xref: /openbmc/linux/drivers/dma/idxd/init.c (revision fe7498ef)
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 MODULE_IMPORT_NS(IDXD);
30 
31 static bool sva = true;
32 module_param(sva, bool, 0644);
33 MODULE_PARM_DESC(sva, "Toggle SVA support on/off");
34 
35 bool tc_override;
36 module_param(tc_override, bool, 0644);
37 MODULE_PARM_DESC(tc_override, "Override traffic class defaults");
38 
39 #define DRV_NAME "idxd"
40 
41 bool support_enqcmd;
42 DEFINE_IDA(idxd_ida);
43 
44 static struct idxd_driver_data idxd_driver_data[] = {
45 	[IDXD_TYPE_DSA] = {
46 		.name_prefix = "dsa",
47 		.type = IDXD_TYPE_DSA,
48 		.compl_size = sizeof(struct dsa_completion_record),
49 		.align = 32,
50 		.dev_type = &dsa_device_type,
51 	},
52 	[IDXD_TYPE_IAX] = {
53 		.name_prefix = "iax",
54 		.type = IDXD_TYPE_IAX,
55 		.compl_size = sizeof(struct iax_completion_record),
56 		.align = 64,
57 		.dev_type = &iax_device_type,
58 	},
59 };
60 
61 static struct pci_device_id idxd_pci_tbl[] = {
62 	/* DSA ver 1.0 platforms */
63 	{ PCI_DEVICE_DATA(INTEL, DSA_SPR0, &idxd_driver_data[IDXD_TYPE_DSA]) },
64 
65 	/* IAX ver 1.0 platforms */
66 	{ PCI_DEVICE_DATA(INTEL, IAX_SPR0, &idxd_driver_data[IDXD_TYPE_IAX]) },
67 	{ 0, }
68 };
69 MODULE_DEVICE_TABLE(pci, idxd_pci_tbl);
70 
71 static int idxd_setup_interrupts(struct idxd_device *idxd)
72 {
73 	struct pci_dev *pdev = idxd->pdev;
74 	struct device *dev = &pdev->dev;
75 	struct idxd_irq_entry *irq_entry;
76 	int i, msixcnt;
77 	int rc = 0;
78 
79 	msixcnt = pci_msix_vec_count(pdev);
80 	if (msixcnt < 0) {
81 		dev_err(dev, "Not MSI-X interrupt capable.\n");
82 		return -ENOSPC;
83 	}
84 
85 	rc = pci_alloc_irq_vectors(pdev, msixcnt, msixcnt, PCI_IRQ_MSIX);
86 	if (rc != msixcnt) {
87 		dev_err(dev, "Failed enabling %d MSIX entries: %d\n", msixcnt, rc);
88 		return -ENOSPC;
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 = kcalloc_node(msixcnt, sizeof(struct idxd_irq_entry),
97 					 GFP_KERNEL, dev_to_node(dev));
98 	if (!idxd->irq_entries) {
99 		rc = -ENOMEM;
100 		goto err_irq_entries;
101 	}
102 
103 	for (i = 0; i < msixcnt; i++) {
104 		idxd->irq_entries[i].id = i;
105 		idxd->irq_entries[i].idxd = idxd;
106 		idxd->irq_entries[i].vector = pci_irq_vector(pdev, i);
107 		spin_lock_init(&idxd->irq_entries[i].list_lock);
108 	}
109 
110 	idxd_msix_perm_setup(idxd);
111 
112 	irq_entry = &idxd->irq_entries[0];
113 	rc = request_threaded_irq(irq_entry->vector, NULL, idxd_misc_thread,
114 				  0, "idxd-misc", irq_entry);
115 	if (rc < 0) {
116 		dev_err(dev, "Failed to allocate misc interrupt.\n");
117 		goto err_misc_irq;
118 	}
119 
120 	dev_dbg(dev, "Allocated idxd-misc handler on msix vector %d\n", irq_entry->vector);
121 
122 	/* first MSI-X entry is not for wq interrupts */
123 	idxd->num_wq_irqs = msixcnt - 1;
124 
125 	for (i = 1; i < msixcnt; i++) {
126 		irq_entry = &idxd->irq_entries[i];
127 
128 		init_llist_head(&idxd->irq_entries[i].pending_llist);
129 		INIT_LIST_HEAD(&idxd->irq_entries[i].work_list);
130 		rc = request_threaded_irq(irq_entry->vector, NULL,
131 					  idxd_wq_thread, 0, "idxd-portal", irq_entry);
132 		if (rc < 0) {
133 			dev_err(dev, "Failed to allocate irq %d.\n", irq_entry->vector);
134 			goto err_wq_irqs;
135 		}
136 
137 		dev_dbg(dev, "Allocated idxd-msix %d for vector %d\n", i, irq_entry->vector);
138 		if (idxd->hw.cmd_cap & BIT(IDXD_CMD_REQUEST_INT_HANDLE)) {
139 			/*
140 			 * The MSIX vector enumeration starts at 1 with vector 0 being the
141 			 * misc interrupt that handles non I/O completion events. The
142 			 * interrupt handles are for IMS enumeration on guest. The misc
143 			 * interrupt vector does not require a handle and therefore we start
144 			 * the int_handles at index 0. Since 'i' starts at 1, the first
145 			 * int_handles index will be 0.
146 			 */
147 			rc = idxd_device_request_int_handle(idxd, i, &idxd->int_handles[i - 1],
148 							    IDXD_IRQ_MSIX);
149 			if (rc < 0) {
150 				free_irq(irq_entry->vector, irq_entry);
151 				goto err_wq_irqs;
152 			}
153 			dev_dbg(dev, "int handle requested: %u\n", idxd->int_handles[i - 1]);
154 		}
155 	}
156 
157 	idxd_unmask_error_interrupts(idxd);
158 	return 0;
159 
160  err_wq_irqs:
161 	while (--i >= 0) {
162 		irq_entry = &idxd->irq_entries[i];
163 		free_irq(irq_entry->vector, irq_entry);
164 		if (i != 0)
165 			idxd_device_release_int_handle(idxd,
166 						       idxd->int_handles[i], IDXD_IRQ_MSIX);
167 	}
168  err_misc_irq:
169 	/* Disable error interrupt generation */
170 	idxd_mask_error_interrupts(idxd);
171 	idxd_msix_perm_clear(idxd);
172  err_irq_entries:
173 	pci_free_irq_vectors(pdev);
174 	dev_err(dev, "No usable interrupts\n");
175 	return rc;
176 }
177 
178 static void idxd_cleanup_interrupts(struct idxd_device *idxd)
179 {
180 	struct pci_dev *pdev = idxd->pdev;
181 	struct idxd_irq_entry *irq_entry;
182 	int i, msixcnt;
183 
184 	msixcnt = pci_msix_vec_count(pdev);
185 	if (msixcnt <= 0)
186 		return;
187 
188 	irq_entry = &idxd->irq_entries[0];
189 	free_irq(irq_entry->vector, irq_entry);
190 
191 	for (i = 1; i < msixcnt; i++) {
192 
193 		irq_entry = &idxd->irq_entries[i];
194 		if (idxd->hw.cmd_cap & BIT(IDXD_CMD_RELEASE_INT_HANDLE))
195 			idxd_device_release_int_handle(idxd, idxd->int_handles[i],
196 						       IDXD_IRQ_MSIX);
197 		free_irq(irq_entry->vector, irq_entry);
198 	}
199 
200 	idxd_mask_error_interrupts(idxd);
201 	pci_free_irq_vectors(pdev);
202 }
203 
204 static int idxd_setup_wqs(struct idxd_device *idxd)
205 {
206 	struct device *dev = &idxd->pdev->dev;
207 	struct idxd_wq *wq;
208 	struct device *conf_dev;
209 	int i, rc;
210 
211 	idxd->wqs = kcalloc_node(idxd->max_wqs, sizeof(struct idxd_wq *),
212 				 GFP_KERNEL, dev_to_node(dev));
213 	if (!idxd->wqs)
214 		return -ENOMEM;
215 
216 	for (i = 0; i < idxd->max_wqs; i++) {
217 		wq = kzalloc_node(sizeof(*wq), GFP_KERNEL, dev_to_node(dev));
218 		if (!wq) {
219 			rc = -ENOMEM;
220 			goto err;
221 		}
222 
223 		idxd_dev_set_type(&wq->idxd_dev, IDXD_DEV_WQ);
224 		conf_dev = wq_confdev(wq);
225 		wq->id = i;
226 		wq->idxd = idxd;
227 		device_initialize(wq_confdev(wq));
228 		conf_dev->parent = idxd_confdev(idxd);
229 		conf_dev->bus = &dsa_bus_type;
230 		conf_dev->type = &idxd_wq_device_type;
231 		rc = dev_set_name(conf_dev, "wq%d.%d", idxd->id, wq->id);
232 		if (rc < 0) {
233 			put_device(conf_dev);
234 			goto err;
235 		}
236 
237 		mutex_init(&wq->wq_lock);
238 		init_waitqueue_head(&wq->err_queue);
239 		init_completion(&wq->wq_dead);
240 		wq->max_xfer_bytes = idxd->max_xfer_bytes;
241 		wq->max_batch_size = idxd->max_batch_size;
242 		wq->wqcfg = kzalloc_node(idxd->wqcfg_size, GFP_KERNEL, dev_to_node(dev));
243 		if (!wq->wqcfg) {
244 			put_device(conf_dev);
245 			rc = -ENOMEM;
246 			goto err;
247 		}
248 		idxd->wqs[i] = wq;
249 	}
250 
251 	return 0;
252 
253  err:
254 	while (--i >= 0) {
255 		wq = idxd->wqs[i];
256 		conf_dev = wq_confdev(wq);
257 		put_device(conf_dev);
258 	}
259 	return rc;
260 }
261 
262 static int idxd_setup_engines(struct idxd_device *idxd)
263 {
264 	struct idxd_engine *engine;
265 	struct device *dev = &idxd->pdev->dev;
266 	struct device *conf_dev;
267 	int i, rc;
268 
269 	idxd->engines = kcalloc_node(idxd->max_engines, sizeof(struct idxd_engine *),
270 				     GFP_KERNEL, dev_to_node(dev));
271 	if (!idxd->engines)
272 		return -ENOMEM;
273 
274 	for (i = 0; i < idxd->max_engines; i++) {
275 		engine = kzalloc_node(sizeof(*engine), GFP_KERNEL, dev_to_node(dev));
276 		if (!engine) {
277 			rc = -ENOMEM;
278 			goto err;
279 		}
280 
281 		idxd_dev_set_type(&engine->idxd_dev, IDXD_DEV_ENGINE);
282 		conf_dev = engine_confdev(engine);
283 		engine->id = i;
284 		engine->idxd = idxd;
285 		device_initialize(conf_dev);
286 		conf_dev->parent = idxd_confdev(idxd);
287 		conf_dev->bus = &dsa_bus_type;
288 		conf_dev->type = &idxd_engine_device_type;
289 		rc = dev_set_name(conf_dev, "engine%d.%d", idxd->id, engine->id);
290 		if (rc < 0) {
291 			put_device(conf_dev);
292 			goto err;
293 		}
294 
295 		idxd->engines[i] = engine;
296 	}
297 
298 	return 0;
299 
300  err:
301 	while (--i >= 0) {
302 		engine = idxd->engines[i];
303 		conf_dev = engine_confdev(engine);
304 		put_device(conf_dev);
305 	}
306 	return rc;
307 }
308 
309 static int idxd_setup_groups(struct idxd_device *idxd)
310 {
311 	struct device *dev = &idxd->pdev->dev;
312 	struct device *conf_dev;
313 	struct idxd_group *group;
314 	int i, rc;
315 
316 	idxd->groups = kcalloc_node(idxd->max_groups, sizeof(struct idxd_group *),
317 				    GFP_KERNEL, dev_to_node(dev));
318 	if (!idxd->groups)
319 		return -ENOMEM;
320 
321 	for (i = 0; i < idxd->max_groups; i++) {
322 		group = kzalloc_node(sizeof(*group), GFP_KERNEL, dev_to_node(dev));
323 		if (!group) {
324 			rc = -ENOMEM;
325 			goto err;
326 		}
327 
328 		idxd_dev_set_type(&group->idxd_dev, IDXD_DEV_GROUP);
329 		conf_dev = group_confdev(group);
330 		group->id = i;
331 		group->idxd = idxd;
332 		device_initialize(conf_dev);
333 		conf_dev->parent = idxd_confdev(idxd);
334 		conf_dev->bus = &dsa_bus_type;
335 		conf_dev->type = &idxd_group_device_type;
336 		rc = dev_set_name(conf_dev, "group%d.%d", idxd->id, group->id);
337 		if (rc < 0) {
338 			put_device(conf_dev);
339 			goto err;
340 		}
341 
342 		idxd->groups[i] = group;
343 		if (idxd->hw.version < DEVICE_VERSION_2 && !tc_override) {
344 			group->tc_a = 1;
345 			group->tc_b = 1;
346 		} else {
347 			group->tc_a = -1;
348 			group->tc_b = -1;
349 		}
350 	}
351 
352 	return 0;
353 
354  err:
355 	while (--i >= 0) {
356 		group = idxd->groups[i];
357 		put_device(group_confdev(group));
358 	}
359 	return rc;
360 }
361 
362 static void idxd_cleanup_internals(struct idxd_device *idxd)
363 {
364 	int i;
365 
366 	for (i = 0; i < idxd->max_groups; i++)
367 		put_device(group_confdev(idxd->groups[i]));
368 	for (i = 0; i < idxd->max_engines; i++)
369 		put_device(engine_confdev(idxd->engines[i]));
370 	for (i = 0; i < idxd->max_wqs; i++)
371 		put_device(wq_confdev(idxd->wqs[i]));
372 	destroy_workqueue(idxd->wq);
373 }
374 
375 static int idxd_setup_internals(struct idxd_device *idxd)
376 {
377 	struct device *dev = &idxd->pdev->dev;
378 	int rc, i;
379 
380 	init_waitqueue_head(&idxd->cmd_waitq);
381 
382 	if (idxd->hw.cmd_cap & BIT(IDXD_CMD_REQUEST_INT_HANDLE)) {
383 		idxd->int_handles = kcalloc_node(idxd->max_wqs, sizeof(int), GFP_KERNEL,
384 						 dev_to_node(dev));
385 		if (!idxd->int_handles)
386 			return -ENOMEM;
387 	}
388 
389 	rc = idxd_setup_wqs(idxd);
390 	if (rc < 0)
391 		goto err_wqs;
392 
393 	rc = idxd_setup_engines(idxd);
394 	if (rc < 0)
395 		goto err_engine;
396 
397 	rc = idxd_setup_groups(idxd);
398 	if (rc < 0)
399 		goto err_group;
400 
401 	idxd->wq = create_workqueue(dev_name(dev));
402 	if (!idxd->wq) {
403 		rc = -ENOMEM;
404 		goto err_wkq_create;
405 	}
406 
407 	return 0;
408 
409  err_wkq_create:
410 	for (i = 0; i < idxd->max_groups; i++)
411 		put_device(group_confdev(idxd->groups[i]));
412  err_group:
413 	for (i = 0; i < idxd->max_engines; i++)
414 		put_device(engine_confdev(idxd->engines[i]));
415  err_engine:
416 	for (i = 0; i < idxd->max_wqs; i++)
417 		put_device(wq_confdev(idxd->wqs[i]));
418  err_wqs:
419 	kfree(idxd->int_handles);
420 	return rc;
421 }
422 
423 static void idxd_read_table_offsets(struct idxd_device *idxd)
424 {
425 	union offsets_reg offsets;
426 	struct device *dev = &idxd->pdev->dev;
427 
428 	offsets.bits[0] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET);
429 	offsets.bits[1] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET + sizeof(u64));
430 	idxd->grpcfg_offset = offsets.grpcfg * IDXD_TABLE_MULT;
431 	dev_dbg(dev, "IDXD Group Config Offset: %#x\n", idxd->grpcfg_offset);
432 	idxd->wqcfg_offset = offsets.wqcfg * IDXD_TABLE_MULT;
433 	dev_dbg(dev, "IDXD Work Queue Config Offset: %#x\n", idxd->wqcfg_offset);
434 	idxd->msix_perm_offset = offsets.msix_perm * IDXD_TABLE_MULT;
435 	dev_dbg(dev, "IDXD MSIX Permission Offset: %#x\n", idxd->msix_perm_offset);
436 	idxd->perfmon_offset = offsets.perfmon * IDXD_TABLE_MULT;
437 	dev_dbg(dev, "IDXD Perfmon Offset: %#x\n", idxd->perfmon_offset);
438 }
439 
440 static void idxd_read_caps(struct idxd_device *idxd)
441 {
442 	struct device *dev = &idxd->pdev->dev;
443 	int i;
444 
445 	/* reading generic capabilities */
446 	idxd->hw.gen_cap.bits = ioread64(idxd->reg_base + IDXD_GENCAP_OFFSET);
447 	dev_dbg(dev, "gen_cap: %#llx\n", idxd->hw.gen_cap.bits);
448 
449 	if (idxd->hw.gen_cap.cmd_cap) {
450 		idxd->hw.cmd_cap = ioread32(idxd->reg_base + IDXD_CMDCAP_OFFSET);
451 		dev_dbg(dev, "cmd_cap: %#x\n", idxd->hw.cmd_cap);
452 	}
453 
454 	idxd->max_xfer_bytes = 1ULL << idxd->hw.gen_cap.max_xfer_shift;
455 	dev_dbg(dev, "max xfer size: %llu bytes\n", idxd->max_xfer_bytes);
456 	idxd->max_batch_size = 1U << idxd->hw.gen_cap.max_batch_shift;
457 	dev_dbg(dev, "max batch size: %u\n", idxd->max_batch_size);
458 	if (idxd->hw.gen_cap.config_en)
459 		set_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags);
460 
461 	/* reading group capabilities */
462 	idxd->hw.group_cap.bits =
463 		ioread64(idxd->reg_base + IDXD_GRPCAP_OFFSET);
464 	dev_dbg(dev, "group_cap: %#llx\n", idxd->hw.group_cap.bits);
465 	idxd->max_groups = idxd->hw.group_cap.num_groups;
466 	dev_dbg(dev, "max groups: %u\n", idxd->max_groups);
467 	idxd->max_tokens = idxd->hw.group_cap.total_tokens;
468 	dev_dbg(dev, "max tokens: %u\n", idxd->max_tokens);
469 	idxd->nr_tokens = idxd->max_tokens;
470 
471 	/* read engine capabilities */
472 	idxd->hw.engine_cap.bits =
473 		ioread64(idxd->reg_base + IDXD_ENGCAP_OFFSET);
474 	dev_dbg(dev, "engine_cap: %#llx\n", idxd->hw.engine_cap.bits);
475 	idxd->max_engines = idxd->hw.engine_cap.num_engines;
476 	dev_dbg(dev, "max engines: %u\n", idxd->max_engines);
477 
478 	/* read workqueue capabilities */
479 	idxd->hw.wq_cap.bits = ioread64(idxd->reg_base + IDXD_WQCAP_OFFSET);
480 	dev_dbg(dev, "wq_cap: %#llx\n", idxd->hw.wq_cap.bits);
481 	idxd->max_wq_size = idxd->hw.wq_cap.total_wq_size;
482 	dev_dbg(dev, "total workqueue size: %u\n", idxd->max_wq_size);
483 	idxd->max_wqs = idxd->hw.wq_cap.num_wqs;
484 	dev_dbg(dev, "max workqueues: %u\n", idxd->max_wqs);
485 	idxd->wqcfg_size = 1 << (idxd->hw.wq_cap.wqcfg_size + IDXD_WQCFG_MIN);
486 	dev_dbg(dev, "wqcfg size: %u\n", idxd->wqcfg_size);
487 
488 	/* reading operation capabilities */
489 	for (i = 0; i < 4; i++) {
490 		idxd->hw.opcap.bits[i] = ioread64(idxd->reg_base +
491 				IDXD_OPCAP_OFFSET + i * sizeof(u64));
492 		dev_dbg(dev, "opcap[%d]: %#llx\n", i, idxd->hw.opcap.bits[i]);
493 	}
494 }
495 
496 static struct idxd_device *idxd_alloc(struct pci_dev *pdev, struct idxd_driver_data *data)
497 {
498 	struct device *dev = &pdev->dev;
499 	struct device *conf_dev;
500 	struct idxd_device *idxd;
501 	int rc;
502 
503 	idxd = kzalloc_node(sizeof(*idxd), GFP_KERNEL, dev_to_node(dev));
504 	if (!idxd)
505 		return NULL;
506 
507 	conf_dev = idxd_confdev(idxd);
508 	idxd->pdev = pdev;
509 	idxd->data = data;
510 	idxd_dev_set_type(&idxd->idxd_dev, idxd->data->type);
511 	idxd->id = ida_alloc(&idxd_ida, GFP_KERNEL);
512 	if (idxd->id < 0)
513 		return NULL;
514 
515 	device_initialize(conf_dev);
516 	conf_dev->parent = dev;
517 	conf_dev->bus = &dsa_bus_type;
518 	conf_dev->type = idxd->data->dev_type;
519 	rc = dev_set_name(conf_dev, "%s%d", idxd->data->name_prefix, idxd->id);
520 	if (rc < 0) {
521 		put_device(conf_dev);
522 		return NULL;
523 	}
524 
525 	spin_lock_init(&idxd->dev_lock);
526 	spin_lock_init(&idxd->cmd_lock);
527 
528 	return idxd;
529 }
530 
531 static int idxd_enable_system_pasid(struct idxd_device *idxd)
532 {
533 	int flags;
534 	unsigned int pasid;
535 	struct iommu_sva *sva;
536 
537 	flags = SVM_FLAG_SUPERVISOR_MODE;
538 
539 	sva = iommu_sva_bind_device(&idxd->pdev->dev, NULL, &flags);
540 	if (IS_ERR(sva)) {
541 		dev_warn(&idxd->pdev->dev,
542 			 "iommu sva bind failed: %ld\n", PTR_ERR(sva));
543 		return PTR_ERR(sva);
544 	}
545 
546 	pasid = iommu_sva_get_pasid(sva);
547 	if (pasid == IOMMU_PASID_INVALID) {
548 		iommu_sva_unbind_device(sva);
549 		return -ENODEV;
550 	}
551 
552 	idxd->sva = sva;
553 	idxd->pasid = pasid;
554 	dev_dbg(&idxd->pdev->dev, "system pasid: %u\n", pasid);
555 	return 0;
556 }
557 
558 static void idxd_disable_system_pasid(struct idxd_device *idxd)
559 {
560 
561 	iommu_sva_unbind_device(idxd->sva);
562 	idxd->sva = NULL;
563 }
564 
565 static int idxd_probe(struct idxd_device *idxd)
566 {
567 	struct pci_dev *pdev = idxd->pdev;
568 	struct device *dev = &pdev->dev;
569 	int rc;
570 
571 	dev_dbg(dev, "%s entered and resetting device\n", __func__);
572 	rc = idxd_device_init_reset(idxd);
573 	if (rc < 0)
574 		return rc;
575 
576 	dev_dbg(dev, "IDXD reset complete\n");
577 
578 	if (IS_ENABLED(CONFIG_INTEL_IDXD_SVM) && sva) {
579 		rc = iommu_dev_enable_feature(dev, IOMMU_DEV_FEAT_SVA);
580 		if (rc == 0) {
581 			rc = idxd_enable_system_pasid(idxd);
582 			if (rc < 0) {
583 				iommu_dev_disable_feature(dev, IOMMU_DEV_FEAT_SVA);
584 				dev_warn(dev, "Failed to enable PASID. No SVA support: %d\n", rc);
585 			} else {
586 				set_bit(IDXD_FLAG_PASID_ENABLED, &idxd->flags);
587 			}
588 		} else {
589 			dev_warn(dev, "Unable to turn on SVA feature.\n");
590 		}
591 	} else if (!sva) {
592 		dev_warn(dev, "User forced SVA off via module param.\n");
593 	}
594 
595 	idxd_read_caps(idxd);
596 	idxd_read_table_offsets(idxd);
597 
598 	rc = idxd_setup_internals(idxd);
599 	if (rc)
600 		goto err;
601 
602 	/* If the configs are readonly, then load them from device */
603 	if (!test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags)) {
604 		dev_dbg(dev, "Loading RO device config\n");
605 		rc = idxd_device_load_config(idxd);
606 		if (rc < 0)
607 			goto err_config;
608 	}
609 
610 	rc = idxd_setup_interrupts(idxd);
611 	if (rc)
612 		goto err_config;
613 
614 	dev_dbg(dev, "IDXD interrupt setup complete.\n");
615 
616 	idxd->major = idxd_cdev_get_major(idxd);
617 
618 	rc = perfmon_pmu_init(idxd);
619 	if (rc < 0)
620 		dev_warn(dev, "Failed to initialize perfmon. No PMU support: %d\n", rc);
621 
622 	dev_dbg(dev, "IDXD device %d probed successfully\n", idxd->id);
623 	return 0;
624 
625  err_config:
626 	idxd_cleanup_internals(idxd);
627  err:
628 	if (device_pasid_enabled(idxd))
629 		idxd_disable_system_pasid(idxd);
630 	iommu_dev_disable_feature(dev, IOMMU_DEV_FEAT_SVA);
631 	return rc;
632 }
633 
634 static void idxd_cleanup(struct idxd_device *idxd)
635 {
636 	struct device *dev = &idxd->pdev->dev;
637 
638 	perfmon_pmu_remove(idxd);
639 	idxd_cleanup_interrupts(idxd);
640 	idxd_cleanup_internals(idxd);
641 	if (device_pasid_enabled(idxd))
642 		idxd_disable_system_pasid(idxd);
643 	iommu_dev_disable_feature(dev, IOMMU_DEV_FEAT_SVA);
644 }
645 
646 static int idxd_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
647 {
648 	struct device *dev = &pdev->dev;
649 	struct idxd_device *idxd;
650 	struct idxd_driver_data *data = (struct idxd_driver_data *)id->driver_data;
651 	int rc;
652 
653 	rc = pci_enable_device(pdev);
654 	if (rc)
655 		return rc;
656 
657 	dev_dbg(dev, "Alloc IDXD context\n");
658 	idxd = idxd_alloc(pdev, data);
659 	if (!idxd) {
660 		rc = -ENOMEM;
661 		goto err_idxd_alloc;
662 	}
663 
664 	dev_dbg(dev, "Mapping BARs\n");
665 	idxd->reg_base = pci_iomap(pdev, IDXD_MMIO_BAR, 0);
666 	if (!idxd->reg_base) {
667 		rc = -ENOMEM;
668 		goto err_iomap;
669 	}
670 
671 	dev_dbg(dev, "Set DMA masks\n");
672 	rc = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
673 	if (rc)
674 		rc = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32));
675 	if (rc)
676 		goto err;
677 
678 	dev_dbg(dev, "Set PCI master\n");
679 	pci_set_master(pdev);
680 	pci_set_drvdata(pdev, idxd);
681 
682 	idxd->hw.version = ioread32(idxd->reg_base + IDXD_VER_OFFSET);
683 	rc = idxd_probe(idxd);
684 	if (rc) {
685 		dev_err(dev, "Intel(R) IDXD DMA Engine init failed\n");
686 		goto err;
687 	}
688 
689 	rc = idxd_register_devices(idxd);
690 	if (rc) {
691 		dev_err(dev, "IDXD sysfs setup failed\n");
692 		goto err_dev_register;
693 	}
694 
695 	dev_info(&pdev->dev, "Intel(R) Accelerator Device (v%x)\n",
696 		 idxd->hw.version);
697 
698 	return 0;
699 
700  err_dev_register:
701 	idxd_cleanup(idxd);
702  err:
703 	pci_iounmap(pdev, idxd->reg_base);
704  err_iomap:
705 	put_device(idxd_confdev(idxd));
706  err_idxd_alloc:
707 	pci_disable_device(pdev);
708 	return rc;
709 }
710 
711 static void idxd_flush_pending_llist(struct idxd_irq_entry *ie)
712 {
713 	struct idxd_desc *desc, *itr;
714 	struct llist_node *head;
715 
716 	head = llist_del_all(&ie->pending_llist);
717 	if (!head)
718 		return;
719 
720 	llist_for_each_entry_safe(desc, itr, head, llnode) {
721 		idxd_dma_complete_txd(desc, IDXD_COMPLETE_ABORT);
722 		idxd_free_desc(desc->wq, desc);
723 	}
724 }
725 
726 static void idxd_flush_work_list(struct idxd_irq_entry *ie)
727 {
728 	struct idxd_desc *desc, *iter;
729 
730 	list_for_each_entry_safe(desc, iter, &ie->work_list, list) {
731 		list_del(&desc->list);
732 		idxd_dma_complete_txd(desc, IDXD_COMPLETE_ABORT);
733 		idxd_free_desc(desc->wq, desc);
734 	}
735 }
736 
737 void idxd_wqs_quiesce(struct idxd_device *idxd)
738 {
739 	struct idxd_wq *wq;
740 	int i;
741 
742 	for (i = 0; i < idxd->max_wqs; i++) {
743 		wq = idxd->wqs[i];
744 		if (wq->state == IDXD_WQ_ENABLED && wq->type == IDXD_WQT_KERNEL)
745 			idxd_wq_quiesce(wq);
746 	}
747 }
748 
749 static void idxd_release_int_handles(struct idxd_device *idxd)
750 {
751 	struct device *dev = &idxd->pdev->dev;
752 	int i, rc;
753 
754 	for (i = 0; i < idxd->num_wq_irqs; i++) {
755 		if (idxd->hw.cmd_cap & BIT(IDXD_CMD_RELEASE_INT_HANDLE)) {
756 			rc = idxd_device_release_int_handle(idxd, idxd->int_handles[i],
757 							    IDXD_IRQ_MSIX);
758 			if (rc < 0)
759 				dev_warn(dev, "irq handle %d release failed\n",
760 					 idxd->int_handles[i]);
761 			else
762 				dev_dbg(dev, "int handle requested: %u\n", idxd->int_handles[i]);
763 		}
764 	}
765 }
766 
767 static void idxd_shutdown(struct pci_dev *pdev)
768 {
769 	struct idxd_device *idxd = pci_get_drvdata(pdev);
770 	int rc, i;
771 	struct idxd_irq_entry *irq_entry;
772 	int msixcnt = pci_msix_vec_count(pdev);
773 
774 	rc = idxd_device_disable(idxd);
775 	if (rc)
776 		dev_err(&pdev->dev, "Disabling device failed\n");
777 
778 	dev_dbg(&pdev->dev, "%s called\n", __func__);
779 	idxd_mask_msix_vectors(idxd);
780 	idxd_mask_error_interrupts(idxd);
781 
782 	for (i = 0; i < msixcnt; i++) {
783 		irq_entry = &idxd->irq_entries[i];
784 		synchronize_irq(irq_entry->vector);
785 		if (i == 0)
786 			continue;
787 		idxd_flush_pending_llist(irq_entry);
788 		idxd_flush_work_list(irq_entry);
789 	}
790 	flush_workqueue(idxd->wq);
791 }
792 
793 static void idxd_remove(struct pci_dev *pdev)
794 {
795 	struct idxd_device *idxd = pci_get_drvdata(pdev);
796 	struct idxd_irq_entry *irq_entry;
797 	int msixcnt = pci_msix_vec_count(pdev);
798 	int i;
799 
800 	idxd_unregister_devices(idxd);
801 	/*
802 	 * When ->release() is called for the idxd->conf_dev, it frees all the memory related
803 	 * to the idxd context. The driver still needs those bits in order to do the rest of
804 	 * the cleanup. However, we do need to unbound the idxd sub-driver. So take a ref
805 	 * on the device here to hold off the freeing while allowing the idxd sub-driver
806 	 * to unbind.
807 	 */
808 	get_device(idxd_confdev(idxd));
809 	device_unregister(idxd_confdev(idxd));
810 	idxd_shutdown(pdev);
811 	if (device_pasid_enabled(idxd))
812 		idxd_disable_system_pasid(idxd);
813 
814 	for (i = 0; i < msixcnt; i++) {
815 		irq_entry = &idxd->irq_entries[i];
816 		free_irq(irq_entry->vector, irq_entry);
817 	}
818 	idxd_msix_perm_clear(idxd);
819 	idxd_release_int_handles(idxd);
820 	pci_free_irq_vectors(pdev);
821 	pci_iounmap(pdev, idxd->reg_base);
822 	iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_SVA);
823 	pci_disable_device(pdev);
824 	destroy_workqueue(idxd->wq);
825 	perfmon_pmu_remove(idxd);
826 	put_device(idxd_confdev(idxd));
827 }
828 
829 static struct pci_driver idxd_pci_driver = {
830 	.name		= DRV_NAME,
831 	.id_table	= idxd_pci_tbl,
832 	.probe		= idxd_pci_probe,
833 	.remove		= idxd_remove,
834 	.shutdown	= idxd_shutdown,
835 };
836 
837 static int __init idxd_init_module(void)
838 {
839 	int err;
840 
841 	/*
842 	 * If the CPU does not support MOVDIR64B or ENQCMDS, there's no point in
843 	 * enumerating the device. We can not utilize it.
844 	 */
845 	if (!cpu_feature_enabled(X86_FEATURE_MOVDIR64B)) {
846 		pr_warn("idxd driver failed to load without MOVDIR64B.\n");
847 		return -ENODEV;
848 	}
849 
850 	if (!cpu_feature_enabled(X86_FEATURE_ENQCMD))
851 		pr_warn("Platform does not have ENQCMD(S) support.\n");
852 	else
853 		support_enqcmd = true;
854 
855 	perfmon_init();
856 
857 	err = idxd_driver_register(&idxd_drv);
858 	if (err < 0)
859 		goto err_idxd_driver_register;
860 
861 	err = idxd_driver_register(&idxd_dmaengine_drv);
862 	if (err < 0)
863 		goto err_idxd_dmaengine_driver_register;
864 
865 	err = idxd_driver_register(&idxd_user_drv);
866 	if (err < 0)
867 		goto err_idxd_user_driver_register;
868 
869 	err = idxd_cdev_register();
870 	if (err)
871 		goto err_cdev_register;
872 
873 	err = pci_register_driver(&idxd_pci_driver);
874 	if (err)
875 		goto err_pci_register;
876 
877 	return 0;
878 
879 err_pci_register:
880 	idxd_cdev_remove();
881 err_cdev_register:
882 	idxd_driver_unregister(&idxd_user_drv);
883 err_idxd_user_driver_register:
884 	idxd_driver_unregister(&idxd_dmaengine_drv);
885 err_idxd_dmaengine_driver_register:
886 	idxd_driver_unregister(&idxd_drv);
887 err_idxd_driver_register:
888 	return err;
889 }
890 module_init(idxd_init_module);
891 
892 static void __exit idxd_exit_module(void)
893 {
894 	idxd_driver_unregister(&idxd_user_drv);
895 	idxd_driver_unregister(&idxd_dmaengine_drv);
896 	idxd_driver_unregister(&idxd_drv);
897 	pci_unregister_driver(&idxd_pci_driver);
898 	idxd_cdev_remove();
899 	perfmon_exit();
900 }
901 module_exit(idxd_exit_module);
902