xref: /openbmc/linux/drivers/infiniband/hw/cxgb4/cq.c (revision 99b7e93c)
1 /*
2  * Copyright (c) 2009-2010 Chelsio, Inc. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *	  copyright notice, this list of conditions and the following
16  *	  disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *	  copyright notice, this list of conditions and the following
20  *	  disclaimer in the documentation and/or other materials
21  *	  provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include "iw_cxgb4.h"
34 
35 static int destroy_cq(struct c4iw_rdev *rdev, struct t4_cq *cq,
36 		      struct c4iw_dev_ucontext *uctx)
37 {
38 	struct fw_ri_res_wr *res_wr;
39 	struct fw_ri_res *res;
40 	int wr_len;
41 	struct c4iw_wr_wait wr_wait;
42 	struct sk_buff *skb;
43 	int ret;
44 
45 	wr_len = sizeof *res_wr + sizeof *res;
46 	skb = alloc_skb(wr_len, GFP_KERNEL);
47 	if (!skb)
48 		return -ENOMEM;
49 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
50 
51 	res_wr = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
52 	memset(res_wr, 0, wr_len);
53 	res_wr->op_nres = cpu_to_be32(
54 			FW_WR_OP_V(FW_RI_RES_WR) |
55 			FW_RI_RES_WR_NRES_V(1) |
56 			FW_WR_COMPL_F);
57 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
58 	res_wr->cookie = (uintptr_t)&wr_wait;
59 	res = res_wr->res;
60 	res->u.cq.restype = FW_RI_RES_TYPE_CQ;
61 	res->u.cq.op = FW_RI_RES_OP_RESET;
62 	res->u.cq.iqid = cpu_to_be32(cq->cqid);
63 
64 	c4iw_init_wr_wait(&wr_wait);
65 	ret = c4iw_ofld_send(rdev, skb);
66 	if (!ret) {
67 		ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, __func__);
68 	}
69 
70 	kfree(cq->sw_queue);
71 	dma_free_coherent(&(rdev->lldi.pdev->dev),
72 			  cq->memsize, cq->queue,
73 			  dma_unmap_addr(cq, mapping));
74 	c4iw_put_cqid(rdev, cq->cqid, uctx);
75 	return ret;
76 }
77 
78 static int create_cq(struct c4iw_rdev *rdev, struct t4_cq *cq,
79 		     struct c4iw_dev_ucontext *uctx)
80 {
81 	struct fw_ri_res_wr *res_wr;
82 	struct fw_ri_res *res;
83 	int wr_len;
84 	int user = (uctx != &rdev->uctx);
85 	struct c4iw_wr_wait wr_wait;
86 	int ret;
87 	struct sk_buff *skb;
88 
89 	cq->cqid = c4iw_get_cqid(rdev, uctx);
90 	if (!cq->cqid) {
91 		ret = -ENOMEM;
92 		goto err1;
93 	}
94 
95 	if (!user) {
96 		cq->sw_queue = kzalloc(cq->memsize, GFP_KERNEL);
97 		if (!cq->sw_queue) {
98 			ret = -ENOMEM;
99 			goto err2;
100 		}
101 	}
102 	cq->queue = dma_alloc_coherent(&rdev->lldi.pdev->dev, cq->memsize,
103 				       &cq->dma_addr, GFP_KERNEL);
104 	if (!cq->queue) {
105 		ret = -ENOMEM;
106 		goto err3;
107 	}
108 	dma_unmap_addr_set(cq, mapping, cq->dma_addr);
109 	memset(cq->queue, 0, cq->memsize);
110 
111 	/* build fw_ri_res_wr */
112 	wr_len = sizeof *res_wr + sizeof *res;
113 
114 	skb = alloc_skb(wr_len, GFP_KERNEL);
115 	if (!skb) {
116 		ret = -ENOMEM;
117 		goto err4;
118 	}
119 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
120 
121 	res_wr = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
122 	memset(res_wr, 0, wr_len);
123 	res_wr->op_nres = cpu_to_be32(
124 			FW_WR_OP_V(FW_RI_RES_WR) |
125 			FW_RI_RES_WR_NRES_V(1) |
126 			FW_WR_COMPL_F);
127 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
128 	res_wr->cookie = (uintptr_t)&wr_wait;
129 	res = res_wr->res;
130 	res->u.cq.restype = FW_RI_RES_TYPE_CQ;
131 	res->u.cq.op = FW_RI_RES_OP_WRITE;
132 	res->u.cq.iqid = cpu_to_be32(cq->cqid);
133 	res->u.cq.iqandst_to_iqandstindex = cpu_to_be32(
134 			FW_RI_RES_WR_IQANUS_V(0) |
135 			FW_RI_RES_WR_IQANUD_V(1) |
136 			FW_RI_RES_WR_IQANDST_F |
137 			FW_RI_RES_WR_IQANDSTINDEX_V(
138 				rdev->lldi.ciq_ids[cq->vector]));
139 	res->u.cq.iqdroprss_to_iqesize = cpu_to_be16(
140 			FW_RI_RES_WR_IQDROPRSS_F |
141 			FW_RI_RES_WR_IQPCIECH_V(2) |
142 			FW_RI_RES_WR_IQINTCNTTHRESH_V(0) |
143 			FW_RI_RES_WR_IQO_F |
144 			FW_RI_RES_WR_IQESIZE_V(1));
145 	res->u.cq.iqsize = cpu_to_be16(cq->size);
146 	res->u.cq.iqaddr = cpu_to_be64(cq->dma_addr);
147 
148 	c4iw_init_wr_wait(&wr_wait);
149 
150 	ret = c4iw_ofld_send(rdev, skb);
151 	if (ret)
152 		goto err4;
153 	PDBG("%s wait_event wr_wait %p\n", __func__, &wr_wait);
154 	ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, __func__);
155 	if (ret)
156 		goto err4;
157 
158 	cq->gen = 1;
159 	cq->rdev = rdev;
160 	if (user) {
161 		u32 off = (cq->cqid << rdev->cqshift) & PAGE_MASK;
162 
163 		cq->ugts = (u64)rdev->bar2_pa + off;
164 	} else if (is_t4(rdev->lldi.adapter_type)) {
165 		cq->gts = rdev->lldi.gts_reg;
166 		cq->qid_mask = -1U;
167 	} else {
168 		u32 off = ((cq->cqid << rdev->cqshift) & PAGE_MASK) + 12;
169 
170 		cq->gts = rdev->bar2_kva + off;
171 		cq->qid_mask = rdev->qpmask;
172 	}
173 	return 0;
174 err4:
175 	dma_free_coherent(&rdev->lldi.pdev->dev, cq->memsize, cq->queue,
176 			  dma_unmap_addr(cq, mapping));
177 err3:
178 	kfree(cq->sw_queue);
179 err2:
180 	c4iw_put_cqid(rdev, cq->cqid, uctx);
181 err1:
182 	return ret;
183 }
184 
185 static void insert_recv_cqe(struct t4_wq *wq, struct t4_cq *cq)
186 {
187 	struct t4_cqe cqe;
188 
189 	PDBG("%s wq %p cq %p sw_cidx %u sw_pidx %u\n", __func__,
190 	     wq, cq, cq->sw_cidx, cq->sw_pidx);
191 	memset(&cqe, 0, sizeof(cqe));
192 	cqe.header = cpu_to_be32(CQE_STATUS_V(T4_ERR_SWFLUSH) |
193 				 CQE_OPCODE_V(FW_RI_SEND) |
194 				 CQE_TYPE_V(0) |
195 				 CQE_SWCQE_V(1) |
196 				 CQE_QPID_V(wq->sq.qid));
197 	cqe.bits_type_ts = cpu_to_be64(CQE_GENBIT_V((u64)cq->gen));
198 	cq->sw_queue[cq->sw_pidx] = cqe;
199 	t4_swcq_produce(cq);
200 }
201 
202 int c4iw_flush_rq(struct t4_wq *wq, struct t4_cq *cq, int count)
203 {
204 	int flushed = 0;
205 	int in_use = wq->rq.in_use - count;
206 
207 	BUG_ON(in_use < 0);
208 	PDBG("%s wq %p cq %p rq.in_use %u skip count %u\n", __func__,
209 	     wq, cq, wq->rq.in_use, count);
210 	while (in_use--) {
211 		insert_recv_cqe(wq, cq);
212 		flushed++;
213 	}
214 	return flushed;
215 }
216 
217 static void insert_sq_cqe(struct t4_wq *wq, struct t4_cq *cq,
218 			  struct t4_swsqe *swcqe)
219 {
220 	struct t4_cqe cqe;
221 
222 	PDBG("%s wq %p cq %p sw_cidx %u sw_pidx %u\n", __func__,
223 	     wq, cq, cq->sw_cidx, cq->sw_pidx);
224 	memset(&cqe, 0, sizeof(cqe));
225 	cqe.header = cpu_to_be32(CQE_STATUS_V(T4_ERR_SWFLUSH) |
226 				 CQE_OPCODE_V(swcqe->opcode) |
227 				 CQE_TYPE_V(1) |
228 				 CQE_SWCQE_V(1) |
229 				 CQE_QPID_V(wq->sq.qid));
230 	CQE_WRID_SQ_IDX(&cqe) = swcqe->idx;
231 	cqe.bits_type_ts = cpu_to_be64(CQE_GENBIT_V((u64)cq->gen));
232 	cq->sw_queue[cq->sw_pidx] = cqe;
233 	t4_swcq_produce(cq);
234 }
235 
236 static void advance_oldest_read(struct t4_wq *wq);
237 
238 int c4iw_flush_sq(struct c4iw_qp *qhp)
239 {
240 	int flushed = 0;
241 	struct t4_wq *wq = &qhp->wq;
242 	struct c4iw_cq *chp = to_c4iw_cq(qhp->ibqp.send_cq);
243 	struct t4_cq *cq = &chp->cq;
244 	int idx;
245 	struct t4_swsqe *swsqe;
246 
247 	if (wq->sq.flush_cidx == -1)
248 		wq->sq.flush_cidx = wq->sq.cidx;
249 	idx = wq->sq.flush_cidx;
250 	BUG_ON(idx >= wq->sq.size);
251 	while (idx != wq->sq.pidx) {
252 		swsqe = &wq->sq.sw_sq[idx];
253 		BUG_ON(swsqe->flushed);
254 		swsqe->flushed = 1;
255 		insert_sq_cqe(wq, cq, swsqe);
256 		if (wq->sq.oldest_read == swsqe) {
257 			BUG_ON(swsqe->opcode != FW_RI_READ_REQ);
258 			advance_oldest_read(wq);
259 		}
260 		flushed++;
261 		if (++idx == wq->sq.size)
262 			idx = 0;
263 	}
264 	wq->sq.flush_cidx += flushed;
265 	if (wq->sq.flush_cidx >= wq->sq.size)
266 		wq->sq.flush_cidx -= wq->sq.size;
267 	return flushed;
268 }
269 
270 static void flush_completed_wrs(struct t4_wq *wq, struct t4_cq *cq)
271 {
272 	struct t4_swsqe *swsqe;
273 	int cidx;
274 
275 	if (wq->sq.flush_cidx == -1)
276 		wq->sq.flush_cidx = wq->sq.cidx;
277 	cidx = wq->sq.flush_cidx;
278 	BUG_ON(cidx > wq->sq.size);
279 
280 	while (cidx != wq->sq.pidx) {
281 		swsqe = &wq->sq.sw_sq[cidx];
282 		if (!swsqe->signaled) {
283 			if (++cidx == wq->sq.size)
284 				cidx = 0;
285 		} else if (swsqe->complete) {
286 
287 			BUG_ON(swsqe->flushed);
288 
289 			/*
290 			 * Insert this completed cqe into the swcq.
291 			 */
292 			PDBG("%s moving cqe into swcq sq idx %u cq idx %u\n",
293 					__func__, cidx, cq->sw_pidx);
294 			swsqe->cqe.header |= htonl(CQE_SWCQE_V(1));
295 			cq->sw_queue[cq->sw_pidx] = swsqe->cqe;
296 			t4_swcq_produce(cq);
297 			swsqe->flushed = 1;
298 			if (++cidx == wq->sq.size)
299 				cidx = 0;
300 			wq->sq.flush_cidx = cidx;
301 		} else
302 			break;
303 	}
304 }
305 
306 static void create_read_req_cqe(struct t4_wq *wq, struct t4_cqe *hw_cqe,
307 		struct t4_cqe *read_cqe)
308 {
309 	read_cqe->u.scqe.cidx = wq->sq.oldest_read->idx;
310 	read_cqe->len = htonl(wq->sq.oldest_read->read_len);
311 	read_cqe->header = htonl(CQE_QPID_V(CQE_QPID(hw_cqe)) |
312 			CQE_SWCQE_V(SW_CQE(hw_cqe)) |
313 			CQE_OPCODE_V(FW_RI_READ_REQ) |
314 			CQE_TYPE_V(1));
315 	read_cqe->bits_type_ts = hw_cqe->bits_type_ts;
316 }
317 
318 static void advance_oldest_read(struct t4_wq *wq)
319 {
320 
321 	u32 rptr = wq->sq.oldest_read - wq->sq.sw_sq + 1;
322 
323 	if (rptr == wq->sq.size)
324 		rptr = 0;
325 	while (rptr != wq->sq.pidx) {
326 		wq->sq.oldest_read = &wq->sq.sw_sq[rptr];
327 
328 		if (wq->sq.oldest_read->opcode == FW_RI_READ_REQ)
329 			return;
330 		if (++rptr == wq->sq.size)
331 			rptr = 0;
332 	}
333 	wq->sq.oldest_read = NULL;
334 }
335 
336 /*
337  * Move all CQEs from the HWCQ into the SWCQ.
338  * Deal with out-of-order and/or completions that complete
339  * prior unsignalled WRs.
340  */
341 void c4iw_flush_hw_cq(struct c4iw_cq *chp)
342 {
343 	struct t4_cqe *hw_cqe, *swcqe, read_cqe;
344 	struct c4iw_qp *qhp;
345 	struct t4_swsqe *swsqe;
346 	int ret;
347 
348 	PDBG("%s  cqid 0x%x\n", __func__, chp->cq.cqid);
349 	ret = t4_next_hw_cqe(&chp->cq, &hw_cqe);
350 
351 	/*
352 	 * This logic is similar to poll_cq(), but not quite the same
353 	 * unfortunately.  Need to move pertinent HW CQEs to the SW CQ but
354 	 * also do any translation magic that poll_cq() normally does.
355 	 */
356 	while (!ret) {
357 		qhp = get_qhp(chp->rhp, CQE_QPID(hw_cqe));
358 
359 		/*
360 		 * drop CQEs with no associated QP
361 		 */
362 		if (qhp == NULL)
363 			goto next_cqe;
364 
365 		if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE)
366 			goto next_cqe;
367 
368 		if (CQE_OPCODE(hw_cqe) == FW_RI_READ_RESP) {
369 
370 			/* If we have reached here because of async
371 			 * event or other error, and have egress error
372 			 * then drop
373 			 */
374 			if (CQE_TYPE(hw_cqe) == 1)
375 				goto next_cqe;
376 
377 			/* drop peer2peer RTR reads.
378 			 */
379 			if (CQE_WRID_STAG(hw_cqe) == 1)
380 				goto next_cqe;
381 
382 			/*
383 			 * Eat completions for unsignaled read WRs.
384 			 */
385 			if (!qhp->wq.sq.oldest_read->signaled) {
386 				advance_oldest_read(&qhp->wq);
387 				goto next_cqe;
388 			}
389 
390 			/*
391 			 * Don't write to the HWCQ, create a new read req CQE
392 			 * in local memory and move it into the swcq.
393 			 */
394 			create_read_req_cqe(&qhp->wq, hw_cqe, &read_cqe);
395 			hw_cqe = &read_cqe;
396 			advance_oldest_read(&qhp->wq);
397 		}
398 
399 		/* if its a SQ completion, then do the magic to move all the
400 		 * unsignaled and now in-order completions into the swcq.
401 		 */
402 		if (SQ_TYPE(hw_cqe)) {
403 			swsqe = &qhp->wq.sq.sw_sq[CQE_WRID_SQ_IDX(hw_cqe)];
404 			swsqe->cqe = *hw_cqe;
405 			swsqe->complete = 1;
406 			flush_completed_wrs(&qhp->wq, &chp->cq);
407 		} else {
408 			swcqe = &chp->cq.sw_queue[chp->cq.sw_pidx];
409 			*swcqe = *hw_cqe;
410 			swcqe->header |= cpu_to_be32(CQE_SWCQE_V(1));
411 			t4_swcq_produce(&chp->cq);
412 		}
413 next_cqe:
414 		t4_hwcq_consume(&chp->cq);
415 		ret = t4_next_hw_cqe(&chp->cq, &hw_cqe);
416 	}
417 }
418 
419 static int cqe_completes_wr(struct t4_cqe *cqe, struct t4_wq *wq)
420 {
421 	if (CQE_OPCODE(cqe) == FW_RI_TERMINATE)
422 		return 0;
423 
424 	if ((CQE_OPCODE(cqe) == FW_RI_RDMA_WRITE) && RQ_TYPE(cqe))
425 		return 0;
426 
427 	if ((CQE_OPCODE(cqe) == FW_RI_READ_RESP) && SQ_TYPE(cqe))
428 		return 0;
429 
430 	if (CQE_SEND_OPCODE(cqe) && RQ_TYPE(cqe) && t4_rq_empty(wq))
431 		return 0;
432 	return 1;
433 }
434 
435 void c4iw_count_rcqes(struct t4_cq *cq, struct t4_wq *wq, int *count)
436 {
437 	struct t4_cqe *cqe;
438 	u32 ptr;
439 
440 	*count = 0;
441 	PDBG("%s count zero %d\n", __func__, *count);
442 	ptr = cq->sw_cidx;
443 	while (ptr != cq->sw_pidx) {
444 		cqe = &cq->sw_queue[ptr];
445 		if (RQ_TYPE(cqe) && (CQE_OPCODE(cqe) != FW_RI_READ_RESP) &&
446 		    (CQE_QPID(cqe) == wq->sq.qid) && cqe_completes_wr(cqe, wq))
447 			(*count)++;
448 		if (++ptr == cq->size)
449 			ptr = 0;
450 	}
451 	PDBG("%s cq %p count %d\n", __func__, cq, *count);
452 }
453 
454 /*
455  * poll_cq
456  *
457  * Caller must:
458  *     check the validity of the first CQE,
459  *     supply the wq assicated with the qpid.
460  *
461  * credit: cq credit to return to sge.
462  * cqe_flushed: 1 iff the CQE is flushed.
463  * cqe: copy of the polled CQE.
464  *
465  * return value:
466  *    0		    CQE returned ok.
467  *    -EAGAIN       CQE skipped, try again.
468  *    -EOVERFLOW    CQ overflow detected.
469  */
470 static int poll_cq(struct t4_wq *wq, struct t4_cq *cq, struct t4_cqe *cqe,
471 		   u8 *cqe_flushed, u64 *cookie, u32 *credit)
472 {
473 	int ret = 0;
474 	struct t4_cqe *hw_cqe, read_cqe;
475 
476 	*cqe_flushed = 0;
477 	*credit = 0;
478 	ret = t4_next_cqe(cq, &hw_cqe);
479 	if (ret)
480 		return ret;
481 
482 	PDBG("%s CQE OVF %u qpid 0x%0x genbit %u type %u status 0x%0x"
483 	     " opcode 0x%0x len 0x%0x wrid_hi_stag 0x%x wrid_low_msn 0x%x\n",
484 	     __func__, CQE_OVFBIT(hw_cqe), CQE_QPID(hw_cqe),
485 	     CQE_GENBIT(hw_cqe), CQE_TYPE(hw_cqe), CQE_STATUS(hw_cqe),
486 	     CQE_OPCODE(hw_cqe), CQE_LEN(hw_cqe), CQE_WRID_HI(hw_cqe),
487 	     CQE_WRID_LOW(hw_cqe));
488 
489 	/*
490 	 * skip cqe's not affiliated with a QP.
491 	 */
492 	if (wq == NULL) {
493 		ret = -EAGAIN;
494 		goto skip_cqe;
495 	}
496 
497 	/*
498 	* skip hw cqe's if the wq is flushed.
499 	*/
500 	if (wq->flushed && !SW_CQE(hw_cqe)) {
501 		ret = -EAGAIN;
502 		goto skip_cqe;
503 	}
504 
505 	/*
506 	 * skip TERMINATE cqes...
507 	 */
508 	if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE) {
509 		ret = -EAGAIN;
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 		PDBG("%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 		PDBG("%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 		PDBG("%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 		PDBG("%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 		PDBG("%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 	PDBG("%s qpid 0x%x type %d opcode %d status 0x%x len %u wrid hi 0x%x "
716 	     "lo 0x%x cookie 0x%llx\n", __func__, CQE_QPID(&cqe),
717 	     CQE_TYPE(&cqe), CQE_OPCODE(&cqe), CQE_STATUS(&cqe), CQE_LEN(&cqe),
718 	     CQE_WRID_HI(&cqe), CQE_WRID_LOW(&cqe), (unsigned long long)cookie);
719 
720 	if (CQE_TYPE(&cqe) == 0) {
721 		if (!CQE_STATUS(&cqe))
722 			wc->byte_len = CQE_LEN(&cqe);
723 		else
724 			wc->byte_len = 0;
725 		wc->opcode = IB_WC_RECV;
726 		if (CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_INV ||
727 		    CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_SE_INV) {
728 			wc->ex.invalidate_rkey = CQE_WRID_STAG(&cqe);
729 			wc->wc_flags |= IB_WC_WITH_INVALIDATE;
730 		}
731 	} else {
732 		switch (CQE_OPCODE(&cqe)) {
733 		case FW_RI_RDMA_WRITE:
734 			wc->opcode = IB_WC_RDMA_WRITE;
735 			break;
736 		case FW_RI_READ_REQ:
737 			wc->opcode = IB_WC_RDMA_READ;
738 			wc->byte_len = CQE_LEN(&cqe);
739 			break;
740 		case FW_RI_SEND_WITH_INV:
741 		case FW_RI_SEND_WITH_SE_INV:
742 			wc->opcode = IB_WC_SEND;
743 			wc->wc_flags |= IB_WC_WITH_INVALIDATE;
744 			break;
745 		case FW_RI_SEND:
746 		case FW_RI_SEND_WITH_SE:
747 			wc->opcode = IB_WC_SEND;
748 			break;
749 		case FW_RI_BIND_MW:
750 			wc->opcode = IB_WC_BIND_MW;
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_FAST_REG_MR;
758 			break;
759 		default:
760 			printk(KERN_ERR MOD "Unexpected opcode %d "
761 			       "in the CQE received for QPID=0x%0x\n",
762 			       CQE_OPCODE(&cqe), CQE_QPID(&cqe));
763 			ret = -EINVAL;
764 			goto out;
765 		}
766 	}
767 
768 	if (cqe_flushed)
769 		wc->status = IB_WC_WR_FLUSH_ERR;
770 	else {
771 
772 		switch (CQE_STATUS(&cqe)) {
773 		case T4_ERR_SUCCESS:
774 			wc->status = IB_WC_SUCCESS;
775 			break;
776 		case T4_ERR_STAG:
777 			wc->status = IB_WC_LOC_ACCESS_ERR;
778 			break;
779 		case T4_ERR_PDID:
780 			wc->status = IB_WC_LOC_PROT_ERR;
781 			break;
782 		case T4_ERR_QPID:
783 		case T4_ERR_ACCESS:
784 			wc->status = IB_WC_LOC_ACCESS_ERR;
785 			break;
786 		case T4_ERR_WRAP:
787 			wc->status = IB_WC_GENERAL_ERR;
788 			break;
789 		case T4_ERR_BOUND:
790 			wc->status = IB_WC_LOC_LEN_ERR;
791 			break;
792 		case T4_ERR_INVALIDATE_SHARED_MR:
793 		case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
794 			wc->status = IB_WC_MW_BIND_ERR;
795 			break;
796 		case T4_ERR_CRC:
797 		case T4_ERR_MARKER:
798 		case T4_ERR_PDU_LEN_ERR:
799 		case T4_ERR_OUT_OF_RQE:
800 		case T4_ERR_DDP_VERSION:
801 		case T4_ERR_RDMA_VERSION:
802 		case T4_ERR_DDP_QUEUE_NUM:
803 		case T4_ERR_MSN:
804 		case T4_ERR_TBIT:
805 		case T4_ERR_MO:
806 		case T4_ERR_MSN_RANGE:
807 		case T4_ERR_IRD_OVERFLOW:
808 		case T4_ERR_OPCODE:
809 		case T4_ERR_INTERNAL_ERR:
810 			wc->status = IB_WC_FATAL_ERR;
811 			break;
812 		case T4_ERR_SWFLUSH:
813 			wc->status = IB_WC_WR_FLUSH_ERR;
814 			break;
815 		default:
816 			printk(KERN_ERR MOD
817 			       "Unexpected cqe_status 0x%x for QPID=0x%0x\n",
818 			       CQE_STATUS(&cqe), CQE_QPID(&cqe));
819 			ret = -EINVAL;
820 		}
821 	}
822 out:
823 	if (wq)
824 		spin_unlock(&qhp->lock);
825 	return ret;
826 }
827 
828 int c4iw_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
829 {
830 	struct c4iw_cq *chp;
831 	unsigned long flags;
832 	int npolled;
833 	int err = 0;
834 
835 	chp = to_c4iw_cq(ibcq);
836 
837 	spin_lock_irqsave(&chp->lock, flags);
838 	for (npolled = 0; npolled < num_entries; ++npolled) {
839 		do {
840 			err = c4iw_poll_cq_one(chp, wc + npolled);
841 		} while (err == -EAGAIN);
842 		if (err)
843 			break;
844 	}
845 	spin_unlock_irqrestore(&chp->lock, flags);
846 	return !err || err == -ENODATA ? npolled : err;
847 }
848 
849 int c4iw_destroy_cq(struct ib_cq *ib_cq)
850 {
851 	struct c4iw_cq *chp;
852 	struct c4iw_ucontext *ucontext;
853 
854 	PDBG("%s ib_cq %p\n", __func__, ib_cq);
855 	chp = to_c4iw_cq(ib_cq);
856 
857 	remove_handle(chp->rhp, &chp->rhp->cqidr, chp->cq.cqid);
858 	atomic_dec(&chp->refcnt);
859 	wait_event(chp->wait, !atomic_read(&chp->refcnt));
860 
861 	ucontext = ib_cq->uobject ? to_c4iw_ucontext(ib_cq->uobject->context)
862 				  : NULL;
863 	destroy_cq(&chp->rhp->rdev, &chp->cq,
864 		   ucontext ? &ucontext->uctx : &chp->cq.rdev->uctx);
865 	kfree(chp);
866 	return 0;
867 }
868 
869 struct ib_cq *c4iw_create_cq(struct ib_device *ibdev, int entries,
870 			     int vector, struct ib_ucontext *ib_context,
871 			     struct ib_udata *udata)
872 {
873 	struct c4iw_dev *rhp;
874 	struct c4iw_cq *chp;
875 	struct c4iw_create_cq_resp uresp;
876 	struct c4iw_ucontext *ucontext = NULL;
877 	int ret;
878 	size_t memsize, hwentries;
879 	struct c4iw_mm_entry *mm, *mm2;
880 
881 	PDBG("%s ib_dev %p entries %d\n", __func__, ibdev, entries);
882 
883 	rhp = to_c4iw_dev(ibdev);
884 
885 	if (vector >= rhp->rdev.lldi.nciq)
886 		return ERR_PTR(-EINVAL);
887 
888 	chp = kzalloc(sizeof(*chp), GFP_KERNEL);
889 	if (!chp)
890 		return ERR_PTR(-ENOMEM);
891 
892 	if (ib_context)
893 		ucontext = to_c4iw_ucontext(ib_context);
894 
895 	/* account for the status page. */
896 	entries++;
897 
898 	/* IQ needs one extra entry to differentiate full vs empty. */
899 	entries++;
900 
901 	/*
902 	 * entries must be multiple of 16 for HW.
903 	 */
904 	entries = roundup(entries, 16);
905 
906 	/*
907 	 * Make actual HW queue 2x to avoid cdix_inc overflows.
908 	 */
909 	hwentries = min(entries * 2, rhp->rdev.hw_queue.t4_max_iq_size);
910 
911 	/*
912 	 * Make HW queue at least 64 entries so GTS updates aren't too
913 	 * frequent.
914 	 */
915 	if (hwentries < 64)
916 		hwentries = 64;
917 
918 	memsize = hwentries * sizeof *chp->cq.queue;
919 
920 	/*
921 	 * memsize must be a multiple of the page size if its a user cq.
922 	 */
923 	if (ucontext)
924 		memsize = roundup(memsize, PAGE_SIZE);
925 	chp->cq.size = hwentries;
926 	chp->cq.memsize = memsize;
927 	chp->cq.vector = vector;
928 
929 	ret = create_cq(&rhp->rdev, &chp->cq,
930 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
931 	if (ret)
932 		goto err1;
933 
934 	chp->rhp = rhp;
935 	chp->cq.size--;				/* status page */
936 	chp->ibcq.cqe = entries - 2;
937 	spin_lock_init(&chp->lock);
938 	spin_lock_init(&chp->comp_handler_lock);
939 	atomic_set(&chp->refcnt, 1);
940 	init_waitqueue_head(&chp->wait);
941 	ret = insert_handle(rhp, &rhp->cqidr, chp, chp->cq.cqid);
942 	if (ret)
943 		goto err2;
944 
945 	if (ucontext) {
946 		mm = kmalloc(sizeof *mm, GFP_KERNEL);
947 		if (!mm)
948 			goto err3;
949 		mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
950 		if (!mm2)
951 			goto err4;
952 
953 		uresp.qid_mask = rhp->rdev.cqmask;
954 		uresp.cqid = chp->cq.cqid;
955 		uresp.size = chp->cq.size;
956 		uresp.memsize = chp->cq.memsize;
957 		spin_lock(&ucontext->mmap_lock);
958 		uresp.key = ucontext->key;
959 		ucontext->key += PAGE_SIZE;
960 		uresp.gts_key = ucontext->key;
961 		ucontext->key += PAGE_SIZE;
962 		spin_unlock(&ucontext->mmap_lock);
963 		ret = ib_copy_to_udata(udata, &uresp,
964 				       sizeof(uresp) - sizeof(uresp.reserved));
965 		if (ret)
966 			goto err5;
967 
968 		mm->key = uresp.key;
969 		mm->addr = virt_to_phys(chp->cq.queue);
970 		mm->len = chp->cq.memsize;
971 		insert_mmap(ucontext, mm);
972 
973 		mm2->key = uresp.gts_key;
974 		mm2->addr = chp->cq.ugts;
975 		mm2->len = PAGE_SIZE;
976 		insert_mmap(ucontext, mm2);
977 	}
978 	PDBG("%s cqid 0x%0x chp %p size %u memsize %zu, dma_addr 0x%0llx\n",
979 	     __func__, chp->cq.cqid, chp, chp->cq.size,
980 	     chp->cq.memsize, (unsigned long long) chp->cq.dma_addr);
981 	return &chp->ibcq;
982 err5:
983 	kfree(mm2);
984 err4:
985 	kfree(mm);
986 err3:
987 	remove_handle(rhp, &rhp->cqidr, chp->cq.cqid);
988 err2:
989 	destroy_cq(&chp->rhp->rdev, &chp->cq,
990 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
991 err1:
992 	kfree(chp);
993 	return ERR_PTR(ret);
994 }
995 
996 int c4iw_resize_cq(struct ib_cq *cq, int cqe, struct ib_udata *udata)
997 {
998 	return -ENOSYS;
999 }
1000 
1001 int c4iw_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
1002 {
1003 	struct c4iw_cq *chp;
1004 	int ret;
1005 	unsigned long flag;
1006 
1007 	chp = to_c4iw_cq(ibcq);
1008 	spin_lock_irqsave(&chp->lock, flag);
1009 	ret = t4_arm_cq(&chp->cq,
1010 			(flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED);
1011 	spin_unlock_irqrestore(&chp->lock, flag);
1012 	if (ret && !(flags & IB_CQ_REPORT_MISSED_EVENTS))
1013 		ret = 0;
1014 	return ret;
1015 }
1016