tipc: improve function tipc_wait_for_rcvmsg()
[muen/linux.git] / net / smc / smc_ib.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
4  *
5  *  IB infrastructure:
6  *  Establish SMC-R as an Infiniband Client to be notified about added and
7  *  removed IB devices of type RDMA.
8  *  Determine device and port characteristics for these IB devices.
9  *
10  *  Copyright IBM Corp. 2016
11  *
12  *  Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
13  */
14
15 #include <linux/random.h>
16 #include <linux/workqueue.h>
17 #include <linux/scatterlist.h>
18 #include <rdma/ib_verbs.h>
19 #include <rdma/ib_cache.h>
20
21 #include "smc_pnet.h"
22 #include "smc_ib.h"
23 #include "smc_core.h"
24 #include "smc_wr.h"
25 #include "smc.h"
26
27 #define SMC_MAX_CQE 32766       /* max. # of completion queue elements */
28
29 #define SMC_QP_MIN_RNR_TIMER            5
30 #define SMC_QP_TIMEOUT                  15 /* 4096 * 2 ** timeout usec */
31 #define SMC_QP_RETRY_CNT                        7 /* 7: infinite */
32 #define SMC_QP_RNR_RETRY                        7 /* 7: infinite */
33
34 struct smc_ib_devices smc_ib_devices = {        /* smc-registered ib devices */
35         .lock = __SPIN_LOCK_UNLOCKED(smc_ib_devices.lock),
36         .list = LIST_HEAD_INIT(smc_ib_devices.list),
37 };
38
39 #define SMC_LOCAL_SYSTEMID_RESET        "%%%%%%%"
40
41 u8 local_systemid[SMC_SYSTEMID_LEN] = SMC_LOCAL_SYSTEMID_RESET; /* unique system
42                                                                  * identifier
43                                                                  */
44
45 static int smc_ib_modify_qp_init(struct smc_link *lnk)
46 {
47         struct ib_qp_attr qp_attr;
48
49         memset(&qp_attr, 0, sizeof(qp_attr));
50         qp_attr.qp_state = IB_QPS_INIT;
51         qp_attr.pkey_index = 0;
52         qp_attr.port_num = lnk->ibport;
53         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE
54                                 | IB_ACCESS_REMOTE_WRITE;
55         return ib_modify_qp(lnk->roce_qp, &qp_attr,
56                             IB_QP_STATE | IB_QP_PKEY_INDEX |
57                             IB_QP_ACCESS_FLAGS | IB_QP_PORT);
58 }
59
60 static int smc_ib_modify_qp_rtr(struct smc_link *lnk)
61 {
62         enum ib_qp_attr_mask qp_attr_mask =
63                 IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN |
64                 IB_QP_RQ_PSN | IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER;
65         struct ib_qp_attr qp_attr;
66
67         memset(&qp_attr, 0, sizeof(qp_attr));
68         qp_attr.qp_state = IB_QPS_RTR;
69         qp_attr.path_mtu = min(lnk->path_mtu, lnk->peer_mtu);
70         qp_attr.ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
71         rdma_ah_set_port_num(&qp_attr.ah_attr, lnk->ibport);
72         rdma_ah_set_grh(&qp_attr.ah_attr, NULL, 0, lnk->sgid_index, 1, 0);
73         rdma_ah_set_dgid_raw(&qp_attr.ah_attr, lnk->peer_gid);
74         memcpy(&qp_attr.ah_attr.roce.dmac, lnk->peer_mac,
75                sizeof(lnk->peer_mac));
76         qp_attr.dest_qp_num = lnk->peer_qpn;
77         qp_attr.rq_psn = lnk->peer_psn; /* starting receive packet seq # */
78         qp_attr.max_dest_rd_atomic = 1; /* max # of resources for incoming
79                                          * requests
80                                          */
81         qp_attr.min_rnr_timer = SMC_QP_MIN_RNR_TIMER;
82
83         return ib_modify_qp(lnk->roce_qp, &qp_attr, qp_attr_mask);
84 }
85
86 int smc_ib_modify_qp_rts(struct smc_link *lnk)
87 {
88         struct ib_qp_attr qp_attr;
89
90         memset(&qp_attr, 0, sizeof(qp_attr));
91         qp_attr.qp_state = IB_QPS_RTS;
92         qp_attr.timeout = SMC_QP_TIMEOUT;       /* local ack timeout */
93         qp_attr.retry_cnt = SMC_QP_RETRY_CNT;   /* retry count */
94         qp_attr.rnr_retry = SMC_QP_RNR_RETRY;   /* RNR retries, 7=infinite */
95         qp_attr.sq_psn = lnk->psn_initial;      /* starting send packet seq # */
96         qp_attr.max_rd_atomic = 1;      /* # of outstanding RDMA reads and
97                                          * atomic ops allowed
98                                          */
99         return ib_modify_qp(lnk->roce_qp, &qp_attr,
100                             IB_QP_STATE | IB_QP_TIMEOUT | IB_QP_RETRY_CNT |
101                             IB_QP_SQ_PSN | IB_QP_RNR_RETRY |
102                             IB_QP_MAX_QP_RD_ATOMIC);
103 }
104
105 int smc_ib_modify_qp_reset(struct smc_link *lnk)
106 {
107         struct ib_qp_attr qp_attr;
108
109         memset(&qp_attr, 0, sizeof(qp_attr));
110         qp_attr.qp_state = IB_QPS_RESET;
111         return ib_modify_qp(lnk->roce_qp, &qp_attr, IB_QP_STATE);
112 }
113
114 int smc_ib_ready_link(struct smc_link *lnk)
115 {
116         struct smc_link_group *lgr = smc_get_lgr(lnk);
117         int rc = 0;
118
119         rc = smc_ib_modify_qp_init(lnk);
120         if (rc)
121                 goto out;
122
123         rc = smc_ib_modify_qp_rtr(lnk);
124         if (rc)
125                 goto out;
126         smc_wr_remember_qp_attr(lnk);
127         rc = ib_req_notify_cq(lnk->smcibdev->roce_cq_recv,
128                               IB_CQ_SOLICITED_MASK);
129         if (rc)
130                 goto out;
131         rc = smc_wr_rx_post_init(lnk);
132         if (rc)
133                 goto out;
134         smc_wr_remember_qp_attr(lnk);
135
136         if (lgr->role == SMC_SERV) {
137                 rc = smc_ib_modify_qp_rts(lnk);
138                 if (rc)
139                         goto out;
140                 smc_wr_remember_qp_attr(lnk);
141         }
142 out:
143         return rc;
144 }
145
146 static int smc_ib_fill_mac(struct smc_ib_device *smcibdev, u8 ibport)
147 {
148         const struct ib_gid_attr *attr;
149         int rc = 0;
150
151         attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, 0);
152         if (IS_ERR(attr))
153                 return -ENODEV;
154
155         if (attr->ndev)
156                 memcpy(smcibdev->mac[ibport - 1], attr->ndev->dev_addr,
157                        ETH_ALEN);
158         else
159                 rc = -ENODEV;
160
161         rdma_put_gid_attr(attr);
162         return rc;
163 }
164
165 /* Create an identifier unique for this instance of SMC-R.
166  * The MAC-address of the first active registered IB device
167  * plus a random 2-byte number is used to create this identifier.
168  * This name is delivered to the peer during connection initialization.
169  */
170 static inline void smc_ib_define_local_systemid(struct smc_ib_device *smcibdev,
171                                                 u8 ibport)
172 {
173         memcpy(&local_systemid[2], &smcibdev->mac[ibport - 1],
174                sizeof(smcibdev->mac[ibport - 1]));
175         get_random_bytes(&local_systemid[0], 2);
176 }
177
178 bool smc_ib_port_active(struct smc_ib_device *smcibdev, u8 ibport)
179 {
180         return smcibdev->pattr[ibport - 1].state == IB_PORT_ACTIVE;
181 }
182
183 /* determine the gid for an ib-device port and vlan id */
184 int smc_ib_determine_gid(struct smc_ib_device *smcibdev, u8 ibport,
185                          unsigned short vlan_id, u8 gid[], u8 *sgid_index)
186 {
187         const struct ib_gid_attr *attr;
188         int i;
189
190         for (i = 0; i < smcibdev->pattr[ibport - 1].gid_tbl_len; i++) {
191                 attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, i);
192                 if (IS_ERR(attr))
193                         continue;
194
195                 if (attr->ndev &&
196                     ((!vlan_id && !is_vlan_dev(attr->ndev)) ||
197                      (vlan_id && is_vlan_dev(attr->ndev) &&
198                       vlan_dev_vlan_id(attr->ndev) == vlan_id)) &&
199                     attr->gid_type == IB_GID_TYPE_ROCE) {
200                         if (gid)
201                                 memcpy(gid, &attr->gid, SMC_GID_SIZE);
202                         if (sgid_index)
203                                 *sgid_index = attr->index;
204                         rdma_put_gid_attr(attr);
205                         return 0;
206                 }
207                 rdma_put_gid_attr(attr);
208         }
209         return -ENODEV;
210 }
211
212 static int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport)
213 {
214         int rc;
215
216         memset(&smcibdev->pattr[ibport - 1], 0,
217                sizeof(smcibdev->pattr[ibport - 1]));
218         rc = ib_query_port(smcibdev->ibdev, ibport,
219                            &smcibdev->pattr[ibport - 1]);
220         if (rc)
221                 goto out;
222         /* the SMC protocol requires specification of the RoCE MAC address */
223         rc = smc_ib_fill_mac(smcibdev, ibport);
224         if (rc)
225                 goto out;
226         if (!strncmp(local_systemid, SMC_LOCAL_SYSTEMID_RESET,
227                      sizeof(local_systemid)) &&
228             smc_ib_port_active(smcibdev, ibport))
229                 /* create unique system identifier */
230                 smc_ib_define_local_systemid(smcibdev, ibport);
231 out:
232         return rc;
233 }
234
235 /* process context wrapper for might_sleep smc_ib_remember_port_attr */
236 static void smc_ib_port_event_work(struct work_struct *work)
237 {
238         struct smc_ib_device *smcibdev = container_of(
239                 work, struct smc_ib_device, port_event_work);
240         u8 port_idx;
241
242         for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) {
243                 smc_ib_remember_port_attr(smcibdev, port_idx + 1);
244                 clear_bit(port_idx, &smcibdev->port_event_mask);
245                 if (!smc_ib_port_active(smcibdev, port_idx + 1))
246                         smc_port_terminate(smcibdev, port_idx + 1);
247         }
248 }
249
250 /* can be called in IRQ context */
251 static void smc_ib_global_event_handler(struct ib_event_handler *handler,
252                                         struct ib_event *ibevent)
253 {
254         struct smc_ib_device *smcibdev;
255         u8 port_idx;
256
257         smcibdev = container_of(handler, struct smc_ib_device, event_handler);
258
259         switch (ibevent->event) {
260         case IB_EVENT_PORT_ERR:
261         case IB_EVENT_DEVICE_FATAL:
262         case IB_EVENT_PORT_ACTIVE:
263                 port_idx = ibevent->element.port_num - 1;
264                 set_bit(port_idx, &smcibdev->port_event_mask);
265                 schedule_work(&smcibdev->port_event_work);
266                 break;
267         default:
268                 break;
269         }
270 }
271
272 void smc_ib_dealloc_protection_domain(struct smc_link *lnk)
273 {
274         if (lnk->roce_pd)
275                 ib_dealloc_pd(lnk->roce_pd);
276         lnk->roce_pd = NULL;
277 }
278
279 int smc_ib_create_protection_domain(struct smc_link *lnk)
280 {
281         int rc;
282
283         lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev, 0);
284         rc = PTR_ERR_OR_ZERO(lnk->roce_pd);
285         if (IS_ERR(lnk->roce_pd))
286                 lnk->roce_pd = NULL;
287         return rc;
288 }
289
290 static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv)
291 {
292         struct smc_link *lnk = (struct smc_link *)priv;
293         struct smc_ib_device *smcibdev = lnk->smcibdev;
294         u8 port_idx;
295
296         switch (ibevent->event) {
297         case IB_EVENT_DEVICE_FATAL:
298         case IB_EVENT_GID_CHANGE:
299         case IB_EVENT_PORT_ERR:
300         case IB_EVENT_QP_ACCESS_ERR:
301                 port_idx = ibevent->element.qp->port - 1;
302                 set_bit(port_idx, &smcibdev->port_event_mask);
303                 schedule_work(&smcibdev->port_event_work);
304                 break;
305         default:
306                 break;
307         }
308 }
309
310 void smc_ib_destroy_queue_pair(struct smc_link *lnk)
311 {
312         if (lnk->roce_qp)
313                 ib_destroy_qp(lnk->roce_qp);
314         lnk->roce_qp = NULL;
315 }
316
317 /* create a queue pair within the protection domain for a link */
318 int smc_ib_create_queue_pair(struct smc_link *lnk)
319 {
320         struct ib_qp_init_attr qp_attr = {
321                 .event_handler = smc_ib_qp_event_handler,
322                 .qp_context = lnk,
323                 .send_cq = lnk->smcibdev->roce_cq_send,
324                 .recv_cq = lnk->smcibdev->roce_cq_recv,
325                 .srq = NULL,
326                 .cap = {
327                                 /* include unsolicited rdma_writes as well,
328                                  * there are max. 2 RDMA_WRITE per 1 WR_SEND
329                                  */
330                         .max_send_wr = SMC_WR_BUF_CNT * 3,
331                         .max_recv_wr = SMC_WR_BUF_CNT * 3,
332                         .max_send_sge = SMC_IB_MAX_SEND_SGE,
333                         .max_recv_sge = 1,
334                 },
335                 .sq_sig_type = IB_SIGNAL_REQ_WR,
336                 .qp_type = IB_QPT_RC,
337         };
338         int rc;
339
340         lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr);
341         rc = PTR_ERR_OR_ZERO(lnk->roce_qp);
342         if (IS_ERR(lnk->roce_qp))
343                 lnk->roce_qp = NULL;
344         else
345                 smc_wr_remember_qp_attr(lnk);
346         return rc;
347 }
348
349 void smc_ib_put_memory_region(struct ib_mr *mr)
350 {
351         ib_dereg_mr(mr);
352 }
353
354 static int smc_ib_map_mr_sg(struct smc_buf_desc *buf_slot)
355 {
356         unsigned int offset = 0;
357         int sg_num;
358
359         /* map the largest prefix of a dma mapped SG list */
360         sg_num = ib_map_mr_sg(buf_slot->mr_rx[SMC_SINGLE_LINK],
361                               buf_slot->sgt[SMC_SINGLE_LINK].sgl,
362                               buf_slot->sgt[SMC_SINGLE_LINK].orig_nents,
363                               &offset, PAGE_SIZE);
364
365         return sg_num;
366 }
367
368 /* Allocate a memory region and map the dma mapped SG list of buf_slot */
369 int smc_ib_get_memory_region(struct ib_pd *pd, int access_flags,
370                              struct smc_buf_desc *buf_slot)
371 {
372         if (buf_slot->mr_rx[SMC_SINGLE_LINK])
373                 return 0; /* already done */
374
375         buf_slot->mr_rx[SMC_SINGLE_LINK] =
376                 ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, 1 << buf_slot->order);
377         if (IS_ERR(buf_slot->mr_rx[SMC_SINGLE_LINK])) {
378                 int rc;
379
380                 rc = PTR_ERR(buf_slot->mr_rx[SMC_SINGLE_LINK]);
381                 buf_slot->mr_rx[SMC_SINGLE_LINK] = NULL;
382                 return rc;
383         }
384
385         if (smc_ib_map_mr_sg(buf_slot) != 1)
386                 return -EINVAL;
387
388         return 0;
389 }
390
391 /* synchronize buffer usage for cpu access */
392 void smc_ib_sync_sg_for_cpu(struct smc_ib_device *smcibdev,
393                             struct smc_buf_desc *buf_slot,
394                             enum dma_data_direction data_direction)
395 {
396         struct scatterlist *sg;
397         unsigned int i;
398
399         /* for now there is just one DMA address */
400         for_each_sg(buf_slot->sgt[SMC_SINGLE_LINK].sgl, sg,
401                     buf_slot->sgt[SMC_SINGLE_LINK].nents, i) {
402                 if (!sg_dma_len(sg))
403                         break;
404                 ib_dma_sync_single_for_cpu(smcibdev->ibdev,
405                                            sg_dma_address(sg),
406                                            sg_dma_len(sg),
407                                            data_direction);
408         }
409 }
410
411 /* synchronize buffer usage for device access */
412 void smc_ib_sync_sg_for_device(struct smc_ib_device *smcibdev,
413                                struct smc_buf_desc *buf_slot,
414                                enum dma_data_direction data_direction)
415 {
416         struct scatterlist *sg;
417         unsigned int i;
418
419         /* for now there is just one DMA address */
420         for_each_sg(buf_slot->sgt[SMC_SINGLE_LINK].sgl, sg,
421                     buf_slot->sgt[SMC_SINGLE_LINK].nents, i) {
422                 if (!sg_dma_len(sg))
423                         break;
424                 ib_dma_sync_single_for_device(smcibdev->ibdev,
425                                               sg_dma_address(sg),
426                                               sg_dma_len(sg),
427                                               data_direction);
428         }
429 }
430
431 /* Map a new TX or RX buffer SG-table to DMA */
432 int smc_ib_buf_map_sg(struct smc_ib_device *smcibdev,
433                       struct smc_buf_desc *buf_slot,
434                       enum dma_data_direction data_direction)
435 {
436         int mapped_nents;
437
438         mapped_nents = ib_dma_map_sg(smcibdev->ibdev,
439                                      buf_slot->sgt[SMC_SINGLE_LINK].sgl,
440                                      buf_slot->sgt[SMC_SINGLE_LINK].orig_nents,
441                                      data_direction);
442         if (!mapped_nents)
443                 return -ENOMEM;
444
445         return mapped_nents;
446 }
447
448 void smc_ib_buf_unmap_sg(struct smc_ib_device *smcibdev,
449                          struct smc_buf_desc *buf_slot,
450                          enum dma_data_direction data_direction)
451 {
452         if (!buf_slot->sgt[SMC_SINGLE_LINK].sgl->dma_address)
453                 return; /* already unmapped */
454
455         ib_dma_unmap_sg(smcibdev->ibdev,
456                         buf_slot->sgt[SMC_SINGLE_LINK].sgl,
457                         buf_slot->sgt[SMC_SINGLE_LINK].orig_nents,
458                         data_direction);
459         buf_slot->sgt[SMC_SINGLE_LINK].sgl->dma_address = 0;
460 }
461
462 long smc_ib_setup_per_ibdev(struct smc_ib_device *smcibdev)
463 {
464         struct ib_cq_init_attr cqattr = {
465                 .cqe = SMC_MAX_CQE, .comp_vector = 0 };
466         int cqe_size_order, smc_order;
467         long rc;
468
469         /* the calculated number of cq entries fits to mlx5 cq allocation */
470         cqe_size_order = cache_line_size() == 128 ? 7 : 6;
471         smc_order = MAX_ORDER - cqe_size_order - 1;
472         if (SMC_MAX_CQE + 2 > (0x00000001 << smc_order) * PAGE_SIZE)
473                 cqattr.cqe = (0x00000001 << smc_order) * PAGE_SIZE - 2;
474         smcibdev->roce_cq_send = ib_create_cq(smcibdev->ibdev,
475                                               smc_wr_tx_cq_handler, NULL,
476                                               smcibdev, &cqattr);
477         rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_send);
478         if (IS_ERR(smcibdev->roce_cq_send)) {
479                 smcibdev->roce_cq_send = NULL;
480                 return rc;
481         }
482         smcibdev->roce_cq_recv = ib_create_cq(smcibdev->ibdev,
483                                               smc_wr_rx_cq_handler, NULL,
484                                               smcibdev, &cqattr);
485         rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_recv);
486         if (IS_ERR(smcibdev->roce_cq_recv)) {
487                 smcibdev->roce_cq_recv = NULL;
488                 goto err;
489         }
490         smc_wr_add_dev(smcibdev);
491         smcibdev->initialized = 1;
492         return rc;
493
494 err:
495         ib_destroy_cq(smcibdev->roce_cq_send);
496         return rc;
497 }
498
499 static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev)
500 {
501         if (!smcibdev->initialized)
502                 return;
503         smcibdev->initialized = 0;
504         smc_wr_remove_dev(smcibdev);
505         ib_destroy_cq(smcibdev->roce_cq_recv);
506         ib_destroy_cq(smcibdev->roce_cq_send);
507 }
508
509 static struct ib_client smc_ib_client;
510
511 /* callback function for ib_register_client() */
512 static void smc_ib_add_dev(struct ib_device *ibdev)
513 {
514         struct smc_ib_device *smcibdev;
515         u8 port_cnt;
516         int i;
517
518         if (ibdev->node_type != RDMA_NODE_IB_CA)
519                 return;
520
521         smcibdev = kzalloc(sizeof(*smcibdev), GFP_KERNEL);
522         if (!smcibdev)
523                 return;
524
525         smcibdev->ibdev = ibdev;
526         INIT_WORK(&smcibdev->port_event_work, smc_ib_port_event_work);
527
528         spin_lock(&smc_ib_devices.lock);
529         list_add_tail(&smcibdev->list, &smc_ib_devices.list);
530         spin_unlock(&smc_ib_devices.lock);
531         ib_set_client_data(ibdev, &smc_ib_client, smcibdev);
532         INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev,
533                               smc_ib_global_event_handler);
534         ib_register_event_handler(&smcibdev->event_handler);
535
536         /* trigger reading of the port attributes */
537         port_cnt = smcibdev->ibdev->phys_port_cnt;
538         for (i = 0;
539              i < min_t(size_t, port_cnt, SMC_MAX_PORTS);
540              i++) {
541                 set_bit(i, &smcibdev->port_event_mask);
542                 /* determine pnetids of the port */
543                 smc_pnetid_by_dev_port(ibdev->dev.parent, i,
544                                        smcibdev->pnetid[i]);
545         }
546         schedule_work(&smcibdev->port_event_work);
547 }
548
549 /* callback function for ib_register_client() */
550 static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data)
551 {
552         struct smc_ib_device *smcibdev;
553
554         smcibdev = ib_get_client_data(ibdev, &smc_ib_client);
555         ib_set_client_data(ibdev, &smc_ib_client, NULL);
556         spin_lock(&smc_ib_devices.lock);
557         list_del_init(&smcibdev->list); /* remove from smc_ib_devices */
558         spin_unlock(&smc_ib_devices.lock);
559         smc_pnet_remove_by_ibdev(smcibdev);
560         smc_ib_cleanup_per_ibdev(smcibdev);
561         ib_unregister_event_handler(&smcibdev->event_handler);
562         kfree(smcibdev);
563 }
564
565 static struct ib_client smc_ib_client = {
566         .name   = "smc_ib",
567         .add    = smc_ib_add_dev,
568         .remove = smc_ib_remove_dev,
569 };
570
571 int __init smc_ib_register_client(void)
572 {
573         return ib_register_client(&smc_ib_client);
574 }
575
576 void smc_ib_unregister_client(void)
577 {
578         ib_unregister_client(&smc_ib_client);
579 }