1 // SPDX-License-Identifier: GPL-2.0 OR BSD-2-Clause
2 /*
3  * Copyright 2018-2019 Amazon.com, Inc. or its affiliates. All rights reserved.
4  */
5 
6 #include <linux/module.h>
7 #include <linux/pci.h>
8 
9 #include <rdma/ib_user_verbs.h>
10 
11 #include "efa.h"
12 
13 #define PCI_DEV_ID_EFA_VF 0xefa0
14 
15 static const struct pci_device_id efa_pci_tbl[] = {
16 	{ PCI_VDEVICE(AMAZON, PCI_DEV_ID_EFA_VF) },
17 	{ }
18 };
19 
20 MODULE_AUTHOR("Amazon.com, Inc. or its affiliates");
21 MODULE_LICENSE("Dual BSD/GPL");
22 MODULE_DESCRIPTION(DEVICE_NAME);
23 MODULE_DEVICE_TABLE(pci, efa_pci_tbl);
24 
25 #define EFA_REG_BAR 0
26 #define EFA_MEM_BAR 2
27 #define EFA_BASE_BAR_MASK (BIT(EFA_REG_BAR) | BIT(EFA_MEM_BAR))
28 
29 #define EFA_AENQ_ENABLED_GROUPS \
30 	(BIT(EFA_ADMIN_FATAL_ERROR) | BIT(EFA_ADMIN_WARNING) | \
31 	 BIT(EFA_ADMIN_NOTIFICATION) | BIT(EFA_ADMIN_KEEP_ALIVE))
32 
33 static void efa_update_network_attr(struct efa_dev *dev,
34 				    struct efa_com_get_network_attr_result *network_attr)
35 {
36 	memcpy(dev->addr, network_attr->addr, sizeof(network_attr->addr));
37 	dev->mtu = network_attr->mtu;
38 
39 	dev_dbg(&dev->pdev->dev, "Full address %pI6\n", dev->addr);
40 }
41 
42 /* This handler will called for unknown event group or unimplemented handlers */
43 static void unimplemented_aenq_handler(void *data,
44 				       struct efa_admin_aenq_entry *aenq_e)
45 {
46 	struct efa_dev *dev = (struct efa_dev *)data;
47 
48 	ibdev_err(&dev->ibdev,
49 		  "Unknown event was received or event with unimplemented handler\n");
50 }
51 
52 static void efa_keep_alive(void *data, struct efa_admin_aenq_entry *aenq_e)
53 {
54 	struct efa_dev *dev = (struct efa_dev *)data;
55 
56 	atomic64_inc(&dev->stats.keep_alive_rcvd);
57 }
58 
59 static struct efa_aenq_handlers aenq_handlers = {
60 	.handlers = {
61 		[EFA_ADMIN_KEEP_ALIVE] = efa_keep_alive,
62 	},
63 	.unimplemented_handler = unimplemented_aenq_handler
64 };
65 
66 static void efa_release_bars(struct efa_dev *dev, int bars_mask)
67 {
68 	struct pci_dev *pdev = dev->pdev;
69 	int release_bars;
70 
71 	release_bars = pci_select_bars(pdev, IORESOURCE_MEM) & bars_mask;
72 	pci_release_selected_regions(pdev, release_bars);
73 }
74 
75 static irqreturn_t efa_intr_msix_mgmnt(int irq, void *data)
76 {
77 	struct efa_dev *dev = data;
78 
79 	efa_com_admin_q_comp_intr_handler(&dev->edev);
80 	efa_com_aenq_intr_handler(&dev->edev, data);
81 
82 	return IRQ_HANDLED;
83 }
84 
85 static int efa_request_mgmnt_irq(struct efa_dev *dev)
86 {
87 	struct efa_irq *irq;
88 	int err;
89 
90 	irq = &dev->admin_irq;
91 	err = request_irq(irq->vector, irq->handler, 0, irq->name,
92 			  irq->data);
93 	if (err) {
94 		dev_err(&dev->pdev->dev, "Failed to request admin irq (%d)\n",
95 			err);
96 		return err;
97 	}
98 
99 	dev_dbg(&dev->pdev->dev, "Set affinity hint of mgmnt irq to %*pbl (irq vector: %d)\n",
100 		nr_cpumask_bits, &irq->affinity_hint_mask, irq->vector);
101 	irq_set_affinity_hint(irq->vector, &irq->affinity_hint_mask);
102 
103 	return 0;
104 }
105 
106 static void efa_setup_mgmnt_irq(struct efa_dev *dev)
107 {
108 	u32 cpu;
109 
110 	snprintf(dev->admin_irq.name, EFA_IRQNAME_SIZE,
111 		 "efa-mgmnt@pci:%s", pci_name(dev->pdev));
112 	dev->admin_irq.handler = efa_intr_msix_mgmnt;
113 	dev->admin_irq.data = dev;
114 	dev->admin_irq.vector =
115 		pci_irq_vector(dev->pdev, dev->admin_msix_vector_idx);
116 	cpu = cpumask_first(cpu_online_mask);
117 	dev->admin_irq.cpu = cpu;
118 	cpumask_set_cpu(cpu,
119 			&dev->admin_irq.affinity_hint_mask);
120 	dev_info(&dev->pdev->dev, "Setup irq:0x%p vector:%d name:%s\n",
121 		 &dev->admin_irq,
122 		 dev->admin_irq.vector,
123 		 dev->admin_irq.name);
124 }
125 
126 static void efa_free_mgmnt_irq(struct efa_dev *dev)
127 {
128 	struct efa_irq *irq;
129 
130 	irq = &dev->admin_irq;
131 	irq_set_affinity_hint(irq->vector, NULL);
132 	free_irq(irq->vector, irq->data);
133 }
134 
135 static int efa_set_mgmnt_irq(struct efa_dev *dev)
136 {
137 	efa_setup_mgmnt_irq(dev);
138 
139 	return efa_request_mgmnt_irq(dev);
140 }
141 
142 static int efa_request_doorbell_bar(struct efa_dev *dev)
143 {
144 	u8 db_bar_idx = dev->dev_attr.db_bar;
145 	struct pci_dev *pdev = dev->pdev;
146 	int bars;
147 	int err;
148 
149 	if (!(BIT(db_bar_idx) & EFA_BASE_BAR_MASK)) {
150 		bars = pci_select_bars(pdev, IORESOURCE_MEM) & BIT(db_bar_idx);
151 
152 		err = pci_request_selected_regions(pdev, bars, DRV_MODULE_NAME);
153 		if (err) {
154 			dev_err(&dev->pdev->dev,
155 				"pci_request_selected_regions for bar %d failed %d\n",
156 				db_bar_idx, err);
157 			return err;
158 		}
159 	}
160 
161 	dev->db_bar_addr = pci_resource_start(dev->pdev, db_bar_idx);
162 	dev->db_bar_len = pci_resource_len(dev->pdev, db_bar_idx);
163 
164 	return 0;
165 }
166 
167 static void efa_release_doorbell_bar(struct efa_dev *dev)
168 {
169 	if (!(BIT(dev->dev_attr.db_bar) & EFA_BASE_BAR_MASK))
170 		efa_release_bars(dev, BIT(dev->dev_attr.db_bar));
171 }
172 
173 static void efa_update_hw_hints(struct efa_dev *dev,
174 				struct efa_com_get_hw_hints_result *hw_hints)
175 {
176 	struct efa_com_dev *edev = &dev->edev;
177 
178 	if (hw_hints->mmio_read_timeout)
179 		edev->mmio_read.mmio_read_timeout =
180 			hw_hints->mmio_read_timeout * 1000;
181 
182 	if (hw_hints->poll_interval)
183 		edev->aq.poll_interval = hw_hints->poll_interval;
184 
185 	if (hw_hints->admin_completion_timeout)
186 		edev->aq.completion_timeout =
187 			hw_hints->admin_completion_timeout;
188 }
189 
190 static void efa_stats_init(struct efa_dev *dev)
191 {
192 	atomic64_t *s = (atomic64_t *)&dev->stats;
193 	int i;
194 
195 	for (i = 0; i < sizeof(dev->stats) / sizeof(*s); i++, s++)
196 		atomic64_set(s, 0);
197 }
198 
199 static const struct ib_device_ops efa_dev_ops = {
200 	.owner = THIS_MODULE,
201 	.driver_id = RDMA_DRIVER_EFA,
202 	.uverbs_abi_ver = EFA_UVERBS_ABI_VERSION,
203 
204 	.alloc_pd = efa_alloc_pd,
205 	.alloc_ucontext = efa_alloc_ucontext,
206 	.create_ah = efa_create_ah,
207 	.create_cq = efa_create_cq,
208 	.create_qp = efa_create_qp,
209 	.dealloc_pd = efa_dealloc_pd,
210 	.dealloc_ucontext = efa_dealloc_ucontext,
211 	.dereg_mr = efa_dereg_mr,
212 	.destroy_ah = efa_destroy_ah,
213 	.destroy_cq = efa_destroy_cq,
214 	.destroy_qp = efa_destroy_qp,
215 	.get_link_layer = efa_port_link_layer,
216 	.get_port_immutable = efa_get_port_immutable,
217 	.mmap = efa_mmap,
218 	.modify_qp = efa_modify_qp,
219 	.query_device = efa_query_device,
220 	.query_gid = efa_query_gid,
221 	.query_pkey = efa_query_pkey,
222 	.query_port = efa_query_port,
223 	.query_qp = efa_query_qp,
224 	.reg_user_mr = efa_reg_mr,
225 
226 	INIT_RDMA_OBJ_SIZE(ib_ah, efa_ah, ibah),
227 	INIT_RDMA_OBJ_SIZE(ib_cq, efa_cq, ibcq),
228 	INIT_RDMA_OBJ_SIZE(ib_pd, efa_pd, ibpd),
229 	INIT_RDMA_OBJ_SIZE(ib_ucontext, efa_ucontext, ibucontext),
230 };
231 
232 static int efa_ib_device_add(struct efa_dev *dev)
233 {
234 	struct efa_com_get_network_attr_result network_attr;
235 	struct efa_com_get_hw_hints_result hw_hints;
236 	struct pci_dev *pdev = dev->pdev;
237 	int err;
238 
239 	efa_stats_init(dev);
240 
241 	err = efa_com_get_device_attr(&dev->edev, &dev->dev_attr);
242 	if (err)
243 		return err;
244 
245 	dev_dbg(&dev->pdev->dev, "Doorbells bar (%d)\n", dev->dev_attr.db_bar);
246 	err = efa_request_doorbell_bar(dev);
247 	if (err)
248 		return err;
249 
250 	err = efa_com_get_network_attr(&dev->edev, &network_attr);
251 	if (err)
252 		goto err_release_doorbell_bar;
253 
254 	efa_update_network_attr(dev, &network_attr);
255 
256 	err = efa_com_get_hw_hints(&dev->edev, &hw_hints);
257 	if (err)
258 		goto err_release_doorbell_bar;
259 
260 	efa_update_hw_hints(dev, &hw_hints);
261 
262 	/* Try to enable all the available aenq groups */
263 	err = efa_com_set_aenq_config(&dev->edev, EFA_AENQ_ENABLED_GROUPS);
264 	if (err)
265 		goto err_release_doorbell_bar;
266 
267 	dev->ibdev.node_type = RDMA_NODE_UNSPECIFIED;
268 	dev->ibdev.phys_port_cnt = 1;
269 	dev->ibdev.num_comp_vectors = 1;
270 	dev->ibdev.dev.parent = &pdev->dev;
271 
272 	dev->ibdev.uverbs_cmd_mask =
273 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT) |
274 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE) |
275 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT) |
276 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD) |
277 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD) |
278 		(1ull << IB_USER_VERBS_CMD_REG_MR) |
279 		(1ull << IB_USER_VERBS_CMD_DEREG_MR) |
280 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
281 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ) |
282 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ) |
283 		(1ull << IB_USER_VERBS_CMD_CREATE_QP) |
284 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP) |
285 		(1ull << IB_USER_VERBS_CMD_QUERY_QP) |
286 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP) |
287 		(1ull << IB_USER_VERBS_CMD_CREATE_AH) |
288 		(1ull << IB_USER_VERBS_CMD_DESTROY_AH);
289 
290 	dev->ibdev.uverbs_ex_cmd_mask =
291 		(1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE);
292 
293 	ib_set_device_ops(&dev->ibdev, &efa_dev_ops);
294 
295 	err = ib_register_device(&dev->ibdev, "efa_%d");
296 	if (err)
297 		goto err_release_doorbell_bar;
298 
299 	ibdev_info(&dev->ibdev, "IB device registered\n");
300 
301 	return 0;
302 
303 err_release_doorbell_bar:
304 	efa_release_doorbell_bar(dev);
305 	return err;
306 }
307 
308 static void efa_ib_device_remove(struct efa_dev *dev)
309 {
310 	efa_com_dev_reset(&dev->edev, EFA_REGS_RESET_NORMAL);
311 	ibdev_info(&dev->ibdev, "Unregister ib device\n");
312 	ib_unregister_device(&dev->ibdev);
313 	efa_release_doorbell_bar(dev);
314 }
315 
316 static void efa_disable_msix(struct efa_dev *dev)
317 {
318 	pci_free_irq_vectors(dev->pdev);
319 }
320 
321 static int efa_enable_msix(struct efa_dev *dev)
322 {
323 	int msix_vecs, irq_num;
324 
325 	/* Reserve the max msix vectors we might need */
326 	msix_vecs = EFA_NUM_MSIX_VEC;
327 	dev_dbg(&dev->pdev->dev, "Trying to enable MSI-X, vectors %d\n",
328 		msix_vecs);
329 
330 	dev->admin_msix_vector_idx = EFA_MGMNT_MSIX_VEC_IDX;
331 	irq_num = pci_alloc_irq_vectors(dev->pdev, msix_vecs,
332 					msix_vecs, PCI_IRQ_MSIX);
333 
334 	if (irq_num < 0) {
335 		dev_err(&dev->pdev->dev, "Failed to enable MSI-X. irq_num %d\n",
336 			irq_num);
337 		return -ENOSPC;
338 	}
339 
340 	if (irq_num != msix_vecs) {
341 		dev_err(&dev->pdev->dev,
342 			"Allocated %d MSI-X (out of %d requested)\n",
343 			irq_num, msix_vecs);
344 		return -ENOSPC;
345 	}
346 
347 	return 0;
348 }
349 
350 static int efa_device_init(struct efa_com_dev *edev, struct pci_dev *pdev)
351 {
352 	int dma_width;
353 	int err;
354 
355 	err = efa_com_dev_reset(edev, EFA_REGS_RESET_NORMAL);
356 	if (err)
357 		return err;
358 
359 	err = efa_com_validate_version(edev);
360 	if (err)
361 		return err;
362 
363 	dma_width = efa_com_get_dma_width(edev);
364 	if (dma_width < 0) {
365 		err = dma_width;
366 		return err;
367 	}
368 
369 	err = pci_set_dma_mask(pdev, DMA_BIT_MASK(dma_width));
370 	if (err) {
371 		dev_err(&pdev->dev, "pci_set_dma_mask failed %d\n", err);
372 		return err;
373 	}
374 
375 	err = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(dma_width));
376 	if (err) {
377 		dev_err(&pdev->dev,
378 			"err_pci_set_consistent_dma_mask failed %d\n",
379 			err);
380 		return err;
381 	}
382 
383 	return 0;
384 }
385 
386 static struct efa_dev *efa_probe_device(struct pci_dev *pdev)
387 {
388 	struct efa_com_dev *edev;
389 	struct efa_dev *dev;
390 	int bars;
391 	int err;
392 
393 	err = pci_enable_device_mem(pdev);
394 	if (err) {
395 		dev_err(&pdev->dev, "pci_enable_device_mem() failed!\n");
396 		return ERR_PTR(err);
397 	}
398 
399 	pci_set_master(pdev);
400 
401 	dev = ib_alloc_device(efa_dev, ibdev);
402 	if (!dev) {
403 		dev_err(&pdev->dev, "Device alloc failed\n");
404 		err = -ENOMEM;
405 		goto err_disable_device;
406 	}
407 
408 	pci_set_drvdata(pdev, dev);
409 	edev = &dev->edev;
410 	edev->efa_dev = dev;
411 	edev->dmadev = &pdev->dev;
412 	dev->pdev = pdev;
413 
414 	bars = pci_select_bars(pdev, IORESOURCE_MEM) & EFA_BASE_BAR_MASK;
415 	err = pci_request_selected_regions(pdev, bars, DRV_MODULE_NAME);
416 	if (err) {
417 		dev_err(&pdev->dev, "pci_request_selected_regions failed %d\n",
418 			err);
419 		goto err_ibdev_destroy;
420 	}
421 
422 	dev->reg_bar_addr = pci_resource_start(pdev, EFA_REG_BAR);
423 	dev->reg_bar_len = pci_resource_len(pdev, EFA_REG_BAR);
424 	dev->mem_bar_addr = pci_resource_start(pdev, EFA_MEM_BAR);
425 	dev->mem_bar_len = pci_resource_len(pdev, EFA_MEM_BAR);
426 
427 	edev->reg_bar = devm_ioremap(&pdev->dev,
428 				     dev->reg_bar_addr,
429 				     dev->reg_bar_len);
430 	if (!edev->reg_bar) {
431 		dev_err(&pdev->dev, "Failed to remap register bar\n");
432 		err = -EFAULT;
433 		goto err_release_bars;
434 	}
435 
436 	err = efa_com_mmio_reg_read_init(edev);
437 	if (err) {
438 		dev_err(&pdev->dev, "Failed to init readless MMIO\n");
439 		goto err_iounmap;
440 	}
441 
442 	err = efa_device_init(edev, pdev);
443 	if (err) {
444 		dev_err(&pdev->dev, "EFA device init failed\n");
445 		if (err == -ETIME)
446 			err = -EPROBE_DEFER;
447 		goto err_reg_read_destroy;
448 	}
449 
450 	err = efa_enable_msix(dev);
451 	if (err)
452 		goto err_reg_read_destroy;
453 
454 	edev->aq.msix_vector_idx = dev->admin_msix_vector_idx;
455 	edev->aenq.msix_vector_idx = dev->admin_msix_vector_idx;
456 
457 	err = efa_set_mgmnt_irq(dev);
458 	if (err)
459 		goto err_disable_msix;
460 
461 	err = efa_com_admin_init(edev, &aenq_handlers);
462 	if (err)
463 		goto err_free_mgmnt_irq;
464 
465 	return dev;
466 
467 err_free_mgmnt_irq:
468 	efa_free_mgmnt_irq(dev);
469 err_disable_msix:
470 	efa_disable_msix(dev);
471 err_reg_read_destroy:
472 	efa_com_mmio_reg_read_destroy(edev);
473 err_iounmap:
474 	devm_iounmap(&pdev->dev, edev->reg_bar);
475 err_release_bars:
476 	efa_release_bars(dev, EFA_BASE_BAR_MASK);
477 err_ibdev_destroy:
478 	ib_dealloc_device(&dev->ibdev);
479 err_disable_device:
480 	pci_disable_device(pdev);
481 	return ERR_PTR(err);
482 }
483 
484 static void efa_remove_device(struct pci_dev *pdev)
485 {
486 	struct efa_dev *dev = pci_get_drvdata(pdev);
487 	struct efa_com_dev *edev;
488 
489 	edev = &dev->edev;
490 	efa_com_admin_destroy(edev);
491 	efa_free_mgmnt_irq(dev);
492 	efa_disable_msix(dev);
493 	efa_com_mmio_reg_read_destroy(edev);
494 	devm_iounmap(&pdev->dev, edev->reg_bar);
495 	efa_release_bars(dev, EFA_BASE_BAR_MASK);
496 	ib_dealloc_device(&dev->ibdev);
497 	pci_disable_device(pdev);
498 }
499 
500 static int efa_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
501 {
502 	struct efa_dev *dev;
503 	int err;
504 
505 	dev = efa_probe_device(pdev);
506 	if (IS_ERR(dev))
507 		return PTR_ERR(dev);
508 
509 	err = efa_ib_device_add(dev);
510 	if (err)
511 		goto err_remove_device;
512 
513 	return 0;
514 
515 err_remove_device:
516 	efa_remove_device(pdev);
517 	return err;
518 }
519 
520 static void efa_remove(struct pci_dev *pdev)
521 {
522 	struct efa_dev *dev = pci_get_drvdata(pdev);
523 
524 	efa_ib_device_remove(dev);
525 	efa_remove_device(pdev);
526 }
527 
528 static struct pci_driver efa_pci_driver = {
529 	.name           = DRV_MODULE_NAME,
530 	.id_table       = efa_pci_tbl,
531 	.probe          = efa_probe,
532 	.remove         = efa_remove,
533 };
534 
535 module_pci_driver(efa_pci_driver);
536