1 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
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 <rdma/ib_verbs.h>
20 #include <rdma/ib_user_verbs.h>
21 #include <rdma/rdma_netlink.h>
22 #include <linux/kthread.h>
25 #include "siw_verbs.h"
27 MODULE_AUTHOR("Bernard Metzler");
28 MODULE_DESCRIPTION("Software iWARP Driver");
29 MODULE_LICENSE("Dual BSD/GPL");
31 /* transmit from user buffer, if possible */
32 const bool zcopy_tx = true;
34 /* Restrict usage of GSO, if hardware peer iwarp is unable to process
35 * large packets. try_gso = true lets siw try to use local GSO,
36 * if peer agrees. Not using GSO severly limits siw maximum tx bandwidth.
40 /* Attach siw also with loopback devices */
41 const bool loopback_enabled = true;
43 /* We try to negotiate CRC on, if true */
44 const bool mpa_crc_required;
46 /* MPA CRC on/off enforced */
47 const bool mpa_crc_strict;
49 /* Control TCP_NODELAY socket option */
50 const bool siw_tcp_nagle;
52 /* Select MPA version to be used during connection setup */
53 u_char mpa_version = MPA_REVISION_2;
55 /* Selects MPA P2P mode (additional handshake during connection
58 const bool peer_to_peer;
60 struct task_struct *siw_tx_thread[NR_CPUS];
61 struct crypto_shash *siw_crypto_shash;
63 static int siw_device_register(struct siw_device *sdev, const char *name)
65 struct ib_device *base_dev = &sdev->base_dev;
66 static int dev_id = 1;
69 rv = ib_register_device(base_dev, name);
71 pr_warn("siw: device registration error %d\n", rv);
74 sdev->vendor_part_id = dev_id++;
76 siw_dbg(base_dev, "HWaddr=%pM\n", sdev->netdev->dev_addr);
81 static void siw_device_cleanup(struct ib_device *base_dev)
83 struct siw_device *sdev = to_siw_dev(base_dev);
85 xa_destroy(&sdev->qp_xa);
86 xa_destroy(&sdev->mem_xa);
89 static int siw_create_tx_threads(void)
91 int cpu, assigned = 0;
93 for_each_online_cpu(cpu) {
95 if (cpu % cpumask_weight(topology_sibling_cpumask(cpu)))
99 kthread_create(siw_run_sq, (unsigned long *)(long)cpu,
101 if (IS_ERR(siw_tx_thread[cpu])) {
102 siw_tx_thread[cpu] = NULL;
105 kthread_bind(siw_tx_thread[cpu], cpu);
107 wake_up_process(siw_tx_thread[cpu]);
113 static int siw_dev_qualified(struct net_device *netdev)
116 * Additional hardware support can be added here
117 * (e.g. ARPHRD_FDDI, ARPHRD_ATM, ...) - see
118 * <linux/if_arp.h> for type identifiers.
120 if (netdev->type == ARPHRD_ETHER || netdev->type == ARPHRD_IEEE802 ||
121 (netdev->type == ARPHRD_LOOPBACK && loopback_enabled))
127 static DEFINE_PER_CPU(atomic_t, siw_use_cnt);
130 struct cpumask **tx_valid_cpus;
134 static int siw_init_cpulist(void)
136 int i, num_nodes = num_possible_nodes();
138 memset(siw_tx_thread, 0, sizeof(siw_tx_thread));
140 siw_cpu_info.num_nodes = num_nodes;
142 siw_cpu_info.tx_valid_cpus =
143 kcalloc(num_nodes, sizeof(struct cpumask *), GFP_KERNEL);
144 if (!siw_cpu_info.tx_valid_cpus) {
145 siw_cpu_info.num_nodes = 0;
148 for (i = 0; i < siw_cpu_info.num_nodes; i++) {
149 siw_cpu_info.tx_valid_cpus[i] =
150 kzalloc(sizeof(struct cpumask), GFP_KERNEL);
151 if (!siw_cpu_info.tx_valid_cpus[i])
154 cpumask_clear(siw_cpu_info.tx_valid_cpus[i]);
156 for_each_possible_cpu(i)
157 cpumask_set_cpu(i, siw_cpu_info.tx_valid_cpus[cpu_to_node(i)]);
162 siw_cpu_info.num_nodes = 0;
164 kfree(siw_cpu_info.tx_valid_cpus[i]);
165 kfree(siw_cpu_info.tx_valid_cpus);
166 siw_cpu_info.tx_valid_cpus = NULL;
171 static void siw_destroy_cpulist(void)
175 while (i < siw_cpu_info.num_nodes)
176 kfree(siw_cpu_info.tx_valid_cpus[i++]);
178 kfree(siw_cpu_info.tx_valid_cpus);
182 * Choose CPU with least number of active QP's from NUMA node of
185 int siw_get_tx_cpu(struct siw_device *sdev)
187 const struct cpumask *tx_cpumask;
188 int i, num_cpus, cpu, min_use, node = sdev->numa_node, tx_cpu = -1;
191 tx_cpumask = cpu_online_mask;
193 tx_cpumask = siw_cpu_info.tx_valid_cpus[node];
195 num_cpus = cpumask_weight(tx_cpumask);
197 /* no CPU on this NUMA node */
198 tx_cpumask = cpu_online_mask;
199 num_cpus = cpumask_weight(tx_cpumask);
204 cpu = cpumask_first(tx_cpumask);
206 for (i = 0, min_use = SIW_MAX_QP; i < num_cpus;
207 i++, cpu = cpumask_next(cpu, tx_cpumask)) {
210 /* Skip any cores which have no TX thread */
211 if (!siw_tx_thread[cpu])
214 usage = atomic_read(&per_cpu(siw_use_cnt, cpu));
215 if (usage <= min_use) {
220 siw_dbg(&sdev->base_dev,
221 "tx cpu %d, node %d, %d qp's\n", tx_cpu, node, min_use);
225 atomic_inc(&per_cpu(siw_use_cnt, tx_cpu));
227 pr_warn("siw: no tx cpu found\n");
232 void siw_put_tx_cpu(int cpu)
234 atomic_dec(&per_cpu(siw_use_cnt, cpu));
237 static struct ib_qp *siw_get_base_qp(struct ib_device *base_dev, int id)
239 struct siw_qp *qp = siw_qp_id2obj(to_siw_dev(base_dev), id);
243 * siw_qp_id2obj() increments object reference count
251 static void siw_verbs_sq_flush(struct ib_qp *base_qp)
253 struct siw_qp *qp = to_siw_qp(base_qp);
255 down_write(&qp->state_lock);
257 up_write(&qp->state_lock);
260 static void siw_verbs_rq_flush(struct ib_qp *base_qp)
262 struct siw_qp *qp = to_siw_qp(base_qp);
264 down_write(&qp->state_lock);
266 up_write(&qp->state_lock);
269 static const struct ib_device_ops siw_device_ops = {
270 .owner = THIS_MODULE,
271 .uverbs_abi_ver = SIW_ABI_VERSION,
272 .driver_id = RDMA_DRIVER_SIW,
274 .alloc_mr = siw_alloc_mr,
275 .alloc_pd = siw_alloc_pd,
276 .alloc_ucontext = siw_alloc_ucontext,
277 .create_cq = siw_create_cq,
278 .create_qp = siw_create_qp,
279 .create_srq = siw_create_srq,
280 .dealloc_driver = siw_device_cleanup,
281 .dealloc_pd = siw_dealloc_pd,
282 .dealloc_ucontext = siw_dealloc_ucontext,
283 .dereg_mr = siw_dereg_mr,
284 .destroy_cq = siw_destroy_cq,
285 .destroy_qp = siw_destroy_qp,
286 .destroy_srq = siw_destroy_srq,
287 .drain_rq = siw_verbs_rq_flush,
288 .drain_sq = siw_verbs_sq_flush,
289 .get_dma_mr = siw_get_dma_mr,
290 .get_port_immutable = siw_get_port_immutable,
291 .iw_accept = siw_accept,
292 .iw_add_ref = siw_qp_get_ref,
293 .iw_connect = siw_connect,
294 .iw_create_listen = siw_create_listen,
295 .iw_destroy_listen = siw_destroy_listen,
296 .iw_get_qp = siw_get_base_qp,
297 .iw_reject = siw_reject,
298 .iw_rem_ref = siw_qp_put_ref,
299 .map_mr_sg = siw_map_mr_sg,
301 .modify_qp = siw_verbs_modify_qp,
302 .modify_srq = siw_modify_srq,
303 .poll_cq = siw_poll_cq,
304 .post_recv = siw_post_receive,
305 .post_send = siw_post_send,
306 .post_srq_recv = siw_post_srq_recv,
307 .query_device = siw_query_device,
308 .query_gid = siw_query_gid,
309 .query_pkey = siw_query_pkey,
310 .query_port = siw_query_port,
311 .query_qp = siw_query_qp,
312 .query_srq = siw_query_srq,
313 .req_notify_cq = siw_req_notify_cq,
314 .reg_user_mr = siw_reg_user_mr,
316 INIT_RDMA_OBJ_SIZE(ib_cq, siw_cq, base_cq),
317 INIT_RDMA_OBJ_SIZE(ib_pd, siw_pd, base_pd),
318 INIT_RDMA_OBJ_SIZE(ib_srq, siw_srq, base_srq),
319 INIT_RDMA_OBJ_SIZE(ib_ucontext, siw_ucontext, base_ucontext),
322 static struct siw_device *siw_device_create(struct net_device *netdev)
324 struct siw_device *sdev = NULL;
325 struct ib_device *base_dev;
326 struct device *parent = netdev->dev.parent;
331 * The loopback device has no parent device,
332 * so it appears as a top-level device. To support
333 * loopback device connectivity, take this device
334 * as the parent device. Skip all other devices
337 if (netdev->type != ARPHRD_LOOPBACK) {
338 pr_warn("siw: device %s error: no parent device\n",
342 parent = &netdev->dev;
344 sdev = ib_alloc_device(siw_device, base_dev);
348 base_dev = &sdev->base_dev;
350 sdev->netdev = netdev;
352 if (netdev->type != ARPHRD_LOOPBACK) {
353 memcpy(&base_dev->node_guid, netdev->dev_addr, 6);
356 * The loopback device does not have a HW address,
357 * but connection mangagement lib expects gid != 0
359 size_t gidlen = min_t(size_t, strlen(base_dev->name), 6);
361 memcpy(&base_dev->node_guid, base_dev->name, gidlen);
363 base_dev->uverbs_cmd_mask =
364 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE) |
365 (1ull << IB_USER_VERBS_CMD_QUERY_PORT) |
366 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT) |
367 (1ull << IB_USER_VERBS_CMD_ALLOC_PD) |
368 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD) |
369 (1ull << IB_USER_VERBS_CMD_REG_MR) |
370 (1ull << IB_USER_VERBS_CMD_DEREG_MR) |
371 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
372 (1ull << IB_USER_VERBS_CMD_CREATE_CQ) |
373 (1ull << IB_USER_VERBS_CMD_POLL_CQ) |
374 (1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ) |
375 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ) |
376 (1ull << IB_USER_VERBS_CMD_CREATE_QP) |
377 (1ull << IB_USER_VERBS_CMD_QUERY_QP) |
378 (1ull << IB_USER_VERBS_CMD_MODIFY_QP) |
379 (1ull << IB_USER_VERBS_CMD_DESTROY_QP) |
380 (1ull << IB_USER_VERBS_CMD_POST_SEND) |
381 (1ull << IB_USER_VERBS_CMD_POST_RECV) |
382 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ) |
383 (1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV) |
384 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ) |
385 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ) |
386 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ);
388 base_dev->node_type = RDMA_NODE_RNIC;
389 memcpy(base_dev->node_desc, SIW_NODE_DESC_COMMON,
390 sizeof(SIW_NODE_DESC_COMMON));
393 * Current model (one-to-one device association):
394 * One Softiwarp device per net_device or, equivalently,
397 base_dev->phys_port_cnt = 1;
398 base_dev->dev.parent = parent;
399 base_dev->dev.dma_ops = &dma_virt_ops;
400 base_dev->num_comp_vectors = num_possible_cpus();
402 ib_set_device_ops(base_dev, &siw_device_ops);
403 rv = ib_device_set_netdev(base_dev, netdev, 1);
407 memcpy(base_dev->iw_ifname, netdev->name,
408 sizeof(base_dev->iw_ifname));
410 /* Disable TCP port mapping */
411 base_dev->iw_driver_flags = IW_F_NO_PORT_MAP,
413 sdev->attrs.max_qp = SIW_MAX_QP;
414 sdev->attrs.max_qp_wr = SIW_MAX_QP_WR;
415 sdev->attrs.max_ord = SIW_MAX_ORD_QP;
416 sdev->attrs.max_ird = SIW_MAX_IRD_QP;
417 sdev->attrs.max_sge = SIW_MAX_SGE;
418 sdev->attrs.max_sge_rd = SIW_MAX_SGE_RD;
419 sdev->attrs.max_cq = SIW_MAX_CQ;
420 sdev->attrs.max_cqe = SIW_MAX_CQE;
421 sdev->attrs.max_mr = SIW_MAX_MR;
422 sdev->attrs.max_pd = SIW_MAX_PD;
423 sdev->attrs.max_mw = SIW_MAX_MW;
424 sdev->attrs.max_fmr = SIW_MAX_FMR;
425 sdev->attrs.max_srq = SIW_MAX_SRQ;
426 sdev->attrs.max_srq_wr = SIW_MAX_SRQ_WR;
427 sdev->attrs.max_srq_sge = SIW_MAX_SGE;
429 xa_init_flags(&sdev->qp_xa, XA_FLAGS_ALLOC1);
430 xa_init_flags(&sdev->mem_xa, XA_FLAGS_ALLOC1);
432 INIT_LIST_HEAD(&sdev->cep_list);
433 INIT_LIST_HEAD(&sdev->qp_list);
435 atomic_set(&sdev->num_ctx, 0);
436 atomic_set(&sdev->num_srq, 0);
437 atomic_set(&sdev->num_qp, 0);
438 atomic_set(&sdev->num_cq, 0);
439 atomic_set(&sdev->num_mr, 0);
440 atomic_set(&sdev->num_pd, 0);
442 sdev->numa_node = dev_to_node(parent);
443 spin_lock_init(&sdev->lock);
447 ib_dealloc_device(base_dev);
453 * Network link becomes unavailable. Mark all
454 * affected QP's accordingly.
456 static void siw_netdev_down(struct work_struct *work)
458 struct siw_device *sdev =
459 container_of(work, struct siw_device, netdev_down);
461 struct siw_qp_attrs qp_attrs;
462 struct list_head *pos, *tmp;
464 memset(&qp_attrs, 0, sizeof(qp_attrs));
465 qp_attrs.state = SIW_QP_STATE_ERROR;
467 list_for_each_safe(pos, tmp, &sdev->qp_list) {
468 struct siw_qp *qp = list_entry(pos, struct siw_qp, devq);
470 down_write(&qp->state_lock);
471 WARN_ON(siw_qp_modify(qp, &qp_attrs, SIW_QP_ATTR_STATE));
472 up_write(&qp->state_lock);
474 ib_device_put(&sdev->base_dev);
477 static void siw_device_goes_down(struct siw_device *sdev)
479 if (ib_device_try_get(&sdev->base_dev)) {
480 INIT_WORK(&sdev->netdev_down, siw_netdev_down);
481 schedule_work(&sdev->netdev_down);
485 static int siw_netdev_event(struct notifier_block *nb, unsigned long event,
488 struct net_device *netdev = netdev_notifier_info_to_dev(arg);
489 struct ib_device *base_dev;
490 struct siw_device *sdev;
492 dev_dbg(&netdev->dev, "siw: event %lu\n", event);
494 if (dev_net(netdev) != &init_net)
497 base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW);
501 sdev = to_siw_dev(base_dev);
505 sdev->state = IB_PORT_ACTIVE;
506 siw_port_event(sdev, 1, IB_EVENT_PORT_ACTIVE);
509 case NETDEV_GOING_DOWN:
510 siw_device_goes_down(sdev);
514 sdev->state = IB_PORT_DOWN;
515 siw_port_event(sdev, 1, IB_EVENT_PORT_ERR);
518 case NETDEV_REGISTER:
520 * Device registration now handled only by
521 * rdma netlink commands. So it shall be impossible
522 * to end up here with a valid siw device.
524 siw_dbg(base_dev, "unexpected NETDEV_REGISTER event\n");
527 case NETDEV_UNREGISTER:
528 ib_unregister_device_queued(&sdev->base_dev);
531 case NETDEV_CHANGEADDR:
532 siw_port_event(sdev, 1, IB_EVENT_LID_CHANGE);
535 * Todo: Below netdev events are currently not handled.
537 case NETDEV_CHANGEMTU:
544 ib_device_put(&sdev->base_dev);
549 static struct notifier_block siw_netdev_nb = {
550 .notifier_call = siw_netdev_event,
553 static int siw_newlink(const char *basedev_name, struct net_device *netdev)
555 struct ib_device *base_dev;
556 struct siw_device *sdev = NULL;
559 if (!siw_dev_qualified(netdev))
562 base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW);
564 ib_device_put(base_dev);
567 sdev = siw_device_create(netdev);
569 dev_dbg(&netdev->dev, "siw: new device\n");
571 if (netif_running(netdev) && netif_carrier_ok(netdev))
572 sdev->state = IB_PORT_ACTIVE;
574 sdev->state = IB_PORT_DOWN;
576 rv = siw_device_register(sdev, basedev_name);
578 ib_dealloc_device(&sdev->base_dev);
583 static struct rdma_link_ops siw_link_ops = {
585 .newlink = siw_newlink,
589 * siw_init_module - Initialize Softiwarp module and register with netdev
592 static __init int siw_init_module(void)
597 if (SENDPAGE_THRESH < SIW_MAX_INLINE) {
598 pr_info("siw: sendpage threshold too small: %u\n",
599 (int)SENDPAGE_THRESH);
603 rv = siw_init_cpulist();
611 if (!siw_create_tx_threads()) {
612 pr_info("siw: Could not start any TX thread\n");
617 * Locate CRC32 algorithm. If unsuccessful, fail
618 * loading siw only, if CRC is required.
620 siw_crypto_shash = crypto_alloc_shash("crc32c", 0, 0);
621 if (IS_ERR(siw_crypto_shash)) {
622 pr_info("siw: Loading CRC32c failed: %ld\n",
623 PTR_ERR(siw_crypto_shash));
624 siw_crypto_shash = NULL;
625 if (mpa_crc_required) {
630 rv = register_netdevice_notifier(&siw_netdev_nb);
634 rdma_link_register(&siw_link_ops);
636 pr_info("SoftiWARP attached\n");
640 for (nr_cpu = 0; nr_cpu < nr_cpu_ids; nr_cpu++) {
641 if (siw_tx_thread[nr_cpu]) {
642 siw_stop_tx_thread(nr_cpu);
643 siw_tx_thread[nr_cpu] = NULL;
646 if (siw_crypto_shash)
647 crypto_free_shash(siw_crypto_shash);
649 pr_info("SoftIWARP attach failed. Error: %d\n", rv);
652 siw_destroy_cpulist();
657 static void __exit siw_exit_module(void)
661 for_each_possible_cpu(cpu) {
662 if (siw_tx_thread[cpu]) {
663 siw_stop_tx_thread(cpu);
664 siw_tx_thread[cpu] = NULL;
667 unregister_netdevice_notifier(&siw_netdev_nb);
668 rdma_link_unregister(&siw_link_ops);
669 ib_unregister_driver(RDMA_DRIVER_SIW);
673 siw_destroy_cpulist();
675 if (siw_crypto_shash)
676 crypto_free_shash(siw_crypto_shash);
678 pr_info("SoftiWARP detached\n");
681 module_init(siw_init_module);
682 module_exit(siw_exit_module);
684 MODULE_ALIAS_RDMA_LINK("siw");