xref: /linux/drivers/dma/idxd/init.c (revision 034b3290ba257f1a3c8730f3fba72e11645e7b50)
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 #include "perfmon.h"
25 
26 MODULE_VERSION(IDXD_DRIVER_VERSION);
27 MODULE_LICENSE("GPL v2");
28 MODULE_AUTHOR("Intel Corporation");
29 
30 static bool sva = true;
31 module_param(sva, bool, 0644);
32 MODULE_PARM_DESC(sva, "Toggle SVA support on/off");
33 
34 #define DRV_NAME "idxd"
35 
36 bool support_enqcmd;
37 DEFINE_IDA(idxd_ida);
38 
39 static struct idxd_driver_data idxd_driver_data[] = {
40 	[IDXD_TYPE_DSA] = {
41 		.name_prefix = "dsa",
42 		.type = IDXD_TYPE_DSA,
43 		.compl_size = sizeof(struct dsa_completion_record),
44 		.align = 32,
45 		.dev_type = &dsa_device_type,
46 	},
47 	[IDXD_TYPE_IAX] = {
48 		.name_prefix = "iax",
49 		.type = IDXD_TYPE_IAX,
50 		.compl_size = sizeof(struct iax_completion_record),
51 		.align = 64,
52 		.dev_type = &iax_device_type,
53 	},
54 };
55 
56 static struct pci_device_id idxd_pci_tbl[] = {
57 	/* DSA ver 1.0 platforms */
58 	{ PCI_DEVICE_DATA(INTEL, DSA_SPR0, &idxd_driver_data[IDXD_TYPE_DSA]) },
59 
60 	/* IAX ver 1.0 platforms */
61 	{ PCI_DEVICE_DATA(INTEL, IAX_SPR0, &idxd_driver_data[IDXD_TYPE_IAX]) },
62 	{ 0, }
63 };
64 MODULE_DEVICE_TABLE(pci, idxd_pci_tbl);
65 
66 static int idxd_setup_interrupts(struct idxd_device *idxd)
67 {
68 	struct pci_dev *pdev = idxd->pdev;
69 	struct device *dev = &pdev->dev;
70 	struct idxd_irq_entry *irq_entry;
71 	int i, msixcnt;
72 	int rc = 0;
73 
74 	msixcnt = pci_msix_vec_count(pdev);
75 	if (msixcnt < 0) {
76 		dev_err(dev, "Not MSI-X interrupt capable.\n");
77 		return -ENOSPC;
78 	}
79 
80 	rc = pci_alloc_irq_vectors(pdev, msixcnt, msixcnt, PCI_IRQ_MSIX);
81 	if (rc != msixcnt) {
82 		dev_err(dev, "Failed enabling %d MSIX entries: %d\n", msixcnt, rc);
83 		return -ENOSPC;
84 	}
85 	dev_dbg(dev, "Enabled %d msix vectors\n", msixcnt);
86 
87 	/*
88 	 * We implement 1 completion list per MSI-X entry except for
89 	 * entry 0, which is for errors and others.
90 	 */
91 	idxd->irq_entries = kcalloc_node(msixcnt, sizeof(struct idxd_irq_entry),
92 					 GFP_KERNEL, dev_to_node(dev));
93 	if (!idxd->irq_entries) {
94 		rc = -ENOMEM;
95 		goto err_irq_entries;
96 	}
97 
98 	for (i = 0; i < msixcnt; i++) {
99 		idxd->irq_entries[i].id = i;
100 		idxd->irq_entries[i].idxd = idxd;
101 		idxd->irq_entries[i].vector = pci_irq_vector(pdev, i);
102 		spin_lock_init(&idxd->irq_entries[i].list_lock);
103 	}
104 
105 	idxd_msix_perm_setup(idxd);
106 
107 	irq_entry = &idxd->irq_entries[0];
108 	rc = request_threaded_irq(irq_entry->vector, NULL, idxd_misc_thread,
109 				  0, "idxd-misc", irq_entry);
110 	if (rc < 0) {
111 		dev_err(dev, "Failed to allocate misc interrupt.\n");
112 		goto err_misc_irq;
113 	}
114 
115 	dev_dbg(dev, "Allocated idxd-misc handler on msix vector %d\n", irq_entry->vector);
116 
117 	/* first MSI-X entry is not for wq interrupts */
118 	idxd->num_wq_irqs = msixcnt - 1;
119 
120 	for (i = 1; i < msixcnt; i++) {
121 		irq_entry = &idxd->irq_entries[i];
122 
123 		init_llist_head(&idxd->irq_entries[i].pending_llist);
124 		INIT_LIST_HEAD(&idxd->irq_entries[i].work_list);
125 		rc = request_threaded_irq(irq_entry->vector, NULL,
126 					  idxd_wq_thread, 0, "idxd-portal", irq_entry);
127 		if (rc < 0) {
128 			dev_err(dev, "Failed to allocate irq %d.\n", irq_entry->vector);
129 			goto err_wq_irqs;
130 		}
131 
132 		dev_dbg(dev, "Allocated idxd-msix %d for vector %d\n", i, irq_entry->vector);
133 		if (idxd->hw.cmd_cap & BIT(IDXD_CMD_REQUEST_INT_HANDLE)) {
134 			/*
135 			 * The MSIX vector enumeration starts at 1 with vector 0 being the
136 			 * misc interrupt that handles non I/O completion events. The
137 			 * interrupt handles are for IMS enumeration on guest. The misc
138 			 * interrupt vector does not require a handle and therefore we start
139 			 * the int_handles at index 0. Since 'i' starts at 1, the first
140 			 * int_handles index will be 0.
141 			 */
142 			rc = idxd_device_request_int_handle(idxd, i, &idxd->int_handles[i - 1],
143 							    IDXD_IRQ_MSIX);
144 			if (rc < 0) {
145 				free_irq(irq_entry->vector, irq_entry);
146 				goto err_wq_irqs;
147 			}
148 			dev_dbg(dev, "int handle requested: %u\n", idxd->int_handles[i - 1]);
149 		}
150 	}
151 
152 	idxd_unmask_error_interrupts(idxd);
153 	return 0;
154 
155  err_wq_irqs:
156 	while (--i >= 0) {
157 		irq_entry = &idxd->irq_entries[i];
158 		free_irq(irq_entry->vector, irq_entry);
159 		if (i != 0)
160 			idxd_device_release_int_handle(idxd,
161 						       idxd->int_handles[i], IDXD_IRQ_MSIX);
162 	}
163  err_misc_irq:
164 	/* Disable error interrupt generation */
165 	idxd_mask_error_interrupts(idxd);
166 	idxd_msix_perm_clear(idxd);
167  err_irq_entries:
168 	pci_free_irq_vectors(pdev);
169 	dev_err(dev, "No usable interrupts\n");
170 	return rc;
171 }
172 
173 static void idxd_cleanup_interrupts(struct idxd_device *idxd)
174 {
175 	struct pci_dev *pdev = idxd->pdev;
176 	struct idxd_irq_entry *irq_entry;
177 	int i, msixcnt;
178 
179 	msixcnt = pci_msix_vec_count(pdev);
180 	if (msixcnt <= 0)
181 		return;
182 
183 	irq_entry = &idxd->irq_entries[0];
184 	free_irq(irq_entry->vector, irq_entry);
185 
186 	for (i = 1; i < msixcnt; i++) {
187 
188 		irq_entry = &idxd->irq_entries[i];
189 		if (idxd->hw.cmd_cap & BIT(IDXD_CMD_RELEASE_INT_HANDLE))
190 			idxd_device_release_int_handle(idxd, idxd->int_handles[i],
191 						       IDXD_IRQ_MSIX);
192 		free_irq(irq_entry->vector, irq_entry);
193 	}
194 
195 	idxd_mask_error_interrupts(idxd);
196 	pci_free_irq_vectors(pdev);
197 }
198 
199 static int idxd_setup_wqs(struct idxd_device *idxd)
200 {
201 	struct device *dev = &idxd->pdev->dev;
202 	struct idxd_wq *wq;
203 	struct device *conf_dev;
204 	int i, rc;
205 
206 	idxd->wqs = kcalloc_node(idxd->max_wqs, sizeof(struct idxd_wq *),
207 				 GFP_KERNEL, dev_to_node(dev));
208 	if (!idxd->wqs)
209 		return -ENOMEM;
210 
211 	for (i = 0; i < idxd->max_wqs; i++) {
212 		wq = kzalloc_node(sizeof(*wq), GFP_KERNEL, dev_to_node(dev));
213 		if (!wq) {
214 			rc = -ENOMEM;
215 			goto err;
216 		}
217 
218 		idxd_dev_set_type(&wq->idxd_dev, IDXD_DEV_WQ);
219 		conf_dev = wq_confdev(wq);
220 		wq->id = i;
221 		wq->idxd = idxd;
222 		device_initialize(wq_confdev(wq));
223 		conf_dev->parent = idxd_confdev(idxd);
224 		conf_dev->bus = &dsa_bus_type;
225 		conf_dev->type = &idxd_wq_device_type;
226 		rc = dev_set_name(conf_dev, "wq%d.%d", idxd->id, wq->id);
227 		if (rc < 0) {
228 			put_device(conf_dev);
229 			goto err;
230 		}
231 
232 		mutex_init(&wq->wq_lock);
233 		init_waitqueue_head(&wq->err_queue);
234 		init_completion(&wq->wq_dead);
235 		wq->max_xfer_bytes = idxd->max_xfer_bytes;
236 		wq->max_batch_size = idxd->max_batch_size;
237 		wq->wqcfg = kzalloc_node(idxd->wqcfg_size, GFP_KERNEL, dev_to_node(dev));
238 		if (!wq->wqcfg) {
239 			put_device(conf_dev);
240 			rc = -ENOMEM;
241 			goto err;
242 		}
243 		idxd->wqs[i] = wq;
244 	}
245 
246 	return 0;
247 
248  err:
249 	while (--i >= 0) {
250 		wq = idxd->wqs[i];
251 		conf_dev = wq_confdev(wq);
252 		put_device(conf_dev);
253 	}
254 	return rc;
255 }
256 
257 static int idxd_setup_engines(struct idxd_device *idxd)
258 {
259 	struct idxd_engine *engine;
260 	struct device *dev = &idxd->pdev->dev;
261 	struct device *conf_dev;
262 	int i, rc;
263 
264 	idxd->engines = kcalloc_node(idxd->max_engines, sizeof(struct idxd_engine *),
265 				     GFP_KERNEL, dev_to_node(dev));
266 	if (!idxd->engines)
267 		return -ENOMEM;
268 
269 	for (i = 0; i < idxd->max_engines; i++) {
270 		engine = kzalloc_node(sizeof(*engine), GFP_KERNEL, dev_to_node(dev));
271 		if (!engine) {
272 			rc = -ENOMEM;
273 			goto err;
274 		}
275 
276 		idxd_dev_set_type(&engine->idxd_dev, IDXD_DEV_ENGINE);
277 		conf_dev = engine_confdev(engine);
278 		engine->id = i;
279 		engine->idxd = idxd;
280 		device_initialize(conf_dev);
281 		conf_dev->parent = idxd_confdev(idxd);
282 		conf_dev->bus = &dsa_bus_type;
283 		conf_dev->type = &idxd_engine_device_type;
284 		rc = dev_set_name(conf_dev, "engine%d.%d", idxd->id, engine->id);
285 		if (rc < 0) {
286 			put_device(conf_dev);
287 			goto err;
288 		}
289 
290 		idxd->engines[i] = engine;
291 	}
292 
293 	return 0;
294 
295  err:
296 	while (--i >= 0) {
297 		engine = idxd->engines[i];
298 		conf_dev = engine_confdev(engine);
299 		put_device(conf_dev);
300 	}
301 	return rc;
302 }
303 
304 static int idxd_setup_groups(struct idxd_device *idxd)
305 {
306 	struct device *dev = &idxd->pdev->dev;
307 	struct device *conf_dev;
308 	struct idxd_group *group;
309 	int i, rc;
310 
311 	idxd->groups = kcalloc_node(idxd->max_groups, sizeof(struct idxd_group *),
312 				    GFP_KERNEL, dev_to_node(dev));
313 	if (!idxd->groups)
314 		return -ENOMEM;
315 
316 	for (i = 0; i < idxd->max_groups; i++) {
317 		group = kzalloc_node(sizeof(*group), GFP_KERNEL, dev_to_node(dev));
318 		if (!group) {
319 			rc = -ENOMEM;
320 			goto err;
321 		}
322 
323 		idxd_dev_set_type(&group->idxd_dev, IDXD_DEV_GROUP);
324 		conf_dev = group_confdev(group);
325 		group->id = i;
326 		group->idxd = idxd;
327 		device_initialize(conf_dev);
328 		conf_dev->parent = idxd_confdev(idxd);
329 		conf_dev->bus = &dsa_bus_type;
330 		conf_dev->type = &idxd_group_device_type;
331 		rc = dev_set_name(conf_dev, "group%d.%d", idxd->id, group->id);
332 		if (rc < 0) {
333 			put_device(conf_dev);
334 			goto err;
335 		}
336 
337 		idxd->groups[i] = group;
338 		group->tc_a = -1;
339 		group->tc_b = -1;
340 	}
341 
342 	return 0;
343 
344  err:
345 	while (--i >= 0) {
346 		group = idxd->groups[i];
347 		put_device(group_confdev(group));
348 	}
349 	return rc;
350 }
351 
352 static void idxd_cleanup_internals(struct idxd_device *idxd)
353 {
354 	int i;
355 
356 	for (i = 0; i < idxd->max_groups; i++)
357 		put_device(group_confdev(idxd->groups[i]));
358 	for (i = 0; i < idxd->max_engines; i++)
359 		put_device(engine_confdev(idxd->engines[i]));
360 	for (i = 0; i < idxd->max_wqs; i++)
361 		put_device(wq_confdev(idxd->wqs[i]));
362 	destroy_workqueue(idxd->wq);
363 }
364 
365 static int idxd_setup_internals(struct idxd_device *idxd)
366 {
367 	struct device *dev = &idxd->pdev->dev;
368 	int rc, i;
369 
370 	init_waitqueue_head(&idxd->cmd_waitq);
371 
372 	if (idxd->hw.cmd_cap & BIT(IDXD_CMD_REQUEST_INT_HANDLE)) {
373 		idxd->int_handles = kcalloc_node(idxd->max_wqs, sizeof(int), GFP_KERNEL,
374 						 dev_to_node(dev));
375 		if (!idxd->int_handles)
376 			return -ENOMEM;
377 	}
378 
379 	rc = idxd_setup_wqs(idxd);
380 	if (rc < 0)
381 		goto err_wqs;
382 
383 	rc = idxd_setup_engines(idxd);
384 	if (rc < 0)
385 		goto err_engine;
386 
387 	rc = idxd_setup_groups(idxd);
388 	if (rc < 0)
389 		goto err_group;
390 
391 	idxd->wq = create_workqueue(dev_name(dev));
392 	if (!idxd->wq) {
393 		rc = -ENOMEM;
394 		goto err_wkq_create;
395 	}
396 
397 	return 0;
398 
399  err_wkq_create:
400 	for (i = 0; i < idxd->max_groups; i++)
401 		put_device(group_confdev(idxd->groups[i]));
402  err_group:
403 	for (i = 0; i < idxd->max_engines; i++)
404 		put_device(engine_confdev(idxd->engines[i]));
405  err_engine:
406 	for (i = 0; i < idxd->max_wqs; i++)
407 		put_device(wq_confdev(idxd->wqs[i]));
408  err_wqs:
409 	kfree(idxd->int_handles);
410 	return rc;
411 }
412 
413 static void idxd_read_table_offsets(struct idxd_device *idxd)
414 {
415 	union offsets_reg offsets;
416 	struct device *dev = &idxd->pdev->dev;
417 
418 	offsets.bits[0] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET);
419 	offsets.bits[1] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET + sizeof(u64));
420 	idxd->grpcfg_offset = offsets.grpcfg * IDXD_TABLE_MULT;
421 	dev_dbg(dev, "IDXD Group Config Offset: %#x\n", idxd->grpcfg_offset);
422 	idxd->wqcfg_offset = offsets.wqcfg * IDXD_TABLE_MULT;
423 	dev_dbg(dev, "IDXD Work Queue Config Offset: %#x\n", idxd->wqcfg_offset);
424 	idxd->msix_perm_offset = offsets.msix_perm * IDXD_TABLE_MULT;
425 	dev_dbg(dev, "IDXD MSIX Permission Offset: %#x\n", idxd->msix_perm_offset);
426 	idxd->perfmon_offset = offsets.perfmon * IDXD_TABLE_MULT;
427 	dev_dbg(dev, "IDXD Perfmon Offset: %#x\n", idxd->perfmon_offset);
428 }
429 
430 static void idxd_read_caps(struct idxd_device *idxd)
431 {
432 	struct device *dev = &idxd->pdev->dev;
433 	int i;
434 
435 	/* reading generic capabilities */
436 	idxd->hw.gen_cap.bits = ioread64(idxd->reg_base + IDXD_GENCAP_OFFSET);
437 	dev_dbg(dev, "gen_cap: %#llx\n", idxd->hw.gen_cap.bits);
438 
439 	if (idxd->hw.gen_cap.cmd_cap) {
440 		idxd->hw.cmd_cap = ioread32(idxd->reg_base + IDXD_CMDCAP_OFFSET);
441 		dev_dbg(dev, "cmd_cap: %#x\n", idxd->hw.cmd_cap);
442 	}
443 
444 	idxd->max_xfer_bytes = 1ULL << idxd->hw.gen_cap.max_xfer_shift;
445 	dev_dbg(dev, "max xfer size: %llu bytes\n", idxd->max_xfer_bytes);
446 	idxd->max_batch_size = 1U << idxd->hw.gen_cap.max_batch_shift;
447 	dev_dbg(dev, "max batch size: %u\n", idxd->max_batch_size);
448 	if (idxd->hw.gen_cap.config_en)
449 		set_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags);
450 
451 	/* reading group capabilities */
452 	idxd->hw.group_cap.bits =
453 		ioread64(idxd->reg_base + IDXD_GRPCAP_OFFSET);
454 	dev_dbg(dev, "group_cap: %#llx\n", idxd->hw.group_cap.bits);
455 	idxd->max_groups = idxd->hw.group_cap.num_groups;
456 	dev_dbg(dev, "max groups: %u\n", idxd->max_groups);
457 	idxd->max_tokens = idxd->hw.group_cap.total_tokens;
458 	dev_dbg(dev, "max tokens: %u\n", idxd->max_tokens);
459 	idxd->nr_tokens = idxd->max_tokens;
460 
461 	/* read engine capabilities */
462 	idxd->hw.engine_cap.bits =
463 		ioread64(idxd->reg_base + IDXD_ENGCAP_OFFSET);
464 	dev_dbg(dev, "engine_cap: %#llx\n", idxd->hw.engine_cap.bits);
465 	idxd->max_engines = idxd->hw.engine_cap.num_engines;
466 	dev_dbg(dev, "max engines: %u\n", idxd->max_engines);
467 
468 	/* read workqueue capabilities */
469 	idxd->hw.wq_cap.bits = ioread64(idxd->reg_base + IDXD_WQCAP_OFFSET);
470 	dev_dbg(dev, "wq_cap: %#llx\n", idxd->hw.wq_cap.bits);
471 	idxd->max_wq_size = idxd->hw.wq_cap.total_wq_size;
472 	dev_dbg(dev, "total workqueue size: %u\n", idxd->max_wq_size);
473 	idxd->max_wqs = idxd->hw.wq_cap.num_wqs;
474 	dev_dbg(dev, "max workqueues: %u\n", idxd->max_wqs);
475 	idxd->wqcfg_size = 1 << (idxd->hw.wq_cap.wqcfg_size + IDXD_WQCFG_MIN);
476 	dev_dbg(dev, "wqcfg size: %u\n", idxd->wqcfg_size);
477 
478 	/* reading operation capabilities */
479 	for (i = 0; i < 4; i++) {
480 		idxd->hw.opcap.bits[i] = ioread64(idxd->reg_base +
481 				IDXD_OPCAP_OFFSET + i * sizeof(u64));
482 		dev_dbg(dev, "opcap[%d]: %#llx\n", i, idxd->hw.opcap.bits[i]);
483 	}
484 }
485 
486 static struct idxd_device *idxd_alloc(struct pci_dev *pdev, struct idxd_driver_data *data)
487 {
488 	struct device *dev = &pdev->dev;
489 	struct device *conf_dev;
490 	struct idxd_device *idxd;
491 	int rc;
492 
493 	idxd = kzalloc_node(sizeof(*idxd), GFP_KERNEL, dev_to_node(dev));
494 	if (!idxd)
495 		return NULL;
496 
497 	conf_dev = idxd_confdev(idxd);
498 	idxd->pdev = pdev;
499 	idxd->data = data;
500 	idxd_dev_set_type(&idxd->idxd_dev, idxd->data->type);
501 	idxd->id = ida_alloc(&idxd_ida, GFP_KERNEL);
502 	if (idxd->id < 0)
503 		return NULL;
504 
505 	device_initialize(conf_dev);
506 	conf_dev->parent = dev;
507 	conf_dev->bus = &dsa_bus_type;
508 	conf_dev->type = idxd->data->dev_type;
509 	rc = dev_set_name(conf_dev, "%s%d", idxd->data->name_prefix, idxd->id);
510 	if (rc < 0) {
511 		put_device(conf_dev);
512 		return NULL;
513 	}
514 
515 	spin_lock_init(&idxd->dev_lock);
516 	spin_lock_init(&idxd->cmd_lock);
517 
518 	return idxd;
519 }
520 
521 static int idxd_enable_system_pasid(struct idxd_device *idxd)
522 {
523 	int flags;
524 	unsigned int pasid;
525 	struct iommu_sva *sva;
526 
527 	flags = SVM_FLAG_SUPERVISOR_MODE;
528 
529 	sva = iommu_sva_bind_device(&idxd->pdev->dev, NULL, &flags);
530 	if (IS_ERR(sva)) {
531 		dev_warn(&idxd->pdev->dev,
532 			 "iommu sva bind failed: %ld\n", PTR_ERR(sva));
533 		return PTR_ERR(sva);
534 	}
535 
536 	pasid = iommu_sva_get_pasid(sva);
537 	if (pasid == IOMMU_PASID_INVALID) {
538 		iommu_sva_unbind_device(sva);
539 		return -ENODEV;
540 	}
541 
542 	idxd->sva = sva;
543 	idxd->pasid = pasid;
544 	dev_dbg(&idxd->pdev->dev, "system pasid: %u\n", pasid);
545 	return 0;
546 }
547 
548 static void idxd_disable_system_pasid(struct idxd_device *idxd)
549 {
550 
551 	iommu_sva_unbind_device(idxd->sva);
552 	idxd->sva = NULL;
553 }
554 
555 static int idxd_probe(struct idxd_device *idxd)
556 {
557 	struct pci_dev *pdev = idxd->pdev;
558 	struct device *dev = &pdev->dev;
559 	int rc;
560 
561 	dev_dbg(dev, "%s entered and resetting device\n", __func__);
562 	rc = idxd_device_init_reset(idxd);
563 	if (rc < 0)
564 		return rc;
565 
566 	dev_dbg(dev, "IDXD reset complete\n");
567 
568 	if (IS_ENABLED(CONFIG_INTEL_IDXD_SVM) && sva) {
569 		rc = iommu_dev_enable_feature(dev, IOMMU_DEV_FEAT_SVA);
570 		if (rc == 0) {
571 			rc = idxd_enable_system_pasid(idxd);
572 			if (rc < 0) {
573 				iommu_dev_disable_feature(dev, IOMMU_DEV_FEAT_SVA);
574 				dev_warn(dev, "Failed to enable PASID. No SVA support: %d\n", rc);
575 			} else {
576 				set_bit(IDXD_FLAG_PASID_ENABLED, &idxd->flags);
577 			}
578 		} else {
579 			dev_warn(dev, "Unable to turn on SVA feature.\n");
580 		}
581 	} else if (!sva) {
582 		dev_warn(dev, "User forced SVA off via module param.\n");
583 	}
584 
585 	idxd_read_caps(idxd);
586 	idxd_read_table_offsets(idxd);
587 
588 	rc = idxd_setup_internals(idxd);
589 	if (rc)
590 		goto err;
591 
592 	/* If the configs are readonly, then load them from device */
593 	if (!test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags)) {
594 		dev_dbg(dev, "Loading RO device config\n");
595 		rc = idxd_device_load_config(idxd);
596 		if (rc < 0)
597 			goto err_config;
598 	}
599 
600 	rc = idxd_setup_interrupts(idxd);
601 	if (rc)
602 		goto err_config;
603 
604 	dev_dbg(dev, "IDXD interrupt setup complete.\n");
605 
606 	idxd->major = idxd_cdev_get_major(idxd);
607 
608 	rc = perfmon_pmu_init(idxd);
609 	if (rc < 0)
610 		dev_warn(dev, "Failed to initialize perfmon. No PMU support: %d\n", rc);
611 
612 	dev_dbg(dev, "IDXD device %d probed successfully\n", idxd->id);
613 	return 0;
614 
615  err_config:
616 	idxd_cleanup_internals(idxd);
617  err:
618 	if (device_pasid_enabled(idxd))
619 		idxd_disable_system_pasid(idxd);
620 	iommu_dev_disable_feature(dev, IOMMU_DEV_FEAT_SVA);
621 	return rc;
622 }
623 
624 static void idxd_cleanup(struct idxd_device *idxd)
625 {
626 	struct device *dev = &idxd->pdev->dev;
627 
628 	perfmon_pmu_remove(idxd);
629 	idxd_cleanup_interrupts(idxd);
630 	idxd_cleanup_internals(idxd);
631 	if (device_pasid_enabled(idxd))
632 		idxd_disable_system_pasid(idxd);
633 	iommu_dev_disable_feature(dev, IOMMU_DEV_FEAT_SVA);
634 }
635 
636 static int idxd_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
637 {
638 	struct device *dev = &pdev->dev;
639 	struct idxd_device *idxd;
640 	struct idxd_driver_data *data = (struct idxd_driver_data *)id->driver_data;
641 	int rc;
642 
643 	rc = pci_enable_device(pdev);
644 	if (rc)
645 		return rc;
646 
647 	dev_dbg(dev, "Alloc IDXD context\n");
648 	idxd = idxd_alloc(pdev, data);
649 	if (!idxd) {
650 		rc = -ENOMEM;
651 		goto err_idxd_alloc;
652 	}
653 
654 	dev_dbg(dev, "Mapping BARs\n");
655 	idxd->reg_base = pci_iomap(pdev, IDXD_MMIO_BAR, 0);
656 	if (!idxd->reg_base) {
657 		rc = -ENOMEM;
658 		goto err_iomap;
659 	}
660 
661 	dev_dbg(dev, "Set DMA masks\n");
662 	rc = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
663 	if (rc)
664 		rc = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32));
665 	if (rc)
666 		goto err;
667 
668 	dev_dbg(dev, "Set PCI master\n");
669 	pci_set_master(pdev);
670 	pci_set_drvdata(pdev, idxd);
671 
672 	idxd->hw.version = ioread32(idxd->reg_base + IDXD_VER_OFFSET);
673 	rc = idxd_probe(idxd);
674 	if (rc) {
675 		dev_err(dev, "Intel(R) IDXD DMA Engine init failed\n");
676 		goto err;
677 	}
678 
679 	rc = idxd_register_devices(idxd);
680 	if (rc) {
681 		dev_err(dev, "IDXD sysfs setup failed\n");
682 		goto err_dev_register;
683 	}
684 
685 	dev_info(&pdev->dev, "Intel(R) Accelerator Device (v%x)\n",
686 		 idxd->hw.version);
687 
688 	return 0;
689 
690  err_dev_register:
691 	idxd_cleanup(idxd);
692  err:
693 	pci_iounmap(pdev, idxd->reg_base);
694  err_iomap:
695 	put_device(idxd_confdev(idxd));
696  err_idxd_alloc:
697 	pci_disable_device(pdev);
698 	return rc;
699 }
700 
701 static void idxd_flush_pending_llist(struct idxd_irq_entry *ie)
702 {
703 	struct idxd_desc *desc, *itr;
704 	struct llist_node *head;
705 
706 	head = llist_del_all(&ie->pending_llist);
707 	if (!head)
708 		return;
709 
710 	llist_for_each_entry_safe(desc, itr, head, llnode) {
711 		idxd_dma_complete_txd(desc, IDXD_COMPLETE_ABORT);
712 		idxd_free_desc(desc->wq, desc);
713 	}
714 }
715 
716 static void idxd_flush_work_list(struct idxd_irq_entry *ie)
717 {
718 	struct idxd_desc *desc, *iter;
719 
720 	list_for_each_entry_safe(desc, iter, &ie->work_list, list) {
721 		list_del(&desc->list);
722 		idxd_dma_complete_txd(desc, IDXD_COMPLETE_ABORT);
723 		idxd_free_desc(desc->wq, desc);
724 	}
725 }
726 
727 void idxd_wqs_quiesce(struct idxd_device *idxd)
728 {
729 	struct idxd_wq *wq;
730 	int i;
731 
732 	for (i = 0; i < idxd->max_wqs; i++) {
733 		wq = idxd->wqs[i];
734 		if (wq->state == IDXD_WQ_ENABLED && wq->type == IDXD_WQT_KERNEL)
735 			idxd_wq_quiesce(wq);
736 	}
737 }
738 
739 static void idxd_release_int_handles(struct idxd_device *idxd)
740 {
741 	struct device *dev = &idxd->pdev->dev;
742 	int i, rc;
743 
744 	for (i = 0; i < idxd->num_wq_irqs; i++) {
745 		if (idxd->hw.cmd_cap & BIT(IDXD_CMD_RELEASE_INT_HANDLE)) {
746 			rc = idxd_device_release_int_handle(idxd, idxd->int_handles[i],
747 							    IDXD_IRQ_MSIX);
748 			if (rc < 0)
749 				dev_warn(dev, "irq handle %d release failed\n",
750 					 idxd->int_handles[i]);
751 			else
752 				dev_dbg(dev, "int handle requested: %u\n", idxd->int_handles[i]);
753 		}
754 	}
755 }
756 
757 static void idxd_shutdown(struct pci_dev *pdev)
758 {
759 	struct idxd_device *idxd = pci_get_drvdata(pdev);
760 	int rc, i;
761 	struct idxd_irq_entry *irq_entry;
762 	int msixcnt = pci_msix_vec_count(pdev);
763 
764 	rc = idxd_device_disable(idxd);
765 	if (rc)
766 		dev_err(&pdev->dev, "Disabling device failed\n");
767 
768 	dev_dbg(&pdev->dev, "%s called\n", __func__);
769 	idxd_mask_msix_vectors(idxd);
770 	idxd_mask_error_interrupts(idxd);
771 
772 	for (i = 0; i < msixcnt; i++) {
773 		irq_entry = &idxd->irq_entries[i];
774 		synchronize_irq(irq_entry->vector);
775 		if (i == 0)
776 			continue;
777 		idxd_flush_pending_llist(irq_entry);
778 		idxd_flush_work_list(irq_entry);
779 	}
780 	flush_workqueue(idxd->wq);
781 }
782 
783 static void idxd_remove(struct pci_dev *pdev)
784 {
785 	struct idxd_device *idxd = pci_get_drvdata(pdev);
786 	struct idxd_irq_entry *irq_entry;
787 	int msixcnt = pci_msix_vec_count(pdev);
788 	int i;
789 
790 	dev_dbg(&pdev->dev, "%s called\n", __func__);
791 	idxd_shutdown(pdev);
792 	if (device_pasid_enabled(idxd))
793 		idxd_disable_system_pasid(idxd);
794 	idxd_unregister_devices(idxd);
795 
796 	for (i = 0; i < msixcnt; i++) {
797 		irq_entry = &idxd->irq_entries[i];
798 		free_irq(irq_entry->vector, irq_entry);
799 	}
800 	idxd_msix_perm_clear(idxd);
801 	idxd_release_int_handles(idxd);
802 	pci_free_irq_vectors(pdev);
803 	pci_iounmap(pdev, idxd->reg_base);
804 	iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_SVA);
805 	pci_disable_device(pdev);
806 	destroy_workqueue(idxd->wq);
807 	perfmon_pmu_remove(idxd);
808 	device_unregister(idxd_confdev(idxd));
809 }
810 
811 static struct pci_driver idxd_pci_driver = {
812 	.name		= DRV_NAME,
813 	.id_table	= idxd_pci_tbl,
814 	.probe		= idxd_pci_probe,
815 	.remove		= idxd_remove,
816 	.shutdown	= idxd_shutdown,
817 };
818 
819 static int __init idxd_init_module(void)
820 {
821 	int err;
822 
823 	/*
824 	 * If the CPU does not support MOVDIR64B or ENQCMDS, there's no point in
825 	 * enumerating the device. We can not utilize it.
826 	 */
827 	if (!cpu_feature_enabled(X86_FEATURE_MOVDIR64B)) {
828 		pr_warn("idxd driver failed to load without MOVDIR64B.\n");
829 		return -ENODEV;
830 	}
831 
832 	if (!cpu_feature_enabled(X86_FEATURE_ENQCMD))
833 		pr_warn("Platform does not have ENQCMD(S) support.\n");
834 	else
835 		support_enqcmd = true;
836 
837 	perfmon_init();
838 
839 	err = idxd_register_bus_type();
840 	if (err < 0)
841 		return err;
842 
843 	err = idxd_driver_register(&idxd_drv);
844 	if (err < 0)
845 		goto err_idxd_driver_register;
846 
847 	err = idxd_driver_register(&dsa_drv);
848 	if (err < 0)
849 		goto err_dsa_driver_register;
850 
851 	err = idxd_cdev_register();
852 	if (err)
853 		goto err_cdev_register;
854 
855 	err = pci_register_driver(&idxd_pci_driver);
856 	if (err)
857 		goto err_pci_register;
858 
859 	return 0;
860 
861 err_pci_register:
862 	idxd_cdev_remove();
863 err_cdev_register:
864 	idxd_driver_unregister(&dsa_drv);
865 err_dsa_driver_register:
866 	idxd_driver_unregister(&idxd_drv);
867 err_idxd_driver_register:
868 	idxd_unregister_bus_type();
869 	return err;
870 }
871 module_init(idxd_init_module);
872 
873 static void __exit idxd_exit_module(void)
874 {
875 	idxd_driver_unregister(&idxd_drv);
876 	idxd_driver_unregister(&dsa_drv);
877 	pci_unregister_driver(&idxd_pci_driver);
878 	idxd_cdev_remove();
879 	idxd_unregister_bus_type();
880 	perfmon_exit();
881 }
882 module_exit(idxd_exit_module);
883 
884 int __idxd_driver_register(struct idxd_device_driver *idxd_drv, struct module *owner,
885 			   const char *mod_name)
886 {
887 	struct device_driver *drv = &idxd_drv->drv;
888 
889 	if (!idxd_drv->type) {
890 		pr_debug("driver type not set (%ps)\n", __builtin_return_address(0));
891 		return -EINVAL;
892 	}
893 
894 	drv->name = idxd_drv->name;
895 	drv->bus = &dsa_bus_type;
896 	drv->owner = owner;
897 	drv->mod_name = mod_name;
898 
899 	return driver_register(drv);
900 }
901 
902 void idxd_driver_unregister(struct idxd_device_driver *idxd_drv)
903 {
904 	driver_unregister(&idxd_drv->drv);
905 }
906