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
15 static int post_one_recv(struct rxe_rq *rq, const struct ib_recv_wr *ibwr);
16
17 /* dev */
rxe_query_device(struct ib_device * ibdev,struct ib_device_attr * attr,struct ib_udata * udata)18 static int rxe_query_device(struct ib_device *ibdev,
19 struct ib_device_attr *attr,
20 struct ib_udata *udata)
21 {
22 struct rxe_dev *rxe = to_rdev(ibdev);
23 int err;
24
25 err = ib_no_udata_io(udata);
26 if (err)
27 return err;
28
29 memcpy(attr, &rxe->attr, sizeof(*attr));
30
31 return 0;
32 }
33
rxe_query_port(struct ib_device * ibdev,u32 port_num,struct ib_port_attr * attr)34 static int rxe_query_port(struct ib_device *ibdev,
35 u32 port_num, struct ib_port_attr *attr)
36 {
37 struct rxe_dev *rxe = to_rdev(ibdev);
38 struct net_device *ndev;
39 int err, ret;
40
41 if (port_num != 1) {
42 err = -EINVAL;
43 rxe_dbg_dev(rxe, "bad port_num = %d\n", port_num);
44 goto err_out;
45 }
46
47 ndev = ib_device_get_netdev(ibdev, 1);
48 if (!ndev) {
49 err = -ENODEV;
50 goto err_out;
51 }
52
53 memcpy(attr, &rxe->port.attr, sizeof(*attr));
54
55 mutex_lock(&rxe->usdev_lock);
56 ret = ib_get_eth_speed(ibdev, port_num, &attr->active_speed,
57 &attr->active_width);
58
59 attr->state = ib_get_curr_port_state(ndev);
60 if (attr->state == IB_PORT_ACTIVE)
61 attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
62 else if (netif_get_flags(ndev) & IFF_UP)
63 attr->phys_state = IB_PORT_PHYS_STATE_POLLING;
64 else
65 attr->phys_state = IB_PORT_PHYS_STATE_DISABLED;
66
67 mutex_unlock(&rxe->usdev_lock);
68
69 dev_put(ndev);
70 return ret;
71
72 err_out:
73 rxe_err_dev(rxe, "returned err = %d\n", err);
74 return err;
75 }
76
rxe_query_gid(struct ib_device * ibdev,u32 port,int idx,union ib_gid * gid)77 static int rxe_query_gid(struct ib_device *ibdev, u32 port, int idx,
78 union ib_gid *gid)
79 {
80 struct rxe_dev *rxe = to_rdev(ibdev);
81
82 /* subnet_prefix == interface_id == 0; */
83 memset(gid, 0, sizeof(*gid));
84 memcpy(gid->raw, rxe->raw_gid, ETH_ALEN);
85
86 return 0;
87 }
88
rxe_query_pkey(struct ib_device * ibdev,u32 port_num,u16 index,u16 * pkey)89 static int rxe_query_pkey(struct ib_device *ibdev,
90 u32 port_num, u16 index, u16 *pkey)
91 {
92 struct rxe_dev *rxe = to_rdev(ibdev);
93 int err;
94
95 if (index != 0) {
96 err = -EINVAL;
97 rxe_dbg_dev(rxe, "bad pkey index = %d\n", index);
98 goto err_out;
99 }
100
101 *pkey = IB_DEFAULT_PKEY_FULL;
102 return 0;
103
104 err_out:
105 rxe_err_dev(rxe, "returned err = %d\n", err);
106 return err;
107 }
108
rxe_modify_device(struct ib_device * ibdev,int mask,struct ib_device_modify * attr)109 static int rxe_modify_device(struct ib_device *ibdev,
110 int mask, struct ib_device_modify *attr)
111 {
112 struct rxe_dev *rxe = to_rdev(ibdev);
113 int err;
114
115 if (mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
116 IB_DEVICE_MODIFY_NODE_DESC)) {
117 err = -EOPNOTSUPP;
118 rxe_dbg_dev(rxe, "unsupported mask = 0x%x\n", mask);
119 goto err_out;
120 }
121
122 if (mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID)
123 rxe->attr.sys_image_guid = cpu_to_be64(attr->sys_image_guid);
124
125 if (mask & IB_DEVICE_MODIFY_NODE_DESC) {
126 memcpy(rxe->ib_dev.node_desc,
127 attr->node_desc, sizeof(rxe->ib_dev.node_desc));
128 }
129
130 return 0;
131
132 err_out:
133 rxe_err_dev(rxe, "returned err = %d\n", err);
134 return err;
135 }
136
rxe_modify_port(struct ib_device * ibdev,u32 port_num,int mask,struct ib_port_modify * attr)137 static int rxe_modify_port(struct ib_device *ibdev, u32 port_num,
138 int mask, struct ib_port_modify *attr)
139 {
140 struct rxe_dev *rxe = to_rdev(ibdev);
141 struct rxe_port *port;
142 int err;
143
144 if (port_num != 1) {
145 err = -EINVAL;
146 rxe_dbg_dev(rxe, "bad port_num = %d\n", port_num);
147 goto err_out;
148 }
149
150 //TODO is shutdown useful
151 if (mask & ~(IB_PORT_RESET_QKEY_CNTR)) {
152 err = -EOPNOTSUPP;
153 rxe_dbg_dev(rxe, "unsupported mask = 0x%x\n", mask);
154 goto err_out;
155 }
156
157 port = &rxe->port;
158 port->attr.port_cap_flags |= attr->set_port_cap_mask;
159 port->attr.port_cap_flags &= ~attr->clr_port_cap_mask;
160
161 if (mask & IB_PORT_RESET_QKEY_CNTR)
162 port->attr.qkey_viol_cntr = 0;
163
164 return 0;
165
166 err_out:
167 rxe_err_dev(rxe, "returned err = %d\n", err);
168 return err;
169 }
170
rxe_get_link_layer(struct ib_device * ibdev,u32 port_num)171 static enum rdma_link_layer rxe_get_link_layer(struct ib_device *ibdev,
172 u32 port_num)
173 {
174 struct rxe_dev *rxe = to_rdev(ibdev);
175 int err;
176
177 if (port_num != 1) {
178 err = -EINVAL;
179 rxe_dbg_dev(rxe, "bad port_num = %d\n", port_num);
180 goto err_out;
181 }
182
183 return IB_LINK_LAYER_ETHERNET;
184
185 err_out:
186 rxe_err_dev(rxe, "returned err = %d\n", err);
187 return err;
188 }
189
rxe_port_immutable(struct ib_device * ibdev,u32 port_num,struct ib_port_immutable * immutable)190 static int rxe_port_immutable(struct ib_device *ibdev, u32 port_num,
191 struct ib_port_immutable *immutable)
192 {
193 struct rxe_dev *rxe = to_rdev(ibdev);
194 struct ib_port_attr attr = {};
195 int err;
196
197 if (port_num != 1) {
198 err = -EINVAL;
199 rxe_dbg_dev(rxe, "bad port_num = %d\n", port_num);
200 goto err_out;
201 }
202
203 err = ib_query_port(ibdev, port_num, &attr);
204 if (err)
205 goto err_out;
206
207 immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
208 immutable->pkey_tbl_len = attr.pkey_tbl_len;
209 immutable->gid_tbl_len = attr.gid_tbl_len;
210 immutable->max_mad_size = IB_MGMT_MAD_SIZE;
211
212 return 0;
213
214 err_out:
215 rxe_err_dev(rxe, "returned err = %d\n", err);
216 return err;
217 }
218
219 /* uc */
rxe_alloc_ucontext(struct ib_ucontext * ibuc,struct ib_udata * udata)220 static int rxe_alloc_ucontext(struct ib_ucontext *ibuc, struct ib_udata *udata)
221 {
222 struct rxe_dev *rxe = to_rdev(ibuc->device);
223 struct rxe_ucontext *uc = to_ruc(ibuc);
224 int err;
225
226 err = rxe_add_to_pool(&rxe->uc_pool, uc);
227 if (err)
228 rxe_err_dev(rxe, "unable to create uc\n");
229
230 return err;
231 }
232
rxe_dealloc_ucontext(struct ib_ucontext * ibuc)233 static void rxe_dealloc_ucontext(struct ib_ucontext *ibuc)
234 {
235 struct rxe_ucontext *uc = to_ruc(ibuc);
236 int err;
237
238 err = rxe_cleanup(uc);
239 if (err)
240 rxe_err_uc(uc, "cleanup failed, err = %d\n", err);
241 }
242
rxe_disassociate_ucontext(struct ib_ucontext * ibuc)243 static void rxe_disassociate_ucontext(struct ib_ucontext *ibuc)
244 {
245 }
246
247 /* pd */
rxe_alloc_pd(struct ib_pd * ibpd,struct ib_udata * udata)248 static int rxe_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
249 {
250 struct rxe_dev *rxe = to_rdev(ibpd->device);
251 struct rxe_pd *pd = to_rpd(ibpd);
252 int err;
253
254 err = rxe_add_to_pool(&rxe->pd_pool, pd);
255 if (err) {
256 rxe_dbg_dev(rxe, "unable to alloc pd\n");
257 goto err_out;
258 }
259
260 return 0;
261
262 err_out:
263 rxe_err_dev(rxe, "returned err = %d\n", err);
264 return err;
265 }
266
rxe_dealloc_pd(struct ib_pd * ibpd,struct ib_udata * udata)267 static int rxe_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
268 {
269 struct rxe_pd *pd = to_rpd(ibpd);
270 int err;
271
272 err = rxe_cleanup(pd);
273 if (err)
274 rxe_err_pd(pd, "cleanup failed, err = %d\n", err);
275
276 return 0;
277 }
278
279 /* ah */
rxe_create_ah(struct ib_ah * ibah,struct rdma_ah_init_attr * init_attr,struct ib_udata * udata)280 static int rxe_create_ah(struct ib_ah *ibah,
281 struct rdma_ah_init_attr *init_attr,
282 struct ib_udata *udata)
283 {
284 struct rxe_dev *rxe = to_rdev(ibah->device);
285 struct rxe_ah *ah = to_rah(ibah);
286 struct rxe_create_ah_resp __user *uresp = NULL;
287 int err, cleanup_err;
288
289 if (udata) {
290 /* test if new user provider */
291 if (udata->outlen >= sizeof(*uresp))
292 uresp = udata->outbuf;
293 ah->is_user = true;
294 } else {
295 ah->is_user = false;
296 }
297
298 err = rxe_add_to_pool_ah(&rxe->ah_pool, ah,
299 init_attr->flags & RDMA_CREATE_AH_SLEEPABLE);
300 if (err) {
301 rxe_dbg_dev(rxe, "unable to create ah\n");
302 goto err_out;
303 }
304
305 /* create index > 0 */
306 ah->ah_num = ah->elem.index;
307
308 err = rxe_ah_chk_attr(ah, init_attr->ah_attr);
309 if (err) {
310 rxe_dbg_ah(ah, "bad attr\n");
311 goto err_cleanup;
312 }
313
314 if (uresp) {
315 /* only if new user provider */
316 err = copy_to_user(&uresp->ah_num, &ah->ah_num,
317 sizeof(uresp->ah_num));
318 if (err) {
319 err = -EFAULT;
320 rxe_dbg_ah(ah, "unable to copy to user\n");
321 goto err_cleanup;
322 }
323 } else if (ah->is_user) {
324 /* only if old user provider */
325 ah->ah_num = 0;
326 }
327
328 rxe_init_av(init_attr->ah_attr, &ah->av);
329 rxe_finalize(ah);
330
331 return 0;
332
333 err_cleanup:
334 cleanup_err = rxe_cleanup(ah);
335 if (cleanup_err)
336 rxe_err_ah(ah, "cleanup failed, err = %d\n", cleanup_err);
337 err_out:
338 rxe_err_ah(ah, "returned err = %d\n", err);
339 return err;
340 }
341
rxe_modify_ah(struct ib_ah * ibah,struct rdma_ah_attr * attr)342 static int rxe_modify_ah(struct ib_ah *ibah, struct rdma_ah_attr *attr)
343 {
344 struct rxe_ah *ah = to_rah(ibah);
345 int err;
346
347 err = rxe_ah_chk_attr(ah, attr);
348 if (err) {
349 rxe_dbg_ah(ah, "bad attr\n");
350 goto err_out;
351 }
352
353 rxe_init_av(attr, &ah->av);
354
355 return 0;
356
357 err_out:
358 rxe_err_ah(ah, "returned err = %d\n", err);
359 return err;
360 }
361
rxe_query_ah(struct ib_ah * ibah,struct rdma_ah_attr * attr)362 static int rxe_query_ah(struct ib_ah *ibah, struct rdma_ah_attr *attr)
363 {
364 struct rxe_ah *ah = to_rah(ibah);
365
366 memset(attr, 0, sizeof(*attr));
367 attr->type = ibah->type;
368 rxe_av_to_attr(&ah->av, attr);
369
370 return 0;
371 }
372
rxe_destroy_ah(struct ib_ah * ibah,u32 flags)373 static int rxe_destroy_ah(struct ib_ah *ibah, u32 flags)
374 {
375 struct rxe_ah *ah = to_rah(ibah);
376 int err;
377
378 err = rxe_cleanup_ah(ah, flags & RDMA_DESTROY_AH_SLEEPABLE);
379 if (err)
380 rxe_err_ah(ah, "cleanup failed, err = %d\n", err);
381
382 return 0;
383 }
384
385 /* srq */
rxe_create_srq(struct ib_srq * ibsrq,struct ib_srq_init_attr * init,struct ib_udata * udata)386 static int rxe_create_srq(struct ib_srq *ibsrq, struct ib_srq_init_attr *init,
387 struct ib_udata *udata)
388 {
389 struct rxe_dev *rxe = to_rdev(ibsrq->device);
390 struct rxe_pd *pd = to_rpd(ibsrq->pd);
391 struct rxe_srq *srq = to_rsrq(ibsrq);
392 struct rxe_create_srq_resp __user *uresp = NULL;
393 int err, cleanup_err;
394
395 if (udata) {
396 if (udata->outlen < sizeof(*uresp)) {
397 err = -EINVAL;
398 rxe_err_dev(rxe, "malformed udata\n");
399 goto err_out;
400 }
401 uresp = udata->outbuf;
402 }
403
404 if (init->srq_type != IB_SRQT_BASIC) {
405 err = -EOPNOTSUPP;
406 rxe_dbg_dev(rxe, "srq type = %d, not supported\n",
407 init->srq_type);
408 goto err_out;
409 }
410
411 err = rxe_srq_chk_init(rxe, init);
412 if (err) {
413 rxe_dbg_dev(rxe, "invalid init attributes\n");
414 goto err_out;
415 }
416
417 err = rxe_add_to_pool(&rxe->srq_pool, srq);
418 if (err) {
419 rxe_dbg_dev(rxe, "unable to create srq, err = %d\n", err);
420 goto err_out;
421 }
422
423 rxe_get(pd);
424 srq->pd = pd;
425
426 err = rxe_srq_from_init(rxe, srq, init, udata, uresp);
427 if (err) {
428 rxe_dbg_srq(srq, "create srq failed, err = %d\n", err);
429 goto err_cleanup;
430 }
431
432 return 0;
433
434 err_cleanup:
435 cleanup_err = rxe_cleanup(srq);
436 if (cleanup_err)
437 rxe_err_srq(srq, "cleanup failed, err = %d\n", cleanup_err);
438 err_out:
439 rxe_err_dev(rxe, "returned err = %d\n", err);
440 return err;
441 }
442
rxe_modify_srq(struct ib_srq * ibsrq,struct ib_srq_attr * attr,enum ib_srq_attr_mask mask,struct ib_udata * udata)443 static int rxe_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
444 enum ib_srq_attr_mask mask,
445 struct ib_udata *udata)
446 {
447 struct rxe_srq *srq = to_rsrq(ibsrq);
448 struct rxe_dev *rxe = to_rdev(ibsrq->device);
449 struct rxe_modify_srq_cmd cmd = {};
450 int err;
451
452 if (udata) {
453 err = ib_copy_validate_udata_in(udata, cmd, mmap_info_addr);
454 if (err)
455 goto err_out;
456 }
457
458 err = rxe_srq_chk_attr(rxe, srq, attr, mask);
459 if (err) {
460 rxe_dbg_srq(srq, "bad init attributes\n");
461 goto err_out;
462 }
463
464 err = rxe_srq_from_attr(rxe, srq, attr, mask, &cmd, udata);
465 if (err) {
466 rxe_dbg_srq(srq, "bad attr\n");
467 goto err_out;
468 }
469
470 return 0;
471
472 err_out:
473 rxe_err_srq(srq, "returned err = %d\n", err);
474 return err;
475 }
476
rxe_query_srq(struct ib_srq * ibsrq,struct ib_srq_attr * attr)477 static int rxe_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr)
478 {
479 struct rxe_srq *srq = to_rsrq(ibsrq);
480 int err;
481
482 if (srq->error) {
483 err = -EINVAL;
484 rxe_dbg_srq(srq, "srq in error state\n");
485 goto err_out;
486 }
487
488 attr->max_wr = srq->rq.queue->buf->index_mask;
489 attr->max_sge = srq->rq.max_sge;
490 attr->srq_limit = srq->limit;
491 return 0;
492
493 err_out:
494 rxe_err_srq(srq, "returned err = %d\n", err);
495 return err;
496 }
497
rxe_post_srq_recv(struct ib_srq * ibsrq,const struct ib_recv_wr * wr,const struct ib_recv_wr ** bad_wr)498 static int rxe_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
499 const struct ib_recv_wr **bad_wr)
500 {
501 int err = 0;
502 struct rxe_srq *srq = to_rsrq(ibsrq);
503 unsigned long flags;
504
505 spin_lock_irqsave(&srq->rq.producer_lock, flags);
506
507 while (wr) {
508 err = post_one_recv(&srq->rq, wr);
509 if (unlikely(err))
510 break;
511 wr = wr->next;
512 }
513
514 spin_unlock_irqrestore(&srq->rq.producer_lock, flags);
515
516 if (err) {
517 *bad_wr = wr;
518 rxe_err_srq(srq, "returned err = %d\n", err);
519 }
520
521 return err;
522 }
523
rxe_destroy_srq(struct ib_srq * ibsrq,struct ib_udata * udata)524 static int rxe_destroy_srq(struct ib_srq *ibsrq, struct ib_udata *udata)
525 {
526 struct rxe_srq *srq = to_rsrq(ibsrq);
527 int err;
528
529 err = rxe_cleanup(srq);
530 if (err)
531 rxe_err_srq(srq, "cleanup failed, err = %d\n", err);
532
533 return 0;
534 }
535
536 /* qp */
rxe_create_qp(struct ib_qp * ibqp,struct ib_qp_init_attr * init,struct ib_udata * udata)537 static int rxe_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *init,
538 struct ib_udata *udata)
539 {
540 struct rxe_dev *rxe = to_rdev(ibqp->device);
541 struct rxe_pd *pd = to_rpd(ibqp->pd);
542 struct rxe_qp *qp = to_rqp(ibqp);
543 struct rxe_create_qp_resp __user *uresp = NULL;
544 int err, cleanup_err;
545
546 if (udata) {
547 if (udata->inlen) {
548 err = -EINVAL;
549 rxe_dbg_dev(rxe, "malformed udata, err = %d\n", err);
550 goto err_out;
551 }
552
553 if (udata->outlen < sizeof(*uresp)) {
554 err = -EINVAL;
555 rxe_dbg_dev(rxe, "malformed udata, err = %d\n", err);
556 goto err_out;
557 }
558
559 qp->is_user = true;
560 uresp = udata->outbuf;
561 } else {
562 qp->is_user = false;
563 }
564
565 if (init->create_flags) {
566 err = -EOPNOTSUPP;
567 rxe_dbg_dev(rxe, "unsupported create_flags, err = %d\n", err);
568 goto err_out;
569 }
570
571 err = rxe_qp_chk_init(rxe, init);
572 if (err) {
573 rxe_dbg_dev(rxe, "bad init attr, err = %d\n", err);
574 goto err_out;
575 }
576
577 err = rxe_add_to_pool(&rxe->qp_pool, qp);
578 if (err) {
579 rxe_dbg_dev(rxe, "unable to create qp, err = %d\n", err);
580 goto err_out;
581 }
582
583 err = rxe_qp_from_init(rxe, qp, pd, init, uresp, ibqp->pd, udata);
584 if (err) {
585 rxe_dbg_qp(qp, "create qp failed, err = %d\n", err);
586 goto err_cleanup;
587 }
588
589 rxe_finalize(qp);
590 return 0;
591
592 err_cleanup:
593 cleanup_err = rxe_cleanup(qp);
594 if (cleanup_err)
595 rxe_err_qp(qp, "cleanup failed, err = %d\n", cleanup_err);
596 err_out:
597 rxe_err_dev(rxe, "returned err = %d\n", err);
598 return err;
599 }
600
rxe_modify_qp(struct ib_qp * ibqp,struct ib_qp_attr * attr,int mask,struct ib_udata * udata)601 static int rxe_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
602 int mask, struct ib_udata *udata)
603 {
604 struct rxe_dev *rxe = to_rdev(ibqp->device);
605 struct rxe_qp *qp = to_rqp(ibqp);
606 int err;
607
608 if (mask & ~IB_QP_ATTR_STANDARD_BITS) {
609 err = -EOPNOTSUPP;
610 rxe_dbg_qp(qp, "unsupported mask = 0x%x, err = %d\n",
611 mask, err);
612 goto err_out;
613 }
614
615 err = rxe_qp_chk_attr(rxe, qp, attr, mask);
616 if (err) {
617 rxe_dbg_qp(qp, "bad mask/attr, err = %d\n", err);
618 goto err_out;
619 }
620
621 err = rxe_qp_from_attr(qp, attr, mask, udata);
622 if (err) {
623 rxe_dbg_qp(qp, "modify qp failed, err = %d\n", err);
624 goto err_out;
625 }
626
627 if ((mask & IB_QP_AV) && (attr->ah_attr.ah_flags & IB_AH_GRH))
628 qp->src_port = rdma_get_udp_sport(attr->ah_attr.grh.flow_label,
629 qp->ibqp.qp_num,
630 qp->attr.dest_qp_num);
631
632 return 0;
633
634 err_out:
635 rxe_err_qp(qp, "returned err = %d\n", err);
636 return err;
637 }
638
rxe_query_qp(struct ib_qp * ibqp,struct ib_qp_attr * attr,int mask,struct ib_qp_init_attr * init)639 static int rxe_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
640 int mask, struct ib_qp_init_attr *init)
641 {
642 struct rxe_qp *qp = to_rqp(ibqp);
643
644 rxe_qp_to_init(qp, init);
645 rxe_qp_to_attr(qp, attr, mask);
646
647 return 0;
648 }
649
rxe_destroy_qp(struct ib_qp * ibqp,struct ib_udata * udata)650 static int rxe_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
651 {
652 struct rxe_qp *qp = to_rqp(ibqp);
653 int err;
654
655 err = rxe_qp_chk_destroy(qp);
656 if (err) {
657 rxe_dbg_qp(qp, "unable to destroy qp, err = %d\n", err);
658 goto err_out;
659 }
660
661 err = rxe_cleanup(qp);
662 if (err)
663 rxe_err_qp(qp, "cleanup failed, err = %d\n", err);
664
665 return 0;
666
667 err_out:
668 rxe_err_qp(qp, "returned err = %d\n", err);
669 return err;
670 }
671
672 /* send wr */
673
674 /* sanity check incoming send work request */
validate_send_wr(struct rxe_qp * qp,const struct ib_send_wr * ibwr,unsigned int * maskp,unsigned int * lengthp)675 static int validate_send_wr(struct rxe_qp *qp, const struct ib_send_wr *ibwr,
676 unsigned int *maskp, unsigned int *lengthp)
677 {
678 int num_sge = ibwr->num_sge;
679 struct rxe_sq *sq = &qp->sq;
680 unsigned int mask = 0;
681 unsigned long length = 0;
682 int err = -EINVAL;
683 int i;
684
685 do {
686 mask = wr_opcode_mask(ibwr->opcode, qp);
687 if (!mask) {
688 rxe_err_qp(qp, "bad wr opcode for qp type\n");
689 break;
690 }
691
692 if (num_sge > sq->max_sge) {
693 rxe_err_qp(qp, "num_sge > max_sge\n");
694 break;
695 }
696
697 length = 0;
698 for (i = 0; i < ibwr->num_sge; i++)
699 length += ibwr->sg_list[i].length;
700
701 if (length > RXE_PORT_MAX_MSG_SZ) {
702 rxe_err_qp(qp, "message length too long\n");
703 break;
704 }
705
706 if (mask & WR_ATOMIC_MASK) {
707 if (length != 8) {
708 rxe_err_qp(qp, "atomic length != 8\n");
709 break;
710 }
711 if (atomic_wr(ibwr)->remote_addr & 0x7) {
712 rxe_err_qp(qp, "misaligned atomic address\n");
713 break;
714 }
715 }
716 if (ibwr->send_flags & IB_SEND_INLINE) {
717 if (!(mask & WR_INLINE_MASK)) {
718 rxe_err_qp(qp, "opcode doesn't support inline data\n");
719 break;
720 }
721 if (length > sq->max_inline) {
722 rxe_err_qp(qp, "inline length too big\n");
723 break;
724 }
725 }
726
727 err = 0;
728 } while (0);
729
730 *maskp = mask;
731 *lengthp = (int)length;
732
733 return err;
734 }
735
init_send_wr(struct rxe_qp * qp,struct rxe_send_wr * wr,const struct ib_send_wr * ibwr)736 static int init_send_wr(struct rxe_qp *qp, struct rxe_send_wr *wr,
737 const struct ib_send_wr *ibwr)
738 {
739 wr->wr_id = ibwr->wr_id;
740 wr->opcode = ibwr->opcode;
741 wr->send_flags = ibwr->send_flags;
742
743 if (qp_type(qp) == IB_QPT_UD ||
744 qp_type(qp) == IB_QPT_GSI) {
745 struct ib_ah *ibah = ud_wr(ibwr)->ah;
746
747 wr->wr.ud.remote_qpn = ud_wr(ibwr)->remote_qpn;
748 wr->wr.ud.remote_qkey = ud_wr(ibwr)->remote_qkey;
749 wr->wr.ud.ah_num = to_rah(ibah)->ah_num;
750 if (qp_type(qp) == IB_QPT_GSI)
751 wr->wr.ud.pkey_index = ud_wr(ibwr)->pkey_index;
752
753 switch (wr->opcode) {
754 case IB_WR_SEND_WITH_IMM:
755 wr->ex.imm_data = ibwr->ex.imm_data;
756 break;
757 case IB_WR_SEND:
758 break;
759 default:
760 rxe_err_qp(qp, "bad wr opcode %d for UD/GSI QP\n",
761 wr->opcode);
762 return -EINVAL;
763 }
764 } else {
765 switch (wr->opcode) {
766 case IB_WR_RDMA_WRITE_WITH_IMM:
767 wr->ex.imm_data = ibwr->ex.imm_data;
768 fallthrough;
769 case IB_WR_RDMA_READ:
770 case IB_WR_RDMA_WRITE:
771 wr->wr.rdma.remote_addr = rdma_wr(ibwr)->remote_addr;
772 wr->wr.rdma.rkey = rdma_wr(ibwr)->rkey;
773 break;
774 case IB_WR_SEND_WITH_IMM:
775 wr->ex.imm_data = ibwr->ex.imm_data;
776 break;
777 case IB_WR_SEND_WITH_INV:
778 wr->ex.invalidate_rkey = ibwr->ex.invalidate_rkey;
779 break;
780 case IB_WR_RDMA_READ_WITH_INV:
781 wr->ex.invalidate_rkey = ibwr->ex.invalidate_rkey;
782 wr->wr.rdma.remote_addr = rdma_wr(ibwr)->remote_addr;
783 wr->wr.rdma.rkey = rdma_wr(ibwr)->rkey;
784 break;
785 case IB_WR_ATOMIC_CMP_AND_SWP:
786 case IB_WR_ATOMIC_FETCH_AND_ADD:
787 wr->wr.atomic.remote_addr =
788 atomic_wr(ibwr)->remote_addr;
789 wr->wr.atomic.compare_add =
790 atomic_wr(ibwr)->compare_add;
791 wr->wr.atomic.swap = atomic_wr(ibwr)->swap;
792 wr->wr.atomic.rkey = atomic_wr(ibwr)->rkey;
793 break;
794 case IB_WR_LOCAL_INV:
795 wr->ex.invalidate_rkey = ibwr->ex.invalidate_rkey;
796 break;
797 case IB_WR_REG_MR:
798 wr->wr.reg.mr = reg_wr(ibwr)->mr;
799 wr->wr.reg.key = reg_wr(ibwr)->key;
800 wr->wr.reg.access = reg_wr(ibwr)->access;
801 break;
802 case IB_WR_SEND:
803 case IB_WR_BIND_MW:
804 case IB_WR_FLUSH:
805 case IB_WR_ATOMIC_WRITE:
806 break;
807 default:
808 rxe_err_qp(qp, "unsupported wr opcode %d\n",
809 wr->opcode);
810 return -EINVAL;
811 }
812 }
813
814 return 0;
815 }
816
copy_inline_data_to_wqe(struct rxe_send_wqe * wqe,const struct ib_send_wr * ibwr)817 static void copy_inline_data_to_wqe(struct rxe_send_wqe *wqe,
818 const struct ib_send_wr *ibwr)
819 {
820 struct ib_sge *sge = ibwr->sg_list;
821 u8 *p = wqe->dma.inline_data;
822 int i;
823
824 for (i = 0; i < ibwr->num_sge; i++, sge++) {
825 memcpy(p, ib_virt_dma_to_ptr(sge->addr), sge->length);
826 p += sge->length;
827 }
828 }
829
init_send_wqe(struct rxe_qp * qp,const struct ib_send_wr * ibwr,unsigned int mask,unsigned int length,struct rxe_send_wqe * wqe)830 static int init_send_wqe(struct rxe_qp *qp, const struct ib_send_wr *ibwr,
831 unsigned int mask, unsigned int length,
832 struct rxe_send_wqe *wqe)
833 {
834 int num_sge = ibwr->num_sge;
835 int err;
836
837 err = init_send_wr(qp, &wqe->wr, ibwr);
838 if (err)
839 return err;
840
841 /* local operation */
842 if (unlikely(mask & WR_LOCAL_OP_MASK)) {
843 wqe->mask = mask;
844 wqe->state = wqe_state_posted;
845 return 0;
846 }
847
848 if (unlikely(ibwr->send_flags & IB_SEND_INLINE))
849 copy_inline_data_to_wqe(wqe, ibwr);
850 else
851 memcpy(wqe->dma.sge, ibwr->sg_list,
852 num_sge * sizeof(struct ib_sge));
853
854 wqe->iova = mask & WR_ATOMIC_MASK ? atomic_wr(ibwr)->remote_addr :
855 mask & WR_READ_OR_WRITE_MASK ? rdma_wr(ibwr)->remote_addr : 0;
856 wqe->mask = mask;
857 wqe->dma.length = length;
858 wqe->dma.resid = length;
859 wqe->dma.num_sge = num_sge;
860 wqe->dma.cur_sge = 0;
861 wqe->dma.sge_offset = 0;
862 wqe->state = wqe_state_posted;
863 wqe->ssn = atomic_add_return(1, &qp->ssn);
864
865 return 0;
866 }
867
post_one_send(struct rxe_qp * qp,const struct ib_send_wr * ibwr)868 static int post_one_send(struct rxe_qp *qp, const struct ib_send_wr *ibwr)
869 {
870 int err;
871 struct rxe_sq *sq = &qp->sq;
872 struct rxe_send_wqe *send_wqe;
873 unsigned int mask;
874 unsigned int length;
875 int full;
876
877 err = validate_send_wr(qp, ibwr, &mask, &length);
878 if (err)
879 return err;
880
881 full = queue_full(sq->queue, QUEUE_TYPE_FROM_ULP);
882 if (unlikely(full)) {
883 rxe_err_qp(qp, "send queue full\n");
884 return -ENOMEM;
885 }
886
887 send_wqe = queue_producer_addr(sq->queue, QUEUE_TYPE_FROM_ULP);
888 err = init_send_wqe(qp, ibwr, mask, length, send_wqe);
889 if (!err)
890 queue_advance_producer(sq->queue, QUEUE_TYPE_FROM_ULP);
891
892 return err;
893 }
894
rxe_post_send_kernel(struct rxe_qp * qp,const struct ib_send_wr * ibwr,const struct ib_send_wr ** bad_wr)895 static int rxe_post_send_kernel(struct rxe_qp *qp,
896 const struct ib_send_wr *ibwr,
897 const struct ib_send_wr **bad_wr)
898 {
899 int err = 0;
900 unsigned long flags;
901 int good = 0;
902
903 spin_lock_irqsave(&qp->sq.sq_lock, flags);
904 while (ibwr) {
905 err = post_one_send(qp, ibwr);
906 if (err) {
907 *bad_wr = ibwr;
908 break;
909 } else {
910 good++;
911 }
912 ibwr = ibwr->next;
913 }
914 spin_unlock_irqrestore(&qp->sq.sq_lock, flags);
915
916 /* kickoff processing of any posted wqes */
917 if (good)
918 rxe_sched_task(&qp->send_task);
919
920 return err;
921 }
922
rxe_post_send(struct ib_qp * ibqp,const struct ib_send_wr * wr,const struct ib_send_wr ** bad_wr)923 static int rxe_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
924 const struct ib_send_wr **bad_wr)
925 {
926 struct rxe_qp *qp = to_rqp(ibqp);
927 int err;
928 unsigned long flags;
929
930 spin_lock_irqsave(&qp->state_lock, flags);
931 /* caller has already called destroy_qp */
932 if (WARN_ON_ONCE(!qp->valid)) {
933 spin_unlock_irqrestore(&qp->state_lock, flags);
934 rxe_err_qp(qp, "qp has been destroyed\n");
935 return -EINVAL;
936 }
937
938 if (unlikely(qp_state(qp) < IB_QPS_RTS)) {
939 spin_unlock_irqrestore(&qp->state_lock, flags);
940 *bad_wr = wr;
941 rxe_err_qp(qp, "qp not ready to send\n");
942 return -EINVAL;
943 }
944 spin_unlock_irqrestore(&qp->state_lock, flags);
945
946 if (qp->is_user) {
947 /* Utilize process context to do protocol processing */
948 rxe_sched_task(&qp->send_task);
949 } else {
950 err = rxe_post_send_kernel(qp, wr, bad_wr);
951 if (err)
952 return err;
953 }
954
955 return 0;
956 }
957
958 /* recv wr */
post_one_recv(struct rxe_rq * rq,const struct ib_recv_wr * ibwr)959 static int post_one_recv(struct rxe_rq *rq, const struct ib_recv_wr *ibwr)
960 {
961 int i;
962 unsigned long length;
963 struct rxe_recv_wqe *recv_wqe;
964 int num_sge = ibwr->num_sge;
965 int full;
966 int err;
967
968 full = queue_full(rq->queue, QUEUE_TYPE_FROM_ULP);
969 if (unlikely(full)) {
970 err = -ENOMEM;
971 rxe_dbg("queue full\n");
972 goto err_out;
973 }
974
975 if (unlikely(num_sge > rq->max_sge)) {
976 err = -EINVAL;
977 rxe_dbg("bad num_sge > max_sge\n");
978 goto err_out;
979 }
980
981 length = 0;
982 for (i = 0; i < num_sge; i++)
983 length += ibwr->sg_list[i].length;
984
985 if (length > RXE_PORT_MAX_MSG_SZ) {
986 err = -EINVAL;
987 rxe_dbg("message length too long\n");
988 goto err_out;
989 }
990
991 recv_wqe = queue_producer_addr(rq->queue, QUEUE_TYPE_FROM_ULP);
992
993 recv_wqe->wr_id = ibwr->wr_id;
994 recv_wqe->dma.length = length;
995 recv_wqe->dma.resid = length;
996 recv_wqe->dma.num_sge = num_sge;
997 recv_wqe->dma.cur_sge = 0;
998 recv_wqe->dma.sge_offset = 0;
999 memcpy(recv_wqe->dma.sge, ibwr->sg_list,
1000 num_sge * sizeof(struct ib_sge));
1001
1002 queue_advance_producer(rq->queue, QUEUE_TYPE_FROM_ULP);
1003
1004 return 0;
1005
1006 err_out:
1007 rxe_dbg("returned err = %d\n", err);
1008 return err;
1009 }
1010
rxe_post_recv(struct ib_qp * ibqp,const struct ib_recv_wr * wr,const struct ib_recv_wr ** bad_wr)1011 static int rxe_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
1012 const struct ib_recv_wr **bad_wr)
1013 {
1014 int err = 0;
1015 struct rxe_qp *qp = to_rqp(ibqp);
1016 struct rxe_rq *rq = &qp->rq;
1017 unsigned long flags;
1018
1019 spin_lock_irqsave(&qp->state_lock, flags);
1020 /* caller has already called destroy_qp */
1021 if (WARN_ON_ONCE(!qp->valid)) {
1022 spin_unlock_irqrestore(&qp->state_lock, flags);
1023 rxe_err_qp(qp, "qp has been destroyed\n");
1024 return -EINVAL;
1025 }
1026
1027 /* see C10-97.2.1 */
1028 if (unlikely((qp_state(qp) < IB_QPS_INIT))) {
1029 spin_unlock_irqrestore(&qp->state_lock, flags);
1030 *bad_wr = wr;
1031 rxe_dbg_qp(qp, "qp not ready to post recv\n");
1032 return -EINVAL;
1033 }
1034 spin_unlock_irqrestore(&qp->state_lock, flags);
1035
1036 if (unlikely(qp->srq)) {
1037 *bad_wr = wr;
1038 rxe_dbg_qp(qp, "qp has srq, use post_srq_recv instead\n");
1039 return -EINVAL;
1040 }
1041
1042 spin_lock_irqsave(&rq->producer_lock, flags);
1043
1044 while (wr) {
1045 err = post_one_recv(rq, wr);
1046 if (unlikely(err)) {
1047 *bad_wr = wr;
1048 break;
1049 }
1050 wr = wr->next;
1051 }
1052
1053 spin_unlock_irqrestore(&rq->producer_lock, flags);
1054
1055 spin_lock_irqsave(&qp->state_lock, flags);
1056 if (qp_state(qp) == IB_QPS_ERR)
1057 rxe_sched_task(&qp->recv_task);
1058 spin_unlock_irqrestore(&qp->state_lock, flags);
1059
1060 return err;
1061 }
1062
1063 /* cq */
rxe_create_cq(struct ib_cq * ibcq,const struct ib_cq_init_attr * attr,struct uverbs_attr_bundle * attrs)1064 static int rxe_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
1065 struct uverbs_attr_bundle *attrs)
1066 {
1067 struct ib_udata *udata = &attrs->driver_udata;
1068 struct ib_device *dev = ibcq->device;
1069 struct rxe_dev *rxe = to_rdev(dev);
1070 struct rxe_cq *cq = to_rcq(ibcq);
1071 struct rxe_create_cq_resp __user *uresp = NULL;
1072 int err, cleanup_err;
1073
1074 if (udata) {
1075 if (udata->outlen < sizeof(*uresp)) {
1076 err = -EINVAL;
1077 rxe_dbg_dev(rxe, "malformed udata, err = %d\n", err);
1078 goto err_out;
1079 }
1080 uresp = udata->outbuf;
1081 }
1082
1083 if (attr->flags) {
1084 err = -EOPNOTSUPP;
1085 rxe_dbg_dev(rxe, "bad attr->flags, err = %d\n", err);
1086 goto err_out;
1087 }
1088
1089 if (attr->cqe > rxe->attr.max_cqe)
1090 return -EINVAL;
1091
1092 err = rxe_add_to_pool(&rxe->cq_pool, cq);
1093 if (err) {
1094 rxe_dbg_dev(rxe, "unable to create cq, err = %d\n", err);
1095 goto err_out;
1096 }
1097
1098 err = rxe_cq_from_init(rxe, cq, attr->cqe, attr->comp_vector, udata,
1099 uresp);
1100 if (err) {
1101 rxe_dbg_cq(cq, "create cq failed, err = %d\n", err);
1102 goto err_cleanup;
1103 }
1104
1105 return 0;
1106
1107 err_cleanup:
1108 cleanup_err = rxe_cleanup(cq);
1109 if (cleanup_err)
1110 rxe_err_cq(cq, "cleanup failed, err = %d\n", cleanup_err);
1111 err_out:
1112 rxe_err_dev(rxe, "returned err = %d\n", err);
1113 return err;
1114 }
1115
rxe_resize_cq(struct ib_cq * ibcq,unsigned int cqe,struct ib_udata * udata)1116 static int rxe_resize_cq(struct ib_cq *ibcq, unsigned int cqe,
1117 struct ib_udata *udata)
1118 {
1119 struct rxe_cq *cq = to_rcq(ibcq);
1120 struct rxe_dev *rxe = to_rdev(ibcq->device);
1121 struct rxe_resize_cq_resp __user *uresp = NULL;
1122 int err;
1123
1124 if (udata) {
1125 if (udata->outlen < sizeof(*uresp)) {
1126 err = -EINVAL;
1127 rxe_dbg_cq(cq, "malformed udata\n");
1128 goto err_out;
1129 }
1130 uresp = udata->outbuf;
1131 }
1132
1133 if (cqe > rxe->attr.max_cqe ||
1134 cqe < queue_count(cq->queue, QUEUE_TYPE_TO_CLIENT))
1135 return -EINVAL;
1136
1137 err = rxe_cq_resize_queue(cq, cqe, uresp, udata);
1138 if (err) {
1139 rxe_dbg_cq(cq, "resize cq failed, err = %d\n", err);
1140 goto err_out;
1141 }
1142
1143 return 0;
1144
1145 err_out:
1146 rxe_err_cq(cq, "returned err = %d\n", err);
1147 return err;
1148 }
1149
rxe_poll_cq(struct ib_cq * ibcq,int num_entries,struct ib_wc * wc)1150 static int rxe_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
1151 {
1152 int i;
1153 struct rxe_cq *cq = to_rcq(ibcq);
1154 struct rxe_cqe *cqe;
1155 unsigned long flags;
1156
1157 spin_lock_irqsave(&cq->cq_lock, flags);
1158 for (i = 0; i < num_entries; i++) {
1159 cqe = queue_head(cq->queue, QUEUE_TYPE_TO_ULP);
1160 if (!cqe)
1161 break; /* queue empty */
1162
1163 memcpy(wc++, &cqe->ibwc, sizeof(*wc));
1164 queue_advance_consumer(cq->queue, QUEUE_TYPE_TO_ULP);
1165 }
1166 spin_unlock_irqrestore(&cq->cq_lock, flags);
1167
1168 return i;
1169 }
1170
rxe_peek_cq(struct ib_cq * ibcq,int wc_cnt)1171 static int rxe_peek_cq(struct ib_cq *ibcq, int wc_cnt)
1172 {
1173 struct rxe_cq *cq = to_rcq(ibcq);
1174 int count;
1175
1176 count = queue_count(cq->queue, QUEUE_TYPE_TO_ULP);
1177
1178 return (count > wc_cnt) ? wc_cnt : count;
1179 }
1180
rxe_req_notify_cq(struct ib_cq * ibcq,enum ib_cq_notify_flags flags)1181 static int rxe_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
1182 {
1183 struct rxe_cq *cq = to_rcq(ibcq);
1184 int ret = 0;
1185 int empty;
1186 unsigned long irq_flags;
1187
1188 spin_lock_irqsave(&cq->cq_lock, irq_flags);
1189 cq->notify |= flags & IB_CQ_SOLICITED_MASK;
1190 empty = queue_empty(cq->queue, QUEUE_TYPE_TO_ULP);
1191
1192 if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !empty)
1193 ret = 1;
1194
1195 spin_unlock_irqrestore(&cq->cq_lock, irq_flags);
1196
1197 return ret;
1198 }
1199
rxe_destroy_cq(struct ib_cq * ibcq,struct ib_udata * udata)1200 static int rxe_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
1201 {
1202 struct rxe_cq *cq = to_rcq(ibcq);
1203 int err;
1204
1205 /* See IBA C11-17: The CI shall return an error if this Verb is
1206 * invoked while a Work Queue is still associated with the CQ.
1207 */
1208 if (atomic_read(&cq->num_wq)) {
1209 err = -EINVAL;
1210 rxe_dbg_cq(cq, "still in use\n");
1211 goto err_out;
1212 }
1213
1214 err = rxe_cleanup(cq);
1215 if (err)
1216 rxe_err_cq(cq, "cleanup failed, err = %d\n", err);
1217
1218 return 0;
1219
1220 err_out:
1221 rxe_err_cq(cq, "returned err = %d\n", err);
1222 return err;
1223 }
1224
1225 /* mr */
rxe_get_dma_mr(struct ib_pd * ibpd,int access)1226 static struct ib_mr *rxe_get_dma_mr(struct ib_pd *ibpd, int access)
1227 {
1228 struct rxe_dev *rxe = to_rdev(ibpd->device);
1229 struct rxe_pd *pd = to_rpd(ibpd);
1230 struct rxe_mr *mr;
1231 int err;
1232
1233 mr = kzalloc_obj(*mr);
1234 if (!mr)
1235 return ERR_PTR(-ENOMEM);
1236
1237 err = rxe_add_to_pool(&rxe->mr_pool, mr);
1238 if (err) {
1239 rxe_dbg_dev(rxe, "unable to create mr\n");
1240 goto err_free;
1241 }
1242
1243 rxe_get(pd);
1244 mr->ibmr.pd = ibpd;
1245 mr->ibmr.device = ibpd->device;
1246
1247 rxe_mr_init_dma(access, mr);
1248 rxe_finalize(mr);
1249 return &mr->ibmr;
1250
1251 err_free:
1252 kfree(mr);
1253 rxe_err_pd(pd, "returned err = %d\n", err);
1254 return ERR_PTR(err);
1255 }
1256
rxe_reg_user_mr(struct ib_pd * ibpd,u64 start,u64 length,u64 iova,int access,struct ib_dmah * dmah,struct ib_udata * udata)1257 static struct ib_mr *rxe_reg_user_mr(struct ib_pd *ibpd, u64 start,
1258 u64 length, u64 iova, int access,
1259 struct ib_dmah *dmah,
1260 struct ib_udata *udata)
1261 {
1262 struct rxe_dev *rxe = to_rdev(ibpd->device);
1263 struct rxe_pd *pd = to_rpd(ibpd);
1264 struct rxe_mr *mr;
1265 int err, cleanup_err;
1266
1267 if (dmah)
1268 return ERR_PTR(-EOPNOTSUPP);
1269
1270 if (access & ~RXE_ACCESS_SUPPORTED_MR) {
1271 rxe_err_pd(pd, "access = %#x not supported (%#x)\n", access,
1272 RXE_ACCESS_SUPPORTED_MR);
1273 return ERR_PTR(-EOPNOTSUPP);
1274 }
1275
1276 mr = kzalloc_obj(*mr);
1277 if (!mr)
1278 return ERR_PTR(-ENOMEM);
1279
1280 err = rxe_add_to_pool(&rxe->mr_pool, mr);
1281 if (err) {
1282 rxe_dbg_pd(pd, "unable to create mr\n");
1283 goto err_free;
1284 }
1285
1286 rxe_get(pd);
1287 mr->ibmr.pd = ibpd;
1288 mr->ibmr.device = ibpd->device;
1289
1290 if (access & IB_ACCESS_ON_DEMAND)
1291 err = rxe_odp_mr_init_user(rxe, start, length, iova, access, mr);
1292 else
1293 err = rxe_mr_init_user(rxe, start, length, access, mr);
1294 if (err) {
1295 rxe_dbg_mr(mr, "reg_user_mr failed, err = %d\n", err);
1296 goto err_cleanup;
1297 }
1298
1299 rxe_finalize(mr);
1300 return &mr->ibmr;
1301
1302 err_cleanup:
1303 cleanup_err = rxe_cleanup(mr);
1304 if (cleanup_err)
1305 rxe_err_mr(mr, "cleanup failed, err = %d\n", cleanup_err);
1306 err_free:
1307 kfree(mr);
1308 rxe_err_pd(pd, "returned err = %d\n", err);
1309 return ERR_PTR(err);
1310 }
1311
rxe_rereg_user_mr(struct ib_mr * ibmr,int flags,u64 start,u64 length,u64 iova,int access,struct ib_pd * ibpd,struct ib_udata * udata)1312 static struct ib_mr *rxe_rereg_user_mr(struct ib_mr *ibmr, int flags,
1313 u64 start, u64 length, u64 iova,
1314 int access, struct ib_pd *ibpd,
1315 struct ib_udata *udata)
1316 {
1317 struct rxe_mr *mr = to_rmr(ibmr);
1318 struct rxe_pd *old_pd = to_rpd(ibmr->pd);
1319 struct rxe_pd *pd = to_rpd(ibpd);
1320 int err;
1321
1322 /* for now only support the two easy cases:
1323 * rereg_pd and rereg_access
1324 */
1325 if (flags & ~RXE_MR_REREG_SUPPORTED) {
1326 rxe_err_mr(mr, "flags = %#x not supported\n", flags);
1327 return ERR_PTR(-EOPNOTSUPP);
1328 }
1329
1330 err = ib_umem_check_rereg(mr->umem, flags, access);
1331 if (err)
1332 return ERR_PTR(err);
1333
1334 if ((flags & IB_MR_REREG_ACCESS) &&
1335 (access & ~RXE_ACCESS_SUPPORTED_MR)) {
1336 rxe_err_mr(mr, "access = %#x not supported\n", access);
1337 return ERR_PTR(-EOPNOTSUPP);
1338 }
1339
1340 if (flags & IB_MR_REREG_PD) {
1341 rxe_put(old_pd);
1342 rxe_get(pd);
1343 mr->ibmr.pd = ibpd;
1344 }
1345
1346 if (flags & IB_MR_REREG_ACCESS)
1347 mr->access = access;
1348
1349 return NULL;
1350 }
1351
rxe_alloc_mr(struct ib_pd * ibpd,enum ib_mr_type mr_type,u32 max_num_sg)1352 static struct ib_mr *rxe_alloc_mr(struct ib_pd *ibpd, enum ib_mr_type mr_type,
1353 u32 max_num_sg)
1354 {
1355 struct rxe_dev *rxe = to_rdev(ibpd->device);
1356 struct rxe_pd *pd = to_rpd(ibpd);
1357 struct rxe_mr *mr;
1358 int err, cleanup_err;
1359
1360 if (mr_type != IB_MR_TYPE_MEM_REG) {
1361 err = -EINVAL;
1362 rxe_dbg_pd(pd, "mr type %d not supported, err = %d\n",
1363 mr_type, err);
1364 goto err_out;
1365 }
1366
1367 mr = kzalloc_obj(*mr);
1368 if (!mr)
1369 return ERR_PTR(-ENOMEM);
1370
1371 err = rxe_add_to_pool(&rxe->mr_pool, mr);
1372 if (err)
1373 goto err_free;
1374
1375 rxe_get(pd);
1376 mr->ibmr.pd = ibpd;
1377 mr->ibmr.device = ibpd->device;
1378
1379 err = rxe_mr_init_fast(max_num_sg, mr);
1380 if (err) {
1381 rxe_dbg_mr(mr, "alloc_mr failed, err = %d\n", err);
1382 goto err_cleanup;
1383 }
1384
1385 rxe_finalize(mr);
1386 return &mr->ibmr;
1387
1388 err_cleanup:
1389 cleanup_err = rxe_cleanup(mr);
1390 if (cleanup_err)
1391 rxe_err_mr(mr, "cleanup failed, err = %d\n", err);
1392 err_free:
1393 kfree(mr);
1394 err_out:
1395 rxe_err_pd(pd, "returned err = %d\n", err);
1396 return ERR_PTR(err);
1397 }
1398
rxe_dereg_mr(struct ib_mr * ibmr,struct ib_udata * udata)1399 static int rxe_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
1400 {
1401 struct rxe_mr *mr = to_rmr(ibmr);
1402 int err, cleanup_err;
1403
1404 /* See IBA 10.6.7.2.6 */
1405 if (atomic_read(&mr->num_mw) > 0) {
1406 err = -EINVAL;
1407 rxe_dbg_mr(mr, "mr has mw's bound\n");
1408 goto err_out;
1409 }
1410
1411 cleanup_err = rxe_cleanup(mr);
1412 if (cleanup_err)
1413 rxe_err_mr(mr, "cleanup failed, err = %d\n", cleanup_err);
1414
1415 kfree_rcu_mightsleep(mr);
1416 return 0;
1417
1418 err_out:
1419 rxe_err_mr(mr, "returned err = %d\n", err);
1420 return err;
1421 }
1422
parent_show(struct device * device,struct device_attribute * attr,char * buf)1423 static ssize_t parent_show(struct device *device,
1424 struct device_attribute *attr, char *buf)
1425 {
1426 struct rxe_dev *rxe =
1427 rdma_device_to_drv_device(device, struct rxe_dev, ib_dev);
1428
1429 return sysfs_emit(buf, "%s\n", rxe_parent_name(rxe, 1));
1430 }
1431
1432 static DEVICE_ATTR_RO(parent);
1433
1434 static struct attribute *rxe_dev_attributes[] = {
1435 &dev_attr_parent.attr,
1436 NULL
1437 };
1438
1439 static const struct attribute_group rxe_attr_group = {
1440 .attrs = rxe_dev_attributes,
1441 };
1442
rxe_enable_driver(struct ib_device * ib_dev)1443 static int rxe_enable_driver(struct ib_device *ib_dev)
1444 {
1445 struct rxe_dev *rxe = container_of(ib_dev, struct rxe_dev, ib_dev);
1446 struct net_device *ndev;
1447
1448 ndev = ib_device_get_netdev(ib_dev, 1);
1449 if (!ndev)
1450 return -ENODEV;
1451
1452 rxe_set_port_state(rxe);
1453 dev_info(&rxe->ib_dev.dev, "added %s\n", netdev_name(ndev));
1454
1455 dev_put(ndev);
1456 return 0;
1457 }
1458
1459 static const struct ib_device_ops rxe_dev_ops = {
1460 .owner = THIS_MODULE,
1461 .driver_id = RDMA_DRIVER_RXE,
1462 .uverbs_abi_ver = RXE_UVERBS_ABI_VERSION,
1463
1464 .alloc_hw_port_stats = rxe_ib_alloc_hw_port_stats,
1465 .alloc_mr = rxe_alloc_mr,
1466 .alloc_mw = rxe_alloc_mw,
1467 .alloc_pd = rxe_alloc_pd,
1468 .alloc_ucontext = rxe_alloc_ucontext,
1469 .attach_mcast = rxe_attach_mcast,
1470 .create_ah = rxe_create_ah,
1471 .create_cq = rxe_create_cq,
1472 .create_qp = rxe_create_qp,
1473 .create_srq = rxe_create_srq,
1474 .create_user_ah = rxe_create_ah,
1475 .dealloc_driver = rxe_dealloc,
1476 .dealloc_mw = rxe_dealloc_mw,
1477 .dealloc_pd = rxe_dealloc_pd,
1478 .dealloc_ucontext = rxe_dealloc_ucontext,
1479 .dereg_mr = rxe_dereg_mr,
1480 .destroy_ah = rxe_destroy_ah,
1481 .destroy_cq = rxe_destroy_cq,
1482 .destroy_qp = rxe_destroy_qp,
1483 .destroy_srq = rxe_destroy_srq,
1484 .detach_mcast = rxe_detach_mcast,
1485 .device_group = &rxe_attr_group,
1486 .disassociate_ucontext = rxe_disassociate_ucontext,
1487 .enable_driver = rxe_enable_driver,
1488 .get_dma_mr = rxe_get_dma_mr,
1489 .get_hw_stats = rxe_ib_get_hw_stats,
1490 .get_link_layer = rxe_get_link_layer,
1491 .get_port_immutable = rxe_port_immutable,
1492 .map_mr_sg = rxe_map_mr_sg,
1493 .mmap = rxe_mmap,
1494 .modify_ah = rxe_modify_ah,
1495 .modify_device = rxe_modify_device,
1496 .modify_port = rxe_modify_port,
1497 .modify_qp = rxe_modify_qp,
1498 .modify_srq = rxe_modify_srq,
1499 .peek_cq = rxe_peek_cq,
1500 .poll_cq = rxe_poll_cq,
1501 .post_recv = rxe_post_recv,
1502 .post_send = rxe_post_send,
1503 .post_srq_recv = rxe_post_srq_recv,
1504 .process_mad = rxe_process_mad,
1505 .query_ah = rxe_query_ah,
1506 .query_device = rxe_query_device,
1507 .query_pkey = rxe_query_pkey,
1508 .query_gid = rxe_query_gid,
1509 .query_port = rxe_query_port,
1510 .query_qp = rxe_query_qp,
1511 .query_srq = rxe_query_srq,
1512 .reg_user_mr = rxe_reg_user_mr,
1513 .req_notify_cq = rxe_req_notify_cq,
1514 .rereg_user_mr = rxe_rereg_user_mr,
1515 .resize_user_cq = rxe_resize_cq,
1516
1517 INIT_RDMA_OBJ_SIZE(ib_ah, rxe_ah, ibah),
1518 INIT_RDMA_OBJ_SIZE(ib_cq, rxe_cq, ibcq),
1519 INIT_RDMA_OBJ_SIZE(ib_pd, rxe_pd, ibpd),
1520 INIT_RDMA_OBJ_SIZE(ib_qp, rxe_qp, ibqp),
1521 INIT_RDMA_OBJ_SIZE(ib_srq, rxe_srq, ibsrq),
1522 INIT_RDMA_OBJ_SIZE(ib_ucontext, rxe_ucontext, ibuc),
1523 INIT_RDMA_OBJ_SIZE(ib_mw, rxe_mw, ibmw),
1524 };
1525
rxe_register_device(struct rxe_dev * rxe,const char * ibdev_name,struct net_device * ndev)1526 int rxe_register_device(struct rxe_dev *rxe, const char *ibdev_name,
1527 struct net_device *ndev)
1528 {
1529 int err;
1530 struct ib_device *dev = &rxe->ib_dev;
1531
1532 strscpy(dev->node_desc, "rxe", sizeof(dev->node_desc));
1533
1534 dev->node_type = RDMA_NODE_IB_CA;
1535 dev->phys_port_cnt = 1;
1536 dev->num_comp_vectors = num_possible_cpus();
1537 dev->local_dma_lkey = 0;
1538 addrconf_addr_eui48((unsigned char *)&dev->node_guid,
1539 rxe->raw_gid);
1540
1541 dev->uverbs_cmd_mask |= BIT_ULL(IB_USER_VERBS_CMD_POST_SEND) |
1542 BIT_ULL(IB_USER_VERBS_CMD_REQ_NOTIFY_CQ);
1543
1544 ib_set_device_ops(dev, &rxe_dev_ops);
1545 err = ib_device_set_netdev(&rxe->ib_dev, ndev, 1);
1546 if (err)
1547 return err;
1548
1549 err = ib_register_device(dev, ibdev_name, NULL);
1550 if (err)
1551 rxe_dbg_dev(rxe, "failed with error %d\n", err);
1552
1553 /*
1554 * Note that rxe may be invalid at this point if another thread
1555 * unregistered it.
1556 */
1557 return err;
1558 }
1559