xref: /openbmc/linux/drivers/dma/idxd/init.c (revision 2ae1beb3)
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 	},
65 };
66 
67 static struct pci_device_id idxd_pci_tbl[] = {
68 	/* DSA ver 1.0 platforms */
69 	{ PCI_DEVICE_DATA(INTEL, DSA_SPR0, &idxd_driver_data[IDXD_TYPE_DSA]) },
70 
71 	/* IAX ver 1.0 platforms */
72 	{ PCI_DEVICE_DATA(INTEL, IAX_SPR0, &idxd_driver_data[IDXD_TYPE_IAX]) },
73 	{ 0, }
74 };
75 MODULE_DEVICE_TABLE(pci, idxd_pci_tbl);
76 
77 static int idxd_setup_interrupts(struct idxd_device *idxd)
78 {
79 	struct pci_dev *pdev = idxd->pdev;
80 	struct device *dev = &pdev->dev;
81 	struct idxd_irq_entry *ie;
82 	int i, msixcnt;
83 	int rc = 0;
84 
85 	msixcnt = pci_msix_vec_count(pdev);
86 	if (msixcnt < 0) {
87 		dev_err(dev, "Not MSI-X interrupt capable.\n");
88 		return -ENOSPC;
89 	}
90 	idxd->irq_cnt = msixcnt;
91 
92 	rc = pci_alloc_irq_vectors(pdev, msixcnt, msixcnt, PCI_IRQ_MSIX);
93 	if (rc != msixcnt) {
94 		dev_err(dev, "Failed enabling %d MSIX entries: %d\n", msixcnt, rc);
95 		return -ENOSPC;
96 	}
97 	dev_dbg(dev, "Enabled %d msix vectors\n", msixcnt);
98 
99 
100 	ie = idxd_get_ie(idxd, 0);
101 	ie->vector = pci_irq_vector(pdev, 0);
102 	rc = request_threaded_irq(ie->vector, NULL, idxd_misc_thread, 0, "idxd-misc", ie);
103 	if (rc < 0) {
104 		dev_err(dev, "Failed to allocate misc interrupt.\n");
105 		goto err_misc_irq;
106 	}
107 	dev_dbg(dev, "Requested idxd-misc handler on msix vector %d\n", ie->vector);
108 
109 	for (i = 0; i < idxd->max_wqs; i++) {
110 		int msix_idx = i + 1;
111 
112 		ie = idxd_get_ie(idxd, msix_idx);
113 		ie->id = msix_idx;
114 		ie->int_handle = INVALID_INT_HANDLE;
115 		ie->pasid = IOMMU_PASID_INVALID;
116 
117 		spin_lock_init(&ie->list_lock);
118 		init_llist_head(&ie->pending_llist);
119 		INIT_LIST_HEAD(&ie->work_list);
120 	}
121 
122 	idxd_unmask_error_interrupts(idxd);
123 	return 0;
124 
125  err_misc_irq:
126 	idxd_mask_error_interrupts(idxd);
127 	pci_free_irq_vectors(pdev);
128 	dev_err(dev, "No usable interrupts\n");
129 	return rc;
130 }
131 
132 static void idxd_cleanup_interrupts(struct idxd_device *idxd)
133 {
134 	struct pci_dev *pdev = idxd->pdev;
135 	struct idxd_irq_entry *ie;
136 	int msixcnt;
137 
138 	msixcnt = pci_msix_vec_count(pdev);
139 	if (msixcnt <= 0)
140 		return;
141 
142 	ie = idxd_get_ie(idxd, 0);
143 	idxd_mask_error_interrupts(idxd);
144 	free_irq(ie->vector, ie);
145 	pci_free_irq_vectors(pdev);
146 }
147 
148 static int idxd_setup_wqs(struct idxd_device *idxd)
149 {
150 	struct device *dev = &idxd->pdev->dev;
151 	struct idxd_wq *wq;
152 	struct device *conf_dev;
153 	int i, rc;
154 
155 	idxd->wqs = kcalloc_node(idxd->max_wqs, sizeof(struct idxd_wq *),
156 				 GFP_KERNEL, dev_to_node(dev));
157 	if (!idxd->wqs)
158 		return -ENOMEM;
159 
160 	idxd->wq_enable_map = bitmap_zalloc_node(idxd->max_wqs, GFP_KERNEL, dev_to_node(dev));
161 	if (!idxd->wq_enable_map) {
162 		kfree(idxd->wqs);
163 		return -ENOMEM;
164 	}
165 
166 	for (i = 0; i < idxd->max_wqs; i++) {
167 		wq = kzalloc_node(sizeof(*wq), GFP_KERNEL, dev_to_node(dev));
168 		if (!wq) {
169 			rc = -ENOMEM;
170 			goto err;
171 		}
172 
173 		idxd_dev_set_type(&wq->idxd_dev, IDXD_DEV_WQ);
174 		conf_dev = wq_confdev(wq);
175 		wq->id = i;
176 		wq->idxd = idxd;
177 		device_initialize(wq_confdev(wq));
178 		conf_dev->parent = idxd_confdev(idxd);
179 		conf_dev->bus = &dsa_bus_type;
180 		conf_dev->type = &idxd_wq_device_type;
181 		rc = dev_set_name(conf_dev, "wq%d.%d", idxd->id, wq->id);
182 		if (rc < 0) {
183 			put_device(conf_dev);
184 			goto err;
185 		}
186 
187 		mutex_init(&wq->wq_lock);
188 		init_waitqueue_head(&wq->err_queue);
189 		init_completion(&wq->wq_dead);
190 		init_completion(&wq->wq_resurrect);
191 		wq->max_xfer_bytes = WQ_DEFAULT_MAX_XFER;
192 		idxd_wq_set_max_batch_size(idxd->data->type, wq, WQ_DEFAULT_MAX_BATCH);
193 		wq->enqcmds_retries = IDXD_ENQCMDS_RETRIES;
194 		wq->wqcfg = kzalloc_node(idxd->wqcfg_size, GFP_KERNEL, dev_to_node(dev));
195 		if (!wq->wqcfg) {
196 			put_device(conf_dev);
197 			rc = -ENOMEM;
198 			goto err;
199 		}
200 
201 		if (idxd->hw.wq_cap.op_config) {
202 			wq->opcap_bmap = bitmap_zalloc(IDXD_MAX_OPCAP_BITS, GFP_KERNEL);
203 			if (!wq->opcap_bmap) {
204 				put_device(conf_dev);
205 				rc = -ENOMEM;
206 				goto err;
207 			}
208 			bitmap_copy(wq->opcap_bmap, idxd->opcap_bmap, IDXD_MAX_OPCAP_BITS);
209 		}
210 		mutex_init(&wq->uc_lock);
211 		xa_init(&wq->upasid_xa);
212 		idxd->wqs[i] = wq;
213 	}
214 
215 	return 0;
216 
217  err:
218 	while (--i >= 0) {
219 		wq = idxd->wqs[i];
220 		conf_dev = wq_confdev(wq);
221 		put_device(conf_dev);
222 	}
223 	return rc;
224 }
225 
226 static int idxd_setup_engines(struct idxd_device *idxd)
227 {
228 	struct idxd_engine *engine;
229 	struct device *dev = &idxd->pdev->dev;
230 	struct device *conf_dev;
231 	int i, rc;
232 
233 	idxd->engines = kcalloc_node(idxd->max_engines, sizeof(struct idxd_engine *),
234 				     GFP_KERNEL, dev_to_node(dev));
235 	if (!idxd->engines)
236 		return -ENOMEM;
237 
238 	for (i = 0; i < idxd->max_engines; i++) {
239 		engine = kzalloc_node(sizeof(*engine), GFP_KERNEL, dev_to_node(dev));
240 		if (!engine) {
241 			rc = -ENOMEM;
242 			goto err;
243 		}
244 
245 		idxd_dev_set_type(&engine->idxd_dev, IDXD_DEV_ENGINE);
246 		conf_dev = engine_confdev(engine);
247 		engine->id = i;
248 		engine->idxd = idxd;
249 		device_initialize(conf_dev);
250 		conf_dev->parent = idxd_confdev(idxd);
251 		conf_dev->bus = &dsa_bus_type;
252 		conf_dev->type = &idxd_engine_device_type;
253 		rc = dev_set_name(conf_dev, "engine%d.%d", idxd->id, engine->id);
254 		if (rc < 0) {
255 			put_device(conf_dev);
256 			goto err;
257 		}
258 
259 		idxd->engines[i] = engine;
260 	}
261 
262 	return 0;
263 
264  err:
265 	while (--i >= 0) {
266 		engine = idxd->engines[i];
267 		conf_dev = engine_confdev(engine);
268 		put_device(conf_dev);
269 	}
270 	return rc;
271 }
272 
273 static int idxd_setup_groups(struct idxd_device *idxd)
274 {
275 	struct device *dev = &idxd->pdev->dev;
276 	struct device *conf_dev;
277 	struct idxd_group *group;
278 	int i, rc;
279 
280 	idxd->groups = kcalloc_node(idxd->max_groups, sizeof(struct idxd_group *),
281 				    GFP_KERNEL, dev_to_node(dev));
282 	if (!idxd->groups)
283 		return -ENOMEM;
284 
285 	for (i = 0; i < idxd->max_groups; i++) {
286 		group = kzalloc_node(sizeof(*group), GFP_KERNEL, dev_to_node(dev));
287 		if (!group) {
288 			rc = -ENOMEM;
289 			goto err;
290 		}
291 
292 		idxd_dev_set_type(&group->idxd_dev, IDXD_DEV_GROUP);
293 		conf_dev = group_confdev(group);
294 		group->id = i;
295 		group->idxd = idxd;
296 		device_initialize(conf_dev);
297 		conf_dev->parent = idxd_confdev(idxd);
298 		conf_dev->bus = &dsa_bus_type;
299 		conf_dev->type = &idxd_group_device_type;
300 		rc = dev_set_name(conf_dev, "group%d.%d", idxd->id, group->id);
301 		if (rc < 0) {
302 			put_device(conf_dev);
303 			goto err;
304 		}
305 
306 		idxd->groups[i] = group;
307 		if (idxd->hw.version <= DEVICE_VERSION_2 && !tc_override) {
308 			group->tc_a = 1;
309 			group->tc_b = 1;
310 		} else {
311 			group->tc_a = -1;
312 			group->tc_b = -1;
313 		}
314 		/*
315 		 * The default value is the same as the value of
316 		 * total read buffers in GRPCAP.
317 		 */
318 		group->rdbufs_allowed = idxd->max_rdbufs;
319 	}
320 
321 	return 0;
322 
323  err:
324 	while (--i >= 0) {
325 		group = idxd->groups[i];
326 		put_device(group_confdev(group));
327 	}
328 	return rc;
329 }
330 
331 static void idxd_cleanup_internals(struct idxd_device *idxd)
332 {
333 	int i;
334 
335 	for (i = 0; i < idxd->max_groups; i++)
336 		put_device(group_confdev(idxd->groups[i]));
337 	for (i = 0; i < idxd->max_engines; i++)
338 		put_device(engine_confdev(idxd->engines[i]));
339 	for (i = 0; i < idxd->max_wqs; i++)
340 		put_device(wq_confdev(idxd->wqs[i]));
341 	destroy_workqueue(idxd->wq);
342 }
343 
344 static int idxd_init_evl(struct idxd_device *idxd)
345 {
346 	struct device *dev = &idxd->pdev->dev;
347 	unsigned int evl_cache_size;
348 	struct idxd_evl *evl;
349 	const char *idxd_name;
350 
351 	if (idxd->hw.gen_cap.evl_support == 0)
352 		return 0;
353 
354 	evl = kzalloc_node(sizeof(*evl), GFP_KERNEL, dev_to_node(dev));
355 	if (!evl)
356 		return -ENOMEM;
357 
358 	mutex_init(&evl->lock);
359 	evl->size = IDXD_EVL_SIZE_MIN;
360 
361 	idxd_name = dev_name(idxd_confdev(idxd));
362 	evl_cache_size = sizeof(struct idxd_evl_fault) + evl_ent_size(idxd);
363 	/*
364 	 * Since completion record in evl_cache will be copied to user
365 	 * when handling completion record page fault, need to create
366 	 * the cache suitable for user copy.
367 	 */
368 	idxd->evl_cache = kmem_cache_create_usercopy(idxd_name, evl_cache_size,
369 						     0, 0, 0, evl_cache_size,
370 						     NULL);
371 	if (!idxd->evl_cache) {
372 		kfree(evl);
373 		return -ENOMEM;
374 	}
375 
376 	idxd->evl = evl;
377 	return 0;
378 }
379 
380 static int idxd_setup_internals(struct idxd_device *idxd)
381 {
382 	struct device *dev = &idxd->pdev->dev;
383 	int rc, i;
384 
385 	init_waitqueue_head(&idxd->cmd_waitq);
386 
387 	rc = idxd_setup_wqs(idxd);
388 	if (rc < 0)
389 		goto err_wqs;
390 
391 	rc = idxd_setup_engines(idxd);
392 	if (rc < 0)
393 		goto err_engine;
394 
395 	rc = idxd_setup_groups(idxd);
396 	if (rc < 0)
397 		goto err_group;
398 
399 	idxd->wq = create_workqueue(dev_name(dev));
400 	if (!idxd->wq) {
401 		rc = -ENOMEM;
402 		goto err_wkq_create;
403 	}
404 
405 	rc = idxd_init_evl(idxd);
406 	if (rc < 0)
407 		goto err_evl;
408 
409 	return 0;
410 
411  err_evl:
412 	destroy_workqueue(idxd->wq);
413  err_wkq_create:
414 	for (i = 0; i < idxd->max_groups; i++)
415 		put_device(group_confdev(idxd->groups[i]));
416  err_group:
417 	for (i = 0; i < idxd->max_engines; i++)
418 		put_device(engine_confdev(idxd->engines[i]));
419  err_engine:
420 	for (i = 0; i < idxd->max_wqs; i++)
421 		put_device(wq_confdev(idxd->wqs[i]));
422  err_wqs:
423 	return rc;
424 }
425 
426 static void idxd_read_table_offsets(struct idxd_device *idxd)
427 {
428 	union offsets_reg offsets;
429 	struct device *dev = &idxd->pdev->dev;
430 
431 	offsets.bits[0] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET);
432 	offsets.bits[1] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET + sizeof(u64));
433 	idxd->grpcfg_offset = offsets.grpcfg * IDXD_TABLE_MULT;
434 	dev_dbg(dev, "IDXD Group Config Offset: %#x\n", idxd->grpcfg_offset);
435 	idxd->wqcfg_offset = offsets.wqcfg * IDXD_TABLE_MULT;
436 	dev_dbg(dev, "IDXD Work Queue Config Offset: %#x\n", idxd->wqcfg_offset);
437 	idxd->msix_perm_offset = offsets.msix_perm * IDXD_TABLE_MULT;
438 	dev_dbg(dev, "IDXD MSIX Permission Offset: %#x\n", idxd->msix_perm_offset);
439 	idxd->perfmon_offset = offsets.perfmon * IDXD_TABLE_MULT;
440 	dev_dbg(dev, "IDXD Perfmon Offset: %#x\n", idxd->perfmon_offset);
441 }
442 
443 void multi_u64_to_bmap(unsigned long *bmap, u64 *val, int count)
444 {
445 	int i, j, nr;
446 
447 	for (i = 0, nr = 0; i < count; i++) {
448 		for (j = 0; j < BITS_PER_LONG_LONG; j++) {
449 			if (val[i] & BIT(j))
450 				set_bit(nr, bmap);
451 			nr++;
452 		}
453 	}
454 }
455 
456 static void idxd_read_caps(struct idxd_device *idxd)
457 {
458 	struct device *dev = &idxd->pdev->dev;
459 	int i;
460 
461 	/* reading generic capabilities */
462 	idxd->hw.gen_cap.bits = ioread64(idxd->reg_base + IDXD_GENCAP_OFFSET);
463 	dev_dbg(dev, "gen_cap: %#llx\n", idxd->hw.gen_cap.bits);
464 
465 	if (idxd->hw.gen_cap.cmd_cap) {
466 		idxd->hw.cmd_cap = ioread32(idxd->reg_base + IDXD_CMDCAP_OFFSET);
467 		dev_dbg(dev, "cmd_cap: %#x\n", idxd->hw.cmd_cap);
468 	}
469 
470 	/* reading command capabilities */
471 	if (idxd->hw.cmd_cap & BIT(IDXD_CMD_REQUEST_INT_HANDLE))
472 		idxd->request_int_handles = true;
473 
474 	idxd->max_xfer_bytes = 1ULL << idxd->hw.gen_cap.max_xfer_shift;
475 	dev_dbg(dev, "max xfer size: %llu bytes\n", idxd->max_xfer_bytes);
476 	idxd_set_max_batch_size(idxd->data->type, idxd, 1U << idxd->hw.gen_cap.max_batch_shift);
477 	dev_dbg(dev, "max batch size: %u\n", idxd->max_batch_size);
478 	if (idxd->hw.gen_cap.config_en)
479 		set_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags);
480 
481 	/* reading group capabilities */
482 	idxd->hw.group_cap.bits =
483 		ioread64(idxd->reg_base + IDXD_GRPCAP_OFFSET);
484 	dev_dbg(dev, "group_cap: %#llx\n", idxd->hw.group_cap.bits);
485 	idxd->max_groups = idxd->hw.group_cap.num_groups;
486 	dev_dbg(dev, "max groups: %u\n", idxd->max_groups);
487 	idxd->max_rdbufs = idxd->hw.group_cap.total_rdbufs;
488 	dev_dbg(dev, "max read buffers: %u\n", idxd->max_rdbufs);
489 	idxd->nr_rdbufs = idxd->max_rdbufs;
490 
491 	/* read engine capabilities */
492 	idxd->hw.engine_cap.bits =
493 		ioread64(idxd->reg_base + IDXD_ENGCAP_OFFSET);
494 	dev_dbg(dev, "engine_cap: %#llx\n", idxd->hw.engine_cap.bits);
495 	idxd->max_engines = idxd->hw.engine_cap.num_engines;
496 	dev_dbg(dev, "max engines: %u\n", idxd->max_engines);
497 
498 	/* read workqueue capabilities */
499 	idxd->hw.wq_cap.bits = ioread64(idxd->reg_base + IDXD_WQCAP_OFFSET);
500 	dev_dbg(dev, "wq_cap: %#llx\n", idxd->hw.wq_cap.bits);
501 	idxd->max_wq_size = idxd->hw.wq_cap.total_wq_size;
502 	dev_dbg(dev, "total workqueue size: %u\n", idxd->max_wq_size);
503 	idxd->max_wqs = idxd->hw.wq_cap.num_wqs;
504 	dev_dbg(dev, "max workqueues: %u\n", idxd->max_wqs);
505 	idxd->wqcfg_size = 1 << (idxd->hw.wq_cap.wqcfg_size + IDXD_WQCFG_MIN);
506 	dev_dbg(dev, "wqcfg size: %u\n", idxd->wqcfg_size);
507 
508 	/* reading operation capabilities */
509 	for (i = 0; i < 4; i++) {
510 		idxd->hw.opcap.bits[i] = ioread64(idxd->reg_base +
511 				IDXD_OPCAP_OFFSET + i * sizeof(u64));
512 		dev_dbg(dev, "opcap[%d]: %#llx\n", i, idxd->hw.opcap.bits[i]);
513 	}
514 	multi_u64_to_bmap(idxd->opcap_bmap, &idxd->hw.opcap.bits[0], 4);
515 
516 	/* read iaa cap */
517 	if (idxd->data->type == IDXD_TYPE_IAX && idxd->hw.version >= DEVICE_VERSION_2)
518 		idxd->hw.iaa_cap.bits = ioread64(idxd->reg_base + IDXD_IAACAP_OFFSET);
519 }
520 
521 static struct idxd_device *idxd_alloc(struct pci_dev *pdev, struct idxd_driver_data *data)
522 {
523 	struct device *dev = &pdev->dev;
524 	struct device *conf_dev;
525 	struct idxd_device *idxd;
526 	int rc;
527 
528 	idxd = kzalloc_node(sizeof(*idxd), GFP_KERNEL, dev_to_node(dev));
529 	if (!idxd)
530 		return NULL;
531 
532 	conf_dev = idxd_confdev(idxd);
533 	idxd->pdev = pdev;
534 	idxd->data = data;
535 	idxd_dev_set_type(&idxd->idxd_dev, idxd->data->type);
536 	idxd->id = ida_alloc(&idxd_ida, GFP_KERNEL);
537 	if (idxd->id < 0)
538 		return NULL;
539 
540 	idxd->opcap_bmap = bitmap_zalloc_node(IDXD_MAX_OPCAP_BITS, GFP_KERNEL, dev_to_node(dev));
541 	if (!idxd->opcap_bmap) {
542 		ida_free(&idxd_ida, idxd->id);
543 		return NULL;
544 	}
545 
546 	device_initialize(conf_dev);
547 	conf_dev->parent = dev;
548 	conf_dev->bus = &dsa_bus_type;
549 	conf_dev->type = idxd->data->dev_type;
550 	rc = dev_set_name(conf_dev, "%s%d", idxd->data->name_prefix, idxd->id);
551 	if (rc < 0) {
552 		put_device(conf_dev);
553 		return NULL;
554 	}
555 
556 	spin_lock_init(&idxd->dev_lock);
557 	spin_lock_init(&idxd->cmd_lock);
558 
559 	return idxd;
560 }
561 
562 static int idxd_enable_system_pasid(struct idxd_device *idxd)
563 {
564 	struct pci_dev *pdev = idxd->pdev;
565 	struct device *dev = &pdev->dev;
566 	struct iommu_domain *domain;
567 	ioasid_t pasid;
568 	int ret;
569 
570 	/*
571 	 * Attach a global PASID to the DMA domain so that we can use ENQCMDS
572 	 * to submit work on buffers mapped by DMA API.
573 	 */
574 	domain = iommu_get_domain_for_dev(dev);
575 	if (!domain)
576 		return -EPERM;
577 
578 	pasid = iommu_alloc_global_pasid(dev);
579 	if (pasid == IOMMU_PASID_INVALID)
580 		return -ENOSPC;
581 
582 	/*
583 	 * DMA domain is owned by the driver, it should support all valid
584 	 * types such as DMA-FQ, identity, etc.
585 	 */
586 	ret = iommu_attach_device_pasid(domain, dev, pasid);
587 	if (ret) {
588 		dev_err(dev, "failed to attach device pasid %d, domain type %d",
589 			pasid, domain->type);
590 		iommu_free_global_pasid(pasid);
591 		return ret;
592 	}
593 
594 	/* Since we set user privilege for kernel DMA, enable completion IRQ */
595 	idxd_set_user_intr(idxd, 1);
596 	idxd->pasid = pasid;
597 
598 	return ret;
599 }
600 
601 static void idxd_disable_system_pasid(struct idxd_device *idxd)
602 {
603 	struct pci_dev *pdev = idxd->pdev;
604 	struct device *dev = &pdev->dev;
605 	struct iommu_domain *domain;
606 
607 	domain = iommu_get_domain_for_dev(dev);
608 	if (!domain)
609 		return;
610 
611 	iommu_detach_device_pasid(domain, dev, idxd->pasid);
612 	iommu_free_global_pasid(idxd->pasid);
613 
614 	idxd_set_user_intr(idxd, 0);
615 	idxd->sva = NULL;
616 	idxd->pasid = IOMMU_PASID_INVALID;
617 }
618 
619 static int idxd_enable_sva(struct pci_dev *pdev)
620 {
621 	int ret;
622 
623 	ret = iommu_dev_enable_feature(&pdev->dev, IOMMU_DEV_FEAT_IOPF);
624 	if (ret)
625 		return ret;
626 
627 	ret = iommu_dev_enable_feature(&pdev->dev, IOMMU_DEV_FEAT_SVA);
628 	if (ret)
629 		iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_IOPF);
630 
631 	return ret;
632 }
633 
634 static void idxd_disable_sva(struct pci_dev *pdev)
635 {
636 	iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_SVA);
637 	iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_IOPF);
638 }
639 
640 static int idxd_probe(struct idxd_device *idxd)
641 {
642 	struct pci_dev *pdev = idxd->pdev;
643 	struct device *dev = &pdev->dev;
644 	int rc;
645 
646 	dev_dbg(dev, "%s entered and resetting device\n", __func__);
647 	rc = idxd_device_init_reset(idxd);
648 	if (rc < 0)
649 		return rc;
650 
651 	dev_dbg(dev, "IDXD reset complete\n");
652 
653 	if (IS_ENABLED(CONFIG_INTEL_IDXD_SVM) && sva) {
654 		if (idxd_enable_sva(pdev)) {
655 			dev_warn(dev, "Unable to turn on user SVA feature.\n");
656 		} else {
657 			set_bit(IDXD_FLAG_USER_PASID_ENABLED, &idxd->flags);
658 
659 			rc = idxd_enable_system_pasid(idxd);
660 			if (rc)
661 				dev_warn(dev, "No in-kernel DMA with PASID. %d\n", rc);
662 			else
663 				set_bit(IDXD_FLAG_PASID_ENABLED, &idxd->flags);
664 		}
665 	} else if (!sva) {
666 		dev_warn(dev, "User forced SVA off via module param.\n");
667 	}
668 
669 	idxd_read_caps(idxd);
670 	idxd_read_table_offsets(idxd);
671 
672 	rc = idxd_setup_internals(idxd);
673 	if (rc)
674 		goto err;
675 
676 	/* If the configs are readonly, then load them from device */
677 	if (!test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags)) {
678 		dev_dbg(dev, "Loading RO device config\n");
679 		rc = idxd_device_load_config(idxd);
680 		if (rc < 0)
681 			goto err_config;
682 	}
683 
684 	rc = idxd_setup_interrupts(idxd);
685 	if (rc)
686 		goto err_config;
687 
688 	idxd->major = idxd_cdev_get_major(idxd);
689 
690 	rc = perfmon_pmu_init(idxd);
691 	if (rc < 0)
692 		dev_warn(dev, "Failed to initialize perfmon. No PMU support: %d\n", rc);
693 
694 	dev_dbg(dev, "IDXD device %d probed successfully\n", idxd->id);
695 	return 0;
696 
697  err_config:
698 	idxd_cleanup_internals(idxd);
699  err:
700 	if (device_pasid_enabled(idxd))
701 		idxd_disable_system_pasid(idxd);
702 	if (device_user_pasid_enabled(idxd))
703 		idxd_disable_sva(pdev);
704 	return rc;
705 }
706 
707 static void idxd_cleanup(struct idxd_device *idxd)
708 {
709 	perfmon_pmu_remove(idxd);
710 	idxd_cleanup_interrupts(idxd);
711 	idxd_cleanup_internals(idxd);
712 	if (device_pasid_enabled(idxd))
713 		idxd_disable_system_pasid(idxd);
714 	if (device_user_pasid_enabled(idxd))
715 		idxd_disable_sva(idxd->pdev);
716 }
717 
718 static int idxd_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
719 {
720 	struct device *dev = &pdev->dev;
721 	struct idxd_device *idxd;
722 	struct idxd_driver_data *data = (struct idxd_driver_data *)id->driver_data;
723 	int rc;
724 
725 	rc = pci_enable_device(pdev);
726 	if (rc)
727 		return rc;
728 
729 	dev_dbg(dev, "Alloc IDXD context\n");
730 	idxd = idxd_alloc(pdev, data);
731 	if (!idxd) {
732 		rc = -ENOMEM;
733 		goto err_idxd_alloc;
734 	}
735 
736 	dev_dbg(dev, "Mapping BARs\n");
737 	idxd->reg_base = pci_iomap(pdev, IDXD_MMIO_BAR, 0);
738 	if (!idxd->reg_base) {
739 		rc = -ENOMEM;
740 		goto err_iomap;
741 	}
742 
743 	dev_dbg(dev, "Set DMA masks\n");
744 	rc = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
745 	if (rc)
746 		goto err;
747 
748 	dev_dbg(dev, "Set PCI master\n");
749 	pci_set_master(pdev);
750 	pci_set_drvdata(pdev, idxd);
751 
752 	idxd->hw.version = ioread32(idxd->reg_base + IDXD_VER_OFFSET);
753 	rc = idxd_probe(idxd);
754 	if (rc) {
755 		dev_err(dev, "Intel(R) IDXD DMA Engine init failed\n");
756 		goto err;
757 	}
758 
759 	rc = idxd_register_devices(idxd);
760 	if (rc) {
761 		dev_err(dev, "IDXD sysfs setup failed\n");
762 		goto err_dev_register;
763 	}
764 
765 	rc = idxd_device_init_debugfs(idxd);
766 	if (rc)
767 		dev_warn(dev, "IDXD debugfs failed to setup\n");
768 
769 	dev_info(&pdev->dev, "Intel(R) Accelerator Device (v%x)\n",
770 		 idxd->hw.version);
771 
772 	idxd->user_submission_safe = data->user_submission_safe;
773 
774 	return 0;
775 
776  err_dev_register:
777 	idxd_cleanup(idxd);
778  err:
779 	pci_iounmap(pdev, idxd->reg_base);
780  err_iomap:
781 	put_device(idxd_confdev(idxd));
782  err_idxd_alloc:
783 	pci_disable_device(pdev);
784 	return rc;
785 }
786 
787 void idxd_wqs_quiesce(struct idxd_device *idxd)
788 {
789 	struct idxd_wq *wq;
790 	int i;
791 
792 	for (i = 0; i < idxd->max_wqs; i++) {
793 		wq = idxd->wqs[i];
794 		if (wq->state == IDXD_WQ_ENABLED && wq->type == IDXD_WQT_KERNEL)
795 			idxd_wq_quiesce(wq);
796 	}
797 }
798 
799 static void idxd_shutdown(struct pci_dev *pdev)
800 {
801 	struct idxd_device *idxd = pci_get_drvdata(pdev);
802 	struct idxd_irq_entry *irq_entry;
803 	int rc;
804 
805 	rc = idxd_device_disable(idxd);
806 	if (rc)
807 		dev_err(&pdev->dev, "Disabling device failed\n");
808 
809 	irq_entry = &idxd->ie;
810 	synchronize_irq(irq_entry->vector);
811 	idxd_mask_error_interrupts(idxd);
812 	flush_workqueue(idxd->wq);
813 }
814 
815 static void idxd_remove(struct pci_dev *pdev)
816 {
817 	struct idxd_device *idxd = pci_get_drvdata(pdev);
818 	struct idxd_irq_entry *irq_entry;
819 
820 	idxd_unregister_devices(idxd);
821 	/*
822 	 * When ->release() is called for the idxd->conf_dev, it frees all the memory related
823 	 * to the idxd context. The driver still needs those bits in order to do the rest of
824 	 * the cleanup. However, we do need to unbound the idxd sub-driver. So take a ref
825 	 * on the device here to hold off the freeing while allowing the idxd sub-driver
826 	 * to unbind.
827 	 */
828 	get_device(idxd_confdev(idxd));
829 	device_unregister(idxd_confdev(idxd));
830 	idxd_shutdown(pdev);
831 	if (device_pasid_enabled(idxd))
832 		idxd_disable_system_pasid(idxd);
833 	idxd_device_remove_debugfs(idxd);
834 
835 	irq_entry = idxd_get_ie(idxd, 0);
836 	free_irq(irq_entry->vector, irq_entry);
837 	pci_free_irq_vectors(pdev);
838 	pci_iounmap(pdev, idxd->reg_base);
839 	if (device_user_pasid_enabled(idxd))
840 		idxd_disable_sva(pdev);
841 	pci_disable_device(pdev);
842 	destroy_workqueue(idxd->wq);
843 	perfmon_pmu_remove(idxd);
844 	put_device(idxd_confdev(idxd));
845 }
846 
847 static struct pci_driver idxd_pci_driver = {
848 	.name		= DRV_NAME,
849 	.id_table	= idxd_pci_tbl,
850 	.probe		= idxd_pci_probe,
851 	.remove		= idxd_remove,
852 	.shutdown	= idxd_shutdown,
853 };
854 
855 static int __init idxd_init_module(void)
856 {
857 	int err;
858 
859 	/*
860 	 * If the CPU does not support MOVDIR64B or ENQCMDS, there's no point in
861 	 * enumerating the device. We can not utilize it.
862 	 */
863 	if (!cpu_feature_enabled(X86_FEATURE_MOVDIR64B)) {
864 		pr_warn("idxd driver failed to load without MOVDIR64B.\n");
865 		return -ENODEV;
866 	}
867 
868 	if (!cpu_feature_enabled(X86_FEATURE_ENQCMD))
869 		pr_warn("Platform does not have ENQCMD(S) support.\n");
870 	else
871 		support_enqcmd = true;
872 
873 	perfmon_init();
874 
875 	err = idxd_driver_register(&idxd_drv);
876 	if (err < 0)
877 		goto err_idxd_driver_register;
878 
879 	err = idxd_driver_register(&idxd_dmaengine_drv);
880 	if (err < 0)
881 		goto err_idxd_dmaengine_driver_register;
882 
883 	err = idxd_driver_register(&idxd_user_drv);
884 	if (err < 0)
885 		goto err_idxd_user_driver_register;
886 
887 	err = idxd_cdev_register();
888 	if (err)
889 		goto err_cdev_register;
890 
891 	err = idxd_init_debugfs();
892 	if (err)
893 		goto err_debugfs;
894 
895 	err = pci_register_driver(&idxd_pci_driver);
896 	if (err)
897 		goto err_pci_register;
898 
899 	return 0;
900 
901 err_pci_register:
902 	idxd_remove_debugfs();
903 err_debugfs:
904 	idxd_cdev_remove();
905 err_cdev_register:
906 	idxd_driver_unregister(&idxd_user_drv);
907 err_idxd_user_driver_register:
908 	idxd_driver_unregister(&idxd_dmaengine_drv);
909 err_idxd_dmaengine_driver_register:
910 	idxd_driver_unregister(&idxd_drv);
911 err_idxd_driver_register:
912 	return err;
913 }
914 module_init(idxd_init_module);
915 
916 static void __exit idxd_exit_module(void)
917 {
918 	idxd_driver_unregister(&idxd_user_drv);
919 	idxd_driver_unregister(&idxd_dmaengine_drv);
920 	idxd_driver_unregister(&idxd_drv);
921 	pci_unregister_driver(&idxd_pci_driver);
922 	idxd_cdev_remove();
923 	perfmon_exit();
924 	idxd_remove_debugfs();
925 }
926 module_exit(idxd_exit_module);
927