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 
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->netdev->dev_addr);
79 
80 	return 0;
81 }
82 
83 static 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 
91 static int siw_create_tx_threads(void)
92 {
93 	int cpu, assigned = 0;
94 
95 	for_each_online_cpu(cpu) {
96 		/* Skip HT cores */
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 
115 static int siw_dev_qualified(struct net_device *netdev)
116 {
117 	/*
118 	 * Additional hardware support can be added here
119 	 * (e.g. ARPHRD_FDDI, ARPHRD_ATM, ...) - see
120 	 * <linux/if_arp.h> for type identifiers.
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 
129 static DEFINE_PER_CPU(atomic_t, siw_use_cnt);
130 
131 static struct {
132 	struct cpumask **tx_valid_cpus;
133 	int num_nodes;
134 } siw_cpu_info;
135 
136 static int siw_init_cpulist(void)
137 {
138 	int i, num_nodes = nr_node_ids;
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 
163 out_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 
173 static 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  * Choose CPU with least number of active QP's from NUMA node of
185  * TX interface.
186  */
187 int 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 		/* no CPU on this NUMA node */
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 		/* Skip any cores which have no TX thread */
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 
225 out:
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 
234 void siw_put_tx_cpu(int cpu)
235 {
236 	atomic_dec(&per_cpu(siw_use_cnt, cpu));
237 }
238 
239 static 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 		 * siw_qp_id2obj() increments object reference count
246 		 */
247 		siw_qp_put(qp);
248 		return &qp->base_qp;
249 	}
250 	return NULL;
251 }
252 
253 static 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 
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 
316 	sdev->netdev = netdev;
317 
318 	if (netdev->type != ARPHRD_LOOPBACK) {
319 		addrconf_addr_eui48((unsigned char *)&base_dev->node_guid,
320 				    netdev->dev_addr);
321 	} else {
322 		/*
323 		 * The loopback device does not have a HW address,
324 		 * but connection mangagement lib expects gid != 0
325 		 */
326 		size_t len = min_t(size_t, strlen(base_dev->name), 6);
327 		char addr[6] = { };
328 
329 		memcpy(addr, base_dev->name, len);
330 		addrconf_addr_eui48((unsigned char *)&base_dev->node_guid,
331 				    addr);
332 	}
333 
334 	base_dev->uverbs_cmd_mask |= BIT_ULL(IB_USER_VERBS_CMD_POST_SEND);
335 
336 	base_dev->node_type = RDMA_NODE_RNIC;
337 	memcpy(base_dev->node_desc, SIW_NODE_DESC_COMMON,
338 	       sizeof(SIW_NODE_DESC_COMMON));
339 
340 	/*
341 	 * Current model (one-to-one device association):
342 	 * One Softiwarp device per net_device or, equivalently,
343 	 * per physical port.
344 	 */
345 	base_dev->phys_port_cnt = 1;
346 	base_dev->num_comp_vectors = num_possible_cpus();
347 
348 	xa_init_flags(&sdev->qp_xa, XA_FLAGS_ALLOC1);
349 	xa_init_flags(&sdev->mem_xa, XA_FLAGS_ALLOC1);
350 
351 	ib_set_device_ops(base_dev, &siw_device_ops);
352 	rv = ib_device_set_netdev(base_dev, netdev, 1);
353 	if (rv)
354 		goto error;
355 
356 	memcpy(base_dev->iw_ifname, netdev->name,
357 	       sizeof(base_dev->iw_ifname));
358 
359 	/* Disable TCP port mapping */
360 	base_dev->iw_driver_flags = IW_F_NO_PORT_MAP;
361 
362 	sdev->attrs.max_qp = SIW_MAX_QP;
363 	sdev->attrs.max_qp_wr = SIW_MAX_QP_WR;
364 	sdev->attrs.max_ord = SIW_MAX_ORD_QP;
365 	sdev->attrs.max_ird = SIW_MAX_IRD_QP;
366 	sdev->attrs.max_sge = SIW_MAX_SGE;
367 	sdev->attrs.max_sge_rd = SIW_MAX_SGE_RD;
368 	sdev->attrs.max_cq = SIW_MAX_CQ;
369 	sdev->attrs.max_cqe = SIW_MAX_CQE;
370 	sdev->attrs.max_mr = SIW_MAX_MR;
371 	sdev->attrs.max_pd = SIW_MAX_PD;
372 	sdev->attrs.max_mw = SIW_MAX_MW;
373 	sdev->attrs.max_srq = SIW_MAX_SRQ;
374 	sdev->attrs.max_srq_wr = SIW_MAX_SRQ_WR;
375 	sdev->attrs.max_srq_sge = SIW_MAX_SGE;
376 
377 	INIT_LIST_HEAD(&sdev->cep_list);
378 	INIT_LIST_HEAD(&sdev->qp_list);
379 
380 	atomic_set(&sdev->num_ctx, 0);
381 	atomic_set(&sdev->num_srq, 0);
382 	atomic_set(&sdev->num_qp, 0);
383 	atomic_set(&sdev->num_cq, 0);
384 	atomic_set(&sdev->num_mr, 0);
385 	atomic_set(&sdev->num_pd, 0);
386 
387 	sdev->numa_node = dev_to_node(&netdev->dev);
388 	spin_lock_init(&sdev->lock);
389 
390 	return sdev;
391 error:
392 	ib_dealloc_device(base_dev);
393 
394 	return NULL;
395 }
396 
397 /*
398  * Network link becomes unavailable. Mark all
399  * affected QP's accordingly.
400  */
401 static void siw_netdev_down(struct work_struct *work)
402 {
403 	struct siw_device *sdev =
404 		container_of(work, struct siw_device, netdev_down);
405 
406 	struct siw_qp_attrs qp_attrs;
407 	struct list_head *pos, *tmp;
408 
409 	memset(&qp_attrs, 0, sizeof(qp_attrs));
410 	qp_attrs.state = SIW_QP_STATE_ERROR;
411 
412 	list_for_each_safe(pos, tmp, &sdev->qp_list) {
413 		struct siw_qp *qp = list_entry(pos, struct siw_qp, devq);
414 
415 		down_write(&qp->state_lock);
416 		WARN_ON(siw_qp_modify(qp, &qp_attrs, SIW_QP_ATTR_STATE));
417 		up_write(&qp->state_lock);
418 	}
419 	ib_device_put(&sdev->base_dev);
420 }
421 
422 static void siw_device_goes_down(struct siw_device *sdev)
423 {
424 	if (ib_device_try_get(&sdev->base_dev)) {
425 		INIT_WORK(&sdev->netdev_down, siw_netdev_down);
426 		schedule_work(&sdev->netdev_down);
427 	}
428 }
429 
430 static int siw_netdev_event(struct notifier_block *nb, unsigned long event,
431 			    void *arg)
432 {
433 	struct net_device *netdev = netdev_notifier_info_to_dev(arg);
434 	struct ib_device *base_dev;
435 	struct siw_device *sdev;
436 
437 	dev_dbg(&netdev->dev, "siw: event %lu\n", event);
438 
439 	if (dev_net(netdev) != &init_net)
440 		return NOTIFY_OK;
441 
442 	base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW);
443 	if (!base_dev)
444 		return NOTIFY_OK;
445 
446 	sdev = to_siw_dev(base_dev);
447 
448 	switch (event) {
449 	case NETDEV_UP:
450 		sdev->state = IB_PORT_ACTIVE;
451 		siw_port_event(sdev, 1, IB_EVENT_PORT_ACTIVE);
452 		break;
453 
454 	case NETDEV_GOING_DOWN:
455 		siw_device_goes_down(sdev);
456 		break;
457 
458 	case NETDEV_DOWN:
459 		sdev->state = IB_PORT_DOWN;
460 		siw_port_event(sdev, 1, IB_EVENT_PORT_ERR);
461 		break;
462 
463 	case NETDEV_REGISTER:
464 		/*
465 		 * Device registration now handled only by
466 		 * rdma netlink commands. So it shall be impossible
467 		 * to end up here with a valid siw device.
468 		 */
469 		siw_dbg(base_dev, "unexpected NETDEV_REGISTER event\n");
470 		break;
471 
472 	case NETDEV_UNREGISTER:
473 		ib_unregister_device_queued(&sdev->base_dev);
474 		break;
475 
476 	case NETDEV_CHANGEADDR:
477 		siw_port_event(sdev, 1, IB_EVENT_LID_CHANGE);
478 		break;
479 	/*
480 	 * Todo: Below netdev events are currently not handled.
481 	 */
482 	case NETDEV_CHANGEMTU:
483 	case NETDEV_CHANGE:
484 		break;
485 
486 	default:
487 		break;
488 	}
489 	ib_device_put(&sdev->base_dev);
490 
491 	return NOTIFY_OK;
492 }
493 
494 static struct notifier_block siw_netdev_nb = {
495 	.notifier_call = siw_netdev_event,
496 };
497 
498 static int siw_newlink(const char *basedev_name, struct net_device *netdev)
499 {
500 	struct ib_device *base_dev;
501 	struct siw_device *sdev = NULL;
502 	int rv = -ENOMEM;
503 
504 	if (!siw_dev_qualified(netdev))
505 		return -EINVAL;
506 
507 	base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW);
508 	if (base_dev) {
509 		ib_device_put(base_dev);
510 		return -EEXIST;
511 	}
512 	sdev = siw_device_create(netdev);
513 	if (sdev) {
514 		dev_dbg(&netdev->dev, "siw: new device\n");
515 
516 		if (netif_running(netdev) && netif_carrier_ok(netdev))
517 			sdev->state = IB_PORT_ACTIVE;
518 		else
519 			sdev->state = IB_PORT_DOWN;
520 
521 		rv = siw_device_register(sdev, basedev_name);
522 		if (rv)
523 			ib_dealloc_device(&sdev->base_dev);
524 	}
525 	return rv;
526 }
527 
528 static struct rdma_link_ops siw_link_ops = {
529 	.type = "siw",
530 	.newlink = siw_newlink,
531 };
532 
533 /*
534  * siw_init_module - Initialize Softiwarp module and register with netdev
535  *                   subsystem.
536  */
537 static __init int siw_init_module(void)
538 {
539 	int rv;
540 	int nr_cpu;
541 
542 	if (SENDPAGE_THRESH < SIW_MAX_INLINE) {
543 		pr_info("siw: sendpage threshold too small: %u\n",
544 			(int)SENDPAGE_THRESH);
545 		rv = -EINVAL;
546 		goto out_error;
547 	}
548 	rv = siw_init_cpulist();
549 	if (rv)
550 		goto out_error;
551 
552 	rv = siw_cm_init();
553 	if (rv)
554 		goto out_error;
555 
556 	if (!siw_create_tx_threads()) {
557 		pr_info("siw: Could not start any TX thread\n");
558 		rv = -ENOMEM;
559 		goto out_error;
560 	}
561 	/*
562 	 * Locate CRC32 algorithm. If unsuccessful, fail
563 	 * loading siw only, if CRC is required.
564 	 */
565 	siw_crypto_shash = crypto_alloc_shash("crc32c", 0, 0);
566 	if (IS_ERR(siw_crypto_shash)) {
567 		pr_info("siw: Loading CRC32c failed: %ld\n",
568 			PTR_ERR(siw_crypto_shash));
569 		siw_crypto_shash = NULL;
570 		if (mpa_crc_required) {
571 			rv = -EOPNOTSUPP;
572 			goto out_error;
573 		}
574 	}
575 	rv = register_netdevice_notifier(&siw_netdev_nb);
576 	if (rv)
577 		goto out_error;
578 
579 	rdma_link_register(&siw_link_ops);
580 
581 	pr_info("SoftiWARP attached\n");
582 	return 0;
583 
584 out_error:
585 	for (nr_cpu = 0; nr_cpu < nr_cpu_ids; nr_cpu++) {
586 		if (siw_tx_thread[nr_cpu]) {
587 			siw_stop_tx_thread(nr_cpu);
588 			siw_tx_thread[nr_cpu] = NULL;
589 		}
590 	}
591 	if (siw_crypto_shash)
592 		crypto_free_shash(siw_crypto_shash);
593 
594 	pr_info("SoftIWARP attach failed. Error: %d\n", rv);
595 
596 	siw_cm_exit();
597 	siw_destroy_cpulist();
598 
599 	return rv;
600 }
601 
602 static void __exit siw_exit_module(void)
603 {
604 	int cpu;
605 
606 	for_each_possible_cpu(cpu) {
607 		if (siw_tx_thread[cpu]) {
608 			siw_stop_tx_thread(cpu);
609 			siw_tx_thread[cpu] = NULL;
610 		}
611 	}
612 	unregister_netdevice_notifier(&siw_netdev_nb);
613 	rdma_link_unregister(&siw_link_ops);
614 	ib_unregister_driver(RDMA_DRIVER_SIW);
615 
616 	siw_cm_exit();
617 
618 	siw_destroy_cpulist();
619 
620 	if (siw_crypto_shash)
621 		crypto_free_shash(siw_crypto_shash);
622 
623 	pr_info("SoftiWARP detached\n");
624 }
625 
626 module_init(siw_init_module);
627 module_exit(siw_exit_module);
628 
629 MODULE_ALIAS_RDMA_LINK("siw");
630