xref: /openbmc/linux/drivers/infiniband/hw/cxgb4/qp.c (revision f5b06569)
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 int build_memreg(struct t4_sq *sq, union t4_wr *wqe,
613 			struct ib_reg_wr *wr, u8 *len16, bool dsgl_supported)
614 {
615 	struct c4iw_mr *mhp = to_c4iw_mr(wr->mr);
616 	struct fw_ri_immd *imdp;
617 	__be64 *p;
618 	int i;
619 	int pbllen = roundup(mhp->mpl_len * sizeof(u64), 32);
620 	int rem;
621 
622 	if (mhp->mpl_len > t4_max_fr_depth(dsgl_supported && use_dsgl))
623 		return -EINVAL;
624 
625 	wqe->fr.qpbinde_to_dcacpu = 0;
626 	wqe->fr.pgsz_shift = ilog2(wr->mr->page_size) - 12;
627 	wqe->fr.addr_type = FW_RI_VA_BASED_TO;
628 	wqe->fr.mem_perms = c4iw_ib_to_tpt_access(wr->access);
629 	wqe->fr.len_hi = 0;
630 	wqe->fr.len_lo = cpu_to_be32(mhp->ibmr.length);
631 	wqe->fr.stag = cpu_to_be32(wr->key);
632 	wqe->fr.va_hi = cpu_to_be32(mhp->ibmr.iova >> 32);
633 	wqe->fr.va_lo_fbo = cpu_to_be32(mhp->ibmr.iova &
634 					0xffffffff);
635 
636 	if (dsgl_supported && use_dsgl && (pbllen > max_fr_immd)) {
637 		struct fw_ri_dsgl *sglp;
638 
639 		for (i = 0; i < mhp->mpl_len; i++)
640 			mhp->mpl[i] = (__force u64)cpu_to_be64((u64)mhp->mpl[i]);
641 
642 		sglp = (struct fw_ri_dsgl *)(&wqe->fr + 1);
643 		sglp->op = FW_RI_DATA_DSGL;
644 		sglp->r1 = 0;
645 		sglp->nsge = cpu_to_be16(1);
646 		sglp->addr0 = cpu_to_be64(mhp->mpl_addr);
647 		sglp->len0 = cpu_to_be32(pbllen);
648 
649 		*len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*sglp), 16);
650 	} else {
651 		imdp = (struct fw_ri_immd *)(&wqe->fr + 1);
652 		imdp->op = FW_RI_DATA_IMMD;
653 		imdp->r1 = 0;
654 		imdp->r2 = 0;
655 		imdp->immdlen = cpu_to_be32(pbllen);
656 		p = (__be64 *)(imdp + 1);
657 		rem = pbllen;
658 		for (i = 0; i < mhp->mpl_len; i++) {
659 			*p = cpu_to_be64((u64)mhp->mpl[i]);
660 			rem -= sizeof(*p);
661 			if (++p == (__be64 *)&sq->queue[sq->size])
662 				p = (__be64 *)sq->queue;
663 		}
664 		BUG_ON(rem < 0);
665 		while (rem) {
666 			*p = 0;
667 			rem -= sizeof(*p);
668 			if (++p == (__be64 *)&sq->queue[sq->size])
669 				p = (__be64 *)sq->queue;
670 		}
671 		*len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*imdp)
672 				      + pbllen, 16);
673 	}
674 	return 0;
675 }
676 
677 static int build_inv_stag(union t4_wr *wqe, struct ib_send_wr *wr,
678 			  u8 *len16)
679 {
680 	wqe->inv.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
681 	wqe->inv.r2 = 0;
682 	*len16 = DIV_ROUND_UP(sizeof wqe->inv, 16);
683 	return 0;
684 }
685 
686 void _free_qp(struct kref *kref)
687 {
688 	struct c4iw_qp *qhp;
689 
690 	qhp = container_of(kref, struct c4iw_qp, kref);
691 	PDBG("%s qhp %p\n", __func__, qhp);
692 	kfree(qhp);
693 }
694 
695 void c4iw_qp_add_ref(struct ib_qp *qp)
696 {
697 	PDBG("%s ib_qp %p\n", __func__, qp);
698 	kref_get(&to_c4iw_qp(qp)->kref);
699 }
700 
701 void c4iw_qp_rem_ref(struct ib_qp *qp)
702 {
703 	PDBG("%s ib_qp %p\n", __func__, qp);
704 	kref_put(&to_c4iw_qp(qp)->kref, _free_qp);
705 }
706 
707 static void add_to_fc_list(struct list_head *head, struct list_head *entry)
708 {
709 	if (list_empty(entry))
710 		list_add_tail(entry, head);
711 }
712 
713 static int ring_kernel_sq_db(struct c4iw_qp *qhp, u16 inc)
714 {
715 	unsigned long flags;
716 
717 	spin_lock_irqsave(&qhp->rhp->lock, flags);
718 	spin_lock(&qhp->lock);
719 	if (qhp->rhp->db_state == NORMAL)
720 		t4_ring_sq_db(&qhp->wq, inc, NULL);
721 	else {
722 		add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
723 		qhp->wq.sq.wq_pidx_inc += inc;
724 	}
725 	spin_unlock(&qhp->lock);
726 	spin_unlock_irqrestore(&qhp->rhp->lock, flags);
727 	return 0;
728 }
729 
730 static int ring_kernel_rq_db(struct c4iw_qp *qhp, u16 inc)
731 {
732 	unsigned long flags;
733 
734 	spin_lock_irqsave(&qhp->rhp->lock, flags);
735 	spin_lock(&qhp->lock);
736 	if (qhp->rhp->db_state == NORMAL)
737 		t4_ring_rq_db(&qhp->wq, inc, NULL);
738 	else {
739 		add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
740 		qhp->wq.rq.wq_pidx_inc += inc;
741 	}
742 	spin_unlock(&qhp->lock);
743 	spin_unlock_irqrestore(&qhp->rhp->lock, flags);
744 	return 0;
745 }
746 
747 int c4iw_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
748 		   struct ib_send_wr **bad_wr)
749 {
750 	int err = 0;
751 	u8 len16 = 0;
752 	enum fw_wr_opcodes fw_opcode = 0;
753 	enum fw_ri_wr_flags fw_flags;
754 	struct c4iw_qp *qhp;
755 	union t4_wr *wqe = NULL;
756 	u32 num_wrs;
757 	struct t4_swsqe *swsqe;
758 	unsigned long flag;
759 	u16 idx = 0;
760 
761 	qhp = to_c4iw_qp(ibqp);
762 	spin_lock_irqsave(&qhp->lock, flag);
763 	if (t4_wq_in_error(&qhp->wq)) {
764 		spin_unlock_irqrestore(&qhp->lock, flag);
765 		return -EINVAL;
766 	}
767 	num_wrs = t4_sq_avail(&qhp->wq);
768 	if (num_wrs == 0) {
769 		spin_unlock_irqrestore(&qhp->lock, flag);
770 		return -ENOMEM;
771 	}
772 	while (wr) {
773 		if (num_wrs == 0) {
774 			err = -ENOMEM;
775 			*bad_wr = wr;
776 			break;
777 		}
778 		wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
779 		      qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
780 
781 		fw_flags = 0;
782 		if (wr->send_flags & IB_SEND_SOLICITED)
783 			fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
784 		if (wr->send_flags & IB_SEND_SIGNALED || qhp->sq_sig_all)
785 			fw_flags |= FW_RI_COMPLETION_FLAG;
786 		swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
787 		switch (wr->opcode) {
788 		case IB_WR_SEND_WITH_INV:
789 		case IB_WR_SEND:
790 			if (wr->send_flags & IB_SEND_FENCE)
791 				fw_flags |= FW_RI_READ_FENCE_FLAG;
792 			fw_opcode = FW_RI_SEND_WR;
793 			if (wr->opcode == IB_WR_SEND)
794 				swsqe->opcode = FW_RI_SEND;
795 			else
796 				swsqe->opcode = FW_RI_SEND_WITH_INV;
797 			err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
798 			break;
799 		case IB_WR_RDMA_WRITE:
800 			fw_opcode = FW_RI_RDMA_WRITE_WR;
801 			swsqe->opcode = FW_RI_RDMA_WRITE;
802 			err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
803 			break;
804 		case IB_WR_RDMA_READ:
805 		case IB_WR_RDMA_READ_WITH_INV:
806 			fw_opcode = FW_RI_RDMA_READ_WR;
807 			swsqe->opcode = FW_RI_READ_REQ;
808 			if (wr->opcode == IB_WR_RDMA_READ_WITH_INV)
809 				fw_flags = FW_RI_RDMA_READ_INVALIDATE;
810 			else
811 				fw_flags = 0;
812 			err = build_rdma_read(wqe, wr, &len16);
813 			if (err)
814 				break;
815 			swsqe->read_len = wr->sg_list[0].length;
816 			if (!qhp->wq.sq.oldest_read)
817 				qhp->wq.sq.oldest_read = swsqe;
818 			break;
819 		case IB_WR_REG_MR:
820 			fw_opcode = FW_RI_FR_NSMR_WR;
821 			swsqe->opcode = FW_RI_FAST_REGISTER;
822 			err = build_memreg(&qhp->wq.sq, wqe, reg_wr(wr), &len16,
823 				qhp->rhp->rdev.lldi.ulptx_memwrite_dsgl);
824 			break;
825 		case IB_WR_LOCAL_INV:
826 			if (wr->send_flags & IB_SEND_FENCE)
827 				fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
828 			fw_opcode = FW_RI_INV_LSTAG_WR;
829 			swsqe->opcode = FW_RI_LOCAL_INV;
830 			err = build_inv_stag(wqe, wr, &len16);
831 			break;
832 		default:
833 			PDBG("%s post of type=%d TBD!\n", __func__,
834 			     wr->opcode);
835 			err = -EINVAL;
836 		}
837 		if (err) {
838 			*bad_wr = wr;
839 			break;
840 		}
841 		swsqe->idx = qhp->wq.sq.pidx;
842 		swsqe->complete = 0;
843 		swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED) ||
844 				  qhp->sq_sig_all;
845 		swsqe->flushed = 0;
846 		swsqe->wr_id = wr->wr_id;
847 		if (c4iw_wr_log) {
848 			swsqe->sge_ts = cxgb4_read_sge_timestamp(
849 					qhp->rhp->rdev.lldi.ports[0]);
850 			getnstimeofday(&swsqe->host_ts);
851 		}
852 
853 		init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
854 
855 		PDBG("%s cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u\n",
856 		     __func__, (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
857 		     swsqe->opcode, swsqe->read_len);
858 		wr = wr->next;
859 		num_wrs--;
860 		t4_sq_produce(&qhp->wq, len16);
861 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
862 	}
863 	if (!qhp->rhp->rdev.status_page->db_off) {
864 		t4_ring_sq_db(&qhp->wq, idx, wqe);
865 		spin_unlock_irqrestore(&qhp->lock, flag);
866 	} else {
867 		spin_unlock_irqrestore(&qhp->lock, flag);
868 		ring_kernel_sq_db(qhp, idx);
869 	}
870 	return err;
871 }
872 
873 int c4iw_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
874 		      struct ib_recv_wr **bad_wr)
875 {
876 	int err = 0;
877 	struct c4iw_qp *qhp;
878 	union t4_recv_wr *wqe = NULL;
879 	u32 num_wrs;
880 	u8 len16 = 0;
881 	unsigned long flag;
882 	u16 idx = 0;
883 
884 	qhp = to_c4iw_qp(ibqp);
885 	spin_lock_irqsave(&qhp->lock, flag);
886 	if (t4_wq_in_error(&qhp->wq)) {
887 		spin_unlock_irqrestore(&qhp->lock, flag);
888 		return -EINVAL;
889 	}
890 	num_wrs = t4_rq_avail(&qhp->wq);
891 	if (num_wrs == 0) {
892 		spin_unlock_irqrestore(&qhp->lock, flag);
893 		return -ENOMEM;
894 	}
895 	while (wr) {
896 		if (wr->num_sge > T4_MAX_RECV_SGE) {
897 			err = -EINVAL;
898 			*bad_wr = wr;
899 			break;
900 		}
901 		wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
902 					   qhp->wq.rq.wq_pidx *
903 					   T4_EQ_ENTRY_SIZE);
904 		if (num_wrs)
905 			err = build_rdma_recv(qhp, wqe, wr, &len16);
906 		else
907 			err = -ENOMEM;
908 		if (err) {
909 			*bad_wr = wr;
910 			break;
911 		}
912 
913 		qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
914 		if (c4iw_wr_log) {
915 			qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].sge_ts =
916 				cxgb4_read_sge_timestamp(
917 						qhp->rhp->rdev.lldi.ports[0]);
918 			getnstimeofday(
919 				&qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].host_ts);
920 		}
921 
922 		wqe->recv.opcode = FW_RI_RECV_WR;
923 		wqe->recv.r1 = 0;
924 		wqe->recv.wrid = qhp->wq.rq.pidx;
925 		wqe->recv.r2[0] = 0;
926 		wqe->recv.r2[1] = 0;
927 		wqe->recv.r2[2] = 0;
928 		wqe->recv.len16 = len16;
929 		PDBG("%s cookie 0x%llx pidx %u\n", __func__,
930 		     (unsigned long long) wr->wr_id, qhp->wq.rq.pidx);
931 		t4_rq_produce(&qhp->wq, len16);
932 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
933 		wr = wr->next;
934 		num_wrs--;
935 	}
936 	if (!qhp->rhp->rdev.status_page->db_off) {
937 		t4_ring_rq_db(&qhp->wq, idx, wqe);
938 		spin_unlock_irqrestore(&qhp->lock, flag);
939 	} else {
940 		spin_unlock_irqrestore(&qhp->lock, flag);
941 		ring_kernel_rq_db(qhp, idx);
942 	}
943 	return err;
944 }
945 
946 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
947 				    u8 *ecode)
948 {
949 	int status;
950 	int tagged;
951 	int opcode;
952 	int rqtype;
953 	int send_inv;
954 
955 	if (!err_cqe) {
956 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
957 		*ecode = 0;
958 		return;
959 	}
960 
961 	status = CQE_STATUS(err_cqe);
962 	opcode = CQE_OPCODE(err_cqe);
963 	rqtype = RQ_TYPE(err_cqe);
964 	send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
965 		   (opcode == FW_RI_SEND_WITH_SE_INV);
966 	tagged = (opcode == FW_RI_RDMA_WRITE) ||
967 		 (rqtype && (opcode == FW_RI_READ_RESP));
968 
969 	switch (status) {
970 	case T4_ERR_STAG:
971 		if (send_inv) {
972 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
973 			*ecode = RDMAP_CANT_INV_STAG;
974 		} else {
975 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
976 			*ecode = RDMAP_INV_STAG;
977 		}
978 		break;
979 	case T4_ERR_PDID:
980 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
981 		if ((opcode == FW_RI_SEND_WITH_INV) ||
982 		    (opcode == FW_RI_SEND_WITH_SE_INV))
983 			*ecode = RDMAP_CANT_INV_STAG;
984 		else
985 			*ecode = RDMAP_STAG_NOT_ASSOC;
986 		break;
987 	case T4_ERR_QPID:
988 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
989 		*ecode = RDMAP_STAG_NOT_ASSOC;
990 		break;
991 	case T4_ERR_ACCESS:
992 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
993 		*ecode = RDMAP_ACC_VIOL;
994 		break;
995 	case T4_ERR_WRAP:
996 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
997 		*ecode = RDMAP_TO_WRAP;
998 		break;
999 	case T4_ERR_BOUND:
1000 		if (tagged) {
1001 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1002 			*ecode = DDPT_BASE_BOUNDS;
1003 		} else {
1004 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1005 			*ecode = RDMAP_BASE_BOUNDS;
1006 		}
1007 		break;
1008 	case T4_ERR_INVALIDATE_SHARED_MR:
1009 	case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
1010 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1011 		*ecode = RDMAP_CANT_INV_STAG;
1012 		break;
1013 	case T4_ERR_ECC:
1014 	case T4_ERR_ECC_PSTAG:
1015 	case T4_ERR_INTERNAL_ERR:
1016 		*layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
1017 		*ecode = 0;
1018 		break;
1019 	case T4_ERR_OUT_OF_RQE:
1020 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1021 		*ecode = DDPU_INV_MSN_NOBUF;
1022 		break;
1023 	case T4_ERR_PBL_ADDR_BOUND:
1024 		*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1025 		*ecode = DDPT_BASE_BOUNDS;
1026 		break;
1027 	case T4_ERR_CRC:
1028 		*layer_type = LAYER_MPA|DDP_LLP;
1029 		*ecode = MPA_CRC_ERR;
1030 		break;
1031 	case T4_ERR_MARKER:
1032 		*layer_type = LAYER_MPA|DDP_LLP;
1033 		*ecode = MPA_MARKER_ERR;
1034 		break;
1035 	case T4_ERR_PDU_LEN_ERR:
1036 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1037 		*ecode = DDPU_MSG_TOOBIG;
1038 		break;
1039 	case T4_ERR_DDP_VERSION:
1040 		if (tagged) {
1041 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1042 			*ecode = DDPT_INV_VERS;
1043 		} else {
1044 			*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1045 			*ecode = DDPU_INV_VERS;
1046 		}
1047 		break;
1048 	case T4_ERR_RDMA_VERSION:
1049 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1050 		*ecode = RDMAP_INV_VERS;
1051 		break;
1052 	case T4_ERR_OPCODE:
1053 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1054 		*ecode = RDMAP_INV_OPCODE;
1055 		break;
1056 	case T4_ERR_DDP_QUEUE_NUM:
1057 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1058 		*ecode = DDPU_INV_QN;
1059 		break;
1060 	case T4_ERR_MSN:
1061 	case T4_ERR_MSN_GAP:
1062 	case T4_ERR_MSN_RANGE:
1063 	case T4_ERR_IRD_OVERFLOW:
1064 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1065 		*ecode = DDPU_INV_MSN_RANGE;
1066 		break;
1067 	case T4_ERR_TBIT:
1068 		*layer_type = LAYER_DDP|DDP_LOCAL_CATA;
1069 		*ecode = 0;
1070 		break;
1071 	case T4_ERR_MO:
1072 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1073 		*ecode = DDPU_INV_MO;
1074 		break;
1075 	default:
1076 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
1077 		*ecode = 0;
1078 		break;
1079 	}
1080 }
1081 
1082 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
1083 			   gfp_t gfp)
1084 {
1085 	struct fw_ri_wr *wqe;
1086 	struct sk_buff *skb;
1087 	struct terminate_message *term;
1088 
1089 	PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1090 	     qhp->ep->hwtid);
1091 
1092 	skb = skb_dequeue(&qhp->ep->com.ep_skb_list);
1093 	if (WARN_ON(!skb))
1094 		return;
1095 
1096 	set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1097 
1098 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1099 	memset(wqe, 0, sizeof *wqe);
1100 	wqe->op_compl = cpu_to_be32(FW_WR_OP_V(FW_RI_INIT_WR));
1101 	wqe->flowid_len16 = cpu_to_be32(
1102 		FW_WR_FLOWID_V(qhp->ep->hwtid) |
1103 		FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1104 
1105 	wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
1106 	wqe->u.terminate.immdlen = cpu_to_be32(sizeof *term);
1107 	term = (struct terminate_message *)wqe->u.terminate.termmsg;
1108 	if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
1109 		term->layer_etype = qhp->attr.layer_etype;
1110 		term->ecode = qhp->attr.ecode;
1111 	} else
1112 		build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
1113 	c4iw_ofld_send(&qhp->rhp->rdev, skb);
1114 }
1115 
1116 /*
1117  * Assumes qhp lock is held.
1118  */
1119 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
1120 		       struct c4iw_cq *schp)
1121 {
1122 	int count;
1123 	int rq_flushed, sq_flushed;
1124 	unsigned long flag;
1125 
1126 	PDBG("%s qhp %p rchp %p schp %p\n", __func__, qhp, rchp, schp);
1127 
1128 	/* locking hierarchy: cq lock first, then qp lock. */
1129 	spin_lock_irqsave(&rchp->lock, flag);
1130 	spin_lock(&qhp->lock);
1131 
1132 	if (qhp->wq.flushed) {
1133 		spin_unlock(&qhp->lock);
1134 		spin_unlock_irqrestore(&rchp->lock, flag);
1135 		return;
1136 	}
1137 	qhp->wq.flushed = 1;
1138 
1139 	c4iw_flush_hw_cq(rchp);
1140 	c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
1141 	rq_flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
1142 	spin_unlock(&qhp->lock);
1143 	spin_unlock_irqrestore(&rchp->lock, flag);
1144 
1145 	/* locking hierarchy: cq lock first, then qp lock. */
1146 	spin_lock_irqsave(&schp->lock, flag);
1147 	spin_lock(&qhp->lock);
1148 	if (schp != rchp)
1149 		c4iw_flush_hw_cq(schp);
1150 	sq_flushed = c4iw_flush_sq(qhp);
1151 	spin_unlock(&qhp->lock);
1152 	spin_unlock_irqrestore(&schp->lock, flag);
1153 
1154 	if (schp == rchp) {
1155 		if (t4_clear_cq_armed(&rchp->cq) &&
1156 		    (rq_flushed || sq_flushed)) {
1157 			spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1158 			(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1159 						   rchp->ibcq.cq_context);
1160 			spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1161 		}
1162 	} else {
1163 		if (t4_clear_cq_armed(&rchp->cq) && rq_flushed) {
1164 			spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1165 			(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1166 						   rchp->ibcq.cq_context);
1167 			spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1168 		}
1169 		if (t4_clear_cq_armed(&schp->cq) && sq_flushed) {
1170 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1171 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1172 						   schp->ibcq.cq_context);
1173 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1174 		}
1175 	}
1176 }
1177 
1178 static void flush_qp(struct c4iw_qp *qhp)
1179 {
1180 	struct c4iw_cq *rchp, *schp;
1181 	unsigned long flag;
1182 
1183 	rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
1184 	schp = to_c4iw_cq(qhp->ibqp.send_cq);
1185 
1186 	t4_set_wq_in_error(&qhp->wq);
1187 	if (qhp->ibqp.uobject) {
1188 		t4_set_cq_in_error(&rchp->cq);
1189 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1190 		(*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1191 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1192 		if (schp != rchp) {
1193 			t4_set_cq_in_error(&schp->cq);
1194 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1195 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1196 					schp->ibcq.cq_context);
1197 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1198 		}
1199 		return;
1200 	}
1201 	__flush_qp(qhp, rchp, schp);
1202 }
1203 
1204 static int rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1205 		     struct c4iw_ep *ep)
1206 {
1207 	struct fw_ri_wr *wqe;
1208 	int ret;
1209 	struct sk_buff *skb;
1210 
1211 	PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1212 	     ep->hwtid);
1213 
1214 	skb = skb_dequeue(&ep->com.ep_skb_list);
1215 	if (WARN_ON(!skb))
1216 		return -ENOMEM;
1217 
1218 	set_wr_txq(skb, CPL_PRIORITY_DATA, ep->txq_idx);
1219 
1220 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1221 	memset(wqe, 0, sizeof *wqe);
1222 	wqe->op_compl = cpu_to_be32(
1223 		FW_WR_OP_V(FW_RI_INIT_WR) |
1224 		FW_WR_COMPL_F);
1225 	wqe->flowid_len16 = cpu_to_be32(
1226 		FW_WR_FLOWID_V(ep->hwtid) |
1227 		FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1228 	wqe->cookie = (uintptr_t)&ep->com.wr_wait;
1229 
1230 	wqe->u.fini.type = FW_RI_TYPE_FINI;
1231 	ret = c4iw_ofld_send(&rhp->rdev, skb);
1232 	if (ret)
1233 		goto out;
1234 
1235 	ret = c4iw_wait_for_reply(&rhp->rdev, &ep->com.wr_wait, qhp->ep->hwtid,
1236 			     qhp->wq.sq.qid, __func__);
1237 out:
1238 	PDBG("%s ret %d\n", __func__, ret);
1239 	return ret;
1240 }
1241 
1242 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1243 {
1244 	PDBG("%s p2p_type = %d\n", __func__, p2p_type);
1245 	memset(&init->u, 0, sizeof init->u);
1246 	switch (p2p_type) {
1247 	case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1248 		init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1249 		init->u.write.stag_sink = cpu_to_be32(1);
1250 		init->u.write.to_sink = cpu_to_be64(1);
1251 		init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1252 		init->u.write.len16 = DIV_ROUND_UP(sizeof init->u.write +
1253 						   sizeof(struct fw_ri_immd),
1254 						   16);
1255 		break;
1256 	case FW_RI_INIT_P2PTYPE_READ_REQ:
1257 		init->u.write.opcode = FW_RI_RDMA_READ_WR;
1258 		init->u.read.stag_src = cpu_to_be32(1);
1259 		init->u.read.to_src_lo = cpu_to_be32(1);
1260 		init->u.read.stag_sink = cpu_to_be32(1);
1261 		init->u.read.to_sink_lo = cpu_to_be32(1);
1262 		init->u.read.len16 = DIV_ROUND_UP(sizeof init->u.read, 16);
1263 		break;
1264 	}
1265 }
1266 
1267 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1268 {
1269 	struct fw_ri_wr *wqe;
1270 	int ret;
1271 	struct sk_buff *skb;
1272 
1273 	PDBG("%s qhp %p qid 0x%x tid %u ird %u ord %u\n", __func__, qhp,
1274 	     qhp->wq.sq.qid, qhp->ep->hwtid, qhp->ep->ird, qhp->ep->ord);
1275 
1276 	skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1277 	if (!skb) {
1278 		ret = -ENOMEM;
1279 		goto out;
1280 	}
1281 	ret = alloc_ird(rhp, qhp->attr.max_ird);
1282 	if (ret) {
1283 		qhp->attr.max_ird = 0;
1284 		kfree_skb(skb);
1285 		goto out;
1286 	}
1287 	set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1288 
1289 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1290 	memset(wqe, 0, sizeof *wqe);
1291 	wqe->op_compl = cpu_to_be32(
1292 		FW_WR_OP_V(FW_RI_INIT_WR) |
1293 		FW_WR_COMPL_F);
1294 	wqe->flowid_len16 = cpu_to_be32(
1295 		FW_WR_FLOWID_V(qhp->ep->hwtid) |
1296 		FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1297 
1298 	wqe->cookie = (uintptr_t)&qhp->ep->com.wr_wait;
1299 
1300 	wqe->u.init.type = FW_RI_TYPE_INIT;
1301 	wqe->u.init.mpareqbit_p2ptype =
1302 		FW_RI_WR_MPAREQBIT_V(qhp->attr.mpa_attr.initiator) |
1303 		FW_RI_WR_P2PTYPE_V(qhp->attr.mpa_attr.p2p_type);
1304 	wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1305 	if (qhp->attr.mpa_attr.recv_marker_enabled)
1306 		wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1307 	if (qhp->attr.mpa_attr.xmit_marker_enabled)
1308 		wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1309 	if (qhp->attr.mpa_attr.crc_enabled)
1310 		wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1311 
1312 	wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1313 			    FW_RI_QP_RDMA_WRITE_ENABLE |
1314 			    FW_RI_QP_BIND_ENABLE;
1315 	if (!qhp->ibqp.uobject)
1316 		wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1317 				     FW_RI_QP_STAG0_ENABLE;
1318 	wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1319 	wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1320 	wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1321 	wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1322 	wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1323 	wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1324 	wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1325 	wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1326 	wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1327 	wqe->u.init.iss = cpu_to_be32(qhp->ep->snd_seq);
1328 	wqe->u.init.irs = cpu_to_be32(qhp->ep->rcv_seq);
1329 	wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1330 	wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1331 					 rhp->rdev.lldi.vr->rq.start);
1332 	if (qhp->attr.mpa_attr.initiator)
1333 		build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1334 
1335 	ret = c4iw_ofld_send(&rhp->rdev, skb);
1336 	if (ret)
1337 		goto err1;
1338 
1339 	ret = c4iw_wait_for_reply(&rhp->rdev, &qhp->ep->com.wr_wait,
1340 				  qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
1341 	if (!ret)
1342 		goto out;
1343 err1:
1344 	free_ird(rhp, qhp->attr.max_ird);
1345 out:
1346 	PDBG("%s ret %d\n", __func__, ret);
1347 	return ret;
1348 }
1349 
1350 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1351 		   enum c4iw_qp_attr_mask mask,
1352 		   struct c4iw_qp_attributes *attrs,
1353 		   int internal)
1354 {
1355 	int ret = 0;
1356 	struct c4iw_qp_attributes newattr = qhp->attr;
1357 	int disconnect = 0;
1358 	int terminate = 0;
1359 	int abort = 0;
1360 	int free = 0;
1361 	struct c4iw_ep *ep = NULL;
1362 
1363 	PDBG("%s qhp %p sqid 0x%x rqid 0x%x ep %p state %d -> %d\n", __func__,
1364 	     qhp, qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep, qhp->attr.state,
1365 	     (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1366 
1367 	mutex_lock(&qhp->mutex);
1368 
1369 	/* Process attr changes if in IDLE */
1370 	if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1371 		if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1372 			ret = -EIO;
1373 			goto out;
1374 		}
1375 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1376 			newattr.enable_rdma_read = attrs->enable_rdma_read;
1377 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1378 			newattr.enable_rdma_write = attrs->enable_rdma_write;
1379 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1380 			newattr.enable_bind = attrs->enable_bind;
1381 		if (mask & C4IW_QP_ATTR_MAX_ORD) {
1382 			if (attrs->max_ord > c4iw_max_read_depth) {
1383 				ret = -EINVAL;
1384 				goto out;
1385 			}
1386 			newattr.max_ord = attrs->max_ord;
1387 		}
1388 		if (mask & C4IW_QP_ATTR_MAX_IRD) {
1389 			if (attrs->max_ird > cur_max_read_depth(rhp)) {
1390 				ret = -EINVAL;
1391 				goto out;
1392 			}
1393 			newattr.max_ird = attrs->max_ird;
1394 		}
1395 		qhp->attr = newattr;
1396 	}
1397 
1398 	if (mask & C4IW_QP_ATTR_SQ_DB) {
1399 		ret = ring_kernel_sq_db(qhp, attrs->sq_db_inc);
1400 		goto out;
1401 	}
1402 	if (mask & C4IW_QP_ATTR_RQ_DB) {
1403 		ret = ring_kernel_rq_db(qhp, attrs->rq_db_inc);
1404 		goto out;
1405 	}
1406 
1407 	if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1408 		goto out;
1409 	if (qhp->attr.state == attrs->next_state)
1410 		goto out;
1411 
1412 	switch (qhp->attr.state) {
1413 	case C4IW_QP_STATE_IDLE:
1414 		switch (attrs->next_state) {
1415 		case C4IW_QP_STATE_RTS:
1416 			if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1417 				ret = -EINVAL;
1418 				goto out;
1419 			}
1420 			if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1421 				ret = -EINVAL;
1422 				goto out;
1423 			}
1424 			qhp->attr.mpa_attr = attrs->mpa_attr;
1425 			qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1426 			qhp->ep = qhp->attr.llp_stream_handle;
1427 			set_state(qhp, C4IW_QP_STATE_RTS);
1428 
1429 			/*
1430 			 * Ref the endpoint here and deref when we
1431 			 * disassociate the endpoint from the QP.  This
1432 			 * happens in CLOSING->IDLE transition or *->ERROR
1433 			 * transition.
1434 			 */
1435 			c4iw_get_ep(&qhp->ep->com);
1436 			ret = rdma_init(rhp, qhp);
1437 			if (ret)
1438 				goto err;
1439 			break;
1440 		case C4IW_QP_STATE_ERROR:
1441 			set_state(qhp, C4IW_QP_STATE_ERROR);
1442 			flush_qp(qhp);
1443 			break;
1444 		default:
1445 			ret = -EINVAL;
1446 			goto out;
1447 		}
1448 		break;
1449 	case C4IW_QP_STATE_RTS:
1450 		switch (attrs->next_state) {
1451 		case C4IW_QP_STATE_CLOSING:
1452 			BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
1453 			t4_set_wq_in_error(&qhp->wq);
1454 			set_state(qhp, C4IW_QP_STATE_CLOSING);
1455 			ep = qhp->ep;
1456 			if (!internal) {
1457 				abort = 0;
1458 				disconnect = 1;
1459 				c4iw_get_ep(&qhp->ep->com);
1460 			}
1461 			ret = rdma_fini(rhp, qhp, ep);
1462 			if (ret)
1463 				goto err;
1464 			break;
1465 		case C4IW_QP_STATE_TERMINATE:
1466 			t4_set_wq_in_error(&qhp->wq);
1467 			set_state(qhp, C4IW_QP_STATE_TERMINATE);
1468 			qhp->attr.layer_etype = attrs->layer_etype;
1469 			qhp->attr.ecode = attrs->ecode;
1470 			ep = qhp->ep;
1471 			if (!internal) {
1472 				c4iw_get_ep(&qhp->ep->com);
1473 				terminate = 1;
1474 				disconnect = 1;
1475 			} else {
1476 				terminate = qhp->attr.send_term;
1477 				ret = rdma_fini(rhp, qhp, ep);
1478 				if (ret)
1479 					goto err;
1480 			}
1481 			break;
1482 		case C4IW_QP_STATE_ERROR:
1483 			t4_set_wq_in_error(&qhp->wq);
1484 			set_state(qhp, C4IW_QP_STATE_ERROR);
1485 			if (!internal) {
1486 				abort = 1;
1487 				disconnect = 1;
1488 				ep = qhp->ep;
1489 				c4iw_get_ep(&qhp->ep->com);
1490 			}
1491 			goto err;
1492 			break;
1493 		default:
1494 			ret = -EINVAL;
1495 			goto out;
1496 		}
1497 		break;
1498 	case C4IW_QP_STATE_CLOSING:
1499 		if (!internal) {
1500 			ret = -EINVAL;
1501 			goto out;
1502 		}
1503 		switch (attrs->next_state) {
1504 		case C4IW_QP_STATE_IDLE:
1505 			flush_qp(qhp);
1506 			set_state(qhp, C4IW_QP_STATE_IDLE);
1507 			qhp->attr.llp_stream_handle = NULL;
1508 			c4iw_put_ep(&qhp->ep->com);
1509 			qhp->ep = NULL;
1510 			wake_up(&qhp->wait);
1511 			break;
1512 		case C4IW_QP_STATE_ERROR:
1513 			goto err;
1514 		default:
1515 			ret = -EINVAL;
1516 			goto err;
1517 		}
1518 		break;
1519 	case C4IW_QP_STATE_ERROR:
1520 		if (attrs->next_state != C4IW_QP_STATE_IDLE) {
1521 			ret = -EINVAL;
1522 			goto out;
1523 		}
1524 		if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) {
1525 			ret = -EINVAL;
1526 			goto out;
1527 		}
1528 		set_state(qhp, C4IW_QP_STATE_IDLE);
1529 		break;
1530 	case C4IW_QP_STATE_TERMINATE:
1531 		if (!internal) {
1532 			ret = -EINVAL;
1533 			goto out;
1534 		}
1535 		goto err;
1536 		break;
1537 	default:
1538 		printk(KERN_ERR "%s in a bad state %d\n",
1539 		       __func__, qhp->attr.state);
1540 		ret = -EINVAL;
1541 		goto err;
1542 		break;
1543 	}
1544 	goto out;
1545 err:
1546 	PDBG("%s disassociating ep %p qpid 0x%x\n", __func__, qhp->ep,
1547 	     qhp->wq.sq.qid);
1548 
1549 	/* disassociate the LLP connection */
1550 	qhp->attr.llp_stream_handle = NULL;
1551 	if (!ep)
1552 		ep = qhp->ep;
1553 	qhp->ep = NULL;
1554 	set_state(qhp, C4IW_QP_STATE_ERROR);
1555 	free = 1;
1556 	abort = 1;
1557 	BUG_ON(!ep);
1558 	flush_qp(qhp);
1559 	wake_up(&qhp->wait);
1560 out:
1561 	mutex_unlock(&qhp->mutex);
1562 
1563 	if (terminate)
1564 		post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
1565 
1566 	/*
1567 	 * If disconnect is 1, then we need to initiate a disconnect
1568 	 * on the EP.  This can be a normal close (RTS->CLOSING) or
1569 	 * an abnormal close (RTS/CLOSING->ERROR).
1570 	 */
1571 	if (disconnect) {
1572 		c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC :
1573 							 GFP_KERNEL);
1574 		c4iw_put_ep(&ep->com);
1575 	}
1576 
1577 	/*
1578 	 * If free is 1, then we've disassociated the EP from the QP
1579 	 * and we need to dereference the EP.
1580 	 */
1581 	if (free)
1582 		c4iw_put_ep(&ep->com);
1583 	PDBG("%s exit state %d\n", __func__, qhp->attr.state);
1584 	return ret;
1585 }
1586 
1587 int c4iw_destroy_qp(struct ib_qp *ib_qp)
1588 {
1589 	struct c4iw_dev *rhp;
1590 	struct c4iw_qp *qhp;
1591 	struct c4iw_qp_attributes attrs;
1592 	struct c4iw_ucontext *ucontext;
1593 
1594 	qhp = to_c4iw_qp(ib_qp);
1595 	rhp = qhp->rhp;
1596 
1597 	attrs.next_state = C4IW_QP_STATE_ERROR;
1598 	if (qhp->attr.state == C4IW_QP_STATE_TERMINATE)
1599 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
1600 	else
1601 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
1602 	wait_event(qhp->wait, !qhp->ep);
1603 
1604 	remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1605 
1606 	spin_lock_irq(&rhp->lock);
1607 	if (!list_empty(&qhp->db_fc_entry))
1608 		list_del_init(&qhp->db_fc_entry);
1609 	spin_unlock_irq(&rhp->lock);
1610 	free_ird(rhp, qhp->attr.max_ird);
1611 
1612 	ucontext = ib_qp->uobject ?
1613 		   to_c4iw_ucontext(ib_qp->uobject->context) : NULL;
1614 	destroy_qp(&rhp->rdev, &qhp->wq,
1615 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1616 
1617 	c4iw_qp_rem_ref(ib_qp);
1618 
1619 	PDBG("%s ib_qp %p qpid 0x%0x\n", __func__, ib_qp, qhp->wq.sq.qid);
1620 	return 0;
1621 }
1622 
1623 struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
1624 			     struct ib_udata *udata)
1625 {
1626 	struct c4iw_dev *rhp;
1627 	struct c4iw_qp *qhp;
1628 	struct c4iw_pd *php;
1629 	struct c4iw_cq *schp;
1630 	struct c4iw_cq *rchp;
1631 	struct c4iw_create_qp_resp uresp;
1632 	unsigned int sqsize, rqsize;
1633 	struct c4iw_ucontext *ucontext;
1634 	int ret;
1635 	struct c4iw_mm_entry *sq_key_mm, *rq_key_mm = NULL, *sq_db_key_mm;
1636 	struct c4iw_mm_entry *rq_db_key_mm = NULL, *ma_sync_key_mm = NULL;
1637 
1638 	PDBG("%s ib_pd %p\n", __func__, pd);
1639 
1640 	if (attrs->qp_type != IB_QPT_RC)
1641 		return ERR_PTR(-EINVAL);
1642 
1643 	php = to_c4iw_pd(pd);
1644 	rhp = php->rhp;
1645 	schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
1646 	rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
1647 	if (!schp || !rchp)
1648 		return ERR_PTR(-EINVAL);
1649 
1650 	if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
1651 		return ERR_PTR(-EINVAL);
1652 
1653 	if (attrs->cap.max_recv_wr > rhp->rdev.hw_queue.t4_max_rq_size)
1654 		return ERR_PTR(-E2BIG);
1655 	rqsize = attrs->cap.max_recv_wr + 1;
1656 	if (rqsize < 8)
1657 		rqsize = 8;
1658 
1659 	if (attrs->cap.max_send_wr > rhp->rdev.hw_queue.t4_max_sq_size)
1660 		return ERR_PTR(-E2BIG);
1661 	sqsize = attrs->cap.max_send_wr + 1;
1662 	if (sqsize < 8)
1663 		sqsize = 8;
1664 
1665 	ucontext = pd->uobject ? to_c4iw_ucontext(pd->uobject->context) : NULL;
1666 
1667 	qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
1668 	if (!qhp)
1669 		return ERR_PTR(-ENOMEM);
1670 	qhp->wq.sq.size = sqsize;
1671 	qhp->wq.sq.memsize =
1672 		(sqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1673 		sizeof(*qhp->wq.sq.queue) + 16 * sizeof(__be64);
1674 	qhp->wq.sq.flush_cidx = -1;
1675 	qhp->wq.rq.size = rqsize;
1676 	qhp->wq.rq.memsize =
1677 		(rqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1678 		sizeof(*qhp->wq.rq.queue);
1679 
1680 	if (ucontext) {
1681 		qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
1682 		qhp->wq.rq.memsize = roundup(qhp->wq.rq.memsize, PAGE_SIZE);
1683 	}
1684 
1685 	ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
1686 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1687 	if (ret)
1688 		goto err1;
1689 
1690 	attrs->cap.max_recv_wr = rqsize - 1;
1691 	attrs->cap.max_send_wr = sqsize - 1;
1692 	attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
1693 
1694 	qhp->rhp = rhp;
1695 	qhp->attr.pd = php->pdid;
1696 	qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
1697 	qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
1698 	qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
1699 	qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
1700 	qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
1701 	qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
1702 	qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
1703 	qhp->attr.state = C4IW_QP_STATE_IDLE;
1704 	qhp->attr.next_state = C4IW_QP_STATE_IDLE;
1705 	qhp->attr.enable_rdma_read = 1;
1706 	qhp->attr.enable_rdma_write = 1;
1707 	qhp->attr.enable_bind = 1;
1708 	qhp->attr.max_ord = 0;
1709 	qhp->attr.max_ird = 0;
1710 	qhp->sq_sig_all = attrs->sq_sig_type == IB_SIGNAL_ALL_WR;
1711 	spin_lock_init(&qhp->lock);
1712 	init_completion(&qhp->sq_drained);
1713 	init_completion(&qhp->rq_drained);
1714 	mutex_init(&qhp->mutex);
1715 	init_waitqueue_head(&qhp->wait);
1716 	kref_init(&qhp->kref);
1717 
1718 	ret = insert_handle(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid);
1719 	if (ret)
1720 		goto err2;
1721 
1722 	if (udata) {
1723 		sq_key_mm = kmalloc(sizeof(*sq_key_mm), GFP_KERNEL);
1724 		if (!sq_key_mm) {
1725 			ret = -ENOMEM;
1726 			goto err3;
1727 		}
1728 		rq_key_mm = kmalloc(sizeof(*rq_key_mm), GFP_KERNEL);
1729 		if (!rq_key_mm) {
1730 			ret = -ENOMEM;
1731 			goto err4;
1732 		}
1733 		sq_db_key_mm = kmalloc(sizeof(*sq_db_key_mm), GFP_KERNEL);
1734 		if (!sq_db_key_mm) {
1735 			ret = -ENOMEM;
1736 			goto err5;
1737 		}
1738 		rq_db_key_mm = kmalloc(sizeof(*rq_db_key_mm), GFP_KERNEL);
1739 		if (!rq_db_key_mm) {
1740 			ret = -ENOMEM;
1741 			goto err6;
1742 		}
1743 		if (t4_sq_onchip(&qhp->wq.sq)) {
1744 			ma_sync_key_mm = kmalloc(sizeof(*ma_sync_key_mm),
1745 						 GFP_KERNEL);
1746 			if (!ma_sync_key_mm) {
1747 				ret = -ENOMEM;
1748 				goto err7;
1749 			}
1750 			uresp.flags = C4IW_QPF_ONCHIP;
1751 		} else
1752 			uresp.flags = 0;
1753 		uresp.qid_mask = rhp->rdev.qpmask;
1754 		uresp.sqid = qhp->wq.sq.qid;
1755 		uresp.sq_size = qhp->wq.sq.size;
1756 		uresp.sq_memsize = qhp->wq.sq.memsize;
1757 		uresp.rqid = qhp->wq.rq.qid;
1758 		uresp.rq_size = qhp->wq.rq.size;
1759 		uresp.rq_memsize = qhp->wq.rq.memsize;
1760 		spin_lock(&ucontext->mmap_lock);
1761 		if (ma_sync_key_mm) {
1762 			uresp.ma_sync_key = ucontext->key;
1763 			ucontext->key += PAGE_SIZE;
1764 		} else {
1765 			uresp.ma_sync_key =  0;
1766 		}
1767 		uresp.sq_key = ucontext->key;
1768 		ucontext->key += PAGE_SIZE;
1769 		uresp.rq_key = ucontext->key;
1770 		ucontext->key += PAGE_SIZE;
1771 		uresp.sq_db_gts_key = ucontext->key;
1772 		ucontext->key += PAGE_SIZE;
1773 		uresp.rq_db_gts_key = ucontext->key;
1774 		ucontext->key += PAGE_SIZE;
1775 		spin_unlock(&ucontext->mmap_lock);
1776 		ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
1777 		if (ret)
1778 			goto err8;
1779 		sq_key_mm->key = uresp.sq_key;
1780 		sq_key_mm->addr = qhp->wq.sq.phys_addr;
1781 		sq_key_mm->len = PAGE_ALIGN(qhp->wq.sq.memsize);
1782 		insert_mmap(ucontext, sq_key_mm);
1783 		rq_key_mm->key = uresp.rq_key;
1784 		rq_key_mm->addr = virt_to_phys(qhp->wq.rq.queue);
1785 		rq_key_mm->len = PAGE_ALIGN(qhp->wq.rq.memsize);
1786 		insert_mmap(ucontext, rq_key_mm);
1787 		sq_db_key_mm->key = uresp.sq_db_gts_key;
1788 		sq_db_key_mm->addr = (u64)(unsigned long)qhp->wq.sq.bar2_pa;
1789 		sq_db_key_mm->len = PAGE_SIZE;
1790 		insert_mmap(ucontext, sq_db_key_mm);
1791 		rq_db_key_mm->key = uresp.rq_db_gts_key;
1792 		rq_db_key_mm->addr = (u64)(unsigned long)qhp->wq.rq.bar2_pa;
1793 		rq_db_key_mm->len = PAGE_SIZE;
1794 		insert_mmap(ucontext, rq_db_key_mm);
1795 		if (ma_sync_key_mm) {
1796 			ma_sync_key_mm->key = uresp.ma_sync_key;
1797 			ma_sync_key_mm->addr =
1798 				(pci_resource_start(rhp->rdev.lldi.pdev, 0) +
1799 				PCIE_MA_SYNC_A) & PAGE_MASK;
1800 			ma_sync_key_mm->len = PAGE_SIZE;
1801 			insert_mmap(ucontext, ma_sync_key_mm);
1802 		}
1803 	}
1804 	qhp->ibqp.qp_num = qhp->wq.sq.qid;
1805 	init_timer(&(qhp->timer));
1806 	INIT_LIST_HEAD(&qhp->db_fc_entry);
1807 	PDBG("%s sq id %u size %u memsize %zu num_entries %u "
1808 	     "rq id %u size %u memsize %zu num_entries %u\n", __func__,
1809 	     qhp->wq.sq.qid, qhp->wq.sq.size, qhp->wq.sq.memsize,
1810 	     attrs->cap.max_send_wr, qhp->wq.rq.qid, qhp->wq.rq.size,
1811 	     qhp->wq.rq.memsize, attrs->cap.max_recv_wr);
1812 	return &qhp->ibqp;
1813 err8:
1814 	kfree(ma_sync_key_mm);
1815 err7:
1816 	kfree(rq_db_key_mm);
1817 err6:
1818 	kfree(sq_db_key_mm);
1819 err5:
1820 	kfree(rq_key_mm);
1821 err4:
1822 	kfree(sq_key_mm);
1823 err3:
1824 	remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1825 err2:
1826 	destroy_qp(&rhp->rdev, &qhp->wq,
1827 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1828 err1:
1829 	kfree(qhp);
1830 	return ERR_PTR(ret);
1831 }
1832 
1833 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1834 		      int attr_mask, struct ib_udata *udata)
1835 {
1836 	struct c4iw_dev *rhp;
1837 	struct c4iw_qp *qhp;
1838 	enum c4iw_qp_attr_mask mask = 0;
1839 	struct c4iw_qp_attributes attrs;
1840 
1841 	PDBG("%s ib_qp %p\n", __func__, ibqp);
1842 
1843 	/* iwarp does not support the RTR state */
1844 	if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
1845 		attr_mask &= ~IB_QP_STATE;
1846 
1847 	/* Make sure we still have something left to do */
1848 	if (!attr_mask)
1849 		return 0;
1850 
1851 	memset(&attrs, 0, sizeof attrs);
1852 	qhp = to_c4iw_qp(ibqp);
1853 	rhp = qhp->rhp;
1854 
1855 	attrs.next_state = c4iw_convert_state(attr->qp_state);
1856 	attrs.enable_rdma_read = (attr->qp_access_flags &
1857 			       IB_ACCESS_REMOTE_READ) ?  1 : 0;
1858 	attrs.enable_rdma_write = (attr->qp_access_flags &
1859 				IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
1860 	attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
1861 
1862 
1863 	mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
1864 	mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
1865 			(C4IW_QP_ATTR_ENABLE_RDMA_READ |
1866 			 C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
1867 			 C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
1868 
1869 	/*
1870 	 * Use SQ_PSN and RQ_PSN to pass in IDX_INC values for
1871 	 * ringing the queue db when we're in DB_FULL mode.
1872 	 * Only allow this on T4 devices.
1873 	 */
1874 	attrs.sq_db_inc = attr->sq_psn;
1875 	attrs.rq_db_inc = attr->rq_psn;
1876 	mask |= (attr_mask & IB_QP_SQ_PSN) ? C4IW_QP_ATTR_SQ_DB : 0;
1877 	mask |= (attr_mask & IB_QP_RQ_PSN) ? C4IW_QP_ATTR_RQ_DB : 0;
1878 	if (!is_t4(to_c4iw_qp(ibqp)->rhp->rdev.lldi.adapter_type) &&
1879 	    (mask & (C4IW_QP_ATTR_SQ_DB|C4IW_QP_ATTR_RQ_DB)))
1880 		return -EINVAL;
1881 
1882 	return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
1883 }
1884 
1885 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
1886 {
1887 	PDBG("%s ib_dev %p qpn 0x%x\n", __func__, dev, qpn);
1888 	return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
1889 }
1890 
1891 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1892 		     int attr_mask, struct ib_qp_init_attr *init_attr)
1893 {
1894 	struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
1895 
1896 	memset(attr, 0, sizeof *attr);
1897 	memset(init_attr, 0, sizeof *init_attr);
1898 	attr->qp_state = to_ib_qp_state(qhp->attr.state);
1899 	init_attr->cap.max_send_wr = qhp->attr.sq_num_entries;
1900 	init_attr->cap.max_recv_wr = qhp->attr.rq_num_entries;
1901 	init_attr->cap.max_send_sge = qhp->attr.sq_max_sges;
1902 	init_attr->cap.max_recv_sge = qhp->attr.sq_max_sges;
1903 	init_attr->cap.max_inline_data = T4_MAX_SEND_INLINE;
1904 	init_attr->sq_sig_type = qhp->sq_sig_all ? IB_SIGNAL_ALL_WR : 0;
1905 	return 0;
1906 }
1907 
1908 static void move_qp_to_err(struct c4iw_qp *qp)
1909 {
1910 	struct c4iw_qp_attributes attrs = { .next_state = C4IW_QP_STATE_ERROR };
1911 
1912 	(void)c4iw_modify_qp(qp->rhp, qp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
1913 }
1914 
1915 void c4iw_drain_sq(struct ib_qp *ibqp)
1916 {
1917 	struct c4iw_qp *qp = to_c4iw_qp(ibqp);
1918 	unsigned long flag;
1919 	bool need_to_wait;
1920 
1921 	move_qp_to_err(qp);
1922 	spin_lock_irqsave(&qp->lock, flag);
1923 	need_to_wait = !t4_sq_empty(&qp->wq);
1924 	spin_unlock_irqrestore(&qp->lock, flag);
1925 
1926 	if (need_to_wait)
1927 		wait_for_completion(&qp->sq_drained);
1928 }
1929 
1930 void c4iw_drain_rq(struct ib_qp *ibqp)
1931 {
1932 	struct c4iw_qp *qp = to_c4iw_qp(ibqp);
1933 	unsigned long flag;
1934 	bool need_to_wait;
1935 
1936 	move_qp_to_err(qp);
1937 	spin_lock_irqsave(&qp->lock, flag);
1938 	need_to_wait = !t4_rq_empty(&qp->wq);
1939 	spin_unlock_irqrestore(&qp->lock, flag);
1940 
1941 	if (need_to_wait)
1942 		wait_for_completion(&qp->rq_drained);
1943 }
1944