xref: /freebsd/sys/dev/cxgbe/iw_cxgbe/cq.c (revision 86c9d9918f1db7cdd968b60f8902466887bcd9e9)
1 /*
2  * Copyright (c) 2009-2013 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 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
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 fw_ri_res_wr *res_wr;
57 	struct fw_ri_res *res;
58 	int wr_len;
59 	struct c4iw_wr_wait wr_wait;
60 	struct wrqe *wr;
61 
62 	wr_len = sizeof *res_wr + sizeof *res;
63 	wr = alloc_wrqe(wr_len, &sc->sge.mgmtq);
64                 if (wr == NULL)
65                         return (0);
66         res_wr = wrtod(wr);
67 	memset(res_wr, 0, wr_len);
68 	res_wr->op_nres = cpu_to_be32(
69 			V_FW_WR_OP(FW_RI_RES_WR) |
70 			V_FW_RI_RES_WR_NRES(1) |
71 			F_FW_WR_COMPL);
72 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
73 	res_wr->cookie = (unsigned long) &wr_wait;
74 	res = res_wr->res;
75 	res->u.cq.restype = FW_RI_RES_TYPE_CQ;
76 	res->u.cq.op = FW_RI_RES_OP_RESET;
77 	res->u.cq.iqid = cpu_to_be32(cq->cqid);
78 
79 	c4iw_init_wr_wait(&wr_wait);
80 
81 	t4_wrq_tx(sc, wr);
82 
83 	c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, __func__);
84 
85 	kfree(cq->sw_queue);
86 	contigfree(cq->queue, cq->memsize, M_DEVBUF);
87 	c4iw_put_cqid(rdev, cq->cqid, uctx);
88 	return 0;
89 }
90 
91 static int
92 create_cq(struct c4iw_rdev *rdev, struct t4_cq *cq,
93     struct c4iw_dev_ucontext *uctx)
94 {
95 	struct adapter *sc = rdev->adap;
96 	struct fw_ri_res_wr *res_wr;
97 	struct fw_ri_res *res;
98 	int wr_len;
99 	int user = (uctx != &rdev->uctx);
100 	struct c4iw_wr_wait wr_wait;
101 	int ret;
102 	struct wrqe *wr;
103 
104 	cq->cqid = c4iw_get_cqid(rdev, uctx);
105 	if (!cq->cqid) {
106 		ret = -ENOMEM;
107 		goto err1;
108 	}
109 
110 	if (!user) {
111 		cq->sw_queue = kzalloc(cq->memsize, GFP_KERNEL);
112 		if (!cq->sw_queue) {
113 			ret = -ENOMEM;
114 			goto err2;
115 		}
116 	}
117 
118 	cq->queue = contigmalloc(cq->memsize, M_DEVBUF, M_NOWAIT, 0ul, ~0ul,
119 	    PAGE_SIZE, 0);
120         if (cq->queue)
121                 cq->dma_addr = vtophys(cq->queue);
122         else {
123 		ret = -ENOMEM;
124                 goto err3;
125 	}
126 
127 	pci_unmap_addr_set(cq, mapping, cq->dma_addr);
128 	memset(cq->queue, 0, cq->memsize);
129 
130 	/* build fw_ri_res_wr */
131 	wr_len = sizeof *res_wr + sizeof *res;
132 
133 	wr = alloc_wrqe(wr_len, &sc->sge.mgmtq);
134 	if (wr == NULL)
135         	return (0);
136         res_wr = wrtod(wr);
137 
138 	memset(res_wr, 0, wr_len);
139 	res_wr->op_nres = cpu_to_be32(
140 			V_FW_WR_OP(FW_RI_RES_WR) |
141 			V_FW_RI_RES_WR_NRES(1) |
142 			F_FW_WR_COMPL);
143 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
144 	res_wr->cookie = (unsigned long) &wr_wait;
145 	res = res_wr->res;
146 	res->u.cq.restype = FW_RI_RES_TYPE_CQ;
147 	res->u.cq.op = FW_RI_RES_OP_WRITE;
148 	res->u.cq.iqid = cpu_to_be32(cq->cqid);
149 	//Fixme: Always use first queue id for IQANDSTINDEX. Linux does the same.
150 	res->u.cq.iqandst_to_iqandstindex = cpu_to_be32(
151 			V_FW_RI_RES_WR_IQANUS(0) |
152 			V_FW_RI_RES_WR_IQANUD(1) |
153 			F_FW_RI_RES_WR_IQANDST |
154 			V_FW_RI_RES_WR_IQANDSTINDEX(sc->sge.ofld_rxq[0].iq.abs_id));
155 	res->u.cq.iqdroprss_to_iqesize = cpu_to_be16(
156 			F_FW_RI_RES_WR_IQDROPRSS |
157 			V_FW_RI_RES_WR_IQPCIECH(2) |
158 			V_FW_RI_RES_WR_IQINTCNTTHRESH(0) |
159 			F_FW_RI_RES_WR_IQO |
160 			V_FW_RI_RES_WR_IQESIZE(1));
161 	res->u.cq.iqsize = cpu_to_be16(cq->size);
162 	res->u.cq.iqaddr = cpu_to_be64(cq->dma_addr);
163 
164 	c4iw_init_wr_wait(&wr_wait);
165 
166 	t4_wrq_tx(sc, wr);
167 
168 	CTR2(KTR_IW_CXGBE, "%s wait_event wr_wait %p", __func__, &wr_wait);
169 	ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, __func__);
170 	if (ret)
171 		goto err4;
172 
173 	cq->gen = 1;
174 	cq->gts = (void *)((unsigned long)rman_get_virtual(sc->regs_res) +
175 	    MYPF_REG(SGE_PF_GTS));
176 	cq->rdev = rdev;
177 
178 	if (user) {
179 		cq->ugts = (u64)((char*)rman_get_virtual(sc->udbs_res) +
180 		    (cq->cqid << rdev->cqshift));
181 		cq->ugts &= PAGE_MASK;
182 		CTR5(KTR_IW_CXGBE,
183 		    "%s: UGTS %p cqid %x cqshift %d page_mask %x", __func__,
184 		    cq->ugts, cq->cqid, rdev->cqshift, PAGE_MASK);
185 	}
186 	return 0;
187 err4:
188 	contigfree(cq->queue, cq->memsize, M_DEVBUF);
189 err3:
190 	kfree(cq->sw_queue);
191 err2:
192 	c4iw_put_cqid(rdev, cq->cqid, uctx);
193 err1:
194 	return ret;
195 }
196 
197 static void insert_recv_cqe(struct t4_wq *wq, struct t4_cq *cq)
198 {
199 	struct t4_cqe cqe;
200 
201 	CTR5(KTR_IW_CXGBE, "%s wq %p cq %p sw_cidx %u sw_pidx %u", __func__, wq,
202 	    cq, cq->sw_cidx, cq->sw_pidx);
203 	memset(&cqe, 0, sizeof(cqe));
204 	cqe.header = cpu_to_be32(V_CQE_STATUS(T4_ERR_SWFLUSH) |
205 				 V_CQE_OPCODE(FW_RI_SEND) |
206 				 V_CQE_TYPE(0) |
207 				 V_CQE_SWCQE(1) |
208 				 V_CQE_QPID(wq->sq.qid));
209 	cqe.bits_type_ts = cpu_to_be64(V_CQE_GENBIT((u64)cq->gen));
210 	cq->sw_queue[cq->sw_pidx] = cqe;
211 	t4_swcq_produce(cq);
212 }
213 
214 int c4iw_flush_rq(struct t4_wq *wq, struct t4_cq *cq, int count)
215 {
216 	int flushed = 0;
217 	int in_use = wq->rq.in_use - count;
218 
219 	BUG_ON(in_use < 0);
220 	CTR5(KTR_IW_CXGBE, "%s wq %p cq %p rq.in_use %u skip count %u",
221 	    __func__, wq, cq, wq->rq.in_use, count);
222 	while (in_use--) {
223 		insert_recv_cqe(wq, cq);
224 		flushed++;
225 	}
226 	return flushed;
227 }
228 
229 static void insert_sq_cqe(struct t4_wq *wq, struct t4_cq *cq,
230 			  struct t4_swsqe *swcqe)
231 {
232 	struct t4_cqe cqe;
233 
234 	CTR5(KTR_IW_CXGBE, "%s wq %p cq %p sw_cidx %u sw_pidx %u", __func__, wq,
235 	    cq, cq->sw_cidx, cq->sw_pidx);
236 	memset(&cqe, 0, sizeof(cqe));
237 	cqe.header = cpu_to_be32(V_CQE_STATUS(T4_ERR_SWFLUSH) |
238 				 V_CQE_OPCODE(swcqe->opcode) |
239 				 V_CQE_TYPE(1) |
240 				 V_CQE_SWCQE(1) |
241 				 V_CQE_QPID(wq->sq.qid));
242 	CQE_WRID_SQ_IDX(&cqe) = swcqe->idx;
243 	cqe.bits_type_ts = cpu_to_be64(V_CQE_GENBIT((u64)cq->gen));
244 	cq->sw_queue[cq->sw_pidx] = cqe;
245 	t4_swcq_produce(cq);
246 }
247 
248 int c4iw_flush_sq(struct t4_wq *wq, struct t4_cq *cq, int count)
249 {
250 	int flushed = 0;
251 	struct t4_swsqe *swsqe = &wq->sq.sw_sq[wq->sq.cidx + count];
252 	int in_use = wq->sq.in_use - count;
253 
254 	BUG_ON(in_use < 0);
255 	while (in_use--) {
256 		swsqe->signaled = 0;
257 		insert_sq_cqe(wq, cq, swsqe);
258 		swsqe++;
259 		if (swsqe == (wq->sq.sw_sq + wq->sq.size))
260 			swsqe = wq->sq.sw_sq;
261 		flushed++;
262 	}
263 	return flushed;
264 }
265 
266 /*
267  * Move all CQEs from the HWCQ into the SWCQ.
268  */
269 void c4iw_flush_hw_cq(struct t4_cq *cq)
270 {
271 	struct t4_cqe *cqe = NULL, *swcqe;
272 	int ret;
273 
274 	CTR3(KTR_IW_CXGBE, "%s cq %p cqid 0x%x", __func__, cq, cq->cqid);
275 	ret = t4_next_hw_cqe(cq, &cqe);
276 	while (!ret) {
277 		CTR3(KTR_IW_CXGBE, "%s flushing hwcq cidx 0x%x swcq pidx 0x%x",
278 		    __func__, cq->cidx, cq->sw_pidx);
279 		swcqe = &cq->sw_queue[cq->sw_pidx];
280 		*swcqe = *cqe;
281 		swcqe->header |= cpu_to_be32(V_CQE_SWCQE(1));
282 		t4_swcq_produce(cq);
283 		t4_hwcq_consume(cq);
284 		ret = t4_next_hw_cqe(cq, &cqe);
285 	}
286 }
287 
288 static int cqe_completes_wr(struct t4_cqe *cqe, struct t4_wq *wq)
289 {
290 	if (CQE_OPCODE(cqe) == FW_RI_TERMINATE)
291 		return 0;
292 
293 	if ((CQE_OPCODE(cqe) == FW_RI_RDMA_WRITE) && RQ_TYPE(cqe))
294 		return 0;
295 
296 	if ((CQE_OPCODE(cqe) == FW_RI_READ_RESP) && SQ_TYPE(cqe))
297 		return 0;
298 
299 	if (CQE_SEND_OPCODE(cqe) && RQ_TYPE(cqe) && t4_rq_empty(wq))
300 		return 0;
301 	return 1;
302 }
303 
304 void c4iw_count_scqes(struct t4_cq *cq, struct t4_wq *wq, int *count)
305 {
306 	struct t4_cqe *cqe;
307 	u32 ptr;
308 
309 	*count = 0;
310 	ptr = cq->sw_cidx;
311 	while (ptr != cq->sw_pidx) {
312 		cqe = &cq->sw_queue[ptr];
313 		if ((SQ_TYPE(cqe) || ((CQE_OPCODE(cqe) == FW_RI_READ_RESP) &&
314 				      wq->sq.oldest_read)) &&
315 		    (CQE_QPID(cqe) == wq->sq.qid))
316 			(*count)++;
317 		if (++ptr == cq->size)
318 			ptr = 0;
319 	}
320 	CTR3(KTR_IW_CXGBE, "%s cq %p count %d", __func__, cq, *count);
321 }
322 
323 void c4iw_count_rcqes(struct t4_cq *cq, struct t4_wq *wq, int *count)
324 {
325 	struct t4_cqe *cqe;
326 	u32 ptr;
327 
328 	*count = 0;
329 	CTR2(KTR_IW_CXGBE, "%s count zero %d", __func__, *count);
330 	ptr = cq->sw_cidx;
331 	while (ptr != cq->sw_pidx) {
332 		cqe = &cq->sw_queue[ptr];
333 		if (RQ_TYPE(cqe) && (CQE_OPCODE(cqe) != FW_RI_READ_RESP) &&
334 		    (CQE_QPID(cqe) == wq->sq.qid) && cqe_completes_wr(cqe, wq))
335 			(*count)++;
336 		if (++ptr == cq->size)
337 			ptr = 0;
338 	}
339 	CTR3(KTR_IW_CXGBE, "%s cq %p count %d", __func__, cq, *count);
340 }
341 
342 static void flush_completed_wrs(struct t4_wq *wq, struct t4_cq *cq)
343 {
344 	struct t4_swsqe *swsqe;
345 	u16 ptr = wq->sq.cidx;
346 	int count = wq->sq.in_use;
347 	int unsignaled = 0;
348 
349 	swsqe = &wq->sq.sw_sq[ptr];
350 	while (count--)
351 		if (!swsqe->signaled) {
352 			if (++ptr == wq->sq.size)
353 				ptr = 0;
354 			swsqe = &wq->sq.sw_sq[ptr];
355 			unsignaled++;
356 		} else if (swsqe->complete) {
357 
358 			/*
359 			 * Insert this completed cqe into the swcq.
360 			 */
361 			CTR3(KTR_IW_CXGBE,
362 			    "%s moving cqe into swcq sq idx %u cq idx %u",
363 			    __func__, ptr, cq->sw_pidx);
364 			swsqe->cqe.header |= htonl(V_CQE_SWCQE(1));
365 			cq->sw_queue[cq->sw_pidx] = swsqe->cqe;
366 			t4_swcq_produce(cq);
367 			swsqe->signaled = 0;
368 			wq->sq.in_use -= unsignaled;
369 			break;
370 		} else
371 			break;
372 }
373 
374 static void create_read_req_cqe(struct t4_wq *wq, struct t4_cqe *hw_cqe,
375 				struct t4_cqe *read_cqe)
376 {
377 	read_cqe->u.scqe.cidx = wq->sq.oldest_read->idx;
378 	read_cqe->len = cpu_to_be32(wq->sq.oldest_read->read_len);
379 	read_cqe->header = htonl(V_CQE_QPID(CQE_QPID(hw_cqe)) |
380 				 V_CQE_SWCQE(SW_CQE(hw_cqe)) |
381 				 V_CQE_OPCODE(FW_RI_READ_REQ) |
382 				 V_CQE_TYPE(1));
383 	read_cqe->bits_type_ts = hw_cqe->bits_type_ts;
384 }
385 
386 /*
387  * Return a ptr to the next read wr in the SWSQ or NULL.
388  */
389 static void advance_oldest_read(struct t4_wq *wq)
390 {
391 
392 	u32 rptr = wq->sq.oldest_read - wq->sq.sw_sq + 1;
393 
394 	if (rptr == wq->sq.size)
395 		rptr = 0;
396 	while (rptr != wq->sq.pidx) {
397 		wq->sq.oldest_read = &wq->sq.sw_sq[rptr];
398 
399 		if (wq->sq.oldest_read->opcode == FW_RI_READ_REQ)
400 			return;
401 		if (++rptr == wq->sq.size)
402 			rptr = 0;
403 	}
404 	wq->sq.oldest_read = NULL;
405 }
406 
407 /*
408  * poll_cq
409  *
410  * Caller must:
411  *     check the validity of the first CQE,
412  *     supply the wq assicated with the qpid.
413  *
414  * credit: cq credit to return to sge.
415  * cqe_flushed: 1 iff the CQE is flushed.
416  * cqe: copy of the polled CQE.
417  *
418  * return value:
419  *    0		    CQE returned ok.
420  *    -EAGAIN       CQE skipped, try again.
421  *    -EOVERFLOW    CQ overflow detected.
422  */
423 static int poll_cq(struct t4_wq *wq, struct t4_cq *cq, struct t4_cqe *cqe,
424 		   u8 *cqe_flushed, u64 *cookie, u32 *credit)
425 {
426 	int ret = 0;
427 	struct t4_cqe *hw_cqe, read_cqe;
428 
429 	*cqe_flushed = 0;
430 	*credit = 0;
431 	ret = t4_next_cqe(cq, &hw_cqe);
432 	if (ret)
433 		return ret;
434 
435 	CTR6(KTR_IW_CXGBE,
436 	    "%s CQE OVF %u qpid 0x%0x genbit %u type %u status 0x%0x", __func__,
437 	    CQE_OVFBIT(hw_cqe), CQE_QPID(hw_cqe), CQE_GENBIT(hw_cqe),
438 	    CQE_TYPE(hw_cqe), CQE_STATUS(hw_cqe));
439 	CTR5(KTR_IW_CXGBE,
440 	    "%s opcode 0x%0x len 0x%0x wrid_hi_stag 0x%x wrid_low_msn 0x%x",
441 	    __func__, CQE_OPCODE(hw_cqe), CQE_LEN(hw_cqe), CQE_WRID_HI(hw_cqe),
442 	    CQE_WRID_LOW(hw_cqe));
443 
444 	/*
445 	 * skip cqe's not affiliated with a QP.
446 	 */
447 	if (wq == NULL) {
448 		ret = -EAGAIN;
449 		goto skip_cqe;
450 	}
451 
452 	/*
453 	 * Gotta tweak READ completions:
454 	 *	1) the cqe doesn't contain the sq_wptr from the wr.
455 	 *	2) opcode not reflected from the wr.
456 	 *	3) read_len not reflected from the wr.
457 	 *	4) cq_type is RQ_TYPE not SQ_TYPE.
458 	 */
459 	if (RQ_TYPE(hw_cqe) && (CQE_OPCODE(hw_cqe) == FW_RI_READ_RESP)) {
460 
461 		/*
462 		 * If this is an unsolicited read response, then the read
463 		 * was generated by the kernel driver as part of peer-2-peer
464 		 * connection setup.  So ignore the completion.
465 		 */
466 		if (!wq->sq.oldest_read) {
467 			if (CQE_STATUS(hw_cqe))
468 				t4_set_wq_in_error(wq);
469 			ret = -EAGAIN;
470 			goto skip_cqe;
471 		}
472 
473 		/*
474 		 * Don't write to the HWCQ, so create a new read req CQE
475 		 * in local memory.
476 		 */
477 		create_read_req_cqe(wq, hw_cqe, &read_cqe);
478 		hw_cqe = &read_cqe;
479 		advance_oldest_read(wq);
480 	}
481 
482 	if (CQE_STATUS(hw_cqe) || t4_wq_in_error(wq)) {
483 		*cqe_flushed = t4_wq_in_error(wq);
484 		t4_set_wq_in_error(wq);
485 		goto proc_cqe;
486 	}
487 
488 	if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE) {
489 		ret = -EAGAIN;
490 		goto skip_cqe;
491 	}
492 
493 	/*
494 	 * RECV completion.
495 	 */
496 	if (RQ_TYPE(hw_cqe)) {
497 
498 		/*
499 		 * HW only validates 4 bits of MSN.  So we must validate that
500 		 * the MSN in the SEND is the next expected MSN.  If its not,
501 		 * then we complete this with T4_ERR_MSN and mark the wq in
502 		 * error.
503 		 */
504 
505 		if (t4_rq_empty(wq)) {
506 			t4_set_wq_in_error(wq);
507 			ret = -EAGAIN;
508 			goto skip_cqe;
509 		}
510 		if (unlikely((CQE_WRID_MSN(hw_cqe) != (wq->rq.msn)))) {
511 			t4_set_wq_in_error(wq);
512 			hw_cqe->header |= htonl(V_CQE_STATUS(T4_ERR_MSN));
513 			goto proc_cqe;
514 		}
515 		goto proc_cqe;
516 	}
517 
518 	/*
519 	 * If we get here its a send completion.
520 	 *
521 	 * Handle out of order completion. These get stuffed
522 	 * in the SW SQ. Then the SW SQ is walked to move any
523 	 * now in-order completions into the SW CQ.  This handles
524 	 * 2 cases:
525 	 *	1) reaping unsignaled WRs when the first subsequent
526 	 *	   signaled WR is completed.
527 	 *	2) out of order read completions.
528 	 */
529 	if (!SW_CQE(hw_cqe) && (CQE_WRID_SQ_IDX(hw_cqe) != wq->sq.cidx)) {
530 		struct t4_swsqe *swsqe;
531 
532 		CTR2(KTR_IW_CXGBE,
533 		    "%s out of order completion going in sw_sq at idx %u",
534 		    __func__, CQE_WRID_SQ_IDX(hw_cqe));
535 		swsqe = &wq->sq.sw_sq[CQE_WRID_SQ_IDX(hw_cqe)];
536 		swsqe->cqe = *hw_cqe;
537 		swsqe->complete = 1;
538 		ret = -EAGAIN;
539 		goto flush_wq;
540 	}
541 
542 proc_cqe:
543 	*cqe = *hw_cqe;
544 
545 	/*
546 	 * Reap the associated WR(s) that are freed up with this
547 	 * completion.
548 	 */
549 	if (SQ_TYPE(hw_cqe)) {
550 		wq->sq.cidx = CQE_WRID_SQ_IDX(hw_cqe);
551 		CTR2(KTR_IW_CXGBE, "%s completing sq idx %u",
552 		     __func__, wq->sq.cidx);
553 		*cookie = wq->sq.sw_sq[wq->sq.cidx].wr_id;
554 		t4_sq_consume(wq);
555 	} else {
556 		CTR2(KTR_IW_CXGBE, "%s completing rq idx %u",
557 		     __func__, wq->rq.cidx);
558 		*cookie = wq->rq.sw_rq[wq->rq.cidx].wr_id;
559 		BUG_ON(t4_rq_empty(wq));
560 		t4_rq_consume(wq);
561 	}
562 
563 flush_wq:
564 	/*
565 	 * Flush any completed cqes that are now in-order.
566 	 */
567 	flush_completed_wrs(wq, cq);
568 
569 skip_cqe:
570 	if (SW_CQE(hw_cqe)) {
571 		CTR4(KTR_IW_CXGBE, "%s cq %p cqid 0x%x skip sw cqe cidx %u",
572 		     __func__, cq, cq->cqid, cq->sw_cidx);
573 		t4_swcq_consume(cq);
574 	} else {
575 		CTR4(KTR_IW_CXGBE, "%s cq %p cqid 0x%x skip hw cqe cidx %u",
576 		     __func__, cq, cq->cqid, cq->cidx);
577 		t4_hwcq_consume(cq);
578 	}
579 	return ret;
580 }
581 
582 /*
583  * Get one cq entry from c4iw and map it to openib.
584  *
585  * Returns:
586  *	0			cqe returned
587  *	-ENODATA		EMPTY;
588  *	-EAGAIN			caller must try again
589  *	any other -errno	fatal error
590  */
591 static int c4iw_poll_cq_one(struct c4iw_cq *chp, struct ib_wc *wc)
592 {
593 	struct c4iw_qp *qhp = NULL;
594 	struct t4_cqe cqe = {0, 0}, *rd_cqe;
595 	struct t4_wq *wq;
596 	u32 credit = 0;
597 	u8 cqe_flushed;
598 	u64 cookie = 0;
599 	int ret;
600 
601 	ret = t4_next_cqe(&chp->cq, &rd_cqe);
602 
603 	if (ret)
604 		return ret;
605 
606 	qhp = get_qhp(chp->rhp, CQE_QPID(rd_cqe));
607 	if (!qhp)
608 		wq = NULL;
609 	else {
610 		spin_lock(&qhp->lock);
611 		wq = &(qhp->wq);
612 	}
613 	ret = poll_cq(wq, &(chp->cq), &cqe, &cqe_flushed, &cookie, &credit);
614 	if (ret)
615 		goto out;
616 
617 	wc->wr_id = cookie;
618 	wc->qp = &qhp->ibqp;
619 	wc->vendor_err = CQE_STATUS(&cqe);
620 	wc->wc_flags = 0;
621 
622 	CTR5(KTR_IW_CXGBE, "%s qpid 0x%x type %d opcode %d status 0x%x",
623 	    __func__, CQE_QPID(&cqe), CQE_TYPE(&cqe), CQE_OPCODE(&cqe),
624 	    CQE_STATUS(&cqe));
625 	CTR5(KTR_IW_CXGBE, "%s len %u wrid hi 0x%x lo 0x%x cookie 0x%llx",
626 	    __func__, CQE_LEN(&cqe), CQE_WRID_HI(&cqe), CQE_WRID_LOW(&cqe),
627 	    (unsigned long long)cookie);
628 
629 	if (CQE_TYPE(&cqe) == 0) {
630 		if (!CQE_STATUS(&cqe))
631 			wc->byte_len = CQE_LEN(&cqe);
632 		else
633 			wc->byte_len = 0;
634 		wc->opcode = IB_WC_RECV;
635 		if (CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_INV ||
636 		    CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_SE_INV) {
637 			wc->ex.invalidate_rkey = CQE_WRID_STAG(&cqe);
638 			wc->wc_flags |= IB_WC_WITH_INVALIDATE;
639 		}
640 	} else {
641 		switch (CQE_OPCODE(&cqe)) {
642 		case FW_RI_RDMA_WRITE:
643 			wc->opcode = IB_WC_RDMA_WRITE;
644 			break;
645 		case FW_RI_READ_REQ:
646 			wc->opcode = IB_WC_RDMA_READ;
647 			wc->byte_len = CQE_LEN(&cqe);
648 			break;
649 		case FW_RI_SEND_WITH_INV:
650 		case FW_RI_SEND_WITH_SE_INV:
651 			wc->opcode = IB_WC_SEND;
652 			wc->wc_flags |= IB_WC_WITH_INVALIDATE;
653 			break;
654 		case FW_RI_SEND:
655 		case FW_RI_SEND_WITH_SE:
656 			wc->opcode = IB_WC_SEND;
657 			break;
658 		case FW_RI_BIND_MW:
659 			wc->opcode = IB_WC_BIND_MW;
660 			break;
661 
662 		case FW_RI_LOCAL_INV:
663 			wc->opcode = IB_WC_LOCAL_INV;
664 			break;
665 		case FW_RI_FAST_REGISTER:
666 			wc->opcode = IB_WC_FAST_REG_MR;
667 			break;
668 		default:
669 			printf("Unexpected opcode %d "
670 			       "in the CQE received for QPID = 0x%0x\n",
671 			       CQE_OPCODE(&cqe), CQE_QPID(&cqe));
672 			ret = -EINVAL;
673 			goto out;
674 		}
675 	}
676 
677 	if (cqe_flushed)
678 		wc->status = IB_WC_WR_FLUSH_ERR;
679 	else {
680 
681 		switch (CQE_STATUS(&cqe)) {
682 		case T4_ERR_SUCCESS:
683 			wc->status = IB_WC_SUCCESS;
684 			break;
685 		case T4_ERR_STAG:
686 			wc->status = IB_WC_LOC_ACCESS_ERR;
687 			break;
688 		case T4_ERR_PDID:
689 			wc->status = IB_WC_LOC_PROT_ERR;
690 			break;
691 		case T4_ERR_QPID:
692 		case T4_ERR_ACCESS:
693 			wc->status = IB_WC_LOC_ACCESS_ERR;
694 			break;
695 		case T4_ERR_WRAP:
696 			wc->status = IB_WC_GENERAL_ERR;
697 			break;
698 		case T4_ERR_BOUND:
699 			wc->status = IB_WC_LOC_LEN_ERR;
700 			break;
701 		case T4_ERR_INVALIDATE_SHARED_MR:
702 		case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
703 			wc->status = IB_WC_MW_BIND_ERR;
704 			break;
705 		case T4_ERR_CRC:
706 		case T4_ERR_MARKER:
707 		case T4_ERR_PDU_LEN_ERR:
708 		case T4_ERR_OUT_OF_RQE:
709 		case T4_ERR_DDP_VERSION:
710 		case T4_ERR_RDMA_VERSION:
711 		case T4_ERR_DDP_QUEUE_NUM:
712 		case T4_ERR_MSN:
713 		case T4_ERR_TBIT:
714 		case T4_ERR_MO:
715 		case T4_ERR_MSN_RANGE:
716 		case T4_ERR_IRD_OVERFLOW:
717 		case T4_ERR_OPCODE:
718 		case T4_ERR_INTERNAL_ERR:
719 			wc->status = IB_WC_FATAL_ERR;
720 			break;
721 		case T4_ERR_SWFLUSH:
722 			wc->status = IB_WC_WR_FLUSH_ERR;
723 			break;
724 		default:
725 			printf("Unexpected cqe_status 0x%x for QPID = 0x%0x\n",
726 			       CQE_STATUS(&cqe), CQE_QPID(&cqe));
727 			wc->status = IB_WC_FATAL_ERR;
728 		}
729 	}
730 out:
731 	if (wq)
732 		spin_unlock(&qhp->lock);
733 	return ret;
734 }
735 
736 int c4iw_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
737 {
738 	struct c4iw_cq *chp;
739 	unsigned long flags;
740 	int npolled;
741 	int err = 0;
742 
743 	chp = to_c4iw_cq(ibcq);
744 
745 	spin_lock_irqsave(&chp->lock, flags);
746 	for (npolled = 0; npolled < num_entries; ++npolled) {
747 		do {
748 			err = c4iw_poll_cq_one(chp, wc + npolled);
749 		} while (err == -EAGAIN);
750 		if (err)
751 			break;
752 	}
753 	spin_unlock_irqrestore(&chp->lock, flags);
754 	return !err || err == -ENODATA ? npolled : err;
755 }
756 
757 int c4iw_destroy_cq(struct ib_cq *ib_cq)
758 {
759 	struct c4iw_cq *chp;
760 	struct c4iw_ucontext *ucontext;
761 
762 	CTR2(KTR_IW_CXGBE, "%s ib_cq %p", __func__, ib_cq);
763 	chp = to_c4iw_cq(ib_cq);
764 
765 	remove_handle(chp->rhp, &chp->rhp->cqidr, chp->cq.cqid);
766 	atomic_dec(&chp->refcnt);
767 	wait_event(chp->wait, !atomic_read(&chp->refcnt));
768 
769 	ucontext = ib_cq->uobject ? to_c4iw_ucontext(ib_cq->uobject->context)
770 				  : NULL;
771 	destroy_cq(&chp->rhp->rdev, &chp->cq,
772 		   ucontext ? &ucontext->uctx : &chp->cq.rdev->uctx);
773 	kfree(chp);
774 	return 0;
775 }
776 
777 struct ib_cq *
778 c4iw_create_cq(struct ib_device *ibdev, struct ib_cq_init_attr *attr,
779     struct ib_ucontext *ib_context, struct ib_udata *udata)
780 {
781 	struct c4iw_dev *rhp;
782 	struct c4iw_cq *chp;
783 	struct c4iw_create_cq_resp uresp;
784 	struct c4iw_ucontext *ucontext = NULL;
785 	int ret;
786 	size_t memsize, hwentries;
787 	struct c4iw_mm_entry *mm, *mm2;
788 	int entries = attr->cqe;
789 
790 	CTR3(KTR_IW_CXGBE, "%s ib_dev %p entries %d", __func__, ibdev, entries);
791 
792 	rhp = to_c4iw_dev(ibdev);
793 
794 	chp = kzalloc(sizeof(*chp), GFP_KERNEL);
795 	if (!chp)
796 		return ERR_PTR(-ENOMEM);
797 
798 	if (ib_context)
799 		ucontext = to_c4iw_ucontext(ib_context);
800 
801 	/* account for the status page. */
802 	entries++;
803 
804 	/* IQ needs one extra entry to differentiate full vs empty. */
805 	entries++;
806 
807 	/*
808 	 * entries must be multiple of 16 for HW.
809 	 */
810 	entries = roundup(entries, 16);
811 
812 	/*
813 	 * Make actual HW queue 2x to avoid cidx_inc overflows.
814 	 */
815 	hwentries = entries * 2;
816 
817 	/*
818 	 * Make HW queue at least 64 entries so GTS updates aren't too
819 	 * frequent.
820 	 */
821 	if (hwentries < 64)
822 		hwentries = 64;
823 
824 	memsize = hwentries * sizeof *chp->cq.queue;
825 
826 	/*
827 	 * memsize must be a multiple of the page size if its a user cq.
828 	 */
829 	if (ucontext) {
830 		memsize = roundup(memsize, PAGE_SIZE);
831 		hwentries = memsize / sizeof *chp->cq.queue;
832 		while (hwentries > T4_MAX_IQ_SIZE) {
833 			memsize -= PAGE_SIZE;
834 			hwentries = memsize / sizeof *chp->cq.queue;
835 		}
836 	}
837 	chp->cq.size = hwentries;
838 	chp->cq.memsize = memsize;
839 
840 	ret = create_cq(&rhp->rdev, &chp->cq,
841 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
842 	if (ret)
843 		goto err1;
844 
845 	chp->rhp = rhp;
846 	chp->cq.size--;				/* status page */
847 	chp->ibcq.cqe = entries - 2;
848 	spin_lock_init(&chp->lock);
849 	spin_lock_init(&chp->comp_handler_lock);
850 	atomic_set(&chp->refcnt, 1);
851 	init_waitqueue_head(&chp->wait);
852 	ret = insert_handle(rhp, &rhp->cqidr, chp, chp->cq.cqid);
853 	if (ret)
854 		goto err2;
855 
856 	if (ucontext) {
857 		mm = kmalloc(sizeof *mm, GFP_KERNEL);
858 		if (!mm)
859 			goto err3;
860 		mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
861 		if (!mm2)
862 			goto err4;
863 
864 		memset(&uresp, 0, sizeof(uresp));
865 		uresp.qid_mask = rhp->rdev.cqmask;
866 		uresp.cqid = chp->cq.cqid;
867 		uresp.size = chp->cq.size;
868 		uresp.memsize = chp->cq.memsize;
869 		spin_lock(&ucontext->mmap_lock);
870 		uresp.key = ucontext->key;
871 		ucontext->key += PAGE_SIZE;
872 		uresp.gts_key = ucontext->key;
873 		ucontext->key += PAGE_SIZE;
874 		spin_unlock(&ucontext->mmap_lock);
875 		ret = ib_copy_to_udata(udata, &uresp,
876 					sizeof(uresp) - sizeof(uresp.reserved));
877 		if (ret)
878 			goto err5;
879 
880 		mm->key = uresp.key;
881 		mm->addr = vtophys(chp->cq.queue);
882 		mm->len = chp->cq.memsize;
883 		insert_mmap(ucontext, mm);
884 
885 		mm2->key = uresp.gts_key;
886 		mm2->addr = chp->cq.ugts;
887 		mm2->len = PAGE_SIZE;
888 		insert_mmap(ucontext, mm2);
889 	}
890 	CTR6(KTR_IW_CXGBE,
891 	    "%s cqid 0x%0x chp %p size %u memsize %zu, dma_addr 0x%0llx",
892 	    __func__, chp->cq.cqid, chp, chp->cq.size, chp->cq.memsize,
893 	    (unsigned long long) chp->cq.dma_addr);
894 	return &chp->ibcq;
895 err5:
896 	kfree(mm2);
897 err4:
898 	kfree(mm);
899 err3:
900 	remove_handle(rhp, &rhp->cqidr, chp->cq.cqid);
901 err2:
902 	destroy_cq(&chp->rhp->rdev, &chp->cq,
903 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
904 err1:
905 	kfree(chp);
906 	return ERR_PTR(ret);
907 }
908 
909 int c4iw_resize_cq(struct ib_cq *cq, int cqe, struct ib_udata *udata)
910 {
911 	return -ENOSYS;
912 }
913 
914 int c4iw_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
915 {
916 	struct c4iw_cq *chp;
917 	int ret;
918 	unsigned long flag;
919 
920 	chp = to_c4iw_cq(ibcq);
921 	spin_lock_irqsave(&chp->lock, flag);
922 	ret = t4_arm_cq(&chp->cq,
923 			(flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED);
924 	spin_unlock_irqrestore(&chp->lock, flag);
925 	if (ret && !(flags & IB_CQ_REPORT_MISSED_EVENTS))
926 		ret = 0;
927 	return ret;
928 }
929 #endif
930