xref: /openbmc/linux/drivers/infiniband/hw/cxgb4/cq.c (revision 4da722ca19f30f7db250db808d1ab1703607a932)
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, struct sk_buff *skb)
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 	int ret;
43 
44 	wr_len = sizeof *res_wr + sizeof *res;
45 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
46 
47 	res_wr = __skb_put_zero(skb, wr_len);
48 	res_wr->op_nres = cpu_to_be32(
49 			FW_WR_OP_V(FW_RI_RES_WR) |
50 			FW_RI_RES_WR_NRES_V(1) |
51 			FW_WR_COMPL_F);
52 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
53 	res_wr->cookie = (uintptr_t)&wr_wait;
54 	res = res_wr->res;
55 	res->u.cq.restype = FW_RI_RES_TYPE_CQ;
56 	res->u.cq.op = FW_RI_RES_OP_RESET;
57 	res->u.cq.iqid = cpu_to_be32(cq->cqid);
58 
59 	c4iw_init_wr_wait(&wr_wait);
60 	ret = c4iw_ofld_send(rdev, skb);
61 	if (!ret) {
62 		ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, __func__);
63 	}
64 
65 	kfree(cq->sw_queue);
66 	dma_free_coherent(&(rdev->lldi.pdev->dev),
67 			  cq->memsize, cq->queue,
68 			  dma_unmap_addr(cq, mapping));
69 	c4iw_put_cqid(rdev, cq->cqid, uctx);
70 	return ret;
71 }
72 
73 static int create_cq(struct c4iw_rdev *rdev, struct t4_cq *cq,
74 		     struct c4iw_dev_ucontext *uctx)
75 {
76 	struct fw_ri_res_wr *res_wr;
77 	struct fw_ri_res *res;
78 	int wr_len;
79 	int user = (uctx != &rdev->uctx);
80 	struct c4iw_wr_wait wr_wait;
81 	int ret;
82 	struct sk_buff *skb;
83 
84 	cq->cqid = c4iw_get_cqid(rdev, uctx);
85 	if (!cq->cqid) {
86 		ret = -ENOMEM;
87 		goto err1;
88 	}
89 
90 	if (!user) {
91 		cq->sw_queue = kzalloc(cq->memsize, GFP_KERNEL);
92 		if (!cq->sw_queue) {
93 			ret = -ENOMEM;
94 			goto err2;
95 		}
96 	}
97 	cq->queue = dma_alloc_coherent(&rdev->lldi.pdev->dev, cq->memsize,
98 				       &cq->dma_addr, GFP_KERNEL);
99 	if (!cq->queue) {
100 		ret = -ENOMEM;
101 		goto err3;
102 	}
103 	dma_unmap_addr_set(cq, mapping, cq->dma_addr);
104 	memset(cq->queue, 0, cq->memsize);
105 
106 	/* build fw_ri_res_wr */
107 	wr_len = sizeof *res_wr + sizeof *res;
108 
109 	skb = alloc_skb(wr_len, GFP_KERNEL);
110 	if (!skb) {
111 		ret = -ENOMEM;
112 		goto err4;
113 	}
114 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
115 
116 	res_wr = __skb_put_zero(skb, wr_len);
117 	res_wr->op_nres = cpu_to_be32(
118 			FW_WR_OP_V(FW_RI_RES_WR) |
119 			FW_RI_RES_WR_NRES_V(1) |
120 			FW_WR_COMPL_F);
121 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
122 	res_wr->cookie = (uintptr_t)&wr_wait;
123 	res = res_wr->res;
124 	res->u.cq.restype = FW_RI_RES_TYPE_CQ;
125 	res->u.cq.op = FW_RI_RES_OP_WRITE;
126 	res->u.cq.iqid = cpu_to_be32(cq->cqid);
127 	res->u.cq.iqandst_to_iqandstindex = cpu_to_be32(
128 			FW_RI_RES_WR_IQANUS_V(0) |
129 			FW_RI_RES_WR_IQANUD_V(1) |
130 			FW_RI_RES_WR_IQANDST_F |
131 			FW_RI_RES_WR_IQANDSTINDEX_V(
132 				rdev->lldi.ciq_ids[cq->vector]));
133 	res->u.cq.iqdroprss_to_iqesize = cpu_to_be16(
134 			FW_RI_RES_WR_IQDROPRSS_F |
135 			FW_RI_RES_WR_IQPCIECH_V(2) |
136 			FW_RI_RES_WR_IQINTCNTTHRESH_V(0) |
137 			FW_RI_RES_WR_IQO_F |
138 			FW_RI_RES_WR_IQESIZE_V(1));
139 	res->u.cq.iqsize = cpu_to_be16(cq->size);
140 	res->u.cq.iqaddr = cpu_to_be64(cq->dma_addr);
141 
142 	c4iw_init_wr_wait(&wr_wait);
143 
144 	ret = c4iw_ofld_send(rdev, skb);
145 	if (ret)
146 		goto err4;
147 	pr_debug("%s wait_event wr_wait %p\n", __func__, &wr_wait);
148 	ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, __func__);
149 	if (ret)
150 		goto err4;
151 
152 	cq->gen = 1;
153 	cq->gts = rdev->lldi.gts_reg;
154 	cq->rdev = rdev;
155 
156 	cq->bar2_va = c4iw_bar2_addrs(rdev, cq->cqid, T4_BAR2_QTYPE_INGRESS,
157 				      &cq->bar2_qid,
158 				      user ? &cq->bar2_pa : NULL);
159 	if (user && !cq->bar2_pa) {
160 		pr_warn("%s: cqid %u not in BAR2 range\n",
161 			pci_name(rdev->lldi.pdev), cq->cqid);
162 		ret = -EINVAL;
163 		goto err4;
164 	}
165 	return 0;
166 err4:
167 	dma_free_coherent(&rdev->lldi.pdev->dev, cq->memsize, cq->queue,
168 			  dma_unmap_addr(cq, mapping));
169 err3:
170 	kfree(cq->sw_queue);
171 err2:
172 	c4iw_put_cqid(rdev, cq->cqid, uctx);
173 err1:
174 	return ret;
175 }
176 
177 static void insert_recv_cqe(struct t4_wq *wq, struct t4_cq *cq)
178 {
179 	struct t4_cqe cqe;
180 
181 	pr_debug("%s wq %p cq %p sw_cidx %u sw_pidx %u\n", __func__,
182 		 wq, cq, cq->sw_cidx, cq->sw_pidx);
183 	memset(&cqe, 0, sizeof(cqe));
184 	cqe.header = cpu_to_be32(CQE_STATUS_V(T4_ERR_SWFLUSH) |
185 				 CQE_OPCODE_V(FW_RI_SEND) |
186 				 CQE_TYPE_V(0) |
187 				 CQE_SWCQE_V(1) |
188 				 CQE_QPID_V(wq->sq.qid));
189 	cqe.bits_type_ts = cpu_to_be64(CQE_GENBIT_V((u64)cq->gen));
190 	cq->sw_queue[cq->sw_pidx] = cqe;
191 	t4_swcq_produce(cq);
192 }
193 
194 int c4iw_flush_rq(struct t4_wq *wq, struct t4_cq *cq, int count)
195 {
196 	int flushed = 0;
197 	int in_use = wq->rq.in_use - count;
198 
199 	BUG_ON(in_use < 0);
200 	pr_debug("%s wq %p cq %p rq.in_use %u skip count %u\n", __func__,
201 		 wq, cq, wq->rq.in_use, count);
202 	while (in_use--) {
203 		insert_recv_cqe(wq, cq);
204 		flushed++;
205 	}
206 	return flushed;
207 }
208 
209 static void insert_sq_cqe(struct t4_wq *wq, struct t4_cq *cq,
210 			  struct t4_swsqe *swcqe)
211 {
212 	struct t4_cqe cqe;
213 
214 	pr_debug("%s wq %p cq %p sw_cidx %u sw_pidx %u\n", __func__,
215 		 wq, cq, cq->sw_cidx, cq->sw_pidx);
216 	memset(&cqe, 0, sizeof(cqe));
217 	cqe.header = cpu_to_be32(CQE_STATUS_V(T4_ERR_SWFLUSH) |
218 				 CQE_OPCODE_V(swcqe->opcode) |
219 				 CQE_TYPE_V(1) |
220 				 CQE_SWCQE_V(1) |
221 				 CQE_QPID_V(wq->sq.qid));
222 	CQE_WRID_SQ_IDX(&cqe) = swcqe->idx;
223 	cqe.bits_type_ts = cpu_to_be64(CQE_GENBIT_V((u64)cq->gen));
224 	cq->sw_queue[cq->sw_pidx] = cqe;
225 	t4_swcq_produce(cq);
226 }
227 
228 static void advance_oldest_read(struct t4_wq *wq);
229 
230 int c4iw_flush_sq(struct c4iw_qp *qhp)
231 {
232 	int flushed = 0;
233 	struct t4_wq *wq = &qhp->wq;
234 	struct c4iw_cq *chp = to_c4iw_cq(qhp->ibqp.send_cq);
235 	struct t4_cq *cq = &chp->cq;
236 	int idx;
237 	struct t4_swsqe *swsqe;
238 
239 	if (wq->sq.flush_cidx == -1)
240 		wq->sq.flush_cidx = wq->sq.cidx;
241 	idx = wq->sq.flush_cidx;
242 	BUG_ON(idx >= wq->sq.size);
243 	while (idx != wq->sq.pidx) {
244 		swsqe = &wq->sq.sw_sq[idx];
245 		BUG_ON(swsqe->flushed);
246 		swsqe->flushed = 1;
247 		insert_sq_cqe(wq, cq, swsqe);
248 		if (wq->sq.oldest_read == swsqe) {
249 			BUG_ON(swsqe->opcode != FW_RI_READ_REQ);
250 			advance_oldest_read(wq);
251 		}
252 		flushed++;
253 		if (++idx == wq->sq.size)
254 			idx = 0;
255 	}
256 	wq->sq.flush_cidx += flushed;
257 	if (wq->sq.flush_cidx >= wq->sq.size)
258 		wq->sq.flush_cidx -= wq->sq.size;
259 	return flushed;
260 }
261 
262 static void flush_completed_wrs(struct t4_wq *wq, struct t4_cq *cq)
263 {
264 	struct t4_swsqe *swsqe;
265 	int cidx;
266 
267 	if (wq->sq.flush_cidx == -1)
268 		wq->sq.flush_cidx = wq->sq.cidx;
269 	cidx = wq->sq.flush_cidx;
270 	BUG_ON(cidx > wq->sq.size);
271 
272 	while (cidx != wq->sq.pidx) {
273 		swsqe = &wq->sq.sw_sq[cidx];
274 		if (!swsqe->signaled) {
275 			if (++cidx == wq->sq.size)
276 				cidx = 0;
277 		} else if (swsqe->complete) {
278 
279 			BUG_ON(swsqe->flushed);
280 
281 			/*
282 			 * Insert this completed cqe into the swcq.
283 			 */
284 			pr_debug("%s moving cqe into swcq sq idx %u cq idx %u\n",
285 				 __func__, cidx, cq->sw_pidx);
286 			swsqe->cqe.header |= htonl(CQE_SWCQE_V(1));
287 			cq->sw_queue[cq->sw_pidx] = swsqe->cqe;
288 			t4_swcq_produce(cq);
289 			swsqe->flushed = 1;
290 			if (++cidx == wq->sq.size)
291 				cidx = 0;
292 			wq->sq.flush_cidx = cidx;
293 		} else
294 			break;
295 	}
296 }
297 
298 static void create_read_req_cqe(struct t4_wq *wq, struct t4_cqe *hw_cqe,
299 		struct t4_cqe *read_cqe)
300 {
301 	read_cqe->u.scqe.cidx = wq->sq.oldest_read->idx;
302 	read_cqe->len = htonl(wq->sq.oldest_read->read_len);
303 	read_cqe->header = htonl(CQE_QPID_V(CQE_QPID(hw_cqe)) |
304 			CQE_SWCQE_V(SW_CQE(hw_cqe)) |
305 			CQE_OPCODE_V(FW_RI_READ_REQ) |
306 			CQE_TYPE_V(1));
307 	read_cqe->bits_type_ts = hw_cqe->bits_type_ts;
308 }
309 
310 static void advance_oldest_read(struct t4_wq *wq)
311 {
312 
313 	u32 rptr = wq->sq.oldest_read - wq->sq.sw_sq + 1;
314 
315 	if (rptr == wq->sq.size)
316 		rptr = 0;
317 	while (rptr != wq->sq.pidx) {
318 		wq->sq.oldest_read = &wq->sq.sw_sq[rptr];
319 
320 		if (wq->sq.oldest_read->opcode == FW_RI_READ_REQ)
321 			return;
322 		if (++rptr == wq->sq.size)
323 			rptr = 0;
324 	}
325 	wq->sq.oldest_read = NULL;
326 }
327 
328 /*
329  * Move all CQEs from the HWCQ into the SWCQ.
330  * Deal with out-of-order and/or completions that complete
331  * prior unsignalled WRs.
332  */
333 void c4iw_flush_hw_cq(struct c4iw_cq *chp)
334 {
335 	struct t4_cqe *hw_cqe, *swcqe, read_cqe;
336 	struct c4iw_qp *qhp;
337 	struct t4_swsqe *swsqe;
338 	int ret;
339 
340 	pr_debug("%s  cqid 0x%x\n", __func__, chp->cq.cqid);
341 	ret = t4_next_hw_cqe(&chp->cq, &hw_cqe);
342 
343 	/*
344 	 * This logic is similar to poll_cq(), but not quite the same
345 	 * unfortunately.  Need to move pertinent HW CQEs to the SW CQ but
346 	 * also do any translation magic that poll_cq() normally does.
347 	 */
348 	while (!ret) {
349 		qhp = get_qhp(chp->rhp, CQE_QPID(hw_cqe));
350 
351 		/*
352 		 * drop CQEs with no associated QP
353 		 */
354 		if (qhp == NULL)
355 			goto next_cqe;
356 
357 		if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE)
358 			goto next_cqe;
359 
360 		if (CQE_OPCODE(hw_cqe) == FW_RI_READ_RESP) {
361 
362 			/* If we have reached here because of async
363 			 * event or other error, and have egress error
364 			 * then drop
365 			 */
366 			if (CQE_TYPE(hw_cqe) == 1)
367 				goto next_cqe;
368 
369 			/* drop peer2peer RTR reads.
370 			 */
371 			if (CQE_WRID_STAG(hw_cqe) == 1)
372 				goto next_cqe;
373 
374 			/*
375 			 * Eat completions for unsignaled read WRs.
376 			 */
377 			if (!qhp->wq.sq.oldest_read->signaled) {
378 				advance_oldest_read(&qhp->wq);
379 				goto next_cqe;
380 			}
381 
382 			/*
383 			 * Don't write to the HWCQ, create a new read req CQE
384 			 * in local memory and move it into the swcq.
385 			 */
386 			create_read_req_cqe(&qhp->wq, hw_cqe, &read_cqe);
387 			hw_cqe = &read_cqe;
388 			advance_oldest_read(&qhp->wq);
389 		}
390 
391 		/* if its a SQ completion, then do the magic to move all the
392 		 * unsignaled and now in-order completions into the swcq.
393 		 */
394 		if (SQ_TYPE(hw_cqe)) {
395 			swsqe = &qhp->wq.sq.sw_sq[CQE_WRID_SQ_IDX(hw_cqe)];
396 			swsqe->cqe = *hw_cqe;
397 			swsqe->complete = 1;
398 			flush_completed_wrs(&qhp->wq, &chp->cq);
399 		} else {
400 			swcqe = &chp->cq.sw_queue[chp->cq.sw_pidx];
401 			*swcqe = *hw_cqe;
402 			swcqe->header |= cpu_to_be32(CQE_SWCQE_V(1));
403 			t4_swcq_produce(&chp->cq);
404 		}
405 next_cqe:
406 		t4_hwcq_consume(&chp->cq);
407 		ret = t4_next_hw_cqe(&chp->cq, &hw_cqe);
408 	}
409 }
410 
411 static int cqe_completes_wr(struct t4_cqe *cqe, struct t4_wq *wq)
412 {
413 	if (CQE_OPCODE(cqe) == FW_RI_TERMINATE)
414 		return 0;
415 
416 	if ((CQE_OPCODE(cqe) == FW_RI_RDMA_WRITE) && RQ_TYPE(cqe))
417 		return 0;
418 
419 	if ((CQE_OPCODE(cqe) == FW_RI_READ_RESP) && SQ_TYPE(cqe))
420 		return 0;
421 
422 	if (CQE_SEND_OPCODE(cqe) && RQ_TYPE(cqe) && t4_rq_empty(wq))
423 		return 0;
424 	return 1;
425 }
426 
427 void c4iw_count_rcqes(struct t4_cq *cq, struct t4_wq *wq, int *count)
428 {
429 	struct t4_cqe *cqe;
430 	u32 ptr;
431 
432 	*count = 0;
433 	pr_debug("%s count zero %d\n", __func__, *count);
434 	ptr = cq->sw_cidx;
435 	while (ptr != cq->sw_pidx) {
436 		cqe = &cq->sw_queue[ptr];
437 		if (RQ_TYPE(cqe) && (CQE_OPCODE(cqe) != FW_RI_READ_RESP) &&
438 		    (CQE_QPID(cqe) == wq->sq.qid) && cqe_completes_wr(cqe, wq))
439 			(*count)++;
440 		if (++ptr == cq->size)
441 			ptr = 0;
442 	}
443 	pr_debug("%s cq %p count %d\n", __func__, cq, *count);
444 }
445 
446 /*
447  * poll_cq
448  *
449  * Caller must:
450  *     check the validity of the first CQE,
451  *     supply the wq assicated with the qpid.
452  *
453  * credit: cq credit to return to sge.
454  * cqe_flushed: 1 iff the CQE is flushed.
455  * cqe: copy of the polled CQE.
456  *
457  * return value:
458  *    0		    CQE returned ok.
459  *    -EAGAIN       CQE skipped, try again.
460  *    -EOVERFLOW    CQ overflow detected.
461  */
462 static int poll_cq(struct t4_wq *wq, struct t4_cq *cq, struct t4_cqe *cqe,
463 		   u8 *cqe_flushed, u64 *cookie, u32 *credit)
464 {
465 	int ret = 0;
466 	struct t4_cqe *hw_cqe, read_cqe;
467 
468 	*cqe_flushed = 0;
469 	*credit = 0;
470 	ret = t4_next_cqe(cq, &hw_cqe);
471 	if (ret)
472 		return ret;
473 
474 	pr_debug("%s CQE OVF %u qpid 0x%0x genbit %u type %u status 0x%0x opcode 0x%0x len 0x%0x wrid_hi_stag 0x%x wrid_low_msn 0x%x\n",
475 		 __func__, CQE_OVFBIT(hw_cqe), CQE_QPID(hw_cqe),
476 		 CQE_GENBIT(hw_cqe), CQE_TYPE(hw_cqe), CQE_STATUS(hw_cqe),
477 		 CQE_OPCODE(hw_cqe), CQE_LEN(hw_cqe), CQE_WRID_HI(hw_cqe),
478 		 CQE_WRID_LOW(hw_cqe));
479 
480 	/*
481 	 * skip cqe's not affiliated with a QP.
482 	 */
483 	if (wq == NULL) {
484 		ret = -EAGAIN;
485 		goto skip_cqe;
486 	}
487 
488 	/*
489 	* skip hw cqe's if the wq is flushed.
490 	*/
491 	if (wq->flushed && !SW_CQE(hw_cqe)) {
492 		ret = -EAGAIN;
493 		goto skip_cqe;
494 	}
495 
496 	/*
497 	 * skip TERMINATE cqes...
498 	 */
499 	if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE) {
500 		ret = -EAGAIN;
501 		goto skip_cqe;
502 	}
503 
504 	/*
505 	 * Special cqe for drain WR completions...
506 	 */
507 	if (CQE_OPCODE(hw_cqe) == C4IW_DRAIN_OPCODE) {
508 		*cookie = CQE_DRAIN_COOKIE(hw_cqe);
509 		*cqe = *hw_cqe;
510 		goto skip_cqe;
511 	}
512 
513 	/*
514 	 * Gotta tweak READ completions:
515 	 *	1) the cqe doesn't contain the sq_wptr from the wr.
516 	 *	2) opcode not reflected from the wr.
517 	 *	3) read_len not reflected from the wr.
518 	 *	4) cq_type is RQ_TYPE not SQ_TYPE.
519 	 */
520 	if (RQ_TYPE(hw_cqe) && (CQE_OPCODE(hw_cqe) == FW_RI_READ_RESP)) {
521 
522 		/* If we have reached here because of async
523 		 * event or other error, and have egress error
524 		 * then drop
525 		 */
526 		if (CQE_TYPE(hw_cqe) == 1) {
527 			if (CQE_STATUS(hw_cqe))
528 				t4_set_wq_in_error(wq);
529 			ret = -EAGAIN;
530 			goto skip_cqe;
531 		}
532 
533 		/* If this is an unsolicited read response, then the read
534 		 * was generated by the kernel driver as part of peer-2-peer
535 		 * connection setup.  So ignore the completion.
536 		 */
537 		if (CQE_WRID_STAG(hw_cqe) == 1) {
538 			if (CQE_STATUS(hw_cqe))
539 				t4_set_wq_in_error(wq);
540 			ret = -EAGAIN;
541 			goto skip_cqe;
542 		}
543 
544 		/*
545 		 * Eat completions for unsignaled read WRs.
546 		 */
547 		if (!wq->sq.oldest_read->signaled) {
548 			advance_oldest_read(wq);
549 			ret = -EAGAIN;
550 			goto skip_cqe;
551 		}
552 
553 		/*
554 		 * Don't write to the HWCQ, so create a new read req CQE
555 		 * in local memory.
556 		 */
557 		create_read_req_cqe(wq, hw_cqe, &read_cqe);
558 		hw_cqe = &read_cqe;
559 		advance_oldest_read(wq);
560 	}
561 
562 	if (CQE_STATUS(hw_cqe) || t4_wq_in_error(wq)) {
563 		*cqe_flushed = (CQE_STATUS(hw_cqe) == T4_ERR_SWFLUSH);
564 		t4_set_wq_in_error(wq);
565 	}
566 
567 	/*
568 	 * RECV completion.
569 	 */
570 	if (RQ_TYPE(hw_cqe)) {
571 
572 		/*
573 		 * HW only validates 4 bits of MSN.  So we must validate that
574 		 * the MSN in the SEND is the next expected MSN.  If its not,
575 		 * then we complete this with T4_ERR_MSN and mark the wq in
576 		 * error.
577 		 */
578 
579 		if (t4_rq_empty(wq)) {
580 			t4_set_wq_in_error(wq);
581 			ret = -EAGAIN;
582 			goto skip_cqe;
583 		}
584 		if (unlikely((CQE_WRID_MSN(hw_cqe) != (wq->rq.msn)))) {
585 			t4_set_wq_in_error(wq);
586 			hw_cqe->header |= htonl(CQE_STATUS_V(T4_ERR_MSN));
587 			goto proc_cqe;
588 		}
589 		goto proc_cqe;
590 	}
591 
592 	/*
593 	 * If we get here its a send completion.
594 	 *
595 	 * Handle out of order completion. These get stuffed
596 	 * in the SW SQ. Then the SW SQ is walked to move any
597 	 * now in-order completions into the SW CQ.  This handles
598 	 * 2 cases:
599 	 *	1) reaping unsignaled WRs when the first subsequent
600 	 *	   signaled WR is completed.
601 	 *	2) out of order read completions.
602 	 */
603 	if (!SW_CQE(hw_cqe) && (CQE_WRID_SQ_IDX(hw_cqe) != wq->sq.cidx)) {
604 		struct t4_swsqe *swsqe;
605 
606 		pr_debug("%s out of order completion going in sw_sq at idx %u\n",
607 			 __func__, CQE_WRID_SQ_IDX(hw_cqe));
608 		swsqe = &wq->sq.sw_sq[CQE_WRID_SQ_IDX(hw_cqe)];
609 		swsqe->cqe = *hw_cqe;
610 		swsqe->complete = 1;
611 		ret = -EAGAIN;
612 		goto flush_wq;
613 	}
614 
615 proc_cqe:
616 	*cqe = *hw_cqe;
617 
618 	/*
619 	 * Reap the associated WR(s) that are freed up with this
620 	 * completion.
621 	 */
622 	if (SQ_TYPE(hw_cqe)) {
623 		int idx = CQE_WRID_SQ_IDX(hw_cqe);
624 		BUG_ON(idx >= wq->sq.size);
625 
626 		/*
627 		* Account for any unsignaled completions completed by
628 		* this signaled completion.  In this case, cidx points
629 		* to the first unsignaled one, and idx points to the
630 		* signaled one.  So adjust in_use based on this delta.
631 		* if this is not completing any unsigned wrs, then the
632 		* delta will be 0. Handle wrapping also!
633 		*/
634 		if (idx < wq->sq.cidx)
635 			wq->sq.in_use -= wq->sq.size + idx - wq->sq.cidx;
636 		else
637 			wq->sq.in_use -= idx - wq->sq.cidx;
638 		BUG_ON(wq->sq.in_use <= 0 && wq->sq.in_use >= wq->sq.size);
639 
640 		wq->sq.cidx = (uint16_t)idx;
641 		pr_debug("%s completing sq idx %u\n", __func__, wq->sq.cidx);
642 		*cookie = wq->sq.sw_sq[wq->sq.cidx].wr_id;
643 		if (c4iw_wr_log)
644 			c4iw_log_wr_stats(wq, hw_cqe);
645 		t4_sq_consume(wq);
646 	} else {
647 		pr_debug("%s completing rq idx %u\n", __func__, wq->rq.cidx);
648 		*cookie = wq->rq.sw_rq[wq->rq.cidx].wr_id;
649 		BUG_ON(t4_rq_empty(wq));
650 		if (c4iw_wr_log)
651 			c4iw_log_wr_stats(wq, hw_cqe);
652 		t4_rq_consume(wq);
653 		goto skip_cqe;
654 	}
655 
656 flush_wq:
657 	/*
658 	 * Flush any completed cqes that are now in-order.
659 	 */
660 	flush_completed_wrs(wq, cq);
661 
662 skip_cqe:
663 	if (SW_CQE(hw_cqe)) {
664 		pr_debug("%s cq %p cqid 0x%x skip sw cqe cidx %u\n",
665 			 __func__, cq, cq->cqid, cq->sw_cidx);
666 		t4_swcq_consume(cq);
667 	} else {
668 		pr_debug("%s cq %p cqid 0x%x skip hw cqe cidx %u\n",
669 			 __func__, cq, cq->cqid, cq->cidx);
670 		t4_hwcq_consume(cq);
671 	}
672 	return ret;
673 }
674 
675 /*
676  * Get one cq entry from c4iw and map it to openib.
677  *
678  * Returns:
679  *	0			cqe returned
680  *	-ENODATA		EMPTY;
681  *	-EAGAIN			caller must try again
682  *	any other -errno	fatal error
683  */
684 static int c4iw_poll_cq_one(struct c4iw_cq *chp, struct ib_wc *wc)
685 {
686 	struct c4iw_qp *qhp = NULL;
687 	struct t4_cqe uninitialized_var(cqe), *rd_cqe;
688 	struct t4_wq *wq;
689 	u32 credit = 0;
690 	u8 cqe_flushed;
691 	u64 cookie = 0;
692 	int ret;
693 
694 	ret = t4_next_cqe(&chp->cq, &rd_cqe);
695 
696 	if (ret)
697 		return ret;
698 
699 	qhp = get_qhp(chp->rhp, CQE_QPID(rd_cqe));
700 	if (!qhp)
701 		wq = NULL;
702 	else {
703 		spin_lock(&qhp->lock);
704 		wq = &(qhp->wq);
705 	}
706 	ret = poll_cq(wq, &(chp->cq), &cqe, &cqe_flushed, &cookie, &credit);
707 	if (ret)
708 		goto out;
709 
710 	wc->wr_id = cookie;
711 	wc->qp = &qhp->ibqp;
712 	wc->vendor_err = CQE_STATUS(&cqe);
713 	wc->wc_flags = 0;
714 
715 	pr_debug("%s qpid 0x%x type %d opcode %d status 0x%x len %u wrid hi 0x%x lo 0x%x cookie 0x%llx\n",
716 		 __func__, CQE_QPID(&cqe),
717 		 CQE_TYPE(&cqe), CQE_OPCODE(&cqe),
718 		 CQE_STATUS(&cqe), CQE_LEN(&cqe),
719 		 CQE_WRID_HI(&cqe), CQE_WRID_LOW(&cqe),
720 		 (unsigned long long)cookie);
721 
722 	if (CQE_TYPE(&cqe) == 0) {
723 		if (!CQE_STATUS(&cqe))
724 			wc->byte_len = CQE_LEN(&cqe);
725 		else
726 			wc->byte_len = 0;
727 		wc->opcode = IB_WC_RECV;
728 		if (CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_INV ||
729 		    CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_SE_INV) {
730 			wc->ex.invalidate_rkey = CQE_WRID_STAG(&cqe);
731 			wc->wc_flags |= IB_WC_WITH_INVALIDATE;
732 			c4iw_invalidate_mr(qhp->rhp, wc->ex.invalidate_rkey);
733 		}
734 	} else {
735 		switch (CQE_OPCODE(&cqe)) {
736 		case FW_RI_RDMA_WRITE:
737 			wc->opcode = IB_WC_RDMA_WRITE;
738 			break;
739 		case FW_RI_READ_REQ:
740 			wc->opcode = IB_WC_RDMA_READ;
741 			wc->byte_len = CQE_LEN(&cqe);
742 			break;
743 		case FW_RI_SEND_WITH_INV:
744 		case FW_RI_SEND_WITH_SE_INV:
745 			wc->opcode = IB_WC_SEND;
746 			wc->wc_flags |= IB_WC_WITH_INVALIDATE;
747 			break;
748 		case FW_RI_SEND:
749 		case FW_RI_SEND_WITH_SE:
750 			wc->opcode = IB_WC_SEND;
751 			break;
752 
753 		case FW_RI_LOCAL_INV:
754 			wc->opcode = IB_WC_LOCAL_INV;
755 			break;
756 		case FW_RI_FAST_REGISTER:
757 			wc->opcode = IB_WC_REG_MR;
758 
759 			/* Invalidate the MR if the fastreg failed */
760 			if (CQE_STATUS(&cqe) != T4_ERR_SUCCESS)
761 				c4iw_invalidate_mr(qhp->rhp,
762 						   CQE_WRID_FR_STAG(&cqe));
763 			break;
764 		case C4IW_DRAIN_OPCODE:
765 			wc->opcode = IB_WC_SEND;
766 			break;
767 		default:
768 			pr_err("Unexpected opcode %d in the CQE received for QPID=0x%0x\n",
769 			       CQE_OPCODE(&cqe), CQE_QPID(&cqe));
770 			ret = -EINVAL;
771 			goto out;
772 		}
773 	}
774 
775 	if (cqe_flushed)
776 		wc->status = IB_WC_WR_FLUSH_ERR;
777 	else {
778 
779 		switch (CQE_STATUS(&cqe)) {
780 		case T4_ERR_SUCCESS:
781 			wc->status = IB_WC_SUCCESS;
782 			break;
783 		case T4_ERR_STAG:
784 			wc->status = IB_WC_LOC_ACCESS_ERR;
785 			break;
786 		case T4_ERR_PDID:
787 			wc->status = IB_WC_LOC_PROT_ERR;
788 			break;
789 		case T4_ERR_QPID:
790 		case T4_ERR_ACCESS:
791 			wc->status = IB_WC_LOC_ACCESS_ERR;
792 			break;
793 		case T4_ERR_WRAP:
794 			wc->status = IB_WC_GENERAL_ERR;
795 			break;
796 		case T4_ERR_BOUND:
797 			wc->status = IB_WC_LOC_LEN_ERR;
798 			break;
799 		case T4_ERR_INVALIDATE_SHARED_MR:
800 		case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
801 			wc->status = IB_WC_MW_BIND_ERR;
802 			break;
803 		case T4_ERR_CRC:
804 		case T4_ERR_MARKER:
805 		case T4_ERR_PDU_LEN_ERR:
806 		case T4_ERR_OUT_OF_RQE:
807 		case T4_ERR_DDP_VERSION:
808 		case T4_ERR_RDMA_VERSION:
809 		case T4_ERR_DDP_QUEUE_NUM:
810 		case T4_ERR_MSN:
811 		case T4_ERR_TBIT:
812 		case T4_ERR_MO:
813 		case T4_ERR_MSN_RANGE:
814 		case T4_ERR_IRD_OVERFLOW:
815 		case T4_ERR_OPCODE:
816 		case T4_ERR_INTERNAL_ERR:
817 			wc->status = IB_WC_FATAL_ERR;
818 			break;
819 		case T4_ERR_SWFLUSH:
820 			wc->status = IB_WC_WR_FLUSH_ERR;
821 			break;
822 		default:
823 			pr_err("Unexpected cqe_status 0x%x for QPID=0x%0x\n",
824 			       CQE_STATUS(&cqe), CQE_QPID(&cqe));
825 			wc->status = IB_WC_FATAL_ERR;
826 		}
827 	}
828 out:
829 	if (wq)
830 		spin_unlock(&qhp->lock);
831 	return ret;
832 }
833 
834 int c4iw_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
835 {
836 	struct c4iw_cq *chp;
837 	unsigned long flags;
838 	int npolled;
839 	int err = 0;
840 
841 	chp = to_c4iw_cq(ibcq);
842 
843 	spin_lock_irqsave(&chp->lock, flags);
844 	for (npolled = 0; npolled < num_entries; ++npolled) {
845 		do {
846 			err = c4iw_poll_cq_one(chp, wc + npolled);
847 		} while (err == -EAGAIN);
848 		if (err)
849 			break;
850 	}
851 	spin_unlock_irqrestore(&chp->lock, flags);
852 	return !err || err == -ENODATA ? npolled : err;
853 }
854 
855 int c4iw_destroy_cq(struct ib_cq *ib_cq)
856 {
857 	struct c4iw_cq *chp;
858 	struct c4iw_ucontext *ucontext;
859 
860 	pr_debug("%s ib_cq %p\n", __func__, ib_cq);
861 	chp = to_c4iw_cq(ib_cq);
862 
863 	remove_handle(chp->rhp, &chp->rhp->cqidr, chp->cq.cqid);
864 	atomic_dec(&chp->refcnt);
865 	wait_event(chp->wait, !atomic_read(&chp->refcnt));
866 
867 	ucontext = ib_cq->uobject ? to_c4iw_ucontext(ib_cq->uobject->context)
868 				  : NULL;
869 	destroy_cq(&chp->rhp->rdev, &chp->cq,
870 		   ucontext ? &ucontext->uctx : &chp->cq.rdev->uctx,
871 		   chp->destroy_skb);
872 	chp->destroy_skb = NULL;
873 	kfree(chp);
874 	return 0;
875 }
876 
877 struct ib_cq *c4iw_create_cq(struct ib_device *ibdev,
878 			     const struct ib_cq_init_attr *attr,
879 			     struct ib_ucontext *ib_context,
880 			     struct ib_udata *udata)
881 {
882 	int entries = attr->cqe;
883 	int vector = attr->comp_vector;
884 	struct c4iw_dev *rhp;
885 	struct c4iw_cq *chp;
886 	struct c4iw_create_cq_resp uresp;
887 	struct c4iw_ucontext *ucontext = NULL;
888 	int ret, wr_len;
889 	size_t memsize, hwentries;
890 	struct c4iw_mm_entry *mm, *mm2;
891 
892 	pr_debug("%s ib_dev %p entries %d\n", __func__, ibdev, entries);
893 	if (attr->flags)
894 		return ERR_PTR(-EINVAL);
895 
896 	rhp = to_c4iw_dev(ibdev);
897 
898 	if (vector >= rhp->rdev.lldi.nciq)
899 		return ERR_PTR(-EINVAL);
900 
901 	chp = kzalloc(sizeof(*chp), GFP_KERNEL);
902 	if (!chp)
903 		return ERR_PTR(-ENOMEM);
904 
905 	wr_len = sizeof(struct fw_ri_res_wr) + sizeof(struct fw_ri_res);
906 	chp->destroy_skb = alloc_skb(wr_len, GFP_KERNEL);
907 	if (!chp->destroy_skb) {
908 		ret = -ENOMEM;
909 		goto err1;
910 	}
911 
912 	if (ib_context)
913 		ucontext = to_c4iw_ucontext(ib_context);
914 
915 	/* account for the status page. */
916 	entries++;
917 
918 	/* IQ needs one extra entry to differentiate full vs empty. */
919 	entries++;
920 
921 	/*
922 	 * entries must be multiple of 16 for HW.
923 	 */
924 	entries = roundup(entries, 16);
925 
926 	/*
927 	 * Make actual HW queue 2x to avoid cdix_inc overflows.
928 	 */
929 	hwentries = min(entries * 2, rhp->rdev.hw_queue.t4_max_iq_size);
930 
931 	/*
932 	 * Make HW queue at least 64 entries so GTS updates aren't too
933 	 * frequent.
934 	 */
935 	if (hwentries < 64)
936 		hwentries = 64;
937 
938 	memsize = hwentries * sizeof *chp->cq.queue;
939 
940 	/*
941 	 * memsize must be a multiple of the page size if its a user cq.
942 	 */
943 	if (ucontext)
944 		memsize = roundup(memsize, PAGE_SIZE);
945 	chp->cq.size = hwentries;
946 	chp->cq.memsize = memsize;
947 	chp->cq.vector = vector;
948 
949 	ret = create_cq(&rhp->rdev, &chp->cq,
950 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
951 	if (ret)
952 		goto err2;
953 
954 	chp->rhp = rhp;
955 	chp->cq.size--;				/* status page */
956 	chp->ibcq.cqe = entries - 2;
957 	spin_lock_init(&chp->lock);
958 	spin_lock_init(&chp->comp_handler_lock);
959 	atomic_set(&chp->refcnt, 1);
960 	init_waitqueue_head(&chp->wait);
961 	ret = insert_handle(rhp, &rhp->cqidr, chp, chp->cq.cqid);
962 	if (ret)
963 		goto err3;
964 
965 	if (ucontext) {
966 		mm = kmalloc(sizeof *mm, GFP_KERNEL);
967 		if (!mm)
968 			goto err4;
969 		mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
970 		if (!mm2)
971 			goto err5;
972 
973 		uresp.qid_mask = rhp->rdev.cqmask;
974 		uresp.cqid = chp->cq.cqid;
975 		uresp.size = chp->cq.size;
976 		uresp.memsize = chp->cq.memsize;
977 		spin_lock(&ucontext->mmap_lock);
978 		uresp.key = ucontext->key;
979 		ucontext->key += PAGE_SIZE;
980 		uresp.gts_key = ucontext->key;
981 		ucontext->key += PAGE_SIZE;
982 		spin_unlock(&ucontext->mmap_lock);
983 		ret = ib_copy_to_udata(udata, &uresp,
984 				       sizeof(uresp) - sizeof(uresp.reserved));
985 		if (ret)
986 			goto err6;
987 
988 		mm->key = uresp.key;
989 		mm->addr = virt_to_phys(chp->cq.queue);
990 		mm->len = chp->cq.memsize;
991 		insert_mmap(ucontext, mm);
992 
993 		mm2->key = uresp.gts_key;
994 		mm2->addr = chp->cq.bar2_pa;
995 		mm2->len = PAGE_SIZE;
996 		insert_mmap(ucontext, mm2);
997 	}
998 	pr_debug("%s cqid 0x%0x chp %p size %u memsize %zu, dma_addr 0x%0llx\n",
999 		 __func__, chp->cq.cqid, chp, chp->cq.size,
1000 		 chp->cq.memsize, (unsigned long long)chp->cq.dma_addr);
1001 	return &chp->ibcq;
1002 err6:
1003 	kfree(mm2);
1004 err5:
1005 	kfree(mm);
1006 err4:
1007 	remove_handle(rhp, &rhp->cqidr, chp->cq.cqid);
1008 err3:
1009 	destroy_cq(&chp->rhp->rdev, &chp->cq,
1010 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx,
1011 		   chp->destroy_skb);
1012 err2:
1013 	kfree_skb(chp->destroy_skb);
1014 err1:
1015 	kfree(chp);
1016 	return ERR_PTR(ret);
1017 }
1018 
1019 int c4iw_resize_cq(struct ib_cq *cq, int cqe, struct ib_udata *udata)
1020 {
1021 	return -ENOSYS;
1022 }
1023 
1024 int c4iw_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
1025 {
1026 	struct c4iw_cq *chp;
1027 	int ret = 0;
1028 	unsigned long flag;
1029 
1030 	chp = to_c4iw_cq(ibcq);
1031 	spin_lock_irqsave(&chp->lock, flag);
1032 	t4_arm_cq(&chp->cq,
1033 		  (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED);
1034 	if (flags & IB_CQ_REPORT_MISSED_EVENTS)
1035 		ret = t4_cq_notempty(&chp->cq);
1036 	spin_unlock_irqrestore(&chp->lock, flag);
1037 	return ret;
1038 }
1039