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 <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->raw_gid);
82 static void siw_device_cleanup(struct ib_device *base_dev)
84 struct siw_device *sdev = to_siw_dev(base_dev);
86 xa_destroy(&sdev->qp_xa);
87 xa_destroy(&sdev->mem_xa);
90 static int siw_dev_qualified(struct net_device *netdev)
93 * Additional hardware support can be added here
94 * (e.g. ARPHRD_FDDI, ARPHRD_ATM, ...) - see
95 * <linux/if_arp.h> for type identifiers.
97 if (netdev->type == ARPHRD_ETHER || netdev->type == ARPHRD_IEEE802 ||
98 netdev->type == ARPHRD_NONE ||
99 (netdev->type == ARPHRD_LOOPBACK && loopback_enabled))
105 static DEFINE_PER_CPU(atomic_t, siw_use_cnt);
108 struct cpumask **tx_valid_cpus;
112 static void siw_destroy_cpulist(int number)
117 kfree(siw_cpu_info.tx_valid_cpus[i++]);
119 kfree(siw_cpu_info.tx_valid_cpus);
120 siw_cpu_info.tx_valid_cpus = NULL;
123 static int siw_init_cpulist(void)
125 int i, num_nodes = nr_node_ids;
127 memset(siw_tx_thread, 0, sizeof(siw_tx_thread));
129 siw_cpu_info.num_nodes = num_nodes;
131 siw_cpu_info.tx_valid_cpus =
132 kcalloc(num_nodes, sizeof(struct cpumask *), GFP_KERNEL);
133 if (!siw_cpu_info.tx_valid_cpus) {
134 siw_cpu_info.num_nodes = 0;
137 for (i = 0; i < siw_cpu_info.num_nodes; i++) {
138 siw_cpu_info.tx_valid_cpus[i] =
139 kzalloc(sizeof(struct cpumask), GFP_KERNEL);
140 if (!siw_cpu_info.tx_valid_cpus[i])
143 cpumask_clear(siw_cpu_info.tx_valid_cpus[i]);
145 for_each_possible_cpu(i)
146 cpumask_set_cpu(i, siw_cpu_info.tx_valid_cpus[cpu_to_node(i)]);
151 siw_cpu_info.num_nodes = 0;
152 siw_destroy_cpulist(i);
158 * Choose CPU with least number of active QP's from NUMA node of
161 int siw_get_tx_cpu(struct siw_device *sdev)
163 const struct cpumask *tx_cpumask;
164 int i, num_cpus, cpu, min_use, node = sdev->numa_node, tx_cpu = -1;
167 tx_cpumask = cpu_online_mask;
169 tx_cpumask = siw_cpu_info.tx_valid_cpus[node];
171 num_cpus = cpumask_weight(tx_cpumask);
173 /* no CPU on this NUMA node */
174 tx_cpumask = cpu_online_mask;
175 num_cpus = cpumask_weight(tx_cpumask);
180 cpu = cpumask_first(tx_cpumask);
182 for (i = 0, min_use = SIW_MAX_QP; i < num_cpus;
183 i++, cpu = cpumask_next(cpu, tx_cpumask)) {
186 /* Skip any cores which have no TX thread */
187 if (!siw_tx_thread[cpu])
190 usage = atomic_read(&per_cpu(siw_use_cnt, cpu));
191 if (usage <= min_use) {
196 siw_dbg(&sdev->base_dev,
197 "tx cpu %d, node %d, %d qp's\n", tx_cpu, node, min_use);
201 atomic_inc(&per_cpu(siw_use_cnt, tx_cpu));
203 pr_warn("siw: no tx cpu found\n");
208 void siw_put_tx_cpu(int cpu)
210 atomic_dec(&per_cpu(siw_use_cnt, cpu));
213 static struct ib_qp *siw_get_base_qp(struct ib_device *base_dev, int id)
215 struct siw_qp *qp = siw_qp_id2obj(to_siw_dev(base_dev), id);
219 * siw_qp_id2obj() increments object reference count
227 static const struct ib_device_ops siw_device_ops = {
228 .owner = THIS_MODULE,
229 .uverbs_abi_ver = SIW_ABI_VERSION,
230 .driver_id = RDMA_DRIVER_SIW,
232 .alloc_mr = siw_alloc_mr,
233 .alloc_pd = siw_alloc_pd,
234 .alloc_ucontext = siw_alloc_ucontext,
235 .create_cq = siw_create_cq,
236 .create_qp = siw_create_qp,
237 .create_srq = siw_create_srq,
238 .dealloc_driver = siw_device_cleanup,
239 .dealloc_pd = siw_dealloc_pd,
240 .dealloc_ucontext = siw_dealloc_ucontext,
241 .dereg_mr = siw_dereg_mr,
242 .destroy_cq = siw_destroy_cq,
243 .destroy_qp = siw_destroy_qp,
244 .destroy_srq = siw_destroy_srq,
245 .get_dma_mr = siw_get_dma_mr,
246 .get_port_immutable = siw_get_port_immutable,
247 .iw_accept = siw_accept,
248 .iw_add_ref = siw_qp_get_ref,
249 .iw_connect = siw_connect,
250 .iw_create_listen = siw_create_listen,
251 .iw_destroy_listen = siw_destroy_listen,
252 .iw_get_qp = siw_get_base_qp,
253 .iw_reject = siw_reject,
254 .iw_rem_ref = siw_qp_put_ref,
255 .map_mr_sg = siw_map_mr_sg,
257 .mmap_free = siw_mmap_free,
258 .modify_qp = siw_verbs_modify_qp,
259 .modify_srq = siw_modify_srq,
260 .poll_cq = siw_poll_cq,
261 .post_recv = siw_post_receive,
262 .post_send = siw_post_send,
263 .post_srq_recv = siw_post_srq_recv,
264 .query_device = siw_query_device,
265 .query_gid = siw_query_gid,
266 .query_port = siw_query_port,
267 .query_qp = siw_query_qp,
268 .query_srq = siw_query_srq,
269 .req_notify_cq = siw_req_notify_cq,
270 .reg_user_mr = siw_reg_user_mr,
272 INIT_RDMA_OBJ_SIZE(ib_cq, siw_cq, base_cq),
273 INIT_RDMA_OBJ_SIZE(ib_pd, siw_pd, base_pd),
274 INIT_RDMA_OBJ_SIZE(ib_qp, siw_qp, base_qp),
275 INIT_RDMA_OBJ_SIZE(ib_srq, siw_srq, base_srq),
276 INIT_RDMA_OBJ_SIZE(ib_ucontext, siw_ucontext, base_ucontext),
279 static struct siw_device *siw_device_create(struct net_device *netdev)
281 struct siw_device *sdev = NULL;
282 struct ib_device *base_dev;
285 sdev = ib_alloc_device(siw_device, base_dev);
289 base_dev = &sdev->base_dev;
291 if (netdev->addr_len) {
292 memcpy(sdev->raw_gid, netdev->dev_addr,
293 min_t(unsigned int, netdev->addr_len, ETH_ALEN));
296 * This device does not have a HW address, but
297 * connection mangagement requires a unique gid.
299 eth_random_addr(sdev->raw_gid);
301 addrconf_addr_eui48((u8 *)&base_dev->node_guid, sdev->raw_gid);
303 base_dev->uverbs_cmd_mask |= BIT_ULL(IB_USER_VERBS_CMD_POST_SEND);
305 base_dev->node_type = RDMA_NODE_RNIC;
306 memcpy(base_dev->node_desc, SIW_NODE_DESC_COMMON,
307 sizeof(SIW_NODE_DESC_COMMON));
310 * Current model (one-to-one device association):
311 * One Softiwarp device per net_device or, equivalently,
314 base_dev->phys_port_cnt = 1;
315 base_dev->num_comp_vectors = num_possible_cpus();
317 xa_init_flags(&sdev->qp_xa, XA_FLAGS_ALLOC1);
318 xa_init_flags(&sdev->mem_xa, XA_FLAGS_ALLOC1);
320 ib_set_device_ops(base_dev, &siw_device_ops);
321 rv = ib_device_set_netdev(base_dev, netdev, 1);
325 memcpy(base_dev->iw_ifname, netdev->name,
326 sizeof(base_dev->iw_ifname));
328 /* Disable TCP port mapping */
329 base_dev->iw_driver_flags = IW_F_NO_PORT_MAP;
331 sdev->attrs.max_qp = SIW_MAX_QP;
332 sdev->attrs.max_qp_wr = SIW_MAX_QP_WR;
333 sdev->attrs.max_ord = SIW_MAX_ORD_QP;
334 sdev->attrs.max_ird = SIW_MAX_IRD_QP;
335 sdev->attrs.max_sge = SIW_MAX_SGE;
336 sdev->attrs.max_sge_rd = SIW_MAX_SGE_RD;
337 sdev->attrs.max_cq = SIW_MAX_CQ;
338 sdev->attrs.max_cqe = SIW_MAX_CQE;
339 sdev->attrs.max_mr = SIW_MAX_MR;
340 sdev->attrs.max_pd = SIW_MAX_PD;
341 sdev->attrs.max_mw = SIW_MAX_MW;
342 sdev->attrs.max_srq = SIW_MAX_SRQ;
343 sdev->attrs.max_srq_wr = SIW_MAX_SRQ_WR;
344 sdev->attrs.max_srq_sge = SIW_MAX_SGE;
346 INIT_LIST_HEAD(&sdev->cep_list);
347 INIT_LIST_HEAD(&sdev->qp_list);
349 atomic_set(&sdev->num_ctx, 0);
350 atomic_set(&sdev->num_srq, 0);
351 atomic_set(&sdev->num_qp, 0);
352 atomic_set(&sdev->num_cq, 0);
353 atomic_set(&sdev->num_mr, 0);
354 atomic_set(&sdev->num_pd, 0);
356 sdev->numa_node = dev_to_node(&netdev->dev);
357 spin_lock_init(&sdev->lock);
361 ib_dealloc_device(base_dev);
366 static int siw_netdev_event(struct notifier_block *nb, unsigned long event,
369 struct net_device *netdev = netdev_notifier_info_to_dev(arg);
370 struct ib_device *base_dev;
371 struct siw_device *sdev;
373 dev_dbg(&netdev->dev, "siw: event %lu\n", event);
375 base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW);
379 sdev = to_siw_dev(base_dev);
383 siw_port_event(sdev, 1, IB_EVENT_PORT_ACTIVE);
387 siw_port_event(sdev, 1, IB_EVENT_PORT_ERR);
390 case NETDEV_REGISTER:
392 * Device registration now handled only by
393 * rdma netlink commands. So it shall be impossible
394 * to end up here with a valid siw device.
396 siw_dbg(base_dev, "unexpected NETDEV_REGISTER event\n");
399 case NETDEV_UNREGISTER:
400 ib_unregister_device_queued(&sdev->base_dev);
403 case NETDEV_CHANGEADDR:
404 siw_port_event(sdev, 1, IB_EVENT_LID_CHANGE);
407 * All other events are not handled
412 ib_device_put(&sdev->base_dev);
417 static struct notifier_block siw_netdev_nb = {
418 .notifier_call = siw_netdev_event,
421 static int siw_newlink(const char *basedev_name, struct net_device *netdev)
423 struct ib_device *base_dev;
424 struct siw_device *sdev = NULL;
427 if (!siw_dev_qualified(netdev))
430 base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW);
432 ib_device_put(base_dev);
435 sdev = siw_device_create(netdev);
437 dev_dbg(&netdev->dev, "siw: new device\n");
438 ib_mark_name_assigned_by_user(&sdev->base_dev);
439 rv = siw_device_register(sdev, basedev_name);
441 ib_dealloc_device(&sdev->base_dev);
446 static struct rdma_link_ops siw_link_ops = {
448 .newlink = siw_newlink,
452 * siw_init_module - Initialize Softiwarp module and register with netdev
455 static __init int siw_init_module(void)
459 if (SENDPAGE_THRESH < SIW_MAX_INLINE) {
460 pr_info("siw: sendpage threshold too small: %u\n",
461 (int)SENDPAGE_THRESH);
465 rv = siw_init_cpulist();
473 if (!siw_create_tx_threads()) {
474 pr_info("siw: Could not start any TX thread\n");
479 * Locate CRC32 algorithm. If unsuccessful, fail
480 * loading siw only, if CRC is required.
482 siw_crypto_shash = crypto_alloc_shash("crc32c", 0, 0);
483 if (IS_ERR(siw_crypto_shash)) {
484 pr_info("siw: Loading CRC32c failed: %ld\n",
485 PTR_ERR(siw_crypto_shash));
486 siw_crypto_shash = NULL;
487 if (mpa_crc_required) {
492 rv = register_netdevice_notifier(&siw_netdev_nb);
496 rdma_link_register(&siw_link_ops);
498 pr_info("SoftiWARP attached\n");
502 siw_stop_tx_threads();
504 if (siw_crypto_shash)
505 crypto_free_shash(siw_crypto_shash);
507 pr_info("SoftIWARP attach failed. Error: %d\n", rv);
510 siw_destroy_cpulist(siw_cpu_info.num_nodes);
515 static void __exit siw_exit_module(void)
517 siw_stop_tx_threads();
519 unregister_netdevice_notifier(&siw_netdev_nb);
520 rdma_link_unregister(&siw_link_ops);
521 ib_unregister_driver(RDMA_DRIVER_SIW);
525 siw_destroy_cpulist(siw_cpu_info.num_nodes);
527 if (siw_crypto_shash)
528 crypto_free_shash(siw_crypto_shash);
530 pr_info("SoftiWARP detached\n");
533 module_init(siw_init_module);
534 module_exit(siw_exit_module);
536 MODULE_ALIAS_RDMA_LINK("siw");