xref: /openbmc/linux/drivers/infiniband/hw/cxgb4/cq.c (revision bc5aa3a0)
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 	PDBG("%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(MOD "%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 	PDBG("%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 	PDBG("%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 	PDBG("%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 			PDBG("%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 	PDBG("%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 	PDBG("%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 	PDBG("%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 	PDBG("%s CQE OVF %u qpid 0x%0x genbit %u type %u status 0x%0x"
477 	     " opcode 0x%0x len 0x%0x wrid_hi_stag 0x%x wrid_low_msn 0x%x\n",
478 	     __func__, CQE_OVFBIT(hw_cqe), CQE_QPID(hw_cqe),
479 	     CQE_GENBIT(hw_cqe), CQE_TYPE(hw_cqe), CQE_STATUS(hw_cqe),
480 	     CQE_OPCODE(hw_cqe), CQE_LEN(hw_cqe), CQE_WRID_HI(hw_cqe),
481 	     CQE_WRID_LOW(hw_cqe));
482 
483 	/*
484 	 * skip cqe's not affiliated with a QP.
485 	 */
486 	if (wq == NULL) {
487 		ret = -EAGAIN;
488 		goto skip_cqe;
489 	}
490 
491 	/*
492 	* skip hw cqe's if the wq is flushed.
493 	*/
494 	if (wq->flushed && !SW_CQE(hw_cqe)) {
495 		ret = -EAGAIN;
496 		goto skip_cqe;
497 	}
498 
499 	/*
500 	 * skip TERMINATE cqes...
501 	 */
502 	if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE) {
503 		ret = -EAGAIN;
504 		goto skip_cqe;
505 	}
506 
507 	/*
508 	 * Gotta tweak READ completions:
509 	 *	1) the cqe doesn't contain the sq_wptr from the wr.
510 	 *	2) opcode not reflected from the wr.
511 	 *	3) read_len not reflected from the wr.
512 	 *	4) cq_type is RQ_TYPE not SQ_TYPE.
513 	 */
514 	if (RQ_TYPE(hw_cqe) && (CQE_OPCODE(hw_cqe) == FW_RI_READ_RESP)) {
515 
516 		/* If we have reached here because of async
517 		 * event or other error, and have egress error
518 		 * then drop
519 		 */
520 		if (CQE_TYPE(hw_cqe) == 1) {
521 			if (CQE_STATUS(hw_cqe))
522 				t4_set_wq_in_error(wq);
523 			ret = -EAGAIN;
524 			goto skip_cqe;
525 		}
526 
527 		/* If this is an unsolicited read response, then the read
528 		 * was generated by the kernel driver as part of peer-2-peer
529 		 * connection setup.  So ignore the completion.
530 		 */
531 		if (CQE_WRID_STAG(hw_cqe) == 1) {
532 			if (CQE_STATUS(hw_cqe))
533 				t4_set_wq_in_error(wq);
534 			ret = -EAGAIN;
535 			goto skip_cqe;
536 		}
537 
538 		/*
539 		 * Eat completions for unsignaled read WRs.
540 		 */
541 		if (!wq->sq.oldest_read->signaled) {
542 			advance_oldest_read(wq);
543 			ret = -EAGAIN;
544 			goto skip_cqe;
545 		}
546 
547 		/*
548 		 * Don't write to the HWCQ, so create a new read req CQE
549 		 * in local memory.
550 		 */
551 		create_read_req_cqe(wq, hw_cqe, &read_cqe);
552 		hw_cqe = &read_cqe;
553 		advance_oldest_read(wq);
554 	}
555 
556 	if (CQE_STATUS(hw_cqe) || t4_wq_in_error(wq)) {
557 		*cqe_flushed = (CQE_STATUS(hw_cqe) == T4_ERR_SWFLUSH);
558 		t4_set_wq_in_error(wq);
559 	}
560 
561 	/*
562 	 * RECV completion.
563 	 */
564 	if (RQ_TYPE(hw_cqe)) {
565 
566 		/*
567 		 * HW only validates 4 bits of MSN.  So we must validate that
568 		 * the MSN in the SEND is the next expected MSN.  If its not,
569 		 * then we complete this with T4_ERR_MSN and mark the wq in
570 		 * error.
571 		 */
572 
573 		if (t4_rq_empty(wq)) {
574 			t4_set_wq_in_error(wq);
575 			ret = -EAGAIN;
576 			goto skip_cqe;
577 		}
578 		if (unlikely((CQE_WRID_MSN(hw_cqe) != (wq->rq.msn)))) {
579 			t4_set_wq_in_error(wq);
580 			hw_cqe->header |= htonl(CQE_STATUS_V(T4_ERR_MSN));
581 			goto proc_cqe;
582 		}
583 		goto proc_cqe;
584 	}
585 
586 	/*
587 	 * If we get here its a send completion.
588 	 *
589 	 * Handle out of order completion. These get stuffed
590 	 * in the SW SQ. Then the SW SQ is walked to move any
591 	 * now in-order completions into the SW CQ.  This handles
592 	 * 2 cases:
593 	 *	1) reaping unsignaled WRs when the first subsequent
594 	 *	   signaled WR is completed.
595 	 *	2) out of order read completions.
596 	 */
597 	if (!SW_CQE(hw_cqe) && (CQE_WRID_SQ_IDX(hw_cqe) != wq->sq.cidx)) {
598 		struct t4_swsqe *swsqe;
599 
600 		PDBG("%s out of order completion going in sw_sq at idx %u\n",
601 		     __func__, CQE_WRID_SQ_IDX(hw_cqe));
602 		swsqe = &wq->sq.sw_sq[CQE_WRID_SQ_IDX(hw_cqe)];
603 		swsqe->cqe = *hw_cqe;
604 		swsqe->complete = 1;
605 		ret = -EAGAIN;
606 		goto flush_wq;
607 	}
608 
609 proc_cqe:
610 	*cqe = *hw_cqe;
611 
612 	/*
613 	 * Reap the associated WR(s) that are freed up with this
614 	 * completion.
615 	 */
616 	if (SQ_TYPE(hw_cqe)) {
617 		int idx = CQE_WRID_SQ_IDX(hw_cqe);
618 		BUG_ON(idx >= wq->sq.size);
619 
620 		/*
621 		* Account for any unsignaled completions completed by
622 		* this signaled completion.  In this case, cidx points
623 		* to the first unsignaled one, and idx points to the
624 		* signaled one.  So adjust in_use based on this delta.
625 		* if this is not completing any unsigned wrs, then the
626 		* delta will be 0. Handle wrapping also!
627 		*/
628 		if (idx < wq->sq.cidx)
629 			wq->sq.in_use -= wq->sq.size + idx - wq->sq.cidx;
630 		else
631 			wq->sq.in_use -= idx - wq->sq.cidx;
632 		BUG_ON(wq->sq.in_use <= 0 && wq->sq.in_use >= wq->sq.size);
633 
634 		wq->sq.cidx = (uint16_t)idx;
635 		PDBG("%s completing sq idx %u\n", __func__, wq->sq.cidx);
636 		*cookie = wq->sq.sw_sq[wq->sq.cidx].wr_id;
637 		if (c4iw_wr_log)
638 			c4iw_log_wr_stats(wq, hw_cqe);
639 		t4_sq_consume(wq);
640 	} else {
641 		PDBG("%s completing rq idx %u\n", __func__, wq->rq.cidx);
642 		*cookie = wq->rq.sw_rq[wq->rq.cidx].wr_id;
643 		BUG_ON(t4_rq_empty(wq));
644 		if (c4iw_wr_log)
645 			c4iw_log_wr_stats(wq, hw_cqe);
646 		t4_rq_consume(wq);
647 		goto skip_cqe;
648 	}
649 
650 flush_wq:
651 	/*
652 	 * Flush any completed cqes that are now in-order.
653 	 */
654 	flush_completed_wrs(wq, cq);
655 
656 skip_cqe:
657 	if (SW_CQE(hw_cqe)) {
658 		PDBG("%s cq %p cqid 0x%x skip sw cqe cidx %u\n",
659 		     __func__, cq, cq->cqid, cq->sw_cidx);
660 		t4_swcq_consume(cq);
661 	} else {
662 		PDBG("%s cq %p cqid 0x%x skip hw cqe cidx %u\n",
663 		     __func__, cq, cq->cqid, cq->cidx);
664 		t4_hwcq_consume(cq);
665 	}
666 	return ret;
667 }
668 
669 /*
670  * Get one cq entry from c4iw and map it to openib.
671  *
672  * Returns:
673  *	0			cqe returned
674  *	-ENODATA		EMPTY;
675  *	-EAGAIN			caller must try again
676  *	any other -errno	fatal error
677  */
678 static int c4iw_poll_cq_one(struct c4iw_cq *chp, struct ib_wc *wc)
679 {
680 	struct c4iw_qp *qhp = NULL;
681 	struct t4_cqe uninitialized_var(cqe), *rd_cqe;
682 	struct t4_wq *wq;
683 	u32 credit = 0;
684 	u8 cqe_flushed;
685 	u64 cookie = 0;
686 	int ret;
687 
688 	ret = t4_next_cqe(&chp->cq, &rd_cqe);
689 
690 	if (ret)
691 		return ret;
692 
693 	qhp = get_qhp(chp->rhp, CQE_QPID(rd_cqe));
694 	if (!qhp)
695 		wq = NULL;
696 	else {
697 		spin_lock(&qhp->lock);
698 		wq = &(qhp->wq);
699 	}
700 	ret = poll_cq(wq, &(chp->cq), &cqe, &cqe_flushed, &cookie, &credit);
701 	if (ret)
702 		goto out;
703 
704 	wc->wr_id = cookie;
705 	wc->qp = &qhp->ibqp;
706 	wc->vendor_err = CQE_STATUS(&cqe);
707 	wc->wc_flags = 0;
708 
709 	PDBG("%s qpid 0x%x type %d opcode %d status 0x%x len %u wrid hi 0x%x "
710 	     "lo 0x%x cookie 0x%llx\n", __func__, CQE_QPID(&cqe),
711 	     CQE_TYPE(&cqe), CQE_OPCODE(&cqe), CQE_STATUS(&cqe), CQE_LEN(&cqe),
712 	     CQE_WRID_HI(&cqe), CQE_WRID_LOW(&cqe), (unsigned long long)cookie);
713 
714 	if (CQE_TYPE(&cqe) == 0) {
715 		if (!CQE_STATUS(&cqe))
716 			wc->byte_len = CQE_LEN(&cqe);
717 		else
718 			wc->byte_len = 0;
719 		wc->opcode = IB_WC_RECV;
720 		if (CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_INV ||
721 		    CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_SE_INV) {
722 			wc->ex.invalidate_rkey = CQE_WRID_STAG(&cqe);
723 			wc->wc_flags |= IB_WC_WITH_INVALIDATE;
724 		}
725 	} else {
726 		switch (CQE_OPCODE(&cqe)) {
727 		case FW_RI_RDMA_WRITE:
728 			wc->opcode = IB_WC_RDMA_WRITE;
729 			break;
730 		case FW_RI_READ_REQ:
731 			wc->opcode = IB_WC_RDMA_READ;
732 			wc->byte_len = CQE_LEN(&cqe);
733 			break;
734 		case FW_RI_SEND_WITH_INV:
735 		case FW_RI_SEND_WITH_SE_INV:
736 			wc->opcode = IB_WC_SEND;
737 			wc->wc_flags |= IB_WC_WITH_INVALIDATE;
738 			break;
739 		case FW_RI_SEND:
740 		case FW_RI_SEND_WITH_SE:
741 			wc->opcode = IB_WC_SEND;
742 			break;
743 
744 		case FW_RI_LOCAL_INV:
745 			wc->opcode = IB_WC_LOCAL_INV;
746 			break;
747 		case FW_RI_FAST_REGISTER:
748 			wc->opcode = IB_WC_REG_MR;
749 			break;
750 		default:
751 			printk(KERN_ERR MOD "Unexpected opcode %d "
752 			       "in the CQE received for QPID=0x%0x\n",
753 			       CQE_OPCODE(&cqe), CQE_QPID(&cqe));
754 			ret = -EINVAL;
755 			goto out;
756 		}
757 	}
758 
759 	if (cqe_flushed)
760 		wc->status = IB_WC_WR_FLUSH_ERR;
761 	else {
762 
763 		switch (CQE_STATUS(&cqe)) {
764 		case T4_ERR_SUCCESS:
765 			wc->status = IB_WC_SUCCESS;
766 			break;
767 		case T4_ERR_STAG:
768 			wc->status = IB_WC_LOC_ACCESS_ERR;
769 			break;
770 		case T4_ERR_PDID:
771 			wc->status = IB_WC_LOC_PROT_ERR;
772 			break;
773 		case T4_ERR_QPID:
774 		case T4_ERR_ACCESS:
775 			wc->status = IB_WC_LOC_ACCESS_ERR;
776 			break;
777 		case T4_ERR_WRAP:
778 			wc->status = IB_WC_GENERAL_ERR;
779 			break;
780 		case T4_ERR_BOUND:
781 			wc->status = IB_WC_LOC_LEN_ERR;
782 			break;
783 		case T4_ERR_INVALIDATE_SHARED_MR:
784 		case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
785 			wc->status = IB_WC_MW_BIND_ERR;
786 			break;
787 		case T4_ERR_CRC:
788 		case T4_ERR_MARKER:
789 		case T4_ERR_PDU_LEN_ERR:
790 		case T4_ERR_OUT_OF_RQE:
791 		case T4_ERR_DDP_VERSION:
792 		case T4_ERR_RDMA_VERSION:
793 		case T4_ERR_DDP_QUEUE_NUM:
794 		case T4_ERR_MSN:
795 		case T4_ERR_TBIT:
796 		case T4_ERR_MO:
797 		case T4_ERR_MSN_RANGE:
798 		case T4_ERR_IRD_OVERFLOW:
799 		case T4_ERR_OPCODE:
800 		case T4_ERR_INTERNAL_ERR:
801 			wc->status = IB_WC_FATAL_ERR;
802 			break;
803 		case T4_ERR_SWFLUSH:
804 			wc->status = IB_WC_WR_FLUSH_ERR;
805 			break;
806 		default:
807 			printk(KERN_ERR MOD
808 			       "Unexpected cqe_status 0x%x for QPID=0x%0x\n",
809 			       CQE_STATUS(&cqe), CQE_QPID(&cqe));
810 			wc->status = IB_WC_FATAL_ERR;
811 		}
812 	}
813 out:
814 	if (wq) {
815 		if (unlikely(qhp->attr.state != C4IW_QP_STATE_RTS)) {
816 			if (t4_sq_empty(wq))
817 				complete(&qhp->sq_drained);
818 			if (t4_rq_empty(wq))
819 				complete(&qhp->rq_drained);
820 		}
821 		spin_unlock(&qhp->lock);
822 	}
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 		   chp->destroy_skb);
864 	chp->destroy_skb = NULL;
865 	kfree(chp);
866 	return 0;
867 }
868 
869 struct ib_cq *c4iw_create_cq(struct ib_device *ibdev,
870 			     const struct ib_cq_init_attr *attr,
871 			     struct ib_ucontext *ib_context,
872 			     struct ib_udata *udata)
873 {
874 	int entries = attr->cqe;
875 	int vector = attr->comp_vector;
876 	struct c4iw_dev *rhp;
877 	struct c4iw_cq *chp;
878 	struct c4iw_create_cq_resp uresp;
879 	struct c4iw_ucontext *ucontext = NULL;
880 	int ret, wr_len;
881 	size_t memsize, hwentries;
882 	struct c4iw_mm_entry *mm, *mm2;
883 
884 	PDBG("%s ib_dev %p entries %d\n", __func__, ibdev, entries);
885 	if (attr->flags)
886 		return ERR_PTR(-EINVAL);
887 
888 	rhp = to_c4iw_dev(ibdev);
889 
890 	if (vector >= rhp->rdev.lldi.nciq)
891 		return ERR_PTR(-EINVAL);
892 
893 	chp = kzalloc(sizeof(*chp), GFP_KERNEL);
894 	if (!chp)
895 		return ERR_PTR(-ENOMEM);
896 
897 	wr_len = sizeof(struct fw_ri_res_wr) + sizeof(struct fw_ri_res);
898 	chp->destroy_skb = alloc_skb(wr_len, GFP_KERNEL);
899 	if (!chp->destroy_skb) {
900 		ret = -ENOMEM;
901 		goto err1;
902 	}
903 
904 	if (ib_context)
905 		ucontext = to_c4iw_ucontext(ib_context);
906 
907 	/* account for the status page. */
908 	entries++;
909 
910 	/* IQ needs one extra entry to differentiate full vs empty. */
911 	entries++;
912 
913 	/*
914 	 * entries must be multiple of 16 for HW.
915 	 */
916 	entries = roundup(entries, 16);
917 
918 	/*
919 	 * Make actual HW queue 2x to avoid cdix_inc overflows.
920 	 */
921 	hwentries = min(entries * 2, rhp->rdev.hw_queue.t4_max_iq_size);
922 
923 	/*
924 	 * Make HW queue at least 64 entries so GTS updates aren't too
925 	 * frequent.
926 	 */
927 	if (hwentries < 64)
928 		hwentries = 64;
929 
930 	memsize = hwentries * sizeof *chp->cq.queue;
931 
932 	/*
933 	 * memsize must be a multiple of the page size if its a user cq.
934 	 */
935 	if (ucontext)
936 		memsize = roundup(memsize, PAGE_SIZE);
937 	chp->cq.size = hwentries;
938 	chp->cq.memsize = memsize;
939 	chp->cq.vector = vector;
940 
941 	ret = create_cq(&rhp->rdev, &chp->cq,
942 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
943 	if (ret)
944 		goto err2;
945 
946 	chp->rhp = rhp;
947 	chp->cq.size--;				/* status page */
948 	chp->ibcq.cqe = entries - 2;
949 	spin_lock_init(&chp->lock);
950 	spin_lock_init(&chp->comp_handler_lock);
951 	atomic_set(&chp->refcnt, 1);
952 	init_waitqueue_head(&chp->wait);
953 	ret = insert_handle(rhp, &rhp->cqidr, chp, chp->cq.cqid);
954 	if (ret)
955 		goto err3;
956 
957 	if (ucontext) {
958 		mm = kmalloc(sizeof *mm, GFP_KERNEL);
959 		if (!mm)
960 			goto err4;
961 		mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
962 		if (!mm2)
963 			goto err5;
964 
965 		uresp.qid_mask = rhp->rdev.cqmask;
966 		uresp.cqid = chp->cq.cqid;
967 		uresp.size = chp->cq.size;
968 		uresp.memsize = chp->cq.memsize;
969 		spin_lock(&ucontext->mmap_lock);
970 		uresp.key = ucontext->key;
971 		ucontext->key += PAGE_SIZE;
972 		uresp.gts_key = ucontext->key;
973 		ucontext->key += PAGE_SIZE;
974 		spin_unlock(&ucontext->mmap_lock);
975 		ret = ib_copy_to_udata(udata, &uresp,
976 				       sizeof(uresp) - sizeof(uresp.reserved));
977 		if (ret)
978 			goto err6;
979 
980 		mm->key = uresp.key;
981 		mm->addr = virt_to_phys(chp->cq.queue);
982 		mm->len = chp->cq.memsize;
983 		insert_mmap(ucontext, mm);
984 
985 		mm2->key = uresp.gts_key;
986 		mm2->addr = chp->cq.bar2_pa;
987 		mm2->len = PAGE_SIZE;
988 		insert_mmap(ucontext, mm2);
989 	}
990 	PDBG("%s cqid 0x%0x chp %p size %u memsize %zu, dma_addr 0x%0llx\n",
991 	     __func__, chp->cq.cqid, chp, chp->cq.size,
992 	     chp->cq.memsize, (unsigned long long) chp->cq.dma_addr);
993 	return &chp->ibcq;
994 err6:
995 	kfree(mm2);
996 err5:
997 	kfree(mm);
998 err4:
999 	remove_handle(rhp, &rhp->cqidr, chp->cq.cqid);
1000 err3:
1001 	destroy_cq(&chp->rhp->rdev, &chp->cq,
1002 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx,
1003 		   chp->destroy_skb);
1004 err2:
1005 	kfree_skb(chp->destroy_skb);
1006 err1:
1007 	kfree(chp);
1008 	return ERR_PTR(ret);
1009 }
1010 
1011 int c4iw_resize_cq(struct ib_cq *cq, int cqe, struct ib_udata *udata)
1012 {
1013 	return -ENOSYS;
1014 }
1015 
1016 int c4iw_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
1017 {
1018 	struct c4iw_cq *chp;
1019 	int ret = 0;
1020 	unsigned long flag;
1021 
1022 	chp = to_c4iw_cq(ibcq);
1023 	spin_lock_irqsave(&chp->lock, flag);
1024 	t4_arm_cq(&chp->cq,
1025 		  (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED);
1026 	if (flags & IB_CQ_REPORT_MISSED_EVENTS)
1027 		ret = t4_cq_notempty(&chp->cq);
1028 	spin_unlock_irqrestore(&chp->lock, flag);
1029 	return ret;
1030 }
1031