xref: /linux/drivers/infiniband/hw/cxgb4/qp.c (revision ee6825c80e870fff1a370c718ec77022ade0889b)
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 	return rdev->bar2_kva + bar2_qoffset;
189 }
190 
191 static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
192 		     struct t4_cq *rcq, struct t4_cq *scq,
193 		     struct c4iw_dev_ucontext *uctx)
194 {
195 	int user = (uctx != &rdev->uctx);
196 	struct fw_ri_res_wr *res_wr;
197 	struct fw_ri_res *res;
198 	int wr_len;
199 	struct c4iw_wr_wait wr_wait;
200 	struct sk_buff *skb;
201 	int ret = 0;
202 	int eqsize;
203 
204 	wq->sq.qid = c4iw_get_qpid(rdev, uctx);
205 	if (!wq->sq.qid)
206 		return -ENOMEM;
207 
208 	wq->rq.qid = c4iw_get_qpid(rdev, uctx);
209 	if (!wq->rq.qid) {
210 		ret = -ENOMEM;
211 		goto free_sq_qid;
212 	}
213 
214 	if (!user) {
215 		wq->sq.sw_sq = kzalloc(wq->sq.size * sizeof *wq->sq.sw_sq,
216 				 GFP_KERNEL);
217 		if (!wq->sq.sw_sq) {
218 			ret = -ENOMEM;
219 			goto free_rq_qid;
220 		}
221 
222 		wq->rq.sw_rq = kzalloc(wq->rq.size * sizeof *wq->rq.sw_rq,
223 				 GFP_KERNEL);
224 		if (!wq->rq.sw_rq) {
225 			ret = -ENOMEM;
226 			goto free_sw_sq;
227 		}
228 	}
229 
230 	/*
231 	 * RQT must be a power of 2 and at least 16 deep.
232 	 */
233 	wq->rq.rqt_size = roundup_pow_of_two(max_t(u16, wq->rq.size, 16));
234 	wq->rq.rqt_hwaddr = c4iw_rqtpool_alloc(rdev, wq->rq.rqt_size);
235 	if (!wq->rq.rqt_hwaddr) {
236 		ret = -ENOMEM;
237 		goto free_sw_rq;
238 	}
239 
240 	ret = alloc_sq(rdev, &wq->sq, user);
241 	if (ret)
242 		goto free_hwaddr;
243 	memset(wq->sq.queue, 0, wq->sq.memsize);
244 	dma_unmap_addr_set(&wq->sq, mapping, wq->sq.dma_addr);
245 
246 	wq->rq.queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev),
247 					  wq->rq.memsize, &(wq->rq.dma_addr),
248 					  GFP_KERNEL);
249 	if (!wq->rq.queue) {
250 		ret = -ENOMEM;
251 		goto free_sq;
252 	}
253 	PDBG("%s sq base va 0x%p pa 0x%llx rq base va 0x%p pa 0x%llx\n",
254 		__func__, wq->sq.queue,
255 		(unsigned long long)virt_to_phys(wq->sq.queue),
256 		wq->rq.queue,
257 		(unsigned long long)virt_to_phys(wq->rq.queue));
258 	memset(wq->rq.queue, 0, wq->rq.memsize);
259 	dma_unmap_addr_set(&wq->rq, mapping, wq->rq.dma_addr);
260 
261 	wq->db = rdev->lldi.db_reg;
262 
263 	wq->sq.bar2_va = c4iw_bar2_addrs(rdev, wq->sq.qid, T4_BAR2_QTYPE_EGRESS,
264 					 &wq->sq.bar2_qid,
265 					 user ? &wq->sq.bar2_pa : NULL);
266 	wq->rq.bar2_va = c4iw_bar2_addrs(rdev, wq->rq.qid, T4_BAR2_QTYPE_EGRESS,
267 					 &wq->rq.bar2_qid,
268 					 user ? &wq->rq.bar2_pa : NULL);
269 
270 	/*
271 	 * User mode must have bar2 access.
272 	 */
273 	if (user && (!wq->sq.bar2_va || !wq->rq.bar2_va)) {
274 		pr_warn(MOD "%s: sqid %u or rqid %u not in BAR2 range.\n",
275 			pci_name(rdev->lldi.pdev), wq->sq.qid, wq->rq.qid);
276 		goto free_dma;
277 	}
278 
279 	wq->rdev = rdev;
280 	wq->rq.msn = 1;
281 
282 	/* build fw_ri_res_wr */
283 	wr_len = sizeof *res_wr + 2 * sizeof *res;
284 
285 	skb = alloc_skb(wr_len, GFP_KERNEL);
286 	if (!skb) {
287 		ret = -ENOMEM;
288 		goto free_dma;
289 	}
290 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
291 
292 	res_wr = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
293 	memset(res_wr, 0, wr_len);
294 	res_wr->op_nres = cpu_to_be32(
295 			FW_WR_OP_V(FW_RI_RES_WR) |
296 			FW_RI_RES_WR_NRES_V(2) |
297 			FW_WR_COMPL_F);
298 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
299 	res_wr->cookie = (uintptr_t)&wr_wait;
300 	res = res_wr->res;
301 	res->u.sqrq.restype = FW_RI_RES_TYPE_SQ;
302 	res->u.sqrq.op = FW_RI_RES_OP_WRITE;
303 
304 	/*
305 	 * eqsize is the number of 64B entries plus the status page size.
306 	 */
307 	eqsize = wq->sq.size * T4_SQ_NUM_SLOTS +
308 		rdev->hw_queue.t4_eq_status_entries;
309 
310 	res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
311 		FW_RI_RES_WR_HOSTFCMODE_V(0) |	/* no host cidx updates */
312 		FW_RI_RES_WR_CPRIO_V(0) |	/* don't keep in chip cache */
313 		FW_RI_RES_WR_PCIECHN_V(0) |	/* set by uP at ri_init time */
314 		(t4_sq_onchip(&wq->sq) ? FW_RI_RES_WR_ONCHIP_F : 0) |
315 		FW_RI_RES_WR_IQID_V(scq->cqid));
316 	res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
317 		FW_RI_RES_WR_DCAEN_V(0) |
318 		FW_RI_RES_WR_DCACPU_V(0) |
319 		FW_RI_RES_WR_FBMIN_V(2) |
320 		FW_RI_RES_WR_FBMAX_V(2) |
321 		FW_RI_RES_WR_CIDXFTHRESHO_V(0) |
322 		FW_RI_RES_WR_CIDXFTHRESH_V(0) |
323 		FW_RI_RES_WR_EQSIZE_V(eqsize));
324 	res->u.sqrq.eqid = cpu_to_be32(wq->sq.qid);
325 	res->u.sqrq.eqaddr = cpu_to_be64(wq->sq.dma_addr);
326 	res++;
327 	res->u.sqrq.restype = FW_RI_RES_TYPE_RQ;
328 	res->u.sqrq.op = FW_RI_RES_OP_WRITE;
329 
330 	/*
331 	 * eqsize is the number of 64B entries plus the status page size.
332 	 */
333 	eqsize = wq->rq.size * T4_RQ_NUM_SLOTS +
334 		rdev->hw_queue.t4_eq_status_entries;
335 	res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
336 		FW_RI_RES_WR_HOSTFCMODE_V(0) |	/* no host cidx updates */
337 		FW_RI_RES_WR_CPRIO_V(0) |	/* don't keep in chip cache */
338 		FW_RI_RES_WR_PCIECHN_V(0) |	/* set by uP at ri_init time */
339 		FW_RI_RES_WR_IQID_V(rcq->cqid));
340 	res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
341 		FW_RI_RES_WR_DCAEN_V(0) |
342 		FW_RI_RES_WR_DCACPU_V(0) |
343 		FW_RI_RES_WR_FBMIN_V(2) |
344 		FW_RI_RES_WR_FBMAX_V(2) |
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->rq.qid);
349 	res->u.sqrq.eqaddr = cpu_to_be64(wq->rq.dma_addr);
350 
351 	c4iw_init_wr_wait(&wr_wait);
352 
353 	ret = c4iw_ofld_send(rdev, skb);
354 	if (ret)
355 		goto free_dma;
356 	ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, wq->sq.qid, __func__);
357 	if (ret)
358 		goto free_dma;
359 
360 	PDBG("%s sqid 0x%x rqid 0x%x kdb 0x%p sq_bar2_addr %p rq_bar2_addr %p\n",
361 	     __func__, wq->sq.qid, wq->rq.qid, wq->db,
362 	     wq->sq.bar2_va, wq->rq.bar2_va);
363 
364 	return 0;
365 free_dma:
366 	dma_free_coherent(&(rdev->lldi.pdev->dev),
367 			  wq->rq.memsize, wq->rq.queue,
368 			  dma_unmap_addr(&wq->rq, mapping));
369 free_sq:
370 	dealloc_sq(rdev, &wq->sq);
371 free_hwaddr:
372 	c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
373 free_sw_rq:
374 	kfree(wq->rq.sw_rq);
375 free_sw_sq:
376 	kfree(wq->sq.sw_sq);
377 free_rq_qid:
378 	c4iw_put_qpid(rdev, wq->rq.qid, uctx);
379 free_sq_qid:
380 	c4iw_put_qpid(rdev, wq->sq.qid, uctx);
381 	return ret;
382 }
383 
384 static int build_immd(struct t4_sq *sq, struct fw_ri_immd *immdp,
385 		      struct ib_send_wr *wr, int max, u32 *plenp)
386 {
387 	u8 *dstp, *srcp;
388 	u32 plen = 0;
389 	int i;
390 	int rem, len;
391 
392 	dstp = (u8 *)immdp->data;
393 	for (i = 0; i < wr->num_sge; i++) {
394 		if ((plen + wr->sg_list[i].length) > max)
395 			return -EMSGSIZE;
396 		srcp = (u8 *)(unsigned long)wr->sg_list[i].addr;
397 		plen += wr->sg_list[i].length;
398 		rem = wr->sg_list[i].length;
399 		while (rem) {
400 			if (dstp == (u8 *)&sq->queue[sq->size])
401 				dstp = (u8 *)sq->queue;
402 			if (rem <= (u8 *)&sq->queue[sq->size] - dstp)
403 				len = rem;
404 			else
405 				len = (u8 *)&sq->queue[sq->size] - dstp;
406 			memcpy(dstp, srcp, len);
407 			dstp += len;
408 			srcp += len;
409 			rem -= len;
410 		}
411 	}
412 	len = roundup(plen + sizeof *immdp, 16) - (plen + sizeof *immdp);
413 	if (len)
414 		memset(dstp, 0, len);
415 	immdp->op = FW_RI_DATA_IMMD;
416 	immdp->r1 = 0;
417 	immdp->r2 = 0;
418 	immdp->immdlen = cpu_to_be32(plen);
419 	*plenp = plen;
420 	return 0;
421 }
422 
423 static int build_isgl(__be64 *queue_start, __be64 *queue_end,
424 		      struct fw_ri_isgl *isglp, struct ib_sge *sg_list,
425 		      int num_sge, u32 *plenp)
426 
427 {
428 	int i;
429 	u32 plen = 0;
430 	__be64 *flitp = (__be64 *)isglp->sge;
431 
432 	for (i = 0; i < num_sge; i++) {
433 		if ((plen + sg_list[i].length) < plen)
434 			return -EMSGSIZE;
435 		plen += sg_list[i].length;
436 		*flitp = cpu_to_be64(((u64)sg_list[i].lkey << 32) |
437 				     sg_list[i].length);
438 		if (++flitp == queue_end)
439 			flitp = queue_start;
440 		*flitp = cpu_to_be64(sg_list[i].addr);
441 		if (++flitp == queue_end)
442 			flitp = queue_start;
443 	}
444 	*flitp = (__force __be64)0;
445 	isglp->op = FW_RI_DATA_ISGL;
446 	isglp->r1 = 0;
447 	isglp->nsge = cpu_to_be16(num_sge);
448 	isglp->r2 = 0;
449 	if (plenp)
450 		*plenp = plen;
451 	return 0;
452 }
453 
454 static int build_rdma_send(struct t4_sq *sq, union t4_wr *wqe,
455 			   struct ib_send_wr *wr, u8 *len16)
456 {
457 	u32 plen;
458 	int size;
459 	int ret;
460 
461 	if (wr->num_sge > T4_MAX_SEND_SGE)
462 		return -EINVAL;
463 	switch (wr->opcode) {
464 	case IB_WR_SEND:
465 		if (wr->send_flags & IB_SEND_SOLICITED)
466 			wqe->send.sendop_pkd = cpu_to_be32(
467 				FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_SE));
468 		else
469 			wqe->send.sendop_pkd = cpu_to_be32(
470 				FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND));
471 		wqe->send.stag_inv = 0;
472 		break;
473 	case IB_WR_SEND_WITH_INV:
474 		if (wr->send_flags & IB_SEND_SOLICITED)
475 			wqe->send.sendop_pkd = cpu_to_be32(
476 				FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_SE_INV));
477 		else
478 			wqe->send.sendop_pkd = cpu_to_be32(
479 				FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_INV));
480 		wqe->send.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
481 		break;
482 
483 	default:
484 		return -EINVAL;
485 	}
486 	wqe->send.r3 = 0;
487 	wqe->send.r4 = 0;
488 
489 	plen = 0;
490 	if (wr->num_sge) {
491 		if (wr->send_flags & IB_SEND_INLINE) {
492 			ret = build_immd(sq, wqe->send.u.immd_src, wr,
493 					 T4_MAX_SEND_INLINE, &plen);
494 			if (ret)
495 				return ret;
496 			size = sizeof wqe->send + sizeof(struct fw_ri_immd) +
497 			       plen;
498 		} else {
499 			ret = build_isgl((__be64 *)sq->queue,
500 					 (__be64 *)&sq->queue[sq->size],
501 					 wqe->send.u.isgl_src,
502 					 wr->sg_list, wr->num_sge, &plen);
503 			if (ret)
504 				return ret;
505 			size = sizeof wqe->send + sizeof(struct fw_ri_isgl) +
506 			       wr->num_sge * sizeof(struct fw_ri_sge);
507 		}
508 	} else {
509 		wqe->send.u.immd_src[0].op = FW_RI_DATA_IMMD;
510 		wqe->send.u.immd_src[0].r1 = 0;
511 		wqe->send.u.immd_src[0].r2 = 0;
512 		wqe->send.u.immd_src[0].immdlen = 0;
513 		size = sizeof wqe->send + sizeof(struct fw_ri_immd);
514 		plen = 0;
515 	}
516 	*len16 = DIV_ROUND_UP(size, 16);
517 	wqe->send.plen = cpu_to_be32(plen);
518 	return 0;
519 }
520 
521 static int build_rdma_write(struct t4_sq *sq, union t4_wr *wqe,
522 			    struct ib_send_wr *wr, u8 *len16)
523 {
524 	u32 plen;
525 	int size;
526 	int ret;
527 
528 	if (wr->num_sge > T4_MAX_SEND_SGE)
529 		return -EINVAL;
530 	wqe->write.r2 = 0;
531 	wqe->write.stag_sink = cpu_to_be32(rdma_wr(wr)->rkey);
532 	wqe->write.to_sink = cpu_to_be64(rdma_wr(wr)->remote_addr);
533 	if (wr->num_sge) {
534 		if (wr->send_flags & IB_SEND_INLINE) {
535 			ret = build_immd(sq, wqe->write.u.immd_src, wr,
536 					 T4_MAX_WRITE_INLINE, &plen);
537 			if (ret)
538 				return ret;
539 			size = sizeof wqe->write + sizeof(struct fw_ri_immd) +
540 			       plen;
541 		} else {
542 			ret = build_isgl((__be64 *)sq->queue,
543 					 (__be64 *)&sq->queue[sq->size],
544 					 wqe->write.u.isgl_src,
545 					 wr->sg_list, wr->num_sge, &plen);
546 			if (ret)
547 				return ret;
548 			size = sizeof wqe->write + sizeof(struct fw_ri_isgl) +
549 			       wr->num_sge * sizeof(struct fw_ri_sge);
550 		}
551 	} else {
552 		wqe->write.u.immd_src[0].op = FW_RI_DATA_IMMD;
553 		wqe->write.u.immd_src[0].r1 = 0;
554 		wqe->write.u.immd_src[0].r2 = 0;
555 		wqe->write.u.immd_src[0].immdlen = 0;
556 		size = sizeof wqe->write + sizeof(struct fw_ri_immd);
557 		plen = 0;
558 	}
559 	*len16 = DIV_ROUND_UP(size, 16);
560 	wqe->write.plen = cpu_to_be32(plen);
561 	return 0;
562 }
563 
564 static int build_rdma_read(union t4_wr *wqe, struct ib_send_wr *wr, u8 *len16)
565 {
566 	if (wr->num_sge > 1)
567 		return -EINVAL;
568 	if (wr->num_sge) {
569 		wqe->read.stag_src = cpu_to_be32(rdma_wr(wr)->rkey);
570 		wqe->read.to_src_hi = cpu_to_be32((u32)(rdma_wr(wr)->remote_addr
571 							>> 32));
572 		wqe->read.to_src_lo = cpu_to_be32((u32)rdma_wr(wr)->remote_addr);
573 		wqe->read.stag_sink = cpu_to_be32(wr->sg_list[0].lkey);
574 		wqe->read.plen = cpu_to_be32(wr->sg_list[0].length);
575 		wqe->read.to_sink_hi = cpu_to_be32((u32)(wr->sg_list[0].addr
576 							 >> 32));
577 		wqe->read.to_sink_lo = cpu_to_be32((u32)(wr->sg_list[0].addr));
578 	} else {
579 		wqe->read.stag_src = cpu_to_be32(2);
580 		wqe->read.to_src_hi = 0;
581 		wqe->read.to_src_lo = 0;
582 		wqe->read.stag_sink = cpu_to_be32(2);
583 		wqe->read.plen = 0;
584 		wqe->read.to_sink_hi = 0;
585 		wqe->read.to_sink_lo = 0;
586 	}
587 	wqe->read.r2 = 0;
588 	wqe->read.r5 = 0;
589 	*len16 = DIV_ROUND_UP(sizeof wqe->read, 16);
590 	return 0;
591 }
592 
593 static int build_rdma_recv(struct c4iw_qp *qhp, union t4_recv_wr *wqe,
594 			   struct ib_recv_wr *wr, u8 *len16)
595 {
596 	int ret;
597 
598 	ret = build_isgl((__be64 *)qhp->wq.rq.queue,
599 			 (__be64 *)&qhp->wq.rq.queue[qhp->wq.rq.size],
600 			 &wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL);
601 	if (ret)
602 		return ret;
603 	*len16 = DIV_ROUND_UP(sizeof wqe->recv +
604 			      wr->num_sge * sizeof(struct fw_ri_sge), 16);
605 	return 0;
606 }
607 
608 static int build_memreg(struct t4_sq *sq, union t4_wr *wqe,
609 			struct ib_reg_wr *wr, u8 *len16, bool dsgl_supported)
610 {
611 	struct c4iw_mr *mhp = to_c4iw_mr(wr->mr);
612 	struct fw_ri_immd *imdp;
613 	__be64 *p;
614 	int i;
615 	int pbllen = roundup(mhp->mpl_len * sizeof(u64), 32);
616 	int rem;
617 
618 	if (mhp->mpl_len > t4_max_fr_depth(dsgl_supported && use_dsgl))
619 		return -EINVAL;
620 
621 	wqe->fr.qpbinde_to_dcacpu = 0;
622 	wqe->fr.pgsz_shift = ilog2(wr->mr->page_size) - 12;
623 	wqe->fr.addr_type = FW_RI_VA_BASED_TO;
624 	wqe->fr.mem_perms = c4iw_ib_to_tpt_access(wr->access);
625 	wqe->fr.len_hi = 0;
626 	wqe->fr.len_lo = cpu_to_be32(mhp->ibmr.length);
627 	wqe->fr.stag = cpu_to_be32(wr->key);
628 	wqe->fr.va_hi = cpu_to_be32(mhp->ibmr.iova >> 32);
629 	wqe->fr.va_lo_fbo = cpu_to_be32(mhp->ibmr.iova &
630 					0xffffffff);
631 
632 	if (dsgl_supported && use_dsgl && (pbllen > max_fr_immd)) {
633 		struct fw_ri_dsgl *sglp;
634 
635 		for (i = 0; i < mhp->mpl_len; i++)
636 			mhp->mpl[i] = (__force u64)cpu_to_be64((u64)mhp->mpl[i]);
637 
638 		sglp = (struct fw_ri_dsgl *)(&wqe->fr + 1);
639 		sglp->op = FW_RI_DATA_DSGL;
640 		sglp->r1 = 0;
641 		sglp->nsge = cpu_to_be16(1);
642 		sglp->addr0 = cpu_to_be64(mhp->mpl_addr);
643 		sglp->len0 = cpu_to_be32(pbllen);
644 
645 		*len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*sglp), 16);
646 	} else {
647 		imdp = (struct fw_ri_immd *)(&wqe->fr + 1);
648 		imdp->op = FW_RI_DATA_IMMD;
649 		imdp->r1 = 0;
650 		imdp->r2 = 0;
651 		imdp->immdlen = cpu_to_be32(pbllen);
652 		p = (__be64 *)(imdp + 1);
653 		rem = pbllen;
654 		for (i = 0; i < mhp->mpl_len; i++) {
655 			*p = cpu_to_be64((u64)mhp->mpl[i]);
656 			rem -= sizeof(*p);
657 			if (++p == (__be64 *)&sq->queue[sq->size])
658 				p = (__be64 *)sq->queue;
659 		}
660 		BUG_ON(rem < 0);
661 		while (rem) {
662 			*p = 0;
663 			rem -= sizeof(*p);
664 			if (++p == (__be64 *)&sq->queue[sq->size])
665 				p = (__be64 *)sq->queue;
666 		}
667 		*len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*imdp)
668 				      + pbllen, 16);
669 	}
670 	return 0;
671 }
672 
673 static int build_inv_stag(union t4_wr *wqe, struct ib_send_wr *wr,
674 			  u8 *len16)
675 {
676 	wqe->inv.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
677 	wqe->inv.r2 = 0;
678 	*len16 = DIV_ROUND_UP(sizeof wqe->inv, 16);
679 	return 0;
680 }
681 
682 void c4iw_qp_add_ref(struct ib_qp *qp)
683 {
684 	PDBG("%s ib_qp %p\n", __func__, qp);
685 	atomic_inc(&(to_c4iw_qp(qp)->refcnt));
686 }
687 
688 void c4iw_qp_rem_ref(struct ib_qp *qp)
689 {
690 	PDBG("%s ib_qp %p\n", __func__, qp);
691 	if (atomic_dec_and_test(&(to_c4iw_qp(qp)->refcnt)))
692 		wake_up(&(to_c4iw_qp(qp)->wait));
693 }
694 
695 static void add_to_fc_list(struct list_head *head, struct list_head *entry)
696 {
697 	if (list_empty(entry))
698 		list_add_tail(entry, head);
699 }
700 
701 static int ring_kernel_sq_db(struct c4iw_qp *qhp, u16 inc)
702 {
703 	unsigned long flags;
704 
705 	spin_lock_irqsave(&qhp->rhp->lock, flags);
706 	spin_lock(&qhp->lock);
707 	if (qhp->rhp->db_state == NORMAL)
708 		t4_ring_sq_db(&qhp->wq, inc, NULL);
709 	else {
710 		add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
711 		qhp->wq.sq.wq_pidx_inc += inc;
712 	}
713 	spin_unlock(&qhp->lock);
714 	spin_unlock_irqrestore(&qhp->rhp->lock, flags);
715 	return 0;
716 }
717 
718 static int ring_kernel_rq_db(struct c4iw_qp *qhp, u16 inc)
719 {
720 	unsigned long flags;
721 
722 	spin_lock_irqsave(&qhp->rhp->lock, flags);
723 	spin_lock(&qhp->lock);
724 	if (qhp->rhp->db_state == NORMAL)
725 		t4_ring_rq_db(&qhp->wq, inc, NULL);
726 	else {
727 		add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
728 		qhp->wq.rq.wq_pidx_inc += inc;
729 	}
730 	spin_unlock(&qhp->lock);
731 	spin_unlock_irqrestore(&qhp->rhp->lock, flags);
732 	return 0;
733 }
734 
735 int c4iw_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
736 		   struct ib_send_wr **bad_wr)
737 {
738 	int err = 0;
739 	u8 len16 = 0;
740 	enum fw_wr_opcodes fw_opcode = 0;
741 	enum fw_ri_wr_flags fw_flags;
742 	struct c4iw_qp *qhp;
743 	union t4_wr *wqe = NULL;
744 	u32 num_wrs;
745 	struct t4_swsqe *swsqe;
746 	unsigned long flag;
747 	u16 idx = 0;
748 
749 	qhp = to_c4iw_qp(ibqp);
750 	spin_lock_irqsave(&qhp->lock, flag);
751 	if (t4_wq_in_error(&qhp->wq)) {
752 		spin_unlock_irqrestore(&qhp->lock, flag);
753 		return -EINVAL;
754 	}
755 	num_wrs = t4_sq_avail(&qhp->wq);
756 	if (num_wrs == 0) {
757 		spin_unlock_irqrestore(&qhp->lock, flag);
758 		return -ENOMEM;
759 	}
760 	while (wr) {
761 		if (num_wrs == 0) {
762 			err = -ENOMEM;
763 			*bad_wr = wr;
764 			break;
765 		}
766 		wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
767 		      qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
768 
769 		fw_flags = 0;
770 		if (wr->send_flags & IB_SEND_SOLICITED)
771 			fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
772 		if (wr->send_flags & IB_SEND_SIGNALED || qhp->sq_sig_all)
773 			fw_flags |= FW_RI_COMPLETION_FLAG;
774 		swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
775 		switch (wr->opcode) {
776 		case IB_WR_SEND_WITH_INV:
777 		case IB_WR_SEND:
778 			if (wr->send_flags & IB_SEND_FENCE)
779 				fw_flags |= FW_RI_READ_FENCE_FLAG;
780 			fw_opcode = FW_RI_SEND_WR;
781 			if (wr->opcode == IB_WR_SEND)
782 				swsqe->opcode = FW_RI_SEND;
783 			else
784 				swsqe->opcode = FW_RI_SEND_WITH_INV;
785 			err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
786 			break;
787 		case IB_WR_RDMA_WRITE:
788 			fw_opcode = FW_RI_RDMA_WRITE_WR;
789 			swsqe->opcode = FW_RI_RDMA_WRITE;
790 			err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
791 			break;
792 		case IB_WR_RDMA_READ:
793 		case IB_WR_RDMA_READ_WITH_INV:
794 			fw_opcode = FW_RI_RDMA_READ_WR;
795 			swsqe->opcode = FW_RI_READ_REQ;
796 			if (wr->opcode == IB_WR_RDMA_READ_WITH_INV)
797 				fw_flags = FW_RI_RDMA_READ_INVALIDATE;
798 			else
799 				fw_flags = 0;
800 			err = build_rdma_read(wqe, wr, &len16);
801 			if (err)
802 				break;
803 			swsqe->read_len = wr->sg_list[0].length;
804 			if (!qhp->wq.sq.oldest_read)
805 				qhp->wq.sq.oldest_read = swsqe;
806 			break;
807 		case IB_WR_REG_MR:
808 			fw_opcode = FW_RI_FR_NSMR_WR;
809 			swsqe->opcode = FW_RI_FAST_REGISTER;
810 			err = build_memreg(&qhp->wq.sq, wqe, reg_wr(wr), &len16,
811 				qhp->rhp->rdev.lldi.ulptx_memwrite_dsgl);
812 			break;
813 		case IB_WR_LOCAL_INV:
814 			if (wr->send_flags & IB_SEND_FENCE)
815 				fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
816 			fw_opcode = FW_RI_INV_LSTAG_WR;
817 			swsqe->opcode = FW_RI_LOCAL_INV;
818 			err = build_inv_stag(wqe, wr, &len16);
819 			break;
820 		default:
821 			PDBG("%s post of type=%d TBD!\n", __func__,
822 			     wr->opcode);
823 			err = -EINVAL;
824 		}
825 		if (err) {
826 			*bad_wr = wr;
827 			break;
828 		}
829 		swsqe->idx = qhp->wq.sq.pidx;
830 		swsqe->complete = 0;
831 		swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED) ||
832 				  qhp->sq_sig_all;
833 		swsqe->flushed = 0;
834 		swsqe->wr_id = wr->wr_id;
835 		if (c4iw_wr_log) {
836 			swsqe->sge_ts = cxgb4_read_sge_timestamp(
837 					qhp->rhp->rdev.lldi.ports[0]);
838 			getnstimeofday(&swsqe->host_ts);
839 		}
840 
841 		init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
842 
843 		PDBG("%s cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u\n",
844 		     __func__, (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
845 		     swsqe->opcode, swsqe->read_len);
846 		wr = wr->next;
847 		num_wrs--;
848 		t4_sq_produce(&qhp->wq, len16);
849 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
850 	}
851 	if (!qhp->rhp->rdev.status_page->db_off) {
852 		t4_ring_sq_db(&qhp->wq, idx, wqe);
853 		spin_unlock_irqrestore(&qhp->lock, flag);
854 	} else {
855 		spin_unlock_irqrestore(&qhp->lock, flag);
856 		ring_kernel_sq_db(qhp, idx);
857 	}
858 	return err;
859 }
860 
861 int c4iw_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
862 		      struct ib_recv_wr **bad_wr)
863 {
864 	int err = 0;
865 	struct c4iw_qp *qhp;
866 	union t4_recv_wr *wqe = NULL;
867 	u32 num_wrs;
868 	u8 len16 = 0;
869 	unsigned long flag;
870 	u16 idx = 0;
871 
872 	qhp = to_c4iw_qp(ibqp);
873 	spin_lock_irqsave(&qhp->lock, flag);
874 	if (t4_wq_in_error(&qhp->wq)) {
875 		spin_unlock_irqrestore(&qhp->lock, flag);
876 		return -EINVAL;
877 	}
878 	num_wrs = t4_rq_avail(&qhp->wq);
879 	if (num_wrs == 0) {
880 		spin_unlock_irqrestore(&qhp->lock, flag);
881 		return -ENOMEM;
882 	}
883 	while (wr) {
884 		if (wr->num_sge > T4_MAX_RECV_SGE) {
885 			err = -EINVAL;
886 			*bad_wr = wr;
887 			break;
888 		}
889 		wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
890 					   qhp->wq.rq.wq_pidx *
891 					   T4_EQ_ENTRY_SIZE);
892 		if (num_wrs)
893 			err = build_rdma_recv(qhp, wqe, wr, &len16);
894 		else
895 			err = -ENOMEM;
896 		if (err) {
897 			*bad_wr = wr;
898 			break;
899 		}
900 
901 		qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
902 		if (c4iw_wr_log) {
903 			qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].sge_ts =
904 				cxgb4_read_sge_timestamp(
905 						qhp->rhp->rdev.lldi.ports[0]);
906 			getnstimeofday(
907 				&qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].host_ts);
908 		}
909 
910 		wqe->recv.opcode = FW_RI_RECV_WR;
911 		wqe->recv.r1 = 0;
912 		wqe->recv.wrid = qhp->wq.rq.pidx;
913 		wqe->recv.r2[0] = 0;
914 		wqe->recv.r2[1] = 0;
915 		wqe->recv.r2[2] = 0;
916 		wqe->recv.len16 = len16;
917 		PDBG("%s cookie 0x%llx pidx %u\n", __func__,
918 		     (unsigned long long) wr->wr_id, qhp->wq.rq.pidx);
919 		t4_rq_produce(&qhp->wq, len16);
920 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
921 		wr = wr->next;
922 		num_wrs--;
923 	}
924 	if (!qhp->rhp->rdev.status_page->db_off) {
925 		t4_ring_rq_db(&qhp->wq, idx, wqe);
926 		spin_unlock_irqrestore(&qhp->lock, flag);
927 	} else {
928 		spin_unlock_irqrestore(&qhp->lock, flag);
929 		ring_kernel_rq_db(qhp, idx);
930 	}
931 	return err;
932 }
933 
934 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
935 				    u8 *ecode)
936 {
937 	int status;
938 	int tagged;
939 	int opcode;
940 	int rqtype;
941 	int send_inv;
942 
943 	if (!err_cqe) {
944 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
945 		*ecode = 0;
946 		return;
947 	}
948 
949 	status = CQE_STATUS(err_cqe);
950 	opcode = CQE_OPCODE(err_cqe);
951 	rqtype = RQ_TYPE(err_cqe);
952 	send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
953 		   (opcode == FW_RI_SEND_WITH_SE_INV);
954 	tagged = (opcode == FW_RI_RDMA_WRITE) ||
955 		 (rqtype && (opcode == FW_RI_READ_RESP));
956 
957 	switch (status) {
958 	case T4_ERR_STAG:
959 		if (send_inv) {
960 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
961 			*ecode = RDMAP_CANT_INV_STAG;
962 		} else {
963 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
964 			*ecode = RDMAP_INV_STAG;
965 		}
966 		break;
967 	case T4_ERR_PDID:
968 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
969 		if ((opcode == FW_RI_SEND_WITH_INV) ||
970 		    (opcode == FW_RI_SEND_WITH_SE_INV))
971 			*ecode = RDMAP_CANT_INV_STAG;
972 		else
973 			*ecode = RDMAP_STAG_NOT_ASSOC;
974 		break;
975 	case T4_ERR_QPID:
976 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
977 		*ecode = RDMAP_STAG_NOT_ASSOC;
978 		break;
979 	case T4_ERR_ACCESS:
980 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
981 		*ecode = RDMAP_ACC_VIOL;
982 		break;
983 	case T4_ERR_WRAP:
984 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
985 		*ecode = RDMAP_TO_WRAP;
986 		break;
987 	case T4_ERR_BOUND:
988 		if (tagged) {
989 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
990 			*ecode = DDPT_BASE_BOUNDS;
991 		} else {
992 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
993 			*ecode = RDMAP_BASE_BOUNDS;
994 		}
995 		break;
996 	case T4_ERR_INVALIDATE_SHARED_MR:
997 	case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
998 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
999 		*ecode = RDMAP_CANT_INV_STAG;
1000 		break;
1001 	case T4_ERR_ECC:
1002 	case T4_ERR_ECC_PSTAG:
1003 	case T4_ERR_INTERNAL_ERR:
1004 		*layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
1005 		*ecode = 0;
1006 		break;
1007 	case T4_ERR_OUT_OF_RQE:
1008 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1009 		*ecode = DDPU_INV_MSN_NOBUF;
1010 		break;
1011 	case T4_ERR_PBL_ADDR_BOUND:
1012 		*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1013 		*ecode = DDPT_BASE_BOUNDS;
1014 		break;
1015 	case T4_ERR_CRC:
1016 		*layer_type = LAYER_MPA|DDP_LLP;
1017 		*ecode = MPA_CRC_ERR;
1018 		break;
1019 	case T4_ERR_MARKER:
1020 		*layer_type = LAYER_MPA|DDP_LLP;
1021 		*ecode = MPA_MARKER_ERR;
1022 		break;
1023 	case T4_ERR_PDU_LEN_ERR:
1024 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1025 		*ecode = DDPU_MSG_TOOBIG;
1026 		break;
1027 	case T4_ERR_DDP_VERSION:
1028 		if (tagged) {
1029 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1030 			*ecode = DDPT_INV_VERS;
1031 		} else {
1032 			*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1033 			*ecode = DDPU_INV_VERS;
1034 		}
1035 		break;
1036 	case T4_ERR_RDMA_VERSION:
1037 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1038 		*ecode = RDMAP_INV_VERS;
1039 		break;
1040 	case T4_ERR_OPCODE:
1041 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1042 		*ecode = RDMAP_INV_OPCODE;
1043 		break;
1044 	case T4_ERR_DDP_QUEUE_NUM:
1045 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1046 		*ecode = DDPU_INV_QN;
1047 		break;
1048 	case T4_ERR_MSN:
1049 	case T4_ERR_MSN_GAP:
1050 	case T4_ERR_MSN_RANGE:
1051 	case T4_ERR_IRD_OVERFLOW:
1052 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1053 		*ecode = DDPU_INV_MSN_RANGE;
1054 		break;
1055 	case T4_ERR_TBIT:
1056 		*layer_type = LAYER_DDP|DDP_LOCAL_CATA;
1057 		*ecode = 0;
1058 		break;
1059 	case T4_ERR_MO:
1060 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1061 		*ecode = DDPU_INV_MO;
1062 		break;
1063 	default:
1064 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
1065 		*ecode = 0;
1066 		break;
1067 	}
1068 }
1069 
1070 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
1071 			   gfp_t gfp)
1072 {
1073 	struct fw_ri_wr *wqe;
1074 	struct sk_buff *skb;
1075 	struct terminate_message *term;
1076 
1077 	PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1078 	     qhp->ep->hwtid);
1079 
1080 	skb = alloc_skb(sizeof *wqe, gfp);
1081 	if (!skb)
1082 		return;
1083 	set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1084 
1085 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1086 	memset(wqe, 0, sizeof *wqe);
1087 	wqe->op_compl = cpu_to_be32(FW_WR_OP_V(FW_RI_INIT_WR));
1088 	wqe->flowid_len16 = cpu_to_be32(
1089 		FW_WR_FLOWID_V(qhp->ep->hwtid) |
1090 		FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1091 
1092 	wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
1093 	wqe->u.terminate.immdlen = cpu_to_be32(sizeof *term);
1094 	term = (struct terminate_message *)wqe->u.terminate.termmsg;
1095 	if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
1096 		term->layer_etype = qhp->attr.layer_etype;
1097 		term->ecode = qhp->attr.ecode;
1098 	} else
1099 		build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
1100 	c4iw_ofld_send(&qhp->rhp->rdev, skb);
1101 }
1102 
1103 /*
1104  * Assumes qhp lock is held.
1105  */
1106 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
1107 		       struct c4iw_cq *schp)
1108 {
1109 	int count;
1110 	int rq_flushed, sq_flushed;
1111 	unsigned long flag;
1112 
1113 	PDBG("%s qhp %p rchp %p schp %p\n", __func__, qhp, rchp, schp);
1114 
1115 	/* locking hierarchy: cq lock first, then qp lock. */
1116 	spin_lock_irqsave(&rchp->lock, flag);
1117 	spin_lock(&qhp->lock);
1118 
1119 	if (qhp->wq.flushed) {
1120 		spin_unlock(&qhp->lock);
1121 		spin_unlock_irqrestore(&rchp->lock, flag);
1122 		return;
1123 	}
1124 	qhp->wq.flushed = 1;
1125 
1126 	c4iw_flush_hw_cq(rchp);
1127 	c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
1128 	rq_flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
1129 	spin_unlock(&qhp->lock);
1130 	spin_unlock_irqrestore(&rchp->lock, flag);
1131 
1132 	/* locking hierarchy: cq lock first, then qp lock. */
1133 	spin_lock_irqsave(&schp->lock, flag);
1134 	spin_lock(&qhp->lock);
1135 	if (schp != rchp)
1136 		c4iw_flush_hw_cq(schp);
1137 	sq_flushed = c4iw_flush_sq(qhp);
1138 	spin_unlock(&qhp->lock);
1139 	spin_unlock_irqrestore(&schp->lock, flag);
1140 
1141 	if (schp == rchp) {
1142 		if (t4_clear_cq_armed(&rchp->cq) &&
1143 		    (rq_flushed || sq_flushed)) {
1144 			spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1145 			(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1146 						   rchp->ibcq.cq_context);
1147 			spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1148 		}
1149 	} else {
1150 		if (t4_clear_cq_armed(&rchp->cq) && rq_flushed) {
1151 			spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1152 			(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1153 						   rchp->ibcq.cq_context);
1154 			spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1155 		}
1156 		if (t4_clear_cq_armed(&schp->cq) && sq_flushed) {
1157 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1158 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1159 						   schp->ibcq.cq_context);
1160 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1161 		}
1162 	}
1163 }
1164 
1165 static void flush_qp(struct c4iw_qp *qhp)
1166 {
1167 	struct c4iw_cq *rchp, *schp;
1168 	unsigned long flag;
1169 
1170 	rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
1171 	schp = to_c4iw_cq(qhp->ibqp.send_cq);
1172 
1173 	t4_set_wq_in_error(&qhp->wq);
1174 	if (qhp->ibqp.uobject) {
1175 		t4_set_cq_in_error(&rchp->cq);
1176 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1177 		(*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1178 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1179 		if (schp != rchp) {
1180 			t4_set_cq_in_error(&schp->cq);
1181 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1182 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1183 					schp->ibcq.cq_context);
1184 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1185 		}
1186 		return;
1187 	}
1188 	__flush_qp(qhp, rchp, schp);
1189 }
1190 
1191 static int rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1192 		     struct c4iw_ep *ep)
1193 {
1194 	struct fw_ri_wr *wqe;
1195 	int ret;
1196 	struct sk_buff *skb;
1197 
1198 	PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1199 	     ep->hwtid);
1200 
1201 	skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1202 	if (!skb)
1203 		return -ENOMEM;
1204 	set_wr_txq(skb, CPL_PRIORITY_DATA, ep->txq_idx);
1205 
1206 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1207 	memset(wqe, 0, sizeof *wqe);
1208 	wqe->op_compl = cpu_to_be32(
1209 		FW_WR_OP_V(FW_RI_INIT_WR) |
1210 		FW_WR_COMPL_F);
1211 	wqe->flowid_len16 = cpu_to_be32(
1212 		FW_WR_FLOWID_V(ep->hwtid) |
1213 		FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1214 	wqe->cookie = (uintptr_t)&ep->com.wr_wait;
1215 
1216 	wqe->u.fini.type = FW_RI_TYPE_FINI;
1217 	ret = c4iw_ofld_send(&rhp->rdev, skb);
1218 	if (ret)
1219 		goto out;
1220 
1221 	ret = c4iw_wait_for_reply(&rhp->rdev, &ep->com.wr_wait, qhp->ep->hwtid,
1222 			     qhp->wq.sq.qid, __func__);
1223 out:
1224 	PDBG("%s ret %d\n", __func__, ret);
1225 	return ret;
1226 }
1227 
1228 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1229 {
1230 	PDBG("%s p2p_type = %d\n", __func__, p2p_type);
1231 	memset(&init->u, 0, sizeof init->u);
1232 	switch (p2p_type) {
1233 	case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1234 		init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1235 		init->u.write.stag_sink = cpu_to_be32(1);
1236 		init->u.write.to_sink = cpu_to_be64(1);
1237 		init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1238 		init->u.write.len16 = DIV_ROUND_UP(sizeof init->u.write +
1239 						   sizeof(struct fw_ri_immd),
1240 						   16);
1241 		break;
1242 	case FW_RI_INIT_P2PTYPE_READ_REQ:
1243 		init->u.write.opcode = FW_RI_RDMA_READ_WR;
1244 		init->u.read.stag_src = cpu_to_be32(1);
1245 		init->u.read.to_src_lo = cpu_to_be32(1);
1246 		init->u.read.stag_sink = cpu_to_be32(1);
1247 		init->u.read.to_sink_lo = cpu_to_be32(1);
1248 		init->u.read.len16 = DIV_ROUND_UP(sizeof init->u.read, 16);
1249 		break;
1250 	}
1251 }
1252 
1253 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1254 {
1255 	struct fw_ri_wr *wqe;
1256 	int ret;
1257 	struct sk_buff *skb;
1258 
1259 	PDBG("%s qhp %p qid 0x%x tid %u ird %u ord %u\n", __func__, qhp,
1260 	     qhp->wq.sq.qid, qhp->ep->hwtid, qhp->ep->ird, qhp->ep->ord);
1261 
1262 	skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1263 	if (!skb) {
1264 		ret = -ENOMEM;
1265 		goto out;
1266 	}
1267 	ret = alloc_ird(rhp, qhp->attr.max_ird);
1268 	if (ret) {
1269 		qhp->attr.max_ird = 0;
1270 		kfree_skb(skb);
1271 		goto out;
1272 	}
1273 	set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1274 
1275 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1276 	memset(wqe, 0, sizeof *wqe);
1277 	wqe->op_compl = cpu_to_be32(
1278 		FW_WR_OP_V(FW_RI_INIT_WR) |
1279 		FW_WR_COMPL_F);
1280 	wqe->flowid_len16 = cpu_to_be32(
1281 		FW_WR_FLOWID_V(qhp->ep->hwtid) |
1282 		FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1283 
1284 	wqe->cookie = (uintptr_t)&qhp->ep->com.wr_wait;
1285 
1286 	wqe->u.init.type = FW_RI_TYPE_INIT;
1287 	wqe->u.init.mpareqbit_p2ptype =
1288 		FW_RI_WR_MPAREQBIT_V(qhp->attr.mpa_attr.initiator) |
1289 		FW_RI_WR_P2PTYPE_V(qhp->attr.mpa_attr.p2p_type);
1290 	wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1291 	if (qhp->attr.mpa_attr.recv_marker_enabled)
1292 		wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1293 	if (qhp->attr.mpa_attr.xmit_marker_enabled)
1294 		wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1295 	if (qhp->attr.mpa_attr.crc_enabled)
1296 		wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1297 
1298 	wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1299 			    FW_RI_QP_RDMA_WRITE_ENABLE |
1300 			    FW_RI_QP_BIND_ENABLE;
1301 	if (!qhp->ibqp.uobject)
1302 		wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1303 				     FW_RI_QP_STAG0_ENABLE;
1304 	wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1305 	wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1306 	wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1307 	wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1308 	wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1309 	wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1310 	wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1311 	wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1312 	wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1313 	wqe->u.init.iss = cpu_to_be32(qhp->ep->snd_seq);
1314 	wqe->u.init.irs = cpu_to_be32(qhp->ep->rcv_seq);
1315 	wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1316 	wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1317 					 rhp->rdev.lldi.vr->rq.start);
1318 	if (qhp->attr.mpa_attr.initiator)
1319 		build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1320 
1321 	ret = c4iw_ofld_send(&rhp->rdev, skb);
1322 	if (ret)
1323 		goto err1;
1324 
1325 	ret = c4iw_wait_for_reply(&rhp->rdev, &qhp->ep->com.wr_wait,
1326 				  qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
1327 	if (!ret)
1328 		goto out;
1329 err1:
1330 	free_ird(rhp, qhp->attr.max_ird);
1331 out:
1332 	PDBG("%s ret %d\n", __func__, ret);
1333 	return ret;
1334 }
1335 
1336 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1337 		   enum c4iw_qp_attr_mask mask,
1338 		   struct c4iw_qp_attributes *attrs,
1339 		   int internal)
1340 {
1341 	int ret = 0;
1342 	struct c4iw_qp_attributes newattr = qhp->attr;
1343 	int disconnect = 0;
1344 	int terminate = 0;
1345 	int abort = 0;
1346 	int free = 0;
1347 	struct c4iw_ep *ep = NULL;
1348 
1349 	PDBG("%s qhp %p sqid 0x%x rqid 0x%x ep %p state %d -> %d\n", __func__,
1350 	     qhp, qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep, qhp->attr.state,
1351 	     (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1352 
1353 	mutex_lock(&qhp->mutex);
1354 
1355 	/* Process attr changes if in IDLE */
1356 	if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1357 		if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1358 			ret = -EIO;
1359 			goto out;
1360 		}
1361 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1362 			newattr.enable_rdma_read = attrs->enable_rdma_read;
1363 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1364 			newattr.enable_rdma_write = attrs->enable_rdma_write;
1365 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1366 			newattr.enable_bind = attrs->enable_bind;
1367 		if (mask & C4IW_QP_ATTR_MAX_ORD) {
1368 			if (attrs->max_ord > c4iw_max_read_depth) {
1369 				ret = -EINVAL;
1370 				goto out;
1371 			}
1372 			newattr.max_ord = attrs->max_ord;
1373 		}
1374 		if (mask & C4IW_QP_ATTR_MAX_IRD) {
1375 			if (attrs->max_ird > cur_max_read_depth(rhp)) {
1376 				ret = -EINVAL;
1377 				goto out;
1378 			}
1379 			newattr.max_ird = attrs->max_ird;
1380 		}
1381 		qhp->attr = newattr;
1382 	}
1383 
1384 	if (mask & C4IW_QP_ATTR_SQ_DB) {
1385 		ret = ring_kernel_sq_db(qhp, attrs->sq_db_inc);
1386 		goto out;
1387 	}
1388 	if (mask & C4IW_QP_ATTR_RQ_DB) {
1389 		ret = ring_kernel_rq_db(qhp, attrs->rq_db_inc);
1390 		goto out;
1391 	}
1392 
1393 	if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1394 		goto out;
1395 	if (qhp->attr.state == attrs->next_state)
1396 		goto out;
1397 
1398 	switch (qhp->attr.state) {
1399 	case C4IW_QP_STATE_IDLE:
1400 		switch (attrs->next_state) {
1401 		case C4IW_QP_STATE_RTS:
1402 			if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1403 				ret = -EINVAL;
1404 				goto out;
1405 			}
1406 			if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1407 				ret = -EINVAL;
1408 				goto out;
1409 			}
1410 			qhp->attr.mpa_attr = attrs->mpa_attr;
1411 			qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1412 			qhp->ep = qhp->attr.llp_stream_handle;
1413 			set_state(qhp, C4IW_QP_STATE_RTS);
1414 
1415 			/*
1416 			 * Ref the endpoint here and deref when we
1417 			 * disassociate the endpoint from the QP.  This
1418 			 * happens in CLOSING->IDLE transition or *->ERROR
1419 			 * transition.
1420 			 */
1421 			c4iw_get_ep(&qhp->ep->com);
1422 			ret = rdma_init(rhp, qhp);
1423 			if (ret)
1424 				goto err;
1425 			break;
1426 		case C4IW_QP_STATE_ERROR:
1427 			set_state(qhp, C4IW_QP_STATE_ERROR);
1428 			flush_qp(qhp);
1429 			break;
1430 		default:
1431 			ret = -EINVAL;
1432 			goto out;
1433 		}
1434 		break;
1435 	case C4IW_QP_STATE_RTS:
1436 		switch (attrs->next_state) {
1437 		case C4IW_QP_STATE_CLOSING:
1438 			BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
1439 			t4_set_wq_in_error(&qhp->wq);
1440 			set_state(qhp, C4IW_QP_STATE_CLOSING);
1441 			ep = qhp->ep;
1442 			if (!internal) {
1443 				abort = 0;
1444 				disconnect = 1;
1445 				c4iw_get_ep(&qhp->ep->com);
1446 			}
1447 			ret = rdma_fini(rhp, qhp, ep);
1448 			if (ret)
1449 				goto err;
1450 			break;
1451 		case C4IW_QP_STATE_TERMINATE:
1452 			t4_set_wq_in_error(&qhp->wq);
1453 			set_state(qhp, C4IW_QP_STATE_TERMINATE);
1454 			qhp->attr.layer_etype = attrs->layer_etype;
1455 			qhp->attr.ecode = attrs->ecode;
1456 			ep = qhp->ep;
1457 			if (!internal) {
1458 				c4iw_get_ep(&qhp->ep->com);
1459 				terminate = 1;
1460 				disconnect = 1;
1461 			} else {
1462 				terminate = qhp->attr.send_term;
1463 				ret = rdma_fini(rhp, qhp, ep);
1464 				if (ret)
1465 					goto err;
1466 			}
1467 			break;
1468 		case C4IW_QP_STATE_ERROR:
1469 			t4_set_wq_in_error(&qhp->wq);
1470 			set_state(qhp, C4IW_QP_STATE_ERROR);
1471 			if (!internal) {
1472 				abort = 1;
1473 				disconnect = 1;
1474 				ep = qhp->ep;
1475 				c4iw_get_ep(&qhp->ep->com);
1476 			}
1477 			goto err;
1478 			break;
1479 		default:
1480 			ret = -EINVAL;
1481 			goto out;
1482 		}
1483 		break;
1484 	case C4IW_QP_STATE_CLOSING:
1485 		if (!internal) {
1486 			ret = -EINVAL;
1487 			goto out;
1488 		}
1489 		switch (attrs->next_state) {
1490 		case C4IW_QP_STATE_IDLE:
1491 			flush_qp(qhp);
1492 			set_state(qhp, C4IW_QP_STATE_IDLE);
1493 			qhp->attr.llp_stream_handle = NULL;
1494 			c4iw_put_ep(&qhp->ep->com);
1495 			qhp->ep = NULL;
1496 			wake_up(&qhp->wait);
1497 			break;
1498 		case C4IW_QP_STATE_ERROR:
1499 			goto err;
1500 		default:
1501 			ret = -EINVAL;
1502 			goto err;
1503 		}
1504 		break;
1505 	case C4IW_QP_STATE_ERROR:
1506 		if (attrs->next_state != C4IW_QP_STATE_IDLE) {
1507 			ret = -EINVAL;
1508 			goto out;
1509 		}
1510 		if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) {
1511 			ret = -EINVAL;
1512 			goto out;
1513 		}
1514 		set_state(qhp, C4IW_QP_STATE_IDLE);
1515 		break;
1516 	case C4IW_QP_STATE_TERMINATE:
1517 		if (!internal) {
1518 			ret = -EINVAL;
1519 			goto out;
1520 		}
1521 		goto err;
1522 		break;
1523 	default:
1524 		printk(KERN_ERR "%s in a bad state %d\n",
1525 		       __func__, qhp->attr.state);
1526 		ret = -EINVAL;
1527 		goto err;
1528 		break;
1529 	}
1530 	goto out;
1531 err:
1532 	PDBG("%s disassociating ep %p qpid 0x%x\n", __func__, qhp->ep,
1533 	     qhp->wq.sq.qid);
1534 
1535 	/* disassociate the LLP connection */
1536 	qhp->attr.llp_stream_handle = NULL;
1537 	if (!ep)
1538 		ep = qhp->ep;
1539 	qhp->ep = NULL;
1540 	set_state(qhp, C4IW_QP_STATE_ERROR);
1541 	free = 1;
1542 	abort = 1;
1543 	BUG_ON(!ep);
1544 	flush_qp(qhp);
1545 	wake_up(&qhp->wait);
1546 out:
1547 	mutex_unlock(&qhp->mutex);
1548 
1549 	if (terminate)
1550 		post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
1551 
1552 	/*
1553 	 * If disconnect is 1, then we need to initiate a disconnect
1554 	 * on the EP.  This can be a normal close (RTS->CLOSING) or
1555 	 * an abnormal close (RTS/CLOSING->ERROR).
1556 	 */
1557 	if (disconnect) {
1558 		c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC :
1559 							 GFP_KERNEL);
1560 		c4iw_put_ep(&ep->com);
1561 	}
1562 
1563 	/*
1564 	 * If free is 1, then we've disassociated the EP from the QP
1565 	 * and we need to dereference the EP.
1566 	 */
1567 	if (free)
1568 		c4iw_put_ep(&ep->com);
1569 	PDBG("%s exit state %d\n", __func__, qhp->attr.state);
1570 	return ret;
1571 }
1572 
1573 int c4iw_destroy_qp(struct ib_qp *ib_qp)
1574 {
1575 	struct c4iw_dev *rhp;
1576 	struct c4iw_qp *qhp;
1577 	struct c4iw_qp_attributes attrs;
1578 	struct c4iw_ucontext *ucontext;
1579 
1580 	qhp = to_c4iw_qp(ib_qp);
1581 	rhp = qhp->rhp;
1582 
1583 	attrs.next_state = C4IW_QP_STATE_ERROR;
1584 	if (qhp->attr.state == C4IW_QP_STATE_TERMINATE)
1585 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
1586 	else
1587 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
1588 	wait_event(qhp->wait, !qhp->ep);
1589 
1590 	remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1591 	atomic_dec(&qhp->refcnt);
1592 	wait_event(qhp->wait, !atomic_read(&qhp->refcnt));
1593 
1594 	spin_lock_irq(&rhp->lock);
1595 	if (!list_empty(&qhp->db_fc_entry))
1596 		list_del_init(&qhp->db_fc_entry);
1597 	spin_unlock_irq(&rhp->lock);
1598 	free_ird(rhp, qhp->attr.max_ird);
1599 
1600 	ucontext = ib_qp->uobject ?
1601 		   to_c4iw_ucontext(ib_qp->uobject->context) : NULL;
1602 	destroy_qp(&rhp->rdev, &qhp->wq,
1603 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1604 
1605 	PDBG("%s ib_qp %p qpid 0x%0x\n", __func__, ib_qp, qhp->wq.sq.qid);
1606 	kfree(qhp);
1607 	return 0;
1608 }
1609 
1610 struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
1611 			     struct ib_udata *udata)
1612 {
1613 	struct c4iw_dev *rhp;
1614 	struct c4iw_qp *qhp;
1615 	struct c4iw_pd *php;
1616 	struct c4iw_cq *schp;
1617 	struct c4iw_cq *rchp;
1618 	struct c4iw_create_qp_resp uresp;
1619 	unsigned int sqsize, rqsize;
1620 	struct c4iw_ucontext *ucontext;
1621 	int ret;
1622 	struct c4iw_mm_entry *sq_key_mm, *rq_key_mm = NULL, *sq_db_key_mm;
1623 	struct c4iw_mm_entry *rq_db_key_mm = NULL, *ma_sync_key_mm = NULL;
1624 
1625 	PDBG("%s ib_pd %p\n", __func__, pd);
1626 
1627 	if (attrs->qp_type != IB_QPT_RC)
1628 		return ERR_PTR(-EINVAL);
1629 
1630 	php = to_c4iw_pd(pd);
1631 	rhp = php->rhp;
1632 	schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
1633 	rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
1634 	if (!schp || !rchp)
1635 		return ERR_PTR(-EINVAL);
1636 
1637 	if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
1638 		return ERR_PTR(-EINVAL);
1639 
1640 	if (attrs->cap.max_recv_wr > rhp->rdev.hw_queue.t4_max_rq_size)
1641 		return ERR_PTR(-E2BIG);
1642 	rqsize = attrs->cap.max_recv_wr + 1;
1643 	if (rqsize < 8)
1644 		rqsize = 8;
1645 
1646 	if (attrs->cap.max_send_wr > rhp->rdev.hw_queue.t4_max_sq_size)
1647 		return ERR_PTR(-E2BIG);
1648 	sqsize = attrs->cap.max_send_wr + 1;
1649 	if (sqsize < 8)
1650 		sqsize = 8;
1651 
1652 	ucontext = pd->uobject ? to_c4iw_ucontext(pd->uobject->context) : NULL;
1653 
1654 	qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
1655 	if (!qhp)
1656 		return ERR_PTR(-ENOMEM);
1657 	qhp->wq.sq.size = sqsize;
1658 	qhp->wq.sq.memsize =
1659 		(sqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1660 		sizeof(*qhp->wq.sq.queue) + 16 * sizeof(__be64);
1661 	qhp->wq.sq.flush_cidx = -1;
1662 	qhp->wq.rq.size = rqsize;
1663 	qhp->wq.rq.memsize =
1664 		(rqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1665 		sizeof(*qhp->wq.rq.queue);
1666 
1667 	if (ucontext) {
1668 		qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
1669 		qhp->wq.rq.memsize = roundup(qhp->wq.rq.memsize, PAGE_SIZE);
1670 	}
1671 
1672 	ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
1673 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1674 	if (ret)
1675 		goto err1;
1676 
1677 	attrs->cap.max_recv_wr = rqsize - 1;
1678 	attrs->cap.max_send_wr = sqsize - 1;
1679 	attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
1680 
1681 	qhp->rhp = rhp;
1682 	qhp->attr.pd = php->pdid;
1683 	qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
1684 	qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
1685 	qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
1686 	qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
1687 	qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
1688 	qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
1689 	qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
1690 	qhp->attr.state = C4IW_QP_STATE_IDLE;
1691 	qhp->attr.next_state = C4IW_QP_STATE_IDLE;
1692 	qhp->attr.enable_rdma_read = 1;
1693 	qhp->attr.enable_rdma_write = 1;
1694 	qhp->attr.enable_bind = 1;
1695 	qhp->attr.max_ord = 0;
1696 	qhp->attr.max_ird = 0;
1697 	qhp->sq_sig_all = attrs->sq_sig_type == IB_SIGNAL_ALL_WR;
1698 	spin_lock_init(&qhp->lock);
1699 	init_completion(&qhp->sq_drained);
1700 	init_completion(&qhp->rq_drained);
1701 	mutex_init(&qhp->mutex);
1702 	init_waitqueue_head(&qhp->wait);
1703 	atomic_set(&qhp->refcnt, 1);
1704 
1705 	ret = insert_handle(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid);
1706 	if (ret)
1707 		goto err2;
1708 
1709 	if (udata) {
1710 		sq_key_mm = kmalloc(sizeof(*sq_key_mm), GFP_KERNEL);
1711 		if (!sq_key_mm) {
1712 			ret = -ENOMEM;
1713 			goto err3;
1714 		}
1715 		rq_key_mm = kmalloc(sizeof(*rq_key_mm), GFP_KERNEL);
1716 		if (!rq_key_mm) {
1717 			ret = -ENOMEM;
1718 			goto err4;
1719 		}
1720 		sq_db_key_mm = kmalloc(sizeof(*sq_db_key_mm), GFP_KERNEL);
1721 		if (!sq_db_key_mm) {
1722 			ret = -ENOMEM;
1723 			goto err5;
1724 		}
1725 		rq_db_key_mm = kmalloc(sizeof(*rq_db_key_mm), GFP_KERNEL);
1726 		if (!rq_db_key_mm) {
1727 			ret = -ENOMEM;
1728 			goto err6;
1729 		}
1730 		if (t4_sq_onchip(&qhp->wq.sq)) {
1731 			ma_sync_key_mm = kmalloc(sizeof(*ma_sync_key_mm),
1732 						 GFP_KERNEL);
1733 			if (!ma_sync_key_mm) {
1734 				ret = -ENOMEM;
1735 				goto err7;
1736 			}
1737 			uresp.flags = C4IW_QPF_ONCHIP;
1738 		} else
1739 			uresp.flags = 0;
1740 		uresp.qid_mask = rhp->rdev.qpmask;
1741 		uresp.sqid = qhp->wq.sq.qid;
1742 		uresp.sq_size = qhp->wq.sq.size;
1743 		uresp.sq_memsize = qhp->wq.sq.memsize;
1744 		uresp.rqid = qhp->wq.rq.qid;
1745 		uresp.rq_size = qhp->wq.rq.size;
1746 		uresp.rq_memsize = qhp->wq.rq.memsize;
1747 		spin_lock(&ucontext->mmap_lock);
1748 		if (ma_sync_key_mm) {
1749 			uresp.ma_sync_key = ucontext->key;
1750 			ucontext->key += PAGE_SIZE;
1751 		} else {
1752 			uresp.ma_sync_key =  0;
1753 		}
1754 		uresp.sq_key = ucontext->key;
1755 		ucontext->key += PAGE_SIZE;
1756 		uresp.rq_key = ucontext->key;
1757 		ucontext->key += PAGE_SIZE;
1758 		uresp.sq_db_gts_key = ucontext->key;
1759 		ucontext->key += PAGE_SIZE;
1760 		uresp.rq_db_gts_key = ucontext->key;
1761 		ucontext->key += PAGE_SIZE;
1762 		spin_unlock(&ucontext->mmap_lock);
1763 		ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
1764 		if (ret)
1765 			goto err8;
1766 		sq_key_mm->key = uresp.sq_key;
1767 		sq_key_mm->addr = qhp->wq.sq.phys_addr;
1768 		sq_key_mm->len = PAGE_ALIGN(qhp->wq.sq.memsize);
1769 		insert_mmap(ucontext, sq_key_mm);
1770 		rq_key_mm->key = uresp.rq_key;
1771 		rq_key_mm->addr = virt_to_phys(qhp->wq.rq.queue);
1772 		rq_key_mm->len = PAGE_ALIGN(qhp->wq.rq.memsize);
1773 		insert_mmap(ucontext, rq_key_mm);
1774 		sq_db_key_mm->key = uresp.sq_db_gts_key;
1775 		sq_db_key_mm->addr = (u64)(unsigned long)qhp->wq.sq.bar2_pa;
1776 		sq_db_key_mm->len = PAGE_SIZE;
1777 		insert_mmap(ucontext, sq_db_key_mm);
1778 		rq_db_key_mm->key = uresp.rq_db_gts_key;
1779 		rq_db_key_mm->addr = (u64)(unsigned long)qhp->wq.rq.bar2_pa;
1780 		rq_db_key_mm->len = PAGE_SIZE;
1781 		insert_mmap(ucontext, rq_db_key_mm);
1782 		if (ma_sync_key_mm) {
1783 			ma_sync_key_mm->key = uresp.ma_sync_key;
1784 			ma_sync_key_mm->addr =
1785 				(pci_resource_start(rhp->rdev.lldi.pdev, 0) +
1786 				PCIE_MA_SYNC_A) & PAGE_MASK;
1787 			ma_sync_key_mm->len = PAGE_SIZE;
1788 			insert_mmap(ucontext, ma_sync_key_mm);
1789 		}
1790 	}
1791 	qhp->ibqp.qp_num = qhp->wq.sq.qid;
1792 	init_timer(&(qhp->timer));
1793 	INIT_LIST_HEAD(&qhp->db_fc_entry);
1794 	PDBG("%s sq id %u size %u memsize %zu num_entries %u "
1795 	     "rq id %u size %u memsize %zu num_entries %u\n", __func__,
1796 	     qhp->wq.sq.qid, qhp->wq.sq.size, qhp->wq.sq.memsize,
1797 	     attrs->cap.max_send_wr, qhp->wq.rq.qid, qhp->wq.rq.size,
1798 	     qhp->wq.rq.memsize, attrs->cap.max_recv_wr);
1799 	return &qhp->ibqp;
1800 err8:
1801 	kfree(ma_sync_key_mm);
1802 err7:
1803 	kfree(rq_db_key_mm);
1804 err6:
1805 	kfree(sq_db_key_mm);
1806 err5:
1807 	kfree(rq_key_mm);
1808 err4:
1809 	kfree(sq_key_mm);
1810 err3:
1811 	remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1812 err2:
1813 	destroy_qp(&rhp->rdev, &qhp->wq,
1814 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1815 err1:
1816 	kfree(qhp);
1817 	return ERR_PTR(ret);
1818 }
1819 
1820 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1821 		      int attr_mask, struct ib_udata *udata)
1822 {
1823 	struct c4iw_dev *rhp;
1824 	struct c4iw_qp *qhp;
1825 	enum c4iw_qp_attr_mask mask = 0;
1826 	struct c4iw_qp_attributes attrs;
1827 
1828 	PDBG("%s ib_qp %p\n", __func__, ibqp);
1829 
1830 	/* iwarp does not support the RTR state */
1831 	if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
1832 		attr_mask &= ~IB_QP_STATE;
1833 
1834 	/* Make sure we still have something left to do */
1835 	if (!attr_mask)
1836 		return 0;
1837 
1838 	memset(&attrs, 0, sizeof attrs);
1839 	qhp = to_c4iw_qp(ibqp);
1840 	rhp = qhp->rhp;
1841 
1842 	attrs.next_state = c4iw_convert_state(attr->qp_state);
1843 	attrs.enable_rdma_read = (attr->qp_access_flags &
1844 			       IB_ACCESS_REMOTE_READ) ?  1 : 0;
1845 	attrs.enable_rdma_write = (attr->qp_access_flags &
1846 				IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
1847 	attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
1848 
1849 
1850 	mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
1851 	mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
1852 			(C4IW_QP_ATTR_ENABLE_RDMA_READ |
1853 			 C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
1854 			 C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
1855 
1856 	/*
1857 	 * Use SQ_PSN and RQ_PSN to pass in IDX_INC values for
1858 	 * ringing the queue db when we're in DB_FULL mode.
1859 	 * Only allow this on T4 devices.
1860 	 */
1861 	attrs.sq_db_inc = attr->sq_psn;
1862 	attrs.rq_db_inc = attr->rq_psn;
1863 	mask |= (attr_mask & IB_QP_SQ_PSN) ? C4IW_QP_ATTR_SQ_DB : 0;
1864 	mask |= (attr_mask & IB_QP_RQ_PSN) ? C4IW_QP_ATTR_RQ_DB : 0;
1865 	if (!is_t4(to_c4iw_qp(ibqp)->rhp->rdev.lldi.adapter_type) &&
1866 	    (mask & (C4IW_QP_ATTR_SQ_DB|C4IW_QP_ATTR_RQ_DB)))
1867 		return -EINVAL;
1868 
1869 	return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
1870 }
1871 
1872 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
1873 {
1874 	PDBG("%s ib_dev %p qpn 0x%x\n", __func__, dev, qpn);
1875 	return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
1876 }
1877 
1878 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1879 		     int attr_mask, struct ib_qp_init_attr *init_attr)
1880 {
1881 	struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
1882 
1883 	memset(attr, 0, sizeof *attr);
1884 	memset(init_attr, 0, sizeof *init_attr);
1885 	attr->qp_state = to_ib_qp_state(qhp->attr.state);
1886 	init_attr->cap.max_send_wr = qhp->attr.sq_num_entries;
1887 	init_attr->cap.max_recv_wr = qhp->attr.rq_num_entries;
1888 	init_attr->cap.max_send_sge = qhp->attr.sq_max_sges;
1889 	init_attr->cap.max_recv_sge = qhp->attr.sq_max_sges;
1890 	init_attr->cap.max_inline_data = T4_MAX_SEND_INLINE;
1891 	init_attr->sq_sig_type = qhp->sq_sig_all ? IB_SIGNAL_ALL_WR : 0;
1892 	return 0;
1893 }
1894 
1895 void c4iw_drain_sq(struct ib_qp *ibqp)
1896 {
1897 	struct c4iw_qp *qp = to_c4iw_qp(ibqp);
1898 
1899 	wait_for_completion(&qp->sq_drained);
1900 }
1901 
1902 void c4iw_drain_rq(struct ib_qp *ibqp)
1903 {
1904 	struct c4iw_qp *qp = to_c4iw_qp(ibqp);
1905 
1906 	wait_for_completion(&qp->rq_drained);
1907 }
1908