xref: /freebsd/contrib/ofed/libcxgb4/qp.c (revision 031beb4e239bfce798af17f5fe8dba8bcaf13d99)
1 /*
2  * Copyright (c) 2006-2016 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 <config.h>
33 
34 #include <assert.h>
35 #include <stdlib.h>
36 #include <pthread.h>
37 #include <string.h>
38 #include <stdio.h>
39 #include "libcxgb4.h"
40 
41 #ifdef STATS
42 struct c4iw_stats c4iw_stats;
43 #endif
44 
45 static void copy_wr_to_sq(struct t4_wq *wq, union t4_wr *wqe, u8 len16)
46 {
47 	void *src, *dst;
48 	uintptr_t end;
49 	int total, len;
50 
51 	src = &wqe->flits[0];
52 	dst = &wq->sq.queue->flits[wq->sq.wq_pidx *
53 	    (T4_EQ_ENTRY_SIZE / sizeof(__be64))];
54 	if (t4_sq_onchip(wq)) {
55 		len16 = align(len16, 4);
56 
57 		/* In onchip mode the copy below will be made to WC memory and
58 		 * could trigger DMA. In offchip mode the copy below only
59 		 * queues the WQE, DMA cannot start until t4_ring_sq_db
60 		 * happens */
61 		mmio_wc_start();
62 	}
63 
64 	/* NOTE len16 cannot be large enough to write to the
65 	   same sq.queue memory twice in this loop */
66 	total = len16 * 16;
67 	end = (uintptr_t)&wq->sq.queue[wq->sq.size];
68 	if (__predict_true((uintptr_t)dst + total <= end)) {
69 		/* Won't wrap around. */
70 		memcpy(dst, src, total);
71 	} else {
72 		len = end - (uintptr_t)dst;
73 		memcpy(dst, src, len);
74 		memcpy(wq->sq.queue, src + len, total - len);
75 	}
76 
77 	if (t4_sq_onchip(wq))
78 		mmio_flush_writes();
79 }
80 
81 static void copy_wr_to_rq(struct t4_wq *wq, union t4_recv_wr *wqe, u8 len16)
82 {
83 	void *src, *dst;
84 	uintptr_t end;
85 	int total, len;
86 
87 	src = &wqe->flits[0];
88 	dst = &wq->rq.queue->flits[wq->rq.wq_pidx *
89 	    (T4_EQ_ENTRY_SIZE / sizeof(__be64))];
90 
91 	total = len16 * 16;
92 	end = (uintptr_t)&wq->rq.queue[wq->rq.size];
93 	if (__predict_true((uintptr_t)dst + total <= end)) {
94 		/* Won't wrap around. */
95 		memcpy(dst, src, total);
96 	} else {
97 		len = end - (uintptr_t)dst;
98 		memcpy(dst, src, len);
99 		memcpy(wq->rq.queue, src + len, total - len);
100 	}
101 }
102 
103 static int build_immd(struct t4_sq *sq, struct fw_ri_immd *immdp,
104 		      struct ibv_send_wr *wr, int max, u32 *plenp)
105 {
106 	u8 *dstp, *srcp;
107 	u32 plen = 0;
108 	int i;
109 	int len;
110 
111 	dstp = (u8 *)immdp->data;
112 	for (i = 0; i < wr->num_sge; i++) {
113 		if ((plen + wr->sg_list[i].length) > max)
114 			return -EMSGSIZE;
115 		srcp = (u8 *)(unsigned long)wr->sg_list[i].addr;
116 		plen += wr->sg_list[i].length;
117 		len = wr->sg_list[i].length;
118 		memcpy(dstp, srcp, len);
119 		dstp += len;
120 		srcp += len;
121 	}
122 	len = ROUND_UP(plen + 8, 16) - (plen + 8);
123 	if (len)
124 		memset(dstp, 0, len);
125 	immdp->op = FW_RI_DATA_IMMD;
126 	immdp->r1 = 0;
127 	immdp->r2 = 0;
128 	immdp->immdlen = htobe32(plen);
129 	*plenp = plen;
130 	return 0;
131 }
132 
133 static int build_isgl(struct fw_ri_isgl *isglp, struct ibv_sge *sg_list,
134 		      int num_sge, u32 *plenp)
135 {
136 	int i;
137 	u32 plen = 0;
138 	__be64 *flitp = (__be64 *)isglp->sge;
139 
140 	for (i = 0; i < num_sge; i++) {
141 		if ((plen + sg_list[i].length) < plen)
142 			return -EMSGSIZE;
143 		plen += sg_list[i].length;
144 		*flitp++ = htobe64(((u64)sg_list[i].lkey << 32) |
145 				     sg_list[i].length);
146 		*flitp++ = htobe64(sg_list[i].addr);
147 	}
148 	*flitp = 0;
149 	isglp->op = FW_RI_DATA_ISGL;
150 	isglp->r1 = 0;
151 	isglp->nsge = htobe16(num_sge);
152 	isglp->r2 = 0;
153 	if (plenp)
154 		*plenp = plen;
155 	return 0;
156 }
157 
158 static int build_rdma_send(struct t4_sq *sq, union t4_wr *wqe,
159 			   struct ibv_send_wr *wr, u8 *len16)
160 {
161 	u32 plen;
162 	int size;
163 	int ret;
164 
165 	if (wr->num_sge > T4_MAX_SEND_SGE)
166 		return -EINVAL;
167 	if (wr->send_flags & IBV_SEND_SOLICITED)
168 		wqe->send.sendop_pkd = htobe32(
169 			FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_SE));
170 	else
171 		wqe->send.sendop_pkd = htobe32(
172 			FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND));
173 	wqe->send.stag_inv = 0;
174 	wqe->send.r3 = 0;
175 	wqe->send.r4 = 0;
176 
177 	plen = 0;
178 	if (wr->num_sge) {
179 		if (wr->send_flags & IBV_SEND_INLINE) {
180 			ret = build_immd(sq, wqe->send.u.immd_src, wr,
181 					 T4_MAX_SEND_INLINE, &plen);
182 			if (ret)
183 				return ret;
184 			size = sizeof wqe->send + sizeof(struct fw_ri_immd) +
185 			       plen;
186 		} else {
187 			ret = build_isgl(wqe->send.u.isgl_src,
188 					 wr->sg_list, wr->num_sge, &plen);
189 			if (ret)
190 				return ret;
191 			size = sizeof wqe->send + sizeof(struct fw_ri_isgl) +
192 			       wr->num_sge * sizeof (struct fw_ri_sge);
193 		}
194 	} else {
195 		wqe->send.u.immd_src[0].op = FW_RI_DATA_IMMD;
196 		wqe->send.u.immd_src[0].r1 = 0;
197 		wqe->send.u.immd_src[0].r2 = 0;
198 		wqe->send.u.immd_src[0].immdlen = 0;
199 		size = sizeof wqe->send + sizeof(struct fw_ri_immd);
200 		plen = 0;
201 	}
202 	*len16 = DIV_ROUND_UP(size, 16);
203 	wqe->send.plen = htobe32(plen);
204 	return 0;
205 }
206 
207 static int build_rdma_write(struct t4_sq *sq, union t4_wr *wqe,
208 			    struct ibv_send_wr *wr, u8 *len16)
209 {
210 	u32 plen;
211 	int size;
212 	int ret;
213 
214 	if (wr->num_sge > T4_MAX_SEND_SGE)
215 		return -EINVAL;
216 	wqe->write.r2 = 0;
217 	wqe->write.stag_sink = htobe32(wr->wr.rdma.rkey);
218 	wqe->write.to_sink = htobe64(wr->wr.rdma.remote_addr);
219 	if (wr->num_sge) {
220 		if (wr->send_flags & IBV_SEND_INLINE) {
221 			ret = build_immd(sq, wqe->write.u.immd_src, wr,
222 					 T4_MAX_WRITE_INLINE, &plen);
223 			if (ret)
224 				return ret;
225 			size = sizeof wqe->write + sizeof(struct fw_ri_immd) +
226 			       plen;
227 		} else {
228 			ret = build_isgl(wqe->write.u.isgl_src,
229 					 wr->sg_list, wr->num_sge, &plen);
230 			if (ret)
231 				return ret;
232 			size = sizeof wqe->write + sizeof(struct fw_ri_isgl) +
233 			       wr->num_sge * sizeof (struct fw_ri_sge);
234 		}
235 	} else {
236 		wqe->write.u.immd_src[0].op = FW_RI_DATA_IMMD;
237 		wqe->write.u.immd_src[0].r1 = 0;
238 		wqe->write.u.immd_src[0].r2 = 0;
239 		wqe->write.u.immd_src[0].immdlen = 0;
240 		size = sizeof wqe->write + sizeof(struct fw_ri_immd);
241 		plen = 0;
242 	}
243 	*len16 = DIV_ROUND_UP(size, 16);
244 	wqe->write.plen = htobe32(plen);
245 	return 0;
246 }
247 
248 static int build_rdma_read(union t4_wr *wqe, struct ibv_send_wr *wr, u8 *len16)
249 {
250 	if (wr->num_sge > 1)
251 		return -EINVAL;
252 	if (wr->num_sge) {
253 		wqe->read.stag_src = htobe32(wr->wr.rdma.rkey);
254 		wqe->read.to_src_hi = htobe32((u32)(wr->wr.rdma.remote_addr >>32));
255 		wqe->read.to_src_lo = htobe32((u32)wr->wr.rdma.remote_addr);
256 		wqe->read.stag_sink = htobe32(wr->sg_list[0].lkey);
257 		wqe->read.plen = htobe32(wr->sg_list[0].length);
258 		wqe->read.to_sink_hi = htobe32((u32)(wr->sg_list[0].addr >> 32));
259 		wqe->read.to_sink_lo = htobe32((u32)(wr->sg_list[0].addr));
260 	} else {
261 		wqe->read.stag_src = htobe32(2);
262 		wqe->read.to_src_hi = 0;
263 		wqe->read.to_src_lo = 0;
264 		wqe->read.stag_sink = htobe32(2);
265 		wqe->read.plen = 0;
266 		wqe->read.to_sink_hi = 0;
267 		wqe->read.to_sink_lo = 0;
268 	}
269 	wqe->read.r2 = 0;
270 	wqe->read.r5 = 0;
271 	*len16 = DIV_ROUND_UP(sizeof wqe->read, 16);
272 	return 0;
273 }
274 
275 static int build_rdma_recv(struct c4iw_qp *qhp, union t4_recv_wr *wqe,
276 			   struct ibv_recv_wr *wr, u8 *len16)
277 {
278 	int ret;
279 
280 	ret = build_isgl(&wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL);
281 	if (ret)
282 		return ret;
283 	*len16 = DIV_ROUND_UP(sizeof wqe->recv +
284 			      wr->num_sge * sizeof(struct fw_ri_sge), 16);
285 	return 0;
286 }
287 
288 static void ring_kernel_db(struct c4iw_qp *qhp, u32 qid, u16 idx)
289 {
290 	struct ibv_modify_qp cmd = {};
291 	struct ibv_qp_attr attr;
292 	int mask;
293 	int __attribute__((unused)) ret;
294 
295 	/* FIXME: Why do we need this barrier if the kernel is going to
296 	   trigger the DMA? */
297 	udma_to_device_barrier();
298 	if (qid == qhp->wq.sq.qid) {
299 		attr.sq_psn = idx;
300 		mask = IBV_QP_SQ_PSN;
301 	} else  {
302 		attr.rq_psn = idx;
303 		mask = IBV_QP_RQ_PSN;
304 	}
305 	ret = ibv_cmd_modify_qp(&qhp->ibv_qp, &attr, mask, &cmd, sizeof cmd);
306 	assert(!ret);
307 }
308 
309 int c4iw_post_send(struct ibv_qp *ibqp, struct ibv_send_wr *wr,
310 	           struct ibv_send_wr **bad_wr)
311 {
312 	int err = 0;
313 	u8 len16 = 0;
314 	enum fw_wr_opcodes fw_opcode;
315 	enum fw_ri_wr_flags fw_flags;
316 	struct c4iw_qp *qhp;
317 	union t4_wr *wqe, lwqe;
318 	u32 num_wrs;
319 	struct t4_swsqe *swsqe;
320 	u16 idx = 0;
321 
322 	qhp = to_c4iw_qp(ibqp);
323 	pthread_spin_lock(&qhp->lock);
324 	if (t4_wq_in_error(&qhp->wq)) {
325 		pthread_spin_unlock(&qhp->lock);
326 		*bad_wr = wr;
327 		return -EINVAL;
328 	}
329 	num_wrs = t4_sq_avail(&qhp->wq);
330 	if (num_wrs == 0) {
331 		pthread_spin_unlock(&qhp->lock);
332 		*bad_wr = wr;
333 		return -ENOMEM;
334 	}
335 	while (wr) {
336 		if (num_wrs == 0) {
337 			err = -ENOMEM;
338 			*bad_wr = wr;
339 			break;
340 		}
341 
342 		wqe = &lwqe;
343 		fw_flags = 0;
344 		if (wr->send_flags & IBV_SEND_SOLICITED)
345 			fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
346 		if (wr->send_flags & IBV_SEND_SIGNALED || qhp->sq_sig_all)
347 			fw_flags |= FW_RI_COMPLETION_FLAG;
348 		swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
349 		switch (wr->opcode) {
350 		case IBV_WR_SEND:
351 			INC_STAT(send);
352 			if (wr->send_flags & IBV_SEND_FENCE)
353 				fw_flags |= FW_RI_READ_FENCE_FLAG;
354 			fw_opcode = FW_RI_SEND_WR;
355 			swsqe->opcode = FW_RI_SEND;
356 			err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
357 			break;
358 		case IBV_WR_RDMA_WRITE:
359 			INC_STAT(write);
360 			fw_opcode = FW_RI_RDMA_WRITE_WR;
361 			swsqe->opcode = FW_RI_RDMA_WRITE;
362 			err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
363 			break;
364 		case IBV_WR_RDMA_READ:
365 			INC_STAT(read);
366 			fw_opcode = FW_RI_RDMA_READ_WR;
367 			swsqe->opcode = FW_RI_READ_REQ;
368 			fw_flags = 0;
369 			err = build_rdma_read(wqe, wr, &len16);
370 			if (err)
371 				break;
372 			swsqe->read_len = wr->sg_list ? wr->sg_list[0].length :
373 					  0;
374 			if (!qhp->wq.sq.oldest_read)
375 				qhp->wq.sq.oldest_read = swsqe;
376 			break;
377 		default:
378 			PDBG("%s post of type=%d TBD!\n", __func__,
379 			     wr->opcode);
380 			err = -EINVAL;
381 		}
382 		if (err) {
383 			*bad_wr = wr;
384 			break;
385 		}
386 		swsqe->idx = qhp->wq.sq.pidx;
387 		swsqe->complete = 0;
388 		swsqe->signaled = (wr->send_flags & IBV_SEND_SIGNALED) ||
389 				  qhp->sq_sig_all;
390 		swsqe->flushed = 0;
391 		swsqe->wr_id = wr->wr_id;
392 
393 		init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
394 		PDBG("%s cookie 0x%llx pidx 0x%x opcode 0x%x\n",
395 		     __func__, (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
396 		     swsqe->opcode);
397 		wr = wr->next;
398 		num_wrs--;
399 		copy_wr_to_sq(&qhp->wq, wqe, len16);
400 		t4_sq_produce(&qhp->wq, len16);
401 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
402 	}
403 	if (t4_wq_db_enabled(&qhp->wq)) {
404 		t4_ring_sq_db(&qhp->wq, idx, dev_is_t4(qhp->rhp),
405 			      len16, wqe);
406 	} else
407 		ring_kernel_db(qhp, qhp->wq.sq.qid, idx);
408 	/* This write is only for debugging, the value does not matter for DMA
409 	 */
410 	qhp->wq.sq.queue[qhp->wq.sq.size].status.host_wq_pidx = \
411 			(qhp->wq.sq.wq_pidx);
412 
413 	pthread_spin_unlock(&qhp->lock);
414 	return err;
415 }
416 
417 int c4iw_post_receive(struct ibv_qp *ibqp, struct ibv_recv_wr *wr,
418 			   struct ibv_recv_wr **bad_wr)
419 {
420 	int err = 0;
421 	struct c4iw_qp *qhp;
422 	union t4_recv_wr *wqe, lwqe;
423 	u32 num_wrs;
424 	u8 len16 = 0;
425 	u16 idx = 0;
426 
427 	qhp = to_c4iw_qp(ibqp);
428 	pthread_spin_lock(&qhp->lock);
429 	if (t4_wq_in_error(&qhp->wq)) {
430 		pthread_spin_unlock(&qhp->lock);
431 		*bad_wr = wr;
432 		return -EINVAL;
433 	}
434 	INC_STAT(recv);
435 	num_wrs = t4_rq_avail(&qhp->wq);
436 	if (num_wrs == 0) {
437 		pthread_spin_unlock(&qhp->lock);
438 		*bad_wr = wr;
439 		return -ENOMEM;
440 	}
441 	while (wr) {
442 		if (wr->num_sge > T4_MAX_RECV_SGE) {
443 			err = -EINVAL;
444 			*bad_wr = wr;
445 			break;
446 		}
447 		wqe = &lwqe;
448 		if (num_wrs)
449 			err = build_rdma_recv(qhp, wqe, wr, &len16);
450 		else
451 			err = -ENOMEM;
452 		if (err) {
453 			*bad_wr = wr;
454 			break;
455 		}
456 
457 		qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
458 
459 		wqe->recv.opcode = FW_RI_RECV_WR;
460 		wqe->recv.r1 = 0;
461 		wqe->recv.wrid = qhp->wq.rq.pidx;
462 		wqe->recv.r2[0] = 0;
463 		wqe->recv.r2[1] = 0;
464 		wqe->recv.r2[2] = 0;
465 		wqe->recv.len16 = len16;
466 		PDBG("%s cookie 0x%llx pidx %u\n", __func__,
467 		     (unsigned long long) wr->wr_id, qhp->wq.rq.pidx);
468 		copy_wr_to_rq(&qhp->wq, wqe, len16);
469 		t4_rq_produce(&qhp->wq, len16);
470 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
471 		wr = wr->next;
472 		num_wrs--;
473 	}
474 	if (t4_wq_db_enabled(&qhp->wq))
475 		t4_ring_rq_db(&qhp->wq, idx, dev_is_t4(qhp->rhp),
476 			      len16, wqe);
477 	else
478 		ring_kernel_db(qhp, qhp->wq.rq.qid, idx);
479 	qhp->wq.rq.queue[qhp->wq.rq.size].status.host_wq_pidx = \
480 			(qhp->wq.rq.wq_pidx);
481 	pthread_spin_unlock(&qhp->lock);
482 	return err;
483 }
484 
485 static void update_qp_state(struct c4iw_qp *qhp)
486 {
487 	struct ibv_query_qp cmd;
488 	struct ibv_qp_attr attr;
489 	struct ibv_qp_init_attr iattr;
490 	int ret;
491 
492 	ret = ibv_cmd_query_qp(&qhp->ibv_qp, &attr, IBV_QP_STATE, &iattr,
493 			       &cmd, sizeof cmd);
494 	assert(!ret);
495 	if (!ret)
496 		qhp->ibv_qp.state = attr.qp_state;
497 }
498 
499 /*
500  * Assumes qhp lock is held.
501  */
502 void c4iw_flush_qp(struct c4iw_qp *qhp)
503 {
504 	struct c4iw_cq *rchp, *schp;
505 	int count;
506 
507 	if (qhp->wq.flushed)
508 		return;
509 
510 	update_qp_state(qhp);
511 
512 	rchp = to_c4iw_cq(qhp->ibv_qp.recv_cq);
513 	schp = to_c4iw_cq(qhp->ibv_qp.send_cq);
514 
515 	PDBG("%s qhp %p rchp %p schp %p\n", __func__, qhp, rchp, schp);
516 	qhp->wq.flushed = 1;
517 	pthread_spin_unlock(&qhp->lock);
518 
519 	/* locking heirarchy: cq lock first, then qp lock. */
520 	pthread_spin_lock(&rchp->lock);
521 	pthread_spin_lock(&qhp->lock);
522 	c4iw_flush_hw_cq(rchp);
523 	c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
524 	c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
525 	pthread_spin_unlock(&qhp->lock);
526 	pthread_spin_unlock(&rchp->lock);
527 
528 	/* locking heirarchy: cq lock first, then qp lock. */
529 	pthread_spin_lock(&schp->lock);
530 	pthread_spin_lock(&qhp->lock);
531 	if (schp != rchp)
532 		c4iw_flush_hw_cq(schp);
533 	c4iw_flush_sq(qhp);
534 	pthread_spin_unlock(&qhp->lock);
535 	pthread_spin_unlock(&schp->lock);
536 	pthread_spin_lock(&qhp->lock);
537 }
538 
539 void c4iw_flush_qps(struct c4iw_dev *dev)
540 {
541 	int i;
542 
543 	pthread_spin_lock(&dev->lock);
544 	for (i=0; i < dev->max_qp; i++) {
545 		struct c4iw_qp *qhp = dev->qpid2ptr[i];
546 		if (qhp) {
547 			if (!qhp->wq.flushed && t4_wq_in_error(&qhp->wq)) {
548 				pthread_spin_lock(&qhp->lock);
549 				c4iw_flush_qp(qhp);
550 				pthread_spin_unlock(&qhp->lock);
551 			}
552 		}
553 	}
554 	pthread_spin_unlock(&dev->lock);
555 }
556