xref: /openbmc/linux/drivers/infiniband/hw/cxgb4/cq.c (revision f7777dcc)
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(FW_RI_RES_WR) |
55 			V_FW_RI_RES_WR_NRES(1) |
56 			FW_WR_COMPL(1));
57 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
58 	res_wr->cookie = (unsigned long) &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(FW_RI_RES_WR) |
125 			V_FW_RI_RES_WR_NRES(1) |
126 			FW_WR_COMPL(1));
127 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
128 	res_wr->cookie = (unsigned long) &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 			V_FW_RI_RES_WR_IQANUS(0) |
135 			V_FW_RI_RES_WR_IQANUD(1) |
136 			F_FW_RI_RES_WR_IQANDST |
137 			V_FW_RI_RES_WR_IQANDSTINDEX(*rdev->lldi.rxq_ids));
138 	res->u.cq.iqdroprss_to_iqesize = cpu_to_be16(
139 			F_FW_RI_RES_WR_IQDROPRSS |
140 			V_FW_RI_RES_WR_IQPCIECH(2) |
141 			V_FW_RI_RES_WR_IQINTCNTTHRESH(0) |
142 			F_FW_RI_RES_WR_IQO |
143 			V_FW_RI_RES_WR_IQESIZE(1));
144 	res->u.cq.iqsize = cpu_to_be16(cq->size);
145 	res->u.cq.iqaddr = cpu_to_be64(cq->dma_addr);
146 
147 	c4iw_init_wr_wait(&wr_wait);
148 
149 	ret = c4iw_ofld_send(rdev, skb);
150 	if (ret)
151 		goto err4;
152 	PDBG("%s wait_event wr_wait %p\n", __func__, &wr_wait);
153 	ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, __func__);
154 	if (ret)
155 		goto err4;
156 
157 	cq->gen = 1;
158 	cq->gts = rdev->lldi.gts_reg;
159 	cq->rdev = rdev;
160 	if (user) {
161 		cq->ugts = (u64)pci_resource_start(rdev->lldi.pdev, 2) +
162 					(cq->cqid << rdev->cqshift);
163 		cq->ugts &= PAGE_MASK;
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 	PDBG("%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(V_CQE_STATUS(T4_ERR_SWFLUSH) |
185 				 V_CQE_OPCODE(FW_RI_SEND) |
186 				 V_CQE_TYPE(0) |
187 				 V_CQE_SWCQE(1) |
188 				 V_CQE_QPID(wq->sq.qid));
189 	cqe.bits_type_ts = cpu_to_be64(V_CQE_GENBIT((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 	PDBG("%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 	PDBG("%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(V_CQE_STATUS(T4_ERR_SWFLUSH) |
218 				 V_CQE_OPCODE(swcqe->opcode) |
219 				 V_CQE_TYPE(1) |
220 				 V_CQE_SWCQE(1) |
221 				 V_CQE_QPID(wq->sq.qid));
222 	CQE_WRID_SQ_IDX(&cqe) = swcqe->idx;
223 	cqe.bits_type_ts = cpu_to_be64(V_CQE_GENBIT((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 	int error = (qhp->attr.state != C4IW_QP_STATE_CLOSING &&
239 			qhp->attr.state != C4IW_QP_STATE_IDLE);
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 		if (error) {
247 			swsqe = &wq->sq.sw_sq[idx];
248 			BUG_ON(swsqe->flushed);
249 			swsqe->flushed = 1;
250 			insert_sq_cqe(wq, cq, swsqe);
251 			if (wq->sq.oldest_read == swsqe) {
252 				BUG_ON(swsqe->opcode != FW_RI_READ_REQ);
253 				advance_oldest_read(wq);
254 			}
255 			flushed++;
256 		} else {
257 			t4_sq_consume(wq);
258 		}
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(V_CQE_SWCQE(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(V_CQE_QPID(CQE_QPID(hw_cqe)) |
310 			V_CQE_SWCQE(SW_CQE(hw_cqe)) |
311 			V_CQE_OPCODE(FW_RI_READ_REQ) |
312 			V_CQE_TYPE(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 			/*
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(V_CQE_SWCQE(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 	PDBG("%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 	PDBG("%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 	PDBG("%s CQE OVF %u qpid 0x%0x genbit %u type %u status 0x%0x"
475 	     " opcode 0x%0x len 0x%0x wrid_hi_stag 0x%x wrid_low_msn 0x%x\n",
476 	     __func__, CQE_OVFBIT(hw_cqe), CQE_QPID(hw_cqe),
477 	     CQE_GENBIT(hw_cqe), CQE_TYPE(hw_cqe), CQE_STATUS(hw_cqe),
478 	     CQE_OPCODE(hw_cqe), CQE_LEN(hw_cqe), CQE_WRID_HI(hw_cqe),
479 	     CQE_WRID_LOW(hw_cqe));
480 
481 	/*
482 	 * skip cqe's not affiliated with a QP.
483 	 */
484 	if (wq == NULL) {
485 		ret = -EAGAIN;
486 		goto skip_cqe;
487 	}
488 
489 	/*
490 	* skip hw cqe's if the wq is flushed.
491 	*/
492 	if (wq->flushed && !SW_CQE(hw_cqe)) {
493 		ret = -EAGAIN;
494 		goto skip_cqe;
495 	}
496 
497 	/*
498 	 * skip TERMINATE cqes...
499 	 */
500 	if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE) {
501 		ret = -EAGAIN;
502 		goto skip_cqe;
503 	}
504 
505 	/*
506 	 * Gotta tweak READ completions:
507 	 *	1) the cqe doesn't contain the sq_wptr from the wr.
508 	 *	2) opcode not reflected from the wr.
509 	 *	3) read_len not reflected from the wr.
510 	 *	4) cq_type is RQ_TYPE not SQ_TYPE.
511 	 */
512 	if (RQ_TYPE(hw_cqe) && (CQE_OPCODE(hw_cqe) == FW_RI_READ_RESP)) {
513 
514 		/*
515 		 * If this is an unsolicited read response, then the read
516 		 * was generated by the kernel driver as part of peer-2-peer
517 		 * connection setup.  So ignore the completion.
518 		 */
519 		if (CQE_WRID_STAG(hw_cqe) == 1) {
520 			if (CQE_STATUS(hw_cqe))
521 				t4_set_wq_in_error(wq);
522 			ret = -EAGAIN;
523 			goto skip_cqe;
524 		}
525 
526 		/*
527 		 * Eat completions for unsignaled read WRs.
528 		 */
529 		if (!wq->sq.oldest_read->signaled) {
530 			advance_oldest_read(wq);
531 			ret = -EAGAIN;
532 			goto skip_cqe;
533 		}
534 
535 		/*
536 		 * Don't write to the HWCQ, so create a new read req CQE
537 		 * in local memory.
538 		 */
539 		create_read_req_cqe(wq, hw_cqe, &read_cqe);
540 		hw_cqe = &read_cqe;
541 		advance_oldest_read(wq);
542 	}
543 
544 	if (CQE_STATUS(hw_cqe) || t4_wq_in_error(wq)) {
545 		*cqe_flushed = (CQE_STATUS(hw_cqe) == T4_ERR_SWFLUSH);
546 		t4_set_wq_in_error(wq);
547 	}
548 
549 	/*
550 	 * RECV completion.
551 	 */
552 	if (RQ_TYPE(hw_cqe)) {
553 
554 		/*
555 		 * HW only validates 4 bits of MSN.  So we must validate that
556 		 * the MSN in the SEND is the next expected MSN.  If its not,
557 		 * then we complete this with T4_ERR_MSN and mark the wq in
558 		 * error.
559 		 */
560 
561 		if (t4_rq_empty(wq)) {
562 			t4_set_wq_in_error(wq);
563 			ret = -EAGAIN;
564 			goto skip_cqe;
565 		}
566 		if (unlikely((CQE_WRID_MSN(hw_cqe) != (wq->rq.msn)))) {
567 			t4_set_wq_in_error(wq);
568 			hw_cqe->header |= htonl(V_CQE_STATUS(T4_ERR_MSN));
569 			goto proc_cqe;
570 		}
571 		goto proc_cqe;
572 	}
573 
574 	/*
575 	 * If we get here its a send completion.
576 	 *
577 	 * Handle out of order completion. These get stuffed
578 	 * in the SW SQ. Then the SW SQ is walked to move any
579 	 * now in-order completions into the SW CQ.  This handles
580 	 * 2 cases:
581 	 *	1) reaping unsignaled WRs when the first subsequent
582 	 *	   signaled WR is completed.
583 	 *	2) out of order read completions.
584 	 */
585 	if (!SW_CQE(hw_cqe) && (CQE_WRID_SQ_IDX(hw_cqe) != wq->sq.cidx)) {
586 		struct t4_swsqe *swsqe;
587 
588 		PDBG("%s out of order completion going in sw_sq at idx %u\n",
589 		     __func__, CQE_WRID_SQ_IDX(hw_cqe));
590 		swsqe = &wq->sq.sw_sq[CQE_WRID_SQ_IDX(hw_cqe)];
591 		swsqe->cqe = *hw_cqe;
592 		swsqe->complete = 1;
593 		ret = -EAGAIN;
594 		goto flush_wq;
595 	}
596 
597 proc_cqe:
598 	*cqe = *hw_cqe;
599 
600 	/*
601 	 * Reap the associated WR(s) that are freed up with this
602 	 * completion.
603 	 */
604 	if (SQ_TYPE(hw_cqe)) {
605 		int idx = CQE_WRID_SQ_IDX(hw_cqe);
606 		BUG_ON(idx > wq->sq.size);
607 
608 		/*
609 		* Account for any unsignaled completions completed by
610 		* this signaled completion.  In this case, cidx points
611 		* to the first unsignaled one, and idx points to the
612 		* signaled one.  So adjust in_use based on this delta.
613 		* if this is not completing any unsigned wrs, then the
614 		* delta will be 0. Handle wrapping also!
615 		*/
616 		if (idx < wq->sq.cidx)
617 			wq->sq.in_use -= wq->sq.size + idx - wq->sq.cidx;
618 		else
619 			wq->sq.in_use -= idx - wq->sq.cidx;
620 		BUG_ON(wq->sq.in_use < 0 && wq->sq.in_use < wq->sq.size);
621 
622 		wq->sq.cidx = (uint16_t)idx;
623 		PDBG("%s completing sq idx %u\n", __func__, wq->sq.cidx);
624 		*cookie = wq->sq.sw_sq[wq->sq.cidx].wr_id;
625 		t4_sq_consume(wq);
626 	} else {
627 		PDBG("%s completing rq idx %u\n", __func__, wq->rq.cidx);
628 		*cookie = wq->rq.sw_rq[wq->rq.cidx].wr_id;
629 		BUG_ON(t4_rq_empty(wq));
630 		t4_rq_consume(wq);
631 		goto skip_cqe;
632 	}
633 
634 flush_wq:
635 	/*
636 	 * Flush any completed cqes that are now in-order.
637 	 */
638 	flush_completed_wrs(wq, cq);
639 
640 skip_cqe:
641 	if (SW_CQE(hw_cqe)) {
642 		PDBG("%s cq %p cqid 0x%x skip sw cqe cidx %u\n",
643 		     __func__, cq, cq->cqid, cq->sw_cidx);
644 		t4_swcq_consume(cq);
645 	} else {
646 		PDBG("%s cq %p cqid 0x%x skip hw cqe cidx %u\n",
647 		     __func__, cq, cq->cqid, cq->cidx);
648 		t4_hwcq_consume(cq);
649 	}
650 	return ret;
651 }
652 
653 /*
654  * Get one cq entry from c4iw and map it to openib.
655  *
656  * Returns:
657  *	0			cqe returned
658  *	-ENODATA		EMPTY;
659  *	-EAGAIN			caller must try again
660  *	any other -errno	fatal error
661  */
662 static int c4iw_poll_cq_one(struct c4iw_cq *chp, struct ib_wc *wc)
663 {
664 	struct c4iw_qp *qhp = NULL;
665 	struct t4_cqe cqe = {0, 0}, *rd_cqe;
666 	struct t4_wq *wq;
667 	u32 credit = 0;
668 	u8 cqe_flushed;
669 	u64 cookie = 0;
670 	int ret;
671 
672 	ret = t4_next_cqe(&chp->cq, &rd_cqe);
673 
674 	if (ret)
675 		return ret;
676 
677 	qhp = get_qhp(chp->rhp, CQE_QPID(rd_cqe));
678 	if (!qhp)
679 		wq = NULL;
680 	else {
681 		spin_lock(&qhp->lock);
682 		wq = &(qhp->wq);
683 	}
684 	ret = poll_cq(wq, &(chp->cq), &cqe, &cqe_flushed, &cookie, &credit);
685 	if (ret)
686 		goto out;
687 
688 	wc->wr_id = cookie;
689 	wc->qp = &qhp->ibqp;
690 	wc->vendor_err = CQE_STATUS(&cqe);
691 	wc->wc_flags = 0;
692 
693 	PDBG("%s qpid 0x%x type %d opcode %d status 0x%x len %u wrid hi 0x%x "
694 	     "lo 0x%x cookie 0x%llx\n", __func__, CQE_QPID(&cqe),
695 	     CQE_TYPE(&cqe), CQE_OPCODE(&cqe), CQE_STATUS(&cqe), CQE_LEN(&cqe),
696 	     CQE_WRID_HI(&cqe), CQE_WRID_LOW(&cqe), (unsigned long long)cookie);
697 
698 	if (CQE_TYPE(&cqe) == 0) {
699 		if (!CQE_STATUS(&cqe))
700 			wc->byte_len = CQE_LEN(&cqe);
701 		else
702 			wc->byte_len = 0;
703 		wc->opcode = IB_WC_RECV;
704 		if (CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_INV ||
705 		    CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_SE_INV) {
706 			wc->ex.invalidate_rkey = CQE_WRID_STAG(&cqe);
707 			wc->wc_flags |= IB_WC_WITH_INVALIDATE;
708 		}
709 	} else {
710 		switch (CQE_OPCODE(&cqe)) {
711 		case FW_RI_RDMA_WRITE:
712 			wc->opcode = IB_WC_RDMA_WRITE;
713 			break;
714 		case FW_RI_READ_REQ:
715 			wc->opcode = IB_WC_RDMA_READ;
716 			wc->byte_len = CQE_LEN(&cqe);
717 			break;
718 		case FW_RI_SEND_WITH_INV:
719 		case FW_RI_SEND_WITH_SE_INV:
720 			wc->opcode = IB_WC_SEND;
721 			wc->wc_flags |= IB_WC_WITH_INVALIDATE;
722 			break;
723 		case FW_RI_SEND:
724 		case FW_RI_SEND_WITH_SE:
725 			wc->opcode = IB_WC_SEND;
726 			break;
727 		case FW_RI_BIND_MW:
728 			wc->opcode = IB_WC_BIND_MW;
729 			break;
730 
731 		case FW_RI_LOCAL_INV:
732 			wc->opcode = IB_WC_LOCAL_INV;
733 			break;
734 		case FW_RI_FAST_REGISTER:
735 			wc->opcode = IB_WC_FAST_REG_MR;
736 			break;
737 		default:
738 			printk(KERN_ERR MOD "Unexpected opcode %d "
739 			       "in the CQE received for QPID=0x%0x\n",
740 			       CQE_OPCODE(&cqe), CQE_QPID(&cqe));
741 			ret = -EINVAL;
742 			goto out;
743 		}
744 	}
745 
746 	if (cqe_flushed)
747 		wc->status = IB_WC_WR_FLUSH_ERR;
748 	else {
749 
750 		switch (CQE_STATUS(&cqe)) {
751 		case T4_ERR_SUCCESS:
752 			wc->status = IB_WC_SUCCESS;
753 			break;
754 		case T4_ERR_STAG:
755 			wc->status = IB_WC_LOC_ACCESS_ERR;
756 			break;
757 		case T4_ERR_PDID:
758 			wc->status = IB_WC_LOC_PROT_ERR;
759 			break;
760 		case T4_ERR_QPID:
761 		case T4_ERR_ACCESS:
762 			wc->status = IB_WC_LOC_ACCESS_ERR;
763 			break;
764 		case T4_ERR_WRAP:
765 			wc->status = IB_WC_GENERAL_ERR;
766 			break;
767 		case T4_ERR_BOUND:
768 			wc->status = IB_WC_LOC_LEN_ERR;
769 			break;
770 		case T4_ERR_INVALIDATE_SHARED_MR:
771 		case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
772 			wc->status = IB_WC_MW_BIND_ERR;
773 			break;
774 		case T4_ERR_CRC:
775 		case T4_ERR_MARKER:
776 		case T4_ERR_PDU_LEN_ERR:
777 		case T4_ERR_OUT_OF_RQE:
778 		case T4_ERR_DDP_VERSION:
779 		case T4_ERR_RDMA_VERSION:
780 		case T4_ERR_DDP_QUEUE_NUM:
781 		case T4_ERR_MSN:
782 		case T4_ERR_TBIT:
783 		case T4_ERR_MO:
784 		case T4_ERR_MSN_RANGE:
785 		case T4_ERR_IRD_OVERFLOW:
786 		case T4_ERR_OPCODE:
787 		case T4_ERR_INTERNAL_ERR:
788 			wc->status = IB_WC_FATAL_ERR;
789 			break;
790 		case T4_ERR_SWFLUSH:
791 			wc->status = IB_WC_WR_FLUSH_ERR;
792 			break;
793 		default:
794 			printk(KERN_ERR MOD
795 			       "Unexpected cqe_status 0x%x for QPID=0x%0x\n",
796 			       CQE_STATUS(&cqe), CQE_QPID(&cqe));
797 			ret = -EINVAL;
798 		}
799 	}
800 out:
801 	if (wq)
802 		spin_unlock(&qhp->lock);
803 	return ret;
804 }
805 
806 int c4iw_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
807 {
808 	struct c4iw_cq *chp;
809 	unsigned long flags;
810 	int npolled;
811 	int err = 0;
812 
813 	chp = to_c4iw_cq(ibcq);
814 
815 	spin_lock_irqsave(&chp->lock, flags);
816 	for (npolled = 0; npolled < num_entries; ++npolled) {
817 		do {
818 			err = c4iw_poll_cq_one(chp, wc + npolled);
819 		} while (err == -EAGAIN);
820 		if (err)
821 			break;
822 	}
823 	spin_unlock_irqrestore(&chp->lock, flags);
824 	return !err || err == -ENODATA ? npolled : err;
825 }
826 
827 int c4iw_destroy_cq(struct ib_cq *ib_cq)
828 {
829 	struct c4iw_cq *chp;
830 	struct c4iw_ucontext *ucontext;
831 
832 	PDBG("%s ib_cq %p\n", __func__, ib_cq);
833 	chp = to_c4iw_cq(ib_cq);
834 
835 	remove_handle(chp->rhp, &chp->rhp->cqidr, chp->cq.cqid);
836 	atomic_dec(&chp->refcnt);
837 	wait_event(chp->wait, !atomic_read(&chp->refcnt));
838 
839 	ucontext = ib_cq->uobject ? to_c4iw_ucontext(ib_cq->uobject->context)
840 				  : NULL;
841 	destroy_cq(&chp->rhp->rdev, &chp->cq,
842 		   ucontext ? &ucontext->uctx : &chp->cq.rdev->uctx);
843 	kfree(chp);
844 	return 0;
845 }
846 
847 struct ib_cq *c4iw_create_cq(struct ib_device *ibdev, int entries,
848 			     int vector, struct ib_ucontext *ib_context,
849 			     struct ib_udata *udata)
850 {
851 	struct c4iw_dev *rhp;
852 	struct c4iw_cq *chp;
853 	struct c4iw_create_cq_resp uresp;
854 	struct c4iw_ucontext *ucontext = NULL;
855 	int ret;
856 	size_t memsize, hwentries;
857 	struct c4iw_mm_entry *mm, *mm2;
858 
859 	PDBG("%s ib_dev %p entries %d\n", __func__, ibdev, entries);
860 
861 	rhp = to_c4iw_dev(ibdev);
862 
863 	chp = kzalloc(sizeof(*chp), GFP_KERNEL);
864 	if (!chp)
865 		return ERR_PTR(-ENOMEM);
866 
867 	if (ib_context)
868 		ucontext = to_c4iw_ucontext(ib_context);
869 
870 	/* account for the status page. */
871 	entries++;
872 
873 	/* IQ needs one extra entry to differentiate full vs empty. */
874 	entries++;
875 
876 	/*
877 	 * entries must be multiple of 16 for HW.
878 	 */
879 	entries = roundup(entries, 16);
880 
881 	/*
882 	 * Make actual HW queue 2x to avoid cdix_inc overflows.
883 	 */
884 	hwentries = entries * 2;
885 
886 	/*
887 	 * Make HW queue at least 64 entries so GTS updates aren't too
888 	 * frequent.
889 	 */
890 	if (hwentries < 64)
891 		hwentries = 64;
892 
893 	memsize = hwentries * sizeof *chp->cq.queue;
894 
895 	/*
896 	 * memsize must be a multiple of the page size if its a user cq.
897 	 */
898 	if (ucontext) {
899 		memsize = roundup(memsize, PAGE_SIZE);
900 		hwentries = memsize / sizeof *chp->cq.queue;
901 		while (hwentries > T4_MAX_IQ_SIZE) {
902 			memsize -= PAGE_SIZE;
903 			hwentries = memsize / sizeof *chp->cq.queue;
904 		}
905 	}
906 	chp->cq.size = hwentries;
907 	chp->cq.memsize = memsize;
908 
909 	ret = create_cq(&rhp->rdev, &chp->cq,
910 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
911 	if (ret)
912 		goto err1;
913 
914 	chp->rhp = rhp;
915 	chp->cq.size--;				/* status page */
916 	chp->ibcq.cqe = entries - 2;
917 	spin_lock_init(&chp->lock);
918 	spin_lock_init(&chp->comp_handler_lock);
919 	atomic_set(&chp->refcnt, 1);
920 	init_waitqueue_head(&chp->wait);
921 	ret = insert_handle(rhp, &rhp->cqidr, chp, chp->cq.cqid);
922 	if (ret)
923 		goto err2;
924 
925 	if (ucontext) {
926 		mm = kmalloc(sizeof *mm, GFP_KERNEL);
927 		if (!mm)
928 			goto err3;
929 		mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
930 		if (!mm2)
931 			goto err4;
932 
933 		uresp.qid_mask = rhp->rdev.cqmask;
934 		uresp.cqid = chp->cq.cqid;
935 		uresp.size = chp->cq.size;
936 		uresp.memsize = chp->cq.memsize;
937 		spin_lock(&ucontext->mmap_lock);
938 		uresp.key = ucontext->key;
939 		ucontext->key += PAGE_SIZE;
940 		uresp.gts_key = ucontext->key;
941 		ucontext->key += PAGE_SIZE;
942 		spin_unlock(&ucontext->mmap_lock);
943 		ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
944 		if (ret)
945 			goto err5;
946 
947 		mm->key = uresp.key;
948 		mm->addr = virt_to_phys(chp->cq.queue);
949 		mm->len = chp->cq.memsize;
950 		insert_mmap(ucontext, mm);
951 
952 		mm2->key = uresp.gts_key;
953 		mm2->addr = chp->cq.ugts;
954 		mm2->len = PAGE_SIZE;
955 		insert_mmap(ucontext, mm2);
956 	}
957 	PDBG("%s cqid 0x%0x chp %p size %u memsize %zu, dma_addr 0x%0llx\n",
958 	     __func__, chp->cq.cqid, chp, chp->cq.size,
959 	     chp->cq.memsize,
960 	     (unsigned long long) chp->cq.dma_addr);
961 	return &chp->ibcq;
962 err5:
963 	kfree(mm2);
964 err4:
965 	kfree(mm);
966 err3:
967 	remove_handle(rhp, &rhp->cqidr, chp->cq.cqid);
968 err2:
969 	destroy_cq(&chp->rhp->rdev, &chp->cq,
970 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
971 err1:
972 	kfree(chp);
973 	return ERR_PTR(ret);
974 }
975 
976 int c4iw_resize_cq(struct ib_cq *cq, int cqe, struct ib_udata *udata)
977 {
978 	return -ENOSYS;
979 }
980 
981 int c4iw_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
982 {
983 	struct c4iw_cq *chp;
984 	int ret;
985 	unsigned long flag;
986 
987 	chp = to_c4iw_cq(ibcq);
988 	spin_lock_irqsave(&chp->lock, flag);
989 	ret = t4_arm_cq(&chp->cq,
990 			(flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED);
991 	spin_unlock_irqrestore(&chp->lock, flag);
992 	if (ret && !(flags & IB_CQ_REPORT_MISSED_EVENTS))
993 		ret = 0;
994 	return ret;
995 }
996