xref: /linux/drivers/infiniband/hw/cxgb4/qp.c (revision 32786fdc9506aeba98278c1844d4bfb766863832)
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 
35 #include "iw_cxgb4.h"
36 
37 static int db_delay_usecs = 1;
38 module_param(db_delay_usecs, int, 0644);
39 MODULE_PARM_DESC(db_delay_usecs, "Usecs to delay awaiting db fifo to drain");
40 
41 static int ocqp_support = 1;
42 module_param(ocqp_support, int, 0644);
43 MODULE_PARM_DESC(ocqp_support, "Support on-chip SQs (default=1)");
44 
45 int db_fc_threshold = 1000;
46 module_param(db_fc_threshold, int, 0644);
47 MODULE_PARM_DESC(db_fc_threshold,
48 		 "QP count/threshold that triggers"
49 		 " automatic db flow control mode (default = 1000)");
50 
51 int db_coalescing_threshold;
52 module_param(db_coalescing_threshold, int, 0644);
53 MODULE_PARM_DESC(db_coalescing_threshold,
54 		 "QP count/threshold that triggers"
55 		 " disabling db coalescing (default = 0)");
56 
57 static int max_fr_immd = T4_MAX_FR_IMMD;
58 module_param(max_fr_immd, int, 0644);
59 MODULE_PARM_DESC(max_fr_immd, "fastreg threshold for using DSGL instead of immedate");
60 
61 static int alloc_ird(struct c4iw_dev *dev, u32 ird)
62 {
63 	int ret = 0;
64 
65 	spin_lock_irq(&dev->lock);
66 	if (ird <= dev->avail_ird)
67 		dev->avail_ird -= ird;
68 	else
69 		ret = -ENOMEM;
70 	spin_unlock_irq(&dev->lock);
71 
72 	if (ret)
73 		dev_warn(&dev->rdev.lldi.pdev->dev,
74 			 "device IRD resources exhausted\n");
75 
76 	return ret;
77 }
78 
79 static void free_ird(struct c4iw_dev *dev, int ird)
80 {
81 	spin_lock_irq(&dev->lock);
82 	dev->avail_ird += ird;
83 	spin_unlock_irq(&dev->lock);
84 }
85 
86 static void set_state(struct c4iw_qp *qhp, enum c4iw_qp_state state)
87 {
88 	unsigned long flag;
89 	spin_lock_irqsave(&qhp->lock, flag);
90 	qhp->attr.state = state;
91 	spin_unlock_irqrestore(&qhp->lock, flag);
92 }
93 
94 static void dealloc_oc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
95 {
96 	c4iw_ocqp_pool_free(rdev, sq->dma_addr, sq->memsize);
97 }
98 
99 static void dealloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
100 {
101 	dma_free_coherent(&(rdev->lldi.pdev->dev), sq->memsize, sq->queue,
102 			  pci_unmap_addr(sq, mapping));
103 }
104 
105 static void dealloc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
106 {
107 	if (t4_sq_onchip(sq))
108 		dealloc_oc_sq(rdev, sq);
109 	else
110 		dealloc_host_sq(rdev, sq);
111 }
112 
113 static int alloc_oc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
114 {
115 	if (!ocqp_support || !ocqp_supported(&rdev->lldi))
116 		return -ENOSYS;
117 	sq->dma_addr = c4iw_ocqp_pool_alloc(rdev, sq->memsize);
118 	if (!sq->dma_addr)
119 		return -ENOMEM;
120 	sq->phys_addr = rdev->oc_mw_pa + sq->dma_addr -
121 			rdev->lldi.vr->ocq.start;
122 	sq->queue = (__force union t4_wr *)(rdev->oc_mw_kva + sq->dma_addr -
123 					    rdev->lldi.vr->ocq.start);
124 	sq->flags |= T4_SQ_ONCHIP;
125 	return 0;
126 }
127 
128 static int alloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
129 {
130 	sq->queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev), sq->memsize,
131 				       &(sq->dma_addr), GFP_KERNEL);
132 	if (!sq->queue)
133 		return -ENOMEM;
134 	sq->phys_addr = virt_to_phys(sq->queue);
135 	pci_unmap_addr_set(sq, mapping, sq->dma_addr);
136 	return 0;
137 }
138 
139 static int alloc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq, int user)
140 {
141 	int ret = -ENOSYS;
142 	if (user)
143 		ret = alloc_oc_sq(rdev, sq);
144 	if (ret)
145 		ret = alloc_host_sq(rdev, sq);
146 	return ret;
147 }
148 
149 static int destroy_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
150 		      struct c4iw_dev_ucontext *uctx)
151 {
152 	/*
153 	 * uP clears EQ contexts when the connection exits rdma mode,
154 	 * so no need to post a RESET WR for these EQs.
155 	 */
156 	dma_free_coherent(&(rdev->lldi.pdev->dev),
157 			  wq->rq.memsize, wq->rq.queue,
158 			  dma_unmap_addr(&wq->rq, mapping));
159 	dealloc_sq(rdev, &wq->sq);
160 	c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
161 	kfree(wq->rq.sw_rq);
162 	kfree(wq->sq.sw_sq);
163 	c4iw_put_qpid(rdev, wq->rq.qid, uctx);
164 	c4iw_put_qpid(rdev, wq->sq.qid, uctx);
165 	return 0;
166 }
167 
168 /*
169  * Determine the BAR2 virtual address and qid. If pbar2_pa is not NULL,
170  * then this is a user mapping so compute the page-aligned physical address
171  * for mapping.
172  */
173 void __iomem *c4iw_bar2_addrs(struct c4iw_rdev *rdev, unsigned int qid,
174 			      enum cxgb4_bar2_qtype qtype,
175 			      unsigned int *pbar2_qid, u64 *pbar2_pa)
176 {
177 	u64 bar2_qoffset;
178 	int ret;
179 
180 	ret = cxgb4_bar2_sge_qregs(rdev->lldi.ports[0], qid, qtype,
181 				   pbar2_pa ? 1 : 0,
182 				   &bar2_qoffset, pbar2_qid);
183 	if (ret)
184 		return NULL;
185 
186 	if (pbar2_pa)
187 		*pbar2_pa = (rdev->bar2_pa + bar2_qoffset) & PAGE_MASK;
188 
189 	if (is_t4(rdev->lldi.adapter_type))
190 		return NULL;
191 
192 	return rdev->bar2_kva + bar2_qoffset;
193 }
194 
195 static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
196 		     struct t4_cq *rcq, struct t4_cq *scq,
197 		     struct c4iw_dev_ucontext *uctx)
198 {
199 	int user = (uctx != &rdev->uctx);
200 	struct fw_ri_res_wr *res_wr;
201 	struct fw_ri_res *res;
202 	int wr_len;
203 	struct c4iw_wr_wait wr_wait;
204 	struct sk_buff *skb;
205 	int ret = 0;
206 	int eqsize;
207 
208 	wq->sq.qid = c4iw_get_qpid(rdev, uctx);
209 	if (!wq->sq.qid)
210 		return -ENOMEM;
211 
212 	wq->rq.qid = c4iw_get_qpid(rdev, uctx);
213 	if (!wq->rq.qid) {
214 		ret = -ENOMEM;
215 		goto free_sq_qid;
216 	}
217 
218 	if (!user) {
219 		wq->sq.sw_sq = kzalloc(wq->sq.size * sizeof *wq->sq.sw_sq,
220 				 GFP_KERNEL);
221 		if (!wq->sq.sw_sq) {
222 			ret = -ENOMEM;
223 			goto free_rq_qid;
224 		}
225 
226 		wq->rq.sw_rq = kzalloc(wq->rq.size * sizeof *wq->rq.sw_rq,
227 				 GFP_KERNEL);
228 		if (!wq->rq.sw_rq) {
229 			ret = -ENOMEM;
230 			goto free_sw_sq;
231 		}
232 	}
233 
234 	/*
235 	 * RQT must be a power of 2 and at least 16 deep.
236 	 */
237 	wq->rq.rqt_size = roundup_pow_of_two(max_t(u16, wq->rq.size, 16));
238 	wq->rq.rqt_hwaddr = c4iw_rqtpool_alloc(rdev, wq->rq.rqt_size);
239 	if (!wq->rq.rqt_hwaddr) {
240 		ret = -ENOMEM;
241 		goto free_sw_rq;
242 	}
243 
244 	ret = alloc_sq(rdev, &wq->sq, user);
245 	if (ret)
246 		goto free_hwaddr;
247 	memset(wq->sq.queue, 0, wq->sq.memsize);
248 	dma_unmap_addr_set(&wq->sq, mapping, wq->sq.dma_addr);
249 
250 	wq->rq.queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev),
251 					  wq->rq.memsize, &(wq->rq.dma_addr),
252 					  GFP_KERNEL);
253 	if (!wq->rq.queue) {
254 		ret = -ENOMEM;
255 		goto free_sq;
256 	}
257 	PDBG("%s sq base va 0x%p pa 0x%llx rq base va 0x%p pa 0x%llx\n",
258 		__func__, wq->sq.queue,
259 		(unsigned long long)virt_to_phys(wq->sq.queue),
260 		wq->rq.queue,
261 		(unsigned long long)virt_to_phys(wq->rq.queue));
262 	memset(wq->rq.queue, 0, wq->rq.memsize);
263 	dma_unmap_addr_set(&wq->rq, mapping, wq->rq.dma_addr);
264 
265 	wq->db = rdev->lldi.db_reg;
266 
267 	wq->sq.bar2_va = c4iw_bar2_addrs(rdev, wq->sq.qid, T4_BAR2_QTYPE_EGRESS,
268 					 &wq->sq.bar2_qid,
269 					 user ? &wq->sq.bar2_pa : NULL);
270 	wq->rq.bar2_va = c4iw_bar2_addrs(rdev, wq->rq.qid, T4_BAR2_QTYPE_EGRESS,
271 					 &wq->rq.bar2_qid,
272 					 user ? &wq->rq.bar2_pa : NULL);
273 
274 	/*
275 	 * User mode must have bar2 access.
276 	 */
277 	if (user && (!wq->sq.bar2_pa || !wq->rq.bar2_pa)) {
278 		pr_warn(MOD "%s: sqid %u or rqid %u not in BAR2 range.\n",
279 			pci_name(rdev->lldi.pdev), wq->sq.qid, wq->rq.qid);
280 		goto free_dma;
281 	}
282 
283 	wq->rdev = rdev;
284 	wq->rq.msn = 1;
285 
286 	/* build fw_ri_res_wr */
287 	wr_len = sizeof *res_wr + 2 * sizeof *res;
288 
289 	skb = alloc_skb(wr_len, GFP_KERNEL);
290 	if (!skb) {
291 		ret = -ENOMEM;
292 		goto free_dma;
293 	}
294 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
295 
296 	res_wr = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
297 	memset(res_wr, 0, wr_len);
298 	res_wr->op_nres = cpu_to_be32(
299 			FW_WR_OP_V(FW_RI_RES_WR) |
300 			FW_RI_RES_WR_NRES_V(2) |
301 			FW_WR_COMPL_F);
302 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
303 	res_wr->cookie = (uintptr_t)&wr_wait;
304 	res = res_wr->res;
305 	res->u.sqrq.restype = FW_RI_RES_TYPE_SQ;
306 	res->u.sqrq.op = FW_RI_RES_OP_WRITE;
307 
308 	/*
309 	 * eqsize is the number of 64B entries plus the status page size.
310 	 */
311 	eqsize = wq->sq.size * T4_SQ_NUM_SLOTS +
312 		rdev->hw_queue.t4_eq_status_entries;
313 
314 	res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
315 		FW_RI_RES_WR_HOSTFCMODE_V(0) |	/* no host cidx updates */
316 		FW_RI_RES_WR_CPRIO_V(0) |	/* don't keep in chip cache */
317 		FW_RI_RES_WR_PCIECHN_V(0) |	/* set by uP at ri_init time */
318 		(t4_sq_onchip(&wq->sq) ? FW_RI_RES_WR_ONCHIP_F : 0) |
319 		FW_RI_RES_WR_IQID_V(scq->cqid));
320 	res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
321 		FW_RI_RES_WR_DCAEN_V(0) |
322 		FW_RI_RES_WR_DCACPU_V(0) |
323 		FW_RI_RES_WR_FBMIN_V(2) |
324 		FW_RI_RES_WR_FBMAX_V(2) |
325 		FW_RI_RES_WR_CIDXFTHRESHO_V(0) |
326 		FW_RI_RES_WR_CIDXFTHRESH_V(0) |
327 		FW_RI_RES_WR_EQSIZE_V(eqsize));
328 	res->u.sqrq.eqid = cpu_to_be32(wq->sq.qid);
329 	res->u.sqrq.eqaddr = cpu_to_be64(wq->sq.dma_addr);
330 	res++;
331 	res->u.sqrq.restype = FW_RI_RES_TYPE_RQ;
332 	res->u.sqrq.op = FW_RI_RES_OP_WRITE;
333 
334 	/*
335 	 * eqsize is the number of 64B entries plus the status page size.
336 	 */
337 	eqsize = wq->rq.size * T4_RQ_NUM_SLOTS +
338 		rdev->hw_queue.t4_eq_status_entries;
339 	res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
340 		FW_RI_RES_WR_HOSTFCMODE_V(0) |	/* no host cidx updates */
341 		FW_RI_RES_WR_CPRIO_V(0) |	/* don't keep in chip cache */
342 		FW_RI_RES_WR_PCIECHN_V(0) |	/* set by uP at ri_init time */
343 		FW_RI_RES_WR_IQID_V(rcq->cqid));
344 	res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
345 		FW_RI_RES_WR_DCAEN_V(0) |
346 		FW_RI_RES_WR_DCACPU_V(0) |
347 		FW_RI_RES_WR_FBMIN_V(2) |
348 		FW_RI_RES_WR_FBMAX_V(2) |
349 		FW_RI_RES_WR_CIDXFTHRESHO_V(0) |
350 		FW_RI_RES_WR_CIDXFTHRESH_V(0) |
351 		FW_RI_RES_WR_EQSIZE_V(eqsize));
352 	res->u.sqrq.eqid = cpu_to_be32(wq->rq.qid);
353 	res->u.sqrq.eqaddr = cpu_to_be64(wq->rq.dma_addr);
354 
355 	c4iw_init_wr_wait(&wr_wait);
356 
357 	ret = c4iw_ofld_send(rdev, skb);
358 	if (ret)
359 		goto free_dma;
360 	ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, wq->sq.qid, __func__);
361 	if (ret)
362 		goto free_dma;
363 
364 	PDBG("%s sqid 0x%x rqid 0x%x kdb 0x%p sq_bar2_addr %p rq_bar2_addr %p\n",
365 	     __func__, wq->sq.qid, wq->rq.qid, wq->db,
366 	     wq->sq.bar2_va, wq->rq.bar2_va);
367 
368 	return 0;
369 free_dma:
370 	dma_free_coherent(&(rdev->lldi.pdev->dev),
371 			  wq->rq.memsize, wq->rq.queue,
372 			  dma_unmap_addr(&wq->rq, mapping));
373 free_sq:
374 	dealloc_sq(rdev, &wq->sq);
375 free_hwaddr:
376 	c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
377 free_sw_rq:
378 	kfree(wq->rq.sw_rq);
379 free_sw_sq:
380 	kfree(wq->sq.sw_sq);
381 free_rq_qid:
382 	c4iw_put_qpid(rdev, wq->rq.qid, uctx);
383 free_sq_qid:
384 	c4iw_put_qpid(rdev, wq->sq.qid, uctx);
385 	return ret;
386 }
387 
388 static int build_immd(struct t4_sq *sq, struct fw_ri_immd *immdp,
389 		      struct ib_send_wr *wr, int max, u32 *plenp)
390 {
391 	u8 *dstp, *srcp;
392 	u32 plen = 0;
393 	int i;
394 	int rem, len;
395 
396 	dstp = (u8 *)immdp->data;
397 	for (i = 0; i < wr->num_sge; i++) {
398 		if ((plen + wr->sg_list[i].length) > max)
399 			return -EMSGSIZE;
400 		srcp = (u8 *)(unsigned long)wr->sg_list[i].addr;
401 		plen += wr->sg_list[i].length;
402 		rem = wr->sg_list[i].length;
403 		while (rem) {
404 			if (dstp == (u8 *)&sq->queue[sq->size])
405 				dstp = (u8 *)sq->queue;
406 			if (rem <= (u8 *)&sq->queue[sq->size] - dstp)
407 				len = rem;
408 			else
409 				len = (u8 *)&sq->queue[sq->size] - dstp;
410 			memcpy(dstp, srcp, len);
411 			dstp += len;
412 			srcp += len;
413 			rem -= len;
414 		}
415 	}
416 	len = roundup(plen + sizeof *immdp, 16) - (plen + sizeof *immdp);
417 	if (len)
418 		memset(dstp, 0, len);
419 	immdp->op = FW_RI_DATA_IMMD;
420 	immdp->r1 = 0;
421 	immdp->r2 = 0;
422 	immdp->immdlen = cpu_to_be32(plen);
423 	*plenp = plen;
424 	return 0;
425 }
426 
427 static int build_isgl(__be64 *queue_start, __be64 *queue_end,
428 		      struct fw_ri_isgl *isglp, struct ib_sge *sg_list,
429 		      int num_sge, u32 *plenp)
430 
431 {
432 	int i;
433 	u32 plen = 0;
434 	__be64 *flitp = (__be64 *)isglp->sge;
435 
436 	for (i = 0; i < num_sge; i++) {
437 		if ((plen + sg_list[i].length) < plen)
438 			return -EMSGSIZE;
439 		plen += sg_list[i].length;
440 		*flitp = cpu_to_be64(((u64)sg_list[i].lkey << 32) |
441 				     sg_list[i].length);
442 		if (++flitp == queue_end)
443 			flitp = queue_start;
444 		*flitp = cpu_to_be64(sg_list[i].addr);
445 		if (++flitp == queue_end)
446 			flitp = queue_start;
447 	}
448 	*flitp = (__force __be64)0;
449 	isglp->op = FW_RI_DATA_ISGL;
450 	isglp->r1 = 0;
451 	isglp->nsge = cpu_to_be16(num_sge);
452 	isglp->r2 = 0;
453 	if (plenp)
454 		*plenp = plen;
455 	return 0;
456 }
457 
458 static int build_rdma_send(struct t4_sq *sq, union t4_wr *wqe,
459 			   struct ib_send_wr *wr, u8 *len16)
460 {
461 	u32 plen;
462 	int size;
463 	int ret;
464 
465 	if (wr->num_sge > T4_MAX_SEND_SGE)
466 		return -EINVAL;
467 	switch (wr->opcode) {
468 	case IB_WR_SEND:
469 		if (wr->send_flags & IB_SEND_SOLICITED)
470 			wqe->send.sendop_pkd = cpu_to_be32(
471 				FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_SE));
472 		else
473 			wqe->send.sendop_pkd = cpu_to_be32(
474 				FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND));
475 		wqe->send.stag_inv = 0;
476 		break;
477 	case IB_WR_SEND_WITH_INV:
478 		if (wr->send_flags & IB_SEND_SOLICITED)
479 			wqe->send.sendop_pkd = cpu_to_be32(
480 				FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_SE_INV));
481 		else
482 			wqe->send.sendop_pkd = cpu_to_be32(
483 				FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_INV));
484 		wqe->send.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
485 		break;
486 
487 	default:
488 		return -EINVAL;
489 	}
490 	wqe->send.r3 = 0;
491 	wqe->send.r4 = 0;
492 
493 	plen = 0;
494 	if (wr->num_sge) {
495 		if (wr->send_flags & IB_SEND_INLINE) {
496 			ret = build_immd(sq, wqe->send.u.immd_src, wr,
497 					 T4_MAX_SEND_INLINE, &plen);
498 			if (ret)
499 				return ret;
500 			size = sizeof wqe->send + sizeof(struct fw_ri_immd) +
501 			       plen;
502 		} else {
503 			ret = build_isgl((__be64 *)sq->queue,
504 					 (__be64 *)&sq->queue[sq->size],
505 					 wqe->send.u.isgl_src,
506 					 wr->sg_list, wr->num_sge, &plen);
507 			if (ret)
508 				return ret;
509 			size = sizeof wqe->send + sizeof(struct fw_ri_isgl) +
510 			       wr->num_sge * sizeof(struct fw_ri_sge);
511 		}
512 	} else {
513 		wqe->send.u.immd_src[0].op = FW_RI_DATA_IMMD;
514 		wqe->send.u.immd_src[0].r1 = 0;
515 		wqe->send.u.immd_src[0].r2 = 0;
516 		wqe->send.u.immd_src[0].immdlen = 0;
517 		size = sizeof wqe->send + sizeof(struct fw_ri_immd);
518 		plen = 0;
519 	}
520 	*len16 = DIV_ROUND_UP(size, 16);
521 	wqe->send.plen = cpu_to_be32(plen);
522 	return 0;
523 }
524 
525 static int build_rdma_write(struct t4_sq *sq, union t4_wr *wqe,
526 			    struct ib_send_wr *wr, u8 *len16)
527 {
528 	u32 plen;
529 	int size;
530 	int ret;
531 
532 	if (wr->num_sge > T4_MAX_SEND_SGE)
533 		return -EINVAL;
534 	wqe->write.r2 = 0;
535 	wqe->write.stag_sink = cpu_to_be32(rdma_wr(wr)->rkey);
536 	wqe->write.to_sink = cpu_to_be64(rdma_wr(wr)->remote_addr);
537 	if (wr->num_sge) {
538 		if (wr->send_flags & IB_SEND_INLINE) {
539 			ret = build_immd(sq, wqe->write.u.immd_src, wr,
540 					 T4_MAX_WRITE_INLINE, &plen);
541 			if (ret)
542 				return ret;
543 			size = sizeof wqe->write + sizeof(struct fw_ri_immd) +
544 			       plen;
545 		} else {
546 			ret = build_isgl((__be64 *)sq->queue,
547 					 (__be64 *)&sq->queue[sq->size],
548 					 wqe->write.u.isgl_src,
549 					 wr->sg_list, wr->num_sge, &plen);
550 			if (ret)
551 				return ret;
552 			size = sizeof wqe->write + sizeof(struct fw_ri_isgl) +
553 			       wr->num_sge * sizeof(struct fw_ri_sge);
554 		}
555 	} else {
556 		wqe->write.u.immd_src[0].op = FW_RI_DATA_IMMD;
557 		wqe->write.u.immd_src[0].r1 = 0;
558 		wqe->write.u.immd_src[0].r2 = 0;
559 		wqe->write.u.immd_src[0].immdlen = 0;
560 		size = sizeof wqe->write + sizeof(struct fw_ri_immd);
561 		plen = 0;
562 	}
563 	*len16 = DIV_ROUND_UP(size, 16);
564 	wqe->write.plen = cpu_to_be32(plen);
565 	return 0;
566 }
567 
568 static int build_rdma_read(union t4_wr *wqe, struct ib_send_wr *wr, u8 *len16)
569 {
570 	if (wr->num_sge > 1)
571 		return -EINVAL;
572 	if (wr->num_sge) {
573 		wqe->read.stag_src = cpu_to_be32(rdma_wr(wr)->rkey);
574 		wqe->read.to_src_hi = cpu_to_be32((u32)(rdma_wr(wr)->remote_addr
575 							>> 32));
576 		wqe->read.to_src_lo = cpu_to_be32((u32)rdma_wr(wr)->remote_addr);
577 		wqe->read.stag_sink = cpu_to_be32(wr->sg_list[0].lkey);
578 		wqe->read.plen = cpu_to_be32(wr->sg_list[0].length);
579 		wqe->read.to_sink_hi = cpu_to_be32((u32)(wr->sg_list[0].addr
580 							 >> 32));
581 		wqe->read.to_sink_lo = cpu_to_be32((u32)(wr->sg_list[0].addr));
582 	} else {
583 		wqe->read.stag_src = cpu_to_be32(2);
584 		wqe->read.to_src_hi = 0;
585 		wqe->read.to_src_lo = 0;
586 		wqe->read.stag_sink = cpu_to_be32(2);
587 		wqe->read.plen = 0;
588 		wqe->read.to_sink_hi = 0;
589 		wqe->read.to_sink_lo = 0;
590 	}
591 	wqe->read.r2 = 0;
592 	wqe->read.r5 = 0;
593 	*len16 = DIV_ROUND_UP(sizeof wqe->read, 16);
594 	return 0;
595 }
596 
597 static int build_rdma_recv(struct c4iw_qp *qhp, union t4_recv_wr *wqe,
598 			   struct ib_recv_wr *wr, u8 *len16)
599 {
600 	int ret;
601 
602 	ret = build_isgl((__be64 *)qhp->wq.rq.queue,
603 			 (__be64 *)&qhp->wq.rq.queue[qhp->wq.rq.size],
604 			 &wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL);
605 	if (ret)
606 		return ret;
607 	*len16 = DIV_ROUND_UP(sizeof wqe->recv +
608 			      wr->num_sge * sizeof(struct fw_ri_sge), 16);
609 	return 0;
610 }
611 
612 static void build_tpte_memreg(struct fw_ri_fr_nsmr_tpte_wr *fr,
613 			      struct ib_reg_wr *wr, struct c4iw_mr *mhp,
614 			      u8 *len16)
615 {
616 	__be64 *p = (__be64 *)fr->pbl;
617 
618 	fr->r2 = cpu_to_be32(0);
619 	fr->stag = cpu_to_be32(mhp->ibmr.rkey);
620 
621 	fr->tpte.valid_to_pdid = cpu_to_be32(FW_RI_TPTE_VALID_F |
622 		FW_RI_TPTE_STAGKEY_V((mhp->ibmr.rkey & FW_RI_TPTE_STAGKEY_M)) |
623 		FW_RI_TPTE_STAGSTATE_V(1) |
624 		FW_RI_TPTE_STAGTYPE_V(FW_RI_STAG_NSMR) |
625 		FW_RI_TPTE_PDID_V(mhp->attr.pdid));
626 	fr->tpte.locread_to_qpid = cpu_to_be32(
627 		FW_RI_TPTE_PERM_V(c4iw_ib_to_tpt_access(wr->access)) |
628 		FW_RI_TPTE_ADDRTYPE_V(FW_RI_VA_BASED_TO) |
629 		FW_RI_TPTE_PS_V(ilog2(wr->mr->page_size) - 12));
630 	fr->tpte.nosnoop_pbladdr = cpu_to_be32(FW_RI_TPTE_PBLADDR_V(
631 		PBL_OFF(&mhp->rhp->rdev, mhp->attr.pbl_addr)>>3));
632 	fr->tpte.dca_mwbcnt_pstag = cpu_to_be32(0);
633 	fr->tpte.len_hi = cpu_to_be32(0);
634 	fr->tpte.len_lo = cpu_to_be32(mhp->ibmr.length);
635 	fr->tpte.va_hi = cpu_to_be32(mhp->ibmr.iova >> 32);
636 	fr->tpte.va_lo_fbo = cpu_to_be32(mhp->ibmr.iova & 0xffffffff);
637 
638 	p[0] = cpu_to_be64((u64)mhp->mpl[0]);
639 	p[1] = cpu_to_be64((u64)mhp->mpl[1]);
640 
641 	*len16 = DIV_ROUND_UP(sizeof(*fr), 16);
642 }
643 
644 static int build_memreg(struct t4_sq *sq, union t4_wr *wqe,
645 			struct ib_reg_wr *wr, struct c4iw_mr *mhp, u8 *len16,
646 			bool dsgl_supported)
647 {
648 	struct fw_ri_immd *imdp;
649 	__be64 *p;
650 	int i;
651 	int pbllen = roundup(mhp->mpl_len * sizeof(u64), 32);
652 	int rem;
653 
654 	if (mhp->mpl_len > t4_max_fr_depth(dsgl_supported && use_dsgl))
655 		return -EINVAL;
656 
657 	wqe->fr.qpbinde_to_dcacpu = 0;
658 	wqe->fr.pgsz_shift = ilog2(wr->mr->page_size) - 12;
659 	wqe->fr.addr_type = FW_RI_VA_BASED_TO;
660 	wqe->fr.mem_perms = c4iw_ib_to_tpt_access(wr->access);
661 	wqe->fr.len_hi = 0;
662 	wqe->fr.len_lo = cpu_to_be32(mhp->ibmr.length);
663 	wqe->fr.stag = cpu_to_be32(wr->key);
664 	wqe->fr.va_hi = cpu_to_be32(mhp->ibmr.iova >> 32);
665 	wqe->fr.va_lo_fbo = cpu_to_be32(mhp->ibmr.iova &
666 					0xffffffff);
667 
668 	if (dsgl_supported && use_dsgl && (pbllen > max_fr_immd)) {
669 		struct fw_ri_dsgl *sglp;
670 
671 		for (i = 0; i < mhp->mpl_len; i++)
672 			mhp->mpl[i] = (__force u64)cpu_to_be64((u64)mhp->mpl[i]);
673 
674 		sglp = (struct fw_ri_dsgl *)(&wqe->fr + 1);
675 		sglp->op = FW_RI_DATA_DSGL;
676 		sglp->r1 = 0;
677 		sglp->nsge = cpu_to_be16(1);
678 		sglp->addr0 = cpu_to_be64(mhp->mpl_addr);
679 		sglp->len0 = cpu_to_be32(pbllen);
680 
681 		*len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*sglp), 16);
682 	} else {
683 		imdp = (struct fw_ri_immd *)(&wqe->fr + 1);
684 		imdp->op = FW_RI_DATA_IMMD;
685 		imdp->r1 = 0;
686 		imdp->r2 = 0;
687 		imdp->immdlen = cpu_to_be32(pbllen);
688 		p = (__be64 *)(imdp + 1);
689 		rem = pbllen;
690 		for (i = 0; i < mhp->mpl_len; i++) {
691 			*p = cpu_to_be64((u64)mhp->mpl[i]);
692 			rem -= sizeof(*p);
693 			if (++p == (__be64 *)&sq->queue[sq->size])
694 				p = (__be64 *)sq->queue;
695 		}
696 		BUG_ON(rem < 0);
697 		while (rem) {
698 			*p = 0;
699 			rem -= sizeof(*p);
700 			if (++p == (__be64 *)&sq->queue[sq->size])
701 				p = (__be64 *)sq->queue;
702 		}
703 		*len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*imdp)
704 				      + pbllen, 16);
705 	}
706 	return 0;
707 }
708 
709 static int build_inv_stag(union t4_wr *wqe, struct ib_send_wr *wr, u8 *len16)
710 {
711 	wqe->inv.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
712 	wqe->inv.r2 = 0;
713 	*len16 = DIV_ROUND_UP(sizeof wqe->inv, 16);
714 	return 0;
715 }
716 
717 static void _free_qp(struct kref *kref)
718 {
719 	struct c4iw_qp *qhp;
720 
721 	qhp = container_of(kref, struct c4iw_qp, kref);
722 	PDBG("%s qhp %p\n", __func__, qhp);
723 	kfree(qhp);
724 }
725 
726 void c4iw_qp_add_ref(struct ib_qp *qp)
727 {
728 	PDBG("%s ib_qp %p\n", __func__, qp);
729 	kref_get(&to_c4iw_qp(qp)->kref);
730 }
731 
732 void c4iw_qp_rem_ref(struct ib_qp *qp)
733 {
734 	PDBG("%s ib_qp %p\n", __func__, qp);
735 	kref_put(&to_c4iw_qp(qp)->kref, _free_qp);
736 }
737 
738 static void add_to_fc_list(struct list_head *head, struct list_head *entry)
739 {
740 	if (list_empty(entry))
741 		list_add_tail(entry, head);
742 }
743 
744 static int ring_kernel_sq_db(struct c4iw_qp *qhp, u16 inc)
745 {
746 	unsigned long flags;
747 
748 	spin_lock_irqsave(&qhp->rhp->lock, flags);
749 	spin_lock(&qhp->lock);
750 	if (qhp->rhp->db_state == NORMAL)
751 		t4_ring_sq_db(&qhp->wq, inc, NULL);
752 	else {
753 		add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
754 		qhp->wq.sq.wq_pidx_inc += inc;
755 	}
756 	spin_unlock(&qhp->lock);
757 	spin_unlock_irqrestore(&qhp->rhp->lock, flags);
758 	return 0;
759 }
760 
761 static int ring_kernel_rq_db(struct c4iw_qp *qhp, u16 inc)
762 {
763 	unsigned long flags;
764 
765 	spin_lock_irqsave(&qhp->rhp->lock, flags);
766 	spin_lock(&qhp->lock);
767 	if (qhp->rhp->db_state == NORMAL)
768 		t4_ring_rq_db(&qhp->wq, inc, NULL);
769 	else {
770 		add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
771 		qhp->wq.rq.wq_pidx_inc += inc;
772 	}
773 	spin_unlock(&qhp->lock);
774 	spin_unlock_irqrestore(&qhp->rhp->lock, flags);
775 	return 0;
776 }
777 
778 int c4iw_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
779 		   struct ib_send_wr **bad_wr)
780 {
781 	int err = 0;
782 	u8 len16 = 0;
783 	enum fw_wr_opcodes fw_opcode = 0;
784 	enum fw_ri_wr_flags fw_flags;
785 	struct c4iw_qp *qhp;
786 	union t4_wr *wqe = NULL;
787 	u32 num_wrs;
788 	struct t4_swsqe *swsqe;
789 	unsigned long flag;
790 	u16 idx = 0;
791 
792 	qhp = to_c4iw_qp(ibqp);
793 	spin_lock_irqsave(&qhp->lock, flag);
794 	if (t4_wq_in_error(&qhp->wq)) {
795 		spin_unlock_irqrestore(&qhp->lock, flag);
796 		*bad_wr = wr;
797 		return -EINVAL;
798 	}
799 	num_wrs = t4_sq_avail(&qhp->wq);
800 	if (num_wrs == 0) {
801 		spin_unlock_irqrestore(&qhp->lock, flag);
802 		*bad_wr = wr;
803 		return -ENOMEM;
804 	}
805 	while (wr) {
806 		if (num_wrs == 0) {
807 			err = -ENOMEM;
808 			*bad_wr = wr;
809 			break;
810 		}
811 		wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
812 		      qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
813 
814 		fw_flags = 0;
815 		if (wr->send_flags & IB_SEND_SOLICITED)
816 			fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
817 		if (wr->send_flags & IB_SEND_SIGNALED || qhp->sq_sig_all)
818 			fw_flags |= FW_RI_COMPLETION_FLAG;
819 		swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
820 		switch (wr->opcode) {
821 		case IB_WR_SEND_WITH_INV:
822 		case IB_WR_SEND:
823 			if (wr->send_flags & IB_SEND_FENCE)
824 				fw_flags |= FW_RI_READ_FENCE_FLAG;
825 			fw_opcode = FW_RI_SEND_WR;
826 			if (wr->opcode == IB_WR_SEND)
827 				swsqe->opcode = FW_RI_SEND;
828 			else
829 				swsqe->opcode = FW_RI_SEND_WITH_INV;
830 			err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
831 			break;
832 		case IB_WR_RDMA_WRITE:
833 			fw_opcode = FW_RI_RDMA_WRITE_WR;
834 			swsqe->opcode = FW_RI_RDMA_WRITE;
835 			err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
836 			break;
837 		case IB_WR_RDMA_READ:
838 		case IB_WR_RDMA_READ_WITH_INV:
839 			fw_opcode = FW_RI_RDMA_READ_WR;
840 			swsqe->opcode = FW_RI_READ_REQ;
841 			if (wr->opcode == IB_WR_RDMA_READ_WITH_INV) {
842 				c4iw_invalidate_mr(qhp->rhp,
843 						   wr->sg_list[0].lkey);
844 				fw_flags = FW_RI_RDMA_READ_INVALIDATE;
845 			} else {
846 				fw_flags = 0;
847 			}
848 			err = build_rdma_read(wqe, wr, &len16);
849 			if (err)
850 				break;
851 			swsqe->read_len = wr->sg_list[0].length;
852 			if (!qhp->wq.sq.oldest_read)
853 				qhp->wq.sq.oldest_read = swsqe;
854 			break;
855 		case IB_WR_REG_MR: {
856 			struct c4iw_mr *mhp = to_c4iw_mr(reg_wr(wr)->mr);
857 
858 			swsqe->opcode = FW_RI_FAST_REGISTER;
859 			if (qhp->rhp->rdev.lldi.fr_nsmr_tpte_wr_support &&
860 			    !mhp->attr.state && mhp->mpl_len <= 2) {
861 				fw_opcode = FW_RI_FR_NSMR_TPTE_WR;
862 				build_tpte_memreg(&wqe->fr_tpte, reg_wr(wr),
863 						  mhp, &len16);
864 			} else {
865 				fw_opcode = FW_RI_FR_NSMR_WR;
866 				err = build_memreg(&qhp->wq.sq, wqe, reg_wr(wr),
867 				       mhp, &len16,
868 				       qhp->rhp->rdev.lldi.ulptx_memwrite_dsgl);
869 				if (err)
870 					break;
871 			}
872 			mhp->attr.state = 1;
873 			break;
874 		}
875 		case IB_WR_LOCAL_INV:
876 			if (wr->send_flags & IB_SEND_FENCE)
877 				fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
878 			fw_opcode = FW_RI_INV_LSTAG_WR;
879 			swsqe->opcode = FW_RI_LOCAL_INV;
880 			err = build_inv_stag(wqe, wr, &len16);
881 			c4iw_invalidate_mr(qhp->rhp, wr->ex.invalidate_rkey);
882 			break;
883 		default:
884 			PDBG("%s post of type=%d TBD!\n", __func__,
885 			     wr->opcode);
886 			err = -EINVAL;
887 		}
888 		if (err) {
889 			*bad_wr = wr;
890 			break;
891 		}
892 		swsqe->idx = qhp->wq.sq.pidx;
893 		swsqe->complete = 0;
894 		swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED) ||
895 				  qhp->sq_sig_all;
896 		swsqe->flushed = 0;
897 		swsqe->wr_id = wr->wr_id;
898 		if (c4iw_wr_log) {
899 			swsqe->sge_ts = cxgb4_read_sge_timestamp(
900 					qhp->rhp->rdev.lldi.ports[0]);
901 			getnstimeofday(&swsqe->host_ts);
902 		}
903 
904 		init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
905 
906 		PDBG("%s cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u\n",
907 		     __func__, (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
908 		     swsqe->opcode, swsqe->read_len);
909 		wr = wr->next;
910 		num_wrs--;
911 		t4_sq_produce(&qhp->wq, len16);
912 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
913 	}
914 	if (!qhp->rhp->rdev.status_page->db_off) {
915 		t4_ring_sq_db(&qhp->wq, idx, wqe);
916 		spin_unlock_irqrestore(&qhp->lock, flag);
917 	} else {
918 		spin_unlock_irqrestore(&qhp->lock, flag);
919 		ring_kernel_sq_db(qhp, idx);
920 	}
921 	return err;
922 }
923 
924 int c4iw_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
925 		      struct ib_recv_wr **bad_wr)
926 {
927 	int err = 0;
928 	struct c4iw_qp *qhp;
929 	union t4_recv_wr *wqe = NULL;
930 	u32 num_wrs;
931 	u8 len16 = 0;
932 	unsigned long flag;
933 	u16 idx = 0;
934 
935 	qhp = to_c4iw_qp(ibqp);
936 	spin_lock_irqsave(&qhp->lock, flag);
937 	if (t4_wq_in_error(&qhp->wq)) {
938 		spin_unlock_irqrestore(&qhp->lock, flag);
939 		*bad_wr = wr;
940 		return -EINVAL;
941 	}
942 	num_wrs = t4_rq_avail(&qhp->wq);
943 	if (num_wrs == 0) {
944 		spin_unlock_irqrestore(&qhp->lock, flag);
945 		*bad_wr = wr;
946 		return -ENOMEM;
947 	}
948 	while (wr) {
949 		if (wr->num_sge > T4_MAX_RECV_SGE) {
950 			err = -EINVAL;
951 			*bad_wr = wr;
952 			break;
953 		}
954 		wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
955 					   qhp->wq.rq.wq_pidx *
956 					   T4_EQ_ENTRY_SIZE);
957 		if (num_wrs)
958 			err = build_rdma_recv(qhp, wqe, wr, &len16);
959 		else
960 			err = -ENOMEM;
961 		if (err) {
962 			*bad_wr = wr;
963 			break;
964 		}
965 
966 		qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
967 		if (c4iw_wr_log) {
968 			qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].sge_ts =
969 				cxgb4_read_sge_timestamp(
970 						qhp->rhp->rdev.lldi.ports[0]);
971 			getnstimeofday(
972 				&qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].host_ts);
973 		}
974 
975 		wqe->recv.opcode = FW_RI_RECV_WR;
976 		wqe->recv.r1 = 0;
977 		wqe->recv.wrid = qhp->wq.rq.pidx;
978 		wqe->recv.r2[0] = 0;
979 		wqe->recv.r2[1] = 0;
980 		wqe->recv.r2[2] = 0;
981 		wqe->recv.len16 = len16;
982 		PDBG("%s cookie 0x%llx pidx %u\n", __func__,
983 		     (unsigned long long) wr->wr_id, qhp->wq.rq.pidx);
984 		t4_rq_produce(&qhp->wq, len16);
985 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
986 		wr = wr->next;
987 		num_wrs--;
988 	}
989 	if (!qhp->rhp->rdev.status_page->db_off) {
990 		t4_ring_rq_db(&qhp->wq, idx, wqe);
991 		spin_unlock_irqrestore(&qhp->lock, flag);
992 	} else {
993 		spin_unlock_irqrestore(&qhp->lock, flag);
994 		ring_kernel_rq_db(qhp, idx);
995 	}
996 	return err;
997 }
998 
999 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
1000 				    u8 *ecode)
1001 {
1002 	int status;
1003 	int tagged;
1004 	int opcode;
1005 	int rqtype;
1006 	int send_inv;
1007 
1008 	if (!err_cqe) {
1009 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
1010 		*ecode = 0;
1011 		return;
1012 	}
1013 
1014 	status = CQE_STATUS(err_cqe);
1015 	opcode = CQE_OPCODE(err_cqe);
1016 	rqtype = RQ_TYPE(err_cqe);
1017 	send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
1018 		   (opcode == FW_RI_SEND_WITH_SE_INV);
1019 	tagged = (opcode == FW_RI_RDMA_WRITE) ||
1020 		 (rqtype && (opcode == FW_RI_READ_RESP));
1021 
1022 	switch (status) {
1023 	case T4_ERR_STAG:
1024 		if (send_inv) {
1025 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1026 			*ecode = RDMAP_CANT_INV_STAG;
1027 		} else {
1028 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1029 			*ecode = RDMAP_INV_STAG;
1030 		}
1031 		break;
1032 	case T4_ERR_PDID:
1033 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1034 		if ((opcode == FW_RI_SEND_WITH_INV) ||
1035 		    (opcode == FW_RI_SEND_WITH_SE_INV))
1036 			*ecode = RDMAP_CANT_INV_STAG;
1037 		else
1038 			*ecode = RDMAP_STAG_NOT_ASSOC;
1039 		break;
1040 	case T4_ERR_QPID:
1041 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1042 		*ecode = RDMAP_STAG_NOT_ASSOC;
1043 		break;
1044 	case T4_ERR_ACCESS:
1045 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1046 		*ecode = RDMAP_ACC_VIOL;
1047 		break;
1048 	case T4_ERR_WRAP:
1049 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1050 		*ecode = RDMAP_TO_WRAP;
1051 		break;
1052 	case T4_ERR_BOUND:
1053 		if (tagged) {
1054 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1055 			*ecode = DDPT_BASE_BOUNDS;
1056 		} else {
1057 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1058 			*ecode = RDMAP_BASE_BOUNDS;
1059 		}
1060 		break;
1061 	case T4_ERR_INVALIDATE_SHARED_MR:
1062 	case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
1063 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1064 		*ecode = RDMAP_CANT_INV_STAG;
1065 		break;
1066 	case T4_ERR_ECC:
1067 	case T4_ERR_ECC_PSTAG:
1068 	case T4_ERR_INTERNAL_ERR:
1069 		*layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
1070 		*ecode = 0;
1071 		break;
1072 	case T4_ERR_OUT_OF_RQE:
1073 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1074 		*ecode = DDPU_INV_MSN_NOBUF;
1075 		break;
1076 	case T4_ERR_PBL_ADDR_BOUND:
1077 		*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1078 		*ecode = DDPT_BASE_BOUNDS;
1079 		break;
1080 	case T4_ERR_CRC:
1081 		*layer_type = LAYER_MPA|DDP_LLP;
1082 		*ecode = MPA_CRC_ERR;
1083 		break;
1084 	case T4_ERR_MARKER:
1085 		*layer_type = LAYER_MPA|DDP_LLP;
1086 		*ecode = MPA_MARKER_ERR;
1087 		break;
1088 	case T4_ERR_PDU_LEN_ERR:
1089 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1090 		*ecode = DDPU_MSG_TOOBIG;
1091 		break;
1092 	case T4_ERR_DDP_VERSION:
1093 		if (tagged) {
1094 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1095 			*ecode = DDPT_INV_VERS;
1096 		} else {
1097 			*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1098 			*ecode = DDPU_INV_VERS;
1099 		}
1100 		break;
1101 	case T4_ERR_RDMA_VERSION:
1102 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1103 		*ecode = RDMAP_INV_VERS;
1104 		break;
1105 	case T4_ERR_OPCODE:
1106 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1107 		*ecode = RDMAP_INV_OPCODE;
1108 		break;
1109 	case T4_ERR_DDP_QUEUE_NUM:
1110 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1111 		*ecode = DDPU_INV_QN;
1112 		break;
1113 	case T4_ERR_MSN:
1114 	case T4_ERR_MSN_GAP:
1115 	case T4_ERR_MSN_RANGE:
1116 	case T4_ERR_IRD_OVERFLOW:
1117 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1118 		*ecode = DDPU_INV_MSN_RANGE;
1119 		break;
1120 	case T4_ERR_TBIT:
1121 		*layer_type = LAYER_DDP|DDP_LOCAL_CATA;
1122 		*ecode = 0;
1123 		break;
1124 	case T4_ERR_MO:
1125 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1126 		*ecode = DDPU_INV_MO;
1127 		break;
1128 	default:
1129 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
1130 		*ecode = 0;
1131 		break;
1132 	}
1133 }
1134 
1135 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
1136 			   gfp_t gfp)
1137 {
1138 	struct fw_ri_wr *wqe;
1139 	struct sk_buff *skb;
1140 	struct terminate_message *term;
1141 
1142 	PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1143 	     qhp->ep->hwtid);
1144 
1145 	skb = skb_dequeue(&qhp->ep->com.ep_skb_list);
1146 	if (WARN_ON(!skb))
1147 		return;
1148 
1149 	set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1150 
1151 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1152 	memset(wqe, 0, sizeof *wqe);
1153 	wqe->op_compl = cpu_to_be32(FW_WR_OP_V(FW_RI_INIT_WR));
1154 	wqe->flowid_len16 = cpu_to_be32(
1155 		FW_WR_FLOWID_V(qhp->ep->hwtid) |
1156 		FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1157 
1158 	wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
1159 	wqe->u.terminate.immdlen = cpu_to_be32(sizeof *term);
1160 	term = (struct terminate_message *)wqe->u.terminate.termmsg;
1161 	if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
1162 		term->layer_etype = qhp->attr.layer_etype;
1163 		term->ecode = qhp->attr.ecode;
1164 	} else
1165 		build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
1166 	c4iw_ofld_send(&qhp->rhp->rdev, skb);
1167 }
1168 
1169 /*
1170  * Assumes qhp lock is held.
1171  */
1172 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
1173 		       struct c4iw_cq *schp)
1174 {
1175 	int count;
1176 	int rq_flushed, sq_flushed;
1177 	unsigned long flag;
1178 
1179 	PDBG("%s qhp %p rchp %p schp %p\n", __func__, qhp, rchp, schp);
1180 
1181 	/* locking hierarchy: cq lock first, then qp lock. */
1182 	spin_lock_irqsave(&rchp->lock, flag);
1183 	spin_lock(&qhp->lock);
1184 
1185 	if (qhp->wq.flushed) {
1186 		spin_unlock(&qhp->lock);
1187 		spin_unlock_irqrestore(&rchp->lock, flag);
1188 		return;
1189 	}
1190 	qhp->wq.flushed = 1;
1191 
1192 	c4iw_flush_hw_cq(rchp);
1193 	c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
1194 	rq_flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
1195 	spin_unlock(&qhp->lock);
1196 	spin_unlock_irqrestore(&rchp->lock, flag);
1197 
1198 	/* locking hierarchy: cq lock first, then qp lock. */
1199 	spin_lock_irqsave(&schp->lock, flag);
1200 	spin_lock(&qhp->lock);
1201 	if (schp != rchp)
1202 		c4iw_flush_hw_cq(schp);
1203 	sq_flushed = c4iw_flush_sq(qhp);
1204 	spin_unlock(&qhp->lock);
1205 	spin_unlock_irqrestore(&schp->lock, flag);
1206 
1207 	if (schp == rchp) {
1208 		if (t4_clear_cq_armed(&rchp->cq) &&
1209 		    (rq_flushed || sq_flushed)) {
1210 			spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1211 			(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1212 						   rchp->ibcq.cq_context);
1213 			spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1214 		}
1215 	} else {
1216 		if (t4_clear_cq_armed(&rchp->cq) && rq_flushed) {
1217 			spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1218 			(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1219 						   rchp->ibcq.cq_context);
1220 			spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1221 		}
1222 		if (t4_clear_cq_armed(&schp->cq) && sq_flushed) {
1223 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1224 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1225 						   schp->ibcq.cq_context);
1226 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1227 		}
1228 	}
1229 }
1230 
1231 static void flush_qp(struct c4iw_qp *qhp)
1232 {
1233 	struct c4iw_cq *rchp, *schp;
1234 	unsigned long flag;
1235 
1236 	rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
1237 	schp = to_c4iw_cq(qhp->ibqp.send_cq);
1238 
1239 	t4_set_wq_in_error(&qhp->wq);
1240 	if (qhp->ibqp.uobject) {
1241 		t4_set_cq_in_error(&rchp->cq);
1242 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1243 		(*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1244 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1245 		if (schp != rchp) {
1246 			t4_set_cq_in_error(&schp->cq);
1247 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1248 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1249 					schp->ibcq.cq_context);
1250 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1251 		}
1252 		return;
1253 	}
1254 	__flush_qp(qhp, rchp, schp);
1255 }
1256 
1257 static int rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1258 		     struct c4iw_ep *ep)
1259 {
1260 	struct fw_ri_wr *wqe;
1261 	int ret;
1262 	struct sk_buff *skb;
1263 
1264 	PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1265 	     ep->hwtid);
1266 
1267 	skb = skb_dequeue(&ep->com.ep_skb_list);
1268 	if (WARN_ON(!skb))
1269 		return -ENOMEM;
1270 
1271 	set_wr_txq(skb, CPL_PRIORITY_DATA, ep->txq_idx);
1272 
1273 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1274 	memset(wqe, 0, sizeof *wqe);
1275 	wqe->op_compl = cpu_to_be32(
1276 		FW_WR_OP_V(FW_RI_INIT_WR) |
1277 		FW_WR_COMPL_F);
1278 	wqe->flowid_len16 = cpu_to_be32(
1279 		FW_WR_FLOWID_V(ep->hwtid) |
1280 		FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1281 	wqe->cookie = (uintptr_t)&ep->com.wr_wait;
1282 
1283 	wqe->u.fini.type = FW_RI_TYPE_FINI;
1284 	ret = c4iw_ofld_send(&rhp->rdev, skb);
1285 	if (ret)
1286 		goto out;
1287 
1288 	ret = c4iw_wait_for_reply(&rhp->rdev, &ep->com.wr_wait, qhp->ep->hwtid,
1289 			     qhp->wq.sq.qid, __func__);
1290 out:
1291 	PDBG("%s ret %d\n", __func__, ret);
1292 	return ret;
1293 }
1294 
1295 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1296 {
1297 	PDBG("%s p2p_type = %d\n", __func__, p2p_type);
1298 	memset(&init->u, 0, sizeof init->u);
1299 	switch (p2p_type) {
1300 	case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1301 		init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1302 		init->u.write.stag_sink = cpu_to_be32(1);
1303 		init->u.write.to_sink = cpu_to_be64(1);
1304 		init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1305 		init->u.write.len16 = DIV_ROUND_UP(sizeof init->u.write +
1306 						   sizeof(struct fw_ri_immd),
1307 						   16);
1308 		break;
1309 	case FW_RI_INIT_P2PTYPE_READ_REQ:
1310 		init->u.write.opcode = FW_RI_RDMA_READ_WR;
1311 		init->u.read.stag_src = cpu_to_be32(1);
1312 		init->u.read.to_src_lo = cpu_to_be32(1);
1313 		init->u.read.stag_sink = cpu_to_be32(1);
1314 		init->u.read.to_sink_lo = cpu_to_be32(1);
1315 		init->u.read.len16 = DIV_ROUND_UP(sizeof init->u.read, 16);
1316 		break;
1317 	}
1318 }
1319 
1320 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1321 {
1322 	struct fw_ri_wr *wqe;
1323 	int ret;
1324 	struct sk_buff *skb;
1325 
1326 	PDBG("%s qhp %p qid 0x%x tid %u ird %u ord %u\n", __func__, qhp,
1327 	     qhp->wq.sq.qid, qhp->ep->hwtid, qhp->ep->ird, qhp->ep->ord);
1328 
1329 	skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1330 	if (!skb) {
1331 		ret = -ENOMEM;
1332 		goto out;
1333 	}
1334 	ret = alloc_ird(rhp, qhp->attr.max_ird);
1335 	if (ret) {
1336 		qhp->attr.max_ird = 0;
1337 		kfree_skb(skb);
1338 		goto out;
1339 	}
1340 	set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1341 
1342 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1343 	memset(wqe, 0, sizeof *wqe);
1344 	wqe->op_compl = cpu_to_be32(
1345 		FW_WR_OP_V(FW_RI_INIT_WR) |
1346 		FW_WR_COMPL_F);
1347 	wqe->flowid_len16 = cpu_to_be32(
1348 		FW_WR_FLOWID_V(qhp->ep->hwtid) |
1349 		FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1350 
1351 	wqe->cookie = (uintptr_t)&qhp->ep->com.wr_wait;
1352 
1353 	wqe->u.init.type = FW_RI_TYPE_INIT;
1354 	wqe->u.init.mpareqbit_p2ptype =
1355 		FW_RI_WR_MPAREQBIT_V(qhp->attr.mpa_attr.initiator) |
1356 		FW_RI_WR_P2PTYPE_V(qhp->attr.mpa_attr.p2p_type);
1357 	wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1358 	if (qhp->attr.mpa_attr.recv_marker_enabled)
1359 		wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1360 	if (qhp->attr.mpa_attr.xmit_marker_enabled)
1361 		wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1362 	if (qhp->attr.mpa_attr.crc_enabled)
1363 		wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1364 
1365 	wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1366 			    FW_RI_QP_RDMA_WRITE_ENABLE |
1367 			    FW_RI_QP_BIND_ENABLE;
1368 	if (!qhp->ibqp.uobject)
1369 		wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1370 				     FW_RI_QP_STAG0_ENABLE;
1371 	wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1372 	wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1373 	wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1374 	wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1375 	wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1376 	wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1377 	wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1378 	wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1379 	wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1380 	wqe->u.init.iss = cpu_to_be32(qhp->ep->snd_seq);
1381 	wqe->u.init.irs = cpu_to_be32(qhp->ep->rcv_seq);
1382 	wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1383 	wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1384 					 rhp->rdev.lldi.vr->rq.start);
1385 	if (qhp->attr.mpa_attr.initiator)
1386 		build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1387 
1388 	ret = c4iw_ofld_send(&rhp->rdev, skb);
1389 	if (ret)
1390 		goto err1;
1391 
1392 	ret = c4iw_wait_for_reply(&rhp->rdev, &qhp->ep->com.wr_wait,
1393 				  qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
1394 	if (!ret)
1395 		goto out;
1396 err1:
1397 	free_ird(rhp, qhp->attr.max_ird);
1398 out:
1399 	PDBG("%s ret %d\n", __func__, ret);
1400 	return ret;
1401 }
1402 
1403 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1404 		   enum c4iw_qp_attr_mask mask,
1405 		   struct c4iw_qp_attributes *attrs,
1406 		   int internal)
1407 {
1408 	int ret = 0;
1409 	struct c4iw_qp_attributes newattr = qhp->attr;
1410 	int disconnect = 0;
1411 	int terminate = 0;
1412 	int abort = 0;
1413 	int free = 0;
1414 	struct c4iw_ep *ep = NULL;
1415 
1416 	PDBG("%s qhp %p sqid 0x%x rqid 0x%x ep %p state %d -> %d\n", __func__,
1417 	     qhp, qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep, qhp->attr.state,
1418 	     (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1419 
1420 	mutex_lock(&qhp->mutex);
1421 
1422 	/* Process attr changes if in IDLE */
1423 	if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1424 		if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1425 			ret = -EIO;
1426 			goto out;
1427 		}
1428 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1429 			newattr.enable_rdma_read = attrs->enable_rdma_read;
1430 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1431 			newattr.enable_rdma_write = attrs->enable_rdma_write;
1432 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1433 			newattr.enable_bind = attrs->enable_bind;
1434 		if (mask & C4IW_QP_ATTR_MAX_ORD) {
1435 			if (attrs->max_ord > c4iw_max_read_depth) {
1436 				ret = -EINVAL;
1437 				goto out;
1438 			}
1439 			newattr.max_ord = attrs->max_ord;
1440 		}
1441 		if (mask & C4IW_QP_ATTR_MAX_IRD) {
1442 			if (attrs->max_ird > cur_max_read_depth(rhp)) {
1443 				ret = -EINVAL;
1444 				goto out;
1445 			}
1446 			newattr.max_ird = attrs->max_ird;
1447 		}
1448 		qhp->attr = newattr;
1449 	}
1450 
1451 	if (mask & C4IW_QP_ATTR_SQ_DB) {
1452 		ret = ring_kernel_sq_db(qhp, attrs->sq_db_inc);
1453 		goto out;
1454 	}
1455 	if (mask & C4IW_QP_ATTR_RQ_DB) {
1456 		ret = ring_kernel_rq_db(qhp, attrs->rq_db_inc);
1457 		goto out;
1458 	}
1459 
1460 	if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1461 		goto out;
1462 	if (qhp->attr.state == attrs->next_state)
1463 		goto out;
1464 
1465 	switch (qhp->attr.state) {
1466 	case C4IW_QP_STATE_IDLE:
1467 		switch (attrs->next_state) {
1468 		case C4IW_QP_STATE_RTS:
1469 			if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1470 				ret = -EINVAL;
1471 				goto out;
1472 			}
1473 			if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1474 				ret = -EINVAL;
1475 				goto out;
1476 			}
1477 			qhp->attr.mpa_attr = attrs->mpa_attr;
1478 			qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1479 			qhp->ep = qhp->attr.llp_stream_handle;
1480 			set_state(qhp, C4IW_QP_STATE_RTS);
1481 
1482 			/*
1483 			 * Ref the endpoint here and deref when we
1484 			 * disassociate the endpoint from the QP.  This
1485 			 * happens in CLOSING->IDLE transition or *->ERROR
1486 			 * transition.
1487 			 */
1488 			c4iw_get_ep(&qhp->ep->com);
1489 			ret = rdma_init(rhp, qhp);
1490 			if (ret)
1491 				goto err;
1492 			break;
1493 		case C4IW_QP_STATE_ERROR:
1494 			set_state(qhp, C4IW_QP_STATE_ERROR);
1495 			flush_qp(qhp);
1496 			break;
1497 		default:
1498 			ret = -EINVAL;
1499 			goto out;
1500 		}
1501 		break;
1502 	case C4IW_QP_STATE_RTS:
1503 		switch (attrs->next_state) {
1504 		case C4IW_QP_STATE_CLOSING:
1505 			BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
1506 			t4_set_wq_in_error(&qhp->wq);
1507 			set_state(qhp, C4IW_QP_STATE_CLOSING);
1508 			ep = qhp->ep;
1509 			if (!internal) {
1510 				abort = 0;
1511 				disconnect = 1;
1512 				c4iw_get_ep(&qhp->ep->com);
1513 			}
1514 			ret = rdma_fini(rhp, qhp, ep);
1515 			if (ret)
1516 				goto err;
1517 			break;
1518 		case C4IW_QP_STATE_TERMINATE:
1519 			t4_set_wq_in_error(&qhp->wq);
1520 			set_state(qhp, C4IW_QP_STATE_TERMINATE);
1521 			qhp->attr.layer_etype = attrs->layer_etype;
1522 			qhp->attr.ecode = attrs->ecode;
1523 			ep = qhp->ep;
1524 			if (!internal) {
1525 				c4iw_get_ep(&qhp->ep->com);
1526 				terminate = 1;
1527 				disconnect = 1;
1528 			} else {
1529 				terminate = qhp->attr.send_term;
1530 				ret = rdma_fini(rhp, qhp, ep);
1531 				if (ret)
1532 					goto err;
1533 			}
1534 			break;
1535 		case C4IW_QP_STATE_ERROR:
1536 			t4_set_wq_in_error(&qhp->wq);
1537 			set_state(qhp, C4IW_QP_STATE_ERROR);
1538 			if (!internal) {
1539 				abort = 1;
1540 				disconnect = 1;
1541 				ep = qhp->ep;
1542 				c4iw_get_ep(&qhp->ep->com);
1543 			}
1544 			goto err;
1545 			break;
1546 		default:
1547 			ret = -EINVAL;
1548 			goto out;
1549 		}
1550 		break;
1551 	case C4IW_QP_STATE_CLOSING:
1552 		if (!internal) {
1553 			ret = -EINVAL;
1554 			goto out;
1555 		}
1556 		switch (attrs->next_state) {
1557 		case C4IW_QP_STATE_IDLE:
1558 			flush_qp(qhp);
1559 			set_state(qhp, C4IW_QP_STATE_IDLE);
1560 			qhp->attr.llp_stream_handle = NULL;
1561 			c4iw_put_ep(&qhp->ep->com);
1562 			qhp->ep = NULL;
1563 			wake_up(&qhp->wait);
1564 			break;
1565 		case C4IW_QP_STATE_ERROR:
1566 			goto err;
1567 		default:
1568 			ret = -EINVAL;
1569 			goto err;
1570 		}
1571 		break;
1572 	case C4IW_QP_STATE_ERROR:
1573 		if (attrs->next_state != C4IW_QP_STATE_IDLE) {
1574 			ret = -EINVAL;
1575 			goto out;
1576 		}
1577 		if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) {
1578 			ret = -EINVAL;
1579 			goto out;
1580 		}
1581 		set_state(qhp, C4IW_QP_STATE_IDLE);
1582 		break;
1583 	case C4IW_QP_STATE_TERMINATE:
1584 		if (!internal) {
1585 			ret = -EINVAL;
1586 			goto out;
1587 		}
1588 		goto err;
1589 		break;
1590 	default:
1591 		printk(KERN_ERR "%s in a bad state %d\n",
1592 		       __func__, qhp->attr.state);
1593 		ret = -EINVAL;
1594 		goto err;
1595 		break;
1596 	}
1597 	goto out;
1598 err:
1599 	PDBG("%s disassociating ep %p qpid 0x%x\n", __func__, qhp->ep,
1600 	     qhp->wq.sq.qid);
1601 
1602 	/* disassociate the LLP connection */
1603 	qhp->attr.llp_stream_handle = NULL;
1604 	if (!ep)
1605 		ep = qhp->ep;
1606 	qhp->ep = NULL;
1607 	set_state(qhp, C4IW_QP_STATE_ERROR);
1608 	free = 1;
1609 	abort = 1;
1610 	BUG_ON(!ep);
1611 	flush_qp(qhp);
1612 	wake_up(&qhp->wait);
1613 out:
1614 	mutex_unlock(&qhp->mutex);
1615 
1616 	if (terminate)
1617 		post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
1618 
1619 	/*
1620 	 * If disconnect is 1, then we need to initiate a disconnect
1621 	 * on the EP.  This can be a normal close (RTS->CLOSING) or
1622 	 * an abnormal close (RTS/CLOSING->ERROR).
1623 	 */
1624 	if (disconnect) {
1625 		c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC :
1626 							 GFP_KERNEL);
1627 		c4iw_put_ep(&ep->com);
1628 	}
1629 
1630 	/*
1631 	 * If free is 1, then we've disassociated the EP from the QP
1632 	 * and we need to dereference the EP.
1633 	 */
1634 	if (free)
1635 		c4iw_put_ep(&ep->com);
1636 	PDBG("%s exit state %d\n", __func__, qhp->attr.state);
1637 	return ret;
1638 }
1639 
1640 int c4iw_destroy_qp(struct ib_qp *ib_qp)
1641 {
1642 	struct c4iw_dev *rhp;
1643 	struct c4iw_qp *qhp;
1644 	struct c4iw_qp_attributes attrs;
1645 	struct c4iw_ucontext *ucontext;
1646 
1647 	qhp = to_c4iw_qp(ib_qp);
1648 	rhp = qhp->rhp;
1649 
1650 	attrs.next_state = C4IW_QP_STATE_ERROR;
1651 	if (qhp->attr.state == C4IW_QP_STATE_TERMINATE)
1652 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
1653 	else
1654 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
1655 	wait_event(qhp->wait, !qhp->ep);
1656 
1657 	remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1658 
1659 	spin_lock_irq(&rhp->lock);
1660 	if (!list_empty(&qhp->db_fc_entry))
1661 		list_del_init(&qhp->db_fc_entry);
1662 	spin_unlock_irq(&rhp->lock);
1663 	free_ird(rhp, qhp->attr.max_ird);
1664 
1665 	ucontext = ib_qp->uobject ?
1666 		   to_c4iw_ucontext(ib_qp->uobject->context) : NULL;
1667 	destroy_qp(&rhp->rdev, &qhp->wq,
1668 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1669 
1670 	c4iw_qp_rem_ref(ib_qp);
1671 
1672 	PDBG("%s ib_qp %p qpid 0x%0x\n", __func__, ib_qp, qhp->wq.sq.qid);
1673 	return 0;
1674 }
1675 
1676 struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
1677 			     struct ib_udata *udata)
1678 {
1679 	struct c4iw_dev *rhp;
1680 	struct c4iw_qp *qhp;
1681 	struct c4iw_pd *php;
1682 	struct c4iw_cq *schp;
1683 	struct c4iw_cq *rchp;
1684 	struct c4iw_create_qp_resp uresp;
1685 	unsigned int sqsize, rqsize;
1686 	struct c4iw_ucontext *ucontext;
1687 	int ret;
1688 	struct c4iw_mm_entry *sq_key_mm, *rq_key_mm = NULL, *sq_db_key_mm;
1689 	struct c4iw_mm_entry *rq_db_key_mm = NULL, *ma_sync_key_mm = NULL;
1690 
1691 	PDBG("%s ib_pd %p\n", __func__, pd);
1692 
1693 	if (attrs->qp_type != IB_QPT_RC)
1694 		return ERR_PTR(-EINVAL);
1695 
1696 	php = to_c4iw_pd(pd);
1697 	rhp = php->rhp;
1698 	schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
1699 	rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
1700 	if (!schp || !rchp)
1701 		return ERR_PTR(-EINVAL);
1702 
1703 	if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
1704 		return ERR_PTR(-EINVAL);
1705 
1706 	if (attrs->cap.max_recv_wr > rhp->rdev.hw_queue.t4_max_rq_size)
1707 		return ERR_PTR(-E2BIG);
1708 	rqsize = attrs->cap.max_recv_wr + 1;
1709 	if (rqsize < 8)
1710 		rqsize = 8;
1711 
1712 	if (attrs->cap.max_send_wr > rhp->rdev.hw_queue.t4_max_sq_size)
1713 		return ERR_PTR(-E2BIG);
1714 	sqsize = attrs->cap.max_send_wr + 1;
1715 	if (sqsize < 8)
1716 		sqsize = 8;
1717 
1718 	ucontext = pd->uobject ? to_c4iw_ucontext(pd->uobject->context) : NULL;
1719 
1720 	qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
1721 	if (!qhp)
1722 		return ERR_PTR(-ENOMEM);
1723 	qhp->wq.sq.size = sqsize;
1724 	qhp->wq.sq.memsize =
1725 		(sqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1726 		sizeof(*qhp->wq.sq.queue) + 16 * sizeof(__be64);
1727 	qhp->wq.sq.flush_cidx = -1;
1728 	qhp->wq.rq.size = rqsize;
1729 	qhp->wq.rq.memsize =
1730 		(rqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1731 		sizeof(*qhp->wq.rq.queue);
1732 
1733 	if (ucontext) {
1734 		qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
1735 		qhp->wq.rq.memsize = roundup(qhp->wq.rq.memsize, PAGE_SIZE);
1736 	}
1737 
1738 	ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
1739 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1740 	if (ret)
1741 		goto err1;
1742 
1743 	attrs->cap.max_recv_wr = rqsize - 1;
1744 	attrs->cap.max_send_wr = sqsize - 1;
1745 	attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
1746 
1747 	qhp->rhp = rhp;
1748 	qhp->attr.pd = php->pdid;
1749 	qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
1750 	qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
1751 	qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
1752 	qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
1753 	qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
1754 	qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
1755 	qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
1756 	qhp->attr.state = C4IW_QP_STATE_IDLE;
1757 	qhp->attr.next_state = C4IW_QP_STATE_IDLE;
1758 	qhp->attr.enable_rdma_read = 1;
1759 	qhp->attr.enable_rdma_write = 1;
1760 	qhp->attr.enable_bind = 1;
1761 	qhp->attr.max_ord = 0;
1762 	qhp->attr.max_ird = 0;
1763 	qhp->sq_sig_all = attrs->sq_sig_type == IB_SIGNAL_ALL_WR;
1764 	spin_lock_init(&qhp->lock);
1765 	init_completion(&qhp->sq_drained);
1766 	init_completion(&qhp->rq_drained);
1767 	mutex_init(&qhp->mutex);
1768 	init_waitqueue_head(&qhp->wait);
1769 	kref_init(&qhp->kref);
1770 
1771 	ret = insert_handle(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid);
1772 	if (ret)
1773 		goto err2;
1774 
1775 	if (udata) {
1776 		sq_key_mm = kmalloc(sizeof(*sq_key_mm), GFP_KERNEL);
1777 		if (!sq_key_mm) {
1778 			ret = -ENOMEM;
1779 			goto err3;
1780 		}
1781 		rq_key_mm = kmalloc(sizeof(*rq_key_mm), GFP_KERNEL);
1782 		if (!rq_key_mm) {
1783 			ret = -ENOMEM;
1784 			goto err4;
1785 		}
1786 		sq_db_key_mm = kmalloc(sizeof(*sq_db_key_mm), GFP_KERNEL);
1787 		if (!sq_db_key_mm) {
1788 			ret = -ENOMEM;
1789 			goto err5;
1790 		}
1791 		rq_db_key_mm = kmalloc(sizeof(*rq_db_key_mm), GFP_KERNEL);
1792 		if (!rq_db_key_mm) {
1793 			ret = -ENOMEM;
1794 			goto err6;
1795 		}
1796 		if (t4_sq_onchip(&qhp->wq.sq)) {
1797 			ma_sync_key_mm = kmalloc(sizeof(*ma_sync_key_mm),
1798 						 GFP_KERNEL);
1799 			if (!ma_sync_key_mm) {
1800 				ret = -ENOMEM;
1801 				goto err7;
1802 			}
1803 			uresp.flags = C4IW_QPF_ONCHIP;
1804 		} else
1805 			uresp.flags = 0;
1806 		uresp.qid_mask = rhp->rdev.qpmask;
1807 		uresp.sqid = qhp->wq.sq.qid;
1808 		uresp.sq_size = qhp->wq.sq.size;
1809 		uresp.sq_memsize = qhp->wq.sq.memsize;
1810 		uresp.rqid = qhp->wq.rq.qid;
1811 		uresp.rq_size = qhp->wq.rq.size;
1812 		uresp.rq_memsize = qhp->wq.rq.memsize;
1813 		spin_lock(&ucontext->mmap_lock);
1814 		if (ma_sync_key_mm) {
1815 			uresp.ma_sync_key = ucontext->key;
1816 			ucontext->key += PAGE_SIZE;
1817 		} else {
1818 			uresp.ma_sync_key =  0;
1819 		}
1820 		uresp.sq_key = ucontext->key;
1821 		ucontext->key += PAGE_SIZE;
1822 		uresp.rq_key = ucontext->key;
1823 		ucontext->key += PAGE_SIZE;
1824 		uresp.sq_db_gts_key = ucontext->key;
1825 		ucontext->key += PAGE_SIZE;
1826 		uresp.rq_db_gts_key = ucontext->key;
1827 		ucontext->key += PAGE_SIZE;
1828 		spin_unlock(&ucontext->mmap_lock);
1829 		ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
1830 		if (ret)
1831 			goto err8;
1832 		sq_key_mm->key = uresp.sq_key;
1833 		sq_key_mm->addr = qhp->wq.sq.phys_addr;
1834 		sq_key_mm->len = PAGE_ALIGN(qhp->wq.sq.memsize);
1835 		insert_mmap(ucontext, sq_key_mm);
1836 		rq_key_mm->key = uresp.rq_key;
1837 		rq_key_mm->addr = virt_to_phys(qhp->wq.rq.queue);
1838 		rq_key_mm->len = PAGE_ALIGN(qhp->wq.rq.memsize);
1839 		insert_mmap(ucontext, rq_key_mm);
1840 		sq_db_key_mm->key = uresp.sq_db_gts_key;
1841 		sq_db_key_mm->addr = (u64)(unsigned long)qhp->wq.sq.bar2_pa;
1842 		sq_db_key_mm->len = PAGE_SIZE;
1843 		insert_mmap(ucontext, sq_db_key_mm);
1844 		rq_db_key_mm->key = uresp.rq_db_gts_key;
1845 		rq_db_key_mm->addr = (u64)(unsigned long)qhp->wq.rq.bar2_pa;
1846 		rq_db_key_mm->len = PAGE_SIZE;
1847 		insert_mmap(ucontext, rq_db_key_mm);
1848 		if (ma_sync_key_mm) {
1849 			ma_sync_key_mm->key = uresp.ma_sync_key;
1850 			ma_sync_key_mm->addr =
1851 				(pci_resource_start(rhp->rdev.lldi.pdev, 0) +
1852 				PCIE_MA_SYNC_A) & PAGE_MASK;
1853 			ma_sync_key_mm->len = PAGE_SIZE;
1854 			insert_mmap(ucontext, ma_sync_key_mm);
1855 		}
1856 	}
1857 	qhp->ibqp.qp_num = qhp->wq.sq.qid;
1858 	init_timer(&(qhp->timer));
1859 	INIT_LIST_HEAD(&qhp->db_fc_entry);
1860 	PDBG("%s sq id %u size %u memsize %zu num_entries %u "
1861 	     "rq id %u size %u memsize %zu num_entries %u\n", __func__,
1862 	     qhp->wq.sq.qid, qhp->wq.sq.size, qhp->wq.sq.memsize,
1863 	     attrs->cap.max_send_wr, qhp->wq.rq.qid, qhp->wq.rq.size,
1864 	     qhp->wq.rq.memsize, attrs->cap.max_recv_wr);
1865 	return &qhp->ibqp;
1866 err8:
1867 	kfree(ma_sync_key_mm);
1868 err7:
1869 	kfree(rq_db_key_mm);
1870 err6:
1871 	kfree(sq_db_key_mm);
1872 err5:
1873 	kfree(rq_key_mm);
1874 err4:
1875 	kfree(sq_key_mm);
1876 err3:
1877 	remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1878 err2:
1879 	destroy_qp(&rhp->rdev, &qhp->wq,
1880 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1881 err1:
1882 	kfree(qhp);
1883 	return ERR_PTR(ret);
1884 }
1885 
1886 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1887 		      int attr_mask, struct ib_udata *udata)
1888 {
1889 	struct c4iw_dev *rhp;
1890 	struct c4iw_qp *qhp;
1891 	enum c4iw_qp_attr_mask mask = 0;
1892 	struct c4iw_qp_attributes attrs;
1893 
1894 	PDBG("%s ib_qp %p\n", __func__, ibqp);
1895 
1896 	/* iwarp does not support the RTR state */
1897 	if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
1898 		attr_mask &= ~IB_QP_STATE;
1899 
1900 	/* Make sure we still have something left to do */
1901 	if (!attr_mask)
1902 		return 0;
1903 
1904 	memset(&attrs, 0, sizeof attrs);
1905 	qhp = to_c4iw_qp(ibqp);
1906 	rhp = qhp->rhp;
1907 
1908 	attrs.next_state = c4iw_convert_state(attr->qp_state);
1909 	attrs.enable_rdma_read = (attr->qp_access_flags &
1910 			       IB_ACCESS_REMOTE_READ) ?  1 : 0;
1911 	attrs.enable_rdma_write = (attr->qp_access_flags &
1912 				IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
1913 	attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
1914 
1915 
1916 	mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
1917 	mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
1918 			(C4IW_QP_ATTR_ENABLE_RDMA_READ |
1919 			 C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
1920 			 C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
1921 
1922 	/*
1923 	 * Use SQ_PSN and RQ_PSN to pass in IDX_INC values for
1924 	 * ringing the queue db when we're in DB_FULL mode.
1925 	 * Only allow this on T4 devices.
1926 	 */
1927 	attrs.sq_db_inc = attr->sq_psn;
1928 	attrs.rq_db_inc = attr->rq_psn;
1929 	mask |= (attr_mask & IB_QP_SQ_PSN) ? C4IW_QP_ATTR_SQ_DB : 0;
1930 	mask |= (attr_mask & IB_QP_RQ_PSN) ? C4IW_QP_ATTR_RQ_DB : 0;
1931 	if (!is_t4(to_c4iw_qp(ibqp)->rhp->rdev.lldi.adapter_type) &&
1932 	    (mask & (C4IW_QP_ATTR_SQ_DB|C4IW_QP_ATTR_RQ_DB)))
1933 		return -EINVAL;
1934 
1935 	return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
1936 }
1937 
1938 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
1939 {
1940 	PDBG("%s ib_dev %p qpn 0x%x\n", __func__, dev, qpn);
1941 	return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
1942 }
1943 
1944 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1945 		     int attr_mask, struct ib_qp_init_attr *init_attr)
1946 {
1947 	struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
1948 
1949 	memset(attr, 0, sizeof *attr);
1950 	memset(init_attr, 0, sizeof *init_attr);
1951 	attr->qp_state = to_ib_qp_state(qhp->attr.state);
1952 	init_attr->cap.max_send_wr = qhp->attr.sq_num_entries;
1953 	init_attr->cap.max_recv_wr = qhp->attr.rq_num_entries;
1954 	init_attr->cap.max_send_sge = qhp->attr.sq_max_sges;
1955 	init_attr->cap.max_recv_sge = qhp->attr.sq_max_sges;
1956 	init_attr->cap.max_inline_data = T4_MAX_SEND_INLINE;
1957 	init_attr->sq_sig_type = qhp->sq_sig_all ? IB_SIGNAL_ALL_WR : 0;
1958 	return 0;
1959 }
1960 
1961 static void move_qp_to_err(struct c4iw_qp *qp)
1962 {
1963 	struct c4iw_qp_attributes attrs = { .next_state = C4IW_QP_STATE_ERROR };
1964 
1965 	(void)c4iw_modify_qp(qp->rhp, qp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
1966 }
1967 
1968 void c4iw_drain_sq(struct ib_qp *ibqp)
1969 {
1970 	struct c4iw_qp *qp = to_c4iw_qp(ibqp);
1971 	unsigned long flag;
1972 	bool need_to_wait;
1973 
1974 	move_qp_to_err(qp);
1975 	spin_lock_irqsave(&qp->lock, flag);
1976 	need_to_wait = !t4_sq_empty(&qp->wq);
1977 	spin_unlock_irqrestore(&qp->lock, flag);
1978 
1979 	if (need_to_wait)
1980 		wait_for_completion(&qp->sq_drained);
1981 }
1982 
1983 void c4iw_drain_rq(struct ib_qp *ibqp)
1984 {
1985 	struct c4iw_qp *qp = to_c4iw_qp(ibqp);
1986 	unsigned long flag;
1987 	bool need_to_wait;
1988 
1989 	move_qp_to_err(qp);
1990 	spin_lock_irqsave(&qp->lock, flag);
1991 	need_to_wait = !t4_rq_empty(&qp->wq);
1992 	spin_unlock_irqrestore(&qp->lock, flag);
1993 
1994 	if (need_to_wait)
1995 		wait_for_completion(&qp->rq_drained);
1996 }
1997