1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3  * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
4  * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
5  */
6 
7 #include <linux/dma-mapping.h>
8 #include <net/addrconf.h>
9 #include <rdma/uverbs_ioctl.h>
10 
11 #include "rxe.h"
12 #include "rxe_queue.h"
13 #include "rxe_hw_counters.h"
14 
rxe_query_device(struct ib_device * dev,struct ib_device_attr * attr,struct ib_udata * uhw)15 static int rxe_query_device(struct ib_device *dev,
16 			    struct ib_device_attr *attr,
17 			    struct ib_udata *uhw)
18 {
19 	struct rxe_dev *rxe = to_rdev(dev);
20 
21 	if (uhw->inlen || uhw->outlen)
22 		return -EINVAL;
23 
24 	*attr = rxe->attr;
25 	return 0;
26 }
27 
rxe_query_port(struct ib_device * dev,u32 port_num,struct ib_port_attr * attr)28 static int rxe_query_port(struct ib_device *dev,
29 			  u32 port_num, struct ib_port_attr *attr)
30 {
31 	struct rxe_dev *rxe = to_rdev(dev);
32 	int rc;
33 
34 	/* *attr being zeroed by the caller, avoid zeroing it here */
35 	*attr = rxe->port.attr;
36 
37 	mutex_lock(&rxe->usdev_lock);
38 	rc = ib_get_eth_speed(dev, port_num, &attr->active_speed,
39 			      &attr->active_width);
40 
41 	if (attr->state == IB_PORT_ACTIVE)
42 		attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
43 	else if (dev_get_flags(rxe->ndev) & IFF_UP)
44 		attr->phys_state = IB_PORT_PHYS_STATE_POLLING;
45 	else
46 		attr->phys_state = IB_PORT_PHYS_STATE_DISABLED;
47 
48 	mutex_unlock(&rxe->usdev_lock);
49 
50 	return rc;
51 }
52 
rxe_query_pkey(struct ib_device * device,u32 port_num,u16 index,u16 * pkey)53 static int rxe_query_pkey(struct ib_device *device,
54 			  u32 port_num, u16 index, u16 *pkey)
55 {
56 	if (index > 0)
57 		return -EINVAL;
58 
59 	*pkey = IB_DEFAULT_PKEY_FULL;
60 	return 0;
61 }
62 
rxe_modify_device(struct ib_device * dev,int mask,struct ib_device_modify * attr)63 static int rxe_modify_device(struct ib_device *dev,
64 			     int mask, struct ib_device_modify *attr)
65 {
66 	struct rxe_dev *rxe = to_rdev(dev);
67 
68 	if (mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
69 		     IB_DEVICE_MODIFY_NODE_DESC))
70 		return -EOPNOTSUPP;
71 
72 	if (mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID)
73 		rxe->attr.sys_image_guid = cpu_to_be64(attr->sys_image_guid);
74 
75 	if (mask & IB_DEVICE_MODIFY_NODE_DESC) {
76 		memcpy(rxe->ib_dev.node_desc,
77 		       attr->node_desc, sizeof(rxe->ib_dev.node_desc));
78 	}
79 
80 	return 0;
81 }
82 
rxe_modify_port(struct ib_device * dev,u32 port_num,int mask,struct ib_port_modify * attr)83 static int rxe_modify_port(struct ib_device *dev,
84 			   u32 port_num, int mask, struct ib_port_modify *attr)
85 {
86 	struct rxe_dev *rxe = to_rdev(dev);
87 	struct rxe_port *port;
88 
89 	port = &rxe->port;
90 
91 	port->attr.port_cap_flags |= attr->set_port_cap_mask;
92 	port->attr.port_cap_flags &= ~attr->clr_port_cap_mask;
93 
94 	if (mask & IB_PORT_RESET_QKEY_CNTR)
95 		port->attr.qkey_viol_cntr = 0;
96 
97 	return 0;
98 }
99 
rxe_get_link_layer(struct ib_device * dev,u32 port_num)100 static enum rdma_link_layer rxe_get_link_layer(struct ib_device *dev,
101 					       u32 port_num)
102 {
103 	return IB_LINK_LAYER_ETHERNET;
104 }
105 
rxe_alloc_ucontext(struct ib_ucontext * ibuc,struct ib_udata * udata)106 static int rxe_alloc_ucontext(struct ib_ucontext *ibuc, struct ib_udata *udata)
107 {
108 	struct rxe_dev *rxe = to_rdev(ibuc->device);
109 	struct rxe_ucontext *uc = to_ruc(ibuc);
110 
111 	return rxe_add_to_pool(&rxe->uc_pool, uc);
112 }
113 
rxe_dealloc_ucontext(struct ib_ucontext * ibuc)114 static void rxe_dealloc_ucontext(struct ib_ucontext *ibuc)
115 {
116 	struct rxe_ucontext *uc = to_ruc(ibuc);
117 
118 	rxe_cleanup(uc);
119 }
120 
rxe_port_immutable(struct ib_device * dev,u32 port_num,struct ib_port_immutable * immutable)121 static int rxe_port_immutable(struct ib_device *dev, u32 port_num,
122 			      struct ib_port_immutable *immutable)
123 {
124 	int err;
125 	struct ib_port_attr attr;
126 
127 	immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
128 
129 	err = ib_query_port(dev, port_num, &attr);
130 	if (err)
131 		return err;
132 
133 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
134 	immutable->gid_tbl_len = attr.gid_tbl_len;
135 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
136 
137 	return 0;
138 }
139 
rxe_alloc_pd(struct ib_pd * ibpd,struct ib_udata * udata)140 static int rxe_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
141 {
142 	struct rxe_dev *rxe = to_rdev(ibpd->device);
143 	struct rxe_pd *pd = to_rpd(ibpd);
144 
145 	return rxe_add_to_pool(&rxe->pd_pool, pd);
146 }
147 
rxe_dealloc_pd(struct ib_pd * ibpd,struct ib_udata * udata)148 static int rxe_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
149 {
150 	struct rxe_pd *pd = to_rpd(ibpd);
151 
152 	rxe_cleanup(pd);
153 	return 0;
154 }
155 
rxe_create_ah(struct ib_ah * ibah,struct rdma_ah_init_attr * init_attr,struct ib_udata * udata)156 static int rxe_create_ah(struct ib_ah *ibah,
157 			 struct rdma_ah_init_attr *init_attr,
158 			 struct ib_udata *udata)
159 
160 {
161 	struct rxe_dev *rxe = to_rdev(ibah->device);
162 	struct rxe_ah *ah = to_rah(ibah);
163 	struct rxe_create_ah_resp __user *uresp = NULL;
164 	int err;
165 
166 	if (udata) {
167 		/* test if new user provider */
168 		if (udata->outlen >= sizeof(*uresp))
169 			uresp = udata->outbuf;
170 		ah->is_user = true;
171 	} else {
172 		ah->is_user = false;
173 	}
174 
175 	err = rxe_av_chk_attr(rxe, init_attr->ah_attr);
176 	if (err)
177 		return err;
178 
179 	err = rxe_add_to_pool_ah(&rxe->ah_pool, ah,
180 			init_attr->flags & RDMA_CREATE_AH_SLEEPABLE);
181 	if (err)
182 		return err;
183 
184 	/* create index > 0 */
185 	ah->ah_num = ah->elem.index;
186 
187 	if (uresp) {
188 		/* only if new user provider */
189 		err = copy_to_user(&uresp->ah_num, &ah->ah_num,
190 					 sizeof(uresp->ah_num));
191 		if (err) {
192 			rxe_cleanup(ah);
193 			return -EFAULT;
194 		}
195 	} else if (ah->is_user) {
196 		/* only if old user provider */
197 		ah->ah_num = 0;
198 	}
199 
200 	rxe_init_av(init_attr->ah_attr, &ah->av);
201 	rxe_finalize(ah);
202 
203 	return 0;
204 }
205 
rxe_modify_ah(struct ib_ah * ibah,struct rdma_ah_attr * attr)206 static int rxe_modify_ah(struct ib_ah *ibah, struct rdma_ah_attr *attr)
207 {
208 	int err;
209 	struct rxe_dev *rxe = to_rdev(ibah->device);
210 	struct rxe_ah *ah = to_rah(ibah);
211 
212 	err = rxe_av_chk_attr(rxe, attr);
213 	if (err)
214 		return err;
215 
216 	rxe_init_av(attr, &ah->av);
217 	return 0;
218 }
219 
rxe_query_ah(struct ib_ah * ibah,struct rdma_ah_attr * attr)220 static int rxe_query_ah(struct ib_ah *ibah, struct rdma_ah_attr *attr)
221 {
222 	struct rxe_ah *ah = to_rah(ibah);
223 
224 	memset(attr, 0, sizeof(*attr));
225 	attr->type = ibah->type;
226 	rxe_av_to_attr(&ah->av, attr);
227 	return 0;
228 }
229 
rxe_destroy_ah(struct ib_ah * ibah,u32 flags)230 static int rxe_destroy_ah(struct ib_ah *ibah, u32 flags)
231 {
232 	struct rxe_ah *ah = to_rah(ibah);
233 
234 	rxe_cleanup_ah(ah, flags & RDMA_DESTROY_AH_SLEEPABLE);
235 
236 	return 0;
237 }
238 
post_one_recv(struct rxe_rq * rq,const struct ib_recv_wr * ibwr)239 static int post_one_recv(struct rxe_rq *rq, const struct ib_recv_wr *ibwr)
240 {
241 	int err;
242 	int i;
243 	u32 length;
244 	struct rxe_recv_wqe *recv_wqe;
245 	int num_sge = ibwr->num_sge;
246 	int full;
247 
248 	full = queue_full(rq->queue, QUEUE_TYPE_TO_DRIVER);
249 	if (unlikely(full)) {
250 		err = -ENOMEM;
251 		goto err1;
252 	}
253 
254 	if (unlikely(num_sge > rq->max_sge)) {
255 		err = -EINVAL;
256 		goto err1;
257 	}
258 
259 	length = 0;
260 	for (i = 0; i < num_sge; i++)
261 		length += ibwr->sg_list[i].length;
262 
263 	recv_wqe = queue_producer_addr(rq->queue, QUEUE_TYPE_TO_DRIVER);
264 	recv_wqe->wr_id = ibwr->wr_id;
265 
266 	memcpy(recv_wqe->dma.sge, ibwr->sg_list,
267 	       num_sge * sizeof(struct ib_sge));
268 
269 	recv_wqe->dma.length		= length;
270 	recv_wqe->dma.resid		= length;
271 	recv_wqe->dma.num_sge		= num_sge;
272 	recv_wqe->dma.cur_sge		= 0;
273 	recv_wqe->dma.sge_offset	= 0;
274 
275 	queue_advance_producer(rq->queue, QUEUE_TYPE_TO_DRIVER);
276 
277 	return 0;
278 
279 err1:
280 	return err;
281 }
282 
rxe_create_srq(struct ib_srq * ibsrq,struct ib_srq_init_attr * init,struct ib_udata * udata)283 static int rxe_create_srq(struct ib_srq *ibsrq, struct ib_srq_init_attr *init,
284 			  struct ib_udata *udata)
285 {
286 	int err;
287 	struct rxe_dev *rxe = to_rdev(ibsrq->device);
288 	struct rxe_pd *pd = to_rpd(ibsrq->pd);
289 	struct rxe_srq *srq = to_rsrq(ibsrq);
290 	struct rxe_create_srq_resp __user *uresp = NULL;
291 
292 	if (udata) {
293 		if (udata->outlen < sizeof(*uresp))
294 			return -EINVAL;
295 		uresp = udata->outbuf;
296 	}
297 
298 	if (init->srq_type != IB_SRQT_BASIC)
299 		return -EOPNOTSUPP;
300 
301 	err = rxe_srq_chk_init(rxe, init);
302 	if (err)
303 		return err;
304 
305 	err = rxe_add_to_pool(&rxe->srq_pool, srq);
306 	if (err)
307 		return err;
308 
309 	rxe_get(pd);
310 	srq->pd = pd;
311 
312 	err = rxe_srq_from_init(rxe, srq, init, udata, uresp);
313 	if (err)
314 		goto err_cleanup;
315 
316 	return 0;
317 
318 err_cleanup:
319 	rxe_cleanup(srq);
320 
321 	return err;
322 }
323 
rxe_modify_srq(struct ib_srq * ibsrq,struct ib_srq_attr * attr,enum ib_srq_attr_mask mask,struct ib_udata * udata)324 static int rxe_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
325 			  enum ib_srq_attr_mask mask,
326 			  struct ib_udata *udata)
327 {
328 	int err;
329 	struct rxe_srq *srq = to_rsrq(ibsrq);
330 	struct rxe_dev *rxe = to_rdev(ibsrq->device);
331 	struct rxe_modify_srq_cmd ucmd = {};
332 
333 	if (udata) {
334 		if (udata->inlen < sizeof(ucmd))
335 			return -EINVAL;
336 
337 		err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
338 		if (err)
339 			return err;
340 	}
341 
342 	err = rxe_srq_chk_attr(rxe, srq, attr, mask);
343 	if (err)
344 		return err;
345 
346 	err = rxe_srq_from_attr(rxe, srq, attr, mask, &ucmd, udata);
347 	if (err)
348 		return err;
349 	return 0;
350 }
351 
rxe_query_srq(struct ib_srq * ibsrq,struct ib_srq_attr * attr)352 static int rxe_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr)
353 {
354 	struct rxe_srq *srq = to_rsrq(ibsrq);
355 
356 	if (srq->error)
357 		return -EINVAL;
358 
359 	attr->max_wr = srq->rq.queue->buf->index_mask;
360 	attr->max_sge = srq->rq.max_sge;
361 	attr->srq_limit = srq->limit;
362 	return 0;
363 }
364 
rxe_destroy_srq(struct ib_srq * ibsrq,struct ib_udata * udata)365 static int rxe_destroy_srq(struct ib_srq *ibsrq, struct ib_udata *udata)
366 {
367 	struct rxe_srq *srq = to_rsrq(ibsrq);
368 
369 	rxe_cleanup(srq);
370 	return 0;
371 }
372 
rxe_post_srq_recv(struct ib_srq * ibsrq,const struct ib_recv_wr * wr,const struct ib_recv_wr ** bad_wr)373 static int rxe_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
374 			     const struct ib_recv_wr **bad_wr)
375 {
376 	int err = 0;
377 	struct rxe_srq *srq = to_rsrq(ibsrq);
378 	unsigned long flags;
379 
380 	spin_lock_irqsave(&srq->rq.producer_lock, flags);
381 
382 	while (wr) {
383 		err = post_one_recv(&srq->rq, wr);
384 		if (unlikely(err))
385 			break;
386 		wr = wr->next;
387 	}
388 
389 	spin_unlock_irqrestore(&srq->rq.producer_lock, flags);
390 
391 	if (err)
392 		*bad_wr = wr;
393 
394 	return err;
395 }
396 
rxe_create_qp(struct ib_qp * ibqp,struct ib_qp_init_attr * init,struct ib_udata * udata)397 static int rxe_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *init,
398 			 struct ib_udata *udata)
399 {
400 	int err;
401 	struct rxe_dev *rxe = to_rdev(ibqp->device);
402 	struct rxe_pd *pd = to_rpd(ibqp->pd);
403 	struct rxe_qp *qp = to_rqp(ibqp);
404 	struct rxe_create_qp_resp __user *uresp = NULL;
405 
406 	if (udata) {
407 		if (udata->outlen < sizeof(*uresp))
408 			return -EINVAL;
409 		uresp = udata->outbuf;
410 	}
411 
412 	if (init->create_flags)
413 		return -EOPNOTSUPP;
414 
415 	err = rxe_qp_chk_init(rxe, init);
416 	if (err)
417 		return err;
418 
419 	if (udata) {
420 		if (udata->inlen)
421 			return -EINVAL;
422 
423 		qp->is_user = true;
424 	} else {
425 		qp->is_user = false;
426 	}
427 
428 	err = rxe_add_to_pool(&rxe->qp_pool, qp);
429 	if (err)
430 		return err;
431 
432 	err = rxe_qp_from_init(rxe, qp, pd, init, uresp, ibqp->pd, udata);
433 	if (err)
434 		goto qp_init;
435 
436 	rxe_finalize(qp);
437 	return 0;
438 
439 qp_init:
440 	rxe_cleanup(qp);
441 	return err;
442 }
443 
rxe_modify_qp(struct ib_qp * ibqp,struct ib_qp_attr * attr,int mask,struct ib_udata * udata)444 static int rxe_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
445 			 int mask, struct ib_udata *udata)
446 {
447 	int err;
448 	struct rxe_dev *rxe = to_rdev(ibqp->device);
449 	struct rxe_qp *qp = to_rqp(ibqp);
450 
451 	if (mask & ~IB_QP_ATTR_STANDARD_BITS)
452 		return -EOPNOTSUPP;
453 
454 	err = rxe_qp_chk_attr(rxe, qp, attr, mask);
455 	if (err)
456 		goto err1;
457 
458 	err = rxe_qp_from_attr(qp, attr, mask, udata);
459 	if (err)
460 		goto err1;
461 
462 	if ((mask & IB_QP_AV) && (attr->ah_attr.ah_flags & IB_AH_GRH))
463 		qp->src_port = rdma_get_udp_sport(attr->ah_attr.grh.flow_label,
464 						  qp->ibqp.qp_num,
465 						  qp->attr.dest_qp_num);
466 
467 	return 0;
468 
469 err1:
470 	return err;
471 }
472 
rxe_query_qp(struct ib_qp * ibqp,struct ib_qp_attr * attr,int mask,struct ib_qp_init_attr * init)473 static int rxe_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
474 			int mask, struct ib_qp_init_attr *init)
475 {
476 	struct rxe_qp *qp = to_rqp(ibqp);
477 
478 	rxe_qp_to_init(qp, init);
479 	rxe_qp_to_attr(qp, attr, mask);
480 
481 	return 0;
482 }
483 
rxe_destroy_qp(struct ib_qp * ibqp,struct ib_udata * udata)484 static int rxe_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
485 {
486 	struct rxe_qp *qp = to_rqp(ibqp);
487 	int ret;
488 
489 	ret = rxe_qp_chk_destroy(qp);
490 	if (ret)
491 		return ret;
492 
493 	rxe_cleanup(qp);
494 	return 0;
495 }
496 
validate_send_wr(struct rxe_qp * qp,const struct ib_send_wr * ibwr,unsigned int mask,unsigned int length)497 static int validate_send_wr(struct rxe_qp *qp, const struct ib_send_wr *ibwr,
498 			    unsigned int mask, unsigned int length)
499 {
500 	int num_sge = ibwr->num_sge;
501 	struct rxe_sq *sq = &qp->sq;
502 
503 	if (unlikely(num_sge > sq->max_sge))
504 		goto err1;
505 
506 	if (unlikely(mask & WR_ATOMIC_MASK)) {
507 		if (length < 8)
508 			goto err1;
509 
510 		if (atomic_wr(ibwr)->remote_addr & 0x7)
511 			goto err1;
512 	}
513 
514 	if (unlikely((ibwr->send_flags & IB_SEND_INLINE) &&
515 		     (length > sq->max_inline)))
516 		goto err1;
517 
518 	return 0;
519 
520 err1:
521 	return -EINVAL;
522 }
523 
init_send_wr(struct rxe_qp * qp,struct rxe_send_wr * wr,const struct ib_send_wr * ibwr)524 static void init_send_wr(struct rxe_qp *qp, struct rxe_send_wr *wr,
525 			 const struct ib_send_wr *ibwr)
526 {
527 	wr->wr_id = ibwr->wr_id;
528 	wr->opcode = ibwr->opcode;
529 	wr->send_flags = ibwr->send_flags;
530 
531 	if (qp_type(qp) == IB_QPT_UD ||
532 	    qp_type(qp) == IB_QPT_GSI) {
533 		struct ib_ah *ibah = ud_wr(ibwr)->ah;
534 
535 		wr->wr.ud.remote_qpn = ud_wr(ibwr)->remote_qpn;
536 		wr->wr.ud.remote_qkey = ud_wr(ibwr)->remote_qkey;
537 		wr->wr.ud.ah_num = to_rah(ibah)->ah_num;
538 		if (qp_type(qp) == IB_QPT_GSI)
539 			wr->wr.ud.pkey_index = ud_wr(ibwr)->pkey_index;
540 		if (wr->opcode == IB_WR_SEND_WITH_IMM)
541 			wr->ex.imm_data = ibwr->ex.imm_data;
542 	} else {
543 		switch (wr->opcode) {
544 		case IB_WR_RDMA_WRITE_WITH_IMM:
545 			wr->ex.imm_data = ibwr->ex.imm_data;
546 			fallthrough;
547 		case IB_WR_RDMA_READ:
548 		case IB_WR_RDMA_WRITE:
549 			wr->wr.rdma.remote_addr = rdma_wr(ibwr)->remote_addr;
550 			wr->wr.rdma.rkey	= rdma_wr(ibwr)->rkey;
551 			break;
552 		case IB_WR_SEND_WITH_IMM:
553 			wr->ex.imm_data = ibwr->ex.imm_data;
554 			break;
555 		case IB_WR_SEND_WITH_INV:
556 			wr->ex.invalidate_rkey = ibwr->ex.invalidate_rkey;
557 			break;
558 		case IB_WR_ATOMIC_CMP_AND_SWP:
559 		case IB_WR_ATOMIC_FETCH_AND_ADD:
560 			wr->wr.atomic.remote_addr =
561 				atomic_wr(ibwr)->remote_addr;
562 			wr->wr.atomic.compare_add =
563 				atomic_wr(ibwr)->compare_add;
564 			wr->wr.atomic.swap = atomic_wr(ibwr)->swap;
565 			wr->wr.atomic.rkey = atomic_wr(ibwr)->rkey;
566 			break;
567 		case IB_WR_LOCAL_INV:
568 			wr->ex.invalidate_rkey = ibwr->ex.invalidate_rkey;
569 		break;
570 		case IB_WR_REG_MR:
571 			wr->wr.reg.mr = reg_wr(ibwr)->mr;
572 			wr->wr.reg.key = reg_wr(ibwr)->key;
573 			wr->wr.reg.access = reg_wr(ibwr)->access;
574 		break;
575 		default:
576 			break;
577 		}
578 	}
579 }
580 
copy_inline_data_to_wqe(struct rxe_send_wqe * wqe,const struct ib_send_wr * ibwr)581 static void copy_inline_data_to_wqe(struct rxe_send_wqe *wqe,
582 				    const struct ib_send_wr *ibwr)
583 {
584 	struct ib_sge *sge = ibwr->sg_list;
585 	u8 *p = wqe->dma.inline_data;
586 	int i;
587 
588 	for (i = 0; i < ibwr->num_sge; i++, sge++) {
589 		memcpy(p, (void *)(uintptr_t)sge->addr, sge->length);
590 		p += sge->length;
591 	}
592 }
593 
init_send_wqe(struct rxe_qp * qp,const struct ib_send_wr * ibwr,unsigned int mask,unsigned int length,struct rxe_send_wqe * wqe)594 static void init_send_wqe(struct rxe_qp *qp, const struct ib_send_wr *ibwr,
595 			 unsigned int mask, unsigned int length,
596 			 struct rxe_send_wqe *wqe)
597 {
598 	int num_sge = ibwr->num_sge;
599 
600 	init_send_wr(qp, &wqe->wr, ibwr);
601 
602 	/* local operation */
603 	if (unlikely(mask & WR_LOCAL_OP_MASK)) {
604 		wqe->mask = mask;
605 		wqe->state = wqe_state_posted;
606 		return;
607 	}
608 
609 	if (unlikely(ibwr->send_flags & IB_SEND_INLINE))
610 		copy_inline_data_to_wqe(wqe, ibwr);
611 	else
612 		memcpy(wqe->dma.sge, ibwr->sg_list,
613 		       num_sge * sizeof(struct ib_sge));
614 
615 	wqe->iova = mask & WR_ATOMIC_MASK ? atomic_wr(ibwr)->remote_addr :
616 		mask & WR_READ_OR_WRITE_MASK ? rdma_wr(ibwr)->remote_addr : 0;
617 	wqe->mask		= mask;
618 	wqe->dma.length		= length;
619 	wqe->dma.resid		= length;
620 	wqe->dma.num_sge	= num_sge;
621 	wqe->dma.cur_sge	= 0;
622 	wqe->dma.sge_offset	= 0;
623 	wqe->state		= wqe_state_posted;
624 	wqe->ssn		= atomic_add_return(1, &qp->ssn);
625 }
626 
post_one_send(struct rxe_qp * qp,const struct ib_send_wr * ibwr,unsigned int mask,u32 length)627 static int post_one_send(struct rxe_qp *qp, const struct ib_send_wr *ibwr,
628 			 unsigned int mask, u32 length)
629 {
630 	int err;
631 	struct rxe_sq *sq = &qp->sq;
632 	struct rxe_send_wqe *send_wqe;
633 	unsigned long flags;
634 	int full;
635 
636 	err = validate_send_wr(qp, ibwr, mask, length);
637 	if (err)
638 		return err;
639 
640 	spin_lock_irqsave(&qp->sq.sq_lock, flags);
641 
642 	full = queue_full(sq->queue, QUEUE_TYPE_TO_DRIVER);
643 
644 	if (unlikely(full)) {
645 		spin_unlock_irqrestore(&qp->sq.sq_lock, flags);
646 		return -ENOMEM;
647 	}
648 
649 	send_wqe = queue_producer_addr(sq->queue, QUEUE_TYPE_TO_DRIVER);
650 	init_send_wqe(qp, ibwr, mask, length, send_wqe);
651 
652 	queue_advance_producer(sq->queue, QUEUE_TYPE_TO_DRIVER);
653 
654 	spin_unlock_irqrestore(&qp->sq.sq_lock, flags);
655 
656 	return 0;
657 }
658 
rxe_post_send_kernel(struct rxe_qp * qp,const struct ib_send_wr * wr,const struct ib_send_wr ** bad_wr)659 static int rxe_post_send_kernel(struct rxe_qp *qp, const struct ib_send_wr *wr,
660 				const struct ib_send_wr **bad_wr)
661 {
662 	int err = 0;
663 	unsigned int mask;
664 	unsigned int length = 0;
665 	int i;
666 	struct ib_send_wr *next;
667 
668 	while (wr) {
669 		mask = wr_opcode_mask(wr->opcode, qp);
670 		if (unlikely(!mask)) {
671 			err = -EINVAL;
672 			*bad_wr = wr;
673 			break;
674 		}
675 
676 		if (unlikely((wr->send_flags & IB_SEND_INLINE) &&
677 			     !(mask & WR_INLINE_MASK))) {
678 			err = -EINVAL;
679 			*bad_wr = wr;
680 			break;
681 		}
682 
683 		next = wr->next;
684 
685 		length = 0;
686 		for (i = 0; i < wr->num_sge; i++)
687 			length += wr->sg_list[i].length;
688 
689 		err = post_one_send(qp, wr, mask, length);
690 
691 		if (err) {
692 			*bad_wr = wr;
693 			break;
694 		}
695 		wr = next;
696 	}
697 
698 	rxe_run_task(&qp->req.task, 1);
699 	if (unlikely(qp->req.state == QP_STATE_ERROR))
700 		rxe_run_task(&qp->comp.task, 1);
701 
702 	return err;
703 }
704 
rxe_post_send(struct ib_qp * ibqp,const struct ib_send_wr * wr,const struct ib_send_wr ** bad_wr)705 static int rxe_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
706 			 const struct ib_send_wr **bad_wr)
707 {
708 	struct rxe_qp *qp = to_rqp(ibqp);
709 
710 	if (unlikely(!qp->valid)) {
711 		*bad_wr = wr;
712 		return -EINVAL;
713 	}
714 
715 	if (unlikely(qp->req.state < QP_STATE_READY)) {
716 		*bad_wr = wr;
717 		return -EINVAL;
718 	}
719 
720 	if (qp->is_user) {
721 		/* Utilize process context to do protocol processing */
722 		rxe_run_task(&qp->req.task, 0);
723 		return 0;
724 	} else
725 		return rxe_post_send_kernel(qp, wr, bad_wr);
726 }
727 
rxe_post_recv(struct ib_qp * ibqp,const struct ib_recv_wr * wr,const struct ib_recv_wr ** bad_wr)728 static int rxe_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
729 			 const struct ib_recv_wr **bad_wr)
730 {
731 	int err = 0;
732 	struct rxe_qp *qp = to_rqp(ibqp);
733 	struct rxe_rq *rq = &qp->rq;
734 	unsigned long flags;
735 
736 	if (unlikely((qp_state(qp) < IB_QPS_INIT) || !qp->valid)) {
737 		*bad_wr = wr;
738 		err = -EINVAL;
739 		goto err1;
740 	}
741 
742 	if (unlikely(qp->srq)) {
743 		*bad_wr = wr;
744 		err = -EINVAL;
745 		goto err1;
746 	}
747 
748 	spin_lock_irqsave(&rq->producer_lock, flags);
749 
750 	while (wr) {
751 		err = post_one_recv(rq, wr);
752 		if (unlikely(err)) {
753 			*bad_wr = wr;
754 			break;
755 		}
756 		wr = wr->next;
757 	}
758 
759 	spin_unlock_irqrestore(&rq->producer_lock, flags);
760 
761 	if (qp->resp.state == QP_STATE_ERROR)
762 		rxe_run_task(&qp->resp.task, 1);
763 
764 err1:
765 	return err;
766 }
767 
rxe_create_cq(struct ib_cq * ibcq,const struct ib_cq_init_attr * attr,struct ib_udata * udata)768 static int rxe_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
769 			 struct ib_udata *udata)
770 {
771 	int err;
772 	struct ib_device *dev = ibcq->device;
773 	struct rxe_dev *rxe = to_rdev(dev);
774 	struct rxe_cq *cq = to_rcq(ibcq);
775 	struct rxe_create_cq_resp __user *uresp = NULL;
776 
777 	if (udata) {
778 		if (udata->outlen < sizeof(*uresp))
779 			return -EINVAL;
780 		uresp = udata->outbuf;
781 	}
782 
783 	if (attr->flags)
784 		return -EOPNOTSUPP;
785 
786 	err = rxe_cq_chk_attr(rxe, NULL, attr->cqe, attr->comp_vector);
787 	if (err)
788 		return err;
789 
790 	err = rxe_cq_from_init(rxe, cq, attr->cqe, attr->comp_vector, udata,
791 			       uresp);
792 	if (err)
793 		return err;
794 
795 	return rxe_add_to_pool(&rxe->cq_pool, cq);
796 }
797 
rxe_destroy_cq(struct ib_cq * ibcq,struct ib_udata * udata)798 static int rxe_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
799 {
800 	struct rxe_cq *cq = to_rcq(ibcq);
801 
802 	/* See IBA C11-17: The CI shall return an error if this Verb is
803 	 * invoked while a Work Queue is still associated with the CQ.
804 	 */
805 	if (atomic_read(&cq->num_wq))
806 		return -EINVAL;
807 
808 	rxe_cq_disable(cq);
809 
810 	rxe_cleanup(cq);
811 	return 0;
812 }
813 
rxe_resize_cq(struct ib_cq * ibcq,int cqe,struct ib_udata * udata)814 static int rxe_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata)
815 {
816 	int err;
817 	struct rxe_cq *cq = to_rcq(ibcq);
818 	struct rxe_dev *rxe = to_rdev(ibcq->device);
819 	struct rxe_resize_cq_resp __user *uresp = NULL;
820 
821 	if (udata) {
822 		if (udata->outlen < sizeof(*uresp))
823 			return -EINVAL;
824 		uresp = udata->outbuf;
825 	}
826 
827 	err = rxe_cq_chk_attr(rxe, cq, cqe, 0);
828 	if (err)
829 		goto err1;
830 
831 	err = rxe_cq_resize_queue(cq, cqe, uresp, udata);
832 	if (err)
833 		goto err1;
834 
835 	return 0;
836 
837 err1:
838 	return err;
839 }
840 
rxe_poll_cq(struct ib_cq * ibcq,int num_entries,struct ib_wc * wc)841 static int rxe_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
842 {
843 	int i;
844 	struct rxe_cq *cq = to_rcq(ibcq);
845 	struct rxe_cqe *cqe;
846 	unsigned long flags;
847 
848 	spin_lock_irqsave(&cq->cq_lock, flags);
849 	for (i = 0; i < num_entries; i++) {
850 		cqe = queue_head(cq->queue, QUEUE_TYPE_FROM_DRIVER);
851 		if (!cqe)
852 			break;
853 
854 		memcpy(wc++, &cqe->ibwc, sizeof(*wc));
855 		queue_advance_consumer(cq->queue, QUEUE_TYPE_FROM_DRIVER);
856 	}
857 	spin_unlock_irqrestore(&cq->cq_lock, flags);
858 
859 	return i;
860 }
861 
rxe_peek_cq(struct ib_cq * ibcq,int wc_cnt)862 static int rxe_peek_cq(struct ib_cq *ibcq, int wc_cnt)
863 {
864 	struct rxe_cq *cq = to_rcq(ibcq);
865 	int count;
866 
867 	count = queue_count(cq->queue, QUEUE_TYPE_FROM_DRIVER);
868 
869 	return (count > wc_cnt) ? wc_cnt : count;
870 }
871 
rxe_req_notify_cq(struct ib_cq * ibcq,enum ib_cq_notify_flags flags)872 static int rxe_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
873 {
874 	struct rxe_cq *cq = to_rcq(ibcq);
875 	int ret = 0;
876 	int empty;
877 	unsigned long irq_flags;
878 
879 	spin_lock_irqsave(&cq->cq_lock, irq_flags);
880 	if (cq->notify != IB_CQ_NEXT_COMP)
881 		cq->notify = flags & IB_CQ_SOLICITED_MASK;
882 
883 	empty = queue_empty(cq->queue, QUEUE_TYPE_FROM_DRIVER);
884 
885 	if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !empty)
886 		ret = 1;
887 
888 	spin_unlock_irqrestore(&cq->cq_lock, irq_flags);
889 
890 	return ret;
891 }
892 
rxe_get_dma_mr(struct ib_pd * ibpd,int access)893 static struct ib_mr *rxe_get_dma_mr(struct ib_pd *ibpd, int access)
894 {
895 	struct rxe_dev *rxe = to_rdev(ibpd->device);
896 	struct rxe_pd *pd = to_rpd(ibpd);
897 	struct rxe_mr *mr;
898 
899 	mr = rxe_alloc(&rxe->mr_pool);
900 	if (!mr)
901 		return ERR_PTR(-ENOMEM);
902 
903 	rxe_get(pd);
904 	mr->ibmr.pd = ibpd;
905 
906 	rxe_mr_init_dma(access, mr);
907 	rxe_finalize(mr);
908 
909 	return &mr->ibmr;
910 }
911 
rxe_reg_user_mr(struct ib_pd * ibpd,u64 start,u64 length,u64 iova,int access,struct ib_udata * udata)912 static struct ib_mr *rxe_reg_user_mr(struct ib_pd *ibpd,
913 				     u64 start,
914 				     u64 length,
915 				     u64 iova,
916 				     int access, struct ib_udata *udata)
917 {
918 	int err;
919 	struct rxe_dev *rxe = to_rdev(ibpd->device);
920 	struct rxe_pd *pd = to_rpd(ibpd);
921 	struct rxe_mr *mr;
922 
923 	mr = rxe_alloc(&rxe->mr_pool);
924 	if (!mr) {
925 		err = -ENOMEM;
926 		goto err2;
927 	}
928 
929 
930 	rxe_get(pd);
931 	mr->ibmr.pd = ibpd;
932 
933 	err = rxe_mr_init_user(rxe, start, length, iova, access, mr);
934 	if (err)
935 		goto err3;
936 
937 	rxe_finalize(mr);
938 
939 	return &mr->ibmr;
940 
941 err3:
942 	rxe_cleanup(mr);
943 err2:
944 	return ERR_PTR(err);
945 }
946 
rxe_alloc_mr(struct ib_pd * ibpd,enum ib_mr_type mr_type,u32 max_num_sg)947 static struct ib_mr *rxe_alloc_mr(struct ib_pd *ibpd, enum ib_mr_type mr_type,
948 				  u32 max_num_sg)
949 {
950 	struct rxe_dev *rxe = to_rdev(ibpd->device);
951 	struct rxe_pd *pd = to_rpd(ibpd);
952 	struct rxe_mr *mr;
953 	int err;
954 
955 	if (mr_type != IB_MR_TYPE_MEM_REG)
956 		return ERR_PTR(-EINVAL);
957 
958 	mr = rxe_alloc(&rxe->mr_pool);
959 	if (!mr) {
960 		err = -ENOMEM;
961 		goto err1;
962 	}
963 
964 	rxe_get(pd);
965 	mr->ibmr.pd = ibpd;
966 
967 	err = rxe_mr_init_fast(max_num_sg, mr);
968 	if (err)
969 		goto err2;
970 
971 	rxe_finalize(mr);
972 
973 	return &mr->ibmr;
974 
975 err2:
976 	rxe_cleanup(mr);
977 err1:
978 	return ERR_PTR(err);
979 }
980 
rxe_set_page(struct ib_mr * ibmr,u64 addr)981 static int rxe_set_page(struct ib_mr *ibmr, u64 addr)
982 {
983 	struct rxe_mr *mr = to_rmr(ibmr);
984 	struct rxe_map *map;
985 	struct rxe_phys_buf *buf;
986 
987 	if (unlikely(mr->nbuf == mr->num_buf))
988 		return -ENOMEM;
989 
990 	map = mr->map[mr->nbuf / RXE_BUF_PER_MAP];
991 	buf = &map->buf[mr->nbuf % RXE_BUF_PER_MAP];
992 
993 	buf->addr = addr;
994 	buf->size = ibmr->page_size;
995 	mr->nbuf++;
996 
997 	return 0;
998 }
999 
rxe_map_mr_sg(struct ib_mr * ibmr,struct scatterlist * sg,int sg_nents,unsigned int * sg_offset)1000 static int rxe_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
1001 			 int sg_nents, unsigned int *sg_offset)
1002 {
1003 	struct rxe_mr *mr = to_rmr(ibmr);
1004 	int n;
1005 
1006 	mr->nbuf = 0;
1007 
1008 	n = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, rxe_set_page);
1009 
1010 	mr->page_shift = ilog2(ibmr->page_size);
1011 	mr->page_mask = ibmr->page_size - 1;
1012 	mr->offset = ibmr->iova & mr->page_mask;
1013 
1014 	return n;
1015 }
1016 
parent_show(struct device * device,struct device_attribute * attr,char * buf)1017 static ssize_t parent_show(struct device *device,
1018 			   struct device_attribute *attr, char *buf)
1019 {
1020 	struct rxe_dev *rxe =
1021 		rdma_device_to_drv_device(device, struct rxe_dev, ib_dev);
1022 
1023 	return sysfs_emit(buf, "%s\n", rxe_parent_name(rxe, 1));
1024 }
1025 
1026 static DEVICE_ATTR_RO(parent);
1027 
1028 static struct attribute *rxe_dev_attributes[] = {
1029 	&dev_attr_parent.attr,
1030 	NULL
1031 };
1032 
1033 static const struct attribute_group rxe_attr_group = {
1034 	.attrs = rxe_dev_attributes,
1035 };
1036 
rxe_enable_driver(struct ib_device * ib_dev)1037 static int rxe_enable_driver(struct ib_device *ib_dev)
1038 {
1039 	struct rxe_dev *rxe = container_of(ib_dev, struct rxe_dev, ib_dev);
1040 
1041 	rxe_set_port_state(rxe);
1042 	dev_info(&rxe->ib_dev.dev, "added %s\n", netdev_name(rxe->ndev));
1043 	return 0;
1044 }
1045 
1046 static const struct ib_device_ops rxe_dev_ops = {
1047 	.owner = THIS_MODULE,
1048 	.driver_id = RDMA_DRIVER_RXE,
1049 	.uverbs_abi_ver = RXE_UVERBS_ABI_VERSION,
1050 
1051 	.alloc_hw_port_stats = rxe_ib_alloc_hw_port_stats,
1052 	.alloc_mr = rxe_alloc_mr,
1053 	.alloc_mw = rxe_alloc_mw,
1054 	.alloc_pd = rxe_alloc_pd,
1055 	.alloc_ucontext = rxe_alloc_ucontext,
1056 	.attach_mcast = rxe_attach_mcast,
1057 	.create_ah = rxe_create_ah,
1058 	.create_cq = rxe_create_cq,
1059 	.create_qp = rxe_create_qp,
1060 	.create_srq = rxe_create_srq,
1061 	.create_user_ah = rxe_create_ah,
1062 	.dealloc_driver = rxe_dealloc,
1063 	.dealloc_mw = rxe_dealloc_mw,
1064 	.dealloc_pd = rxe_dealloc_pd,
1065 	.dealloc_ucontext = rxe_dealloc_ucontext,
1066 	.dereg_mr = rxe_dereg_mr,
1067 	.destroy_ah = rxe_destroy_ah,
1068 	.destroy_cq = rxe_destroy_cq,
1069 	.destroy_qp = rxe_destroy_qp,
1070 	.destroy_srq = rxe_destroy_srq,
1071 	.detach_mcast = rxe_detach_mcast,
1072 	.device_group = &rxe_attr_group,
1073 	.enable_driver = rxe_enable_driver,
1074 	.get_dma_mr = rxe_get_dma_mr,
1075 	.get_hw_stats = rxe_ib_get_hw_stats,
1076 	.get_link_layer = rxe_get_link_layer,
1077 	.get_port_immutable = rxe_port_immutable,
1078 	.map_mr_sg = rxe_map_mr_sg,
1079 	.mmap = rxe_mmap,
1080 	.modify_ah = rxe_modify_ah,
1081 	.modify_device = rxe_modify_device,
1082 	.modify_port = rxe_modify_port,
1083 	.modify_qp = rxe_modify_qp,
1084 	.modify_srq = rxe_modify_srq,
1085 	.peek_cq = rxe_peek_cq,
1086 	.poll_cq = rxe_poll_cq,
1087 	.post_recv = rxe_post_recv,
1088 	.post_send = rxe_post_send,
1089 	.post_srq_recv = rxe_post_srq_recv,
1090 	.query_ah = rxe_query_ah,
1091 	.query_device = rxe_query_device,
1092 	.query_pkey = rxe_query_pkey,
1093 	.query_port = rxe_query_port,
1094 	.query_qp = rxe_query_qp,
1095 	.query_srq = rxe_query_srq,
1096 	.reg_user_mr = rxe_reg_user_mr,
1097 	.req_notify_cq = rxe_req_notify_cq,
1098 	.resize_cq = rxe_resize_cq,
1099 
1100 	INIT_RDMA_OBJ_SIZE(ib_ah, rxe_ah, ibah),
1101 	INIT_RDMA_OBJ_SIZE(ib_cq, rxe_cq, ibcq),
1102 	INIT_RDMA_OBJ_SIZE(ib_pd, rxe_pd, ibpd),
1103 	INIT_RDMA_OBJ_SIZE(ib_qp, rxe_qp, ibqp),
1104 	INIT_RDMA_OBJ_SIZE(ib_srq, rxe_srq, ibsrq),
1105 	INIT_RDMA_OBJ_SIZE(ib_ucontext, rxe_ucontext, ibuc),
1106 	INIT_RDMA_OBJ_SIZE(ib_mw, rxe_mw, ibmw),
1107 };
1108 
rxe_register_device(struct rxe_dev * rxe,const char * ibdev_name)1109 int rxe_register_device(struct rxe_dev *rxe, const char *ibdev_name)
1110 {
1111 	int err;
1112 	struct ib_device *dev = &rxe->ib_dev;
1113 
1114 	strscpy(dev->node_desc, "rxe", sizeof(dev->node_desc));
1115 
1116 	dev->node_type = RDMA_NODE_IB_CA;
1117 	dev->phys_port_cnt = 1;
1118 	dev->num_comp_vectors = num_possible_cpus();
1119 	dev->local_dma_lkey = 0;
1120 	addrconf_addr_eui48((unsigned char *)&dev->node_guid,
1121 			    rxe->ndev->dev_addr);
1122 
1123 	dev->uverbs_cmd_mask |= BIT_ULL(IB_USER_VERBS_CMD_POST_SEND) |
1124 				BIT_ULL(IB_USER_VERBS_CMD_REQ_NOTIFY_CQ);
1125 
1126 	ib_set_device_ops(dev, &rxe_dev_ops);
1127 	err = ib_device_set_netdev(&rxe->ib_dev, rxe->ndev, 1);
1128 	if (err)
1129 		return err;
1130 
1131 	err = rxe_icrc_init(rxe);
1132 	if (err)
1133 		return err;
1134 
1135 	err = ib_register_device(dev, ibdev_name, NULL);
1136 	if (err)
1137 		pr_warn("%s failed with error %d\n", __func__, err);
1138 
1139 	/*
1140 	 * Note that rxe may be invalid at this point if another thread
1141 	 * unregistered it.
1142 	 */
1143 	return err;
1144 }
1145