1 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
2
3 /* Authors: Bernard Metzler <bmt@zurich.ibm.com> */
4 /* Copyright (c) 2008-2019, IBM Corporation */
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
28 MODULE_AUTHOR("Bernard Metzler");
29 MODULE_DESCRIPTION("Software iWARP Driver");
30 MODULE_LICENSE("Dual BSD/GPL");
31
32 /* transmit from user buffer, if possible */
33 const bool zcopy_tx = true;
34
35 /* Restrict usage of GSO, if hardware peer iwarp is unable to process
36 * large packets. try_gso = true lets siw try to use local GSO,
37 * if peer agrees. Not using GSO severly limits siw maximum tx bandwidth.
38 */
39 const bool try_gso;
40
41 /* Attach siw also with loopback devices */
42 const bool loopback_enabled = true;
43
44 /* We try to negotiate CRC on, if true */
45 const bool mpa_crc_required;
46
47 /* MPA CRC on/off enforced */
48 const bool mpa_crc_strict;
49
50 /* Control TCP_NODELAY socket option */
51 const bool siw_tcp_nagle;
52
53 /* Select MPA version to be used during connection setup */
54 u_char mpa_version = MPA_REVISION_2;
55
56 /* Selects MPA P2P mode (additional handshake during connection
57 * setup, if true.
58 */
59 const bool peer_to_peer;
60
61 struct task_struct *siw_tx_thread[NR_CPUS];
62 struct crypto_shash *siw_crypto_shash;
63
siw_device_register(struct siw_device * sdev,const char * name)64 static 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, NULL);
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->raw_gid);
79 return 0;
80 }
81
siw_device_cleanup(struct ib_device * base_dev)82 static void siw_device_cleanup(struct ib_device *base_dev)
83 {
84 struct siw_device *sdev = to_siw_dev(base_dev);
85
86 xa_destroy(&sdev->qp_xa);
87 xa_destroy(&sdev->mem_xa);
88 }
89
siw_create_tx_threads(void)90 static int siw_create_tx_threads(void)
91 {
92 int cpu, assigned = 0;
93
94 for_each_online_cpu(cpu) {
95 /* Skip HT cores */
96 if (cpu % cpumask_weight(topology_sibling_cpumask(cpu)))
97 continue;
98
99 siw_tx_thread[cpu] =
100 kthread_run_on_cpu(siw_run_sq,
101 (unsigned long *)(long)cpu,
102 cpu, "siw_tx/%u");
103 if (IS_ERR(siw_tx_thread[cpu])) {
104 siw_tx_thread[cpu] = NULL;
105 continue;
106 }
107
108 assigned++;
109 }
110 return assigned;
111 }
112
siw_dev_qualified(struct net_device * netdev)113 static int siw_dev_qualified(struct net_device *netdev)
114 {
115 /*
116 * Additional hardware support can be added here
117 * (e.g. ARPHRD_FDDI, ARPHRD_ATM, ...) - see
118 * <linux/if_arp.h> for type identifiers.
119 */
120 if (netdev->type == ARPHRD_ETHER || netdev->type == ARPHRD_IEEE802 ||
121 netdev->type == ARPHRD_NONE ||
122 (netdev->type == ARPHRD_LOOPBACK && loopback_enabled))
123 return 1;
124
125 return 0;
126 }
127
128 static DEFINE_PER_CPU(atomic_t, siw_use_cnt);
129
130 static struct {
131 struct cpumask **tx_valid_cpus;
132 int num_nodes;
133 } siw_cpu_info;
134
siw_init_cpulist(void)135 static int siw_init_cpulist(void)
136 {
137 int i, num_nodes = nr_node_ids;
138
139 memset(siw_tx_thread, 0, sizeof(siw_tx_thread));
140
141 siw_cpu_info.num_nodes = num_nodes;
142
143 siw_cpu_info.tx_valid_cpus =
144 kcalloc(num_nodes, sizeof(struct cpumask *), GFP_KERNEL);
145 if (!siw_cpu_info.tx_valid_cpus) {
146 siw_cpu_info.num_nodes = 0;
147 return -ENOMEM;
148 }
149 for (i = 0; i < siw_cpu_info.num_nodes; i++) {
150 siw_cpu_info.tx_valid_cpus[i] =
151 kzalloc(sizeof(struct cpumask), GFP_KERNEL);
152 if (!siw_cpu_info.tx_valid_cpus[i])
153 goto out_err;
154
155 cpumask_clear(siw_cpu_info.tx_valid_cpus[i]);
156 }
157 for_each_possible_cpu(i)
158 cpumask_set_cpu(i, siw_cpu_info.tx_valid_cpus[cpu_to_node(i)]);
159
160 return 0;
161
162 out_err:
163 siw_cpu_info.num_nodes = 0;
164 while (--i >= 0)
165 kfree(siw_cpu_info.tx_valid_cpus[i]);
166 kfree(siw_cpu_info.tx_valid_cpus);
167 siw_cpu_info.tx_valid_cpus = NULL;
168
169 return -ENOMEM;
170 }
171
siw_destroy_cpulist(void)172 static void siw_destroy_cpulist(void)
173 {
174 int i = 0;
175
176 while (i < siw_cpu_info.num_nodes)
177 kfree(siw_cpu_info.tx_valid_cpus[i++]);
178
179 kfree(siw_cpu_info.tx_valid_cpus);
180 }
181
182 /*
183 * Choose CPU with least number of active QP's from NUMA node of
184 * TX interface.
185 */
siw_get_tx_cpu(struct siw_device * sdev)186 int siw_get_tx_cpu(struct siw_device *sdev)
187 {
188 const struct cpumask *tx_cpumask;
189 int i, num_cpus, cpu, min_use, node = sdev->numa_node, tx_cpu = -1;
190
191 if (node < 0)
192 tx_cpumask = cpu_online_mask;
193 else
194 tx_cpumask = siw_cpu_info.tx_valid_cpus[node];
195
196 num_cpus = cpumask_weight(tx_cpumask);
197 if (!num_cpus) {
198 /* no CPU on this NUMA node */
199 tx_cpumask = cpu_online_mask;
200 num_cpus = cpumask_weight(tx_cpumask);
201 }
202 if (!num_cpus)
203 goto out;
204
205 cpu = cpumask_first(tx_cpumask);
206
207 for (i = 0, min_use = SIW_MAX_QP; i < num_cpus;
208 i++, cpu = cpumask_next(cpu, tx_cpumask)) {
209 int usage;
210
211 /* Skip any cores which have no TX thread */
212 if (!siw_tx_thread[cpu])
213 continue;
214
215 usage = atomic_read(&per_cpu(siw_use_cnt, cpu));
216 if (usage <= min_use) {
217 tx_cpu = cpu;
218 min_use = usage;
219 }
220 }
221 siw_dbg(&sdev->base_dev,
222 "tx cpu %d, node %d, %d qp's\n", tx_cpu, node, min_use);
223
224 out:
225 if (tx_cpu >= 0)
226 atomic_inc(&per_cpu(siw_use_cnt, tx_cpu));
227 else
228 pr_warn("siw: no tx cpu found\n");
229
230 return tx_cpu;
231 }
232
siw_put_tx_cpu(int cpu)233 void siw_put_tx_cpu(int cpu)
234 {
235 atomic_dec(&per_cpu(siw_use_cnt, cpu));
236 }
237
siw_get_base_qp(struct ib_device * base_dev,int id)238 static struct ib_qp *siw_get_base_qp(struct ib_device *base_dev, int id)
239 {
240 struct siw_qp *qp = siw_qp_id2obj(to_siw_dev(base_dev), id);
241
242 if (qp) {
243 /*
244 * siw_qp_id2obj() increments object reference count
245 */
246 siw_qp_put(qp);
247 return &qp->base_qp;
248 }
249 return NULL;
250 }
251
252 static const struct ib_device_ops siw_device_ops = {
253 .owner = THIS_MODULE,
254 .uverbs_abi_ver = SIW_ABI_VERSION,
255 .driver_id = RDMA_DRIVER_SIW,
256
257 .alloc_mr = siw_alloc_mr,
258 .alloc_pd = siw_alloc_pd,
259 .alloc_ucontext = siw_alloc_ucontext,
260 .create_cq = siw_create_cq,
261 .create_qp = siw_create_qp,
262 .create_srq = siw_create_srq,
263 .dealloc_driver = siw_device_cleanup,
264 .dealloc_pd = siw_dealloc_pd,
265 .dealloc_ucontext = siw_dealloc_ucontext,
266 .dereg_mr = siw_dereg_mr,
267 .destroy_cq = siw_destroy_cq,
268 .destroy_qp = siw_destroy_qp,
269 .destroy_srq = siw_destroy_srq,
270 .get_dma_mr = siw_get_dma_mr,
271 .get_port_immutable = siw_get_port_immutable,
272 .iw_accept = siw_accept,
273 .iw_add_ref = siw_qp_get_ref,
274 .iw_connect = siw_connect,
275 .iw_create_listen = siw_create_listen,
276 .iw_destroy_listen = siw_destroy_listen,
277 .iw_get_qp = siw_get_base_qp,
278 .iw_reject = siw_reject,
279 .iw_rem_ref = siw_qp_put_ref,
280 .map_mr_sg = siw_map_mr_sg,
281 .mmap = siw_mmap,
282 .mmap_free = siw_mmap_free,
283 .modify_qp = siw_verbs_modify_qp,
284 .modify_srq = siw_modify_srq,
285 .poll_cq = siw_poll_cq,
286 .post_recv = siw_post_receive,
287 .post_send = siw_post_send,
288 .post_srq_recv = siw_post_srq_recv,
289 .query_device = siw_query_device,
290 .query_gid = siw_query_gid,
291 .query_port = siw_query_port,
292 .query_qp = siw_query_qp,
293 .query_srq = siw_query_srq,
294 .req_notify_cq = siw_req_notify_cq,
295 .reg_user_mr = siw_reg_user_mr,
296
297 INIT_RDMA_OBJ_SIZE(ib_cq, siw_cq, base_cq),
298 INIT_RDMA_OBJ_SIZE(ib_pd, siw_pd, base_pd),
299 INIT_RDMA_OBJ_SIZE(ib_qp, siw_qp, base_qp),
300 INIT_RDMA_OBJ_SIZE(ib_srq, siw_srq, base_srq),
301 INIT_RDMA_OBJ_SIZE(ib_ucontext, siw_ucontext, base_ucontext),
302 };
303
siw_device_create(struct net_device * netdev)304 static struct siw_device *siw_device_create(struct net_device *netdev)
305 {
306 struct siw_device *sdev = NULL;
307 struct ib_device *base_dev;
308 int rv;
309
310 sdev = ib_alloc_device(siw_device, base_dev);
311 if (!sdev)
312 return NULL;
313
314 base_dev = &sdev->base_dev;
315 sdev->netdev = netdev;
316
317 if (netdev->addr_len) {
318 memcpy(sdev->raw_gid, netdev->dev_addr,
319 min_t(unsigned int, netdev->addr_len, ETH_ALEN));
320 } else {
321 /*
322 * This device does not have a HW address, but
323 * connection mangagement requires a unique gid.
324 */
325 eth_random_addr(sdev->raw_gid);
326 }
327 addrconf_addr_eui48((u8 *)&base_dev->node_guid, sdev->raw_gid);
328
329 base_dev->uverbs_cmd_mask |= BIT_ULL(IB_USER_VERBS_CMD_POST_SEND);
330
331 base_dev->node_type = RDMA_NODE_RNIC;
332 memcpy(base_dev->node_desc, SIW_NODE_DESC_COMMON,
333 sizeof(SIW_NODE_DESC_COMMON));
334
335 /*
336 * Current model (one-to-one device association):
337 * One Softiwarp device per net_device or, equivalently,
338 * per physical port.
339 */
340 base_dev->phys_port_cnt = 1;
341 base_dev->num_comp_vectors = num_possible_cpus();
342
343 xa_init_flags(&sdev->qp_xa, XA_FLAGS_ALLOC1);
344 xa_init_flags(&sdev->mem_xa, XA_FLAGS_ALLOC1);
345
346 ib_set_device_ops(base_dev, &siw_device_ops);
347 rv = ib_device_set_netdev(base_dev, netdev, 1);
348 if (rv)
349 goto error;
350
351 memcpy(base_dev->iw_ifname, netdev->name,
352 sizeof(base_dev->iw_ifname));
353
354 /* Disable TCP port mapping */
355 base_dev->iw_driver_flags = IW_F_NO_PORT_MAP;
356
357 sdev->attrs.max_qp = SIW_MAX_QP;
358 sdev->attrs.max_qp_wr = SIW_MAX_QP_WR;
359 sdev->attrs.max_ord = SIW_MAX_ORD_QP;
360 sdev->attrs.max_ird = SIW_MAX_IRD_QP;
361 sdev->attrs.max_sge = SIW_MAX_SGE;
362 sdev->attrs.max_sge_rd = SIW_MAX_SGE_RD;
363 sdev->attrs.max_cq = SIW_MAX_CQ;
364 sdev->attrs.max_cqe = SIW_MAX_CQE;
365 sdev->attrs.max_mr = SIW_MAX_MR;
366 sdev->attrs.max_pd = SIW_MAX_PD;
367 sdev->attrs.max_mw = SIW_MAX_MW;
368 sdev->attrs.max_srq = SIW_MAX_SRQ;
369 sdev->attrs.max_srq_wr = SIW_MAX_SRQ_WR;
370 sdev->attrs.max_srq_sge = SIW_MAX_SGE;
371
372 INIT_LIST_HEAD(&sdev->cep_list);
373 INIT_LIST_HEAD(&sdev->qp_list);
374
375 atomic_set(&sdev->num_ctx, 0);
376 atomic_set(&sdev->num_srq, 0);
377 atomic_set(&sdev->num_qp, 0);
378 atomic_set(&sdev->num_cq, 0);
379 atomic_set(&sdev->num_mr, 0);
380 atomic_set(&sdev->num_pd, 0);
381
382 sdev->numa_node = dev_to_node(&netdev->dev);
383 spin_lock_init(&sdev->lock);
384
385 return sdev;
386 error:
387 ib_dealloc_device(base_dev);
388
389 return NULL;
390 }
391
392 /*
393 * Network link becomes unavailable. Mark all
394 * affected QP's accordingly.
395 */
siw_netdev_down(struct work_struct * work)396 static void siw_netdev_down(struct work_struct *work)
397 {
398 struct siw_device *sdev =
399 container_of(work, struct siw_device, netdev_down);
400
401 struct siw_qp_attrs qp_attrs;
402 struct list_head *pos, *tmp;
403
404 memset(&qp_attrs, 0, sizeof(qp_attrs));
405 qp_attrs.state = SIW_QP_STATE_ERROR;
406
407 list_for_each_safe(pos, tmp, &sdev->qp_list) {
408 struct siw_qp *qp = list_entry(pos, struct siw_qp, devq);
409
410 down_write(&qp->state_lock);
411 WARN_ON(siw_qp_modify(qp, &qp_attrs, SIW_QP_ATTR_STATE));
412 up_write(&qp->state_lock);
413 }
414 ib_device_put(&sdev->base_dev);
415 }
416
siw_device_goes_down(struct siw_device * sdev)417 static void siw_device_goes_down(struct siw_device *sdev)
418 {
419 if (ib_device_try_get(&sdev->base_dev)) {
420 INIT_WORK(&sdev->netdev_down, siw_netdev_down);
421 schedule_work(&sdev->netdev_down);
422 }
423 }
424
siw_netdev_event(struct notifier_block * nb,unsigned long event,void * arg)425 static int siw_netdev_event(struct notifier_block *nb, unsigned long event,
426 void *arg)
427 {
428 struct net_device *netdev = netdev_notifier_info_to_dev(arg);
429 struct ib_device *base_dev;
430 struct siw_device *sdev;
431
432 dev_dbg(&netdev->dev, "siw: event %lu\n", event);
433
434 base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW);
435 if (!base_dev)
436 return NOTIFY_OK;
437
438 sdev = to_siw_dev(base_dev);
439
440 switch (event) {
441 case NETDEV_UP:
442 sdev->state = IB_PORT_ACTIVE;
443 siw_port_event(sdev, 1, IB_EVENT_PORT_ACTIVE);
444 break;
445
446 case NETDEV_GOING_DOWN:
447 siw_device_goes_down(sdev);
448 break;
449
450 case NETDEV_DOWN:
451 sdev->state = IB_PORT_DOWN;
452 siw_port_event(sdev, 1, IB_EVENT_PORT_ERR);
453 break;
454
455 case NETDEV_REGISTER:
456 /*
457 * Device registration now handled only by
458 * rdma netlink commands. So it shall be impossible
459 * to end up here with a valid siw device.
460 */
461 siw_dbg(base_dev, "unexpected NETDEV_REGISTER event\n");
462 break;
463
464 case NETDEV_UNREGISTER:
465 ib_unregister_device_queued(&sdev->base_dev);
466 break;
467
468 case NETDEV_CHANGEADDR:
469 siw_port_event(sdev, 1, IB_EVENT_LID_CHANGE);
470 break;
471 /*
472 * Todo: Below netdev events are currently not handled.
473 */
474 case NETDEV_CHANGEMTU:
475 case NETDEV_CHANGE:
476 break;
477
478 default:
479 break;
480 }
481 ib_device_put(&sdev->base_dev);
482
483 return NOTIFY_OK;
484 }
485
486 static struct notifier_block siw_netdev_nb = {
487 .notifier_call = siw_netdev_event,
488 };
489
siw_newlink(const char * basedev_name,struct net_device * netdev)490 static int siw_newlink(const char *basedev_name, struct net_device *netdev)
491 {
492 struct ib_device *base_dev;
493 struct siw_device *sdev = NULL;
494 int rv = -ENOMEM;
495
496 if (!siw_dev_qualified(netdev))
497 return -EINVAL;
498
499 base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW);
500 if (base_dev) {
501 ib_device_put(base_dev);
502 return -EEXIST;
503 }
504 sdev = siw_device_create(netdev);
505 if (sdev) {
506 dev_dbg(&netdev->dev, "siw: new device\n");
507
508 if (netif_running(netdev) && netif_carrier_ok(netdev))
509 sdev->state = IB_PORT_ACTIVE;
510 else
511 sdev->state = IB_PORT_DOWN;
512
513 rv = siw_device_register(sdev, basedev_name);
514 if (rv)
515 ib_dealloc_device(&sdev->base_dev);
516 }
517 return rv;
518 }
519
520 static struct rdma_link_ops siw_link_ops = {
521 .type = "siw",
522 .newlink = siw_newlink,
523 };
524
525 /*
526 * siw_init_module - Initialize Softiwarp module and register with netdev
527 * subsystem.
528 */
siw_init_module(void)529 static __init int siw_init_module(void)
530 {
531 int rv;
532 int nr_cpu;
533
534 if (SENDPAGE_THRESH < SIW_MAX_INLINE) {
535 pr_info("siw: sendpage threshold too small: %u\n",
536 (int)SENDPAGE_THRESH);
537 rv = -EINVAL;
538 goto out_error;
539 }
540 rv = siw_init_cpulist();
541 if (rv)
542 goto out_error;
543
544 rv = siw_cm_init();
545 if (rv)
546 goto out_error;
547
548 if (!siw_create_tx_threads()) {
549 pr_info("siw: Could not start any TX thread\n");
550 rv = -ENOMEM;
551 goto out_error;
552 }
553 /*
554 * Locate CRC32 algorithm. If unsuccessful, fail
555 * loading siw only, if CRC is required.
556 */
557 siw_crypto_shash = crypto_alloc_shash("crc32c", 0, 0);
558 if (IS_ERR(siw_crypto_shash)) {
559 pr_info("siw: Loading CRC32c failed: %ld\n",
560 PTR_ERR(siw_crypto_shash));
561 siw_crypto_shash = NULL;
562 if (mpa_crc_required) {
563 rv = -EOPNOTSUPP;
564 goto out_error;
565 }
566 }
567 rv = register_netdevice_notifier(&siw_netdev_nb);
568 if (rv)
569 goto out_error;
570
571 rdma_link_register(&siw_link_ops);
572
573 pr_info("SoftiWARP attached\n");
574 return 0;
575
576 out_error:
577 for (nr_cpu = 0; nr_cpu < nr_cpu_ids; nr_cpu++) {
578 if (siw_tx_thread[nr_cpu]) {
579 siw_stop_tx_thread(nr_cpu);
580 siw_tx_thread[nr_cpu] = NULL;
581 }
582 }
583 if (siw_crypto_shash)
584 crypto_free_shash(siw_crypto_shash);
585
586 pr_info("SoftIWARP attach failed. Error: %d\n", rv);
587
588 siw_cm_exit();
589 siw_destroy_cpulist();
590
591 return rv;
592 }
593
siw_exit_module(void)594 static void __exit siw_exit_module(void)
595 {
596 int cpu;
597
598 for_each_possible_cpu(cpu) {
599 if (siw_tx_thread[cpu]) {
600 siw_stop_tx_thread(cpu);
601 siw_tx_thread[cpu] = NULL;
602 }
603 }
604 unregister_netdevice_notifier(&siw_netdev_nb);
605 rdma_link_unregister(&siw_link_ops);
606 ib_unregister_driver(RDMA_DRIVER_SIW);
607
608 siw_cm_exit();
609
610 siw_destroy_cpulist();
611
612 if (siw_crypto_shash)
613 crypto_free_shash(siw_crypto_shash);
614
615 pr_info("SoftiWARP detached\n");
616 }
617
618 module_init(siw_init_module);
619 module_exit(siw_exit_module);
620
621 MODULE_ALIAS_RDMA_LINK("siw");
622