1// SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2
3/* Authors: Cheng Xu <chengyou@linux.alibaba.com> */
4/*          Kai Shen <kaishen@linux.alibaba.com> */
5/* Copyright (c) 2020-2022, Alibaba Group. */
6
7#include <linux/module.h>
8#include <net/addrconf.h>
9#include <rdma/erdma-abi.h>
10
11#include "erdma.h"
12#include "erdma_cm.h"
13#include "erdma_verbs.h"
14
15MODULE_AUTHOR("Cheng Xu <chengyou@linux.alibaba.com>");
16MODULE_DESCRIPTION("Alibaba elasticRDMA adapter driver");
17MODULE_LICENSE("Dual BSD/GPL");
18
19static int erdma_netdev_event(struct notifier_block *nb, unsigned long event,
20			      void *arg)
21{
22	struct net_device *netdev = netdev_notifier_info_to_dev(arg);
23	struct erdma_dev *dev = container_of(nb, struct erdma_dev, netdev_nb);
24
25	if (dev->netdev == NULL || dev->netdev != netdev)
26		goto done;
27
28	switch (event) {
29	case NETDEV_UP:
30		dev->state = IB_PORT_ACTIVE;
31		erdma_port_event(dev, IB_EVENT_PORT_ACTIVE);
32		break;
33	case NETDEV_DOWN:
34		dev->state = IB_PORT_DOWN;
35		erdma_port_event(dev, IB_EVENT_PORT_ERR);
36		break;
37	case NETDEV_CHANGEMTU:
38		if (dev->mtu != netdev->mtu) {
39			erdma_set_mtu(dev, netdev->mtu);
40			dev->mtu = netdev->mtu;
41		}
42		break;
43	case NETDEV_REGISTER:
44	case NETDEV_UNREGISTER:
45	case NETDEV_CHANGEADDR:
46	case NETDEV_GOING_DOWN:
47	case NETDEV_CHANGE:
48	default:
49		break;
50	}
51
52done:
53	return NOTIFY_OK;
54}
55
56static int erdma_enum_and_get_netdev(struct erdma_dev *dev)
57{
58	struct net_device *netdev;
59	int ret = -EPROBE_DEFER;
60
61	/* Already binded to a net_device, so we skip. */
62	if (dev->netdev)
63		return 0;
64
65	rtnl_lock();
66	for_each_netdev(&init_net, netdev) {
67		/*
68		 * In erdma, the paired netdev and ibdev should have the same
69		 * MAC address. erdma can get the value from its PCIe bar
70		 * registers. Since erdma can not get the paired netdev
71		 * reference directly, we do a traverse here to get the paired
72		 * netdev.
73		 */
74		if (ether_addr_equal_unaligned(netdev->perm_addr,
75					       dev->attrs.peer_addr)) {
76			ret = ib_device_set_netdev(&dev->ibdev, netdev, 1);
77			if (ret) {
78				rtnl_unlock();
79				ibdev_warn(&dev->ibdev,
80					   "failed (%d) to link netdev", ret);
81				return ret;
82			}
83
84			dev->netdev = netdev;
85			break;
86		}
87	}
88
89	rtnl_unlock();
90
91	return ret;
92}
93
94static int erdma_device_register(struct erdma_dev *dev)
95{
96	struct ib_device *ibdev = &dev->ibdev;
97	int ret;
98
99	ret = erdma_enum_and_get_netdev(dev);
100	if (ret)
101		return ret;
102
103	dev->mtu = dev->netdev->mtu;
104	addrconf_addr_eui48((u8 *)&ibdev->node_guid, dev->netdev->dev_addr);
105
106	ret = ib_register_device(ibdev, "erdma_%d", &dev->pdev->dev);
107	if (ret) {
108		dev_err(&dev->pdev->dev,
109			"ib_register_device failed: ret = %d\n", ret);
110		return ret;
111	}
112
113	dev->netdev_nb.notifier_call = erdma_netdev_event;
114	ret = register_netdevice_notifier(&dev->netdev_nb);
115	if (ret) {
116		ibdev_err(&dev->ibdev, "failed to register notifier.\n");
117		ib_unregister_device(ibdev);
118	}
119
120	return ret;
121}
122
123static irqreturn_t erdma_comm_irq_handler(int irq, void *data)
124{
125	struct erdma_dev *dev = data;
126
127	erdma_cmdq_completion_handler(&dev->cmdq);
128	erdma_aeq_event_handler(dev);
129
130	return IRQ_HANDLED;
131}
132
133static int erdma_request_vectors(struct erdma_dev *dev)
134{
135	int expect_irq_num = min(num_possible_cpus() + 1, ERDMA_NUM_MSIX_VEC);
136	int ret;
137
138	ret = pci_alloc_irq_vectors(dev->pdev, 1, expect_irq_num, PCI_IRQ_MSIX);
139	if (ret < 0) {
140		dev_err(&dev->pdev->dev, "request irq vectors failed(%d)\n",
141			ret);
142		return ret;
143	}
144	dev->attrs.irq_num = ret;
145
146	return 0;
147}
148
149static int erdma_comm_irq_init(struct erdma_dev *dev)
150{
151	snprintf(dev->comm_irq.name, ERDMA_IRQNAME_SIZE, "erdma-common@pci:%s",
152		 pci_name(dev->pdev));
153	dev->comm_irq.msix_vector =
154		pci_irq_vector(dev->pdev, ERDMA_MSIX_VECTOR_CMDQ);
155
156	cpumask_set_cpu(cpumask_first(cpumask_of_pcibus(dev->pdev->bus)),
157			&dev->comm_irq.affinity_hint_mask);
158	irq_set_affinity_hint(dev->comm_irq.msix_vector,
159			      &dev->comm_irq.affinity_hint_mask);
160
161	return request_irq(dev->comm_irq.msix_vector, erdma_comm_irq_handler, 0,
162			   dev->comm_irq.name, dev);
163}
164
165static void erdma_comm_irq_uninit(struct erdma_dev *dev)
166{
167	irq_set_affinity_hint(dev->comm_irq.msix_vector, NULL);
168	free_irq(dev->comm_irq.msix_vector, dev);
169}
170
171static int erdma_device_init(struct erdma_dev *dev, struct pci_dev *pdev)
172{
173	int ret;
174
175	dev->resp_pool = dma_pool_create("erdma_resp_pool", &pdev->dev,
176					 ERDMA_HW_RESP_SIZE, ERDMA_HW_RESP_SIZE,
177					 0);
178	if (!dev->resp_pool)
179		return -ENOMEM;
180
181	ret = dma_set_mask_and_coherent(&pdev->dev,
182					DMA_BIT_MASK(ERDMA_PCI_WIDTH));
183	if (ret)
184		goto destroy_pool;
185
186	dma_set_max_seg_size(&pdev->dev, UINT_MAX);
187
188	return 0;
189
190destroy_pool:
191	dma_pool_destroy(dev->resp_pool);
192
193	return ret;
194}
195
196static void erdma_device_uninit(struct erdma_dev *dev)
197{
198	dma_pool_destroy(dev->resp_pool);
199}
200
201static void erdma_hw_reset(struct erdma_dev *dev)
202{
203	u32 ctrl = FIELD_PREP(ERDMA_REG_DEV_CTRL_RESET_MASK, 1);
204
205	erdma_reg_write32(dev, ERDMA_REGS_DEV_CTRL_REG, ctrl);
206}
207
208static int erdma_wait_hw_init_done(struct erdma_dev *dev)
209{
210	int i;
211
212	erdma_reg_write32(dev, ERDMA_REGS_DEV_CTRL_REG,
213			  FIELD_PREP(ERDMA_REG_DEV_CTRL_INIT_MASK, 1));
214
215	for (i = 0; i < ERDMA_WAIT_DEV_DONE_CNT; i++) {
216		if (erdma_reg_read32_filed(dev, ERDMA_REGS_DEV_ST_REG,
217					   ERDMA_REG_DEV_ST_INIT_DONE_MASK))
218			break;
219
220		msleep(ERDMA_REG_ACCESS_WAIT_MS);
221	}
222
223	if (i == ERDMA_WAIT_DEV_DONE_CNT) {
224		dev_err(&dev->pdev->dev, "wait init done failed.\n");
225		return -ETIMEDOUT;
226	}
227
228	return 0;
229}
230
231static const struct pci_device_id erdma_pci_tbl[] = {
232	{ PCI_DEVICE(PCI_VENDOR_ID_ALIBABA, 0x107f) },
233	{}
234};
235
236static int erdma_probe_dev(struct pci_dev *pdev)
237{
238	struct erdma_dev *dev;
239	int bars, err;
240	u32 version;
241
242	err = pci_enable_device(pdev);
243	if (err) {
244		dev_err(&pdev->dev, "pci_enable_device failed(%d)\n", err);
245		return err;
246	}
247
248	pci_set_master(pdev);
249
250	dev = ib_alloc_device(erdma_dev, ibdev);
251	if (!dev) {
252		dev_err(&pdev->dev, "ib_alloc_device failed\n");
253		err = -ENOMEM;
254		goto err_disable_device;
255	}
256
257	pci_set_drvdata(pdev, dev);
258	dev->pdev = pdev;
259	dev->attrs.numa_node = dev_to_node(&pdev->dev);
260
261	bars = pci_select_bars(pdev, IORESOURCE_MEM);
262	err = pci_request_selected_regions(pdev, bars, DRV_MODULE_NAME);
263	if (bars != ERDMA_BAR_MASK || err) {
264		err = err ? err : -EINVAL;
265		goto err_ib_device_release;
266	}
267
268	dev->func_bar_addr = pci_resource_start(pdev, ERDMA_FUNC_BAR);
269	dev->func_bar_len = pci_resource_len(pdev, ERDMA_FUNC_BAR);
270
271	dev->func_bar =
272		devm_ioremap(&pdev->dev, dev->func_bar_addr, dev->func_bar_len);
273	if (!dev->func_bar) {
274		dev_err(&pdev->dev, "devm_ioremap failed.\n");
275		err = -EFAULT;
276		goto err_release_bars;
277	}
278
279	version = erdma_reg_read32(dev, ERDMA_REGS_VERSION_REG);
280	if (version == 0) {
281		/* we knows that it is a non-functional function. */
282		err = -ENODEV;
283		goto err_iounmap_func_bar;
284	}
285
286	err = erdma_device_init(dev, pdev);
287	if (err)
288		goto err_iounmap_func_bar;
289
290	err = erdma_request_vectors(dev);
291	if (err)
292		goto err_uninit_device;
293
294	err = erdma_comm_irq_init(dev);
295	if (err)
296		goto err_free_vectors;
297
298	err = erdma_aeq_init(dev);
299	if (err)
300		goto err_uninit_comm_irq;
301
302	err = erdma_cmdq_init(dev);
303	if (err)
304		goto err_uninit_aeq;
305
306	err = erdma_wait_hw_init_done(dev);
307	if (err)
308		goto err_uninit_cmdq;
309
310	err = erdma_ceqs_init(dev);
311	if (err)
312		goto err_reset_hw;
313
314	erdma_finish_cmdq_init(dev);
315
316	return 0;
317
318err_reset_hw:
319	erdma_hw_reset(dev);
320
321err_uninit_cmdq:
322	erdma_cmdq_destroy(dev);
323
324err_uninit_aeq:
325	erdma_aeq_destroy(dev);
326
327err_uninit_comm_irq:
328	erdma_comm_irq_uninit(dev);
329
330err_free_vectors:
331	pci_free_irq_vectors(dev->pdev);
332
333err_uninit_device:
334	erdma_device_uninit(dev);
335
336err_iounmap_func_bar:
337	devm_iounmap(&pdev->dev, dev->func_bar);
338
339err_release_bars:
340	pci_release_selected_regions(pdev, bars);
341
342err_ib_device_release:
343	ib_dealloc_device(&dev->ibdev);
344
345err_disable_device:
346	pci_disable_device(pdev);
347
348	return err;
349}
350
351static void erdma_remove_dev(struct pci_dev *pdev)
352{
353	struct erdma_dev *dev = pci_get_drvdata(pdev);
354
355	erdma_ceqs_uninit(dev);
356	erdma_hw_reset(dev);
357	erdma_cmdq_destroy(dev);
358	erdma_aeq_destroy(dev);
359	erdma_comm_irq_uninit(dev);
360	pci_free_irq_vectors(dev->pdev);
361	erdma_device_uninit(dev);
362
363	devm_iounmap(&pdev->dev, dev->func_bar);
364	pci_release_selected_regions(pdev, ERDMA_BAR_MASK);
365
366	ib_dealloc_device(&dev->ibdev);
367
368	pci_disable_device(pdev);
369}
370
371#define ERDMA_GET_CAP(name, cap) FIELD_GET(ERDMA_CMD_DEV_CAP_##name##_MASK, cap)
372
373static int erdma_dev_attrs_init(struct erdma_dev *dev)
374{
375	int err;
376	u64 req_hdr, cap0, cap1;
377
378	erdma_cmdq_build_reqhdr(&req_hdr, CMDQ_SUBMOD_RDMA,
379				CMDQ_OPCODE_QUERY_DEVICE);
380
381	err = erdma_post_cmd_wait(&dev->cmdq, &req_hdr, sizeof(req_hdr), &cap0,
382				  &cap1);
383	if (err)
384		return err;
385
386	dev->attrs.max_cqe = 1 << ERDMA_GET_CAP(MAX_CQE, cap0);
387	dev->attrs.max_mr_size = 1ULL << ERDMA_GET_CAP(MAX_MR_SIZE, cap0);
388	dev->attrs.max_mw = 1 << ERDMA_GET_CAP(MAX_MW, cap1);
389	dev->attrs.max_recv_wr = 1 << ERDMA_GET_CAP(MAX_RECV_WR, cap0);
390	dev->attrs.local_dma_key = ERDMA_GET_CAP(DMA_LOCAL_KEY, cap1);
391	dev->attrs.cc = ERDMA_GET_CAP(DEFAULT_CC, cap1);
392	dev->attrs.max_qp = ERDMA_NQP_PER_QBLOCK * ERDMA_GET_CAP(QBLOCK, cap1);
393	dev->attrs.max_mr = dev->attrs.max_qp << 1;
394	dev->attrs.max_cq = dev->attrs.max_qp << 1;
395	dev->attrs.cap_flags = ERDMA_GET_CAP(FLAGS, cap0);
396
397	dev->attrs.max_send_wr = ERDMA_MAX_SEND_WR;
398	dev->attrs.max_ord = ERDMA_MAX_ORD;
399	dev->attrs.max_ird = ERDMA_MAX_IRD;
400	dev->attrs.max_send_sge = ERDMA_MAX_SEND_SGE;
401	dev->attrs.max_recv_sge = ERDMA_MAX_RECV_SGE;
402	dev->attrs.max_sge_rd = ERDMA_MAX_SGE_RD;
403	dev->attrs.max_pd = ERDMA_MAX_PD;
404
405	dev->res_cb[ERDMA_RES_TYPE_PD].max_cap = ERDMA_MAX_PD;
406	dev->res_cb[ERDMA_RES_TYPE_STAG_IDX].max_cap = dev->attrs.max_mr;
407
408	erdma_cmdq_build_reqhdr(&req_hdr, CMDQ_SUBMOD_COMMON,
409				CMDQ_OPCODE_QUERY_FW_INFO);
410
411	err = erdma_post_cmd_wait(&dev->cmdq, &req_hdr, sizeof(req_hdr), &cap0,
412				  &cap1);
413	if (!err)
414		dev->attrs.fw_version =
415			FIELD_GET(ERDMA_CMD_INFO0_FW_VER_MASK, cap0);
416
417	return err;
418}
419
420static int erdma_device_config(struct erdma_dev *dev)
421{
422	struct erdma_cmdq_config_device_req req = {};
423
424	if (!(dev->attrs.cap_flags & ERDMA_DEV_CAP_FLAGS_EXTEND_DB))
425		return 0;
426
427	erdma_cmdq_build_reqhdr(&req.hdr, CMDQ_SUBMOD_COMMON,
428				CMDQ_OPCODE_CONF_DEVICE);
429
430	req.cfg = FIELD_PREP(ERDMA_CMD_CONFIG_DEVICE_PGSHIFT_MASK, PAGE_SHIFT) |
431		  FIELD_PREP(ERDMA_CMD_CONFIG_DEVICE_PS_EN_MASK, 1);
432
433	return erdma_post_cmd_wait(&dev->cmdq, &req, sizeof(req), NULL, NULL);
434}
435
436static int erdma_res_cb_init(struct erdma_dev *dev)
437{
438	int i, j;
439
440	for (i = 0; i < ERDMA_RES_CNT; i++) {
441		dev->res_cb[i].next_alloc_idx = 1;
442		spin_lock_init(&dev->res_cb[i].lock);
443		dev->res_cb[i].bitmap =
444			bitmap_zalloc(dev->res_cb[i].max_cap, GFP_KERNEL);
445		if (!dev->res_cb[i].bitmap)
446			goto err;
447	}
448
449	return 0;
450
451err:
452	for (j = 0; j < i; j++)
453		bitmap_free(dev->res_cb[j].bitmap);
454
455	return -ENOMEM;
456}
457
458static void erdma_res_cb_free(struct erdma_dev *dev)
459{
460	int i;
461
462	for (i = 0; i < ERDMA_RES_CNT; i++)
463		bitmap_free(dev->res_cb[i].bitmap);
464}
465
466static const struct ib_device_ops erdma_device_ops = {
467	.owner = THIS_MODULE,
468	.driver_id = RDMA_DRIVER_ERDMA,
469	.uverbs_abi_ver = ERDMA_ABI_VERSION,
470
471	.alloc_hw_port_stats = erdma_alloc_hw_port_stats,
472	.alloc_mr = erdma_ib_alloc_mr,
473	.alloc_pd = erdma_alloc_pd,
474	.alloc_ucontext = erdma_alloc_ucontext,
475	.create_cq = erdma_create_cq,
476	.create_qp = erdma_create_qp,
477	.dealloc_pd = erdma_dealloc_pd,
478	.dealloc_ucontext = erdma_dealloc_ucontext,
479	.dereg_mr = erdma_dereg_mr,
480	.destroy_cq = erdma_destroy_cq,
481	.destroy_qp = erdma_destroy_qp,
482	.get_dma_mr = erdma_get_dma_mr,
483	.get_hw_stats = erdma_get_hw_stats,
484	.get_port_immutable = erdma_get_port_immutable,
485	.iw_accept = erdma_accept,
486	.iw_add_ref = erdma_qp_get_ref,
487	.iw_connect = erdma_connect,
488	.iw_create_listen = erdma_create_listen,
489	.iw_destroy_listen = erdma_destroy_listen,
490	.iw_get_qp = erdma_get_ibqp,
491	.iw_reject = erdma_reject,
492	.iw_rem_ref = erdma_qp_put_ref,
493	.map_mr_sg = erdma_map_mr_sg,
494	.mmap = erdma_mmap,
495	.mmap_free = erdma_mmap_free,
496	.modify_qp = erdma_modify_qp,
497	.post_recv = erdma_post_recv,
498	.post_send = erdma_post_send,
499	.poll_cq = erdma_poll_cq,
500	.query_device = erdma_query_device,
501	.query_gid = erdma_query_gid,
502	.query_port = erdma_query_port,
503	.query_qp = erdma_query_qp,
504	.req_notify_cq = erdma_req_notify_cq,
505	.reg_user_mr = erdma_reg_user_mr,
506
507	INIT_RDMA_OBJ_SIZE(ib_cq, erdma_cq, ibcq),
508	INIT_RDMA_OBJ_SIZE(ib_pd, erdma_pd, ibpd),
509	INIT_RDMA_OBJ_SIZE(ib_ucontext, erdma_ucontext, ibucontext),
510	INIT_RDMA_OBJ_SIZE(ib_qp, erdma_qp, ibqp),
511};
512
513static int erdma_ib_device_add(struct pci_dev *pdev)
514{
515	struct erdma_dev *dev = pci_get_drvdata(pdev);
516	struct ib_device *ibdev = &dev->ibdev;
517	u64 mac;
518	int ret;
519
520	ret = erdma_dev_attrs_init(dev);
521	if (ret)
522		return ret;
523
524	ret = erdma_device_config(dev);
525	if (ret)
526		return ret;
527
528	ibdev->node_type = RDMA_NODE_RNIC;
529	memcpy(ibdev->node_desc, ERDMA_NODE_DESC, sizeof(ERDMA_NODE_DESC));
530
531	/*
532	 * Current model (one-to-one device association):
533	 * One ERDMA device per net_device or, equivalently,
534	 * per physical port.
535	 */
536	ibdev->phys_port_cnt = 1;
537	ibdev->num_comp_vectors = dev->attrs.irq_num - 1;
538
539	ib_set_device_ops(ibdev, &erdma_device_ops);
540
541	INIT_LIST_HEAD(&dev->cep_list);
542
543	spin_lock_init(&dev->lock);
544	xa_init_flags(&dev->qp_xa, XA_FLAGS_ALLOC1);
545	xa_init_flags(&dev->cq_xa, XA_FLAGS_ALLOC1);
546	dev->next_alloc_cqn = 1;
547	dev->next_alloc_qpn = 1;
548
549	ret = erdma_res_cb_init(dev);
550	if (ret)
551		return ret;
552
553	atomic_set(&dev->num_ctx, 0);
554
555	mac = erdma_reg_read32(dev, ERDMA_REGS_NETDEV_MAC_L_REG);
556	mac |= (u64)erdma_reg_read32(dev, ERDMA_REGS_NETDEV_MAC_H_REG) << 32;
557
558	u64_to_ether_addr(mac, dev->attrs.peer_addr);
559
560	dev->reflush_wq = alloc_workqueue("erdma-reflush-wq", WQ_UNBOUND,
561					  WQ_UNBOUND_MAX_ACTIVE);
562	if (!dev->reflush_wq) {
563		ret = -ENOMEM;
564		goto err_alloc_workqueue;
565	}
566
567	ret = erdma_device_register(dev);
568	if (ret)
569		goto err_register;
570
571	return 0;
572
573err_register:
574	destroy_workqueue(dev->reflush_wq);
575err_alloc_workqueue:
576	xa_destroy(&dev->qp_xa);
577	xa_destroy(&dev->cq_xa);
578
579	erdma_res_cb_free(dev);
580
581	return ret;
582}
583
584static void erdma_ib_device_remove(struct pci_dev *pdev)
585{
586	struct erdma_dev *dev = pci_get_drvdata(pdev);
587
588	unregister_netdevice_notifier(&dev->netdev_nb);
589	ib_unregister_device(&dev->ibdev);
590
591	destroy_workqueue(dev->reflush_wq);
592	erdma_res_cb_free(dev);
593	xa_destroy(&dev->qp_xa);
594	xa_destroy(&dev->cq_xa);
595}
596
597static int erdma_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
598{
599	int ret;
600
601	ret = erdma_probe_dev(pdev);
602	if (ret)
603		return ret;
604
605	ret = erdma_ib_device_add(pdev);
606	if (ret) {
607		erdma_remove_dev(pdev);
608		return ret;
609	}
610
611	return 0;
612}
613
614static void erdma_remove(struct pci_dev *pdev)
615{
616	erdma_ib_device_remove(pdev);
617	erdma_remove_dev(pdev);
618}
619
620static struct pci_driver erdma_pci_driver = {
621	.name = DRV_MODULE_NAME,
622	.id_table = erdma_pci_tbl,
623	.probe = erdma_probe,
624	.remove = erdma_remove
625};
626
627MODULE_DEVICE_TABLE(pci, erdma_pci_tbl);
628
629static __init int erdma_init_module(void)
630{
631	int ret;
632
633	ret = erdma_cm_init();
634	if (ret)
635		return ret;
636
637	ret = pci_register_driver(&erdma_pci_driver);
638	if (ret)
639		erdma_cm_exit();
640
641	return ret;
642}
643
644static void __exit erdma_exit_module(void)
645{
646	pci_unregister_driver(&erdma_pci_driver);
647
648	erdma_cm_exit();
649}
650
651module_init(erdma_init_module);
652module_exit(erdma_exit_module);
653