siw_main.c

  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
 28MODULE_AUTHOR("Bernard Metzler");
 29MODULE_DESCRIPTION("Software iWARP Driver");
 30MODULE_LICENSE("Dual BSD/GPL");
 31
 32/* transmit from user buffer, if possible */
 33const 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 */
 39const bool try_gso;
 40
 41/* Attach siw also with loopback devices */
 42const bool loopback_enabled = true;
 43
 44/* We try to negotiate CRC on, if true */
 45const bool mpa_crc_required;
 46
 47/* MPA CRC on/off enforced */
 48const bool mpa_crc_strict;
 49
 50/* Control TCP_NODELAY socket option */
 51const bool siw_tcp_nagle;
 52
 53/* Select MPA version to be used during connection setup */
 54u_char mpa_version = MPA_REVISION_2;
 55
 56/* Selects MPA P2P mode (additional handshake during connection
 57 * setup, if true.
 58 */
 59const bool peer_to_peer;
 60
 61struct task_struct *siw_tx_thread[NR_CPUS];
 62
 63static int siw_device_register(struct siw_device *sdev, const char *name)
 64{
 65	struct ib_device *base_dev = &sdev->base_dev;
 66	static int dev_id = 1;
 67	int rv;
 68
 69	sdev->vendor_part_id = dev_id++;
 70
 71	rv = ib_register_device(base_dev, name, NULL);
 72	if (rv) {
 73		pr_warn("siw: device registration error %d\n", rv);
 74		return rv;
 75	}
 76
 77	siw_dbg(base_dev, "HWaddr=%pM\n", sdev->raw_gid);
 78	return 0;
 79}
 80
 81static void siw_device_cleanup(struct ib_device *base_dev)
 82{
 83	struct siw_device *sdev = to_siw_dev(base_dev);
 84
 85	xa_destroy(&sdev->qp_xa);
 86	xa_destroy(&sdev->mem_xa);
 87}
 88
 89static int siw_dev_qualified(struct net_device *netdev)
 90{
 91	/*
 92	 * Additional hardware support can be added here
 93	 * (e.g. ARPHRD_FDDI, ARPHRD_ATM, ...) - see
 94	 * <linux/if_arp.h> for type identifiers.
 95	 */
 96	if (netdev->type == ARPHRD_ETHER || netdev->type == ARPHRD_IEEE802 ||
 97	    netdev->type == ARPHRD_NONE ||
 98	    (netdev->type == ARPHRD_LOOPBACK && loopback_enabled))
 99		return 1;
100
101	return 0;
102}
103
104static DEFINE_PER_CPU(atomic_t, siw_use_cnt);
105
106static struct {
107	struct cpumask **tx_valid_cpus;
108	int num_nodes;
109} siw_cpu_info;
110
111static void siw_destroy_cpulist(int number)
112{
113	int i = 0;
114
115	while (i < number)
116		kfree(siw_cpu_info.tx_valid_cpus[i++]);
117
118	kfree(siw_cpu_info.tx_valid_cpus);
119	siw_cpu_info.tx_valid_cpus = NULL;
120}
121
122static int siw_init_cpulist(void)
123{
124	int i, num_nodes = nr_node_ids;
125
126	memset(siw_tx_thread, 0, sizeof(siw_tx_thread));
127
128	siw_cpu_info.num_nodes = num_nodes;
129
130	siw_cpu_info.tx_valid_cpus =
131		kcalloc(num_nodes, sizeof(struct cpumask *), GFP_KERNEL);
132	if (!siw_cpu_info.tx_valid_cpus) {
133		siw_cpu_info.num_nodes = 0;
134		return -ENOMEM;
135	}
136	for (i = 0; i < siw_cpu_info.num_nodes; i++) {
137		siw_cpu_info.tx_valid_cpus[i] =
138			kzalloc(sizeof(struct cpumask), GFP_KERNEL);
139		if (!siw_cpu_info.tx_valid_cpus[i])
140			goto out_err;
141
142		cpumask_clear(siw_cpu_info.tx_valid_cpus[i]);
143	}
144	for_each_possible_cpu(i)
145		cpumask_set_cpu(i, siw_cpu_info.tx_valid_cpus[cpu_to_node(i)]);
146
147	return 0;
148
149out_err:
150	siw_cpu_info.num_nodes = 0;
151	siw_destroy_cpulist(i);
152
153	return -ENOMEM;
154}
155
156/*
157 * Choose CPU with least number of active QP's from NUMA node of
158 * TX interface.
159 */
160int siw_get_tx_cpu(struct siw_device *sdev)
161{
162	const struct cpumask *tx_cpumask;
163	int i, num_cpus, cpu, min_use, node = sdev->numa_node, tx_cpu = -1;
164
165	if (node < 0)
166		tx_cpumask = cpu_online_mask;
167	else
168		tx_cpumask = siw_cpu_info.tx_valid_cpus[node];
169
170	num_cpus = cpumask_weight(tx_cpumask);
171	if (!num_cpus) {
172		/* no CPU on this NUMA node */
173		tx_cpumask = cpu_online_mask;
174		num_cpus = cpumask_weight(tx_cpumask);
175	}
176	if (!num_cpus)
177		goto out;
178
179	cpu = cpumask_first(tx_cpumask);
180
181	for (i = 0, min_use = SIW_MAX_QP; i < num_cpus;
182	     i++, cpu = cpumask_next(cpu, tx_cpumask)) {
183		int usage;
184
185		/* Skip any cores which have no TX thread */
186		if (!siw_tx_thread[cpu])
187			continue;
188
189		usage = atomic_read(&per_cpu(siw_use_cnt, cpu));
190		if (usage <= min_use) {
191			tx_cpu = cpu;
192			min_use = usage;
193		}
194	}
195	siw_dbg(&sdev->base_dev,
196		"tx cpu %d, node %d, %d qp's\n", tx_cpu, node, min_use);
197
198out:
199	if (tx_cpu >= 0)
200		atomic_inc(&per_cpu(siw_use_cnt, tx_cpu));
201	else
202		pr_warn("siw: no tx cpu found\n");
203
204	return tx_cpu;
205}
206
207void siw_put_tx_cpu(int cpu)
208{
209	atomic_dec(&per_cpu(siw_use_cnt, cpu));
210}
211
212static struct ib_qp *siw_get_base_qp(struct ib_device *base_dev, int id)
213{
214	struct siw_qp *qp = siw_qp_id2obj(to_siw_dev(base_dev), id);
215
216	if (qp) {
217		/*
218		 * siw_qp_id2obj() increments object reference count
219		 */
220		siw_qp_put(qp);
221		return &qp->base_qp;
222	}
223	return NULL;
224}
225
226static const struct ib_device_ops siw_device_ops = {
227	.owner = THIS_MODULE,
228	.uverbs_abi_ver = SIW_ABI_VERSION,
229	.driver_id = RDMA_DRIVER_SIW,
230
231	.alloc_mr = siw_alloc_mr,
232	.alloc_pd = siw_alloc_pd,
233	.alloc_ucontext = siw_alloc_ucontext,
234	.create_cq = siw_create_cq,
235	.create_qp = siw_create_qp,
236	.create_srq = siw_create_srq,
237	.dealloc_driver = siw_device_cleanup,
238	.dealloc_pd = siw_dealloc_pd,
239	.dealloc_ucontext = siw_dealloc_ucontext,
240	.dereg_mr = siw_dereg_mr,
241	.destroy_cq = siw_destroy_cq,
242	.destroy_qp = siw_destroy_qp,
243	.destroy_srq = siw_destroy_srq,
244	.get_dma_mr = siw_get_dma_mr,
245	.get_port_immutable = siw_get_port_immutable,
246	.iw_accept = siw_accept,
247	.iw_add_ref = siw_qp_get_ref,
248	.iw_connect = siw_connect,
249	.iw_create_listen = siw_create_listen,
250	.iw_destroy_listen = siw_destroy_listen,
251	.iw_get_qp = siw_get_base_qp,
252	.iw_reject = siw_reject,
253	.iw_rem_ref = siw_qp_put_ref,
254	.map_mr_sg = siw_map_mr_sg,
255	.mmap = siw_mmap,
256	.mmap_free = siw_mmap_free,
257	.modify_qp = siw_verbs_modify_qp,
258	.modify_srq = siw_modify_srq,
259	.poll_cq = siw_poll_cq,
260	.post_recv = siw_post_receive,
261	.post_send = siw_post_send,
262	.post_srq_recv = siw_post_srq_recv,
263	.query_device = siw_query_device,
264	.query_gid = siw_query_gid,
265	.query_port = siw_query_port,
266	.query_qp = siw_query_qp,
267	.query_srq = siw_query_srq,
268	.req_notify_cq = siw_req_notify_cq,
269	.reg_user_mr = siw_reg_user_mr,
270
271	INIT_RDMA_OBJ_SIZE(ib_cq, siw_cq, base_cq),
272	INIT_RDMA_OBJ_SIZE(ib_pd, siw_pd, base_pd),
273	INIT_RDMA_OBJ_SIZE(ib_qp, siw_qp, base_qp),
274	INIT_RDMA_OBJ_SIZE(ib_srq, siw_srq, base_srq),
275	INIT_RDMA_OBJ_SIZE(ib_ucontext, siw_ucontext, base_ucontext),
276};
277
278static struct siw_device *siw_device_create(struct net_device *netdev)
279{
280	struct siw_device *sdev = NULL;
281	struct ib_device *base_dev;
282	int rv;
283
284	sdev = ib_alloc_device(siw_device, base_dev);
285	if (!sdev)
286		return NULL;
287
288	base_dev = &sdev->base_dev;
289
290	if (netdev->addr_len) {
291		memcpy(sdev->raw_gid, netdev->dev_addr,
292		       min_t(unsigned int, netdev->addr_len, ETH_ALEN));
293	} else {
294		/*
295		 * This device does not have a HW address, but
296		 * connection mangagement requires a unique gid.
297		 */
298		eth_random_addr(sdev->raw_gid);
299	}
300	addrconf_addr_eui48((u8 *)&base_dev->node_guid, sdev->raw_gid);
301
302	base_dev->uverbs_cmd_mask |= BIT_ULL(IB_USER_VERBS_CMD_POST_SEND);
303
304	base_dev->node_type = RDMA_NODE_RNIC;
305	memcpy(base_dev->node_desc, SIW_NODE_DESC_COMMON,
306	       sizeof(SIW_NODE_DESC_COMMON));
307
308	/*
309	 * Current model (one-to-one device association):
310	 * One Softiwarp device per net_device or, equivalently,
311	 * per physical port.
312	 */
313	base_dev->phys_port_cnt = 1;
314	base_dev->num_comp_vectors = num_possible_cpus();
315
316	xa_init_flags(&sdev->qp_xa, XA_FLAGS_ALLOC1);
317	xa_init_flags(&sdev->mem_xa, XA_FLAGS_ALLOC1);
318
319	ib_set_device_ops(base_dev, &siw_device_ops);
320	rv = ib_device_set_netdev(base_dev, netdev, 1);
321	if (rv)
322		goto error;
323
324	memcpy(base_dev->iw_ifname, netdev->name,
325	       sizeof(base_dev->iw_ifname));
326
327	/* Disable TCP port mapping */
328	base_dev->iw_driver_flags = IW_F_NO_PORT_MAP;
329
330	sdev->attrs.max_qp = SIW_MAX_QP;
331	sdev->attrs.max_qp_wr = SIW_MAX_QP_WR;
332	sdev->attrs.max_ord = SIW_MAX_ORD_QP;
333	sdev->attrs.max_ird = SIW_MAX_IRD_QP;
334	sdev->attrs.max_sge = SIW_MAX_SGE;
335	sdev->attrs.max_sge_rd = SIW_MAX_SGE_RD;
336	sdev->attrs.max_cq = SIW_MAX_CQ;
337	sdev->attrs.max_cqe = SIW_MAX_CQE;
338	sdev->attrs.max_mr = SIW_MAX_MR;
339	sdev->attrs.max_pd = SIW_MAX_PD;
340	sdev->attrs.max_mw = SIW_MAX_MW;
341	sdev->attrs.max_srq = SIW_MAX_SRQ;
342	sdev->attrs.max_srq_wr = SIW_MAX_SRQ_WR;
343	sdev->attrs.max_srq_sge = SIW_MAX_SGE;
344
345	INIT_LIST_HEAD(&sdev->cep_list);
346	INIT_LIST_HEAD(&sdev->qp_list);
347
348	atomic_set(&sdev->num_ctx, 0);
349	atomic_set(&sdev->num_srq, 0);
350	atomic_set(&sdev->num_qp, 0);
351	atomic_set(&sdev->num_cq, 0);
352	atomic_set(&sdev->num_mr, 0);
353	atomic_set(&sdev->num_pd, 0);
354
355	sdev->numa_node = dev_to_node(&netdev->dev);
356	spin_lock_init(&sdev->lock);
357
358	return sdev;
359error:
360	ib_dealloc_device(base_dev);
361
362	return NULL;
363}
364
365static int siw_netdev_event(struct notifier_block *nb, unsigned long event,
366			    void *arg)
367{
368	struct net_device *netdev = netdev_notifier_info_to_dev(arg);
369	struct ib_device *base_dev;
370	struct siw_device *sdev;
371
372	dev_dbg(&netdev->dev, "siw: event %lu\n", event);
373
374	base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW);
375	if (!base_dev)
376		return NOTIFY_OK;
377
378	sdev = to_siw_dev(base_dev);
379
380	switch (event) {
381	case NETDEV_REGISTER:
382		/*
383		 * Device registration now handled only by
384		 * rdma netlink commands. So it shall be impossible
385		 * to end up here with a valid siw device.
386		 */
387		siw_dbg(base_dev, "unexpected NETDEV_REGISTER event\n");
388		break;
389
390	case NETDEV_UNREGISTER:
391		ib_unregister_device_queued(&sdev->base_dev);
392		break;
393
394	case NETDEV_CHANGEADDR:
395		siw_port_event(sdev, 1, IB_EVENT_LID_CHANGE);
396		break;
397	/*
398	 * All other events are not handled
399	 */
400	default:
401		break;
402	}
403	ib_device_put(&sdev->base_dev);
404
405	return NOTIFY_OK;
406}
407
408static struct notifier_block siw_netdev_nb = {
409	.notifier_call = siw_netdev_event,
410};
411
412static int siw_newlink(const char *basedev_name, struct net_device *netdev)
413{
414	struct ib_device *base_dev;
415	struct siw_device *sdev = NULL;
416	int rv = -ENOMEM;
417
418	if (!siw_dev_qualified(netdev))
419		return -EINVAL;
420
421	base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW);
422	if (base_dev) {
423		ib_device_put(base_dev);
424		return -EEXIST;
425	}
426	sdev = siw_device_create(netdev);
427	if (sdev) {
428		dev_dbg(&netdev->dev, "siw: new device\n");
429		ib_mark_name_assigned_by_user(&sdev->base_dev);
430		rv = siw_device_register(sdev, basedev_name);
431		if (rv)
432			ib_dealloc_device(&sdev->base_dev);
433	}
434	return rv;
435}
436
437static struct rdma_link_ops siw_link_ops = {
438	.type = "siw",
439	.newlink = siw_newlink,
440};
441
442/*
443 * siw_init_module - Initialize Softiwarp module and register with netdev
444 *                   subsystem.
445 */
446static __init int siw_init_module(void)
447{
448	int rv;
449
450	if (SENDPAGE_THRESH < SIW_MAX_INLINE) {
451		pr_info("siw: sendpage threshold too small: %u\n",
452			(int)SENDPAGE_THRESH);
453		rv = -EINVAL;
454		goto out_error;
455	}
456	rv = siw_init_cpulist();
457	if (rv)
458		goto out_error;
459
460	rv = siw_cm_init();
461	if (rv)
462		goto out_error;
463
464	if (!siw_create_tx_threads()) {
465		pr_info("siw: Could not start any TX thread\n");
466		rv = -ENOMEM;
467		goto out_error;
468	}
469
470	rv = register_netdevice_notifier(&siw_netdev_nb);
471	if (rv)
472		goto out_error;
473
474	rdma_link_register(&siw_link_ops);
475
476	pr_info("SoftiWARP attached\n");
477	return 0;
478
479out_error:
480	siw_stop_tx_threads();
481
482	pr_info("SoftIWARP attach failed. Error: %d\n", rv);
483
484	siw_cm_exit();
485	siw_destroy_cpulist(siw_cpu_info.num_nodes);
486
487	return rv;
488}
489
490static void __exit siw_exit_module(void)
491{
492	siw_stop_tx_threads();
493
494	unregister_netdevice_notifier(&siw_netdev_nb);
495	rdma_link_unregister(&siw_link_ops);
496	ib_unregister_driver(RDMA_DRIVER_SIW);
497
498	siw_cm_exit();
499
500	siw_destroy_cpulist(siw_cpu_info.num_nodes);
501
502	pr_info("SoftiWARP detached\n");
503}
504
505module_init(siw_init_module);
506module_exit(siw_exit_module);
507
508MODULE_ALIAS_RDMA_LINK("siw");