xref: /linux/drivers/infiniband/hw/cxgb4/cq.c (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
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 "iw_cxgb4.h"
34 
35 static int destroy_cq(struct c4iw_rdev *rdev, struct t4_cq *cq,
36 		      struct c4iw_dev_ucontext *uctx)
37 {
38 	struct fw_ri_res_wr *res_wr;
39 	struct fw_ri_res *res;
40 	int wr_len;
41 	struct c4iw_wr_wait wr_wait;
42 	struct sk_buff *skb;
43 	int ret;
44 
45 	wr_len = sizeof *res_wr + sizeof *res;
46 	skb = alloc_skb(wr_len, GFP_KERNEL);
47 	if (!skb)
48 		return -ENOMEM;
49 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
50 
51 	res_wr = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
52 	memset(res_wr, 0, wr_len);
53 	res_wr->op_nres = cpu_to_be32(
54 			FW_WR_OP_V(FW_RI_RES_WR) |
55 			FW_RI_RES_WR_NRES_V(1) |
56 			FW_WR_COMPL_F);
57 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
58 	res_wr->cookie = (uintptr_t)&wr_wait;
59 	res = res_wr->res;
60 	res->u.cq.restype = FW_RI_RES_TYPE_CQ;
61 	res->u.cq.op = FW_RI_RES_OP_RESET;
62 	res->u.cq.iqid = cpu_to_be32(cq->cqid);
63 
64 	c4iw_init_wr_wait(&wr_wait);
65 	ret = c4iw_ofld_send(rdev, skb);
66 	if (!ret) {
67 		ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, __func__);
68 	}
69 
70 	kfree(cq->sw_queue);
71 	dma_free_coherent(&(rdev->lldi.pdev->dev),
72 			  cq->memsize, cq->queue,
73 			  dma_unmap_addr(cq, mapping));
74 	c4iw_put_cqid(rdev, cq->cqid, uctx);
75 	return ret;
76 }
77 
78 static int create_cq(struct c4iw_rdev *rdev, struct t4_cq *cq,
79 		     struct c4iw_dev_ucontext *uctx)
80 {
81 	struct fw_ri_res_wr *res_wr;
82 	struct fw_ri_res *res;
83 	int wr_len;
84 	int user = (uctx != &rdev->uctx);
85 	struct c4iw_wr_wait wr_wait;
86 	int ret;
87 	struct sk_buff *skb;
88 
89 	cq->cqid = c4iw_get_cqid(rdev, uctx);
90 	if (!cq->cqid) {
91 		ret = -ENOMEM;
92 		goto err1;
93 	}
94 
95 	if (!user) {
96 		cq->sw_queue = kzalloc(cq->memsize, GFP_KERNEL);
97 		if (!cq->sw_queue) {
98 			ret = -ENOMEM;
99 			goto err2;
100 		}
101 	}
102 	cq->queue = dma_alloc_coherent(&rdev->lldi.pdev->dev, cq->memsize,
103 				       &cq->dma_addr, GFP_KERNEL);
104 	if (!cq->queue) {
105 		ret = -ENOMEM;
106 		goto err3;
107 	}
108 	dma_unmap_addr_set(cq, mapping, cq->dma_addr);
109 	memset(cq->queue, 0, cq->memsize);
110 
111 	/* build fw_ri_res_wr */
112 	wr_len = sizeof *res_wr + sizeof *res;
113 
114 	skb = alloc_skb(wr_len, GFP_KERNEL);
115 	if (!skb) {
116 		ret = -ENOMEM;
117 		goto err4;
118 	}
119 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
120 
121 	res_wr = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
122 	memset(res_wr, 0, wr_len);
123 	res_wr->op_nres = cpu_to_be32(
124 			FW_WR_OP_V(FW_RI_RES_WR) |
125 			FW_RI_RES_WR_NRES_V(1) |
126 			FW_WR_COMPL_F);
127 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
128 	res_wr->cookie = (uintptr_t)&wr_wait;
129 	res = res_wr->res;
130 	res->u.cq.restype = FW_RI_RES_TYPE_CQ;
131 	res->u.cq.op = FW_RI_RES_OP_WRITE;
132 	res->u.cq.iqid = cpu_to_be32(cq->cqid);
133 	res->u.cq.iqandst_to_iqandstindex = cpu_to_be32(
134 			FW_RI_RES_WR_IQANUS_V(0) |
135 			FW_RI_RES_WR_IQANUD_V(1) |
136 			FW_RI_RES_WR_IQANDST_F |
137 			FW_RI_RES_WR_IQANDSTINDEX_V(
138 				rdev->lldi.ciq_ids[cq->vector]));
139 	res->u.cq.iqdroprss_to_iqesize = cpu_to_be16(
140 			FW_RI_RES_WR_IQDROPRSS_F |
141 			FW_RI_RES_WR_IQPCIECH_V(2) |
142 			FW_RI_RES_WR_IQINTCNTTHRESH_V(0) |
143 			FW_RI_RES_WR_IQO_F |
144 			FW_RI_RES_WR_IQESIZE_V(1));
145 	res->u.cq.iqsize = cpu_to_be16(cq->size);
146 	res->u.cq.iqaddr = cpu_to_be64(cq->dma_addr);
147 
148 	c4iw_init_wr_wait(&wr_wait);
149 
150 	ret = c4iw_ofld_send(rdev, skb);
151 	if (ret)
152 		goto err4;
153 	PDBG("%s wait_event wr_wait %p\n", __func__, &wr_wait);
154 	ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, __func__);
155 	if (ret)
156 		goto err4;
157 
158 	cq->gen = 1;
159 	cq->gts = rdev->lldi.gts_reg;
160 	cq->rdev = rdev;
161 
162 	cq->bar2_va = c4iw_bar2_addrs(rdev, cq->cqid, T4_BAR2_QTYPE_INGRESS,
163 				      &cq->bar2_qid,
164 				      user ? &cq->bar2_pa : NULL);
165 	if (user && !cq->bar2_va) {
166 		pr_warn(MOD "%s: cqid %u not in BAR2 range.\n",
167 			pci_name(rdev->lldi.pdev), cq->cqid);
168 		ret = -EINVAL;
169 		goto err4;
170 	}
171 	return 0;
172 err4:
173 	dma_free_coherent(&rdev->lldi.pdev->dev, cq->memsize, cq->queue,
174 			  dma_unmap_addr(cq, mapping));
175 err3:
176 	kfree(cq->sw_queue);
177 err2:
178 	c4iw_put_cqid(rdev, cq->cqid, uctx);
179 err1:
180 	return ret;
181 }
182 
183 static void insert_recv_cqe(struct t4_wq *wq, struct t4_cq *cq)
184 {
185 	struct t4_cqe cqe;
186 
187 	PDBG("%s wq %p cq %p sw_cidx %u sw_pidx %u\n", __func__,
188 	     wq, cq, cq->sw_cidx, cq->sw_pidx);
189 	memset(&cqe, 0, sizeof(cqe));
190 	cqe.header = cpu_to_be32(CQE_STATUS_V(T4_ERR_SWFLUSH) |
191 				 CQE_OPCODE_V(FW_RI_SEND) |
192 				 CQE_TYPE_V(0) |
193 				 CQE_SWCQE_V(1) |
194 				 CQE_QPID_V(wq->sq.qid));
195 	cqe.bits_type_ts = cpu_to_be64(CQE_GENBIT_V((u64)cq->gen));
196 	cq->sw_queue[cq->sw_pidx] = cqe;
197 	t4_swcq_produce(cq);
198 }
199 
200 int c4iw_flush_rq(struct t4_wq *wq, struct t4_cq *cq, int count)
201 {
202 	int flushed = 0;
203 	int in_use = wq->rq.in_use - count;
204 
205 	BUG_ON(in_use < 0);
206 	PDBG("%s wq %p cq %p rq.in_use %u skip count %u\n", __func__,
207 	     wq, cq, wq->rq.in_use, count);
208 	while (in_use--) {
209 		insert_recv_cqe(wq, cq);
210 		flushed++;
211 	}
212 	return flushed;
213 }
214 
215 static void insert_sq_cqe(struct t4_wq *wq, struct t4_cq *cq,
216 			  struct t4_swsqe *swcqe)
217 {
218 	struct t4_cqe cqe;
219 
220 	PDBG("%s wq %p cq %p sw_cidx %u sw_pidx %u\n", __func__,
221 	     wq, cq, cq->sw_cidx, cq->sw_pidx);
222 	memset(&cqe, 0, sizeof(cqe));
223 	cqe.header = cpu_to_be32(CQE_STATUS_V(T4_ERR_SWFLUSH) |
224 				 CQE_OPCODE_V(swcqe->opcode) |
225 				 CQE_TYPE_V(1) |
226 				 CQE_SWCQE_V(1) |
227 				 CQE_QPID_V(wq->sq.qid));
228 	CQE_WRID_SQ_IDX(&cqe) = swcqe->idx;
229 	cqe.bits_type_ts = cpu_to_be64(CQE_GENBIT_V((u64)cq->gen));
230 	cq->sw_queue[cq->sw_pidx] = cqe;
231 	t4_swcq_produce(cq);
232 }
233 
234 static void advance_oldest_read(struct t4_wq *wq);
235 
236 int c4iw_flush_sq(struct c4iw_qp *qhp)
237 {
238 	int flushed = 0;
239 	struct t4_wq *wq = &qhp->wq;
240 	struct c4iw_cq *chp = to_c4iw_cq(qhp->ibqp.send_cq);
241 	struct t4_cq *cq = &chp->cq;
242 	int idx;
243 	struct t4_swsqe *swsqe;
244 
245 	if (wq->sq.flush_cidx == -1)
246 		wq->sq.flush_cidx = wq->sq.cidx;
247 	idx = wq->sq.flush_cidx;
248 	BUG_ON(idx >= wq->sq.size);
249 	while (idx != wq->sq.pidx) {
250 		swsqe = &wq->sq.sw_sq[idx];
251 		BUG_ON(swsqe->flushed);
252 		swsqe->flushed = 1;
253 		insert_sq_cqe(wq, cq, swsqe);
254 		if (wq->sq.oldest_read == swsqe) {
255 			BUG_ON(swsqe->opcode != FW_RI_READ_REQ);
256 			advance_oldest_read(wq);
257 		}
258 		flushed++;
259 		if (++idx == wq->sq.size)
260 			idx = 0;
261 	}
262 	wq->sq.flush_cidx += flushed;
263 	if (wq->sq.flush_cidx >= wq->sq.size)
264 		wq->sq.flush_cidx -= wq->sq.size;
265 	return flushed;
266 }
267 
268 static void flush_completed_wrs(struct t4_wq *wq, struct t4_cq *cq)
269 {
270 	struct t4_swsqe *swsqe;
271 	int cidx;
272 
273 	if (wq->sq.flush_cidx == -1)
274 		wq->sq.flush_cidx = wq->sq.cidx;
275 	cidx = wq->sq.flush_cidx;
276 	BUG_ON(cidx > wq->sq.size);
277 
278 	while (cidx != wq->sq.pidx) {
279 		swsqe = &wq->sq.sw_sq[cidx];
280 		if (!swsqe->signaled) {
281 			if (++cidx == wq->sq.size)
282 				cidx = 0;
283 		} else if (swsqe->complete) {
284 
285 			BUG_ON(swsqe->flushed);
286 
287 			/*
288 			 * Insert this completed cqe into the swcq.
289 			 */
290 			PDBG("%s moving cqe into swcq sq idx %u cq idx %u\n",
291 					__func__, cidx, cq->sw_pidx);
292 			swsqe->cqe.header |= htonl(CQE_SWCQE_V(1));
293 			cq->sw_queue[cq->sw_pidx] = swsqe->cqe;
294 			t4_swcq_produce(cq);
295 			swsqe->flushed = 1;
296 			if (++cidx == wq->sq.size)
297 				cidx = 0;
298 			wq->sq.flush_cidx = cidx;
299 		} else
300 			break;
301 	}
302 }
303 
304 static void create_read_req_cqe(struct t4_wq *wq, struct t4_cqe *hw_cqe,
305 		struct t4_cqe *read_cqe)
306 {
307 	read_cqe->u.scqe.cidx = wq->sq.oldest_read->idx;
308 	read_cqe->len = htonl(wq->sq.oldest_read->read_len);
309 	read_cqe->header = htonl(CQE_QPID_V(CQE_QPID(hw_cqe)) |
310 			CQE_SWCQE_V(SW_CQE(hw_cqe)) |
311 			CQE_OPCODE_V(FW_RI_READ_REQ) |
312 			CQE_TYPE_V(1));
313 	read_cqe->bits_type_ts = hw_cqe->bits_type_ts;
314 }
315 
316 static void advance_oldest_read(struct t4_wq *wq)
317 {
318 
319 	u32 rptr = wq->sq.oldest_read - wq->sq.sw_sq + 1;
320 
321 	if (rptr == wq->sq.size)
322 		rptr = 0;
323 	while (rptr != wq->sq.pidx) {
324 		wq->sq.oldest_read = &wq->sq.sw_sq[rptr];
325 
326 		if (wq->sq.oldest_read->opcode == FW_RI_READ_REQ)
327 			return;
328 		if (++rptr == wq->sq.size)
329 			rptr = 0;
330 	}
331 	wq->sq.oldest_read = NULL;
332 }
333 
334 /*
335  * Move all CQEs from the HWCQ into the SWCQ.
336  * Deal with out-of-order and/or completions that complete
337  * prior unsignalled WRs.
338  */
339 void c4iw_flush_hw_cq(struct c4iw_cq *chp)
340 {
341 	struct t4_cqe *hw_cqe, *swcqe, read_cqe;
342 	struct c4iw_qp *qhp;
343 	struct t4_swsqe *swsqe;
344 	int ret;
345 
346 	PDBG("%s  cqid 0x%x\n", __func__, chp->cq.cqid);
347 	ret = t4_next_hw_cqe(&chp->cq, &hw_cqe);
348 
349 	/*
350 	 * This logic is similar to poll_cq(), but not quite the same
351 	 * unfortunately.  Need to move pertinent HW CQEs to the SW CQ but
352 	 * also do any translation magic that poll_cq() normally does.
353 	 */
354 	while (!ret) {
355 		qhp = get_qhp(chp->rhp, CQE_QPID(hw_cqe));
356 
357 		/*
358 		 * drop CQEs with no associated QP
359 		 */
360 		if (qhp == NULL)
361 			goto next_cqe;
362 
363 		if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE)
364 			goto next_cqe;
365 
366 		if (CQE_OPCODE(hw_cqe) == FW_RI_READ_RESP) {
367 
368 			/* If we have reached here because of async
369 			 * event or other error, and have egress error
370 			 * then drop
371 			 */
372 			if (CQE_TYPE(hw_cqe) == 1)
373 				goto next_cqe;
374 
375 			/* drop peer2peer RTR reads.
376 			 */
377 			if (CQE_WRID_STAG(hw_cqe) == 1)
378 				goto next_cqe;
379 
380 			/*
381 			 * Eat completions for unsignaled read WRs.
382 			 */
383 			if (!qhp->wq.sq.oldest_read->signaled) {
384 				advance_oldest_read(&qhp->wq);
385 				goto next_cqe;
386 			}
387 
388 			/*
389 			 * Don't write to the HWCQ, create a new read req CQE
390 			 * in local memory and move it into the swcq.
391 			 */
392 			create_read_req_cqe(&qhp->wq, hw_cqe, &read_cqe);
393 			hw_cqe = &read_cqe;
394 			advance_oldest_read(&qhp->wq);
395 		}
396 
397 		/* if its a SQ completion, then do the magic to move all the
398 		 * unsignaled and now in-order completions into the swcq.
399 		 */
400 		if (SQ_TYPE(hw_cqe)) {
401 			swsqe = &qhp->wq.sq.sw_sq[CQE_WRID_SQ_IDX(hw_cqe)];
402 			swsqe->cqe = *hw_cqe;
403 			swsqe->complete = 1;
404 			flush_completed_wrs(&qhp->wq, &chp->cq);
405 		} else {
406 			swcqe = &chp->cq.sw_queue[chp->cq.sw_pidx];
407 			*swcqe = *hw_cqe;
408 			swcqe->header |= cpu_to_be32(CQE_SWCQE_V(1));
409 			t4_swcq_produce(&chp->cq);
410 		}
411 next_cqe:
412 		t4_hwcq_consume(&chp->cq);
413 		ret = t4_next_hw_cqe(&chp->cq, &hw_cqe);
414 	}
415 }
416 
417 static int cqe_completes_wr(struct t4_cqe *cqe, struct t4_wq *wq)
418 {
419 	if (CQE_OPCODE(cqe) == FW_RI_TERMINATE)
420 		return 0;
421 
422 	if ((CQE_OPCODE(cqe) == FW_RI_RDMA_WRITE) && RQ_TYPE(cqe))
423 		return 0;
424 
425 	if ((CQE_OPCODE(cqe) == FW_RI_READ_RESP) && SQ_TYPE(cqe))
426 		return 0;
427 
428 	if (CQE_SEND_OPCODE(cqe) && RQ_TYPE(cqe) && t4_rq_empty(wq))
429 		return 0;
430 	return 1;
431 }
432 
433 void c4iw_count_rcqes(struct t4_cq *cq, struct t4_wq *wq, int *count)
434 {
435 	struct t4_cqe *cqe;
436 	u32 ptr;
437 
438 	*count = 0;
439 	PDBG("%s count zero %d\n", __func__, *count);
440 	ptr = cq->sw_cidx;
441 	while (ptr != cq->sw_pidx) {
442 		cqe = &cq->sw_queue[ptr];
443 		if (RQ_TYPE(cqe) && (CQE_OPCODE(cqe) != FW_RI_READ_RESP) &&
444 		    (CQE_QPID(cqe) == wq->sq.qid) && cqe_completes_wr(cqe, wq))
445 			(*count)++;
446 		if (++ptr == cq->size)
447 			ptr = 0;
448 	}
449 	PDBG("%s cq %p count %d\n", __func__, cq, *count);
450 }
451 
452 /*
453  * poll_cq
454  *
455  * Caller must:
456  *     check the validity of the first CQE,
457  *     supply the wq assicated with the qpid.
458  *
459  * credit: cq credit to return to sge.
460  * cqe_flushed: 1 iff the CQE is flushed.
461  * cqe: copy of the polled CQE.
462  *
463  * return value:
464  *    0		    CQE returned ok.
465  *    -EAGAIN       CQE skipped, try again.
466  *    -EOVERFLOW    CQ overflow detected.
467  */
468 static int poll_cq(struct t4_wq *wq, struct t4_cq *cq, struct t4_cqe *cqe,
469 		   u8 *cqe_flushed, u64 *cookie, u32 *credit)
470 {
471 	int ret = 0;
472 	struct t4_cqe *hw_cqe, read_cqe;
473 
474 	*cqe_flushed = 0;
475 	*credit = 0;
476 	ret = t4_next_cqe(cq, &hw_cqe);
477 	if (ret)
478 		return ret;
479 
480 	PDBG("%s CQE OVF %u qpid 0x%0x genbit %u type %u status 0x%0x"
481 	     " opcode 0x%0x len 0x%0x wrid_hi_stag 0x%x wrid_low_msn 0x%x\n",
482 	     __func__, CQE_OVFBIT(hw_cqe), CQE_QPID(hw_cqe),
483 	     CQE_GENBIT(hw_cqe), CQE_TYPE(hw_cqe), CQE_STATUS(hw_cqe),
484 	     CQE_OPCODE(hw_cqe), CQE_LEN(hw_cqe), CQE_WRID_HI(hw_cqe),
485 	     CQE_WRID_LOW(hw_cqe));
486 
487 	/*
488 	 * skip cqe's not affiliated with a QP.
489 	 */
490 	if (wq == NULL) {
491 		ret = -EAGAIN;
492 		goto skip_cqe;
493 	}
494 
495 	/*
496 	* skip hw cqe's if the wq is flushed.
497 	*/
498 	if (wq->flushed && !SW_CQE(hw_cqe)) {
499 		ret = -EAGAIN;
500 		goto skip_cqe;
501 	}
502 
503 	/*
504 	 * skip TERMINATE cqes...
505 	 */
506 	if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE) {
507 		ret = -EAGAIN;
508 		goto skip_cqe;
509 	}
510 
511 	/*
512 	 * Gotta tweak READ completions:
513 	 *	1) the cqe doesn't contain the sq_wptr from the wr.
514 	 *	2) opcode not reflected from the wr.
515 	 *	3) read_len not reflected from the wr.
516 	 *	4) cq_type is RQ_TYPE not SQ_TYPE.
517 	 */
518 	if (RQ_TYPE(hw_cqe) && (CQE_OPCODE(hw_cqe) == FW_RI_READ_RESP)) {
519 
520 		/* If we have reached here because of async
521 		 * event or other error, and have egress error
522 		 * then drop
523 		 */
524 		if (CQE_TYPE(hw_cqe) == 1) {
525 			if (CQE_STATUS(hw_cqe))
526 				t4_set_wq_in_error(wq);
527 			ret = -EAGAIN;
528 			goto skip_cqe;
529 		}
530 
531 		/* If this is an unsolicited read response, then the read
532 		 * was generated by the kernel driver as part of peer-2-peer
533 		 * connection setup.  So ignore the completion.
534 		 */
535 		if (CQE_WRID_STAG(hw_cqe) == 1) {
536 			if (CQE_STATUS(hw_cqe))
537 				t4_set_wq_in_error(wq);
538 			ret = -EAGAIN;
539 			goto skip_cqe;
540 		}
541 
542 		/*
543 		 * Eat completions for unsignaled read WRs.
544 		 */
545 		if (!wq->sq.oldest_read->signaled) {
546 			advance_oldest_read(wq);
547 			ret = -EAGAIN;
548 			goto skip_cqe;
549 		}
550 
551 		/*
552 		 * Don't write to the HWCQ, so create a new read req CQE
553 		 * in local memory.
554 		 */
555 		create_read_req_cqe(wq, hw_cqe, &read_cqe);
556 		hw_cqe = &read_cqe;
557 		advance_oldest_read(wq);
558 	}
559 
560 	if (CQE_STATUS(hw_cqe) || t4_wq_in_error(wq)) {
561 		*cqe_flushed = (CQE_STATUS(hw_cqe) == T4_ERR_SWFLUSH);
562 		t4_set_wq_in_error(wq);
563 	}
564 
565 	/*
566 	 * RECV completion.
567 	 */
568 	if (RQ_TYPE(hw_cqe)) {
569 
570 		/*
571 		 * HW only validates 4 bits of MSN.  So we must validate that
572 		 * the MSN in the SEND is the next expected MSN.  If its not,
573 		 * then we complete this with T4_ERR_MSN and mark the wq in
574 		 * error.
575 		 */
576 
577 		if (t4_rq_empty(wq)) {
578 			t4_set_wq_in_error(wq);
579 			ret = -EAGAIN;
580 			goto skip_cqe;
581 		}
582 		if (unlikely((CQE_WRID_MSN(hw_cqe) != (wq->rq.msn)))) {
583 			t4_set_wq_in_error(wq);
584 			hw_cqe->header |= htonl(CQE_STATUS_V(T4_ERR_MSN));
585 			goto proc_cqe;
586 		}
587 		goto proc_cqe;
588 	}
589 
590 	/*
591 	 * If we get here its a send completion.
592 	 *
593 	 * Handle out of order completion. These get stuffed
594 	 * in the SW SQ. Then the SW SQ is walked to move any
595 	 * now in-order completions into the SW CQ.  This handles
596 	 * 2 cases:
597 	 *	1) reaping unsignaled WRs when the first subsequent
598 	 *	   signaled WR is completed.
599 	 *	2) out of order read completions.
600 	 */
601 	if (!SW_CQE(hw_cqe) && (CQE_WRID_SQ_IDX(hw_cqe) != wq->sq.cidx)) {
602 		struct t4_swsqe *swsqe;
603 
604 		PDBG("%s out of order completion going in sw_sq at idx %u\n",
605 		     __func__, CQE_WRID_SQ_IDX(hw_cqe));
606 		swsqe = &wq->sq.sw_sq[CQE_WRID_SQ_IDX(hw_cqe)];
607 		swsqe->cqe = *hw_cqe;
608 		swsqe->complete = 1;
609 		ret = -EAGAIN;
610 		goto flush_wq;
611 	}
612 
613 proc_cqe:
614 	*cqe = *hw_cqe;
615 
616 	/*
617 	 * Reap the associated WR(s) that are freed up with this
618 	 * completion.
619 	 */
620 	if (SQ_TYPE(hw_cqe)) {
621 		int idx = CQE_WRID_SQ_IDX(hw_cqe);
622 		BUG_ON(idx >= wq->sq.size);
623 
624 		/*
625 		* Account for any unsignaled completions completed by
626 		* this signaled completion.  In this case, cidx points
627 		* to the first unsignaled one, and idx points to the
628 		* signaled one.  So adjust in_use based on this delta.
629 		* if this is not completing any unsigned wrs, then the
630 		* delta will be 0. Handle wrapping also!
631 		*/
632 		if (idx < wq->sq.cidx)
633 			wq->sq.in_use -= wq->sq.size + idx - wq->sq.cidx;
634 		else
635 			wq->sq.in_use -= idx - wq->sq.cidx;
636 		BUG_ON(wq->sq.in_use <= 0 && wq->sq.in_use >= wq->sq.size);
637 
638 		wq->sq.cidx = (uint16_t)idx;
639 		PDBG("%s completing sq idx %u\n", __func__, wq->sq.cidx);
640 		*cookie = wq->sq.sw_sq[wq->sq.cidx].wr_id;
641 		if (c4iw_wr_log)
642 			c4iw_log_wr_stats(wq, hw_cqe);
643 		t4_sq_consume(wq);
644 	} else {
645 		PDBG("%s completing rq idx %u\n", __func__, wq->rq.cidx);
646 		*cookie = wq->rq.sw_rq[wq->rq.cidx].wr_id;
647 		BUG_ON(t4_rq_empty(wq));
648 		if (c4iw_wr_log)
649 			c4iw_log_wr_stats(wq, hw_cqe);
650 		t4_rq_consume(wq);
651 		goto skip_cqe;
652 	}
653 
654 flush_wq:
655 	/*
656 	 * Flush any completed cqes that are now in-order.
657 	 */
658 	flush_completed_wrs(wq, cq);
659 
660 skip_cqe:
661 	if (SW_CQE(hw_cqe)) {
662 		PDBG("%s cq %p cqid 0x%x skip sw cqe cidx %u\n",
663 		     __func__, cq, cq->cqid, cq->sw_cidx);
664 		t4_swcq_consume(cq);
665 	} else {
666 		PDBG("%s cq %p cqid 0x%x skip hw cqe cidx %u\n",
667 		     __func__, cq, cq->cqid, cq->cidx);
668 		t4_hwcq_consume(cq);
669 	}
670 	return ret;
671 }
672 
673 /*
674  * Get one cq entry from c4iw and map it to openib.
675  *
676  * Returns:
677  *	0			cqe returned
678  *	-ENODATA		EMPTY;
679  *	-EAGAIN			caller must try again
680  *	any other -errno	fatal error
681  */
682 static int c4iw_poll_cq_one(struct c4iw_cq *chp, struct ib_wc *wc)
683 {
684 	struct c4iw_qp *qhp = NULL;
685 	struct t4_cqe uninitialized_var(cqe), *rd_cqe;
686 	struct t4_wq *wq;
687 	u32 credit = 0;
688 	u8 cqe_flushed;
689 	u64 cookie = 0;
690 	int ret;
691 
692 	ret = t4_next_cqe(&chp->cq, &rd_cqe);
693 
694 	if (ret)
695 		return ret;
696 
697 	qhp = get_qhp(chp->rhp, CQE_QPID(rd_cqe));
698 	if (!qhp)
699 		wq = NULL;
700 	else {
701 		spin_lock(&qhp->lock);
702 		wq = &(qhp->wq);
703 	}
704 	ret = poll_cq(wq, &(chp->cq), &cqe, &cqe_flushed, &cookie, &credit);
705 	if (ret)
706 		goto out;
707 
708 	wc->wr_id = cookie;
709 	wc->qp = &qhp->ibqp;
710 	wc->vendor_err = CQE_STATUS(&cqe);
711 	wc->wc_flags = 0;
712 
713 	PDBG("%s qpid 0x%x type %d opcode %d status 0x%x len %u wrid hi 0x%x "
714 	     "lo 0x%x cookie 0x%llx\n", __func__, CQE_QPID(&cqe),
715 	     CQE_TYPE(&cqe), CQE_OPCODE(&cqe), CQE_STATUS(&cqe), CQE_LEN(&cqe),
716 	     CQE_WRID_HI(&cqe), CQE_WRID_LOW(&cqe), (unsigned long long)cookie);
717 
718 	if (CQE_TYPE(&cqe) == 0) {
719 		if (!CQE_STATUS(&cqe))
720 			wc->byte_len = CQE_LEN(&cqe);
721 		else
722 			wc->byte_len = 0;
723 		wc->opcode = IB_WC_RECV;
724 		if (CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_INV ||
725 		    CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_SE_INV) {
726 			wc->ex.invalidate_rkey = CQE_WRID_STAG(&cqe);
727 			wc->wc_flags |= IB_WC_WITH_INVALIDATE;
728 		}
729 	} else {
730 		switch (CQE_OPCODE(&cqe)) {
731 		case FW_RI_RDMA_WRITE:
732 			wc->opcode = IB_WC_RDMA_WRITE;
733 			break;
734 		case FW_RI_READ_REQ:
735 			wc->opcode = IB_WC_RDMA_READ;
736 			wc->byte_len = CQE_LEN(&cqe);
737 			break;
738 		case FW_RI_SEND_WITH_INV:
739 		case FW_RI_SEND_WITH_SE_INV:
740 			wc->opcode = IB_WC_SEND;
741 			wc->wc_flags |= IB_WC_WITH_INVALIDATE;
742 			break;
743 		case FW_RI_SEND:
744 		case FW_RI_SEND_WITH_SE:
745 			wc->opcode = IB_WC_SEND;
746 			break;
747 		case FW_RI_BIND_MW:
748 			wc->opcode = IB_WC_BIND_MW;
749 			break;
750 
751 		case FW_RI_LOCAL_INV:
752 			wc->opcode = IB_WC_LOCAL_INV;
753 			break;
754 		case FW_RI_FAST_REGISTER:
755 			wc->opcode = IB_WC_FAST_REG_MR;
756 			break;
757 		default:
758 			printk(KERN_ERR MOD "Unexpected opcode %d "
759 			       "in the CQE received for QPID=0x%0x\n",
760 			       CQE_OPCODE(&cqe), CQE_QPID(&cqe));
761 			ret = -EINVAL;
762 			goto out;
763 		}
764 	}
765 
766 	if (cqe_flushed)
767 		wc->status = IB_WC_WR_FLUSH_ERR;
768 	else {
769 
770 		switch (CQE_STATUS(&cqe)) {
771 		case T4_ERR_SUCCESS:
772 			wc->status = IB_WC_SUCCESS;
773 			break;
774 		case T4_ERR_STAG:
775 			wc->status = IB_WC_LOC_ACCESS_ERR;
776 			break;
777 		case T4_ERR_PDID:
778 			wc->status = IB_WC_LOC_PROT_ERR;
779 			break;
780 		case T4_ERR_QPID:
781 		case T4_ERR_ACCESS:
782 			wc->status = IB_WC_LOC_ACCESS_ERR;
783 			break;
784 		case T4_ERR_WRAP:
785 			wc->status = IB_WC_GENERAL_ERR;
786 			break;
787 		case T4_ERR_BOUND:
788 			wc->status = IB_WC_LOC_LEN_ERR;
789 			break;
790 		case T4_ERR_INVALIDATE_SHARED_MR:
791 		case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
792 			wc->status = IB_WC_MW_BIND_ERR;
793 			break;
794 		case T4_ERR_CRC:
795 		case T4_ERR_MARKER:
796 		case T4_ERR_PDU_LEN_ERR:
797 		case T4_ERR_OUT_OF_RQE:
798 		case T4_ERR_DDP_VERSION:
799 		case T4_ERR_RDMA_VERSION:
800 		case T4_ERR_DDP_QUEUE_NUM:
801 		case T4_ERR_MSN:
802 		case T4_ERR_TBIT:
803 		case T4_ERR_MO:
804 		case T4_ERR_MSN_RANGE:
805 		case T4_ERR_IRD_OVERFLOW:
806 		case T4_ERR_OPCODE:
807 		case T4_ERR_INTERNAL_ERR:
808 			wc->status = IB_WC_FATAL_ERR;
809 			break;
810 		case T4_ERR_SWFLUSH:
811 			wc->status = IB_WC_WR_FLUSH_ERR;
812 			break;
813 		default:
814 			printk(KERN_ERR MOD
815 			       "Unexpected cqe_status 0x%x for QPID=0x%0x\n",
816 			       CQE_STATUS(&cqe), CQE_QPID(&cqe));
817 			wc->status = IB_WC_FATAL_ERR;
818 		}
819 	}
820 out:
821 	if (wq)
822 		spin_unlock(&qhp->lock);
823 	return ret;
824 }
825 
826 int c4iw_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
827 {
828 	struct c4iw_cq *chp;
829 	unsigned long flags;
830 	int npolled;
831 	int err = 0;
832 
833 	chp = to_c4iw_cq(ibcq);
834 
835 	spin_lock_irqsave(&chp->lock, flags);
836 	for (npolled = 0; npolled < num_entries; ++npolled) {
837 		do {
838 			err = c4iw_poll_cq_one(chp, wc + npolled);
839 		} while (err == -EAGAIN);
840 		if (err)
841 			break;
842 	}
843 	spin_unlock_irqrestore(&chp->lock, flags);
844 	return !err || err == -ENODATA ? npolled : err;
845 }
846 
847 int c4iw_destroy_cq(struct ib_cq *ib_cq)
848 {
849 	struct c4iw_cq *chp;
850 	struct c4iw_ucontext *ucontext;
851 
852 	PDBG("%s ib_cq %p\n", __func__, ib_cq);
853 	chp = to_c4iw_cq(ib_cq);
854 
855 	remove_handle(chp->rhp, &chp->rhp->cqidr, chp->cq.cqid);
856 	atomic_dec(&chp->refcnt);
857 	wait_event(chp->wait, !atomic_read(&chp->refcnt));
858 
859 	ucontext = ib_cq->uobject ? to_c4iw_ucontext(ib_cq->uobject->context)
860 				  : NULL;
861 	destroy_cq(&chp->rhp->rdev, &chp->cq,
862 		   ucontext ? &ucontext->uctx : &chp->cq.rdev->uctx);
863 	kfree(chp);
864 	return 0;
865 }
866 
867 struct ib_cq *c4iw_create_cq(struct ib_device *ibdev,
868 			     const struct ib_cq_init_attr *attr,
869 			     struct ib_ucontext *ib_context,
870 			     struct ib_udata *udata)
871 {
872 	int entries = attr->cqe;
873 	int vector = attr->comp_vector;
874 	struct c4iw_dev *rhp;
875 	struct c4iw_cq *chp;
876 	struct c4iw_create_cq_resp uresp;
877 	struct c4iw_ucontext *ucontext = NULL;
878 	int ret;
879 	size_t memsize, hwentries;
880 	struct c4iw_mm_entry *mm, *mm2;
881 
882 	PDBG("%s ib_dev %p entries %d\n", __func__, ibdev, entries);
883 	if (attr->flags)
884 		return ERR_PTR(-EINVAL);
885 
886 	rhp = to_c4iw_dev(ibdev);
887 
888 	if (vector >= rhp->rdev.lldi.nciq)
889 		return ERR_PTR(-EINVAL);
890 
891 	chp = kzalloc(sizeof(*chp), GFP_KERNEL);
892 	if (!chp)
893 		return ERR_PTR(-ENOMEM);
894 
895 	if (ib_context)
896 		ucontext = to_c4iw_ucontext(ib_context);
897 
898 	/* account for the status page. */
899 	entries++;
900 
901 	/* IQ needs one extra entry to differentiate full vs empty. */
902 	entries++;
903 
904 	/*
905 	 * entries must be multiple of 16 for HW.
906 	 */
907 	entries = roundup(entries, 16);
908 
909 	/*
910 	 * Make actual HW queue 2x to avoid cdix_inc overflows.
911 	 */
912 	hwentries = min(entries * 2, rhp->rdev.hw_queue.t4_max_iq_size);
913 
914 	/*
915 	 * Make HW queue at least 64 entries so GTS updates aren't too
916 	 * frequent.
917 	 */
918 	if (hwentries < 64)
919 		hwentries = 64;
920 
921 	memsize = hwentries * sizeof *chp->cq.queue;
922 
923 	/*
924 	 * memsize must be a multiple of the page size if its a user cq.
925 	 */
926 	if (ucontext)
927 		memsize = roundup(memsize, PAGE_SIZE);
928 	chp->cq.size = hwentries;
929 	chp->cq.memsize = memsize;
930 	chp->cq.vector = vector;
931 
932 	ret = create_cq(&rhp->rdev, &chp->cq,
933 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
934 	if (ret)
935 		goto err1;
936 
937 	chp->rhp = rhp;
938 	chp->cq.size--;				/* status page */
939 	chp->ibcq.cqe = entries - 2;
940 	spin_lock_init(&chp->lock);
941 	spin_lock_init(&chp->comp_handler_lock);
942 	atomic_set(&chp->refcnt, 1);
943 	init_waitqueue_head(&chp->wait);
944 	ret = insert_handle(rhp, &rhp->cqidr, chp, chp->cq.cqid);
945 	if (ret)
946 		goto err2;
947 
948 	if (ucontext) {
949 		mm = kmalloc(sizeof *mm, GFP_KERNEL);
950 		if (!mm)
951 			goto err3;
952 		mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
953 		if (!mm2)
954 			goto err4;
955 
956 		uresp.qid_mask = rhp->rdev.cqmask;
957 		uresp.cqid = chp->cq.cqid;
958 		uresp.size = chp->cq.size;
959 		uresp.memsize = chp->cq.memsize;
960 		spin_lock(&ucontext->mmap_lock);
961 		uresp.key = ucontext->key;
962 		ucontext->key += PAGE_SIZE;
963 		uresp.gts_key = ucontext->key;
964 		ucontext->key += PAGE_SIZE;
965 		spin_unlock(&ucontext->mmap_lock);
966 		ret = ib_copy_to_udata(udata, &uresp,
967 				       sizeof(uresp) - sizeof(uresp.reserved));
968 		if (ret)
969 			goto err5;
970 
971 		mm->key = uresp.key;
972 		mm->addr = virt_to_phys(chp->cq.queue);
973 		mm->len = chp->cq.memsize;
974 		insert_mmap(ucontext, mm);
975 
976 		mm2->key = uresp.gts_key;
977 		mm2->addr = chp->cq.bar2_pa;
978 		mm2->len = PAGE_SIZE;
979 		insert_mmap(ucontext, mm2);
980 	}
981 	PDBG("%s cqid 0x%0x chp %p size %u memsize %zu, dma_addr 0x%0llx\n",
982 	     __func__, chp->cq.cqid, chp, chp->cq.size,
983 	     chp->cq.memsize, (unsigned long long) chp->cq.dma_addr);
984 	return &chp->ibcq;
985 err5:
986 	kfree(mm2);
987 err4:
988 	kfree(mm);
989 err3:
990 	remove_handle(rhp, &rhp->cqidr, chp->cq.cqid);
991 err2:
992 	destroy_cq(&chp->rhp->rdev, &chp->cq,
993 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
994 err1:
995 	kfree(chp);
996 	return ERR_PTR(ret);
997 }
998 
999 int c4iw_resize_cq(struct ib_cq *cq, int cqe, struct ib_udata *udata)
1000 {
1001 	return -ENOSYS;
1002 }
1003 
1004 int c4iw_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
1005 {
1006 	struct c4iw_cq *chp;
1007 	int ret;
1008 	unsigned long flag;
1009 
1010 	chp = to_c4iw_cq(ibcq);
1011 	spin_lock_irqsave(&chp->lock, flag);
1012 	ret = t4_arm_cq(&chp->cq,
1013 			(flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED);
1014 	spin_unlock_irqrestore(&chp->lock, flag);
1015 	if (ret && !(flags & IB_CQ_REPORT_MISSED_EVENTS))
1016 		ret = 0;
1017 	return ret;
1018 }
1019