1
2
3
4
5
6#include <linux/init.h>
7#include <linux/errno.h>
8#include <linux/netdevice.h>
9#include <linux/inetdevice.h>
10#include <net/net_namespace.h>
11#include <linux/rtnetlink.h>
12#include <linux/if_arp.h>
13#include <linux/list.h>
14#include <linux/kernel.h>
15#include <linux/sched.h>
16#include <linux/module.h>
17#include <linux/dma-mapping.h>
18
19#include <net/addrconf.h>
20#include <rdma/ib_verbs.h>
21#include <rdma/ib_user_verbs.h>
22#include <rdma/rdma_netlink.h>
23#include <linux/kthread.h>
24
25#include "siw.h"
26#include "siw_verbs.h"
27
28MODULE_AUTHOR("Bernard Metzler");
29MODULE_DESCRIPTION("Software iWARP Driver");
30MODULE_LICENSE("Dual BSD/GPL");
31
32
33const bool zcopy_tx = true;
34
35
36
37
38
39const bool try_gso;
40
41
42const bool loopback_enabled = true;
43
44
45const bool mpa_crc_required;
46
47
48const bool mpa_crc_strict;
49
50
51const bool siw_tcp_nagle;
52
53
54u_char mpa_version = MPA_REVISION_2;
55
56
57
58
59const bool peer_to_peer;
60
61struct task_struct *siw_tx_thread[NR_CPUS];
62struct crypto_shash *siw_crypto_shash;
63
64static int siw_device_register(struct siw_device *sdev, const char *name)
65{
66 struct ib_device *base_dev = &sdev->base_dev;
67 static int dev_id = 1;
68 int rv;
69
70 sdev->vendor_part_id = dev_id++;
71
72 rv = ib_register_device(base_dev, name);
73 if (rv) {
74 pr_warn("siw: device registration error %d\n", rv);
75 return rv;
76 }
77
78 siw_dbg(base_dev, "HWaddr=%pM\n", sdev->netdev->dev_addr);
79
80 return 0;
81}
82
83static void siw_device_cleanup(struct ib_device *base_dev)
84{
85 struct siw_device *sdev = to_siw_dev(base_dev);
86
87 xa_destroy(&sdev->qp_xa);
88 xa_destroy(&sdev->mem_xa);
89}
90
91static int siw_create_tx_threads(void)
92{
93 int cpu, assigned = 0;
94
95 for_each_online_cpu(cpu) {
96
97 if (cpu % cpumask_weight(topology_sibling_cpumask(cpu)))
98 continue;
99
100 siw_tx_thread[cpu] =
101 kthread_create(siw_run_sq, (unsigned long *)(long)cpu,
102 "siw_tx/%d", cpu);
103 if (IS_ERR(siw_tx_thread[cpu])) {
104 siw_tx_thread[cpu] = NULL;
105 continue;
106 }
107 kthread_bind(siw_tx_thread[cpu], cpu);
108
109 wake_up_process(siw_tx_thread[cpu]);
110 assigned++;
111 }
112 return assigned;
113}
114
115static int siw_dev_qualified(struct net_device *netdev)
116{
117
118
119
120
121
122 if (netdev->type == ARPHRD_ETHER || netdev->type == ARPHRD_IEEE802 ||
123 (netdev->type == ARPHRD_LOOPBACK && loopback_enabled))
124 return 1;
125
126 return 0;
127}
128
129static DEFINE_PER_CPU(atomic_t, siw_use_cnt);
130
131static struct {
132 struct cpumask **tx_valid_cpus;
133 int num_nodes;
134} siw_cpu_info;
135
136static int siw_init_cpulist(void)
137{
138 int i, num_nodes = num_possible_nodes();
139
140 memset(siw_tx_thread, 0, sizeof(siw_tx_thread));
141
142 siw_cpu_info.num_nodes = num_nodes;
143
144 siw_cpu_info.tx_valid_cpus =
145 kcalloc(num_nodes, sizeof(struct cpumask *), GFP_KERNEL);
146 if (!siw_cpu_info.tx_valid_cpus) {
147 siw_cpu_info.num_nodes = 0;
148 return -ENOMEM;
149 }
150 for (i = 0; i < siw_cpu_info.num_nodes; i++) {
151 siw_cpu_info.tx_valid_cpus[i] =
152 kzalloc(sizeof(struct cpumask), GFP_KERNEL);
153 if (!siw_cpu_info.tx_valid_cpus[i])
154 goto out_err;
155
156 cpumask_clear(siw_cpu_info.tx_valid_cpus[i]);
157 }
158 for_each_possible_cpu(i)
159 cpumask_set_cpu(i, siw_cpu_info.tx_valid_cpus[cpu_to_node(i)]);
160
161 return 0;
162
163out_err:
164 siw_cpu_info.num_nodes = 0;
165 while (--i >= 0)
166 kfree(siw_cpu_info.tx_valid_cpus[i]);
167 kfree(siw_cpu_info.tx_valid_cpus);
168 siw_cpu_info.tx_valid_cpus = NULL;
169
170 return -ENOMEM;
171}
172
173static void siw_destroy_cpulist(void)
174{
175 int i = 0;
176
177 while (i < siw_cpu_info.num_nodes)
178 kfree(siw_cpu_info.tx_valid_cpus[i++]);
179
180 kfree(siw_cpu_info.tx_valid_cpus);
181}
182
183
184
185
186
187int siw_get_tx_cpu(struct siw_device *sdev)
188{
189 const struct cpumask *tx_cpumask;
190 int i, num_cpus, cpu, min_use, node = sdev->numa_node, tx_cpu = -1;
191
192 if (node < 0)
193 tx_cpumask = cpu_online_mask;
194 else
195 tx_cpumask = siw_cpu_info.tx_valid_cpus[node];
196
197 num_cpus = cpumask_weight(tx_cpumask);
198 if (!num_cpus) {
199
200 tx_cpumask = cpu_online_mask;
201 num_cpus = cpumask_weight(tx_cpumask);
202 }
203 if (!num_cpus)
204 goto out;
205
206 cpu = cpumask_first(tx_cpumask);
207
208 for (i = 0, min_use = SIW_MAX_QP; i < num_cpus;
209 i++, cpu = cpumask_next(cpu, tx_cpumask)) {
210 int usage;
211
212
213 if (!siw_tx_thread[cpu])
214 continue;
215
216 usage = atomic_read(&per_cpu(siw_use_cnt, cpu));
217 if (usage <= min_use) {
218 tx_cpu = cpu;
219 min_use = usage;
220 }
221 }
222 siw_dbg(&sdev->base_dev,
223 "tx cpu %d, node %d, %d qp's\n", tx_cpu, node, min_use);
224
225out:
226 if (tx_cpu >= 0)
227 atomic_inc(&per_cpu(siw_use_cnt, tx_cpu));
228 else
229 pr_warn("siw: no tx cpu found\n");
230
231 return tx_cpu;
232}
233
234void siw_put_tx_cpu(int cpu)
235{
236 atomic_dec(&per_cpu(siw_use_cnt, cpu));
237}
238
239static struct ib_qp *siw_get_base_qp(struct ib_device *base_dev, int id)
240{
241 struct siw_qp *qp = siw_qp_id2obj(to_siw_dev(base_dev), id);
242
243 if (qp) {
244
245
246
247 siw_qp_put(qp);
248 return &qp->base_qp;
249 }
250 return NULL;
251}
252
253static const struct ib_device_ops siw_device_ops = {
254 .owner = THIS_MODULE,
255 .uverbs_abi_ver = SIW_ABI_VERSION,
256 .driver_id = RDMA_DRIVER_SIW,
257
258 .alloc_mr = siw_alloc_mr,
259 .alloc_pd = siw_alloc_pd,
260 .alloc_ucontext = siw_alloc_ucontext,
261 .create_cq = siw_create_cq,
262 .create_qp = siw_create_qp,
263 .create_srq = siw_create_srq,
264 .dealloc_driver = siw_device_cleanup,
265 .dealloc_pd = siw_dealloc_pd,
266 .dealloc_ucontext = siw_dealloc_ucontext,
267 .dereg_mr = siw_dereg_mr,
268 .destroy_cq = siw_destroy_cq,
269 .destroy_qp = siw_destroy_qp,
270 .destroy_srq = siw_destroy_srq,
271 .get_dma_mr = siw_get_dma_mr,
272 .get_port_immutable = siw_get_port_immutable,
273 .iw_accept = siw_accept,
274 .iw_add_ref = siw_qp_get_ref,
275 .iw_connect = siw_connect,
276 .iw_create_listen = siw_create_listen,
277 .iw_destroy_listen = siw_destroy_listen,
278 .iw_get_qp = siw_get_base_qp,
279 .iw_reject = siw_reject,
280 .iw_rem_ref = siw_qp_put_ref,
281 .map_mr_sg = siw_map_mr_sg,
282 .mmap = siw_mmap,
283 .mmap_free = siw_mmap_free,
284 .modify_qp = siw_verbs_modify_qp,
285 .modify_srq = siw_modify_srq,
286 .poll_cq = siw_poll_cq,
287 .post_recv = siw_post_receive,
288 .post_send = siw_post_send,
289 .post_srq_recv = siw_post_srq_recv,
290 .query_device = siw_query_device,
291 .query_gid = siw_query_gid,
292 .query_port = siw_query_port,
293 .query_qp = siw_query_qp,
294 .query_srq = siw_query_srq,
295 .req_notify_cq = siw_req_notify_cq,
296 .reg_user_mr = siw_reg_user_mr,
297
298 INIT_RDMA_OBJ_SIZE(ib_cq, siw_cq, base_cq),
299 INIT_RDMA_OBJ_SIZE(ib_pd, siw_pd, base_pd),
300 INIT_RDMA_OBJ_SIZE(ib_srq, siw_srq, base_srq),
301 INIT_RDMA_OBJ_SIZE(ib_ucontext, siw_ucontext, base_ucontext),
302};
303
304static struct siw_device *siw_device_create(struct net_device *netdev)
305{
306 struct siw_device *sdev = NULL;
307 struct ib_device *base_dev;
308 struct device *parent = netdev->dev.parent;
309 int rv;
310
311 if (!parent) {
312
313
314
315
316
317
318
319 if (netdev->type != ARPHRD_LOOPBACK) {
320 pr_warn("siw: device %s error: no parent device\n",
321 netdev->name);
322 return NULL;
323 }
324 parent = &netdev->dev;
325 }
326 sdev = ib_alloc_device(siw_device, base_dev);
327 if (!sdev)
328 return NULL;
329
330 base_dev = &sdev->base_dev;
331
332 sdev->netdev = netdev;
333
334 if (netdev->type != ARPHRD_LOOPBACK) {
335 addrconf_addr_eui48((unsigned char *)&base_dev->node_guid,
336 netdev->dev_addr);
337 } else {
338
339
340
341
342 size_t len = min_t(size_t, strlen(base_dev->name), 6);
343 char addr[6] = { };
344
345 memcpy(addr, base_dev->name, len);
346 addrconf_addr_eui48((unsigned char *)&base_dev->node_guid,
347 addr);
348 }
349 base_dev->uverbs_cmd_mask =
350 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE) |
351 (1ull << IB_USER_VERBS_CMD_QUERY_PORT) |
352 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT) |
353 (1ull << IB_USER_VERBS_CMD_ALLOC_PD) |
354 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD) |
355 (1ull << IB_USER_VERBS_CMD_REG_MR) |
356 (1ull << IB_USER_VERBS_CMD_DEREG_MR) |
357 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
358 (1ull << IB_USER_VERBS_CMD_CREATE_CQ) |
359 (1ull << IB_USER_VERBS_CMD_POLL_CQ) |
360 (1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ) |
361 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ) |
362 (1ull << IB_USER_VERBS_CMD_CREATE_QP) |
363 (1ull << IB_USER_VERBS_CMD_QUERY_QP) |
364 (1ull << IB_USER_VERBS_CMD_MODIFY_QP) |
365 (1ull << IB_USER_VERBS_CMD_DESTROY_QP) |
366 (1ull << IB_USER_VERBS_CMD_POST_SEND) |
367 (1ull << IB_USER_VERBS_CMD_POST_RECV) |
368 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ) |
369 (1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV) |
370 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ) |
371 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ) |
372 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ);
373
374 base_dev->node_type = RDMA_NODE_RNIC;
375 memcpy(base_dev->node_desc, SIW_NODE_DESC_COMMON,
376 sizeof(SIW_NODE_DESC_COMMON));
377
378
379
380
381
382
383 base_dev->phys_port_cnt = 1;
384 base_dev->dev.parent = parent;
385 base_dev->dev.dma_ops = &dma_virt_ops;
386 base_dev->dev.dma_parms = &sdev->dma_parms;
387 sdev->dma_parms = (struct device_dma_parameters)
388 { .max_segment_size = SZ_2G };
389 base_dev->num_comp_vectors = num_possible_cpus();
390
391 xa_init_flags(&sdev->qp_xa, XA_FLAGS_ALLOC1);
392 xa_init_flags(&sdev->mem_xa, XA_FLAGS_ALLOC1);
393
394 ib_set_device_ops(base_dev, &siw_device_ops);
395 rv = ib_device_set_netdev(base_dev, netdev, 1);
396 if (rv)
397 goto error;
398
399 memcpy(base_dev->iw_ifname, netdev->name,
400 sizeof(base_dev->iw_ifname));
401
402
403 base_dev->iw_driver_flags = IW_F_NO_PORT_MAP,
404
405 sdev->attrs.max_qp = SIW_MAX_QP;
406 sdev->attrs.max_qp_wr = SIW_MAX_QP_WR;
407 sdev->attrs.max_ord = SIW_MAX_ORD_QP;
408 sdev->attrs.max_ird = SIW_MAX_IRD_QP;
409 sdev->attrs.max_sge = SIW_MAX_SGE;
410 sdev->attrs.max_sge_rd = SIW_MAX_SGE_RD;
411 sdev->attrs.max_cq = SIW_MAX_CQ;
412 sdev->attrs.max_cqe = SIW_MAX_CQE;
413 sdev->attrs.max_mr = SIW_MAX_MR;
414 sdev->attrs.max_pd = SIW_MAX_PD;
415 sdev->attrs.max_mw = SIW_MAX_MW;
416 sdev->attrs.max_srq = SIW_MAX_SRQ;
417 sdev->attrs.max_srq_wr = SIW_MAX_SRQ_WR;
418 sdev->attrs.max_srq_sge = SIW_MAX_SGE;
419
420 INIT_LIST_HEAD(&sdev->cep_list);
421 INIT_LIST_HEAD(&sdev->qp_list);
422
423 atomic_set(&sdev->num_ctx, 0);
424 atomic_set(&sdev->num_srq, 0);
425 atomic_set(&sdev->num_qp, 0);
426 atomic_set(&sdev->num_cq, 0);
427 atomic_set(&sdev->num_mr, 0);
428 atomic_set(&sdev->num_pd, 0);
429
430 sdev->numa_node = dev_to_node(parent);
431 spin_lock_init(&sdev->lock);
432
433 return sdev;
434error:
435 ib_dealloc_device(base_dev);
436
437 return NULL;
438}
439
440
441
442
443
444static void siw_netdev_down(struct work_struct *work)
445{
446 struct siw_device *sdev =
447 container_of(work, struct siw_device, netdev_down);
448
449 struct siw_qp_attrs qp_attrs;
450 struct list_head *pos, *tmp;
451
452 memset(&qp_attrs, 0, sizeof(qp_attrs));
453 qp_attrs.state = SIW_QP_STATE_ERROR;
454
455 list_for_each_safe(pos, tmp, &sdev->qp_list) {
456 struct siw_qp *qp = list_entry(pos, struct siw_qp, devq);
457
458 down_write(&qp->state_lock);
459 WARN_ON(siw_qp_modify(qp, &qp_attrs, SIW_QP_ATTR_STATE));
460 up_write(&qp->state_lock);
461 }
462 ib_device_put(&sdev->base_dev);
463}
464
465static void siw_device_goes_down(struct siw_device *sdev)
466{
467 if (ib_device_try_get(&sdev->base_dev)) {
468 INIT_WORK(&sdev->netdev_down, siw_netdev_down);
469 schedule_work(&sdev->netdev_down);
470 }
471}
472
473static int siw_netdev_event(struct notifier_block *nb, unsigned long event,
474 void *arg)
475{
476 struct net_device *netdev = netdev_notifier_info_to_dev(arg);
477 struct ib_device *base_dev;
478 struct siw_device *sdev;
479
480 dev_dbg(&netdev->dev, "siw: event %lu\n", event);
481
482 if (dev_net(netdev) != &init_net)
483 return NOTIFY_OK;
484
485 base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW);
486 if (!base_dev)
487 return NOTIFY_OK;
488
489 sdev = to_siw_dev(base_dev);
490
491 switch (event) {
492 case NETDEV_UP:
493 sdev->state = IB_PORT_ACTIVE;
494 siw_port_event(sdev, 1, IB_EVENT_PORT_ACTIVE);
495 break;
496
497 case NETDEV_GOING_DOWN:
498 siw_device_goes_down(sdev);
499 break;
500
501 case NETDEV_DOWN:
502 sdev->state = IB_PORT_DOWN;
503 siw_port_event(sdev, 1, IB_EVENT_PORT_ERR);
504 break;
505
506 case NETDEV_REGISTER:
507
508
509
510
511
512 siw_dbg(base_dev, "unexpected NETDEV_REGISTER event\n");
513 break;
514
515 case NETDEV_UNREGISTER:
516 ib_unregister_device_queued(&sdev->base_dev);
517 break;
518
519 case NETDEV_CHANGEADDR:
520 siw_port_event(sdev, 1, IB_EVENT_LID_CHANGE);
521 break;
522
523
524
525 case NETDEV_CHANGEMTU:
526 case NETDEV_CHANGE:
527 break;
528
529 default:
530 break;
531 }
532 ib_device_put(&sdev->base_dev);
533
534 return NOTIFY_OK;
535}
536
537static struct notifier_block siw_netdev_nb = {
538 .notifier_call = siw_netdev_event,
539};
540
541static int siw_newlink(const char *basedev_name, struct net_device *netdev)
542{
543 struct ib_device *base_dev;
544 struct siw_device *sdev = NULL;
545 int rv = -ENOMEM;
546
547 if (!siw_dev_qualified(netdev))
548 return -EINVAL;
549
550 base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW);
551 if (base_dev) {
552 ib_device_put(base_dev);
553 return -EEXIST;
554 }
555 sdev = siw_device_create(netdev);
556 if (sdev) {
557 dev_dbg(&netdev->dev, "siw: new device\n");
558
559 if (netif_running(netdev) && netif_carrier_ok(netdev))
560 sdev->state = IB_PORT_ACTIVE;
561 else
562 sdev->state = IB_PORT_DOWN;
563
564 rv = siw_device_register(sdev, basedev_name);
565 if (rv)
566 ib_dealloc_device(&sdev->base_dev);
567 }
568 return rv;
569}
570
571static struct rdma_link_ops siw_link_ops = {
572 .type = "siw",
573 .newlink = siw_newlink,
574};
575
576
577
578
579
580static __init int siw_init_module(void)
581{
582 int rv;
583 int nr_cpu;
584
585 if (SENDPAGE_THRESH < SIW_MAX_INLINE) {
586 pr_info("siw: sendpage threshold too small: %u\n",
587 (int)SENDPAGE_THRESH);
588 rv = -EINVAL;
589 goto out_error;
590 }
591 rv = siw_init_cpulist();
592 if (rv)
593 goto out_error;
594
595 rv = siw_cm_init();
596 if (rv)
597 goto out_error;
598
599 if (!siw_create_tx_threads()) {
600 pr_info("siw: Could not start any TX thread\n");
601 rv = -ENOMEM;
602 goto out_error;
603 }
604
605
606
607
608 siw_crypto_shash = crypto_alloc_shash("crc32c", 0, 0);
609 if (IS_ERR(siw_crypto_shash)) {
610 pr_info("siw: Loading CRC32c failed: %ld\n",
611 PTR_ERR(siw_crypto_shash));
612 siw_crypto_shash = NULL;
613 if (mpa_crc_required) {
614 rv = -EOPNOTSUPP;
615 goto out_error;
616 }
617 }
618 rv = register_netdevice_notifier(&siw_netdev_nb);
619 if (rv)
620 goto out_error;
621
622 rdma_link_register(&siw_link_ops);
623
624 pr_info("SoftiWARP attached\n");
625 return 0;
626
627out_error:
628 for (nr_cpu = 0; nr_cpu < nr_cpu_ids; nr_cpu++) {
629 if (siw_tx_thread[nr_cpu]) {
630 siw_stop_tx_thread(nr_cpu);
631 siw_tx_thread[nr_cpu] = NULL;
632 }
633 }
634 if (siw_crypto_shash)
635 crypto_free_shash(siw_crypto_shash);
636
637 pr_info("SoftIWARP attach failed. Error: %d\n", rv);
638
639 siw_cm_exit();
640 siw_destroy_cpulist();
641
642 return rv;
643}
644
645static void __exit siw_exit_module(void)
646{
647 int cpu;
648
649 for_each_possible_cpu(cpu) {
650 if (siw_tx_thread[cpu]) {
651 siw_stop_tx_thread(cpu);
652 siw_tx_thread[cpu] = NULL;
653 }
654 }
655 unregister_netdevice_notifier(&siw_netdev_nb);
656 rdma_link_unregister(&siw_link_ops);
657 ib_unregister_driver(RDMA_DRIVER_SIW);
658
659 siw_cm_exit();
660
661 siw_destroy_cpulist();
662
663 if (siw_crypto_shash)
664 crypto_free_shash(siw_crypto_shash);
665
666 pr_info("SoftiWARP detached\n");
667}
668
669module_init(siw_init_module);
670module_exit(siw_exit_module);
671
672MODULE_ALIAS_RDMA_LINK("siw");
673