xref: /openbmc/linux/drivers/dma/idxd/init.c (revision 79e790ff)
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 	irq_entry = &idxd->irq_entries[0];
106 	rc = request_threaded_irq(irq_entry->vector, NULL, idxd_misc_thread,
107 				  0, "idxd-misc", irq_entry);
108 	if (rc < 0) {
109 		dev_err(dev, "Failed to allocate misc interrupt.\n");
110 		goto err_misc_irq;
111 	}
112 
113 	dev_dbg(dev, "Allocated idxd-misc handler on msix vector %d\n", irq_entry->vector);
114 
115 	/* first MSI-X entry is not for wq interrupts */
116 	idxd->num_wq_irqs = msixcnt - 1;
117 
118 	for (i = 1; i < msixcnt; i++) {
119 		irq_entry = &idxd->irq_entries[i];
120 
121 		init_llist_head(&idxd->irq_entries[i].pending_llist);
122 		INIT_LIST_HEAD(&idxd->irq_entries[i].work_list);
123 		rc = request_threaded_irq(irq_entry->vector, NULL,
124 					  idxd_wq_thread, 0, "idxd-portal", irq_entry);
125 		if (rc < 0) {
126 			dev_err(dev, "Failed to allocate irq %d.\n", irq_entry->vector);
127 			goto err_wq_irqs;
128 		}
129 
130 		dev_dbg(dev, "Allocated idxd-msix %d for vector %d\n", i, irq_entry->vector);
131 		if (idxd->hw.cmd_cap & BIT(IDXD_CMD_REQUEST_INT_HANDLE)) {
132 			/*
133 			 * The MSIX vector enumeration starts at 1 with vector 0 being the
134 			 * misc interrupt that handles non I/O completion events. The
135 			 * interrupt handles are for IMS enumeration on guest. The misc
136 			 * interrupt vector does not require a handle and therefore we start
137 			 * the int_handles at index 0. Since 'i' starts at 1, the first
138 			 * int_handles index will be 0.
139 			 */
140 			rc = idxd_device_request_int_handle(idxd, i, &idxd->int_handles[i - 1],
141 							    IDXD_IRQ_MSIX);
142 			if (rc < 0) {
143 				free_irq(irq_entry->vector, irq_entry);
144 				goto err_wq_irqs;
145 			}
146 			dev_dbg(dev, "int handle requested: %u\n", idxd->int_handles[i - 1]);
147 		}
148 	}
149 
150 	idxd_unmask_error_interrupts(idxd);
151 	idxd_msix_perm_setup(idxd);
152 	return 0;
153 
154  err_wq_irqs:
155 	while (--i >= 0) {
156 		irq_entry = &idxd->irq_entries[i];
157 		free_irq(irq_entry->vector, irq_entry);
158 		if (i != 0)
159 			idxd_device_release_int_handle(idxd,
160 						       idxd->int_handles[i], IDXD_IRQ_MSIX);
161 	}
162  err_misc_irq:
163 	/* Disable error interrupt generation */
164 	idxd_mask_error_interrupts(idxd);
165  err_irq_entries:
166 	pci_free_irq_vectors(pdev);
167 	dev_err(dev, "No usable interrupts\n");
168 	return rc;
169 }
170 
171 static int idxd_setup_wqs(struct idxd_device *idxd)
172 {
173 	struct device *dev = &idxd->pdev->dev;
174 	struct idxd_wq *wq;
175 	int i, rc;
176 
177 	idxd->wqs = kcalloc_node(idxd->max_wqs, sizeof(struct idxd_wq *),
178 				 GFP_KERNEL, dev_to_node(dev));
179 	if (!idxd->wqs)
180 		return -ENOMEM;
181 
182 	for (i = 0; i < idxd->max_wqs; i++) {
183 		wq = kzalloc_node(sizeof(*wq), GFP_KERNEL, dev_to_node(dev));
184 		if (!wq) {
185 			rc = -ENOMEM;
186 			goto err;
187 		}
188 
189 		wq->id = i;
190 		wq->idxd = idxd;
191 		device_initialize(&wq->conf_dev);
192 		wq->conf_dev.parent = &idxd->conf_dev;
193 		wq->conf_dev.bus = &dsa_bus_type;
194 		wq->conf_dev.type = &idxd_wq_device_type;
195 		rc = dev_set_name(&wq->conf_dev, "wq%d.%d", idxd->id, wq->id);
196 		if (rc < 0) {
197 			put_device(&wq->conf_dev);
198 			goto err;
199 		}
200 
201 		mutex_init(&wq->wq_lock);
202 		init_waitqueue_head(&wq->err_queue);
203 		init_completion(&wq->wq_dead);
204 		wq->max_xfer_bytes = idxd->max_xfer_bytes;
205 		wq->max_batch_size = idxd->max_batch_size;
206 		wq->wqcfg = kzalloc_node(idxd->wqcfg_size, GFP_KERNEL, dev_to_node(dev));
207 		if (!wq->wqcfg) {
208 			put_device(&wq->conf_dev);
209 			rc = -ENOMEM;
210 			goto err;
211 		}
212 		idxd->wqs[i] = wq;
213 	}
214 
215 	return 0;
216 
217  err:
218 	while (--i >= 0)
219 		put_device(&idxd->wqs[i]->conf_dev);
220 	return rc;
221 }
222 
223 static int idxd_setup_engines(struct idxd_device *idxd)
224 {
225 	struct idxd_engine *engine;
226 	struct device *dev = &idxd->pdev->dev;
227 	int i, rc;
228 
229 	idxd->engines = kcalloc_node(idxd->max_engines, sizeof(struct idxd_engine *),
230 				     GFP_KERNEL, dev_to_node(dev));
231 	if (!idxd->engines)
232 		return -ENOMEM;
233 
234 	for (i = 0; i < idxd->max_engines; i++) {
235 		engine = kzalloc_node(sizeof(*engine), GFP_KERNEL, dev_to_node(dev));
236 		if (!engine) {
237 			rc = -ENOMEM;
238 			goto err;
239 		}
240 
241 		engine->id = i;
242 		engine->idxd = idxd;
243 		device_initialize(&engine->conf_dev);
244 		engine->conf_dev.parent = &idxd->conf_dev;
245 		engine->conf_dev.type = &idxd_engine_device_type;
246 		rc = dev_set_name(&engine->conf_dev, "engine%d.%d", idxd->id, engine->id);
247 		if (rc < 0) {
248 			put_device(&engine->conf_dev);
249 			goto err;
250 		}
251 
252 		idxd->engines[i] = engine;
253 	}
254 
255 	return 0;
256 
257  err:
258 	while (--i >= 0)
259 		put_device(&idxd->engines[i]->conf_dev);
260 	return rc;
261 }
262 
263 static int idxd_setup_groups(struct idxd_device *idxd)
264 {
265 	struct device *dev = &idxd->pdev->dev;
266 	struct idxd_group *group;
267 	int i, rc;
268 
269 	idxd->groups = kcalloc_node(idxd->max_groups, sizeof(struct idxd_group *),
270 				    GFP_KERNEL, dev_to_node(dev));
271 	if (!idxd->groups)
272 		return -ENOMEM;
273 
274 	for (i = 0; i < idxd->max_groups; i++) {
275 		group = kzalloc_node(sizeof(*group), GFP_KERNEL, dev_to_node(dev));
276 		if (!group) {
277 			rc = -ENOMEM;
278 			goto err;
279 		}
280 
281 		group->id = i;
282 		group->idxd = idxd;
283 		device_initialize(&group->conf_dev);
284 		group->conf_dev.parent = &idxd->conf_dev;
285 		group->conf_dev.bus = &dsa_bus_type;
286 		group->conf_dev.type = &idxd_group_device_type;
287 		rc = dev_set_name(&group->conf_dev, "group%d.%d", idxd->id, group->id);
288 		if (rc < 0) {
289 			put_device(&group->conf_dev);
290 			goto err;
291 		}
292 
293 		idxd->groups[i] = group;
294 		group->tc_a = -1;
295 		group->tc_b = -1;
296 	}
297 
298 	return 0;
299 
300  err:
301 	while (--i >= 0)
302 		put_device(&idxd->groups[i]->conf_dev);
303 	return rc;
304 }
305 
306 static int idxd_setup_internals(struct idxd_device *idxd)
307 {
308 	struct device *dev = &idxd->pdev->dev;
309 	int rc, i;
310 
311 	init_waitqueue_head(&idxd->cmd_waitq);
312 
313 	if (idxd->hw.cmd_cap & BIT(IDXD_CMD_REQUEST_INT_HANDLE)) {
314 		idxd->int_handles = devm_kcalloc(dev, idxd->max_wqs, sizeof(int), GFP_KERNEL);
315 		if (!idxd->int_handles)
316 			return -ENOMEM;
317 	}
318 
319 	rc = idxd_setup_wqs(idxd);
320 	if (rc < 0)
321 		goto err_wqs;
322 
323 	rc = idxd_setup_engines(idxd);
324 	if (rc < 0)
325 		goto err_engine;
326 
327 	rc = idxd_setup_groups(idxd);
328 	if (rc < 0)
329 		goto err_group;
330 
331 	idxd->wq = create_workqueue(dev_name(dev));
332 	if (!idxd->wq) {
333 		rc = -ENOMEM;
334 		goto err_wkq_create;
335 	}
336 
337 	return 0;
338 
339  err_wkq_create:
340 	for (i = 0; i < idxd->max_groups; i++)
341 		put_device(&idxd->groups[i]->conf_dev);
342  err_group:
343 	for (i = 0; i < idxd->max_engines; i++)
344 		put_device(&idxd->engines[i]->conf_dev);
345  err_engine:
346 	for (i = 0; i < idxd->max_wqs; i++)
347 		put_device(&idxd->wqs[i]->conf_dev);
348  err_wqs:
349 	kfree(idxd->int_handles);
350 	return rc;
351 }
352 
353 static void idxd_read_table_offsets(struct idxd_device *idxd)
354 {
355 	union offsets_reg offsets;
356 	struct device *dev = &idxd->pdev->dev;
357 
358 	offsets.bits[0] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET);
359 	offsets.bits[1] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET + sizeof(u64));
360 	idxd->grpcfg_offset = offsets.grpcfg * IDXD_TABLE_MULT;
361 	dev_dbg(dev, "IDXD Group Config Offset: %#x\n", idxd->grpcfg_offset);
362 	idxd->wqcfg_offset = offsets.wqcfg * IDXD_TABLE_MULT;
363 	dev_dbg(dev, "IDXD Work Queue Config Offset: %#x\n", idxd->wqcfg_offset);
364 	idxd->msix_perm_offset = offsets.msix_perm * IDXD_TABLE_MULT;
365 	dev_dbg(dev, "IDXD MSIX Permission Offset: %#x\n", idxd->msix_perm_offset);
366 	idxd->perfmon_offset = offsets.perfmon * IDXD_TABLE_MULT;
367 	dev_dbg(dev, "IDXD Perfmon Offset: %#x\n", idxd->perfmon_offset);
368 }
369 
370 static void idxd_read_caps(struct idxd_device *idxd)
371 {
372 	struct device *dev = &idxd->pdev->dev;
373 	int i;
374 
375 	/* reading generic capabilities */
376 	idxd->hw.gen_cap.bits = ioread64(idxd->reg_base + IDXD_GENCAP_OFFSET);
377 	dev_dbg(dev, "gen_cap: %#llx\n", idxd->hw.gen_cap.bits);
378 
379 	if (idxd->hw.gen_cap.cmd_cap) {
380 		idxd->hw.cmd_cap = ioread32(idxd->reg_base + IDXD_CMDCAP_OFFSET);
381 		dev_dbg(dev, "cmd_cap: %#x\n", idxd->hw.cmd_cap);
382 	}
383 
384 	idxd->max_xfer_bytes = 1ULL << idxd->hw.gen_cap.max_xfer_shift;
385 	dev_dbg(dev, "max xfer size: %llu bytes\n", idxd->max_xfer_bytes);
386 	idxd->max_batch_size = 1U << idxd->hw.gen_cap.max_batch_shift;
387 	dev_dbg(dev, "max batch size: %u\n", idxd->max_batch_size);
388 	if (idxd->hw.gen_cap.config_en)
389 		set_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags);
390 
391 	/* reading group capabilities */
392 	idxd->hw.group_cap.bits =
393 		ioread64(idxd->reg_base + IDXD_GRPCAP_OFFSET);
394 	dev_dbg(dev, "group_cap: %#llx\n", idxd->hw.group_cap.bits);
395 	idxd->max_groups = idxd->hw.group_cap.num_groups;
396 	dev_dbg(dev, "max groups: %u\n", idxd->max_groups);
397 	idxd->max_tokens = idxd->hw.group_cap.total_tokens;
398 	dev_dbg(dev, "max tokens: %u\n", idxd->max_tokens);
399 	idxd->nr_tokens = idxd->max_tokens;
400 
401 	/* read engine capabilities */
402 	idxd->hw.engine_cap.bits =
403 		ioread64(idxd->reg_base + IDXD_ENGCAP_OFFSET);
404 	dev_dbg(dev, "engine_cap: %#llx\n", idxd->hw.engine_cap.bits);
405 	idxd->max_engines = idxd->hw.engine_cap.num_engines;
406 	dev_dbg(dev, "max engines: %u\n", idxd->max_engines);
407 
408 	/* read workqueue capabilities */
409 	idxd->hw.wq_cap.bits = ioread64(idxd->reg_base + IDXD_WQCAP_OFFSET);
410 	dev_dbg(dev, "wq_cap: %#llx\n", idxd->hw.wq_cap.bits);
411 	idxd->max_wq_size = idxd->hw.wq_cap.total_wq_size;
412 	dev_dbg(dev, "total workqueue size: %u\n", idxd->max_wq_size);
413 	idxd->max_wqs = idxd->hw.wq_cap.num_wqs;
414 	dev_dbg(dev, "max workqueues: %u\n", idxd->max_wqs);
415 	idxd->wqcfg_size = 1 << (idxd->hw.wq_cap.wqcfg_size + IDXD_WQCFG_MIN);
416 	dev_dbg(dev, "wqcfg size: %u\n", idxd->wqcfg_size);
417 
418 	/* reading operation capabilities */
419 	for (i = 0; i < 4; i++) {
420 		idxd->hw.opcap.bits[i] = ioread64(idxd->reg_base +
421 				IDXD_OPCAP_OFFSET + i * sizeof(u64));
422 		dev_dbg(dev, "opcap[%d]: %#llx\n", i, idxd->hw.opcap.bits[i]);
423 	}
424 }
425 
426 static struct idxd_device *idxd_alloc(struct pci_dev *pdev, struct idxd_driver_data *data)
427 {
428 	struct device *dev = &pdev->dev;
429 	struct idxd_device *idxd;
430 	int rc;
431 
432 	idxd = kzalloc_node(sizeof(*idxd), GFP_KERNEL, dev_to_node(dev));
433 	if (!idxd)
434 		return NULL;
435 
436 	idxd->pdev = pdev;
437 	idxd->data = data;
438 	idxd->id = ida_alloc(&idxd_ida, GFP_KERNEL);
439 	if (idxd->id < 0)
440 		return NULL;
441 
442 	device_initialize(&idxd->conf_dev);
443 	idxd->conf_dev.parent = dev;
444 	idxd->conf_dev.bus = &dsa_bus_type;
445 	idxd->conf_dev.type = idxd->data->dev_type;
446 	rc = dev_set_name(&idxd->conf_dev, "%s%d", idxd->data->name_prefix, idxd->id);
447 	if (rc < 0) {
448 		put_device(&idxd->conf_dev);
449 		return NULL;
450 	}
451 
452 	spin_lock_init(&idxd->dev_lock);
453 	spin_lock_init(&idxd->cmd_lock);
454 
455 	return idxd;
456 }
457 
458 static int idxd_enable_system_pasid(struct idxd_device *idxd)
459 {
460 	int flags;
461 	unsigned int pasid;
462 	struct iommu_sva *sva;
463 
464 	flags = SVM_FLAG_SUPERVISOR_MODE;
465 
466 	sva = iommu_sva_bind_device(&idxd->pdev->dev, NULL, &flags);
467 	if (IS_ERR(sva)) {
468 		dev_warn(&idxd->pdev->dev,
469 			 "iommu sva bind failed: %ld\n", PTR_ERR(sva));
470 		return PTR_ERR(sva);
471 	}
472 
473 	pasid = iommu_sva_get_pasid(sva);
474 	if (pasid == IOMMU_PASID_INVALID) {
475 		iommu_sva_unbind_device(sva);
476 		return -ENODEV;
477 	}
478 
479 	idxd->sva = sva;
480 	idxd->pasid = pasid;
481 	dev_dbg(&idxd->pdev->dev, "system pasid: %u\n", pasid);
482 	return 0;
483 }
484 
485 static void idxd_disable_system_pasid(struct idxd_device *idxd)
486 {
487 
488 	iommu_sva_unbind_device(idxd->sva);
489 	idxd->sva = NULL;
490 }
491 
492 static int idxd_probe(struct idxd_device *idxd)
493 {
494 	struct pci_dev *pdev = idxd->pdev;
495 	struct device *dev = &pdev->dev;
496 	int rc;
497 
498 	dev_dbg(dev, "%s entered and resetting device\n", __func__);
499 	rc = idxd_device_init_reset(idxd);
500 	if (rc < 0)
501 		return rc;
502 
503 	dev_dbg(dev, "IDXD reset complete\n");
504 
505 	if (IS_ENABLED(CONFIG_INTEL_IDXD_SVM) && sva) {
506 		rc = iommu_dev_enable_feature(dev, IOMMU_DEV_FEAT_SVA);
507 		if (rc == 0) {
508 			rc = idxd_enable_system_pasid(idxd);
509 			if (rc < 0) {
510 				iommu_dev_disable_feature(dev, IOMMU_DEV_FEAT_SVA);
511 				dev_warn(dev, "Failed to enable PASID. No SVA support: %d\n", rc);
512 			} else {
513 				set_bit(IDXD_FLAG_PASID_ENABLED, &idxd->flags);
514 			}
515 		} else {
516 			dev_warn(dev, "Unable to turn on SVA feature.\n");
517 		}
518 	} else if (!sva) {
519 		dev_warn(dev, "User forced SVA off via module param.\n");
520 	}
521 
522 	idxd_read_caps(idxd);
523 	idxd_read_table_offsets(idxd);
524 
525 	rc = idxd_setup_internals(idxd);
526 	if (rc)
527 		goto err;
528 
529 	/* If the configs are readonly, then load them from device */
530 	if (!test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags)) {
531 		dev_dbg(dev, "Loading RO device config\n");
532 		rc = idxd_device_load_config(idxd);
533 		if (rc < 0)
534 			goto err;
535 	}
536 
537 	rc = idxd_setup_interrupts(idxd);
538 	if (rc)
539 		goto err;
540 
541 	dev_dbg(dev, "IDXD interrupt setup complete.\n");
542 
543 	idxd->major = idxd_cdev_get_major(idxd);
544 
545 	rc = perfmon_pmu_init(idxd);
546 	if (rc < 0)
547 		dev_warn(dev, "Failed to initialize perfmon. No PMU support: %d\n", rc);
548 
549 	dev_dbg(dev, "IDXD device %d probed successfully\n", idxd->id);
550 	return 0;
551 
552  err:
553 	if (device_pasid_enabled(idxd))
554 		idxd_disable_system_pasid(idxd);
555 	iommu_dev_disable_feature(dev, IOMMU_DEV_FEAT_SVA);
556 	return rc;
557 }
558 
559 static int idxd_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
560 {
561 	struct device *dev = &pdev->dev;
562 	struct idxd_device *idxd;
563 	struct idxd_driver_data *data = (struct idxd_driver_data *)id->driver_data;
564 	int rc;
565 
566 	rc = pci_enable_device(pdev);
567 	if (rc)
568 		return rc;
569 
570 	dev_dbg(dev, "Alloc IDXD context\n");
571 	idxd = idxd_alloc(pdev, data);
572 	if (!idxd) {
573 		rc = -ENOMEM;
574 		goto err_idxd_alloc;
575 	}
576 
577 	dev_dbg(dev, "Mapping BARs\n");
578 	idxd->reg_base = pci_iomap(pdev, IDXD_MMIO_BAR, 0);
579 	if (!idxd->reg_base) {
580 		rc = -ENOMEM;
581 		goto err_iomap;
582 	}
583 
584 	dev_dbg(dev, "Set DMA masks\n");
585 	rc = pci_set_dma_mask(pdev, DMA_BIT_MASK(64));
586 	if (rc)
587 		rc = pci_set_dma_mask(pdev, DMA_BIT_MASK(32));
588 	if (rc)
589 		goto err;
590 
591 	rc = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
592 	if (rc)
593 		rc = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(32));
594 	if (rc)
595 		goto err;
596 
597 	dev_dbg(dev, "Set PCI master\n");
598 	pci_set_master(pdev);
599 	pci_set_drvdata(pdev, idxd);
600 
601 	idxd->hw.version = ioread32(idxd->reg_base + IDXD_VER_OFFSET);
602 	rc = idxd_probe(idxd);
603 	if (rc) {
604 		dev_err(dev, "Intel(R) IDXD DMA Engine init failed\n");
605 		goto err;
606 	}
607 
608 	rc = idxd_register_devices(idxd);
609 	if (rc) {
610 		dev_err(dev, "IDXD sysfs setup failed\n");
611 		goto err;
612 	}
613 
614 	idxd->state = IDXD_DEV_CONF_READY;
615 
616 	dev_info(&pdev->dev, "Intel(R) Accelerator Device (v%x)\n",
617 		 idxd->hw.version);
618 
619 	return 0;
620 
621  err:
622 	pci_iounmap(pdev, idxd->reg_base);
623  err_iomap:
624 	put_device(&idxd->conf_dev);
625  err_idxd_alloc:
626 	pci_disable_device(pdev);
627 	return rc;
628 }
629 
630 static void idxd_flush_pending_llist(struct idxd_irq_entry *ie)
631 {
632 	struct idxd_desc *desc, *itr;
633 	struct llist_node *head;
634 
635 	head = llist_del_all(&ie->pending_llist);
636 	if (!head)
637 		return;
638 
639 	llist_for_each_entry_safe(desc, itr, head, llnode) {
640 		idxd_dma_complete_txd(desc, IDXD_COMPLETE_ABORT);
641 		idxd_free_desc(desc->wq, desc);
642 	}
643 }
644 
645 static void idxd_flush_work_list(struct idxd_irq_entry *ie)
646 {
647 	struct idxd_desc *desc, *iter;
648 
649 	list_for_each_entry_safe(desc, iter, &ie->work_list, list) {
650 		list_del(&desc->list);
651 		idxd_dma_complete_txd(desc, IDXD_COMPLETE_ABORT);
652 		idxd_free_desc(desc->wq, desc);
653 	}
654 }
655 
656 void idxd_wqs_quiesce(struct idxd_device *idxd)
657 {
658 	struct idxd_wq *wq;
659 	int i;
660 
661 	for (i = 0; i < idxd->max_wqs; i++) {
662 		wq = idxd->wqs[i];
663 		if (wq->state == IDXD_WQ_ENABLED && wq->type == IDXD_WQT_KERNEL)
664 			idxd_wq_quiesce(wq);
665 	}
666 }
667 
668 static void idxd_release_int_handles(struct idxd_device *idxd)
669 {
670 	struct device *dev = &idxd->pdev->dev;
671 	int i, rc;
672 
673 	for (i = 0; i < idxd->num_wq_irqs; i++) {
674 		if (idxd->hw.cmd_cap & BIT(IDXD_CMD_RELEASE_INT_HANDLE)) {
675 			rc = idxd_device_release_int_handle(idxd, idxd->int_handles[i],
676 							    IDXD_IRQ_MSIX);
677 			if (rc < 0)
678 				dev_warn(dev, "irq handle %d release failed\n",
679 					 idxd->int_handles[i]);
680 			else
681 				dev_dbg(dev, "int handle requested: %u\n", idxd->int_handles[i]);
682 		}
683 	}
684 }
685 
686 static void idxd_shutdown(struct pci_dev *pdev)
687 {
688 	struct idxd_device *idxd = pci_get_drvdata(pdev);
689 	int rc, i;
690 	struct idxd_irq_entry *irq_entry;
691 	int msixcnt = pci_msix_vec_count(pdev);
692 
693 	rc = idxd_device_disable(idxd);
694 	if (rc)
695 		dev_err(&pdev->dev, "Disabling device failed\n");
696 
697 	dev_dbg(&pdev->dev, "%s called\n", __func__);
698 	idxd_mask_msix_vectors(idxd);
699 	idxd_mask_error_interrupts(idxd);
700 
701 	for (i = 0; i < msixcnt; i++) {
702 		irq_entry = &idxd->irq_entries[i];
703 		synchronize_irq(irq_entry->vector);
704 		free_irq(irq_entry->vector, irq_entry);
705 		if (i == 0)
706 			continue;
707 		idxd_flush_pending_llist(irq_entry);
708 		idxd_flush_work_list(irq_entry);
709 	}
710 
711 	idxd_msix_perm_clear(idxd);
712 	idxd_release_int_handles(idxd);
713 	pci_free_irq_vectors(pdev);
714 	pci_iounmap(pdev, idxd->reg_base);
715 	pci_disable_device(pdev);
716 	destroy_workqueue(idxd->wq);
717 }
718 
719 static void idxd_remove(struct pci_dev *pdev)
720 {
721 	struct idxd_device *idxd = pci_get_drvdata(pdev);
722 
723 	dev_dbg(&pdev->dev, "%s called\n", __func__);
724 	idxd_shutdown(pdev);
725 	if (device_pasid_enabled(idxd))
726 		idxd_disable_system_pasid(idxd);
727 	idxd_unregister_devices(idxd);
728 	perfmon_pmu_remove(idxd);
729 	iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_SVA);
730 }
731 
732 static struct pci_driver idxd_pci_driver = {
733 	.name		= DRV_NAME,
734 	.id_table	= idxd_pci_tbl,
735 	.probe		= idxd_pci_probe,
736 	.remove		= idxd_remove,
737 	.shutdown	= idxd_shutdown,
738 };
739 
740 static int __init idxd_init_module(void)
741 {
742 	int err;
743 
744 	/*
745 	 * If the CPU does not support MOVDIR64B or ENQCMDS, there's no point in
746 	 * enumerating the device. We can not utilize it.
747 	 */
748 	if (!boot_cpu_has(X86_FEATURE_MOVDIR64B)) {
749 		pr_warn("idxd driver failed to load without MOVDIR64B.\n");
750 		return -ENODEV;
751 	}
752 
753 	if (!boot_cpu_has(X86_FEATURE_ENQCMD))
754 		pr_warn("Platform does not have ENQCMD(S) support.\n");
755 	else
756 		support_enqcmd = true;
757 
758 	perfmon_init();
759 
760 	err = idxd_register_bus_type();
761 	if (err < 0)
762 		return err;
763 
764 	err = idxd_register_driver();
765 	if (err < 0)
766 		goto err_idxd_driver_register;
767 
768 	err = idxd_cdev_register();
769 	if (err)
770 		goto err_cdev_register;
771 
772 	err = pci_register_driver(&idxd_pci_driver);
773 	if (err)
774 		goto err_pci_register;
775 
776 	return 0;
777 
778 err_pci_register:
779 	idxd_cdev_remove();
780 err_cdev_register:
781 	idxd_unregister_driver();
782 err_idxd_driver_register:
783 	idxd_unregister_bus_type();
784 	return err;
785 }
786 module_init(idxd_init_module);
787 
788 static void __exit idxd_exit_module(void)
789 {
790 	pci_unregister_driver(&idxd_pci_driver);
791 	idxd_cdev_remove();
792 	idxd_unregister_bus_type();
793 	perfmon_exit();
794 }
795 module_exit(idxd_exit_module);
796