xref: /openbmc/linux/drivers/dma/idxd/init.c (revision 04c53de5)
1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright(c) 2019 Intel Corporation. All rights rsvd. */
3 #include <linux/init.h>
4 #include <linux/kernel.h>
5 #include <linux/module.h>
6 #include <linux/slab.h>
7 #include <linux/pci.h>
8 #include <linux/interrupt.h>
9 #include <linux/delay.h>
10 #include <linux/dma-mapping.h>
11 #include <linux/workqueue.h>
12 #include <linux/aer.h>
13 #include <linux/fs.h>
14 #include <linux/io-64-nonatomic-lo-hi.h>
15 #include <linux/device.h>
16 #include <linux/idr.h>
17 #include <linux/intel-svm.h>
18 #include <linux/iommu.h>
19 #include <uapi/linux/idxd.h>
20 #include <linux/dmaengine.h>
21 #include "../dmaengine.h"
22 #include "registers.h"
23 #include "idxd.h"
24 
25 MODULE_VERSION(IDXD_DRIVER_VERSION);
26 MODULE_LICENSE("GPL v2");
27 MODULE_AUTHOR("Intel Corporation");
28 
29 static bool sva = true;
30 module_param(sva, bool, 0644);
31 MODULE_PARM_DESC(sva, "Toggle SVA support on/off");
32 
33 #define DRV_NAME "idxd"
34 
35 bool support_enqcmd;
36 
37 static struct idr idxd_idrs[IDXD_TYPE_MAX];
38 static DEFINE_MUTEX(idxd_idr_lock);
39 
40 static struct pci_device_id idxd_pci_tbl[] = {
41 	/* DSA ver 1.0 platforms */
42 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_DSA_SPR0) },
43 
44 	/* IAX ver 1.0 platforms */
45 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IAX_SPR0) },
46 	{ 0, }
47 };
48 MODULE_DEVICE_TABLE(pci, idxd_pci_tbl);
49 
50 static char *idxd_name[] = {
51 	"dsa",
52 	"iax"
53 };
54 
55 const char *idxd_get_dev_name(struct idxd_device *idxd)
56 {
57 	return idxd_name[idxd->type];
58 }
59 
60 static int idxd_setup_interrupts(struct idxd_device *idxd)
61 {
62 	struct pci_dev *pdev = idxd->pdev;
63 	struct device *dev = &pdev->dev;
64 	struct msix_entry *msix;
65 	struct idxd_irq_entry *irq_entry;
66 	int i, msixcnt;
67 	int rc = 0;
68 
69 	msixcnt = pci_msix_vec_count(pdev);
70 	if (msixcnt < 0) {
71 		dev_err(dev, "Not MSI-X interrupt capable.\n");
72 		goto err_no_irq;
73 	}
74 
75 	idxd->msix_entries = devm_kzalloc(dev, sizeof(struct msix_entry) *
76 			msixcnt, GFP_KERNEL);
77 	if (!idxd->msix_entries) {
78 		rc = -ENOMEM;
79 		goto err_no_irq;
80 	}
81 
82 	for (i = 0; i < msixcnt; i++)
83 		idxd->msix_entries[i].entry = i;
84 
85 	rc = pci_enable_msix_exact(pdev, idxd->msix_entries, msixcnt);
86 	if (rc) {
87 		dev_err(dev, "Failed enabling %d MSIX entries.\n", msixcnt);
88 		goto err_no_irq;
89 	}
90 	dev_dbg(dev, "Enabled %d msix vectors\n", msixcnt);
91 
92 	/*
93 	 * We implement 1 completion list per MSI-X entry except for
94 	 * entry 0, which is for errors and others.
95 	 */
96 	idxd->irq_entries = devm_kcalloc(dev, msixcnt,
97 					 sizeof(struct idxd_irq_entry),
98 					 GFP_KERNEL);
99 	if (!idxd->irq_entries) {
100 		rc = -ENOMEM;
101 		goto err_no_irq;
102 	}
103 
104 	for (i = 0; i < msixcnt; i++) {
105 		idxd->irq_entries[i].id = i;
106 		idxd->irq_entries[i].idxd = idxd;
107 		spin_lock_init(&idxd->irq_entries[i].list_lock);
108 	}
109 
110 	msix = &idxd->msix_entries[0];
111 	irq_entry = &idxd->irq_entries[0];
112 	rc = devm_request_threaded_irq(dev, msix->vector, idxd_irq_handler,
113 				       idxd_misc_thread, 0, "idxd-misc",
114 				       irq_entry);
115 	if (rc < 0) {
116 		dev_err(dev, "Failed to allocate misc interrupt.\n");
117 		goto err_no_irq;
118 	}
119 
120 	dev_dbg(dev, "Allocated idxd-misc handler on msix vector %d\n",
121 		msix->vector);
122 
123 	/* first MSI-X entry is not for wq interrupts */
124 	idxd->num_wq_irqs = msixcnt - 1;
125 
126 	for (i = 1; i < msixcnt; i++) {
127 		msix = &idxd->msix_entries[i];
128 		irq_entry = &idxd->irq_entries[i];
129 
130 		init_llist_head(&idxd->irq_entries[i].pending_llist);
131 		INIT_LIST_HEAD(&idxd->irq_entries[i].work_list);
132 		rc = devm_request_threaded_irq(dev, msix->vector,
133 					       idxd_irq_handler,
134 					       idxd_wq_thread, 0,
135 					       "idxd-portal", irq_entry);
136 		if (rc < 0) {
137 			dev_err(dev, "Failed to allocate irq %d.\n",
138 				msix->vector);
139 			goto err_no_irq;
140 		}
141 		dev_dbg(dev, "Allocated idxd-msix %d for vector %d\n",
142 			i, msix->vector);
143 	}
144 
145 	idxd_unmask_error_interrupts(idxd);
146 	idxd_msix_perm_setup(idxd);
147 	return 0;
148 
149  err_no_irq:
150 	/* Disable error interrupt generation */
151 	idxd_mask_error_interrupts(idxd);
152 	pci_disable_msix(pdev);
153 	dev_err(dev, "No usable interrupts\n");
154 	return rc;
155 }
156 
157 static int idxd_setup_internals(struct idxd_device *idxd)
158 {
159 	struct device *dev = &idxd->pdev->dev;
160 	int i;
161 
162 	init_waitqueue_head(&idxd->cmd_waitq);
163 	idxd->groups = devm_kcalloc(dev, idxd->max_groups,
164 				    sizeof(struct idxd_group), GFP_KERNEL);
165 	if (!idxd->groups)
166 		return -ENOMEM;
167 
168 	for (i = 0; i < idxd->max_groups; i++) {
169 		idxd->groups[i].idxd = idxd;
170 		idxd->groups[i].id = i;
171 		idxd->groups[i].tc_a = -1;
172 		idxd->groups[i].tc_b = -1;
173 	}
174 
175 	idxd->wqs = devm_kcalloc(dev, idxd->max_wqs, sizeof(struct idxd_wq),
176 				 GFP_KERNEL);
177 	if (!idxd->wqs)
178 		return -ENOMEM;
179 
180 	idxd->engines = devm_kcalloc(dev, idxd->max_engines,
181 				     sizeof(struct idxd_engine), GFP_KERNEL);
182 	if (!idxd->engines)
183 		return -ENOMEM;
184 
185 	for (i = 0; i < idxd->max_wqs; i++) {
186 		struct idxd_wq *wq = &idxd->wqs[i];
187 
188 		wq->id = i;
189 		wq->idxd = idxd;
190 		mutex_init(&wq->wq_lock);
191 		wq->idxd_cdev.minor = -1;
192 		wq->max_xfer_bytes = idxd->max_xfer_bytes;
193 		wq->max_batch_size = idxd->max_batch_size;
194 		wq->wqcfg = devm_kzalloc(dev, idxd->wqcfg_size, GFP_KERNEL);
195 		if (!wq->wqcfg)
196 			return -ENOMEM;
197 	}
198 
199 	for (i = 0; i < idxd->max_engines; i++) {
200 		idxd->engines[i].idxd = idxd;
201 		idxd->engines[i].id = i;
202 	}
203 
204 	idxd->wq = create_workqueue(dev_name(dev));
205 	if (!idxd->wq)
206 		return -ENOMEM;
207 
208 	return 0;
209 }
210 
211 static void idxd_read_table_offsets(struct idxd_device *idxd)
212 {
213 	union offsets_reg offsets;
214 	struct device *dev = &idxd->pdev->dev;
215 
216 	offsets.bits[0] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET);
217 	offsets.bits[1] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET + sizeof(u64));
218 	idxd->grpcfg_offset = offsets.grpcfg * IDXD_TABLE_MULT;
219 	dev_dbg(dev, "IDXD Group Config Offset: %#x\n", idxd->grpcfg_offset);
220 	idxd->wqcfg_offset = offsets.wqcfg * IDXD_TABLE_MULT;
221 	dev_dbg(dev, "IDXD Work Queue Config Offset: %#x\n", idxd->wqcfg_offset);
222 	idxd->msix_perm_offset = offsets.msix_perm * IDXD_TABLE_MULT;
223 	dev_dbg(dev, "IDXD MSIX Permission Offset: %#x\n", idxd->msix_perm_offset);
224 	idxd->perfmon_offset = offsets.perfmon * IDXD_TABLE_MULT;
225 	dev_dbg(dev, "IDXD Perfmon Offset: %#x\n", idxd->perfmon_offset);
226 }
227 
228 static void idxd_read_caps(struct idxd_device *idxd)
229 {
230 	struct device *dev = &idxd->pdev->dev;
231 	int i;
232 
233 	/* reading generic capabilities */
234 	idxd->hw.gen_cap.bits = ioread64(idxd->reg_base + IDXD_GENCAP_OFFSET);
235 	dev_dbg(dev, "gen_cap: %#llx\n", idxd->hw.gen_cap.bits);
236 	idxd->max_xfer_bytes = 1ULL << idxd->hw.gen_cap.max_xfer_shift;
237 	dev_dbg(dev, "max xfer size: %llu bytes\n", idxd->max_xfer_bytes);
238 	idxd->max_batch_size = 1U << idxd->hw.gen_cap.max_batch_shift;
239 	dev_dbg(dev, "max batch size: %u\n", idxd->max_batch_size);
240 	if (idxd->hw.gen_cap.config_en)
241 		set_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags);
242 
243 	/* reading group capabilities */
244 	idxd->hw.group_cap.bits =
245 		ioread64(idxd->reg_base + IDXD_GRPCAP_OFFSET);
246 	dev_dbg(dev, "group_cap: %#llx\n", idxd->hw.group_cap.bits);
247 	idxd->max_groups = idxd->hw.group_cap.num_groups;
248 	dev_dbg(dev, "max groups: %u\n", idxd->max_groups);
249 	idxd->max_tokens = idxd->hw.group_cap.total_tokens;
250 	dev_dbg(dev, "max tokens: %u\n", idxd->max_tokens);
251 	idxd->nr_tokens = idxd->max_tokens;
252 
253 	/* read engine capabilities */
254 	idxd->hw.engine_cap.bits =
255 		ioread64(idxd->reg_base + IDXD_ENGCAP_OFFSET);
256 	dev_dbg(dev, "engine_cap: %#llx\n", idxd->hw.engine_cap.bits);
257 	idxd->max_engines = idxd->hw.engine_cap.num_engines;
258 	dev_dbg(dev, "max engines: %u\n", idxd->max_engines);
259 
260 	/* read workqueue capabilities */
261 	idxd->hw.wq_cap.bits = ioread64(idxd->reg_base + IDXD_WQCAP_OFFSET);
262 	dev_dbg(dev, "wq_cap: %#llx\n", idxd->hw.wq_cap.bits);
263 	idxd->max_wq_size = idxd->hw.wq_cap.total_wq_size;
264 	dev_dbg(dev, "total workqueue size: %u\n", idxd->max_wq_size);
265 	idxd->max_wqs = idxd->hw.wq_cap.num_wqs;
266 	dev_dbg(dev, "max workqueues: %u\n", idxd->max_wqs);
267 	idxd->wqcfg_size = 1 << (idxd->hw.wq_cap.wqcfg_size + IDXD_WQCFG_MIN);
268 	dev_dbg(dev, "wqcfg size: %u\n", idxd->wqcfg_size);
269 
270 	/* reading operation capabilities */
271 	for (i = 0; i < 4; i++) {
272 		idxd->hw.opcap.bits[i] = ioread64(idxd->reg_base +
273 				IDXD_OPCAP_OFFSET + i * sizeof(u64));
274 		dev_dbg(dev, "opcap[%d]: %#llx\n", i, idxd->hw.opcap.bits[i]);
275 	}
276 }
277 
278 static struct idxd_device *idxd_alloc(struct pci_dev *pdev)
279 {
280 	struct device *dev = &pdev->dev;
281 	struct idxd_device *idxd;
282 
283 	idxd = devm_kzalloc(dev, sizeof(struct idxd_device), GFP_KERNEL);
284 	if (!idxd)
285 		return NULL;
286 
287 	idxd->pdev = pdev;
288 	spin_lock_init(&idxd->dev_lock);
289 
290 	return idxd;
291 }
292 
293 static int idxd_enable_system_pasid(struct idxd_device *idxd)
294 {
295 	int flags;
296 	unsigned int pasid;
297 	struct iommu_sva *sva;
298 
299 	flags = SVM_FLAG_SUPERVISOR_MODE;
300 
301 	sva = iommu_sva_bind_device(&idxd->pdev->dev, NULL, &flags);
302 	if (IS_ERR(sva)) {
303 		dev_warn(&idxd->pdev->dev,
304 			 "iommu sva bind failed: %ld\n", PTR_ERR(sva));
305 		return PTR_ERR(sva);
306 	}
307 
308 	pasid = iommu_sva_get_pasid(sva);
309 	if (pasid == IOMMU_PASID_INVALID) {
310 		iommu_sva_unbind_device(sva);
311 		return -ENODEV;
312 	}
313 
314 	idxd->sva = sva;
315 	idxd->pasid = pasid;
316 	dev_dbg(&idxd->pdev->dev, "system pasid: %u\n", pasid);
317 	return 0;
318 }
319 
320 static void idxd_disable_system_pasid(struct idxd_device *idxd)
321 {
322 
323 	iommu_sva_unbind_device(idxd->sva);
324 	idxd->sva = NULL;
325 }
326 
327 static int idxd_probe(struct idxd_device *idxd)
328 {
329 	struct pci_dev *pdev = idxd->pdev;
330 	struct device *dev = &pdev->dev;
331 	int rc;
332 
333 	dev_dbg(dev, "%s entered and resetting device\n", __func__);
334 	rc = idxd_device_init_reset(idxd);
335 	if (rc < 0)
336 		return rc;
337 
338 	dev_dbg(dev, "IDXD reset complete\n");
339 
340 	if (IS_ENABLED(CONFIG_INTEL_IDXD_SVM) && sva) {
341 		rc = idxd_enable_system_pasid(idxd);
342 		if (rc < 0)
343 			dev_warn(dev, "Failed to enable PASID. No SVA support: %d\n", rc);
344 		else
345 			set_bit(IDXD_FLAG_PASID_ENABLED, &idxd->flags);
346 	} else if (!sva) {
347 		dev_warn(dev, "User forced SVA off via module param.\n");
348 	}
349 
350 	idxd_read_caps(idxd);
351 	idxd_read_table_offsets(idxd);
352 
353 	rc = idxd_setup_internals(idxd);
354 	if (rc)
355 		goto err_setup;
356 
357 	rc = idxd_setup_interrupts(idxd);
358 	if (rc)
359 		goto err_setup;
360 
361 	dev_dbg(dev, "IDXD interrupt setup complete.\n");
362 
363 	mutex_lock(&idxd_idr_lock);
364 	idxd->id = idr_alloc(&idxd_idrs[idxd->type], idxd, 0, 0, GFP_KERNEL);
365 	mutex_unlock(&idxd_idr_lock);
366 	if (idxd->id < 0) {
367 		rc = -ENOMEM;
368 		goto err_idr_fail;
369 	}
370 
371 	idxd->major = idxd_cdev_get_major(idxd);
372 
373 	dev_dbg(dev, "IDXD device %d probed successfully\n", idxd->id);
374 	return 0;
375 
376  err_idr_fail:
377 	idxd_mask_error_interrupts(idxd);
378 	idxd_mask_msix_vectors(idxd);
379  err_setup:
380 	if (device_pasid_enabled(idxd))
381 		idxd_disable_system_pasid(idxd);
382 	return rc;
383 }
384 
385 static void idxd_type_init(struct idxd_device *idxd)
386 {
387 	if (idxd->type == IDXD_TYPE_DSA)
388 		idxd->compl_size = sizeof(struct dsa_completion_record);
389 	else if (idxd->type == IDXD_TYPE_IAX)
390 		idxd->compl_size = sizeof(struct iax_completion_record);
391 }
392 
393 static int idxd_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
394 {
395 	struct device *dev = &pdev->dev;
396 	struct idxd_device *idxd;
397 	int rc;
398 
399 	rc = pcim_enable_device(pdev);
400 	if (rc)
401 		return rc;
402 
403 	dev_dbg(dev, "Alloc IDXD context\n");
404 	idxd = idxd_alloc(pdev);
405 	if (!idxd)
406 		return -ENOMEM;
407 
408 	dev_dbg(dev, "Mapping BARs\n");
409 	idxd->reg_base = pcim_iomap(pdev, IDXD_MMIO_BAR, 0);
410 	if (!idxd->reg_base)
411 		return -ENOMEM;
412 
413 	dev_dbg(dev, "Set DMA masks\n");
414 	rc = pci_set_dma_mask(pdev, DMA_BIT_MASK(64));
415 	if (rc)
416 		rc = pci_set_dma_mask(pdev, DMA_BIT_MASK(32));
417 	if (rc)
418 		return rc;
419 
420 	rc = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
421 	if (rc)
422 		rc = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(32));
423 	if (rc)
424 		return rc;
425 
426 	idxd_set_type(idxd);
427 
428 	idxd_type_init(idxd);
429 
430 	dev_dbg(dev, "Set PCI master\n");
431 	pci_set_master(pdev);
432 	pci_set_drvdata(pdev, idxd);
433 
434 	idxd->hw.version = ioread32(idxd->reg_base + IDXD_VER_OFFSET);
435 	rc = idxd_probe(idxd);
436 	if (rc) {
437 		dev_err(dev, "Intel(R) IDXD DMA Engine init failed\n");
438 		return -ENODEV;
439 	}
440 
441 	rc = idxd_setup_sysfs(idxd);
442 	if (rc) {
443 		dev_err(dev, "IDXD sysfs setup failed\n");
444 		return -ENODEV;
445 	}
446 
447 	idxd->state = IDXD_DEV_CONF_READY;
448 
449 	dev_info(&pdev->dev, "Intel(R) Accelerator Device (v%x)\n",
450 		 idxd->hw.version);
451 
452 	return 0;
453 }
454 
455 static void idxd_flush_pending_llist(struct idxd_irq_entry *ie)
456 {
457 	struct idxd_desc *desc, *itr;
458 	struct llist_node *head;
459 
460 	head = llist_del_all(&ie->pending_llist);
461 	if (!head)
462 		return;
463 
464 	llist_for_each_entry_safe(desc, itr, head, llnode) {
465 		idxd_dma_complete_txd(desc, IDXD_COMPLETE_ABORT);
466 		idxd_free_desc(desc->wq, desc);
467 	}
468 }
469 
470 static void idxd_flush_work_list(struct idxd_irq_entry *ie)
471 {
472 	struct idxd_desc *desc, *iter;
473 
474 	list_for_each_entry_safe(desc, iter, &ie->work_list, list) {
475 		list_del(&desc->list);
476 		idxd_dma_complete_txd(desc, IDXD_COMPLETE_ABORT);
477 		idxd_free_desc(desc->wq, desc);
478 	}
479 }
480 
481 static void idxd_shutdown(struct pci_dev *pdev)
482 {
483 	struct idxd_device *idxd = pci_get_drvdata(pdev);
484 	int rc, i;
485 	struct idxd_irq_entry *irq_entry;
486 	int msixcnt = pci_msix_vec_count(pdev);
487 
488 	rc = idxd_device_disable(idxd);
489 	if (rc)
490 		dev_err(&pdev->dev, "Disabling device failed\n");
491 
492 	dev_dbg(&pdev->dev, "%s called\n", __func__);
493 	idxd_mask_msix_vectors(idxd);
494 	idxd_mask_error_interrupts(idxd);
495 
496 	for (i = 0; i < msixcnt; i++) {
497 		irq_entry = &idxd->irq_entries[i];
498 		synchronize_irq(idxd->msix_entries[i].vector);
499 		if (i == 0)
500 			continue;
501 		idxd_flush_pending_llist(irq_entry);
502 		idxd_flush_work_list(irq_entry);
503 	}
504 
505 	idxd_msix_perm_clear(idxd);
506 	destroy_workqueue(idxd->wq);
507 }
508 
509 static void idxd_remove(struct pci_dev *pdev)
510 {
511 	struct idxd_device *idxd = pci_get_drvdata(pdev);
512 
513 	dev_dbg(&pdev->dev, "%s called\n", __func__);
514 	idxd_cleanup_sysfs(idxd);
515 	idxd_shutdown(pdev);
516 	if (device_pasid_enabled(idxd))
517 		idxd_disable_system_pasid(idxd);
518 	mutex_lock(&idxd_idr_lock);
519 	idr_remove(&idxd_idrs[idxd->type], idxd->id);
520 	mutex_unlock(&idxd_idr_lock);
521 }
522 
523 static struct pci_driver idxd_pci_driver = {
524 	.name		= DRV_NAME,
525 	.id_table	= idxd_pci_tbl,
526 	.probe		= idxd_pci_probe,
527 	.remove		= idxd_remove,
528 	.shutdown	= idxd_shutdown,
529 };
530 
531 static int __init idxd_init_module(void)
532 {
533 	int err, i;
534 
535 	/*
536 	 * If the CPU does not support MOVDIR64B or ENQCMDS, there's no point in
537 	 * enumerating the device. We can not utilize it.
538 	 */
539 	if (!boot_cpu_has(X86_FEATURE_MOVDIR64B)) {
540 		pr_warn("idxd driver failed to load without MOVDIR64B.\n");
541 		return -ENODEV;
542 	}
543 
544 	if (!boot_cpu_has(X86_FEATURE_ENQCMD))
545 		pr_warn("Platform does not have ENQCMD(S) support.\n");
546 	else
547 		support_enqcmd = true;
548 
549 	for (i = 0; i < IDXD_TYPE_MAX; i++)
550 		idr_init(&idxd_idrs[i]);
551 
552 	err = idxd_register_bus_type();
553 	if (err < 0)
554 		return err;
555 
556 	err = idxd_register_driver();
557 	if (err < 0)
558 		goto err_idxd_driver_register;
559 
560 	err = idxd_cdev_register();
561 	if (err)
562 		goto err_cdev_register;
563 
564 	err = pci_register_driver(&idxd_pci_driver);
565 	if (err)
566 		goto err_pci_register;
567 
568 	return 0;
569 
570 err_pci_register:
571 	idxd_cdev_remove();
572 err_cdev_register:
573 	idxd_unregister_driver();
574 err_idxd_driver_register:
575 	idxd_unregister_bus_type();
576 	return err;
577 }
578 module_init(idxd_init_module);
579 
580 static void __exit idxd_exit_module(void)
581 {
582 	pci_unregister_driver(&idxd_pci_driver);
583 	idxd_cdev_remove();
584 	idxd_unregister_bus_type();
585 }
586 module_exit(idxd_exit_module);
587