1 // SPDX-License-Identifier: GPL-2.0 OR BSD-2-Clause
3 * Copyright 2018-2024 Amazon.com, Inc. or its affiliates. All rights reserved.
6 #include <linux/module.h>
8 #include <linux/utsname.h>
9 #include <linux/version.h>
11 #include <rdma/ib_user_verbs.h>
12 #include <rdma/uverbs_ioctl.h>
16 #define PCI_DEV_ID_EFA0_VF 0xefa0
17 #define PCI_DEV_ID_EFA1_VF 0xefa1
18 #define PCI_DEV_ID_EFA2_VF 0xefa2
20 static const struct pci_device_id efa_pci_tbl[] = {
21 { PCI_VDEVICE(AMAZON, PCI_DEV_ID_EFA0_VF) },
22 { PCI_VDEVICE(AMAZON, PCI_DEV_ID_EFA1_VF) },
23 { PCI_VDEVICE(AMAZON, PCI_DEV_ID_EFA2_VF) },
27 MODULE_AUTHOR("Amazon.com, Inc. or its affiliates");
28 MODULE_LICENSE("Dual BSD/GPL");
29 MODULE_DESCRIPTION(DEVICE_NAME);
30 MODULE_DEVICE_TABLE(pci, efa_pci_tbl);
34 #define EFA_BASE_BAR_MASK (BIT(EFA_REG_BAR) | BIT(EFA_MEM_BAR))
36 #define EFA_AENQ_ENABLED_GROUPS \
37 (BIT(EFA_ADMIN_FATAL_ERROR) | BIT(EFA_ADMIN_WARNING) | \
38 BIT(EFA_ADMIN_NOTIFICATION) | BIT(EFA_ADMIN_KEEP_ALIVE))
40 extern const struct uapi_definition efa_uapi_defs[];
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)
46 struct efa_dev *dev = (struct efa_dev *)data;
48 ibdev_err(&dev->ibdev,
49 "Unknown event was received or event with unimplemented handler\n");
52 static void efa_keep_alive(void *data, struct efa_admin_aenq_entry *aenq_e)
54 struct efa_dev *dev = (struct efa_dev *)data;
56 atomic64_inc(&dev->stats.keep_alive_rcvd);
59 static struct efa_aenq_handlers aenq_handlers = {
61 [EFA_ADMIN_KEEP_ALIVE] = efa_keep_alive,
63 .unimplemented_handler = unimplemented_aenq_handler
66 static void efa_release_bars(struct efa_dev *dev, int bars_mask)
68 struct pci_dev *pdev = dev->pdev;
71 release_bars = pci_select_bars(pdev, IORESOURCE_MEM) & bars_mask;
72 pci_release_selected_regions(pdev, release_bars);
75 static void efa_process_comp_eqe(struct efa_dev *dev, struct efa_admin_eqe *eqe)
77 u16 cqn = eqe->u.comp_event.cqn;
80 /* Safe to load as we're in irq and removal calls synchronize_irq() */
81 cq = xa_load(&dev->cqs_xa, cqn);
83 ibdev_err_ratelimited(&dev->ibdev,
84 "Completion event on non-existent CQ[%u]",
89 cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
92 static void efa_process_eqe(struct efa_com_eq *eeq, struct efa_admin_eqe *eqe)
94 struct efa_dev *dev = container_of(eeq->edev, struct efa_dev, edev);
96 if (likely(EFA_GET(&eqe->common, EFA_ADMIN_EQE_EVENT_TYPE) ==
97 EFA_ADMIN_EQE_EVENT_TYPE_COMPLETION))
98 efa_process_comp_eqe(dev, eqe);
100 ibdev_err_ratelimited(&dev->ibdev,
101 "Unknown event type received %lu",
102 EFA_GET(&eqe->common,
103 EFA_ADMIN_EQE_EVENT_TYPE));
106 static irqreturn_t efa_intr_msix_comp(int irq, void *data)
108 struct efa_eq *eq = data;
109 struct efa_com_dev *edev = eq->eeq.edev;
111 efa_com_eq_comp_intr_handler(edev, &eq->eeq);
116 static irqreturn_t efa_intr_msix_mgmnt(int irq, void *data)
118 struct efa_dev *dev = data;
120 efa_com_admin_q_comp_intr_handler(&dev->edev);
121 efa_com_aenq_intr_handler(&dev->edev, data);
126 static int efa_request_irq(struct efa_dev *dev, struct efa_irq *irq)
130 err = request_irq(irq->irqn, irq->handler, 0, irq->name, irq->data);
132 dev_err(&dev->pdev->dev, "Failed to request irq %s (%d)\n",
137 irq_set_affinity_hint(irq->irqn, &irq->affinity_hint_mask);
142 static void efa_setup_comp_irq(struct efa_dev *dev, struct efa_eq *eq,
147 cpu = vector - EFA_COMP_EQS_VEC_BASE;
148 snprintf(eq->irq.name, EFA_IRQNAME_SIZE, "efa-comp%d@pci:%s", cpu,
149 pci_name(dev->pdev));
150 eq->irq.handler = efa_intr_msix_comp;
152 eq->irq.vector = vector;
153 eq->irq.irqn = pci_irq_vector(dev->pdev, vector);
154 cpumask_set_cpu(cpu, &eq->irq.affinity_hint_mask);
157 static void efa_free_irq(struct efa_dev *dev, struct efa_irq *irq)
159 irq_set_affinity_hint(irq->irqn, NULL);
160 free_irq(irq->irqn, irq->data);
163 static void efa_setup_mgmnt_irq(struct efa_dev *dev)
167 snprintf(dev->admin_irq.name, EFA_IRQNAME_SIZE,
168 "efa-mgmnt@pci:%s", pci_name(dev->pdev));
169 dev->admin_irq.handler = efa_intr_msix_mgmnt;
170 dev->admin_irq.data = dev;
171 dev->admin_irq.vector = dev->admin_msix_vector_idx;
172 dev->admin_irq.irqn = pci_irq_vector(dev->pdev,
173 dev->admin_msix_vector_idx);
174 cpu = cpumask_first(cpu_online_mask);
176 &dev->admin_irq.affinity_hint_mask);
177 dev_info(&dev->pdev->dev, "Setup irq:%d name:%s\n",
179 dev->admin_irq.name);
182 static int efa_set_mgmnt_irq(struct efa_dev *dev)
184 efa_setup_mgmnt_irq(dev);
186 return efa_request_irq(dev, &dev->admin_irq);
189 static int efa_request_doorbell_bar(struct efa_dev *dev)
191 u8 db_bar_idx = dev->dev_attr.db_bar;
192 struct pci_dev *pdev = dev->pdev;
196 if (!(BIT(db_bar_idx) & EFA_BASE_BAR_MASK)) {
197 bars = pci_select_bars(pdev, IORESOURCE_MEM) & BIT(db_bar_idx);
199 err = pci_request_selected_regions(pdev, bars, DRV_MODULE_NAME);
201 dev_err(&dev->pdev->dev,
202 "pci_request_selected_regions for bar %d failed %d\n",
208 dev->db_bar_addr = pci_resource_start(dev->pdev, db_bar_idx);
209 dev->db_bar_len = pci_resource_len(dev->pdev, db_bar_idx);
214 static void efa_release_doorbell_bar(struct efa_dev *dev)
216 if (!(BIT(dev->dev_attr.db_bar) & EFA_BASE_BAR_MASK))
217 efa_release_bars(dev, BIT(dev->dev_attr.db_bar));
220 static void efa_update_hw_hints(struct efa_dev *dev,
221 struct efa_com_get_hw_hints_result *hw_hints)
223 struct efa_com_dev *edev = &dev->edev;
225 if (hw_hints->mmio_read_timeout)
226 edev->mmio_read.mmio_read_timeout =
227 hw_hints->mmio_read_timeout * 1000;
229 if (hw_hints->poll_interval)
230 edev->aq.poll_interval = hw_hints->poll_interval;
232 if (hw_hints->admin_completion_timeout)
233 edev->aq.completion_timeout =
234 hw_hints->admin_completion_timeout;
237 static void efa_stats_init(struct efa_dev *dev)
239 atomic64_t *s = (atomic64_t *)&dev->stats;
242 for (i = 0; i < sizeof(dev->stats) / sizeof(*s); i++, s++)
246 static void efa_set_host_info(struct efa_dev *dev)
248 struct efa_admin_set_feature_resp resp = {};
249 struct efa_admin_set_feature_cmd cmd = {};
250 struct efa_admin_host_info *hinf;
251 u32 bufsz = sizeof(*hinf);
254 if (!efa_com_check_supported_feature_id(&dev->edev,
255 EFA_ADMIN_HOST_INFO))
258 /* Failures in host info set shall not disturb probe */
259 hinf = dma_alloc_coherent(&dev->pdev->dev, bufsz, &hinf_dma,
264 strscpy(hinf->os_dist_str, utsname()->release,
265 sizeof(hinf->os_dist_str));
266 hinf->os_type = EFA_ADMIN_OS_LINUX;
267 strscpy(hinf->kernel_ver_str, utsname()->version,
268 sizeof(hinf->kernel_ver_str));
269 hinf->kernel_ver = LINUX_VERSION_CODE;
270 EFA_SET(&hinf->driver_ver, EFA_ADMIN_HOST_INFO_DRIVER_MAJOR, 0);
271 EFA_SET(&hinf->driver_ver, EFA_ADMIN_HOST_INFO_DRIVER_MINOR, 0);
272 EFA_SET(&hinf->driver_ver, EFA_ADMIN_HOST_INFO_DRIVER_SUB_MINOR, 0);
273 EFA_SET(&hinf->driver_ver, EFA_ADMIN_HOST_INFO_DRIVER_MODULE_TYPE, 0);
274 EFA_SET(&hinf->bdf, EFA_ADMIN_HOST_INFO_BUS, dev->pdev->bus->number);
275 EFA_SET(&hinf->bdf, EFA_ADMIN_HOST_INFO_DEVICE,
276 PCI_SLOT(dev->pdev->devfn));
277 EFA_SET(&hinf->bdf, EFA_ADMIN_HOST_INFO_FUNCTION,
278 PCI_FUNC(dev->pdev->devfn));
279 EFA_SET(&hinf->spec_ver, EFA_ADMIN_HOST_INFO_SPEC_MAJOR,
280 EFA_COMMON_SPEC_VERSION_MAJOR);
281 EFA_SET(&hinf->spec_ver, EFA_ADMIN_HOST_INFO_SPEC_MINOR,
282 EFA_COMMON_SPEC_VERSION_MINOR);
283 EFA_SET(&hinf->flags, EFA_ADMIN_HOST_INFO_INTREE, 1);
284 EFA_SET(&hinf->flags, EFA_ADMIN_HOST_INFO_GDR, 0);
286 efa_com_set_feature_ex(&dev->edev, &resp, &cmd, EFA_ADMIN_HOST_INFO,
289 dma_free_coherent(&dev->pdev->dev, bufsz, hinf, hinf_dma);
292 static void efa_destroy_eq(struct efa_dev *dev, struct efa_eq *eq)
294 efa_com_eq_destroy(&dev->edev, &eq->eeq);
295 efa_free_irq(dev, &eq->irq);
298 static int efa_create_eq(struct efa_dev *dev, struct efa_eq *eq, u8 msix_vec)
302 efa_setup_comp_irq(dev, eq, msix_vec);
303 err = efa_request_irq(dev, &eq->irq);
307 err = efa_com_eq_init(&dev->edev, &eq->eeq, efa_process_eqe,
308 dev->dev_attr.max_eq_depth, msix_vec);
310 goto err_free_comp_irq;
315 efa_free_irq(dev, &eq->irq);
319 static int efa_create_eqs(struct efa_dev *dev)
321 unsigned int neqs = dev->dev_attr.max_eq;
325 neqs = min_t(unsigned int, neqs,
326 dev->num_irq_vectors - EFA_COMP_EQS_VEC_BASE);
329 dev->eqs = kcalloc(neqs, sizeof(*dev->eqs), GFP_KERNEL);
333 for (i = 0; i < neqs; i++) {
334 err = efa_create_eq(dev, &dev->eqs[i],
335 i + EFA_COMP_EQS_VEC_BASE);
337 goto err_destroy_eqs;
343 for (i--; i >= 0; i--)
344 efa_destroy_eq(dev, &dev->eqs[i]);
350 static void efa_destroy_eqs(struct efa_dev *dev)
354 for (i = 0; i < dev->neqs; i++)
355 efa_destroy_eq(dev, &dev->eqs[i]);
360 static const struct ib_device_ops efa_dev_ops = {
361 .owner = THIS_MODULE,
362 .driver_id = RDMA_DRIVER_EFA,
363 .uverbs_abi_ver = EFA_UVERBS_ABI_VERSION,
365 .alloc_hw_port_stats = efa_alloc_hw_port_stats,
366 .alloc_hw_device_stats = efa_alloc_hw_device_stats,
367 .alloc_pd = efa_alloc_pd,
368 .alloc_ucontext = efa_alloc_ucontext,
369 .create_cq = efa_create_cq,
370 .create_qp = efa_create_qp,
371 .create_user_ah = efa_create_ah,
372 .dealloc_pd = efa_dealloc_pd,
373 .dealloc_ucontext = efa_dealloc_ucontext,
374 .dereg_mr = efa_dereg_mr,
375 .destroy_ah = efa_destroy_ah,
376 .destroy_cq = efa_destroy_cq,
377 .destroy_qp = efa_destroy_qp,
378 .get_hw_stats = efa_get_hw_stats,
379 .get_link_layer = efa_port_link_layer,
380 .get_port_immutable = efa_get_port_immutable,
382 .mmap_free = efa_mmap_free,
383 .modify_qp = efa_modify_qp,
384 .query_device = efa_query_device,
385 .query_gid = efa_query_gid,
386 .query_pkey = efa_query_pkey,
387 .query_port = efa_query_port,
388 .query_qp = efa_query_qp,
389 .reg_user_mr = efa_reg_mr,
390 .reg_user_mr_dmabuf = efa_reg_user_mr_dmabuf,
392 INIT_RDMA_OBJ_SIZE(ib_ah, efa_ah, ibah),
393 INIT_RDMA_OBJ_SIZE(ib_cq, efa_cq, ibcq),
394 INIT_RDMA_OBJ_SIZE(ib_pd, efa_pd, ibpd),
395 INIT_RDMA_OBJ_SIZE(ib_qp, efa_qp, ibqp),
396 INIT_RDMA_OBJ_SIZE(ib_ucontext, efa_ucontext, ibucontext),
399 static int efa_ib_device_add(struct efa_dev *dev)
401 struct efa_com_get_hw_hints_result hw_hints;
402 struct pci_dev *pdev = dev->pdev;
407 err = efa_com_get_device_attr(&dev->edev, &dev->dev_attr);
411 dev_dbg(&dev->pdev->dev, "Doorbells bar (%d)\n", dev->dev_attr.db_bar);
412 err = efa_request_doorbell_bar(dev);
416 err = efa_com_get_hw_hints(&dev->edev, &hw_hints);
418 goto err_release_doorbell_bar;
420 efa_update_hw_hints(dev, &hw_hints);
422 /* Try to enable all the available aenq groups */
423 err = efa_com_set_aenq_config(&dev->edev, EFA_AENQ_ENABLED_GROUPS);
425 goto err_release_doorbell_bar;
427 err = efa_create_eqs(dev);
429 goto err_release_doorbell_bar;
431 efa_set_host_info(dev);
433 dev->ibdev.node_type = RDMA_NODE_UNSPECIFIED;
434 dev->ibdev.phys_port_cnt = 1;
435 dev->ibdev.num_comp_vectors = dev->neqs ?: 1;
436 dev->ibdev.dev.parent = &pdev->dev;
438 ib_set_device_ops(&dev->ibdev, &efa_dev_ops);
440 dev->ibdev.driver_def = efa_uapi_defs;
442 err = ib_register_device(&dev->ibdev, "efa_%d", &pdev->dev);
444 goto err_destroy_eqs;
446 ibdev_info(&dev->ibdev, "IB device registered\n");
451 efa_destroy_eqs(dev);
452 err_release_doorbell_bar:
453 efa_release_doorbell_bar(dev);
457 static void efa_ib_device_remove(struct efa_dev *dev)
459 ibdev_info(&dev->ibdev, "Unregister ib device\n");
460 ib_unregister_device(&dev->ibdev);
461 efa_destroy_eqs(dev);
462 efa_com_dev_reset(&dev->edev, EFA_REGS_RESET_NORMAL);
463 efa_release_doorbell_bar(dev);
466 static void efa_disable_msix(struct efa_dev *dev)
468 pci_free_irq_vectors(dev->pdev);
471 static int efa_enable_msix(struct efa_dev *dev)
473 int max_vecs, num_vecs;
476 * Reserve the max msix vectors we might need, one vector is reserved
479 max_vecs = min_t(int, pci_msix_vec_count(dev->pdev),
480 num_online_cpus() + 1);
481 dev_dbg(&dev->pdev->dev, "Trying to enable MSI-X, vectors %d\n",
484 dev->admin_msix_vector_idx = EFA_MGMNT_MSIX_VEC_IDX;
485 num_vecs = pci_alloc_irq_vectors(dev->pdev, 1,
486 max_vecs, PCI_IRQ_MSIX);
489 dev_err(&dev->pdev->dev, "Failed to enable MSI-X. error %d\n",
494 dev_dbg(&dev->pdev->dev, "Allocated %d MSI-X vectors\n", num_vecs);
496 dev->num_irq_vectors = num_vecs;
501 static int efa_device_init(struct efa_com_dev *edev, struct pci_dev *pdev)
506 err = efa_com_dev_reset(edev, EFA_REGS_RESET_NORMAL);
510 err = efa_com_validate_version(edev);
514 dma_width = efa_com_get_dma_width(edev);
520 err = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(dma_width));
522 dev_err(&pdev->dev, "dma_set_mask_and_coherent failed %d\n", err);
526 dma_set_max_seg_size(&pdev->dev, UINT_MAX);
530 static struct efa_dev *efa_probe_device(struct pci_dev *pdev)
532 struct efa_com_dev *edev;
537 err = pci_enable_device_mem(pdev);
539 dev_err(&pdev->dev, "pci_enable_device_mem() failed!\n");
543 pci_set_master(pdev);
545 dev = ib_alloc_device(efa_dev, ibdev);
547 dev_err(&pdev->dev, "Device alloc failed\n");
549 goto err_disable_device;
552 pci_set_drvdata(pdev, dev);
555 edev->dmadev = &pdev->dev;
557 xa_init(&dev->cqs_xa);
559 bars = pci_select_bars(pdev, IORESOURCE_MEM) & EFA_BASE_BAR_MASK;
560 err = pci_request_selected_regions(pdev, bars, DRV_MODULE_NAME);
562 dev_err(&pdev->dev, "pci_request_selected_regions failed %d\n",
564 goto err_ibdev_destroy;
567 dev->reg_bar_addr = pci_resource_start(pdev, EFA_REG_BAR);
568 dev->reg_bar_len = pci_resource_len(pdev, EFA_REG_BAR);
569 dev->mem_bar_addr = pci_resource_start(pdev, EFA_MEM_BAR);
570 dev->mem_bar_len = pci_resource_len(pdev, EFA_MEM_BAR);
572 edev->reg_bar = devm_ioremap(&pdev->dev,
575 if (!edev->reg_bar) {
576 dev_err(&pdev->dev, "Failed to remap register bar\n");
578 goto err_release_bars;
581 err = efa_com_mmio_reg_read_init(edev);
583 dev_err(&pdev->dev, "Failed to init readless MMIO\n");
587 err = efa_device_init(edev, pdev);
589 dev_err(&pdev->dev, "EFA device init failed\n");
592 goto err_reg_read_destroy;
595 err = efa_enable_msix(dev);
597 goto err_reg_read_destroy;
599 edev->aq.msix_vector_idx = dev->admin_msix_vector_idx;
600 edev->aenq.msix_vector_idx = dev->admin_msix_vector_idx;
602 err = efa_set_mgmnt_irq(dev);
604 goto err_disable_msix;
606 err = efa_com_admin_init(edev, &aenq_handlers);
608 goto err_free_mgmnt_irq;
613 efa_free_irq(dev, &dev->admin_irq);
615 efa_disable_msix(dev);
616 err_reg_read_destroy:
617 efa_com_mmio_reg_read_destroy(edev);
619 devm_iounmap(&pdev->dev, edev->reg_bar);
621 efa_release_bars(dev, EFA_BASE_BAR_MASK);
623 ib_dealloc_device(&dev->ibdev);
625 pci_disable_device(pdev);
629 static void efa_remove_device(struct pci_dev *pdev)
631 struct efa_dev *dev = pci_get_drvdata(pdev);
632 struct efa_com_dev *edev;
635 efa_com_admin_destroy(edev);
636 efa_free_irq(dev, &dev->admin_irq);
637 efa_disable_msix(dev);
638 efa_com_mmio_reg_read_destroy(edev);
639 devm_iounmap(&pdev->dev, edev->reg_bar);
640 efa_release_bars(dev, EFA_BASE_BAR_MASK);
641 xa_destroy(&dev->cqs_xa);
642 ib_dealloc_device(&dev->ibdev);
643 pci_disable_device(pdev);
646 static int efa_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
651 dev = efa_probe_device(pdev);
655 err = efa_ib_device_add(dev);
657 goto err_remove_device;
662 efa_remove_device(pdev);
666 static void efa_remove(struct pci_dev *pdev)
668 struct efa_dev *dev = pci_get_drvdata(pdev);
670 efa_ib_device_remove(dev);
671 efa_remove_device(pdev);
674 static struct pci_driver efa_pci_driver = {
675 .name = DRV_MODULE_NAME,
676 .id_table = efa_pci_tbl,
678 .remove = efa_remove,
681 module_pci_driver(efa_pci_driver);