xref: /linux/drivers/dma/idxd/init.c (revision 20dfee95936413708701eb151f419597fdd9d948)
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/fs.h>
13 #include <linux/io-64-nonatomic-lo-hi.h>
14 #include <linux/device.h>
15 #include <linux/idr.h>
16 #include <linux/iommu.h>
17 #include <uapi/linux/idxd.h>
18 #include <linux/dmaengine.h>
19 #include "../dmaengine.h"
20 #include "registers.h"
21 #include "idxd.h"
22 #include "perfmon.h"
23 
24 MODULE_VERSION(IDXD_DRIVER_VERSION);
25 MODULE_LICENSE("GPL v2");
26 MODULE_AUTHOR("Intel Corporation");
27 MODULE_IMPORT_NS(IDXD);
28 
29 static bool sva = true;
30 module_param(sva, bool, 0644);
31 MODULE_PARM_DESC(sva, "Toggle SVA support on/off");
32 
33 bool tc_override;
34 module_param(tc_override, bool, 0644);
35 MODULE_PARM_DESC(tc_override, "Override traffic class defaults");
36 
37 #define DRV_NAME "idxd"
38 
39 bool support_enqcmd;
40 DEFINE_IDA(idxd_ida);
41 
42 static struct idxd_driver_data idxd_driver_data[] = {
43 	[IDXD_TYPE_DSA] = {
44 		.name_prefix = "dsa",
45 		.type = IDXD_TYPE_DSA,
46 		.compl_size = sizeof(struct dsa_completion_record),
47 		.align = 32,
48 		.dev_type = &dsa_device_type,
49 		.evl_cr_off = offsetof(struct dsa_evl_entry, cr),
50 		.user_submission_safe = false, /* See INTEL-SA-01084 security advisory */
51 		.cr_status_off = offsetof(struct dsa_completion_record, status),
52 		.cr_result_off = offsetof(struct dsa_completion_record, result),
53 	},
54 	[IDXD_TYPE_IAX] = {
55 		.name_prefix = "iax",
56 		.type = IDXD_TYPE_IAX,
57 		.compl_size = sizeof(struct iax_completion_record),
58 		.align = 64,
59 		.dev_type = &iax_device_type,
60 		.evl_cr_off = offsetof(struct iax_evl_entry, cr),
61 		.user_submission_safe = false, /* See INTEL-SA-01084 security advisory */
62 		.cr_status_off = offsetof(struct iax_completion_record, status),
63 		.cr_result_off = offsetof(struct iax_completion_record, error_code),
64 		.load_device_defaults = idxd_load_iaa_device_defaults,
65 	},
66 };
67 
68 static struct pci_device_id idxd_pci_tbl[] = {
69 	/* DSA ver 1.0 platforms */
70 	{ PCI_DEVICE_DATA(INTEL, DSA_SPR0, &idxd_driver_data[IDXD_TYPE_DSA]) },
71 
72 	/* IAX ver 1.0 platforms */
73 	{ PCI_DEVICE_DATA(INTEL, IAX_SPR0, &idxd_driver_data[IDXD_TYPE_IAX]) },
74 	{ 0, }
75 };
76 MODULE_DEVICE_TABLE(pci, idxd_pci_tbl);
77 
78 static int idxd_setup_interrupts(struct idxd_device *idxd)
79 {
80 	struct pci_dev *pdev = idxd->pdev;
81 	struct device *dev = &pdev->dev;
82 	struct idxd_irq_entry *ie;
83 	int i, msixcnt;
84 	int rc = 0;
85 
86 	msixcnt = pci_msix_vec_count(pdev);
87 	if (msixcnt < 0) {
88 		dev_err(dev, "Not MSI-X interrupt capable.\n");
89 		return -ENOSPC;
90 	}
91 	idxd->irq_cnt = msixcnt;
92 
93 	rc = pci_alloc_irq_vectors(pdev, msixcnt, msixcnt, PCI_IRQ_MSIX);
94 	if (rc != msixcnt) {
95 		dev_err(dev, "Failed enabling %d MSIX entries: %d\n", msixcnt, rc);
96 		return -ENOSPC;
97 	}
98 	dev_dbg(dev, "Enabled %d msix vectors\n", msixcnt);
99 
100 
101 	ie = idxd_get_ie(idxd, 0);
102 	ie->vector = pci_irq_vector(pdev, 0);
103 	rc = request_threaded_irq(ie->vector, NULL, idxd_misc_thread, 0, "idxd-misc", ie);
104 	if (rc < 0) {
105 		dev_err(dev, "Failed to allocate misc interrupt.\n");
106 		goto err_misc_irq;
107 	}
108 	dev_dbg(dev, "Requested idxd-misc handler on msix vector %d\n", ie->vector);
109 
110 	for (i = 0; i < idxd->max_wqs; i++) {
111 		int msix_idx = i + 1;
112 
113 		ie = idxd_get_ie(idxd, msix_idx);
114 		ie->id = msix_idx;
115 		ie->int_handle = INVALID_INT_HANDLE;
116 		ie->pasid = IOMMU_PASID_INVALID;
117 
118 		spin_lock_init(&ie->list_lock);
119 		init_llist_head(&ie->pending_llist);
120 		INIT_LIST_HEAD(&ie->work_list);
121 	}
122 
123 	idxd_unmask_error_interrupts(idxd);
124 	return 0;
125 
126  err_misc_irq:
127 	idxd_mask_error_interrupts(idxd);
128 	pci_free_irq_vectors(pdev);
129 	dev_err(dev, "No usable interrupts\n");
130 	return rc;
131 }
132 
133 static void idxd_cleanup_interrupts(struct idxd_device *idxd)
134 {
135 	struct pci_dev *pdev = idxd->pdev;
136 	struct idxd_irq_entry *ie;
137 	int msixcnt;
138 
139 	msixcnt = pci_msix_vec_count(pdev);
140 	if (msixcnt <= 0)
141 		return;
142 
143 	ie = idxd_get_ie(idxd, 0);
144 	idxd_mask_error_interrupts(idxd);
145 	free_irq(ie->vector, ie);
146 	pci_free_irq_vectors(pdev);
147 }
148 
149 static int idxd_setup_wqs(struct idxd_device *idxd)
150 {
151 	struct device *dev = &idxd->pdev->dev;
152 	struct idxd_wq *wq;
153 	struct device *conf_dev;
154 	int i, rc;
155 
156 	idxd->wqs = kcalloc_node(idxd->max_wqs, sizeof(struct idxd_wq *),
157 				 GFP_KERNEL, dev_to_node(dev));
158 	if (!idxd->wqs)
159 		return -ENOMEM;
160 
161 	idxd->wq_enable_map = bitmap_zalloc_node(idxd->max_wqs, GFP_KERNEL, dev_to_node(dev));
162 	if (!idxd->wq_enable_map) {
163 		kfree(idxd->wqs);
164 		return -ENOMEM;
165 	}
166 
167 	for (i = 0; i < idxd->max_wqs; i++) {
168 		wq = kzalloc_node(sizeof(*wq), GFP_KERNEL, dev_to_node(dev));
169 		if (!wq) {
170 			rc = -ENOMEM;
171 			goto err;
172 		}
173 
174 		idxd_dev_set_type(&wq->idxd_dev, IDXD_DEV_WQ);
175 		conf_dev = wq_confdev(wq);
176 		wq->id = i;
177 		wq->idxd = idxd;
178 		device_initialize(wq_confdev(wq));
179 		conf_dev->parent = idxd_confdev(idxd);
180 		conf_dev->bus = &dsa_bus_type;
181 		conf_dev->type = &idxd_wq_device_type;
182 		rc = dev_set_name(conf_dev, "wq%d.%d", idxd->id, wq->id);
183 		if (rc < 0) {
184 			put_device(conf_dev);
185 			goto err;
186 		}
187 
188 		mutex_init(&wq->wq_lock);
189 		init_waitqueue_head(&wq->err_queue);
190 		init_completion(&wq->wq_dead);
191 		init_completion(&wq->wq_resurrect);
192 		wq->max_xfer_bytes = WQ_DEFAULT_MAX_XFER;
193 		idxd_wq_set_max_batch_size(idxd->data->type, wq, WQ_DEFAULT_MAX_BATCH);
194 		wq->enqcmds_retries = IDXD_ENQCMDS_RETRIES;
195 		wq->wqcfg = kzalloc_node(idxd->wqcfg_size, GFP_KERNEL, dev_to_node(dev));
196 		if (!wq->wqcfg) {
197 			put_device(conf_dev);
198 			rc = -ENOMEM;
199 			goto err;
200 		}
201 
202 		if (idxd->hw.wq_cap.op_config) {
203 			wq->opcap_bmap = bitmap_zalloc(IDXD_MAX_OPCAP_BITS, GFP_KERNEL);
204 			if (!wq->opcap_bmap) {
205 				put_device(conf_dev);
206 				rc = -ENOMEM;
207 				goto err;
208 			}
209 			bitmap_copy(wq->opcap_bmap, idxd->opcap_bmap, IDXD_MAX_OPCAP_BITS);
210 		}
211 		mutex_init(&wq->uc_lock);
212 		xa_init(&wq->upasid_xa);
213 		idxd->wqs[i] = wq;
214 	}
215 
216 	return 0;
217 
218  err:
219 	while (--i >= 0) {
220 		wq = idxd->wqs[i];
221 		conf_dev = wq_confdev(wq);
222 		put_device(conf_dev);
223 	}
224 	return rc;
225 }
226 
227 static int idxd_setup_engines(struct idxd_device *idxd)
228 {
229 	struct idxd_engine *engine;
230 	struct device *dev = &idxd->pdev->dev;
231 	struct device *conf_dev;
232 	int i, rc;
233 
234 	idxd->engines = kcalloc_node(idxd->max_engines, sizeof(struct idxd_engine *),
235 				     GFP_KERNEL, dev_to_node(dev));
236 	if (!idxd->engines)
237 		return -ENOMEM;
238 
239 	for (i = 0; i < idxd->max_engines; i++) {
240 		engine = kzalloc_node(sizeof(*engine), GFP_KERNEL, dev_to_node(dev));
241 		if (!engine) {
242 			rc = -ENOMEM;
243 			goto err;
244 		}
245 
246 		idxd_dev_set_type(&engine->idxd_dev, IDXD_DEV_ENGINE);
247 		conf_dev = engine_confdev(engine);
248 		engine->id = i;
249 		engine->idxd = idxd;
250 		device_initialize(conf_dev);
251 		conf_dev->parent = idxd_confdev(idxd);
252 		conf_dev->bus = &dsa_bus_type;
253 		conf_dev->type = &idxd_engine_device_type;
254 		rc = dev_set_name(conf_dev, "engine%d.%d", idxd->id, engine->id);
255 		if (rc < 0) {
256 			put_device(conf_dev);
257 			goto err;
258 		}
259 
260 		idxd->engines[i] = engine;
261 	}
262 
263 	return 0;
264 
265  err:
266 	while (--i >= 0) {
267 		engine = idxd->engines[i];
268 		conf_dev = engine_confdev(engine);
269 		put_device(conf_dev);
270 	}
271 	return rc;
272 }
273 
274 static int idxd_setup_groups(struct idxd_device *idxd)
275 {
276 	struct device *dev = &idxd->pdev->dev;
277 	struct device *conf_dev;
278 	struct idxd_group *group;
279 	int i, rc;
280 
281 	idxd->groups = kcalloc_node(idxd->max_groups, sizeof(struct idxd_group *),
282 				    GFP_KERNEL, dev_to_node(dev));
283 	if (!idxd->groups)
284 		return -ENOMEM;
285 
286 	for (i = 0; i < idxd->max_groups; i++) {
287 		group = kzalloc_node(sizeof(*group), GFP_KERNEL, dev_to_node(dev));
288 		if (!group) {
289 			rc = -ENOMEM;
290 			goto err;
291 		}
292 
293 		idxd_dev_set_type(&group->idxd_dev, IDXD_DEV_GROUP);
294 		conf_dev = group_confdev(group);
295 		group->id = i;
296 		group->idxd = idxd;
297 		device_initialize(conf_dev);
298 		conf_dev->parent = idxd_confdev(idxd);
299 		conf_dev->bus = &dsa_bus_type;
300 		conf_dev->type = &idxd_group_device_type;
301 		rc = dev_set_name(conf_dev, "group%d.%d", idxd->id, group->id);
302 		if (rc < 0) {
303 			put_device(conf_dev);
304 			goto err;
305 		}
306 
307 		idxd->groups[i] = group;
308 		if (idxd->hw.version <= DEVICE_VERSION_2 && !tc_override) {
309 			group->tc_a = 1;
310 			group->tc_b = 1;
311 		} else {
312 			group->tc_a = -1;
313 			group->tc_b = -1;
314 		}
315 		/*
316 		 * The default value is the same as the value of
317 		 * total read buffers in GRPCAP.
318 		 */
319 		group->rdbufs_allowed = idxd->max_rdbufs;
320 	}
321 
322 	return 0;
323 
324  err:
325 	while (--i >= 0) {
326 		group = idxd->groups[i];
327 		put_device(group_confdev(group));
328 	}
329 	return rc;
330 }
331 
332 static void idxd_cleanup_internals(struct idxd_device *idxd)
333 {
334 	int i;
335 
336 	for (i = 0; i < idxd->max_groups; i++)
337 		put_device(group_confdev(idxd->groups[i]));
338 	for (i = 0; i < idxd->max_engines; i++)
339 		put_device(engine_confdev(idxd->engines[i]));
340 	for (i = 0; i < idxd->max_wqs; i++)
341 		put_device(wq_confdev(idxd->wqs[i]));
342 	destroy_workqueue(idxd->wq);
343 }
344 
345 static int idxd_init_evl(struct idxd_device *idxd)
346 {
347 	struct device *dev = &idxd->pdev->dev;
348 	unsigned int evl_cache_size;
349 	struct idxd_evl *evl;
350 	const char *idxd_name;
351 
352 	if (idxd->hw.gen_cap.evl_support == 0)
353 		return 0;
354 
355 	evl = kzalloc_node(sizeof(*evl), GFP_KERNEL, dev_to_node(dev));
356 	if (!evl)
357 		return -ENOMEM;
358 
359 	mutex_init(&evl->lock);
360 	evl->size = IDXD_EVL_SIZE_MIN;
361 
362 	idxd_name = dev_name(idxd_confdev(idxd));
363 	evl_cache_size = sizeof(struct idxd_evl_fault) + evl_ent_size(idxd);
364 	/*
365 	 * Since completion record in evl_cache will be copied to user
366 	 * when handling completion record page fault, need to create
367 	 * the cache suitable for user copy.
368 	 */
369 	idxd->evl_cache = kmem_cache_create_usercopy(idxd_name, evl_cache_size,
370 						     0, 0, 0, evl_cache_size,
371 						     NULL);
372 	if (!idxd->evl_cache) {
373 		kfree(evl);
374 		return -ENOMEM;
375 	}
376 
377 	idxd->evl = evl;
378 	return 0;
379 }
380 
381 static int idxd_setup_internals(struct idxd_device *idxd)
382 {
383 	struct device *dev = &idxd->pdev->dev;
384 	int rc, i;
385 
386 	init_waitqueue_head(&idxd->cmd_waitq);
387 
388 	rc = idxd_setup_wqs(idxd);
389 	if (rc < 0)
390 		goto err_wqs;
391 
392 	rc = idxd_setup_engines(idxd);
393 	if (rc < 0)
394 		goto err_engine;
395 
396 	rc = idxd_setup_groups(idxd);
397 	if (rc < 0)
398 		goto err_group;
399 
400 	idxd->wq = create_workqueue(dev_name(dev));
401 	if (!idxd->wq) {
402 		rc = -ENOMEM;
403 		goto err_wkq_create;
404 	}
405 
406 	rc = idxd_init_evl(idxd);
407 	if (rc < 0)
408 		goto err_evl;
409 
410 	return 0;
411 
412  err_evl:
413 	destroy_workqueue(idxd->wq);
414  err_wkq_create:
415 	for (i = 0; i < idxd->max_groups; i++)
416 		put_device(group_confdev(idxd->groups[i]));
417  err_group:
418 	for (i = 0; i < idxd->max_engines; i++)
419 		put_device(engine_confdev(idxd->engines[i]));
420  err_engine:
421 	for (i = 0; i < idxd->max_wqs; i++)
422 		put_device(wq_confdev(idxd->wqs[i]));
423  err_wqs:
424 	return rc;
425 }
426 
427 static void idxd_read_table_offsets(struct idxd_device *idxd)
428 {
429 	union offsets_reg offsets;
430 	struct device *dev = &idxd->pdev->dev;
431 
432 	offsets.bits[0] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET);
433 	offsets.bits[1] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET + sizeof(u64));
434 	idxd->grpcfg_offset = offsets.grpcfg * IDXD_TABLE_MULT;
435 	dev_dbg(dev, "IDXD Group Config Offset: %#x\n", idxd->grpcfg_offset);
436 	idxd->wqcfg_offset = offsets.wqcfg * IDXD_TABLE_MULT;
437 	dev_dbg(dev, "IDXD Work Queue Config Offset: %#x\n", idxd->wqcfg_offset);
438 	idxd->msix_perm_offset = offsets.msix_perm * IDXD_TABLE_MULT;
439 	dev_dbg(dev, "IDXD MSIX Permission Offset: %#x\n", idxd->msix_perm_offset);
440 	idxd->perfmon_offset = offsets.perfmon * IDXD_TABLE_MULT;
441 	dev_dbg(dev, "IDXD Perfmon Offset: %#x\n", idxd->perfmon_offset);
442 }
443 
444 void multi_u64_to_bmap(unsigned long *bmap, u64 *val, int count)
445 {
446 	int i, j, nr;
447 
448 	for (i = 0, nr = 0; i < count; i++) {
449 		for (j = 0; j < BITS_PER_LONG_LONG; j++) {
450 			if (val[i] & BIT(j))
451 				set_bit(nr, bmap);
452 			nr++;
453 		}
454 	}
455 }
456 
457 static void idxd_read_caps(struct idxd_device *idxd)
458 {
459 	struct device *dev = &idxd->pdev->dev;
460 	int i;
461 
462 	/* reading generic capabilities */
463 	idxd->hw.gen_cap.bits = ioread64(idxd->reg_base + IDXD_GENCAP_OFFSET);
464 	dev_dbg(dev, "gen_cap: %#llx\n", idxd->hw.gen_cap.bits);
465 
466 	if (idxd->hw.gen_cap.cmd_cap) {
467 		idxd->hw.cmd_cap = ioread32(idxd->reg_base + IDXD_CMDCAP_OFFSET);
468 		dev_dbg(dev, "cmd_cap: %#x\n", idxd->hw.cmd_cap);
469 	}
470 
471 	/* reading command capabilities */
472 	if (idxd->hw.cmd_cap & BIT(IDXD_CMD_REQUEST_INT_HANDLE))
473 		idxd->request_int_handles = true;
474 
475 	idxd->max_xfer_bytes = 1ULL << idxd->hw.gen_cap.max_xfer_shift;
476 	dev_dbg(dev, "max xfer size: %llu bytes\n", idxd->max_xfer_bytes);
477 	idxd_set_max_batch_size(idxd->data->type, idxd, 1U << idxd->hw.gen_cap.max_batch_shift);
478 	dev_dbg(dev, "max batch size: %u\n", idxd->max_batch_size);
479 	if (idxd->hw.gen_cap.config_en)
480 		set_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags);
481 
482 	/* reading group capabilities */
483 	idxd->hw.group_cap.bits =
484 		ioread64(idxd->reg_base + IDXD_GRPCAP_OFFSET);
485 	dev_dbg(dev, "group_cap: %#llx\n", idxd->hw.group_cap.bits);
486 	idxd->max_groups = idxd->hw.group_cap.num_groups;
487 	dev_dbg(dev, "max groups: %u\n", idxd->max_groups);
488 	idxd->max_rdbufs = idxd->hw.group_cap.total_rdbufs;
489 	dev_dbg(dev, "max read buffers: %u\n", idxd->max_rdbufs);
490 	idxd->nr_rdbufs = idxd->max_rdbufs;
491 
492 	/* read engine capabilities */
493 	idxd->hw.engine_cap.bits =
494 		ioread64(idxd->reg_base + IDXD_ENGCAP_OFFSET);
495 	dev_dbg(dev, "engine_cap: %#llx\n", idxd->hw.engine_cap.bits);
496 	idxd->max_engines = idxd->hw.engine_cap.num_engines;
497 	dev_dbg(dev, "max engines: %u\n", idxd->max_engines);
498 
499 	/* read workqueue capabilities */
500 	idxd->hw.wq_cap.bits = ioread64(idxd->reg_base + IDXD_WQCAP_OFFSET);
501 	dev_dbg(dev, "wq_cap: %#llx\n", idxd->hw.wq_cap.bits);
502 	idxd->max_wq_size = idxd->hw.wq_cap.total_wq_size;
503 	dev_dbg(dev, "total workqueue size: %u\n", idxd->max_wq_size);
504 	idxd->max_wqs = idxd->hw.wq_cap.num_wqs;
505 	dev_dbg(dev, "max workqueues: %u\n", idxd->max_wqs);
506 	idxd->wqcfg_size = 1 << (idxd->hw.wq_cap.wqcfg_size + IDXD_WQCFG_MIN);
507 	dev_dbg(dev, "wqcfg size: %u\n", idxd->wqcfg_size);
508 
509 	/* reading operation capabilities */
510 	for (i = 0; i < 4; i++) {
511 		idxd->hw.opcap.bits[i] = ioread64(idxd->reg_base +
512 				IDXD_OPCAP_OFFSET + i * sizeof(u64));
513 		dev_dbg(dev, "opcap[%d]: %#llx\n", i, idxd->hw.opcap.bits[i]);
514 	}
515 	multi_u64_to_bmap(idxd->opcap_bmap, &idxd->hw.opcap.bits[0], 4);
516 
517 	/* read iaa cap */
518 	if (idxd->data->type == IDXD_TYPE_IAX && idxd->hw.version >= DEVICE_VERSION_2)
519 		idxd->hw.iaa_cap.bits = ioread64(idxd->reg_base + IDXD_IAACAP_OFFSET);
520 }
521 
522 static struct idxd_device *idxd_alloc(struct pci_dev *pdev, struct idxd_driver_data *data)
523 {
524 	struct device *dev = &pdev->dev;
525 	struct device *conf_dev;
526 	struct idxd_device *idxd;
527 	int rc;
528 
529 	idxd = kzalloc_node(sizeof(*idxd), GFP_KERNEL, dev_to_node(dev));
530 	if (!idxd)
531 		return NULL;
532 
533 	conf_dev = idxd_confdev(idxd);
534 	idxd->pdev = pdev;
535 	idxd->data = data;
536 	idxd_dev_set_type(&idxd->idxd_dev, idxd->data->type);
537 	idxd->id = ida_alloc(&idxd_ida, GFP_KERNEL);
538 	if (idxd->id < 0)
539 		return NULL;
540 
541 	idxd->opcap_bmap = bitmap_zalloc_node(IDXD_MAX_OPCAP_BITS, GFP_KERNEL, dev_to_node(dev));
542 	if (!idxd->opcap_bmap) {
543 		ida_free(&idxd_ida, idxd->id);
544 		return NULL;
545 	}
546 
547 	device_initialize(conf_dev);
548 	conf_dev->parent = dev;
549 	conf_dev->bus = &dsa_bus_type;
550 	conf_dev->type = idxd->data->dev_type;
551 	rc = dev_set_name(conf_dev, "%s%d", idxd->data->name_prefix, idxd->id);
552 	if (rc < 0) {
553 		put_device(conf_dev);
554 		return NULL;
555 	}
556 
557 	spin_lock_init(&idxd->dev_lock);
558 	spin_lock_init(&idxd->cmd_lock);
559 
560 	return idxd;
561 }
562 
563 static int idxd_enable_system_pasid(struct idxd_device *idxd)
564 {
565 	struct pci_dev *pdev = idxd->pdev;
566 	struct device *dev = &pdev->dev;
567 	struct iommu_domain *domain;
568 	ioasid_t pasid;
569 	int ret;
570 
571 	/*
572 	 * Attach a global PASID to the DMA domain so that we can use ENQCMDS
573 	 * to submit work on buffers mapped by DMA API.
574 	 */
575 	domain = iommu_get_domain_for_dev(dev);
576 	if (!domain)
577 		return -EPERM;
578 
579 	pasid = iommu_alloc_global_pasid(dev);
580 	if (pasid == IOMMU_PASID_INVALID)
581 		return -ENOSPC;
582 
583 	/*
584 	 * DMA domain is owned by the driver, it should support all valid
585 	 * types such as DMA-FQ, identity, etc.
586 	 */
587 	ret = iommu_attach_device_pasid(domain, dev, pasid);
588 	if (ret) {
589 		dev_err(dev, "failed to attach device pasid %d, domain type %d",
590 			pasid, domain->type);
591 		iommu_free_global_pasid(pasid);
592 		return ret;
593 	}
594 
595 	/* Since we set user privilege for kernel DMA, enable completion IRQ */
596 	idxd_set_user_intr(idxd, 1);
597 	idxd->pasid = pasid;
598 
599 	return ret;
600 }
601 
602 static void idxd_disable_system_pasid(struct idxd_device *idxd)
603 {
604 	struct pci_dev *pdev = idxd->pdev;
605 	struct device *dev = &pdev->dev;
606 	struct iommu_domain *domain;
607 
608 	domain = iommu_get_domain_for_dev(dev);
609 	if (!domain)
610 		return;
611 
612 	iommu_detach_device_pasid(domain, dev, idxd->pasid);
613 	iommu_free_global_pasid(idxd->pasid);
614 
615 	idxd_set_user_intr(idxd, 0);
616 	idxd->sva = NULL;
617 	idxd->pasid = IOMMU_PASID_INVALID;
618 }
619 
620 static int idxd_enable_sva(struct pci_dev *pdev)
621 {
622 	int ret;
623 
624 	ret = iommu_dev_enable_feature(&pdev->dev, IOMMU_DEV_FEAT_IOPF);
625 	if (ret)
626 		return ret;
627 
628 	ret = iommu_dev_enable_feature(&pdev->dev, IOMMU_DEV_FEAT_SVA);
629 	if (ret)
630 		iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_IOPF);
631 
632 	return ret;
633 }
634 
635 static void idxd_disable_sva(struct pci_dev *pdev)
636 {
637 	iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_SVA);
638 	iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_IOPF);
639 }
640 
641 static int idxd_probe(struct idxd_device *idxd)
642 {
643 	struct pci_dev *pdev = idxd->pdev;
644 	struct device *dev = &pdev->dev;
645 	int rc;
646 
647 	dev_dbg(dev, "%s entered and resetting device\n", __func__);
648 	rc = idxd_device_init_reset(idxd);
649 	if (rc < 0)
650 		return rc;
651 
652 	dev_dbg(dev, "IDXD reset complete\n");
653 
654 	if (IS_ENABLED(CONFIG_INTEL_IDXD_SVM) && sva) {
655 		if (idxd_enable_sva(pdev)) {
656 			dev_warn(dev, "Unable to turn on user SVA feature.\n");
657 		} else {
658 			set_bit(IDXD_FLAG_USER_PASID_ENABLED, &idxd->flags);
659 
660 			rc = idxd_enable_system_pasid(idxd);
661 			if (rc)
662 				dev_warn(dev, "No in-kernel DMA with PASID. %d\n", rc);
663 			else
664 				set_bit(IDXD_FLAG_PASID_ENABLED, &idxd->flags);
665 		}
666 	} else if (!sva) {
667 		dev_warn(dev, "User forced SVA off via module param.\n");
668 	}
669 
670 	idxd_read_caps(idxd);
671 	idxd_read_table_offsets(idxd);
672 
673 	rc = idxd_setup_internals(idxd);
674 	if (rc)
675 		goto err;
676 
677 	/* If the configs are readonly, then load them from device */
678 	if (!test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags)) {
679 		dev_dbg(dev, "Loading RO device config\n");
680 		rc = idxd_device_load_config(idxd);
681 		if (rc < 0)
682 			goto err_config;
683 	}
684 
685 	rc = idxd_setup_interrupts(idxd);
686 	if (rc)
687 		goto err_config;
688 
689 	idxd->major = idxd_cdev_get_major(idxd);
690 
691 	rc = perfmon_pmu_init(idxd);
692 	if (rc < 0)
693 		dev_warn(dev, "Failed to initialize perfmon. No PMU support: %d\n", rc);
694 
695 	dev_dbg(dev, "IDXD device %d probed successfully\n", idxd->id);
696 	return 0;
697 
698  err_config:
699 	idxd_cleanup_internals(idxd);
700  err:
701 	if (device_pasid_enabled(idxd))
702 		idxd_disable_system_pasid(idxd);
703 	if (device_user_pasid_enabled(idxd))
704 		idxd_disable_sva(pdev);
705 	return rc;
706 }
707 
708 static void idxd_cleanup(struct idxd_device *idxd)
709 {
710 	perfmon_pmu_remove(idxd);
711 	idxd_cleanup_interrupts(idxd);
712 	idxd_cleanup_internals(idxd);
713 	if (device_pasid_enabled(idxd))
714 		idxd_disable_system_pasid(idxd);
715 	if (device_user_pasid_enabled(idxd))
716 		idxd_disable_sva(idxd->pdev);
717 }
718 
719 static int idxd_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
720 {
721 	struct device *dev = &pdev->dev;
722 	struct idxd_device *idxd;
723 	struct idxd_driver_data *data = (struct idxd_driver_data *)id->driver_data;
724 	int rc;
725 
726 	rc = pci_enable_device(pdev);
727 	if (rc)
728 		return rc;
729 
730 	dev_dbg(dev, "Alloc IDXD context\n");
731 	idxd = idxd_alloc(pdev, data);
732 	if (!idxd) {
733 		rc = -ENOMEM;
734 		goto err_idxd_alloc;
735 	}
736 
737 	dev_dbg(dev, "Mapping BARs\n");
738 	idxd->reg_base = pci_iomap(pdev, IDXD_MMIO_BAR, 0);
739 	if (!idxd->reg_base) {
740 		rc = -ENOMEM;
741 		goto err_iomap;
742 	}
743 
744 	dev_dbg(dev, "Set DMA masks\n");
745 	rc = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
746 	if (rc)
747 		goto err;
748 
749 	dev_dbg(dev, "Set PCI master\n");
750 	pci_set_master(pdev);
751 	pci_set_drvdata(pdev, idxd);
752 
753 	idxd->hw.version = ioread32(idxd->reg_base + IDXD_VER_OFFSET);
754 	rc = idxd_probe(idxd);
755 	if (rc) {
756 		dev_err(dev, "Intel(R) IDXD DMA Engine init failed\n");
757 		goto err;
758 	}
759 
760 	if (data->load_device_defaults) {
761 		rc = data->load_device_defaults(idxd);
762 		if (rc)
763 			dev_warn(dev, "IDXD loading device defaults failed\n");
764 	}
765 
766 	rc = idxd_register_devices(idxd);
767 	if (rc) {
768 		dev_err(dev, "IDXD sysfs setup failed\n");
769 		goto err_dev_register;
770 	}
771 
772 	rc = idxd_device_init_debugfs(idxd);
773 	if (rc)
774 		dev_warn(dev, "IDXD debugfs failed to setup\n");
775 
776 	dev_info(&pdev->dev, "Intel(R) Accelerator Device (v%x)\n",
777 		 idxd->hw.version);
778 
779 	idxd->user_submission_safe = data->user_submission_safe;
780 
781 	return 0;
782 
783  err_dev_register:
784 	idxd_cleanup(idxd);
785  err:
786 	pci_iounmap(pdev, idxd->reg_base);
787  err_iomap:
788 	put_device(idxd_confdev(idxd));
789  err_idxd_alloc:
790 	pci_disable_device(pdev);
791 	return rc;
792 }
793 
794 void idxd_wqs_quiesce(struct idxd_device *idxd)
795 {
796 	struct idxd_wq *wq;
797 	int i;
798 
799 	for (i = 0; i < idxd->max_wqs; i++) {
800 		wq = idxd->wqs[i];
801 		if (wq->state == IDXD_WQ_ENABLED && wq->type == IDXD_WQT_KERNEL)
802 			idxd_wq_quiesce(wq);
803 	}
804 }
805 
806 static void idxd_shutdown(struct pci_dev *pdev)
807 {
808 	struct idxd_device *idxd = pci_get_drvdata(pdev);
809 	struct idxd_irq_entry *irq_entry;
810 	int rc;
811 
812 	rc = idxd_device_disable(idxd);
813 	if (rc)
814 		dev_err(&pdev->dev, "Disabling device failed\n");
815 
816 	irq_entry = &idxd->ie;
817 	synchronize_irq(irq_entry->vector);
818 	idxd_mask_error_interrupts(idxd);
819 	flush_workqueue(idxd->wq);
820 }
821 
822 static void idxd_remove(struct pci_dev *pdev)
823 {
824 	struct idxd_device *idxd = pci_get_drvdata(pdev);
825 	struct idxd_irq_entry *irq_entry;
826 
827 	idxd_unregister_devices(idxd);
828 	/*
829 	 * When ->release() is called for the idxd->conf_dev, it frees all the memory related
830 	 * to the idxd context. The driver still needs those bits in order to do the rest of
831 	 * the cleanup. However, we do need to unbound the idxd sub-driver. So take a ref
832 	 * on the device here to hold off the freeing while allowing the idxd sub-driver
833 	 * to unbind.
834 	 */
835 	get_device(idxd_confdev(idxd));
836 	device_unregister(idxd_confdev(idxd));
837 	idxd_shutdown(pdev);
838 	if (device_pasid_enabled(idxd))
839 		idxd_disable_system_pasid(idxd);
840 	idxd_device_remove_debugfs(idxd);
841 
842 	irq_entry = idxd_get_ie(idxd, 0);
843 	free_irq(irq_entry->vector, irq_entry);
844 	pci_free_irq_vectors(pdev);
845 	pci_iounmap(pdev, idxd->reg_base);
846 	if (device_user_pasid_enabled(idxd))
847 		idxd_disable_sva(pdev);
848 	pci_disable_device(pdev);
849 	destroy_workqueue(idxd->wq);
850 	perfmon_pmu_remove(idxd);
851 	put_device(idxd_confdev(idxd));
852 }
853 
854 static struct pci_driver idxd_pci_driver = {
855 	.name		= DRV_NAME,
856 	.id_table	= idxd_pci_tbl,
857 	.probe		= idxd_pci_probe,
858 	.remove		= idxd_remove,
859 	.shutdown	= idxd_shutdown,
860 };
861 
862 static int __init idxd_init_module(void)
863 {
864 	int err;
865 
866 	/*
867 	 * If the CPU does not support MOVDIR64B or ENQCMDS, there's no point in
868 	 * enumerating the device. We can not utilize it.
869 	 */
870 	if (!cpu_feature_enabled(X86_FEATURE_MOVDIR64B)) {
871 		pr_warn("idxd driver failed to load without MOVDIR64B.\n");
872 		return -ENODEV;
873 	}
874 
875 	if (!cpu_feature_enabled(X86_FEATURE_ENQCMD))
876 		pr_warn("Platform does not have ENQCMD(S) support.\n");
877 	else
878 		support_enqcmd = true;
879 
880 	perfmon_init();
881 
882 	err = idxd_driver_register(&idxd_drv);
883 	if (err < 0)
884 		goto err_idxd_driver_register;
885 
886 	err = idxd_driver_register(&idxd_dmaengine_drv);
887 	if (err < 0)
888 		goto err_idxd_dmaengine_driver_register;
889 
890 	err = idxd_driver_register(&idxd_user_drv);
891 	if (err < 0)
892 		goto err_idxd_user_driver_register;
893 
894 	err = idxd_cdev_register();
895 	if (err)
896 		goto err_cdev_register;
897 
898 	err = idxd_init_debugfs();
899 	if (err)
900 		goto err_debugfs;
901 
902 	err = pci_register_driver(&idxd_pci_driver);
903 	if (err)
904 		goto err_pci_register;
905 
906 	return 0;
907 
908 err_pci_register:
909 	idxd_remove_debugfs();
910 err_debugfs:
911 	idxd_cdev_remove();
912 err_cdev_register:
913 	idxd_driver_unregister(&idxd_user_drv);
914 err_idxd_user_driver_register:
915 	idxd_driver_unregister(&idxd_dmaengine_drv);
916 err_idxd_dmaengine_driver_register:
917 	idxd_driver_unregister(&idxd_drv);
918 err_idxd_driver_register:
919 	return err;
920 }
921 module_init(idxd_init_module);
922 
923 static void __exit idxd_exit_module(void)
924 {
925 	idxd_driver_unregister(&idxd_user_drv);
926 	idxd_driver_unregister(&idxd_dmaengine_drv);
927 	idxd_driver_unregister(&idxd_drv);
928 	pci_unregister_driver(&idxd_pci_driver);
929 	idxd_cdev_remove();
930 	perfmon_exit();
931 	idxd_remove_debugfs();
932 }
933 module_exit(idxd_exit_module);
934