));
2529 	c4iw_rqtpool_free(rdev, wq->rqt_hwaddr, wq->rqt_size);
2530 	kfree(wq->sw_rq);
2531 	c4iw_put_qpid(rdev, wq->qid, uctx);
2532 }
2533 
2534 static int alloc_srq_queue(struct c4iw_srq *srq, struct c4iw_dev_ucontext *uctx,
2535 			   struct c4iw_wr_wait *wr_waitp)
2536 {
2537 	struct c4iw_rdev *rdev = &srq->rhp->rdev;
2538 	int user = (uctx != &rdev->uctx);
2539 	struct t4_srq *wq = &srq->wq;
2540 	struct fw_ri_res_wr *res_wr;
2541 	struct fw_ri_res *res;
2542 	struct sk_buff *skb;
2543 	int wr_len;
2544 	int eqsize;
2545 	int ret = -ENOMEM;
2546 
2547 	wq->qid = c4iw_get_qpid(rdev, uctx);
2548 	if (!wq->qid)
2549 		goto err;
2550 
2551 	if (!user) {
2552 		wq->sw_rq = kcalloc(wq->size, sizeof(*wq->sw_rq),
2553 				    GFP_KERNEL);
2554 		if (!wq->sw_rq)
2555 			goto err_put_qpid;
2556 		wq->pending_wrs = kcalloc(srq->wq.size,
2557 					  sizeof(*srq->wq.pending_wrs),
2558 					  GFP_KERNEL);
2559 		if (!wq->pending_wrs)
2560 			goto err_free_sw_rq;
2561 	}
2562 
2563 	wq->rqt_size = wq->size;
2564 	wq->rqt_hwaddr = c4iw_rqtpool_alloc(rdev, wq->rqt_size);
2565 	if (!wq->rqt_hwaddr)
2566 		goto err_free_pending_wrs;
2567 	wq->rqt_abs_idx = (wq->rqt_hwaddr - rdev->lldi.vr->rq.start) >>
2568 		T4_RQT_ENTRY_SHIFT;
2569 
2570 	wq->queue = dma_alloc_coherent(&rdev->lldi.pdev->dev, wq->memsize,
2571 				       &wq->dma_addr, GFP_KERNEL);
2572 	if (!wq->queue)
2573 		goto err_free_rqtpool;
2574 
2575 	dma_unmap_addr_set(wq, mapping, wq->dma_addr);
2576 
2577 	wq->bar2_va = c4iw_bar2_addrs(rdev, wq->qid, CXGB4_BAR2_QTYPE_EGRESS,
2578 				      &wq->bar2_qid,
2579 			user ? &wq->bar2_pa : NULL);
2580 
2581 	/*
2582 	 * User mode must have bar2 access.
2583 	 */
2584 
2585 	if (user && !wq->bar2_va) {
2586 		pr_warn(MOD "%s: srqid %u not in BAR2 range.\n",
2587 			pci_name(rdev->lldi.pdev), wq->qid);
2588 		ret = -EINVAL;
2589 		goto err_free_queue;
2590 	}
2591 
2592 	/* build fw_ri_res_wr */
2593 	wr_len = sizeof(*res_wr) + sizeof(*res);
2594 
2595 	skb = alloc_skb(wr_len, GFP_KERNEL);
2596 	if (!skb)
2597 		goto err_free_queue;
2598 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
2599 
2600 	res_wr = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
2601 	memset(res_wr, 0, wr_len);
2602 	res_wr->op_nres = cpu_to_be32(FW_WR_OP_V(FW_RI_RES_WR) |
2603 			FW_RI_RES_WR_NRES_V(1) |
2604 			FW_WR_COMPL_F);
2605 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
2606 	res_wr->cookie = (uintptr_t)wr_waitp;
2607 	res = res_wr->res;
2608 	res->u.srq.restype = FW_RI_RES_TYPE_SRQ;
2609 	res->u.srq.op = FW_RI_RES_OP_WRITE;
2610 
2611 	/*
2612 	 * eqsize is the number of 64B entries plus the status page size.
2613 	 */
2614 	eqsize = wq->size * T4_RQ_NUM_SLOTS +
2615 		rdev->hw_queue.t4_eq_status_entries;
2616 	res->u.srq.eqid = cpu_to_be32(wq->qid);
2617 	res->u.srq.fetchszm_to_iqid =
2618 						/* no host cidx updates */
2619 		cpu_to_be32(FW_RI_RES_WR_HOSTFCMODE_V(0) |
2620 		FW_RI_RES_WR_CPRIO_V(0) |       /* don't keep in chip cache */
2621 		FW_RI_RES_WR_PCIECHN_V(0) |     /* set by uP at ri_init time */
2622 		FW_RI_RES_WR_FETCHRO_V(0));     /* relaxed_ordering */
2623 	res->u.srq.dcaen_to_eqsize =
2624 		cpu_to_be32(FW_RI_RES_WR_DCAEN_V(0) |
2625 		FW_RI_RES_WR_DCACPU_V(0) |
2626 		FW_RI_RES_WR_FBMIN_V(2) |
2627 		FW_RI_RES_WR_FBMAX_V(3) |
2628 		FW_RI_RES_WR_CIDXFTHRESHO_V(0) |
2629 		FW_RI_RES_WR_CIDXFTHRESH_V(0) |
2630 		FW_RI_RES_WR_EQSIZE_V(eqsize));
2631 	res->u.srq.eqaddr = cpu_to_be64(wq->dma_addr);
2632 	res->u.srq.srqid = cpu_to_be32(srq->idx);
2633 	res->u.srq.pdid = cpu_to_be32(srq->pdid);
2634 	res->u.srq.hwsrqsize = cpu_to_be32(wq->rqt_size);
2635 	res->u.srq.hwsrqaddr = cpu_to_be32(wq->rqt_hwaddr -
2636 			rdev->lldi.vr->rq.start);
2637 
2638 	c4iw_init_wr_wait(wr_waitp);
2639 
2640 	ret = c4iw_ref_send_wait(rdev, skb, wr_waitp, 0, wq->qid, __func__);
2641 	if (ret)
2642 		goto err_free_queue;
2643 
2644 	pr_debug("%s srq %u eqid %u pdid %u queue va %p pa 0x%llx\n"
2645 			" bar2_addr %p rqt addr 0x%x size %d\n",
2646 			__func__, srq->idx, wq->qid, srq->pdid, wq->queue,
2647 			(u64)virt_to_phys(wq->queue), wq->bar2_va,
2648 			wq->rqt_hwaddr, wq->rqt_size);
2649 
2650 	return 0;
2651 err_free_queue:
2652 	dma_free_coherent(&rdev->lldi.pdev->dev,
2653 			  wq->memsize, wq->queue,
2654 			dma_unmap_addr(wq, mapping));
2655 err_free_rqtpool:
2656 	c4iw_rqtpool_free(rdev, wq->rqt_hwaddr, wq->rqt_size);
2657 err_free_pending_wrs:
2658 	if (!user)
2659 		kfree(wq->pending_wrs);
2660 err_free_sw_rq:
2661 	if (!user)
2662 		kfree(wq->sw_rq);
2663 err_put_qpid:
2664 	c4iw_put_qpid(rdev, wq->qid, uctx);
2665 err:
2666 	return ret;
2667 }
2668 
2669 void c4iw_copy_wr_to_srq(struct t4_srq *srq, union t4_recv_wr *wqe, u8 len16)
2670 {
2671 	u64 *src, *dst;
2672 
2673 	src = (u64 *)wqe;
2674 	dst = (u64 *)((u8 *)srq->queue + srq->wq_pidx * T4_EQ_ENTRY_SIZE);
2675 	while (len16) {
2676 		*dst++ = *src++;
2677 		if (dst >= (u64 *)&srq->queue[srq->size])
2678 			dst = (u64 *)srq->queue;
2679 		*dst++ = *src++;
2680 		if (dst >= (u64 *)&srq->queue[srq->size])
2681 			dst = (u64 *)srq->queue;
2682 		len16--;
2683 	}
2684 }
2685 
2686 int c4iw_create_srq(struct ib_srq *ib_srq, struct ib_srq_init_attr *attrs,
2687 			       struct ib_udata *udata)
2688 {
2689 	struct ib_pd *pd = ib_srq->pd;
2690 	struct c4iw_dev *rhp;
2691 	struct c4iw_srq *srq = to_c4iw_srq(ib_srq);
2692 	struct c4iw_pd *php;
2693 	struct c4iw_create_srq_resp uresp;
2694 	struct c4iw_ucontext *ucontext;
2695 	struct c4iw_mm_entry *srq_key_mm, *srq_db_key_mm;
2696 	int rqsize;
2697 	int ret;
2698 	int wr_len;
2699 
2700 	pr_debug("%s ib_pd %p\n", __func__, pd);
2701 
2702 	php = to_c4iw_pd(pd);
2703 	rhp = php->rhp;
2704 
2705 	if (!rhp->rdev.lldi.vr->srq.size)
2706 		return -EINVAL;
2707 	if (attrs->attr.max_wr > rhp->rdev.hw_queue.t4_max_rq_size)
2708 		return -E2BIG;
2709 	if (attrs->attr.max_sge > T4_MAX_RECV_SGE)
2710 		return -E2BIG;
2711 
2712 	/*
2713 	 * SRQ RQT and RQ must be a power of 2 and at least 16 deep.
2714 	 */
2715 	rqsize = attrs->attr.max_wr + 1;
2716 	rqsize = roundup_pow_of_two(max_t(u16, rqsize, 16));
2717 
2718 	ucontext = rdma_udata_to_drv_context(udata, struct c4iw_ucontext,
2719 					     ibucontext);
2720 
2721 	srq->wr_waitp = c4iw_alloc_wr_wait(GFP_KERNEL);
2722 	if (!srq->wr_waitp)
2723 		return -ENOMEM;
2724 
2725 	srq->idx = c4iw_alloc_srq_idx(&rhp->rdev);
2726 	if (srq->idx < 0) {
2727 		ret = -ENOMEM;
2728 		goto err_free_wr_wait;
2729 	}
2730 
2731 	wr_len = sizeof(struct fw_ri_res_wr) + sizeof(struct fw_ri_res);
2732 	srq->destroy_skb = alloc_skb(wr_len, GFP_KERNEL);
2733 	if (!srq->destroy_skb) {
2734 		ret = -ENOMEM;
2735 		goto err_free_srq_idx;
2736 	}
2737 
2738 	srq->rhp = rhp;
2739 	srq->pdid = php->pdid;
2740 
2741 	srq->wq.size = rqsize;
2742 	srq->wq.memsize =
2743 		(rqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
2744 		sizeof(*srq->wq.queue);
2745 	if (ucontext)
2746 		srq->wq.memsize = roundup(srq->wq.memsize, PAGE_SIZE);
2747 
2748 	ret = alloc_srq_queue(srq, ucontext ? &ucontext->uctx :
2749 			&rhp->rdev.uctx, srq->wr_waitp);
2750 	if (ret)
2751 		goto err_free_skb;
2752 	attrs->attr.max_wr = rqsize - 1;
2753 
2754 	if (CHELSIO_CHIP_VERSION(rhp->rdev.lldi.adapter_type) > CHELSIO_T6)
2755 		srq->flags = T4_SRQ_LIMIT_SUPPORT;
2756 
2757 	ret = xa_insert_irq(&rhp->qps, srq->wq.qid, srq, GFP_KERNEL);
2758 	if (ret)
2759 		goto err_free_queue;
2760 
2761 	if (udata) {
2762 		srq_key_mm = kmalloc(sizeof(*srq_key_mm), GFP_KERNEL);
2763 		if (!srq_key_mm) {
2764 			ret = -ENOMEM;
2765 			goto err_remove_handle;
2766 		}
2767 		srq_db_key_mm = kmalloc(sizeof(*srq_db_key_mm), GFP_KERNEL);
2768 		if (!srq_db_key_mm) {
2769 			ret = -ENOMEM;
2770 			goto err_free_srq_key_mm;
2771 		}
2772 		memset(&uresp, 0, sizeof(uresp));
2773 		uresp.flags = srq->flags;
2774 		uresp.qid_mask = rhp->rdev.qpmask;
2775 		uresp.srqid = srq->wq.qid;
2776 		uresp.srq_size = srq->wq.size;
2777 		uresp.srq_memsize = srq->wq.memsize;
2778 		uresp.rqt_abs_idx = srq->wq.rqt_abs_idx;
2779 		spin_lock(&ucontext->mmap_lock);
2780 		uresp.srq_key = ucontext->key;
2781 		ucontext->key += PAGE_SIZE;
2782 		uresp.srq_db_gts_key = ucontext->key;
2783 		ucontext->key += PAGE_SIZE;
2784 		spin_unlock(&ucontext->mmap_lock);
2785 		ret = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
2786 		if (ret)
2787 			goto err_free_srq_db_key_mm;
2788 		srq_key_mm->key = uresp.srq_key;
2789 		srq_key_mm->addr = virt_to_phys(srq->wq.queue);
2790 		srq_key_mm->len = PAGE_ALIGN(srq->wq.memsize);
2791 		insert_mmap(ucontext, srq_key_mm);
2792 		srq_db_key_mm->key = uresp.srq_db_gts_key;
2793 		srq_db_key_mm->addr = (u64)(unsigned long)srq->wq.bar2_pa;
2794 		srq_db_key_mm->len = PAGE_SIZE;
2795 		insert_mmap(ucontext, srq_db_key_mm);
2796 	}
2797 
2798 	pr_debug("%s srq qid %u idx %u size %u memsize %lu num_entries %u\n",
2799 		 __func__, srq->wq.qid, srq->idx, srq->wq.size,
2800 			(unsigned long)srq->wq.memsize, attrs->attr.max_wr);
2801 
2802 	spin_lock_init(&srq->lock);
2803 	return 0;
2804 
2805 err_free_srq_db_key_mm:
2806 	kfree(srq_db_key_mm);
2807 err_free_srq_key_mm:
2808 	kfree(srq_key_mm);
2809 err_remove_handle:
2810 	xa_erase_irq(&rhp->qps, srq->wq.qid);
2811 err_free_queue:
2812 	free_srq_queue(srq, ucontext ? &ucontext->uctx : &rhp->rdev.uctx,
2813 		       srq->wr_waitp);
2814 err_free_skb:
2815 	kfree_skb(srq->destroy_skb);
2816 err_free_srq_idx:
2817 	c4iw_free_srq_idx(&rhp->rdev, srq->idx);
2818 err_free_wr_wait:
2819 	c4iw_put_wr_wait(srq->wr_waitp);
2820 	return ret;
2821 }
2822 
2823 void c4iw_destroy_srq(struct ib_srq *ibsrq, struct ib_udata *udata)
2824 {
2825 	struct c4iw_dev *rhp;
2826 	struct c4iw_srq *srq;
2827 	struct c4iw_ucontext *ucontext;
2828 
2829 	srq = to_c4iw_srq(ibsrq);
2830 	rhp = srq->rhp;
2831 
2832 	pr_debug("%s id %d\n", __func__, srq->wq.qid);
2833 
2834 	xa_erase_irq(&rhp->qps, srq->wq.qid);
2835 	ucontext = rdma_udata_to_drv_context(udata, struct c4iw_ucontext,
2836 					     ibucontext);
2837 	free_srq_queue(srq, ucontext ? &ucontext->uctx : &rhp->rdev.uctx,
2838 		       srq->wr_waitp);
2839 	c4iw_free_srq_idx(&rhp->rdev, srq->idx);
2840 	c4iw_put_wr_wait(srq->wr_waitp);
2841 }
2842