1 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
3 /* Authors: Bernard Metzler <bmt@zurich.ibm.com> */
4 /* Copyright (c) 2008-2019, IBM Corporation */
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>
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>
26 #include "siw_verbs.h"
28 MODULE_AUTHOR("Bernard Metzler");
29 MODULE_DESCRIPTION("Software iWARP Driver");
30 MODULE_LICENSE("Dual BSD/GPL");
32 /* transmit from user buffer, if possible */
33 const bool zcopy_tx = true;
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.
41 /* Attach siw also with loopback devices */
42 const bool loopback_enabled = true;
44 /* We try to negotiate CRC on, if true */
45 const bool mpa_crc_required;
47 /* MPA CRC on/off enforced */
48 const bool mpa_crc_strict;
50 /* Control TCP_NODELAY socket option */
51 const bool siw_tcp_nagle;
53 /* Select MPA version to be used during connection setup */
54 u_char mpa_version = MPA_REVISION_2;
56 /* Selects MPA P2P mode (additional handshake during connection
59 const bool peer_to_peer;
61 struct task_struct *siw_tx_thread[NR_CPUS];
62 struct crypto_shash *siw_crypto_shash;
64 static int siw_device_register(struct siw_device *sdev, const char *name)
66 struct ib_device *base_dev = &sdev->base_dev;
67 static int dev_id = 1;
70 sdev->vendor_part_id = dev_id++;
72 rv = ib_register_device(base_dev, name, NULL);
74 pr_warn("siw: device registration error %d\n", rv);
78 siw_dbg(base_dev, "HWaddr=%pM\n", sdev->netdev->dev_addr);
83 static void siw_device_cleanup(struct ib_device *base_dev)
85 struct siw_device *sdev = to_siw_dev(base_dev);
87 xa_destroy(&sdev->qp_xa);
88 xa_destroy(&sdev->mem_xa);
91 static int siw_create_tx_threads(void)
93 int cpu, assigned = 0;
95 for_each_online_cpu(cpu) {
97 if (cpu % cpumask_weight(topology_sibling_cpumask(cpu)))
101 kthread_run_on_cpu(siw_run_sq,
102 (unsigned long *)(long)cpu,
104 if (IS_ERR(siw_tx_thread[cpu])) {
105 siw_tx_thread[cpu] = NULL;
114 static int siw_dev_qualified(struct net_device *netdev)
117 * Additional hardware support can be added here
118 * (e.g. ARPHRD_FDDI, ARPHRD_ATM, ...) - see
119 * <linux/if_arp.h> for type identifiers.
121 if (netdev->type == ARPHRD_ETHER || netdev->type == ARPHRD_IEEE802 ||
122 netdev->type == ARPHRD_NONE ||
123 (netdev->type == ARPHRD_LOOPBACK && loopback_enabled))
129 static DEFINE_PER_CPU(atomic_t, siw_use_cnt);
132 struct cpumask **tx_valid_cpus;
136 static int siw_init_cpulist(void)
138 int i, num_nodes = nr_node_ids;
140 memset(siw_tx_thread, 0, sizeof(siw_tx_thread));
142 siw_cpu_info.num_nodes = num_nodes;
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;
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])
156 cpumask_clear(siw_cpu_info.tx_valid_cpus[i]);
158 for_each_possible_cpu(i)
159 cpumask_set_cpu(i, siw_cpu_info.tx_valid_cpus[cpu_to_node(i)]);
164 siw_cpu_info.num_nodes = 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;
173 static void siw_destroy_cpulist(void)
177 while (i < siw_cpu_info.num_nodes)
178 kfree(siw_cpu_info.tx_valid_cpus[i++]);
180 kfree(siw_cpu_info.tx_valid_cpus);
184 * Choose CPU with least number of active QP's from NUMA node of
187 int siw_get_tx_cpu(struct siw_device *sdev)
189 const struct cpumask *tx_cpumask;
190 int i, num_cpus, cpu, min_use, node = sdev->numa_node, tx_cpu = -1;
193 tx_cpumask = cpu_online_mask;
195 tx_cpumask = siw_cpu_info.tx_valid_cpus[node];
197 num_cpus = cpumask_weight(tx_cpumask);
199 /* no CPU on this NUMA node */
200 tx_cpumask = cpu_online_mask;
201 num_cpus = cpumask_weight(tx_cpumask);
206 cpu = cpumask_first(tx_cpumask);
208 for (i = 0, min_use = SIW_MAX_QP; i < num_cpus;
209 i++, cpu = cpumask_next(cpu, tx_cpumask)) {
212 /* Skip any cores which have no TX thread */
213 if (!siw_tx_thread[cpu])
216 usage = atomic_read(&per_cpu(siw_use_cnt, cpu));
217 if (usage <= min_use) {
222 siw_dbg(&sdev->base_dev,
223 "tx cpu %d, node %d, %d qp's\n", tx_cpu, node, min_use);
227 atomic_inc(&per_cpu(siw_use_cnt, tx_cpu));
229 pr_warn("siw: no tx cpu found\n");
234 void siw_put_tx_cpu(int cpu)
236 atomic_dec(&per_cpu(siw_use_cnt, cpu));
239 static struct ib_qp *siw_get_base_qp(struct ib_device *base_dev, int id)
241 struct siw_qp *qp = siw_qp_id2obj(to_siw_dev(base_dev), id);
245 * siw_qp_id2obj() increments object reference count
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,
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,
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,
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_qp, siw_qp, base_qp),
301 INIT_RDMA_OBJ_SIZE(ib_srq, siw_srq, base_srq),
302 INIT_RDMA_OBJ_SIZE(ib_ucontext, siw_ucontext, base_ucontext),
305 static struct siw_device *siw_device_create(struct net_device *netdev)
307 struct siw_device *sdev = NULL;
308 struct ib_device *base_dev;
311 sdev = ib_alloc_device(siw_device, base_dev);
315 base_dev = &sdev->base_dev;
317 sdev->netdev = netdev;
319 if (netdev->type != ARPHRD_LOOPBACK && netdev->type != ARPHRD_NONE) {
320 addrconf_addr_eui48((unsigned char *)&base_dev->node_guid,
324 * This device does not have a HW address,
325 * but connection mangagement lib expects gid != 0
327 size_t len = min_t(size_t, strlen(base_dev->name), 6);
330 memcpy(addr, base_dev->name, len);
331 addrconf_addr_eui48((unsigned char *)&base_dev->node_guid,
335 base_dev->uverbs_cmd_mask |= BIT_ULL(IB_USER_VERBS_CMD_POST_SEND);
337 base_dev->node_type = RDMA_NODE_RNIC;
338 memcpy(base_dev->node_desc, SIW_NODE_DESC_COMMON,
339 sizeof(SIW_NODE_DESC_COMMON));
342 * Current model (one-to-one device association):
343 * One Softiwarp device per net_device or, equivalently,
346 base_dev->phys_port_cnt = 1;
347 base_dev->num_comp_vectors = num_possible_cpus();
349 xa_init_flags(&sdev->qp_xa, XA_FLAGS_ALLOC1);
350 xa_init_flags(&sdev->mem_xa, XA_FLAGS_ALLOC1);
352 ib_set_device_ops(base_dev, &siw_device_ops);
353 rv = ib_device_set_netdev(base_dev, netdev, 1);
357 memcpy(base_dev->iw_ifname, netdev->name,
358 sizeof(base_dev->iw_ifname));
360 /* Disable TCP port mapping */
361 base_dev->iw_driver_flags = IW_F_NO_PORT_MAP;
363 sdev->attrs.max_qp = SIW_MAX_QP;
364 sdev->attrs.max_qp_wr = SIW_MAX_QP_WR;
365 sdev->attrs.max_ord = SIW_MAX_ORD_QP;
366 sdev->attrs.max_ird = SIW_MAX_IRD_QP;
367 sdev->attrs.max_sge = SIW_MAX_SGE;
368 sdev->attrs.max_sge_rd = SIW_MAX_SGE_RD;
369 sdev->attrs.max_cq = SIW_MAX_CQ;
370 sdev->attrs.max_cqe = SIW_MAX_CQE;
371 sdev->attrs.max_mr = SIW_MAX_MR;
372 sdev->attrs.max_pd = SIW_MAX_PD;
373 sdev->attrs.max_mw = SIW_MAX_MW;
374 sdev->attrs.max_srq = SIW_MAX_SRQ;
375 sdev->attrs.max_srq_wr = SIW_MAX_SRQ_WR;
376 sdev->attrs.max_srq_sge = SIW_MAX_SGE;
378 INIT_LIST_HEAD(&sdev->cep_list);
379 INIT_LIST_HEAD(&sdev->qp_list);
381 atomic_set(&sdev->num_ctx, 0);
382 atomic_set(&sdev->num_srq, 0);
383 atomic_set(&sdev->num_qp, 0);
384 atomic_set(&sdev->num_cq, 0);
385 atomic_set(&sdev->num_mr, 0);
386 atomic_set(&sdev->num_pd, 0);
388 sdev->numa_node = dev_to_node(&netdev->dev);
389 spin_lock_init(&sdev->lock);
393 ib_dealloc_device(base_dev);
399 * Network link becomes unavailable. Mark all
400 * affected QP's accordingly.
402 static void siw_netdev_down(struct work_struct *work)
404 struct siw_device *sdev =
405 container_of(work, struct siw_device, netdev_down);
407 struct siw_qp_attrs qp_attrs;
408 struct list_head *pos, *tmp;
410 memset(&qp_attrs, 0, sizeof(qp_attrs));
411 qp_attrs.state = SIW_QP_STATE_ERROR;
413 list_for_each_safe(pos, tmp, &sdev->qp_list) {
414 struct siw_qp *qp = list_entry(pos, struct siw_qp, devq);
416 down_write(&qp->state_lock);
417 WARN_ON(siw_qp_modify(qp, &qp_attrs, SIW_QP_ATTR_STATE));
418 up_write(&qp->state_lock);
420 ib_device_put(&sdev->base_dev);
423 static void siw_device_goes_down(struct siw_device *sdev)
425 if (ib_device_try_get(&sdev->base_dev)) {
426 INIT_WORK(&sdev->netdev_down, siw_netdev_down);
427 schedule_work(&sdev->netdev_down);
431 static int siw_netdev_event(struct notifier_block *nb, unsigned long event,
434 struct net_device *netdev = netdev_notifier_info_to_dev(arg);
435 struct ib_device *base_dev;
436 struct siw_device *sdev;
438 dev_dbg(&netdev->dev, "siw: event %lu\n", event);
440 if (dev_net(netdev) != &init_net)
443 base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW);
447 sdev = to_siw_dev(base_dev);
451 sdev->state = IB_PORT_ACTIVE;
452 siw_port_event(sdev, 1, IB_EVENT_PORT_ACTIVE);
455 case NETDEV_GOING_DOWN:
456 siw_device_goes_down(sdev);
460 sdev->state = IB_PORT_DOWN;
461 siw_port_event(sdev, 1, IB_EVENT_PORT_ERR);
464 case NETDEV_REGISTER:
466 * Device registration now handled only by
467 * rdma netlink commands. So it shall be impossible
468 * to end up here with a valid siw device.
470 siw_dbg(base_dev, "unexpected NETDEV_REGISTER event\n");
473 case NETDEV_UNREGISTER:
474 ib_unregister_device_queued(&sdev->base_dev);
477 case NETDEV_CHANGEADDR:
478 siw_port_event(sdev, 1, IB_EVENT_LID_CHANGE);
481 * Todo: Below netdev events are currently not handled.
483 case NETDEV_CHANGEMTU:
490 ib_device_put(&sdev->base_dev);
495 static struct notifier_block siw_netdev_nb = {
496 .notifier_call = siw_netdev_event,
499 static int siw_newlink(const char *basedev_name, struct net_device *netdev)
501 struct ib_device *base_dev;
502 struct siw_device *sdev = NULL;
505 if (!siw_dev_qualified(netdev))
508 base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW);
510 ib_device_put(base_dev);
513 sdev = siw_device_create(netdev);
515 dev_dbg(&netdev->dev, "siw: new device\n");
517 if (netif_running(netdev) && netif_carrier_ok(netdev))
518 sdev->state = IB_PORT_ACTIVE;
520 sdev->state = IB_PORT_DOWN;
522 rv = siw_device_register(sdev, basedev_name);
524 ib_dealloc_device(&sdev->base_dev);
529 static struct rdma_link_ops siw_link_ops = {
531 .newlink = siw_newlink,
535 * siw_init_module - Initialize Softiwarp module and register with netdev
538 static __init int siw_init_module(void)
543 if (SENDPAGE_THRESH < SIW_MAX_INLINE) {
544 pr_info("siw: sendpage threshold too small: %u\n",
545 (int)SENDPAGE_THRESH);
549 rv = siw_init_cpulist();
557 if (!siw_create_tx_threads()) {
558 pr_info("siw: Could not start any TX thread\n");
563 * Locate CRC32 algorithm. If unsuccessful, fail
564 * loading siw only, if CRC is required.
566 siw_crypto_shash = crypto_alloc_shash("crc32c", 0, 0);
567 if (IS_ERR(siw_crypto_shash)) {
568 pr_info("siw: Loading CRC32c failed: %ld\n",
569 PTR_ERR(siw_crypto_shash));
570 siw_crypto_shash = NULL;
571 if (mpa_crc_required) {
576 rv = register_netdevice_notifier(&siw_netdev_nb);
580 rdma_link_register(&siw_link_ops);
582 pr_info("SoftiWARP attached\n");
586 for (nr_cpu = 0; nr_cpu < nr_cpu_ids; nr_cpu++) {
587 if (siw_tx_thread[nr_cpu]) {
588 siw_stop_tx_thread(nr_cpu);
589 siw_tx_thread[nr_cpu] = NULL;
592 if (siw_crypto_shash)
593 crypto_free_shash(siw_crypto_shash);
595 pr_info("SoftIWARP attach failed. Error: %d\n", rv);
598 siw_destroy_cpulist();
603 static void __exit siw_exit_module(void)
607 for_each_possible_cpu(cpu) {
608 if (siw_tx_thread[cpu]) {
609 siw_stop_tx_thread(cpu);
610 siw_tx_thread[cpu] = NULL;
613 unregister_netdevice_notifier(&siw_netdev_nb);
614 rdma_link_unregister(&siw_link_ops);
615 ib_unregister_driver(RDMA_DRIVER_SIW);
619 siw_destroy_cpulist();
621 if (siw_crypto_shash)
622 crypto_free_shash(siw_crypto_shash);
624 pr_info("SoftiWARP detached\n");
627 module_init(siw_init_module);
628 module_exit(siw_exit_module);
630 MODULE_ALIAS_RDMA_LINK("siw");