xref: /freebsd/sys/dev/cxgbe/iw_cxgbe/cq.c (revision 7fdf597e96a02165cfe22ff357b857d5fa15ed8a)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2009-2013 Chelsio, Inc. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *	  copyright notice, this list of conditions and the following
18  *	  disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *	  copyright notice, this list of conditions and the following
22  *	  disclaimer in the documentation and/or other materials
23  *	  provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  */
34 #include <sys/cdefs.h>
35 #include "opt_inet.h"
36 
37 #ifdef TCP_OFFLOAD
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/kernel.h>
41 #include <sys/ktr.h>
42 #include <sys/bus.h>
43 #include <sys/lock.h>
44 #include <sys/mutex.h>
45 #include <sys/rwlock.h>
46 #include <sys/socket.h>
47 #include <sys/sbuf.h>
48 
49 #include "iw_cxgbe.h"
50 #include "user.h"
51 
52 static int destroy_cq(struct c4iw_rdev *rdev, struct t4_cq *cq,
53 		      struct c4iw_dev_ucontext *uctx)
54 {
55 	struct adapter *sc = rdev->adap;
56 	struct c4iw_dev *rhp = rdev_to_c4iw_dev(rdev);
57 	struct fw_ri_res_wr *res_wr;
58 	struct fw_ri_res *res;
59 	int wr_len;
60 	struct c4iw_wr_wait wr_wait;
61 	struct wrqe *wr;
62 
63 	wr_len = sizeof *res_wr + sizeof *res;
64 	wr = alloc_wrqe(wr_len, &sc->sge.ctrlq[0]);
65                 if (wr == NULL)
66                         return (0);
67         res_wr = wrtod(wr);
68 	memset(res_wr, 0, wr_len);
69 	res_wr->op_nres = cpu_to_be32(
70 			V_FW_WR_OP(FW_RI_RES_WR) |
71 			V_FW_RI_RES_WR_NRES(1) |
72 			F_FW_WR_COMPL);
73 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
74 	res_wr->cookie = (unsigned long) &wr_wait;
75 	res = res_wr->res;
76 	res->u.cq.restype = FW_RI_RES_TYPE_CQ;
77 	res->u.cq.op = FW_RI_RES_OP_RESET;
78 	res->u.cq.iqid = cpu_to_be32(cq->cqid);
79 
80 	c4iw_init_wr_wait(&wr_wait);
81 
82 	t4_wrq_tx(sc, wr);
83 
84 	c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, NULL, __func__);
85 
86 	kfree(cq->sw_queue);
87 	dma_free_coherent(rhp->ibdev.dma_device,
88 			  cq->memsize, cq->queue,
89 			  dma_unmap_addr(cq, mapping));
90 	c4iw_put_cqid(rdev, cq->cqid, uctx);
91 	return 0;
92 }
93 
94 static int
95 create_cq(struct c4iw_rdev *rdev, struct t4_cq *cq,
96     struct c4iw_dev_ucontext *uctx)
97 {
98 	struct adapter *sc = rdev->adap;
99 	struct c4iw_dev *rhp = rdev_to_c4iw_dev(rdev);
100 	struct fw_ri_res_wr *res_wr;
101 	struct fw_ri_res *res;
102 	int wr_len;
103 	int user = (uctx != &rdev->uctx);
104 	struct c4iw_wr_wait wr_wait;
105 	int ret;
106 	struct wrqe *wr;
107 	u64 cq_bar2_qoffset = 0;
108 
109 	if (__predict_false(c4iw_stopped(rdev)))
110 		return -EIO;
111 	cq->cqid = c4iw_get_cqid(rdev, uctx);
112 	if (!cq->cqid) {
113 		ret = -ENOMEM;
114 		goto err1;
115 	}
116 
117 	if (!user) {
118 		cq->sw_queue = kzalloc(cq->memsize, GFP_KERNEL);
119 		if (!cq->sw_queue) {
120 			ret = -ENOMEM;
121 			goto err2;
122 		}
123 	}
124 	cq->queue = dma_alloc_coherent(rhp->ibdev.dma_device, cq->memsize,
125 				       &cq->dma_addr, GFP_KERNEL);
126 	if (!cq->queue) {
127 		ret = -ENOMEM;
128 		goto err3;
129 	}
130 	dma_unmap_addr_set(cq, mapping, cq->dma_addr);
131 	memset(cq->queue, 0, cq->memsize);
132 
133 	/* build fw_ri_res_wr */
134 	wr_len = sizeof *res_wr + sizeof *res;
135 
136 	wr = alloc_wrqe(wr_len, &sc->sge.ctrlq[0]);
137 	if (wr == NULL)
138 		return (0);
139         res_wr = wrtod(wr);
140 
141 	memset(res_wr, 0, wr_len);
142 	res_wr->op_nres = cpu_to_be32(
143 			V_FW_WR_OP(FW_RI_RES_WR) |
144 			V_FW_RI_RES_WR_NRES(1) |
145 			F_FW_WR_COMPL);
146 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
147 	res_wr->cookie = (unsigned long) &wr_wait;
148 	res = res_wr->res;
149 	res->u.cq.restype = FW_RI_RES_TYPE_CQ;
150 	res->u.cq.op = FW_RI_RES_OP_WRITE;
151 	res->u.cq.iqid = cpu_to_be32(cq->cqid);
152 	//Fixme: Always use first queue id for IQANDSTINDEX. Linux does the same.
153 	res->u.cq.iqandst_to_iqandstindex = cpu_to_be32(
154 			V_FW_RI_RES_WR_IQANUS(0) |
155 			V_FW_RI_RES_WR_IQANUD(1) |
156 			F_FW_RI_RES_WR_IQANDST |
157 			V_FW_RI_RES_WR_IQANDSTINDEX(sc->sge.ofld_rxq[0].iq.abs_id));
158 	res->u.cq.iqdroprss_to_iqesize = cpu_to_be16(
159 			F_FW_RI_RES_WR_IQDROPRSS |
160 			V_FW_RI_RES_WR_IQPCIECH(2) |
161 			V_FW_RI_RES_WR_IQINTCNTTHRESH(0) |
162 			F_FW_RI_RES_WR_IQO |
163 			V_FW_RI_RES_WR_IQESIZE(1));
164 	res->u.cq.iqsize = cpu_to_be16(cq->size);
165 	res->u.cq.iqaddr = cpu_to_be64(cq->dma_addr);
166 
167 	c4iw_init_wr_wait(&wr_wait);
168 
169 	t4_wrq_tx(sc, wr);
170 
171 	CTR2(KTR_IW_CXGBE, "%s wait_event wr_wait %p", __func__, &wr_wait);
172 	ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, NULL, __func__);
173 	if (ret)
174 		goto err4;
175 
176 	cq->gen = 1;
177 	cq->rdev = rdev;
178 
179 	/* Determine the BAR2 queue offset and qid. */
180 	t4_bar2_sge_qregs(rdev->adap, cq->cqid, T4_BAR2_QTYPE_INGRESS, user,
181 			&cq_bar2_qoffset, &cq->bar2_qid);
182 
183 	/* If user mapping then compute the page-aligned physical
184 	 * address for mapping.
185 	 */
186 	if (user)
187 		cq->bar2_pa = (rdev->bar2_pa + cq_bar2_qoffset) & PAGE_MASK;
188 	else
189 		cq->bar2_va = (void __iomem *)((u64)rdev->bar2_kva +
190 			cq_bar2_qoffset);
191 
192 	return 0;
193 err4:
194 	dma_free_coherent(rhp->ibdev.dma_device, cq->memsize, cq->queue,
195 			  dma_unmap_addr(cq, mapping));
196 err3:
197 	kfree(cq->sw_queue);
198 err2:
199 	c4iw_put_cqid(rdev, cq->cqid, uctx);
200 err1:
201 	return ret;
202 }
203 
204 static void insert_recv_cqe(struct t4_wq *wq, struct t4_cq *cq)
205 {
206 	struct t4_cqe cqe;
207 
208 	CTR5(KTR_IW_CXGBE, "%s wq %p cq %p sw_cidx %u sw_pidx %u", __func__, wq,
209 	    cq, cq->sw_cidx, cq->sw_pidx);
210 	memset(&cqe, 0, sizeof(cqe));
211 	cqe.header = cpu_to_be32(V_CQE_STATUS(T4_ERR_SWFLUSH) |
212 				 V_CQE_OPCODE(FW_RI_SEND) |
213 				 V_CQE_TYPE(0) |
214 				 V_CQE_SWCQE(1) |
215 				 V_CQE_QPID(wq->sq.qid));
216 	cqe.bits_type_ts = cpu_to_be64(V_CQE_GENBIT((u64)cq->gen));
217 	cq->sw_queue[cq->sw_pidx] = cqe;
218 	t4_swcq_produce(cq);
219 }
220 
221 int c4iw_flush_rq(struct t4_wq *wq, struct t4_cq *cq, int count)
222 {
223 	int flushed = 0;
224 	int in_use = wq->rq.in_use - count;
225 
226 	BUG_ON(in_use < 0);
227 	CTR5(KTR_IW_CXGBE, "%s wq %p cq %p rq.in_use %u skip count %u",
228 	    __func__, wq, cq, wq->rq.in_use, count);
229 	while (in_use--) {
230 		insert_recv_cqe(wq, cq);
231 		flushed++;
232 	}
233 	return flushed;
234 }
235 
236 static void insert_sq_cqe(struct t4_wq *wq, struct t4_cq *cq,
237 			  struct t4_swsqe *swcqe)
238 {
239 	struct t4_cqe cqe;
240 
241 	CTR5(KTR_IW_CXGBE, "%s wq %p cq %p sw_cidx %u sw_pidx %u", __func__, wq,
242 	    cq, cq->sw_cidx, cq->sw_pidx);
243 	memset(&cqe, 0, sizeof(cqe));
244 	cqe.header = cpu_to_be32(V_CQE_STATUS(T4_ERR_SWFLUSH) |
245 				 V_CQE_OPCODE(swcqe->opcode) |
246 				 V_CQE_TYPE(1) |
247 				 V_CQE_SWCQE(1) |
248 				 V_CQE_QPID(wq->sq.qid));
249 	CQE_WRID_SQ_IDX(&cqe) = swcqe->idx;
250 	cqe.bits_type_ts = cpu_to_be64(V_CQE_GENBIT((u64)cq->gen));
251 	cq->sw_queue[cq->sw_pidx] = cqe;
252 	t4_swcq_produce(cq);
253 }
254 
255 static void advance_oldest_read(struct t4_wq *wq);
256 
257 int c4iw_flush_sq(struct c4iw_qp *qhp)
258 {
259 	int flushed = 0;
260 	struct t4_wq *wq = &qhp->wq;
261 	struct c4iw_cq *chp = to_c4iw_cq(qhp->ibqp.send_cq);
262 	struct t4_cq *cq = &chp->cq;
263 	int idx;
264 	struct t4_swsqe *swsqe;
265 
266 	if (wq->sq.flush_cidx == -1)
267 		wq->sq.flush_cidx = wq->sq.cidx;
268 	idx = wq->sq.flush_cidx;
269 	BUG_ON(idx >= wq->sq.size);
270 	while (idx != wq->sq.pidx) {
271 		swsqe = &wq->sq.sw_sq[idx];
272 		BUG_ON(swsqe->flushed);
273 		swsqe->flushed = 1;
274 		insert_sq_cqe(wq, cq, swsqe);
275 		if (wq->sq.oldest_read == swsqe) {
276 			BUG_ON(swsqe->opcode != FW_RI_READ_REQ);
277 			advance_oldest_read(wq);
278 		}
279 		flushed++;
280 		if (++idx == wq->sq.size)
281 			idx = 0;
282 	}
283 	wq->sq.flush_cidx += flushed;
284 	if (wq->sq.flush_cidx >= wq->sq.size)
285 		wq->sq.flush_cidx -= wq->sq.size;
286 	return flushed;
287 }
288 
289 static void flush_completed_wrs(struct t4_wq *wq, struct t4_cq *cq)
290 {
291 	struct t4_swsqe *swsqe;
292 	int cidx;
293 
294 	if (wq->sq.flush_cidx == -1)
295 		wq->sq.flush_cidx = wq->sq.cidx;
296 	cidx = wq->sq.flush_cidx;
297 	BUG_ON(cidx > wq->sq.size);
298 
299 	while (cidx != wq->sq.pidx) {
300 		swsqe = &wq->sq.sw_sq[cidx];
301 		if (!swsqe->signaled) {
302 			if (++cidx == wq->sq.size)
303 				cidx = 0;
304 		} else if (swsqe->complete) {
305 
306 			BUG_ON(swsqe->flushed);
307 
308 			/*
309 			 * Insert this completed cqe into the swcq.
310 			 */
311 			CTR3(KTR_IW_CXGBE,
312 				"%s moving cqe into swcq sq idx %u cq idx %u\n",
313 				__func__, cidx, cq->sw_pidx);
314 			swsqe->cqe.header |= htonl(V_CQE_SWCQE(1));
315 			cq->sw_queue[cq->sw_pidx] = swsqe->cqe;
316 			t4_swcq_produce(cq);
317 			swsqe->flushed = 1;
318 			if (++cidx == wq->sq.size)
319 				cidx = 0;
320 			wq->sq.flush_cidx = cidx;
321 		} else
322 			break;
323 	}
324 }
325 
326 static void create_read_req_cqe(struct t4_wq *wq, struct t4_cqe *hw_cqe,
327 		struct t4_cqe *read_cqe)
328 {
329 	read_cqe->u.scqe.cidx = wq->sq.oldest_read->idx;
330 	read_cqe->len = htonl(wq->sq.oldest_read->read_len);
331 	read_cqe->header = htonl(V_CQE_QPID(CQE_QPID(hw_cqe)) |
332 			V_CQE_SWCQE(SW_CQE(hw_cqe)) |
333 			V_CQE_OPCODE(FW_RI_READ_REQ) |
334 			V_CQE_TYPE(1));
335 	read_cqe->bits_type_ts = hw_cqe->bits_type_ts;
336 }
337 
338 static void advance_oldest_read(struct t4_wq *wq)
339 {
340 
341 	u32 rptr = wq->sq.oldest_read - wq->sq.sw_sq + 1;
342 
343 	if (rptr == wq->sq.size)
344 		rptr = 0;
345 	while (rptr != wq->sq.pidx) {
346 		wq->sq.oldest_read = &wq->sq.sw_sq[rptr];
347 
348 		if (wq->sq.oldest_read->opcode == FW_RI_READ_REQ)
349 			return;
350 		if (++rptr == wq->sq.size)
351 			rptr = 0;
352 	}
353 	wq->sq.oldest_read = NULL;
354 }
355 
356 /*
357  * Move all CQEs from the HWCQ into the SWCQ.
358  * Deal with out-of-order and/or completions that complete
359  * prior unsignalled WRs.
360  */
361 void c4iw_flush_hw_cq(struct c4iw_cq *chp)
362 {
363 	struct t4_cqe *hw_cqe, *swcqe, read_cqe;
364 	struct c4iw_qp *qhp;
365 	struct t4_swsqe *swsqe;
366 	int ret;
367 
368 	CTR3(KTR_IW_CXGBE, "%s cq %p cqid 0x%x", __func__, &chp->cq,
369 			chp->cq.cqid);
370 	ret = t4_next_hw_cqe(&chp->cq, &hw_cqe);
371 
372 	/*
373 	 * This logic is similar to poll_cq(), but not quite the same
374 	 * unfortunately.  Need to move pertinent HW CQEs to the SW CQ but
375 	 * also do any translation magic that poll_cq() normally does.
376 	 */
377 	while (!ret) {
378 		qhp = get_qhp(chp->rhp, CQE_QPID(hw_cqe));
379 
380 		/*
381 		 * drop CQEs with no associated QP
382 		 */
383 		if (qhp == NULL)
384 			goto next_cqe;
385 
386 		if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE)
387 			goto next_cqe;
388 
389 		if (CQE_OPCODE(hw_cqe) == FW_RI_READ_RESP) {
390 
391 			/* If we have reached here because of async
392 			 * event or other error, and have egress error
393 			 * then drop
394 			 */
395 			if (CQE_TYPE(hw_cqe) == 1)
396 				goto next_cqe;
397 
398 			/* drop peer2peer RTR reads.
399 			 */
400 			if (CQE_WRID_STAG(hw_cqe) == 1)
401 				goto next_cqe;
402 
403 			/*
404 			 * Eat completions for unsignaled read WRs.
405 			 */
406 			if (!qhp->wq.sq.oldest_read->signaled) {
407 				advance_oldest_read(&qhp->wq);
408 				goto next_cqe;
409 			}
410 
411 			/*
412 			 * Don't write to the HWCQ, create a new read req CQE
413 			 * in local memory and move it into the swcq.
414 			 */
415 			create_read_req_cqe(&qhp->wq, hw_cqe, &read_cqe);
416 			hw_cqe = &read_cqe;
417 			advance_oldest_read(&qhp->wq);
418 		}
419 
420 		/* if its a SQ completion, then do the magic to move all the
421 		 * unsignaled and now in-order completions into the swcq.
422 		 */
423 		if (SQ_TYPE(hw_cqe)) {
424 			swsqe = &qhp->wq.sq.sw_sq[CQE_WRID_SQ_IDX(hw_cqe)];
425 			swsqe->cqe = *hw_cqe;
426 			swsqe->complete = 1;
427 			flush_completed_wrs(&qhp->wq, &chp->cq);
428 		} else {
429 			swcqe = &chp->cq.sw_queue[chp->cq.sw_pidx];
430 			*swcqe = *hw_cqe;
431 			swcqe->header |= cpu_to_be32(V_CQE_SWCQE(1));
432 			t4_swcq_produce(&chp->cq);
433 		}
434 next_cqe:
435 		t4_hwcq_consume(&chp->cq);
436 		ret = t4_next_hw_cqe(&chp->cq, &hw_cqe);
437 	}
438 }
439 
440 static int cqe_completes_wr(struct t4_cqe *cqe, struct t4_wq *wq)
441 {
442 	if (CQE_OPCODE(cqe) == FW_RI_TERMINATE)
443 		return 0;
444 
445 	if ((CQE_OPCODE(cqe) == FW_RI_RDMA_WRITE) && RQ_TYPE(cqe))
446 		return 0;
447 
448 	if ((CQE_OPCODE(cqe) == FW_RI_READ_RESP) && SQ_TYPE(cqe))
449 		return 0;
450 
451 	if (CQE_SEND_OPCODE(cqe) && RQ_TYPE(cqe) && t4_rq_empty(wq))
452 		return 0;
453 	return 1;
454 }
455 
456 void c4iw_count_rcqes(struct t4_cq *cq, struct t4_wq *wq, int *count)
457 {
458 	struct t4_cqe *cqe;
459 	u32 ptr;
460 
461 	*count = 0;
462 	CTR2(KTR_IW_CXGBE, "%s count zero %d", __func__, *count);
463 	ptr = cq->sw_cidx;
464 	while (ptr != cq->sw_pidx) {
465 		cqe = &cq->sw_queue[ptr];
466 		if (RQ_TYPE(cqe) && (CQE_OPCODE(cqe) != FW_RI_READ_RESP) &&
467 		    (CQE_QPID(cqe) == wq->sq.qid) && cqe_completes_wr(cqe, wq))
468 			(*count)++;
469 		if (++ptr == cq->size)
470 			ptr = 0;
471 	}
472 	CTR3(KTR_IW_CXGBE, "%s cq %p count %d", __func__, cq, *count);
473 }
474 
475 /*
476  * poll_cq
477  *
478  * Caller must:
479  *     check the validity of the first CQE,
480  *     supply the wq assicated with the qpid.
481  *
482  * credit: cq credit to return to sge.
483  * cqe_flushed: 1 iff the CQE is flushed.
484  * cqe: copy of the polled CQE.
485  *
486  * return value:
487  *    0		    CQE returned ok.
488  *    -EAGAIN       CQE skipped, try again.
489  *    -EOVERFLOW    CQ overflow detected.
490  */
491 static int poll_cq(struct t4_wq *wq, struct t4_cq *cq, struct t4_cqe *cqe,
492 		   u8 *cqe_flushed, u64 *cookie, u32 *credit)
493 {
494 	int ret = 0;
495 	struct t4_cqe *hw_cqe, read_cqe;
496 
497 	*cqe_flushed = 0;
498 	*credit = 0;
499 	ret = t4_next_cqe(cq, &hw_cqe);
500 	if (ret)
501 		return ret;
502 
503 	CTR6(KTR_IW_CXGBE,
504 	    "%s CQE OVF %u qpid 0x%0x genbit %u type %u status 0x%0x", __func__,
505 	    CQE_OVFBIT(hw_cqe), CQE_QPID(hw_cqe), CQE_GENBIT(hw_cqe),
506 	    CQE_TYPE(hw_cqe), CQE_STATUS(hw_cqe));
507 	CTR5(KTR_IW_CXGBE,
508 	    "%s opcode 0x%0x len 0x%0x wrid_hi_stag 0x%x wrid_low_msn 0x%x",
509 	    __func__, CQE_OPCODE(hw_cqe), CQE_LEN(hw_cqe), CQE_WRID_HI(hw_cqe),
510 	    CQE_WRID_LOW(hw_cqe));
511 
512 	/*
513 	 * skip cqe's not affiliated with a QP.
514 	 */
515 	if (wq == NULL) {
516 		ret = -EAGAIN;
517 		goto skip_cqe;
518 	}
519 
520 	/*
521 	* skip hw cqe's if the wq is flushed.
522 	*/
523 	if (wq->flushed && !SW_CQE(hw_cqe)) {
524 		ret = -EAGAIN;
525 		goto skip_cqe;
526 	}
527 
528 	/*
529 	 * skip TERMINATE cqes...
530 	 */
531 	if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE) {
532 		ret = -EAGAIN;
533 		goto skip_cqe;
534 	}
535 
536 	/*
537 	 * Special cqe for drain WR completions...
538 	 */
539 	if (CQE_OPCODE(hw_cqe) == C4IW_DRAIN_OPCODE) {
540 		*cookie = CQE_DRAIN_COOKIE(hw_cqe);
541 		*cqe = *hw_cqe;
542 		goto skip_cqe;
543 	}
544 
545 	/*
546 	 * Gotta tweak READ completions:
547 	 *	1) the cqe doesn't contain the sq_wptr from the wr.
548 	 *	2) opcode not reflected from the wr.
549 	 *	3) read_len not reflected from the wr.
550 	 *	4) cq_type is RQ_TYPE not SQ_TYPE.
551 	 */
552 	if (RQ_TYPE(hw_cqe) && (CQE_OPCODE(hw_cqe) == FW_RI_READ_RESP)) {
553 
554 		/* If we have reached here because of async
555 		 * event or other error, and have egress error
556 		 * then drop
557 		 */
558 		if (CQE_TYPE(hw_cqe) == 1) {
559 			if (CQE_STATUS(hw_cqe))
560 				t4_set_wq_in_error(wq);
561 			ret = -EAGAIN;
562 			goto skip_cqe;
563 		}
564 
565 		/* If this is an unsolicited read response, then the read
566 		 * was generated by the kernel driver as part of peer-2-peer
567 		 * connection setup.  So ignore the completion.
568 		 */
569 		if (CQE_WRID_STAG(hw_cqe) == 1) {
570 			if (CQE_STATUS(hw_cqe))
571 				t4_set_wq_in_error(wq);
572 			ret = -EAGAIN;
573 			goto skip_cqe;
574 		}
575 
576 		/*
577 		 * Eat completions for unsignaled read WRs.
578 		 */
579 		if (!wq->sq.oldest_read->signaled) {
580 			advance_oldest_read(wq);
581 			ret = -EAGAIN;
582 			goto skip_cqe;
583 		}
584 
585 		/*
586 		 * Don't write to the HWCQ, so create a new read req CQE
587 		 * in local memory.
588 		 */
589 		create_read_req_cqe(wq, hw_cqe, &read_cqe);
590 		hw_cqe = &read_cqe;
591 		advance_oldest_read(wq);
592 	}
593 
594 	if (CQE_STATUS(hw_cqe) || t4_wq_in_error(wq)) {
595 		*cqe_flushed = (CQE_STATUS(hw_cqe) == T4_ERR_SWFLUSH);
596 		t4_set_wq_in_error(wq);
597 	}
598 
599 	/*
600 	 * RECV completion.
601 	 */
602 	if (RQ_TYPE(hw_cqe)) {
603 
604 		/*
605 		 * HW only validates 4 bits of MSN.  So we must validate that
606 		 * the MSN in the SEND is the next expected MSN.  If its not,
607 		 * then we complete this with T4_ERR_MSN and mark the wq in
608 		 * error.
609 		 */
610 
611 		if (t4_rq_empty(wq)) {
612 			t4_set_wq_in_error(wq);
613 			ret = -EAGAIN;
614 			goto skip_cqe;
615 		}
616 		if (unlikely((CQE_WRID_MSN(hw_cqe) != (wq->rq.msn)))) {
617 			t4_set_wq_in_error(wq);
618 			hw_cqe->header |= htonl(V_CQE_STATUS(T4_ERR_MSN));
619 			goto proc_cqe;
620 		}
621 		goto proc_cqe;
622 	}
623 
624 	/*
625 	 * If we get here its a send completion.
626 	 *
627 	 * Handle out of order completion. These get stuffed
628 	 * in the SW SQ. Then the SW SQ is walked to move any
629 	 * now in-order completions into the SW CQ.  This handles
630 	 * 2 cases:
631 	 *	1) reaping unsignaled WRs when the first subsequent
632 	 *	   signaled WR is completed.
633 	 *	2) out of order read completions.
634 	 */
635 	if (!SW_CQE(hw_cqe) && (CQE_WRID_SQ_IDX(hw_cqe) != wq->sq.cidx)) {
636 		struct t4_swsqe *swsqe;
637 
638 		CTR2(KTR_IW_CXGBE,
639 		    "%s out of order completion going in sw_sq at idx %u",
640 		    __func__, CQE_WRID_SQ_IDX(hw_cqe));
641 		swsqe = &wq->sq.sw_sq[CQE_WRID_SQ_IDX(hw_cqe)];
642 		swsqe->cqe = *hw_cqe;
643 		swsqe->complete = 1;
644 		ret = -EAGAIN;
645 		goto flush_wq;
646 	}
647 
648 proc_cqe:
649 	*cqe = *hw_cqe;
650 
651 	/*
652 	 * Reap the associated WR(s) that are freed up with this
653 	 * completion.
654 	 */
655 	if (SQ_TYPE(hw_cqe)) {
656 		int idx = CQE_WRID_SQ_IDX(hw_cqe);
657 		BUG_ON(idx >= wq->sq.size);
658 
659 		/*
660 		* Account for any unsignaled completions completed by
661 		* this signaled completion.  In this case, cidx points
662 		* to the first unsignaled one, and idx points to the
663 		* signaled one.  So adjust in_use based on this delta.
664 		* if this is not completing any unsigned wrs, then the
665 		* delta will be 0. Handle wrapping also!
666 		*/
667 		if (idx < wq->sq.cidx)
668 			wq->sq.in_use -= wq->sq.size + idx - wq->sq.cidx;
669 		else
670 			wq->sq.in_use -= idx - wq->sq.cidx;
671 		BUG_ON(wq->sq.in_use <= 0 && wq->sq.in_use >= wq->sq.size);
672 
673 		wq->sq.cidx = (uint16_t)idx;
674 		CTR2(KTR_IW_CXGBE, "%s completing sq idx %u",
675 				__func__, wq->sq.cidx);
676 		*cookie = wq->sq.sw_sq[wq->sq.cidx].wr_id;
677 		t4_sq_consume(wq);
678 	} else {
679 		CTR2(KTR_IW_CXGBE, "%s completing rq idx %u",
680 		     __func__, wq->rq.cidx);
681 		*cookie = wq->rq.sw_rq[wq->rq.cidx].wr_id;
682 		BUG_ON(t4_rq_empty(wq));
683 		t4_rq_consume(wq);
684 		goto skip_cqe;
685 	}
686 
687 flush_wq:
688 	/*
689 	 * Flush any completed cqes that are now in-order.
690 	 */
691 	flush_completed_wrs(wq, cq);
692 
693 skip_cqe:
694 	if (SW_CQE(hw_cqe)) {
695 		CTR4(KTR_IW_CXGBE, "%s cq %p cqid 0x%x skip sw cqe cidx %u",
696 		     __func__, cq, cq->cqid, cq->sw_cidx);
697 		t4_swcq_consume(cq);
698 	} else {
699 		CTR4(KTR_IW_CXGBE, "%s cq %p cqid 0x%x skip hw cqe cidx %u",
700 		     __func__, cq, cq->cqid, cq->cidx);
701 		t4_hwcq_consume(cq);
702 	}
703 	return ret;
704 }
705 
706 /*
707  * Get one cq entry from c4iw and map it to openib.
708  *
709  * Returns:
710  *	0			cqe returned
711  *	-ENODATA		EMPTY;
712  *	-EAGAIN			caller must try again
713  *	any other -errno	fatal error
714  */
715 static int c4iw_poll_cq_one(struct c4iw_cq *chp, struct ib_wc *wc)
716 {
717 	struct c4iw_qp *qhp = NULL;
718 	struct t4_cqe cqe = {0, 0}, *rd_cqe;
719 	struct t4_wq *wq;
720 	u32 credit = 0;
721 	u8 cqe_flushed;
722 	u64 cookie = 0;
723 	int ret;
724 
725 	ret = t4_next_cqe(&chp->cq, &rd_cqe);
726 
727 	if (ret)
728 		return ret;
729 
730 	qhp = get_qhp(chp->rhp, CQE_QPID(rd_cqe));
731 	if (!qhp)
732 		wq = NULL;
733 	else {
734 		spin_lock(&qhp->lock);
735 		wq = &(qhp->wq);
736 	}
737 	ret = poll_cq(wq, &(chp->cq), &cqe, &cqe_flushed, &cookie, &credit);
738 	if (ret)
739 		goto out;
740 
741 	wc->wr_id = cookie;
742 	wc->qp = &qhp->ibqp;
743 	wc->vendor_err = CQE_STATUS(&cqe);
744 	wc->wc_flags = 0;
745 
746 	CTR5(KTR_IW_CXGBE, "%s qpid 0x%x type %d opcode %d status 0x%x",
747 	    __func__, CQE_QPID(&cqe), CQE_TYPE(&cqe), CQE_OPCODE(&cqe),
748 	    CQE_STATUS(&cqe));
749 	CTR5(KTR_IW_CXGBE, "%s len %u wrid hi 0x%x lo 0x%x cookie 0x%llx",
750 	    __func__, CQE_LEN(&cqe), CQE_WRID_HI(&cqe), CQE_WRID_LOW(&cqe),
751 	    (unsigned long long)cookie);
752 
753 	if (CQE_TYPE(&cqe) == 0) {
754 		if (!CQE_STATUS(&cqe))
755 			wc->byte_len = CQE_LEN(&cqe);
756 		else
757 			wc->byte_len = 0;
758 		wc->opcode = IB_WC_RECV;
759 		if (CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_INV ||
760 		    CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_SE_INV) {
761 			wc->ex.invalidate_rkey = CQE_WRID_STAG(&cqe);
762 			wc->wc_flags |= IB_WC_WITH_INVALIDATE;
763 			c4iw_invalidate_mr(qhp->rhp, wc->ex.invalidate_rkey);
764 		}
765 	} else {
766 		switch (CQE_OPCODE(&cqe)) {
767 		case FW_RI_RDMA_WRITE:
768 			wc->opcode = IB_WC_RDMA_WRITE;
769 			break;
770 		case FW_RI_READ_REQ:
771 			wc->opcode = IB_WC_RDMA_READ;
772 			wc->byte_len = CQE_LEN(&cqe);
773 			break;
774 		case FW_RI_SEND_WITH_INV:
775 		case FW_RI_SEND_WITH_SE_INV:
776 			wc->opcode = IB_WC_SEND;
777 			wc->wc_flags |= IB_WC_WITH_INVALIDATE;
778 			break;
779 		case FW_RI_SEND:
780 		case FW_RI_SEND_WITH_SE:
781 			wc->opcode = IB_WC_SEND;
782 			break;
783 		case FW_RI_LOCAL_INV:
784 			wc->opcode = IB_WC_LOCAL_INV;
785 			break;
786 		case FW_RI_FAST_REGISTER:
787 			wc->opcode = IB_WC_REG_MR;
788 
789 			/* Invalidate the MR if the fastreg failed */
790 			if (CQE_STATUS(&cqe) != T4_ERR_SUCCESS)
791 				c4iw_invalidate_mr(qhp->rhp,
792 						   CQE_WRID_FR_STAG(&cqe));
793 			break;
794 		case C4IW_DRAIN_OPCODE:
795 			wc->opcode = IB_WC_SEND;
796 			break;
797 		default:
798 			printf("Unexpected opcode %d "
799 			       "in the CQE received for QPID = 0x%0x\n",
800 			       CQE_OPCODE(&cqe), CQE_QPID(&cqe));
801 			ret = -EINVAL;
802 			goto out;
803 		}
804 	}
805 
806 	if (cqe_flushed)
807 		wc->status = IB_WC_WR_FLUSH_ERR;
808 	else {
809 
810 		switch (CQE_STATUS(&cqe)) {
811 		case T4_ERR_SUCCESS:
812 			wc->status = IB_WC_SUCCESS;
813 			break;
814 		case T4_ERR_STAG:
815 			wc->status = IB_WC_LOC_ACCESS_ERR;
816 			break;
817 		case T4_ERR_PDID:
818 			wc->status = IB_WC_LOC_PROT_ERR;
819 			break;
820 		case T4_ERR_QPID:
821 		case T4_ERR_ACCESS:
822 			wc->status = IB_WC_LOC_ACCESS_ERR;
823 			break;
824 		case T4_ERR_WRAP:
825 			wc->status = IB_WC_GENERAL_ERR;
826 			break;
827 		case T4_ERR_BOUND:
828 			wc->status = IB_WC_LOC_LEN_ERR;
829 			break;
830 		case T4_ERR_INVALIDATE_SHARED_MR:
831 		case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
832 			wc->status = IB_WC_MW_BIND_ERR;
833 			break;
834 		case T4_ERR_CRC:
835 		case T4_ERR_MARKER:
836 		case T4_ERR_PDU_LEN_ERR:
837 		case T4_ERR_OUT_OF_RQE:
838 		case T4_ERR_DDP_VERSION:
839 		case T4_ERR_RDMA_VERSION:
840 		case T4_ERR_DDP_QUEUE_NUM:
841 		case T4_ERR_MSN:
842 		case T4_ERR_TBIT:
843 		case T4_ERR_MO:
844 		case T4_ERR_MSN_RANGE:
845 		case T4_ERR_IRD_OVERFLOW:
846 		case T4_ERR_OPCODE:
847 		case T4_ERR_INTERNAL_ERR:
848 			wc->status = IB_WC_FATAL_ERR;
849 			break;
850 		case T4_ERR_SWFLUSH:
851 			wc->status = IB_WC_WR_FLUSH_ERR;
852 			break;
853 		default:
854 			printf("Unexpected cqe_status 0x%x for QPID = 0x%0x\n",
855 			       CQE_STATUS(&cqe), CQE_QPID(&cqe));
856 			wc->status = IB_WC_FATAL_ERR;
857 		}
858 	}
859 out:
860 	if (wq)
861 		spin_unlock(&qhp->lock);
862 	return ret;
863 }
864 
865 int c4iw_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
866 {
867 	struct c4iw_cq *chp;
868 	unsigned long flags;
869 	int npolled;
870 	int err = 0;
871 
872 	chp = to_c4iw_cq(ibcq);
873 
874 	spin_lock_irqsave(&chp->lock, flags);
875 	for (npolled = 0; npolled < num_entries; ++npolled) {
876 		do {
877 			err = c4iw_poll_cq_one(chp, wc + npolled);
878 		} while (err == -EAGAIN);
879 		if (err)
880 			break;
881 	}
882 	spin_unlock_irqrestore(&chp->lock, flags);
883 	return !err || err == -ENODATA ? npolled : err;
884 }
885 
886 void c4iw_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata)
887 {
888 	struct c4iw_cq *chp;
889 	struct c4iw_ucontext *ucontext;
890 
891 	CTR2(KTR_IW_CXGBE, "%s ib_cq %p", __func__, ib_cq);
892 	chp = to_c4iw_cq(ib_cq);
893 
894 	remove_handle(chp->rhp, &chp->rhp->cqidr, chp->cq.cqid);
895 	atomic_dec(&chp->refcnt);
896 	wait_event(chp->wait, !atomic_read(&chp->refcnt));
897 
898 	ucontext = rdma_udata_to_drv_context(udata, struct c4iw_ucontext,
899 	    ibucontext);
900 	destroy_cq(&chp->rhp->rdev, &chp->cq,
901 		   ucontext ? &ucontext->uctx : &chp->cq.rdev->uctx);
902 }
903 
904 int c4iw_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
905 		   struct ib_udata *udata)
906 {
907 	struct ib_device *ibdev = ibcq->device;
908 	int entries = attr->cqe;
909 	int vector = attr->comp_vector;
910 	struct c4iw_dev *rhp;
911 	struct c4iw_cq *chp = to_c4iw_cq(ibcq);
912 	struct c4iw_create_cq_resp uresp;
913 	struct c4iw_ucontext *ucontext = NULL;
914 	int ret;
915 	size_t memsize, hwentries;
916 	struct c4iw_mm_entry *mm, *mm2;
917 
918 	CTR3(KTR_IW_CXGBE, "%s ib_dev %p entries %d", __func__, ibdev, entries);
919 	if (attr->flags)
920 		return -EINVAL;
921 
922 	rhp = to_c4iw_dev(ibdev);
923 
924 	ucontext = rdma_udata_to_drv_context(udata, struct c4iw_ucontext,
925 	    ibucontext);
926 
927 	/* account for the status page. */
928 	entries++;
929 
930 	/* IQ needs one extra entry to differentiate full vs empty. */
931 	entries++;
932 
933 	/*
934 	 * entries must be multiple of 16 for HW.
935 	 */
936 	entries = roundup(entries, 16);
937 
938 	/*
939 	 * Make actual HW queue 2x to avoid cdix_inc overflows.
940 	 */
941 	hwentries = min(entries * 2, rhp->rdev.hw_queue.t4_max_iq_size);
942 
943 	/*
944 	 * Make HW queue at least 64 entries so GTS updates aren't too
945 	 * frequent.
946 	 */
947 	if (hwentries < 64)
948 		hwentries = 64;
949 
950 	memsize = hwentries * sizeof *chp->cq.queue;
951 
952 	/*
953 	 * memsize must be a multiple of the page size if its a user cq.
954 	 */
955 	if (ucontext)
956 		memsize = roundup(memsize, PAGE_SIZE);
957 	chp->cq.size = hwentries;
958 	chp->cq.memsize = memsize;
959 	chp->cq.vector = vector;
960 
961 	ret = create_cq(&rhp->rdev, &chp->cq,
962 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
963 	if (ret)
964 		goto err1;
965 
966 	chp->rhp = rhp;
967 	chp->cq.size--;				/* status page */
968 	chp->ibcq.cqe = entries - 2;
969 	spin_lock_init(&chp->lock);
970 	spin_lock_init(&chp->comp_handler_lock);
971 	atomic_set(&chp->refcnt, 1);
972 	init_waitqueue_head(&chp->wait);
973 	ret = insert_handle(rhp, &rhp->cqidr, chp, chp->cq.cqid);
974 	if (ret)
975 		goto err2;
976 
977 	if (ucontext) {
978 		ret = -ENOMEM;
979 		mm = kmalloc(sizeof *mm, GFP_KERNEL);
980 		if (!mm)
981 			goto err3;
982 		mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
983 		if (!mm2)
984 			goto err4;
985 
986 		memset(&uresp, 0, sizeof(uresp));
987 		uresp.qid_mask = rhp->rdev.cqmask;
988 		uresp.cqid = chp->cq.cqid;
989 		uresp.size = chp->cq.size;
990 		uresp.memsize = chp->cq.memsize;
991 		spin_lock(&ucontext->mmap_lock);
992 		uresp.key = ucontext->key;
993 		ucontext->key += PAGE_SIZE;
994 		uresp.gts_key = ucontext->key;
995 		ucontext->key += PAGE_SIZE;
996 		spin_unlock(&ucontext->mmap_lock);
997 		ret = ib_copy_to_udata(udata, &uresp,
998 					sizeof(uresp) - sizeof(uresp.reserved));
999 		if (ret)
1000 			goto err5;
1001 
1002 		mm->key = uresp.key;
1003 		mm->addr = vtophys(chp->cq.queue);
1004 		mm->len = chp->cq.memsize;
1005 		insert_mmap(ucontext, mm);
1006 
1007 		mm2->key = uresp.gts_key;
1008 		mm2->addr = chp->cq.bar2_pa;
1009 		mm2->len = PAGE_SIZE;
1010 		insert_mmap(ucontext, mm2);
1011 	}
1012 	CTR6(KTR_IW_CXGBE,
1013 	    "%s cqid 0x%0x chp %p size %u memsize %zu, dma_addr 0x%0llx",
1014 	    __func__, chp->cq.cqid, chp, chp->cq.size, chp->cq.memsize,
1015 	    (unsigned long long) chp->cq.dma_addr);
1016 	return 0;
1017 err5:
1018 	kfree(mm2);
1019 err4:
1020 	kfree(mm);
1021 err3:
1022 	remove_handle(rhp, &rhp->cqidr, chp->cq.cqid);
1023 err2:
1024 	destroy_cq(&chp->rhp->rdev, &chp->cq,
1025 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1026 err1:
1027 	return ret;
1028 }
1029 
1030 int c4iw_resize_cq(struct ib_cq *cq, int cqe, struct ib_udata *udata)
1031 {
1032 	return -ENOSYS;
1033 }
1034 
1035 int c4iw_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
1036 {
1037 	struct c4iw_cq *chp;
1038 	int ret = 0;
1039 	unsigned long flag;
1040 
1041 	chp = to_c4iw_cq(ibcq);
1042 	if (__predict_false(c4iw_stopped(chp->cq.rdev)))
1043 		return -EIO;
1044 	spin_lock_irqsave(&chp->lock, flag);
1045 	t4_arm_cq(&chp->cq,
1046 		  (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED);
1047 	if (flags & IB_CQ_REPORT_MISSED_EVENTS)
1048 		ret = t4_cq_notempty(&chp->cq);
1049 	spin_unlock_irqrestore(&chp->lock, flag);
1050 	return ret;
1051 }
1052 #endif
1053