xref: /openbmc/linux/drivers/infiniband/hw/cxgb4/cq.c (revision d236d361)
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 = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
48 	memset(res_wr, 0, wr_len);
49 	res_wr->op_nres = cpu_to_be32(
50 			FW_WR_OP_V(FW_RI_RES_WR) |
51 			FW_RI_RES_WR_NRES_V(1) |
52 			FW_WR_COMPL_F);
53 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
54 	res_wr->cookie = (uintptr_t)&wr_wait;
55 	res = res_wr->res;
56 	res->u.cq.restype = FW_RI_RES_TYPE_CQ;
57 	res->u.cq.op = FW_RI_RES_OP_RESET;
58 	res->u.cq.iqid = cpu_to_be32(cq->cqid);
59 
60 	c4iw_init_wr_wait(&wr_wait);
61 	ret = c4iw_ofld_send(rdev, skb);
62 	if (!ret) {
63 		ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, __func__);
64 	}
65 
66 	kfree(cq->sw_queue);
67 	dma_free_coherent(&(rdev->lldi.pdev->dev),
68 			  cq->memsize, cq->queue,
69 			  dma_unmap_addr(cq, mapping));
70 	c4iw_put_cqid(rdev, cq->cqid, uctx);
71 	return ret;
72 }
73 
74 static int create_cq(struct c4iw_rdev *rdev, struct t4_cq *cq,
75 		     struct c4iw_dev_ucontext *uctx)
76 {
77 	struct fw_ri_res_wr *res_wr;
78 	struct fw_ri_res *res;
79 	int wr_len;
80 	int user = (uctx != &rdev->uctx);
81 	struct c4iw_wr_wait wr_wait;
82 	int ret;
83 	struct sk_buff *skb;
84 
85 	cq->cqid = c4iw_get_cqid(rdev, uctx);
86 	if (!cq->cqid) {
87 		ret = -ENOMEM;
88 		goto err1;
89 	}
90 
91 	if (!user) {
92 		cq->sw_queue = kzalloc(cq->memsize, GFP_KERNEL);
93 		if (!cq->sw_queue) {
94 			ret = -ENOMEM;
95 			goto err2;
96 		}
97 	}
98 	cq->queue = dma_alloc_coherent(&rdev->lldi.pdev->dev, cq->memsize,
99 				       &cq->dma_addr, GFP_KERNEL);
100 	if (!cq->queue) {
101 		ret = -ENOMEM;
102 		goto err3;
103 	}
104 	dma_unmap_addr_set(cq, mapping, cq->dma_addr);
105 	memset(cq->queue, 0, cq->memsize);
106 
107 	/* build fw_ri_res_wr */
108 	wr_len = sizeof *res_wr + sizeof *res;
109 
110 	skb = alloc_skb(wr_len, GFP_KERNEL);
111 	if (!skb) {
112 		ret = -ENOMEM;
113 		goto err4;
114 	}
115 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
116 
117 	res_wr = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
118 	memset(res_wr, 0, wr_len);
119 	res_wr->op_nres = cpu_to_be32(
120 			FW_WR_OP_V(FW_RI_RES_WR) |
121 			FW_RI_RES_WR_NRES_V(1) |
122 			FW_WR_COMPL_F);
123 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
124 	res_wr->cookie = (uintptr_t)&wr_wait;
125 	res = res_wr->res;
126 	res->u.cq.restype = FW_RI_RES_TYPE_CQ;
127 	res->u.cq.op = FW_RI_RES_OP_WRITE;
128 	res->u.cq.iqid = cpu_to_be32(cq->cqid);
129 	res->u.cq.iqandst_to_iqandstindex = cpu_to_be32(
130 			FW_RI_RES_WR_IQANUS_V(0) |
131 			FW_RI_RES_WR_IQANUD_V(1) |
132 			FW_RI_RES_WR_IQANDST_F |
133 			FW_RI_RES_WR_IQANDSTINDEX_V(
134 				rdev->lldi.ciq_ids[cq->vector]));
135 	res->u.cq.iqdroprss_to_iqesize = cpu_to_be16(
136 			FW_RI_RES_WR_IQDROPRSS_F |
137 			FW_RI_RES_WR_IQPCIECH_V(2) |
138 			FW_RI_RES_WR_IQINTCNTTHRESH_V(0) |
139 			FW_RI_RES_WR_IQO_F |
140 			FW_RI_RES_WR_IQESIZE_V(1));
141 	res->u.cq.iqsize = cpu_to_be16(cq->size);
142 	res->u.cq.iqaddr = cpu_to_be64(cq->dma_addr);
143 
144 	c4iw_init_wr_wait(&wr_wait);
145 
146 	ret = c4iw_ofld_send(rdev, skb);
147 	if (ret)
148 		goto err4;
149 	pr_debug("%s wait_event wr_wait %p\n", __func__, &wr_wait);
150 	ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, __func__);
151 	if (ret)
152 		goto err4;
153 
154 	cq->gen = 1;
155 	cq->gts = rdev->lldi.gts_reg;
156 	cq->rdev = rdev;
157 
158 	cq->bar2_va = c4iw_bar2_addrs(rdev, cq->cqid, T4_BAR2_QTYPE_INGRESS,
159 				      &cq->bar2_qid,
160 				      user ? &cq->bar2_pa : NULL);
161 	if (user && !cq->bar2_pa) {
162 		pr_warn("%s: cqid %u not in BAR2 range\n",
163 			pci_name(rdev->lldi.pdev), cq->cqid);
164 		ret = -EINVAL;
165 		goto err4;
166 	}
167 	return 0;
168 err4:
169 	dma_free_coherent(&rdev->lldi.pdev->dev, cq->memsize, cq->queue,
170 			  dma_unmap_addr(cq, mapping));
171 err3:
172 	kfree(cq->sw_queue);
173 err2:
174 	c4iw_put_cqid(rdev, cq->cqid, uctx);
175 err1:
176 	return ret;
177 }
178 
179 static void insert_recv_cqe(struct t4_wq *wq, struct t4_cq *cq)
180 {
181 	struct t4_cqe cqe;
182 
183 	pr_debug("%s wq %p cq %p sw_cidx %u sw_pidx %u\n", __func__,
184 		 wq, cq, cq->sw_cidx, cq->sw_pidx);
185 	memset(&cqe, 0, sizeof(cqe));
186 	cqe.header = cpu_to_be32(CQE_STATUS_V(T4_ERR_SWFLUSH) |
187 				 CQE_OPCODE_V(FW_RI_SEND) |
188 				 CQE_TYPE_V(0) |
189 				 CQE_SWCQE_V(1) |
190 				 CQE_QPID_V(wq->sq.qid));
191 	cqe.bits_type_ts = cpu_to_be64(CQE_GENBIT_V((u64)cq->gen));
192 	cq->sw_queue[cq->sw_pidx] = cqe;
193 	t4_swcq_produce(cq);
194 }
195 
196 int c4iw_flush_rq(struct t4_wq *wq, struct t4_cq *cq, int count)
197 {
198 	int flushed = 0;
199 	int in_use = wq->rq.in_use - count;
200 
201 	BUG_ON(in_use < 0);
202 	pr_debug("%s wq %p cq %p rq.in_use %u skip count %u\n", __func__,
203 		 wq, cq, wq->rq.in_use, count);
204 	while (in_use--) {
205 		insert_recv_cqe(wq, cq);
206 		flushed++;
207 	}
208 	return flushed;
209 }
210 
211 static void insert_sq_cqe(struct t4_wq *wq, struct t4_cq *cq,
212 			  struct t4_swsqe *swcqe)
213 {
214 	struct t4_cqe cqe;
215 
216 	pr_debug("%s wq %p cq %p sw_cidx %u sw_pidx %u\n", __func__,
217 		 wq, cq, cq->sw_cidx, cq->sw_pidx);
218 	memset(&cqe, 0, sizeof(cqe));
219 	cqe.header = cpu_to_be32(CQE_STATUS_V(T4_ERR_SWFLUSH) |
220 				 CQE_OPCODE_V(swcqe->opcode) |
221 				 CQE_TYPE_V(1) |
222 				 CQE_SWCQE_V(1) |
223 				 CQE_QPID_V(wq->sq.qid));
224 	CQE_WRID_SQ_IDX(&cqe) = swcqe->idx;
225 	cqe.bits_type_ts = cpu_to_be64(CQE_GENBIT_V((u64)cq->gen));
226 	cq->sw_queue[cq->sw_pidx] = cqe;
227 	t4_swcq_produce(cq);
228 }
229 
230 static void advance_oldest_read(struct t4_wq *wq);
231 
232 int c4iw_flush_sq(struct c4iw_qp *qhp)
233 {
234 	int flushed = 0;
235 	struct t4_wq *wq = &qhp->wq;
236 	struct c4iw_cq *chp = to_c4iw_cq(qhp->ibqp.send_cq);
237 	struct t4_cq *cq = &chp->cq;
238 	int idx;
239 	struct t4_swsqe *swsqe;
240 
241 	if (wq->sq.flush_cidx == -1)
242 		wq->sq.flush_cidx = wq->sq.cidx;
243 	idx = wq->sq.flush_cidx;
244 	BUG_ON(idx >= wq->sq.size);
245 	while (idx != wq->sq.pidx) {
246 		swsqe = &wq->sq.sw_sq[idx];
247 		BUG_ON(swsqe->flushed);
248 		swsqe->flushed = 1;
249 		insert_sq_cqe(wq, cq, swsqe);
250 		if (wq->sq.oldest_read == swsqe) {
251 			BUG_ON(swsqe->opcode != FW_RI_READ_REQ);
252 			advance_oldest_read(wq);
253 		}
254 		flushed++;
255 		if (++idx == wq->sq.size)
256 			idx = 0;
257 	}
258 	wq->sq.flush_cidx += flushed;
259 	if (wq->sq.flush_cidx >= wq->sq.size)
260 		wq->sq.flush_cidx -= wq->sq.size;
261 	return flushed;
262 }
263 
264 static void flush_completed_wrs(struct t4_wq *wq, struct t4_cq *cq)
265 {
266 	struct t4_swsqe *swsqe;
267 	int cidx;
268 
269 	if (wq->sq.flush_cidx == -1)
270 		wq->sq.flush_cidx = wq->sq.cidx;
271 	cidx = wq->sq.flush_cidx;
272 	BUG_ON(cidx > wq->sq.size);
273 
274 	while (cidx != wq->sq.pidx) {
275 		swsqe = &wq->sq.sw_sq[cidx];
276 		if (!swsqe->signaled) {
277 			if (++cidx == wq->sq.size)
278 				cidx = 0;
279 		} else if (swsqe->complete) {
280 
281 			BUG_ON(swsqe->flushed);
282 
283 			/*
284 			 * Insert this completed cqe into the swcq.
285 			 */
286 			pr_debug("%s moving cqe into swcq sq idx %u cq idx %u\n",
287 				 __func__, cidx, cq->sw_pidx);
288 			swsqe->cqe.header |= htonl(CQE_SWCQE_V(1));
289 			cq->sw_queue[cq->sw_pidx] = swsqe->cqe;
290 			t4_swcq_produce(cq);
291 			swsqe->flushed = 1;
292 			if (++cidx == wq->sq.size)
293 				cidx = 0;
294 			wq->sq.flush_cidx = cidx;
295 		} else
296 			break;
297 	}
298 }
299 
300 static void create_read_req_cqe(struct t4_wq *wq, struct t4_cqe *hw_cqe,
301 		struct t4_cqe *read_cqe)
302 {
303 	read_cqe->u.scqe.cidx = wq->sq.oldest_read->idx;
304 	read_cqe->len = htonl(wq->sq.oldest_read->read_len);
305 	read_cqe->header = htonl(CQE_QPID_V(CQE_QPID(hw_cqe)) |
306 			CQE_SWCQE_V(SW_CQE(hw_cqe)) |
307 			CQE_OPCODE_V(FW_RI_READ_REQ) |
308 			CQE_TYPE_V(1));
309 	read_cqe->bits_type_ts = hw_cqe->bits_type_ts;
310 }
311 
312 static void advance_oldest_read(struct t4_wq *wq)
313 {
314 
315 	u32 rptr = wq->sq.oldest_read - wq->sq.sw_sq + 1;
316 
317 	if (rptr == wq->sq.size)
318 		rptr = 0;
319 	while (rptr != wq->sq.pidx) {
320 		wq->sq.oldest_read = &wq->sq.sw_sq[rptr];
321 
322 		if (wq->sq.oldest_read->opcode == FW_RI_READ_REQ)
323 			return;
324 		if (++rptr == wq->sq.size)
325 			rptr = 0;
326 	}
327 	wq->sq.oldest_read = NULL;
328 }
329 
330 /*
331  * Move all CQEs from the HWCQ into the SWCQ.
332  * Deal with out-of-order and/or completions that complete
333  * prior unsignalled WRs.
334  */
335 void c4iw_flush_hw_cq(struct c4iw_cq *chp)
336 {
337 	struct t4_cqe *hw_cqe, *swcqe, read_cqe;
338 	struct c4iw_qp *qhp;
339 	struct t4_swsqe *swsqe;
340 	int ret;
341 
342 	pr_debug("%s  cqid 0x%x\n", __func__, chp->cq.cqid);
343 	ret = t4_next_hw_cqe(&chp->cq, &hw_cqe);
344 
345 	/*
346 	 * This logic is similar to poll_cq(), but not quite the same
347 	 * unfortunately.  Need to move pertinent HW CQEs to the SW CQ but
348 	 * also do any translation magic that poll_cq() normally does.
349 	 */
350 	while (!ret) {
351 		qhp = get_qhp(chp->rhp, CQE_QPID(hw_cqe));
352 
353 		/*
354 		 * drop CQEs with no associated QP
355 		 */
356 		if (qhp == NULL)
357 			goto next_cqe;
358 
359 		if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE)
360 			goto next_cqe;
361 
362 		if (CQE_OPCODE(hw_cqe) == FW_RI_READ_RESP) {
363 
364 			/* If we have reached here because of async
365 			 * event or other error, and have egress error
366 			 * then drop
367 			 */
368 			if (CQE_TYPE(hw_cqe) == 1)
369 				goto next_cqe;
370 
371 			/* drop peer2peer RTR reads.
372 			 */
373 			if (CQE_WRID_STAG(hw_cqe) == 1)
374 				goto next_cqe;
375 
376 			/*
377 			 * Eat completions for unsignaled read WRs.
378 			 */
379 			if (!qhp->wq.sq.oldest_read->signaled) {
380 				advance_oldest_read(&qhp->wq);
381 				goto next_cqe;
382 			}
383 
384 			/*
385 			 * Don't write to the HWCQ, create a new read req CQE
386 			 * in local memory and move it into the swcq.
387 			 */
388 			create_read_req_cqe(&qhp->wq, hw_cqe, &read_cqe);
389 			hw_cqe = &read_cqe;
390 			advance_oldest_read(&qhp->wq);
391 		}
392 
393 		/* if its a SQ completion, then do the magic to move all the
394 		 * unsignaled and now in-order completions into the swcq.
395 		 */
396 		if (SQ_TYPE(hw_cqe)) {
397 			swsqe = &qhp->wq.sq.sw_sq[CQE_WRID_SQ_IDX(hw_cqe)];
398 			swsqe->cqe = *hw_cqe;
399 			swsqe->complete = 1;
400 			flush_completed_wrs(&qhp->wq, &chp->cq);
401 		} else {
402 			swcqe = &chp->cq.sw_queue[chp->cq.sw_pidx];
403 			*swcqe = *hw_cqe;
404 			swcqe->header |= cpu_to_be32(CQE_SWCQE_V(1));
405 			t4_swcq_produce(&chp->cq);
406 		}
407 next_cqe:
408 		t4_hwcq_consume(&chp->cq);
409 		ret = t4_next_hw_cqe(&chp->cq, &hw_cqe);
410 	}
411 }
412 
413 static int cqe_completes_wr(struct t4_cqe *cqe, struct t4_wq *wq)
414 {
415 	if (CQE_OPCODE(cqe) == FW_RI_TERMINATE)
416 		return 0;
417 
418 	if ((CQE_OPCODE(cqe) == FW_RI_RDMA_WRITE) && RQ_TYPE(cqe))
419 		return 0;
420 
421 	if ((CQE_OPCODE(cqe) == FW_RI_READ_RESP) && SQ_TYPE(cqe))
422 		return 0;
423 
424 	if (CQE_SEND_OPCODE(cqe) && RQ_TYPE(cqe) && t4_rq_empty(wq))
425 		return 0;
426 	return 1;
427 }
428 
429 void c4iw_count_rcqes(struct t4_cq *cq, struct t4_wq *wq, int *count)
430 {
431 	struct t4_cqe *cqe;
432 	u32 ptr;
433 
434 	*count = 0;
435 	pr_debug("%s count zero %d\n", __func__, *count);
436 	ptr = cq->sw_cidx;
437 	while (ptr != cq->sw_pidx) {
438 		cqe = &cq->sw_queue[ptr];
439 		if (RQ_TYPE(cqe) && (CQE_OPCODE(cqe) != FW_RI_READ_RESP) &&
440 		    (CQE_QPID(cqe) == wq->sq.qid) && cqe_completes_wr(cqe, wq))
441 			(*count)++;
442 		if (++ptr == cq->size)
443 			ptr = 0;
444 	}
445 	pr_debug("%s cq %p count %d\n", __func__, cq, *count);
446 }
447 
448 /*
449  * poll_cq
450  *
451  * Caller must:
452  *     check the validity of the first CQE,
453  *     supply the wq assicated with the qpid.
454  *
455  * credit: cq credit to return to sge.
456  * cqe_flushed: 1 iff the CQE is flushed.
457  * cqe: copy of the polled CQE.
458  *
459  * return value:
460  *    0		    CQE returned ok.
461  *    -EAGAIN       CQE skipped, try again.
462  *    -EOVERFLOW    CQ overflow detected.
463  */
464 static int poll_cq(struct t4_wq *wq, struct t4_cq *cq, struct t4_cqe *cqe,
465 		   u8 *cqe_flushed, u64 *cookie, u32 *credit)
466 {
467 	int ret = 0;
468 	struct t4_cqe *hw_cqe, read_cqe;
469 
470 	*cqe_flushed = 0;
471 	*credit = 0;
472 	ret = t4_next_cqe(cq, &hw_cqe);
473 	if (ret)
474 		return ret;
475 
476 	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",
477 		 __func__, CQE_OVFBIT(hw_cqe), CQE_QPID(hw_cqe),
478 		 CQE_GENBIT(hw_cqe), CQE_TYPE(hw_cqe), CQE_STATUS(hw_cqe),
479 		 CQE_OPCODE(hw_cqe), CQE_LEN(hw_cqe), CQE_WRID_HI(hw_cqe),
480 		 CQE_WRID_LOW(hw_cqe));
481 
482 	/*
483 	 * skip cqe's not affiliated with a QP.
484 	 */
485 	if (wq == NULL) {
486 		ret = -EAGAIN;
487 		goto skip_cqe;
488 	}
489 
490 	/*
491 	* skip hw cqe's if the wq is flushed.
492 	*/
493 	if (wq->flushed && !SW_CQE(hw_cqe)) {
494 		ret = -EAGAIN;
495 		goto skip_cqe;
496 	}
497 
498 	/*
499 	 * skip TERMINATE cqes...
500 	 */
501 	if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE) {
502 		ret = -EAGAIN;
503 		goto skip_cqe;
504 	}
505 
506 	/*
507 	 * Special cqe for drain WR completions...
508 	 */
509 	if (CQE_OPCODE(hw_cqe) == C4IW_DRAIN_OPCODE) {
510 		*cookie = CQE_DRAIN_COOKIE(hw_cqe);
511 		*cqe = *hw_cqe;
512 		goto skip_cqe;
513 	}
514 
515 	/*
516 	 * Gotta tweak READ completions:
517 	 *	1) the cqe doesn't contain the sq_wptr from the wr.
518 	 *	2) opcode not reflected from the wr.
519 	 *	3) read_len not reflected from the wr.
520 	 *	4) cq_type is RQ_TYPE not SQ_TYPE.
521 	 */
522 	if (RQ_TYPE(hw_cqe) && (CQE_OPCODE(hw_cqe) == FW_RI_READ_RESP)) {
523 
524 		/* If we have reached here because of async
525 		 * event or other error, and have egress error
526 		 * then drop
527 		 */
528 		if (CQE_TYPE(hw_cqe) == 1) {
529 			if (CQE_STATUS(hw_cqe))
530 				t4_set_wq_in_error(wq);
531 			ret = -EAGAIN;
532 			goto skip_cqe;
533 		}
534 
535 		/* If this is an unsolicited read response, then the read
536 		 * was generated by the kernel driver as part of peer-2-peer
537 		 * connection setup.  So ignore the completion.
538 		 */
539 		if (CQE_WRID_STAG(hw_cqe) == 1) {
540 			if (CQE_STATUS(hw_cqe))
541 				t4_set_wq_in_error(wq);
542 			ret = -EAGAIN;
543 			goto skip_cqe;
544 		}
545 
546 		/*
547 		 * Eat completions for unsignaled read WRs.
548 		 */
549 		if (!wq->sq.oldest_read->signaled) {
550 			advance_oldest_read(wq);
551 			ret = -EAGAIN;
552 			goto skip_cqe;
553 		}
554 
555 		/*
556 		 * Don't write to the HWCQ, so create a new read req CQE
557 		 * in local memory.
558 		 */
559 		create_read_req_cqe(wq, hw_cqe, &read_cqe);
560 		hw_cqe = &read_cqe;
561 		advance_oldest_read(wq);
562 	}
563 
564 	if (CQE_STATUS(hw_cqe) || t4_wq_in_error(wq)) {
565 		*cqe_flushed = (CQE_STATUS(hw_cqe) == T4_ERR_SWFLUSH);
566 		t4_set_wq_in_error(wq);
567 	}
568 
569 	/*
570 	 * RECV completion.
571 	 */
572 	if (RQ_TYPE(hw_cqe)) {
573 
574 		/*
575 		 * HW only validates 4 bits of MSN.  So we must validate that
576 		 * the MSN in the SEND is the next expected MSN.  If its not,
577 		 * then we complete this with T4_ERR_MSN and mark the wq in
578 		 * error.
579 		 */
580 
581 		if (t4_rq_empty(wq)) {
582 			t4_set_wq_in_error(wq);
583 			ret = -EAGAIN;
584 			goto skip_cqe;
585 		}
586 		if (unlikely((CQE_WRID_MSN(hw_cqe) != (wq->rq.msn)))) {
587 			t4_set_wq_in_error(wq);
588 			hw_cqe->header |= htonl(CQE_STATUS_V(T4_ERR_MSN));
589 			goto proc_cqe;
590 		}
591 		goto proc_cqe;
592 	}
593 
594 	/*
595 	 * If we get here its a send completion.
596 	 *
597 	 * Handle out of order completion. These get stuffed
598 	 * in the SW SQ. Then the SW SQ is walked to move any
599 	 * now in-order completions into the SW CQ.  This handles
600 	 * 2 cases:
601 	 *	1) reaping unsignaled WRs when the first subsequent
602 	 *	   signaled WR is completed.
603 	 *	2) out of order read completions.
604 	 */
605 	if (!SW_CQE(hw_cqe) && (CQE_WRID_SQ_IDX(hw_cqe) != wq->sq.cidx)) {
606 		struct t4_swsqe *swsqe;
607 
608 		pr_debug("%s out of order completion going in sw_sq at idx %u\n",
609 			 __func__, CQE_WRID_SQ_IDX(hw_cqe));
610 		swsqe = &wq->sq.sw_sq[CQE_WRID_SQ_IDX(hw_cqe)];
611 		swsqe->cqe = *hw_cqe;
612 		swsqe->complete = 1;
613 		ret = -EAGAIN;
614 		goto flush_wq;
615 	}
616 
617 proc_cqe:
618 	*cqe = *hw_cqe;
619 
620 	/*
621 	 * Reap the associated WR(s) that are freed up with this
622 	 * completion.
623 	 */
624 	if (SQ_TYPE(hw_cqe)) {
625 		int idx = CQE_WRID_SQ_IDX(hw_cqe);
626 		BUG_ON(idx >= wq->sq.size);
627 
628 		/*
629 		* Account for any unsignaled completions completed by
630 		* this signaled completion.  In this case, cidx points
631 		* to the first unsignaled one, and idx points to the
632 		* signaled one.  So adjust in_use based on this delta.
633 		* if this is not completing any unsigned wrs, then the
634 		* delta will be 0. Handle wrapping also!
635 		*/
636 		if (idx < wq->sq.cidx)
637 			wq->sq.in_use -= wq->sq.size + idx - wq->sq.cidx;
638 		else
639 			wq->sq.in_use -= idx - wq->sq.cidx;
640 		BUG_ON(wq->sq.in_use <= 0 && wq->sq.in_use >= wq->sq.size);
641 
642 		wq->sq.cidx = (uint16_t)idx;
643 		pr_debug("%s completing sq idx %u\n", __func__, wq->sq.cidx);
644 		*cookie = wq->sq.sw_sq[wq->sq.cidx].wr_id;
645 		if (c4iw_wr_log)
646 			c4iw_log_wr_stats(wq, hw_cqe);
647 		t4_sq_consume(wq);
648 	} else {
649 		pr_debug("%s completing rq idx %u\n", __func__, wq->rq.cidx);
650 		*cookie = wq->rq.sw_rq[wq->rq.cidx].wr_id;
651 		BUG_ON(t4_rq_empty(wq));
652 		if (c4iw_wr_log)
653 			c4iw_log_wr_stats(wq, hw_cqe);
654 		t4_rq_consume(wq);
655 		goto skip_cqe;
656 	}
657 
658 flush_wq:
659 	/*
660 	 * Flush any completed cqes that are now in-order.
661 	 */
662 	flush_completed_wrs(wq, cq);
663 
664 skip_cqe:
665 	if (SW_CQE(hw_cqe)) {
666 		pr_debug("%s cq %p cqid 0x%x skip sw cqe cidx %u\n",
667 			 __func__, cq, cq->cqid, cq->sw_cidx);
668 		t4_swcq_consume(cq);
669 	} else {
670 		pr_debug("%s cq %p cqid 0x%x skip hw cqe cidx %u\n",
671 			 __func__, cq, cq->cqid, cq->cidx);
672 		t4_hwcq_consume(cq);
673 	}
674 	return ret;
675 }
676 
677 /*
678  * Get one cq entry from c4iw and map it to openib.
679  *
680  * Returns:
681  *	0			cqe returned
682  *	-ENODATA		EMPTY;
683  *	-EAGAIN			caller must try again
684  *	any other -errno	fatal error
685  */
686 static int c4iw_poll_cq_one(struct c4iw_cq *chp, struct ib_wc *wc)
687 {
688 	struct c4iw_qp *qhp = NULL;
689 	struct t4_cqe uninitialized_var(cqe), *rd_cqe;
690 	struct t4_wq *wq;
691 	u32 credit = 0;
692 	u8 cqe_flushed;
693 	u64 cookie = 0;
694 	int ret;
695 
696 	ret = t4_next_cqe(&chp->cq, &rd_cqe);
697 
698 	if (ret)
699 		return ret;
700 
701 	qhp = get_qhp(chp->rhp, CQE_QPID(rd_cqe));
702 	if (!qhp)
703 		wq = NULL;
704 	else {
705 		spin_lock(&qhp->lock);
706 		wq = &(qhp->wq);
707 	}
708 	ret = poll_cq(wq, &(chp->cq), &cqe, &cqe_flushed, &cookie, &credit);
709 	if (ret)
710 		goto out;
711 
712 	wc->wr_id = cookie;
713 	wc->qp = &qhp->ibqp;
714 	wc->vendor_err = CQE_STATUS(&cqe);
715 	wc->wc_flags = 0;
716 
717 	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",
718 		 __func__, CQE_QPID(&cqe),
719 		 CQE_TYPE(&cqe), CQE_OPCODE(&cqe),
720 		 CQE_STATUS(&cqe), CQE_LEN(&cqe),
721 		 CQE_WRID_HI(&cqe), CQE_WRID_LOW(&cqe),
722 		 (unsigned long long)cookie);
723 
724 	if (CQE_TYPE(&cqe) == 0) {
725 		if (!CQE_STATUS(&cqe))
726 			wc->byte_len = CQE_LEN(&cqe);
727 		else
728 			wc->byte_len = 0;
729 		wc->opcode = IB_WC_RECV;
730 		if (CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_INV ||
731 		    CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_SE_INV) {
732 			wc->ex.invalidate_rkey = CQE_WRID_STAG(&cqe);
733 			wc->wc_flags |= IB_WC_WITH_INVALIDATE;
734 			c4iw_invalidate_mr(qhp->rhp, wc->ex.invalidate_rkey);
735 		}
736 	} else {
737 		switch (CQE_OPCODE(&cqe)) {
738 		case FW_RI_RDMA_WRITE:
739 			wc->opcode = IB_WC_RDMA_WRITE;
740 			break;
741 		case FW_RI_READ_REQ:
742 			wc->opcode = IB_WC_RDMA_READ;
743 			wc->byte_len = CQE_LEN(&cqe);
744 			break;
745 		case FW_RI_SEND_WITH_INV:
746 		case FW_RI_SEND_WITH_SE_INV:
747 			wc->opcode = IB_WC_SEND;
748 			wc->wc_flags |= IB_WC_WITH_INVALIDATE;
749 			break;
750 		case FW_RI_SEND:
751 		case FW_RI_SEND_WITH_SE:
752 			wc->opcode = IB_WC_SEND;
753 			break;
754 
755 		case FW_RI_LOCAL_INV:
756 			wc->opcode = IB_WC_LOCAL_INV;
757 			break;
758 		case FW_RI_FAST_REGISTER:
759 			wc->opcode = IB_WC_REG_MR;
760 
761 			/* Invalidate the MR if the fastreg failed */
762 			if (CQE_STATUS(&cqe) != T4_ERR_SUCCESS)
763 				c4iw_invalidate_mr(qhp->rhp,
764 						   CQE_WRID_FR_STAG(&cqe));
765 			break;
766 		case C4IW_DRAIN_OPCODE:
767 			wc->opcode = IB_WC_SEND;
768 			break;
769 		default:
770 			pr_err("Unexpected opcode %d in the CQE received for QPID=0x%0x\n",
771 			       CQE_OPCODE(&cqe), CQE_QPID(&cqe));
772 			ret = -EINVAL;
773 			goto out;
774 		}
775 	}
776 
777 	if (cqe_flushed)
778 		wc->status = IB_WC_WR_FLUSH_ERR;
779 	else {
780 
781 		switch (CQE_STATUS(&cqe)) {
782 		case T4_ERR_SUCCESS:
783 			wc->status = IB_WC_SUCCESS;
784 			break;
785 		case T4_ERR_STAG:
786 			wc->status = IB_WC_LOC_ACCESS_ERR;
787 			break;
788 		case T4_ERR_PDID:
789 			wc->status = IB_WC_LOC_PROT_ERR;
790 			break;
791 		case T4_ERR_QPID:
792 		case T4_ERR_ACCESS:
793 			wc->status = IB_WC_LOC_ACCESS_ERR;
794 			break;
795 		case T4_ERR_WRAP:
796 			wc->status = IB_WC_GENERAL_ERR;
797 			break;
798 		case T4_ERR_BOUND:
799 			wc->status = IB_WC_LOC_LEN_ERR;
800 			break;
801 		case T4_ERR_INVALIDATE_SHARED_MR:
802 		case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
803 			wc->status = IB_WC_MW_BIND_ERR;
804 			break;
805 		case T4_ERR_CRC:
806 		case T4_ERR_MARKER:
807 		case T4_ERR_PDU_LEN_ERR:
808 		case T4_ERR_OUT_OF_RQE:
809 		case T4_ERR_DDP_VERSION:
810 		case T4_ERR_RDMA_VERSION:
811 		case T4_ERR_DDP_QUEUE_NUM:
812 		case T4_ERR_MSN:
813 		case T4_ERR_TBIT:
814 		case T4_ERR_MO:
815 		case T4_ERR_MSN_RANGE:
816 		case T4_ERR_IRD_OVERFLOW:
817 		case T4_ERR_OPCODE:
818 		case T4_ERR_INTERNAL_ERR:
819 			wc->status = IB_WC_FATAL_ERR;
820 			break;
821 		case T4_ERR_SWFLUSH:
822 			wc->status = IB_WC_WR_FLUSH_ERR;
823 			break;
824 		default:
825 			pr_err("Unexpected cqe_status 0x%x for QPID=0x%0x\n",
826 			       CQE_STATUS(&cqe), CQE_QPID(&cqe));
827 			wc->status = IB_WC_FATAL_ERR;
828 		}
829 	}
830 out:
831 	if (wq)
832 		spin_unlock(&qhp->lock);
833 	return ret;
834 }
835 
836 int c4iw_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
837 {
838 	struct c4iw_cq *chp;
839 	unsigned long flags;
840 	int npolled;
841 	int err = 0;
842 
843 	chp = to_c4iw_cq(ibcq);
844 
845 	spin_lock_irqsave(&chp->lock, flags);
846 	for (npolled = 0; npolled < num_entries; ++npolled) {
847 		do {
848 			err = c4iw_poll_cq_one(chp, wc + npolled);
849 		} while (err == -EAGAIN);
850 		if (err)
851 			break;
852 	}
853 	spin_unlock_irqrestore(&chp->lock, flags);
854 	return !err || err == -ENODATA ? npolled : err;
855 }
856 
857 int c4iw_destroy_cq(struct ib_cq *ib_cq)
858 {
859 	struct c4iw_cq *chp;
860 	struct c4iw_ucontext *ucontext;
861 
862 	pr_debug("%s ib_cq %p\n", __func__, ib_cq);
863 	chp = to_c4iw_cq(ib_cq);
864 
865 	remove_handle(chp->rhp, &chp->rhp->cqidr, chp->cq.cqid);
866 	atomic_dec(&chp->refcnt);
867 	wait_event(chp->wait, !atomic_read(&chp->refcnt));
868 
869 	ucontext = ib_cq->uobject ? to_c4iw_ucontext(ib_cq->uobject->context)
870 				  : NULL;
871 	destroy_cq(&chp->rhp->rdev, &chp->cq,
872 		   ucontext ? &ucontext->uctx : &chp->cq.rdev->uctx,
873 		   chp->destroy_skb);
874 	chp->destroy_skb = NULL;
875 	kfree(chp);
876 	return 0;
877 }
878 
879 struct ib_cq *c4iw_create_cq(struct ib_device *ibdev,
880 			     const struct ib_cq_init_attr *attr,
881 			     struct ib_ucontext *ib_context,
882 			     struct ib_udata *udata)
883 {
884 	int entries = attr->cqe;
885 	int vector = attr->comp_vector;
886 	struct c4iw_dev *rhp;
887 	struct c4iw_cq *chp;
888 	struct c4iw_create_cq_resp uresp;
889 	struct c4iw_ucontext *ucontext = NULL;
890 	int ret, wr_len;
891 	size_t memsize, hwentries;
892 	struct c4iw_mm_entry *mm, *mm2;
893 
894 	pr_debug("%s ib_dev %p entries %d\n", __func__, ibdev, entries);
895 	if (attr->flags)
896 		return ERR_PTR(-EINVAL);
897 
898 	rhp = to_c4iw_dev(ibdev);
899 
900 	if (vector >= rhp->rdev.lldi.nciq)
901 		return ERR_PTR(-EINVAL);
902 
903 	chp = kzalloc(sizeof(*chp), GFP_KERNEL);
904 	if (!chp)
905 		return ERR_PTR(-ENOMEM);
906 
907 	wr_len = sizeof(struct fw_ri_res_wr) + sizeof(struct fw_ri_res);
908 	chp->destroy_skb = alloc_skb(wr_len, GFP_KERNEL);
909 	if (!chp->destroy_skb) {
910 		ret = -ENOMEM;
911 		goto err1;
912 	}
913 
914 	if (ib_context)
915 		ucontext = to_c4iw_ucontext(ib_context);
916 
917 	/* account for the status page. */
918 	entries++;
919 
920 	/* IQ needs one extra entry to differentiate full vs empty. */
921 	entries++;
922 
923 	/*
924 	 * entries must be multiple of 16 for HW.
925 	 */
926 	entries = roundup(entries, 16);
927 
928 	/*
929 	 * Make actual HW queue 2x to avoid cdix_inc overflows.
930 	 */
931 	hwentries = min(entries * 2, rhp->rdev.hw_queue.t4_max_iq_size);
932 
933 	/*
934 	 * Make HW queue at least 64 entries so GTS updates aren't too
935 	 * frequent.
936 	 */
937 	if (hwentries < 64)
938 		hwentries = 64;
939 
940 	memsize = hwentries * sizeof *chp->cq.queue;
941 
942 	/*
943 	 * memsize must be a multiple of the page size if its a user cq.
944 	 */
945 	if (ucontext)
946 		memsize = roundup(memsize, PAGE_SIZE);
947 	chp->cq.size = hwentries;
948 	chp->cq.memsize = memsize;
949 	chp->cq.vector = vector;
950 
951 	ret = create_cq(&rhp->rdev, &chp->cq,
952 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
953 	if (ret)
954 		goto err2;
955 
956 	chp->rhp = rhp;
957 	chp->cq.size--;				/* status page */
958 	chp->ibcq.cqe = entries - 2;
959 	spin_lock_init(&chp->lock);
960 	spin_lock_init(&chp->comp_handler_lock);
961 	atomic_set(&chp->refcnt, 1);
962 	init_waitqueue_head(&chp->wait);
963 	ret = insert_handle(rhp, &rhp->cqidr, chp, chp->cq.cqid);
964 	if (ret)
965 		goto err3;
966 
967 	if (ucontext) {
968 		mm = kmalloc(sizeof *mm, GFP_KERNEL);
969 		if (!mm)
970 			goto err4;
971 		mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
972 		if (!mm2)
973 			goto err5;
974 
975 		uresp.qid_mask = rhp->rdev.cqmask;
976 		uresp.cqid = chp->cq.cqid;
977 		uresp.size = chp->cq.size;
978 		uresp.memsize = chp->cq.memsize;
979 		spin_lock(&ucontext->mmap_lock);
980 		uresp.key = ucontext->key;
981 		ucontext->key += PAGE_SIZE;
982 		uresp.gts_key = ucontext->key;
983 		ucontext->key += PAGE_SIZE;
984 		spin_unlock(&ucontext->mmap_lock);
985 		ret = ib_copy_to_udata(udata, &uresp,
986 				       sizeof(uresp) - sizeof(uresp.reserved));
987 		if (ret)
988 			goto err6;
989 
990 		mm->key = uresp.key;
991 		mm->addr = virt_to_phys(chp->cq.queue);
992 		mm->len = chp->cq.memsize;
993 		insert_mmap(ucontext, mm);
994 
995 		mm2->key = uresp.gts_key;
996 		mm2->addr = chp->cq.bar2_pa;
997 		mm2->len = PAGE_SIZE;
998 		insert_mmap(ucontext, mm2);
999 	}
1000 	pr_debug("%s cqid 0x%0x chp %p size %u memsize %zu, dma_addr 0x%0llx\n",
1001 		 __func__, chp->cq.cqid, chp, chp->cq.size,
1002 		 chp->cq.memsize, (unsigned long long)chp->cq.dma_addr);
1003 	return &chp->ibcq;
1004 err6:
1005 	kfree(mm2);
1006 err5:
1007 	kfree(mm);
1008 err4:
1009 	remove_handle(rhp, &rhp->cqidr, chp->cq.cqid);
1010 err3:
1011 	destroy_cq(&chp->rhp->rdev, &chp->cq,
1012 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx,
1013 		   chp->destroy_skb);
1014 err2:
1015 	kfree_skb(chp->destroy_skb);
1016 err1:
1017 	kfree(chp);
1018 	return ERR_PTR(ret);
1019 }
1020 
1021 int c4iw_resize_cq(struct ib_cq *cq, int cqe, struct ib_udata *udata)
1022 {
1023 	return -ENOSYS;
1024 }
1025 
1026 int c4iw_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
1027 {
1028 	struct c4iw_cq *chp;
1029 	int ret = 0;
1030 	unsigned long flag;
1031 
1032 	chp = to_c4iw_cq(ibcq);
1033 	spin_lock_irqsave(&chp->lock, flag);
1034 	t4_arm_cq(&chp->cq,
1035 		  (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED);
1036 	if (flags & IB_CQ_REPORT_MISSED_EVENTS)
1037 		ret = t4_cq_notempty(&chp->cq);
1038 	spin_unlock_irqrestore(&chp->lock, flag);
1039 	return ret;
1040 }
1041