xref: /freebsd/sys/dev/cxgbe/iw_cxgbe/cq.c (revision 06c3fb2749bda94cb5201f81ffdb8fa6c3161b2e)
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 	cq->cqid = c4iw_get_cqid(rdev, uctx);
110 	if (!cq->cqid) {
111 		ret = -ENOMEM;
112 		goto err1;
113 	}
114 
115 	if (!user) {
116 		cq->sw_queue = kzalloc(cq->memsize, GFP_KERNEL);
117 		if (!cq->sw_queue) {
118 			ret = -ENOMEM;
119 			goto err2;
120 		}
121 	}
122 	cq->queue = dma_alloc_coherent(rhp->ibdev.dma_device, cq->memsize,
123 				       &cq->dma_addr, GFP_KERNEL);
124 	if (!cq->queue) {
125 		ret = -ENOMEM;
126 		goto err3;
127 	}
128 	dma_unmap_addr_set(cq, mapping, cq->dma_addr);
129 	memset(cq->queue, 0, cq->memsize);
130 
131 	/* build fw_ri_res_wr */
132 	wr_len = sizeof *res_wr + sizeof *res;
133 
134 	wr = alloc_wrqe(wr_len, &sc->sge.ctrlq[0]);
135 	if (wr == NULL)
136 		return (0);
137         res_wr = wrtod(wr);
138 
139 	memset(res_wr, 0, wr_len);
140 	res_wr->op_nres = cpu_to_be32(
141 			V_FW_WR_OP(FW_RI_RES_WR) |
142 			V_FW_RI_RES_WR_NRES(1) |
143 			F_FW_WR_COMPL);
144 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
145 	res_wr->cookie = (unsigned long) &wr_wait;
146 	res = res_wr->res;
147 	res->u.cq.restype = FW_RI_RES_TYPE_CQ;
148 	res->u.cq.op = FW_RI_RES_OP_WRITE;
149 	res->u.cq.iqid = cpu_to_be32(cq->cqid);
150 	//Fixme: Always use first queue id for IQANDSTINDEX. Linux does the same.
151 	res->u.cq.iqandst_to_iqandstindex = cpu_to_be32(
152 			V_FW_RI_RES_WR_IQANUS(0) |
153 			V_FW_RI_RES_WR_IQANUD(1) |
154 			F_FW_RI_RES_WR_IQANDST |
155 			V_FW_RI_RES_WR_IQANDSTINDEX(sc->sge.ofld_rxq[0].iq.abs_id));
156 	res->u.cq.iqdroprss_to_iqesize = cpu_to_be16(
157 			F_FW_RI_RES_WR_IQDROPRSS |
158 			V_FW_RI_RES_WR_IQPCIECH(2) |
159 			V_FW_RI_RES_WR_IQINTCNTTHRESH(0) |
160 			F_FW_RI_RES_WR_IQO |
161 			V_FW_RI_RES_WR_IQESIZE(1));
162 	res->u.cq.iqsize = cpu_to_be16(cq->size);
163 	res->u.cq.iqaddr = cpu_to_be64(cq->dma_addr);
164 
165 	c4iw_init_wr_wait(&wr_wait);
166 
167 	t4_wrq_tx(sc, wr);
168 
169 	CTR2(KTR_IW_CXGBE, "%s wait_event wr_wait %p", __func__, &wr_wait);
170 	ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, NULL, __func__);
171 	if (ret)
172 		goto err4;
173 
174 	cq->gen = 1;
175 	cq->rdev = rdev;
176 
177 	/* Determine the BAR2 queue offset and qid. */
178 	t4_bar2_sge_qregs(rdev->adap, cq->cqid, T4_BAR2_QTYPE_INGRESS, user,
179 			&cq_bar2_qoffset, &cq->bar2_qid);
180 
181 	/* If user mapping then compute the page-aligned physical
182 	 * address for mapping.
183 	 */
184 	if (user)
185 		cq->bar2_pa = (rdev->bar2_pa + cq_bar2_qoffset) & PAGE_MASK;
186 	else
187 		cq->bar2_va = (void __iomem *)((u64)rdev->bar2_kva +
188 			cq_bar2_qoffset);
189 
190 	return 0;
191 err4:
192 	dma_free_coherent(rhp->ibdev.dma_device, cq->memsize, cq->queue,
193 			  dma_unmap_addr(cq, mapping));
194 err3:
195 	kfree(cq->sw_queue);
196 err2:
197 	c4iw_put_cqid(rdev, cq->cqid, uctx);
198 err1:
199 	return ret;
200 }
201 
202 static void insert_recv_cqe(struct t4_wq *wq, struct t4_cq *cq)
203 {
204 	struct t4_cqe cqe;
205 
206 	CTR5(KTR_IW_CXGBE, "%s wq %p cq %p sw_cidx %u sw_pidx %u", __func__, wq,
207 	    cq, cq->sw_cidx, cq->sw_pidx);
208 	memset(&cqe, 0, sizeof(cqe));
209 	cqe.header = cpu_to_be32(V_CQE_STATUS(T4_ERR_SWFLUSH) |
210 				 V_CQE_OPCODE(FW_RI_SEND) |
211 				 V_CQE_TYPE(0) |
212 				 V_CQE_SWCQE(1) |
213 				 V_CQE_QPID(wq->sq.qid));
214 	cqe.bits_type_ts = cpu_to_be64(V_CQE_GENBIT((u64)cq->gen));
215 	cq->sw_queue[cq->sw_pidx] = cqe;
216 	t4_swcq_produce(cq);
217 }
218 
219 int c4iw_flush_rq(struct t4_wq *wq, struct t4_cq *cq, int count)
220 {
221 	int flushed = 0;
222 	int in_use = wq->rq.in_use - count;
223 
224 	BUG_ON(in_use < 0);
225 	CTR5(KTR_IW_CXGBE, "%s wq %p cq %p rq.in_use %u skip count %u",
226 	    __func__, wq, cq, wq->rq.in_use, count);
227 	while (in_use--) {
228 		insert_recv_cqe(wq, cq);
229 		flushed++;
230 	}
231 	return flushed;
232 }
233 
234 static void insert_sq_cqe(struct t4_wq *wq, struct t4_cq *cq,
235 			  struct t4_swsqe *swcqe)
236 {
237 	struct t4_cqe cqe;
238 
239 	CTR5(KTR_IW_CXGBE, "%s wq %p cq %p sw_cidx %u sw_pidx %u", __func__, wq,
240 	    cq, cq->sw_cidx, cq->sw_pidx);
241 	memset(&cqe, 0, sizeof(cqe));
242 	cqe.header = cpu_to_be32(V_CQE_STATUS(T4_ERR_SWFLUSH) |
243 				 V_CQE_OPCODE(swcqe->opcode) |
244 				 V_CQE_TYPE(1) |
245 				 V_CQE_SWCQE(1) |
246 				 V_CQE_QPID(wq->sq.qid));
247 	CQE_WRID_SQ_IDX(&cqe) = swcqe->idx;
248 	cqe.bits_type_ts = cpu_to_be64(V_CQE_GENBIT((u64)cq->gen));
249 	cq->sw_queue[cq->sw_pidx] = cqe;
250 	t4_swcq_produce(cq);
251 }
252 
253 static void advance_oldest_read(struct t4_wq *wq);
254 
255 int c4iw_flush_sq(struct c4iw_qp *qhp)
256 {
257 	int flushed = 0;
258 	struct t4_wq *wq = &qhp->wq;
259 	struct c4iw_cq *chp = to_c4iw_cq(qhp->ibqp.send_cq);
260 	struct t4_cq *cq = &chp->cq;
261 	int idx;
262 	struct t4_swsqe *swsqe;
263 
264 	if (wq->sq.flush_cidx == -1)
265 		wq->sq.flush_cidx = wq->sq.cidx;
266 	idx = wq->sq.flush_cidx;
267 	BUG_ON(idx >= wq->sq.size);
268 	while (idx != wq->sq.pidx) {
269 		swsqe = &wq->sq.sw_sq[idx];
270 		BUG_ON(swsqe->flushed);
271 		swsqe->flushed = 1;
272 		insert_sq_cqe(wq, cq, swsqe);
273 		if (wq->sq.oldest_read == swsqe) {
274 			BUG_ON(swsqe->opcode != FW_RI_READ_REQ);
275 			advance_oldest_read(wq);
276 		}
277 		flushed++;
278 		if (++idx == wq->sq.size)
279 			idx = 0;
280 	}
281 	wq->sq.flush_cidx += flushed;
282 	if (wq->sq.flush_cidx >= wq->sq.size)
283 		wq->sq.flush_cidx -= wq->sq.size;
284 	return flushed;
285 }
286 
287 static void flush_completed_wrs(struct t4_wq *wq, struct t4_cq *cq)
288 {
289 	struct t4_swsqe *swsqe;
290 	int cidx;
291 
292 	if (wq->sq.flush_cidx == -1)
293 		wq->sq.flush_cidx = wq->sq.cidx;
294 	cidx = wq->sq.flush_cidx;
295 	BUG_ON(cidx > wq->sq.size);
296 
297 	while (cidx != wq->sq.pidx) {
298 		swsqe = &wq->sq.sw_sq[cidx];
299 		if (!swsqe->signaled) {
300 			if (++cidx == wq->sq.size)
301 				cidx = 0;
302 		} else if (swsqe->complete) {
303 
304 			BUG_ON(swsqe->flushed);
305 
306 			/*
307 			 * Insert this completed cqe into the swcq.
308 			 */
309 			CTR3(KTR_IW_CXGBE,
310 				"%s moving cqe into swcq sq idx %u cq idx %u\n",
311 				__func__, cidx, cq->sw_pidx);
312 			swsqe->cqe.header |= htonl(V_CQE_SWCQE(1));
313 			cq->sw_queue[cq->sw_pidx] = swsqe->cqe;
314 			t4_swcq_produce(cq);
315 			swsqe->flushed = 1;
316 			if (++cidx == wq->sq.size)
317 				cidx = 0;
318 			wq->sq.flush_cidx = cidx;
319 		} else
320 			break;
321 	}
322 }
323 
324 static void create_read_req_cqe(struct t4_wq *wq, struct t4_cqe *hw_cqe,
325 		struct t4_cqe *read_cqe)
326 {
327 	read_cqe->u.scqe.cidx = wq->sq.oldest_read->idx;
328 	read_cqe->len = htonl(wq->sq.oldest_read->read_len);
329 	read_cqe->header = htonl(V_CQE_QPID(CQE_QPID(hw_cqe)) |
330 			V_CQE_SWCQE(SW_CQE(hw_cqe)) |
331 			V_CQE_OPCODE(FW_RI_READ_REQ) |
332 			V_CQE_TYPE(1));
333 	read_cqe->bits_type_ts = hw_cqe->bits_type_ts;
334 }
335 
336 static void advance_oldest_read(struct t4_wq *wq)
337 {
338 
339 	u32 rptr = wq->sq.oldest_read - wq->sq.sw_sq + 1;
340 
341 	if (rptr == wq->sq.size)
342 		rptr = 0;
343 	while (rptr != wq->sq.pidx) {
344 		wq->sq.oldest_read = &wq->sq.sw_sq[rptr];
345 
346 		if (wq->sq.oldest_read->opcode == FW_RI_READ_REQ)
347 			return;
348 		if (++rptr == wq->sq.size)
349 			rptr = 0;
350 	}
351 	wq->sq.oldest_read = NULL;
352 }
353 
354 /*
355  * Move all CQEs from the HWCQ into the SWCQ.
356  * Deal with out-of-order and/or completions that complete
357  * prior unsignalled WRs.
358  */
359 void c4iw_flush_hw_cq(struct c4iw_cq *chp)
360 {
361 	struct t4_cqe *hw_cqe, *swcqe, read_cqe;
362 	struct c4iw_qp *qhp;
363 	struct t4_swsqe *swsqe;
364 	int ret;
365 
366 	CTR3(KTR_IW_CXGBE, "%s cq %p cqid 0x%x", __func__, &chp->cq,
367 			chp->cq.cqid);
368 	ret = t4_next_hw_cqe(&chp->cq, &hw_cqe);
369 
370 	/*
371 	 * This logic is similar to poll_cq(), but not quite the same
372 	 * unfortunately.  Need to move pertinent HW CQEs to the SW CQ but
373 	 * also do any translation magic that poll_cq() normally does.
374 	 */
375 	while (!ret) {
376 		qhp = get_qhp(chp->rhp, CQE_QPID(hw_cqe));
377 
378 		/*
379 		 * drop CQEs with no associated QP
380 		 */
381 		if (qhp == NULL)
382 			goto next_cqe;
383 
384 		if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE)
385 			goto next_cqe;
386 
387 		if (CQE_OPCODE(hw_cqe) == FW_RI_READ_RESP) {
388 
389 			/* If we have reached here because of async
390 			 * event or other error, and have egress error
391 			 * then drop
392 			 */
393 			if (CQE_TYPE(hw_cqe) == 1)
394 				goto next_cqe;
395 
396 			/* drop peer2peer RTR reads.
397 			 */
398 			if (CQE_WRID_STAG(hw_cqe) == 1)
399 				goto next_cqe;
400 
401 			/*
402 			 * Eat completions for unsignaled read WRs.
403 			 */
404 			if (!qhp->wq.sq.oldest_read->signaled) {
405 				advance_oldest_read(&qhp->wq);
406 				goto next_cqe;
407 			}
408 
409 			/*
410 			 * Don't write to the HWCQ, create a new read req CQE
411 			 * in local memory and move it into the swcq.
412 			 */
413 			create_read_req_cqe(&qhp->wq, hw_cqe, &read_cqe);
414 			hw_cqe = &read_cqe;
415 			advance_oldest_read(&qhp->wq);
416 		}
417 
418 		/* if its a SQ completion, then do the magic to move all the
419 		 * unsignaled and now in-order completions into the swcq.
420 		 */
421 		if (SQ_TYPE(hw_cqe)) {
422 			swsqe = &qhp->wq.sq.sw_sq[CQE_WRID_SQ_IDX(hw_cqe)];
423 			swsqe->cqe = *hw_cqe;
424 			swsqe->complete = 1;
425 			flush_completed_wrs(&qhp->wq, &chp->cq);
426 		} else {
427 			swcqe = &chp->cq.sw_queue[chp->cq.sw_pidx];
428 			*swcqe = *hw_cqe;
429 			swcqe->header |= cpu_to_be32(V_CQE_SWCQE(1));
430 			t4_swcq_produce(&chp->cq);
431 		}
432 next_cqe:
433 		t4_hwcq_consume(&chp->cq);
434 		ret = t4_next_hw_cqe(&chp->cq, &hw_cqe);
435 	}
436 }
437 
438 static int cqe_completes_wr(struct t4_cqe *cqe, struct t4_wq *wq)
439 {
440 	if (CQE_OPCODE(cqe) == FW_RI_TERMINATE)
441 		return 0;
442 
443 	if ((CQE_OPCODE(cqe) == FW_RI_RDMA_WRITE) && RQ_TYPE(cqe))
444 		return 0;
445 
446 	if ((CQE_OPCODE(cqe) == FW_RI_READ_RESP) && SQ_TYPE(cqe))
447 		return 0;
448 
449 	if (CQE_SEND_OPCODE(cqe) && RQ_TYPE(cqe) && t4_rq_empty(wq))
450 		return 0;
451 	return 1;
452 }
453 
454 void c4iw_count_rcqes(struct t4_cq *cq, struct t4_wq *wq, int *count)
455 {
456 	struct t4_cqe *cqe;
457 	u32 ptr;
458 
459 	*count = 0;
460 	CTR2(KTR_IW_CXGBE, "%s count zero %d", __func__, *count);
461 	ptr = cq->sw_cidx;
462 	while (ptr != cq->sw_pidx) {
463 		cqe = &cq->sw_queue[ptr];
464 		if (RQ_TYPE(cqe) && (CQE_OPCODE(cqe) != FW_RI_READ_RESP) &&
465 		    (CQE_QPID(cqe) == wq->sq.qid) && cqe_completes_wr(cqe, wq))
466 			(*count)++;
467 		if (++ptr == cq->size)
468 			ptr = 0;
469 	}
470 	CTR3(KTR_IW_CXGBE, "%s cq %p count %d", __func__, cq, *count);
471 }
472 
473 /*
474  * poll_cq
475  *
476  * Caller must:
477  *     check the validity of the first CQE,
478  *     supply the wq assicated with the qpid.
479  *
480  * credit: cq credit to return to sge.
481  * cqe_flushed: 1 iff the CQE is flushed.
482  * cqe: copy of the polled CQE.
483  *
484  * return value:
485  *    0		    CQE returned ok.
486  *    -EAGAIN       CQE skipped, try again.
487  *    -EOVERFLOW    CQ overflow detected.
488  */
489 static int poll_cq(struct t4_wq *wq, struct t4_cq *cq, struct t4_cqe *cqe,
490 		   u8 *cqe_flushed, u64 *cookie, u32 *credit)
491 {
492 	int ret = 0;
493 	struct t4_cqe *hw_cqe, read_cqe;
494 
495 	*cqe_flushed = 0;
496 	*credit = 0;
497 	ret = t4_next_cqe(cq, &hw_cqe);
498 	if (ret)
499 		return ret;
500 
501 	CTR6(KTR_IW_CXGBE,
502 	    "%s CQE OVF %u qpid 0x%0x genbit %u type %u status 0x%0x", __func__,
503 	    CQE_OVFBIT(hw_cqe), CQE_QPID(hw_cqe), CQE_GENBIT(hw_cqe),
504 	    CQE_TYPE(hw_cqe), CQE_STATUS(hw_cqe));
505 	CTR5(KTR_IW_CXGBE,
506 	    "%s opcode 0x%0x len 0x%0x wrid_hi_stag 0x%x wrid_low_msn 0x%x",
507 	    __func__, CQE_OPCODE(hw_cqe), CQE_LEN(hw_cqe), CQE_WRID_HI(hw_cqe),
508 	    CQE_WRID_LOW(hw_cqe));
509 
510 	/*
511 	 * skip cqe's not affiliated with a QP.
512 	 */
513 	if (wq == NULL) {
514 		ret = -EAGAIN;
515 		goto skip_cqe;
516 	}
517 
518 	/*
519 	* skip hw cqe's if the wq is flushed.
520 	*/
521 	if (wq->flushed && !SW_CQE(hw_cqe)) {
522 		ret = -EAGAIN;
523 		goto skip_cqe;
524 	}
525 
526 	/*
527 	 * skip TERMINATE cqes...
528 	 */
529 	if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE) {
530 		ret = -EAGAIN;
531 		goto skip_cqe;
532 	}
533 
534 	/*
535 	 * Special cqe for drain WR completions...
536 	 */
537 	if (CQE_OPCODE(hw_cqe) == C4IW_DRAIN_OPCODE) {
538 		*cookie = CQE_DRAIN_COOKIE(hw_cqe);
539 		*cqe = *hw_cqe;
540 		goto skip_cqe;
541 	}
542 
543 	/*
544 	 * Gotta tweak READ completions:
545 	 *	1) the cqe doesn't contain the sq_wptr from the wr.
546 	 *	2) opcode not reflected from the wr.
547 	 *	3) read_len not reflected from the wr.
548 	 *	4) cq_type is RQ_TYPE not SQ_TYPE.
549 	 */
550 	if (RQ_TYPE(hw_cqe) && (CQE_OPCODE(hw_cqe) == FW_RI_READ_RESP)) {
551 
552 		/* If we have reached here because of async
553 		 * event or other error, and have egress error
554 		 * then drop
555 		 */
556 		if (CQE_TYPE(hw_cqe) == 1) {
557 			if (CQE_STATUS(hw_cqe))
558 				t4_set_wq_in_error(wq);
559 			ret = -EAGAIN;
560 			goto skip_cqe;
561 		}
562 
563 		/* If this is an unsolicited read response, then the read
564 		 * was generated by the kernel driver as part of peer-2-peer
565 		 * connection setup.  So ignore the completion.
566 		 */
567 		if (CQE_WRID_STAG(hw_cqe) == 1) {
568 			if (CQE_STATUS(hw_cqe))
569 				t4_set_wq_in_error(wq);
570 			ret = -EAGAIN;
571 			goto skip_cqe;
572 		}
573 
574 		/*
575 		 * Eat completions for unsignaled read WRs.
576 		 */
577 		if (!wq->sq.oldest_read->signaled) {
578 			advance_oldest_read(wq);
579 			ret = -EAGAIN;
580 			goto skip_cqe;
581 		}
582 
583 		/*
584 		 * Don't write to the HWCQ, so create a new read req CQE
585 		 * in local memory.
586 		 */
587 		create_read_req_cqe(wq, hw_cqe, &read_cqe);
588 		hw_cqe = &read_cqe;
589 		advance_oldest_read(wq);
590 	}
591 
592 	if (CQE_STATUS(hw_cqe) || t4_wq_in_error(wq)) {
593 		*cqe_flushed = (CQE_STATUS(hw_cqe) == T4_ERR_SWFLUSH);
594 		t4_set_wq_in_error(wq);
595 	}
596 
597 	/*
598 	 * RECV completion.
599 	 */
600 	if (RQ_TYPE(hw_cqe)) {
601 
602 		/*
603 		 * HW only validates 4 bits of MSN.  So we must validate that
604 		 * the MSN in the SEND is the next expected MSN.  If its not,
605 		 * then we complete this with T4_ERR_MSN and mark the wq in
606 		 * error.
607 		 */
608 
609 		if (t4_rq_empty(wq)) {
610 			t4_set_wq_in_error(wq);
611 			ret = -EAGAIN;
612 			goto skip_cqe;
613 		}
614 		if (unlikely((CQE_WRID_MSN(hw_cqe) != (wq->rq.msn)))) {
615 			t4_set_wq_in_error(wq);
616 			hw_cqe->header |= htonl(V_CQE_STATUS(T4_ERR_MSN));
617 			goto proc_cqe;
618 		}
619 		goto proc_cqe;
620 	}
621 
622 	/*
623 	 * If we get here its a send completion.
624 	 *
625 	 * Handle out of order completion. These get stuffed
626 	 * in the SW SQ. Then the SW SQ is walked to move any
627 	 * now in-order completions into the SW CQ.  This handles
628 	 * 2 cases:
629 	 *	1) reaping unsignaled WRs when the first subsequent
630 	 *	   signaled WR is completed.
631 	 *	2) out of order read completions.
632 	 */
633 	if (!SW_CQE(hw_cqe) && (CQE_WRID_SQ_IDX(hw_cqe) != wq->sq.cidx)) {
634 		struct t4_swsqe *swsqe;
635 
636 		CTR2(KTR_IW_CXGBE,
637 		    "%s out of order completion going in sw_sq at idx %u",
638 		    __func__, CQE_WRID_SQ_IDX(hw_cqe));
639 		swsqe = &wq->sq.sw_sq[CQE_WRID_SQ_IDX(hw_cqe)];
640 		swsqe->cqe = *hw_cqe;
641 		swsqe->complete = 1;
642 		ret = -EAGAIN;
643 		goto flush_wq;
644 	}
645 
646 proc_cqe:
647 	*cqe = *hw_cqe;
648 
649 	/*
650 	 * Reap the associated WR(s) that are freed up with this
651 	 * completion.
652 	 */
653 	if (SQ_TYPE(hw_cqe)) {
654 		int idx = CQE_WRID_SQ_IDX(hw_cqe);
655 		BUG_ON(idx >= wq->sq.size);
656 
657 		/*
658 		* Account for any unsignaled completions completed by
659 		* this signaled completion.  In this case, cidx points
660 		* to the first unsignaled one, and idx points to the
661 		* signaled one.  So adjust in_use based on this delta.
662 		* if this is not completing any unsigned wrs, then the
663 		* delta will be 0. Handle wrapping also!
664 		*/
665 		if (idx < wq->sq.cidx)
666 			wq->sq.in_use -= wq->sq.size + idx - wq->sq.cidx;
667 		else
668 			wq->sq.in_use -= idx - wq->sq.cidx;
669 		BUG_ON(wq->sq.in_use <= 0 && wq->sq.in_use >= wq->sq.size);
670 
671 		wq->sq.cidx = (uint16_t)idx;
672 		CTR2(KTR_IW_CXGBE, "%s completing sq idx %u",
673 				__func__, wq->sq.cidx);
674 		*cookie = wq->sq.sw_sq[wq->sq.cidx].wr_id;
675 		t4_sq_consume(wq);
676 	} else {
677 		CTR2(KTR_IW_CXGBE, "%s completing rq idx %u",
678 		     __func__, wq->rq.cidx);
679 		*cookie = wq->rq.sw_rq[wq->rq.cidx].wr_id;
680 		BUG_ON(t4_rq_empty(wq));
681 		t4_rq_consume(wq);
682 		goto skip_cqe;
683 	}
684 
685 flush_wq:
686 	/*
687 	 * Flush any completed cqes that are now in-order.
688 	 */
689 	flush_completed_wrs(wq, cq);
690 
691 skip_cqe:
692 	if (SW_CQE(hw_cqe)) {
693 		CTR4(KTR_IW_CXGBE, "%s cq %p cqid 0x%x skip sw cqe cidx %u",
694 		     __func__, cq, cq->cqid, cq->sw_cidx);
695 		t4_swcq_consume(cq);
696 	} else {
697 		CTR4(KTR_IW_CXGBE, "%s cq %p cqid 0x%x skip hw cqe cidx %u",
698 		     __func__, cq, cq->cqid, cq->cidx);
699 		t4_hwcq_consume(cq);
700 	}
701 	return ret;
702 }
703 
704 /*
705  * Get one cq entry from c4iw and map it to openib.
706  *
707  * Returns:
708  *	0			cqe returned
709  *	-ENODATA		EMPTY;
710  *	-EAGAIN			caller must try again
711  *	any other -errno	fatal error
712  */
713 static int c4iw_poll_cq_one(struct c4iw_cq *chp, struct ib_wc *wc)
714 {
715 	struct c4iw_qp *qhp = NULL;
716 	struct t4_cqe cqe = {0, 0}, *rd_cqe;
717 	struct t4_wq *wq;
718 	u32 credit = 0;
719 	u8 cqe_flushed;
720 	u64 cookie = 0;
721 	int ret;
722 
723 	ret = t4_next_cqe(&chp->cq, &rd_cqe);
724 
725 	if (ret)
726 		return ret;
727 
728 	qhp = get_qhp(chp->rhp, CQE_QPID(rd_cqe));
729 	if (!qhp)
730 		wq = NULL;
731 	else {
732 		spin_lock(&qhp->lock);
733 		wq = &(qhp->wq);
734 	}
735 	ret = poll_cq(wq, &(chp->cq), &cqe, &cqe_flushed, &cookie, &credit);
736 	if (ret)
737 		goto out;
738 
739 	wc->wr_id = cookie;
740 	wc->qp = &qhp->ibqp;
741 	wc->vendor_err = CQE_STATUS(&cqe);
742 	wc->wc_flags = 0;
743 
744 	CTR5(KTR_IW_CXGBE, "%s qpid 0x%x type %d opcode %d status 0x%x",
745 	    __func__, CQE_QPID(&cqe), CQE_TYPE(&cqe), CQE_OPCODE(&cqe),
746 	    CQE_STATUS(&cqe));
747 	CTR5(KTR_IW_CXGBE, "%s len %u wrid hi 0x%x lo 0x%x cookie 0x%llx",
748 	    __func__, CQE_LEN(&cqe), CQE_WRID_HI(&cqe), CQE_WRID_LOW(&cqe),
749 	    (unsigned long long)cookie);
750 
751 	if (CQE_TYPE(&cqe) == 0) {
752 		if (!CQE_STATUS(&cqe))
753 			wc->byte_len = CQE_LEN(&cqe);
754 		else
755 			wc->byte_len = 0;
756 		wc->opcode = IB_WC_RECV;
757 		if (CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_INV ||
758 		    CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_SE_INV) {
759 			wc->ex.invalidate_rkey = CQE_WRID_STAG(&cqe);
760 			wc->wc_flags |= IB_WC_WITH_INVALIDATE;
761 			c4iw_invalidate_mr(qhp->rhp, wc->ex.invalidate_rkey);
762 		}
763 	} else {
764 		switch (CQE_OPCODE(&cqe)) {
765 		case FW_RI_RDMA_WRITE:
766 			wc->opcode = IB_WC_RDMA_WRITE;
767 			break;
768 		case FW_RI_READ_REQ:
769 			wc->opcode = IB_WC_RDMA_READ;
770 			wc->byte_len = CQE_LEN(&cqe);
771 			break;
772 		case FW_RI_SEND_WITH_INV:
773 		case FW_RI_SEND_WITH_SE_INV:
774 			wc->opcode = IB_WC_SEND;
775 			wc->wc_flags |= IB_WC_WITH_INVALIDATE;
776 			break;
777 		case FW_RI_SEND:
778 		case FW_RI_SEND_WITH_SE:
779 			wc->opcode = IB_WC_SEND;
780 			break;
781 		case FW_RI_LOCAL_INV:
782 			wc->opcode = IB_WC_LOCAL_INV;
783 			break;
784 		case FW_RI_FAST_REGISTER:
785 			wc->opcode = IB_WC_REG_MR;
786 
787 			/* Invalidate the MR if the fastreg failed */
788 			if (CQE_STATUS(&cqe) != T4_ERR_SUCCESS)
789 				c4iw_invalidate_mr(qhp->rhp,
790 						   CQE_WRID_FR_STAG(&cqe));
791 			break;
792 		case C4IW_DRAIN_OPCODE:
793 			wc->opcode = IB_WC_SEND;
794 			break;
795 		default:
796 			printf("Unexpected opcode %d "
797 			       "in the CQE received for QPID = 0x%0x\n",
798 			       CQE_OPCODE(&cqe), CQE_QPID(&cqe));
799 			ret = -EINVAL;
800 			goto out;
801 		}
802 	}
803 
804 	if (cqe_flushed)
805 		wc->status = IB_WC_WR_FLUSH_ERR;
806 	else {
807 
808 		switch (CQE_STATUS(&cqe)) {
809 		case T4_ERR_SUCCESS:
810 			wc->status = IB_WC_SUCCESS;
811 			break;
812 		case T4_ERR_STAG:
813 			wc->status = IB_WC_LOC_ACCESS_ERR;
814 			break;
815 		case T4_ERR_PDID:
816 			wc->status = IB_WC_LOC_PROT_ERR;
817 			break;
818 		case T4_ERR_QPID:
819 		case T4_ERR_ACCESS:
820 			wc->status = IB_WC_LOC_ACCESS_ERR;
821 			break;
822 		case T4_ERR_WRAP:
823 			wc->status = IB_WC_GENERAL_ERR;
824 			break;
825 		case T4_ERR_BOUND:
826 			wc->status = IB_WC_LOC_LEN_ERR;
827 			break;
828 		case T4_ERR_INVALIDATE_SHARED_MR:
829 		case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
830 			wc->status = IB_WC_MW_BIND_ERR;
831 			break;
832 		case T4_ERR_CRC:
833 		case T4_ERR_MARKER:
834 		case T4_ERR_PDU_LEN_ERR:
835 		case T4_ERR_OUT_OF_RQE:
836 		case T4_ERR_DDP_VERSION:
837 		case T4_ERR_RDMA_VERSION:
838 		case T4_ERR_DDP_QUEUE_NUM:
839 		case T4_ERR_MSN:
840 		case T4_ERR_TBIT:
841 		case T4_ERR_MO:
842 		case T4_ERR_MSN_RANGE:
843 		case T4_ERR_IRD_OVERFLOW:
844 		case T4_ERR_OPCODE:
845 		case T4_ERR_INTERNAL_ERR:
846 			wc->status = IB_WC_FATAL_ERR;
847 			break;
848 		case T4_ERR_SWFLUSH:
849 			wc->status = IB_WC_WR_FLUSH_ERR;
850 			break;
851 		default:
852 			printf("Unexpected cqe_status 0x%x for QPID = 0x%0x\n",
853 			       CQE_STATUS(&cqe), CQE_QPID(&cqe));
854 			wc->status = IB_WC_FATAL_ERR;
855 		}
856 	}
857 out:
858 	if (wq)
859 		spin_unlock(&qhp->lock);
860 	return ret;
861 }
862 
863 int c4iw_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
864 {
865 	struct c4iw_cq *chp;
866 	unsigned long flags;
867 	int npolled;
868 	int err = 0;
869 
870 	chp = to_c4iw_cq(ibcq);
871 
872 	spin_lock_irqsave(&chp->lock, flags);
873 	for (npolled = 0; npolled < num_entries; ++npolled) {
874 		do {
875 			err = c4iw_poll_cq_one(chp, wc + npolled);
876 		} while (err == -EAGAIN);
877 		if (err)
878 			break;
879 	}
880 	spin_unlock_irqrestore(&chp->lock, flags);
881 	return !err || err == -ENODATA ? npolled : err;
882 }
883 
884 void c4iw_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata)
885 {
886 	struct c4iw_cq *chp;
887 	struct c4iw_ucontext *ucontext;
888 
889 	CTR2(KTR_IW_CXGBE, "%s ib_cq %p", __func__, ib_cq);
890 	chp = to_c4iw_cq(ib_cq);
891 
892 	remove_handle(chp->rhp, &chp->rhp->cqidr, chp->cq.cqid);
893 	atomic_dec(&chp->refcnt);
894 	wait_event(chp->wait, !atomic_read(&chp->refcnt));
895 
896 	ucontext = rdma_udata_to_drv_context(udata, struct c4iw_ucontext,
897 	    ibucontext);
898 	destroy_cq(&chp->rhp->rdev, &chp->cq,
899 		   ucontext ? &ucontext->uctx : &chp->cq.rdev->uctx);
900 }
901 
902 int c4iw_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
903 		   struct ib_udata *udata)
904 {
905 	struct ib_device *ibdev = ibcq->device;
906 	int entries = attr->cqe;
907 	int vector = attr->comp_vector;
908 	struct c4iw_dev *rhp;
909 	struct c4iw_cq *chp = to_c4iw_cq(ibcq);
910 	struct c4iw_create_cq_resp uresp;
911 	struct c4iw_ucontext *ucontext = NULL;
912 	int ret;
913 	size_t memsize, hwentries;
914 	struct c4iw_mm_entry *mm, *mm2;
915 
916 	CTR3(KTR_IW_CXGBE, "%s ib_dev %p entries %d", __func__, ibdev, entries);
917 	if (attr->flags)
918 		return -EINVAL;
919 
920 	rhp = to_c4iw_dev(ibdev);
921 
922 	ucontext = rdma_udata_to_drv_context(udata, struct c4iw_ucontext,
923 	    ibucontext);
924 
925 	/* account for the status page. */
926 	entries++;
927 
928 	/* IQ needs one extra entry to differentiate full vs empty. */
929 	entries++;
930 
931 	/*
932 	 * entries must be multiple of 16 for HW.
933 	 */
934 	entries = roundup(entries, 16);
935 
936 	/*
937 	 * Make actual HW queue 2x to avoid cdix_inc overflows.
938 	 */
939 	hwentries = min(entries * 2, rhp->rdev.hw_queue.t4_max_iq_size);
940 
941 	/*
942 	 * Make HW queue at least 64 entries so GTS updates aren't too
943 	 * frequent.
944 	 */
945 	if (hwentries < 64)
946 		hwentries = 64;
947 
948 	memsize = hwentries * sizeof *chp->cq.queue;
949 
950 	/*
951 	 * memsize must be a multiple of the page size if its a user cq.
952 	 */
953 	if (ucontext)
954 		memsize = roundup(memsize, PAGE_SIZE);
955 	chp->cq.size = hwentries;
956 	chp->cq.memsize = memsize;
957 	chp->cq.vector = vector;
958 
959 	ret = create_cq(&rhp->rdev, &chp->cq,
960 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
961 	if (ret)
962 		goto err1;
963 
964 	chp->rhp = rhp;
965 	chp->cq.size--;				/* status page */
966 	chp->ibcq.cqe = entries - 2;
967 	spin_lock_init(&chp->lock);
968 	spin_lock_init(&chp->comp_handler_lock);
969 	atomic_set(&chp->refcnt, 1);
970 	init_waitqueue_head(&chp->wait);
971 	ret = insert_handle(rhp, &rhp->cqidr, chp, chp->cq.cqid);
972 	if (ret)
973 		goto err2;
974 
975 	if (ucontext) {
976 		ret = -ENOMEM;
977 		mm = kmalloc(sizeof *mm, GFP_KERNEL);
978 		if (!mm)
979 			goto err3;
980 		mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
981 		if (!mm2)
982 			goto err4;
983 
984 		memset(&uresp, 0, sizeof(uresp));
985 		uresp.qid_mask = rhp->rdev.cqmask;
986 		uresp.cqid = chp->cq.cqid;
987 		uresp.size = chp->cq.size;
988 		uresp.memsize = chp->cq.memsize;
989 		spin_lock(&ucontext->mmap_lock);
990 		uresp.key = ucontext->key;
991 		ucontext->key += PAGE_SIZE;
992 		uresp.gts_key = ucontext->key;
993 		ucontext->key += PAGE_SIZE;
994 		spin_unlock(&ucontext->mmap_lock);
995 		ret = ib_copy_to_udata(udata, &uresp,
996 					sizeof(uresp) - sizeof(uresp.reserved));
997 		if (ret)
998 			goto err5;
999 
1000 		mm->key = uresp.key;
1001 		mm->addr = vtophys(chp->cq.queue);
1002 		mm->len = chp->cq.memsize;
1003 		insert_mmap(ucontext, mm);
1004 
1005 		mm2->key = uresp.gts_key;
1006 		mm2->addr = chp->cq.bar2_pa;
1007 		mm2->len = PAGE_SIZE;
1008 		insert_mmap(ucontext, mm2);
1009 	}
1010 	CTR6(KTR_IW_CXGBE,
1011 	    "%s cqid 0x%0x chp %p size %u memsize %zu, dma_addr 0x%0llx",
1012 	    __func__, chp->cq.cqid, chp, chp->cq.size, chp->cq.memsize,
1013 	    (unsigned long long) chp->cq.dma_addr);
1014 	return 0;
1015 err5:
1016 	kfree(mm2);
1017 err4:
1018 	kfree(mm);
1019 err3:
1020 	remove_handle(rhp, &rhp->cqidr, chp->cq.cqid);
1021 err2:
1022 	destroy_cq(&chp->rhp->rdev, &chp->cq,
1023 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1024 err1:
1025 	return ret;
1026 }
1027 
1028 int c4iw_resize_cq(struct ib_cq *cq, int cqe, struct ib_udata *udata)
1029 {
1030 	return -ENOSYS;
1031 }
1032 
1033 int c4iw_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
1034 {
1035 	struct c4iw_cq *chp;
1036 	int ret = 0;
1037 	unsigned long flag;
1038 
1039 	chp = to_c4iw_cq(ibcq);
1040 	spin_lock_irqsave(&chp->lock, flag);
1041 	t4_arm_cq(&chp->cq,
1042 		  (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED);
1043 	if (flags & IB_CQ_REPORT_MISSED_EVENTS)
1044 		ret = t4_cq_notempty(&chp->cq);
1045 	spin_unlock_irqrestore(&chp->lock, flag);
1046 	return ret;
1047 }
1048 #endif
1049