xref: /linux/drivers/dma/idxd/init.c (revision 78c3925c048c752334873f56c3a3d1c9d53e0416)
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 		.cr_status_off = offsetof(struct dsa_completion_record, status),
51 		.cr_result_off = offsetof(struct dsa_completion_record, result),
52 	},
53 	[IDXD_TYPE_IAX] = {
54 		.name_prefix = "iax",
55 		.type = IDXD_TYPE_IAX,
56 		.compl_size = sizeof(struct iax_completion_record),
57 		.align = 64,
58 		.dev_type = &iax_device_type,
59 		.evl_cr_off = offsetof(struct iax_evl_entry, cr),
60 		.cr_status_off = offsetof(struct iax_completion_record, status),
61 		.cr_result_off = offsetof(struct iax_completion_record, error_code),
62 		.load_device_defaults = idxd_load_iaa_device_defaults,
63 	},
64 };
65 
66 static struct pci_device_id idxd_pci_tbl[] = {
67 	/* DSA ver 1.0 platforms */
68 	{ PCI_DEVICE_DATA(INTEL, DSA_SPR0, &idxd_driver_data[IDXD_TYPE_DSA]) },
69 
70 	/* IAX ver 1.0 platforms */
71 	{ PCI_DEVICE_DATA(INTEL, IAX_SPR0, &idxd_driver_data[IDXD_TYPE_IAX]) },
72 	{ 0, }
73 };
74 MODULE_DEVICE_TABLE(pci, idxd_pci_tbl);
75 
76 static int idxd_setup_interrupts(struct idxd_device *idxd)
77 {
78 	struct pci_dev *pdev = idxd->pdev;
79 	struct device *dev = &pdev->dev;
80 	struct idxd_irq_entry *ie;
81 	int i, msixcnt;
82 	int rc = 0;
83 
84 	msixcnt = pci_msix_vec_count(pdev);
85 	if (msixcnt < 0) {
86 		dev_err(dev, "Not MSI-X interrupt capable.\n");
87 		return -ENOSPC;
88 	}
89 	idxd->irq_cnt = msixcnt;
90 
91 	rc = pci_alloc_irq_vectors(pdev, msixcnt, msixcnt, PCI_IRQ_MSIX);
92 	if (rc != msixcnt) {
93 		dev_err(dev, "Failed enabling %d MSIX entries: %d\n", msixcnt, rc);
94 		return -ENOSPC;
95 	}
96 	dev_dbg(dev, "Enabled %d msix vectors\n", msixcnt);
97 
98 
99 	ie = idxd_get_ie(idxd, 0);
100 	ie->vector = pci_irq_vector(pdev, 0);
101 	rc = request_threaded_irq(ie->vector, NULL, idxd_misc_thread, 0, "idxd-misc", ie);
102 	if (rc < 0) {
103 		dev_err(dev, "Failed to allocate misc interrupt.\n");
104 		goto err_misc_irq;
105 	}
106 	dev_dbg(dev, "Requested idxd-misc handler on msix vector %d\n", ie->vector);
107 
108 	for (i = 0; i < idxd->max_wqs; i++) {
109 		int msix_idx = i + 1;
110 
111 		ie = idxd_get_ie(idxd, msix_idx);
112 		ie->id = msix_idx;
113 		ie->int_handle = INVALID_INT_HANDLE;
114 		ie->pasid = IOMMU_PASID_INVALID;
115 
116 		spin_lock_init(&ie->list_lock);
117 		init_llist_head(&ie->pending_llist);
118 		INIT_LIST_HEAD(&ie->work_list);
119 	}
120 
121 	idxd_unmask_error_interrupts(idxd);
122 	return 0;
123 
124  err_misc_irq:
125 	idxd_mask_error_interrupts(idxd);
126 	pci_free_irq_vectors(pdev);
127 	dev_err(dev, "No usable interrupts\n");
128 	return rc;
129 }
130 
131 static void idxd_cleanup_interrupts(struct idxd_device *idxd)
132 {
133 	struct pci_dev *pdev = idxd->pdev;
134 	struct idxd_irq_entry *ie;
135 	int msixcnt;
136 
137 	msixcnt = pci_msix_vec_count(pdev);
138 	if (msixcnt <= 0)
139 		return;
140 
141 	ie = idxd_get_ie(idxd, 0);
142 	idxd_mask_error_interrupts(idxd);
143 	free_irq(ie->vector, ie);
144 	pci_free_irq_vectors(pdev);
145 }
146 
147 static int idxd_setup_wqs(struct idxd_device *idxd)
148 {
149 	struct device *dev = &idxd->pdev->dev;
150 	struct idxd_wq *wq;
151 	struct device *conf_dev;
152 	int i, rc;
153 
154 	idxd->wqs = kcalloc_node(idxd->max_wqs, sizeof(struct idxd_wq *),
155 				 GFP_KERNEL, dev_to_node(dev));
156 	if (!idxd->wqs)
157 		return -ENOMEM;
158 
159 	idxd->wq_enable_map = bitmap_zalloc_node(idxd->max_wqs, GFP_KERNEL, dev_to_node(dev));
160 	if (!idxd->wq_enable_map) {
161 		kfree(idxd->wqs);
162 		return -ENOMEM;
163 	}
164 
165 	for (i = 0; i < idxd->max_wqs; i++) {
166 		wq = kzalloc_node(sizeof(*wq), GFP_KERNEL, dev_to_node(dev));
167 		if (!wq) {
168 			rc = -ENOMEM;
169 			goto err;
170 		}
171 
172 		idxd_dev_set_type(&wq->idxd_dev, IDXD_DEV_WQ);
173 		conf_dev = wq_confdev(wq);
174 		wq->id = i;
175 		wq->idxd = idxd;
176 		device_initialize(wq_confdev(wq));
177 		conf_dev->parent = idxd_confdev(idxd);
178 		conf_dev->bus = &dsa_bus_type;
179 		conf_dev->type = &idxd_wq_device_type;
180 		rc = dev_set_name(conf_dev, "wq%d.%d", idxd->id, wq->id);
181 		if (rc < 0) {
182 			put_device(conf_dev);
183 			goto err;
184 		}
185 
186 		mutex_init(&wq->wq_lock);
187 		init_waitqueue_head(&wq->err_queue);
188 		init_completion(&wq->wq_dead);
189 		init_completion(&wq->wq_resurrect);
190 		wq->max_xfer_bytes = WQ_DEFAULT_MAX_XFER;
191 		idxd_wq_set_max_batch_size(idxd->data->type, wq, WQ_DEFAULT_MAX_BATCH);
192 		wq->enqcmds_retries = IDXD_ENQCMDS_RETRIES;
193 		wq->wqcfg = kzalloc_node(idxd->wqcfg_size, GFP_KERNEL, dev_to_node(dev));
194 		if (!wq->wqcfg) {
195 			put_device(conf_dev);
196 			rc = -ENOMEM;
197 			goto err;
198 		}
199 
200 		if (idxd->hw.wq_cap.op_config) {
201 			wq->opcap_bmap = bitmap_zalloc(IDXD_MAX_OPCAP_BITS, GFP_KERNEL);
202 			if (!wq->opcap_bmap) {
203 				put_device(conf_dev);
204 				rc = -ENOMEM;
205 				goto err;
206 			}
207 			bitmap_copy(wq->opcap_bmap, idxd->opcap_bmap, IDXD_MAX_OPCAP_BITS);
208 		}
209 		mutex_init(&wq->uc_lock);
210 		xa_init(&wq->upasid_xa);
211 		idxd->wqs[i] = wq;
212 	}
213 
214 	return 0;
215 
216  err:
217 	while (--i >= 0) {
218 		wq = idxd->wqs[i];
219 		conf_dev = wq_confdev(wq);
220 		put_device(conf_dev);
221 	}
222 	return rc;
223 }
224 
225 static int idxd_setup_engines(struct idxd_device *idxd)
226 {
227 	struct idxd_engine *engine;
228 	struct device *dev = &idxd->pdev->dev;
229 	struct device *conf_dev;
230 	int i, rc;
231 
232 	idxd->engines = kcalloc_node(idxd->max_engines, sizeof(struct idxd_engine *),
233 				     GFP_KERNEL, dev_to_node(dev));
234 	if (!idxd->engines)
235 		return -ENOMEM;
236 
237 	for (i = 0; i < idxd->max_engines; i++) {
238 		engine = kzalloc_node(sizeof(*engine), GFP_KERNEL, dev_to_node(dev));
239 		if (!engine) {
240 			rc = -ENOMEM;
241 			goto err;
242 		}
243 
244 		idxd_dev_set_type(&engine->idxd_dev, IDXD_DEV_ENGINE);
245 		conf_dev = engine_confdev(engine);
246 		engine->id = i;
247 		engine->idxd = idxd;
248 		device_initialize(conf_dev);
249 		conf_dev->parent = idxd_confdev(idxd);
250 		conf_dev->bus = &dsa_bus_type;
251 		conf_dev->type = &idxd_engine_device_type;
252 		rc = dev_set_name(conf_dev, "engine%d.%d", idxd->id, engine->id);
253 		if (rc < 0) {
254 			put_device(conf_dev);
255 			goto err;
256 		}
257 
258 		idxd->engines[i] = engine;
259 	}
260 
261 	return 0;
262 
263  err:
264 	while (--i >= 0) {
265 		engine = idxd->engines[i];
266 		conf_dev = engine_confdev(engine);
267 		put_device(conf_dev);
268 	}
269 	return rc;
270 }
271 
272 static int idxd_setup_groups(struct idxd_device *idxd)
273 {
274 	struct device *dev = &idxd->pdev->dev;
275 	struct device *conf_dev;
276 	struct idxd_group *group;
277 	int i, rc;
278 
279 	idxd->groups = kcalloc_node(idxd->max_groups, sizeof(struct idxd_group *),
280 				    GFP_KERNEL, dev_to_node(dev));
281 	if (!idxd->groups)
282 		return -ENOMEM;
283 
284 	for (i = 0; i < idxd->max_groups; i++) {
285 		group = kzalloc_node(sizeof(*group), GFP_KERNEL, dev_to_node(dev));
286 		if (!group) {
287 			rc = -ENOMEM;
288 			goto err;
289 		}
290 
291 		idxd_dev_set_type(&group->idxd_dev, IDXD_DEV_GROUP);
292 		conf_dev = group_confdev(group);
293 		group->id = i;
294 		group->idxd = idxd;
295 		device_initialize(conf_dev);
296 		conf_dev->parent = idxd_confdev(idxd);
297 		conf_dev->bus = &dsa_bus_type;
298 		conf_dev->type = &idxd_group_device_type;
299 		rc = dev_set_name(conf_dev, "group%d.%d", idxd->id, group->id);
300 		if (rc < 0) {
301 			put_device(conf_dev);
302 			goto err;
303 		}
304 
305 		idxd->groups[i] = group;
306 		if (idxd->hw.version <= DEVICE_VERSION_2 && !tc_override) {
307 			group->tc_a = 1;
308 			group->tc_b = 1;
309 		} else {
310 			group->tc_a = -1;
311 			group->tc_b = -1;
312 		}
313 		/*
314 		 * The default value is the same as the value of
315 		 * total read buffers in GRPCAP.
316 		 */
317 		group->rdbufs_allowed = idxd->max_rdbufs;
318 	}
319 
320 	return 0;
321 
322  err:
323 	while (--i >= 0) {
324 		group = idxd->groups[i];
325 		put_device(group_confdev(group));
326 	}
327 	return rc;
328 }
329 
330 static void idxd_cleanup_internals(struct idxd_device *idxd)
331 {
332 	int i;
333 
334 	for (i = 0; i < idxd->max_groups; i++)
335 		put_device(group_confdev(idxd->groups[i]));
336 	for (i = 0; i < idxd->max_engines; i++)
337 		put_device(engine_confdev(idxd->engines[i]));
338 	for (i = 0; i < idxd->max_wqs; i++)
339 		put_device(wq_confdev(idxd->wqs[i]));
340 	destroy_workqueue(idxd->wq);
341 }
342 
343 static int idxd_init_evl(struct idxd_device *idxd)
344 {
345 	struct device *dev = &idxd->pdev->dev;
346 	unsigned int evl_cache_size;
347 	struct idxd_evl *evl;
348 	const char *idxd_name;
349 
350 	if (idxd->hw.gen_cap.evl_support == 0)
351 		return 0;
352 
353 	evl = kzalloc_node(sizeof(*evl), GFP_KERNEL, dev_to_node(dev));
354 	if (!evl)
355 		return -ENOMEM;
356 
357 	spin_lock_init(&evl->lock);
358 	evl->size = IDXD_EVL_SIZE_MIN;
359 
360 	idxd_name = dev_name(idxd_confdev(idxd));
361 	evl_cache_size = sizeof(struct idxd_evl_fault) + evl_ent_size(idxd);
362 	/*
363 	 * Since completion record in evl_cache will be copied to user
364 	 * when handling completion record page fault, need to create
365 	 * the cache suitable for user copy.
366 	 */
367 	idxd->evl_cache = kmem_cache_create_usercopy(idxd_name, evl_cache_size,
368 						     0, 0, 0, evl_cache_size,
369 						     NULL);
370 	if (!idxd->evl_cache) {
371 		kfree(evl);
372 		return -ENOMEM;
373 	}
374 
375 	idxd->evl = evl;
376 	return 0;
377 }
378 
379 static int idxd_setup_internals(struct idxd_device *idxd)
380 {
381 	struct device *dev = &idxd->pdev->dev;
382 	int rc, i;
383 
384 	init_waitqueue_head(&idxd->cmd_waitq);
385 
386 	rc = idxd_setup_wqs(idxd);
387 	if (rc < 0)
388 		goto err_wqs;
389 
390 	rc = idxd_setup_engines(idxd);
391 	if (rc < 0)
392 		goto err_engine;
393 
394 	rc = idxd_setup_groups(idxd);
395 	if (rc < 0)
396 		goto err_group;
397 
398 	idxd->wq = create_workqueue(dev_name(dev));
399 	if (!idxd->wq) {
400 		rc = -ENOMEM;
401 		goto err_wkq_create;
402 	}
403 
404 	rc = idxd_init_evl(idxd);
405 	if (rc < 0)
406 		goto err_evl;
407 
408 	return 0;
409 
410  err_evl:
411 	destroy_workqueue(idxd->wq);
412  err_wkq_create:
413 	for (i = 0; i < idxd->max_groups; i++)
414 		put_device(group_confdev(idxd->groups[i]));
415  err_group:
416 	for (i = 0; i < idxd->max_engines; i++)
417 		put_device(engine_confdev(idxd->engines[i]));
418  err_engine:
419 	for (i = 0; i < idxd->max_wqs; i++)
420 		put_device(wq_confdev(idxd->wqs[i]));
421  err_wqs:
422 	return rc;
423 }
424 
425 static void idxd_read_table_offsets(struct idxd_device *idxd)
426 {
427 	union offsets_reg offsets;
428 	struct device *dev = &idxd->pdev->dev;
429 
430 	offsets.bits[0] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET);
431 	offsets.bits[1] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET + sizeof(u64));
432 	idxd->grpcfg_offset = offsets.grpcfg * IDXD_TABLE_MULT;
433 	dev_dbg(dev, "IDXD Group Config Offset: %#x\n", idxd->grpcfg_offset);
434 	idxd->wqcfg_offset = offsets.wqcfg * IDXD_TABLE_MULT;
435 	dev_dbg(dev, "IDXD Work Queue Config Offset: %#x\n", idxd->wqcfg_offset);
436 	idxd->msix_perm_offset = offsets.msix_perm * IDXD_TABLE_MULT;
437 	dev_dbg(dev, "IDXD MSIX Permission Offset: %#x\n", idxd->msix_perm_offset);
438 	idxd->perfmon_offset = offsets.perfmon * IDXD_TABLE_MULT;
439 	dev_dbg(dev, "IDXD Perfmon Offset: %#x\n", idxd->perfmon_offset);
440 }
441 
442 void multi_u64_to_bmap(unsigned long *bmap, u64 *val, int count)
443 {
444 	int i, j, nr;
445 
446 	for (i = 0, nr = 0; i < count; i++) {
447 		for (j = 0; j < BITS_PER_LONG_LONG; j++) {
448 			if (val[i] & BIT(j))
449 				set_bit(nr, bmap);
450 			nr++;
451 		}
452 	}
453 }
454 
455 static void idxd_read_caps(struct idxd_device *idxd)
456 {
457 	struct device *dev = &idxd->pdev->dev;
458 	int i;
459 
460 	/* reading generic capabilities */
461 	idxd->hw.gen_cap.bits = ioread64(idxd->reg_base + IDXD_GENCAP_OFFSET);
462 	dev_dbg(dev, "gen_cap: %#llx\n", idxd->hw.gen_cap.bits);
463 
464 	if (idxd->hw.gen_cap.cmd_cap) {
465 		idxd->hw.cmd_cap = ioread32(idxd->reg_base + IDXD_CMDCAP_OFFSET);
466 		dev_dbg(dev, "cmd_cap: %#x\n", idxd->hw.cmd_cap);
467 	}
468 
469 	/* reading command capabilities */
470 	if (idxd->hw.cmd_cap & BIT(IDXD_CMD_REQUEST_INT_HANDLE))
471 		idxd->request_int_handles = true;
472 
473 	idxd->max_xfer_bytes = 1ULL << idxd->hw.gen_cap.max_xfer_shift;
474 	dev_dbg(dev, "max xfer size: %llu bytes\n", idxd->max_xfer_bytes);
475 	idxd_set_max_batch_size(idxd->data->type, idxd, 1U << idxd->hw.gen_cap.max_batch_shift);
476 	dev_dbg(dev, "max batch size: %u\n", idxd->max_batch_size);
477 	if (idxd->hw.gen_cap.config_en)
478 		set_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags);
479 
480 	/* reading group capabilities */
481 	idxd->hw.group_cap.bits =
482 		ioread64(idxd->reg_base + IDXD_GRPCAP_OFFSET);
483 	dev_dbg(dev, "group_cap: %#llx\n", idxd->hw.group_cap.bits);
484 	idxd->max_groups = idxd->hw.group_cap.num_groups;
485 	dev_dbg(dev, "max groups: %u\n", idxd->max_groups);
486 	idxd->max_rdbufs = idxd->hw.group_cap.total_rdbufs;
487 	dev_dbg(dev, "max read buffers: %u\n", idxd->max_rdbufs);
488 	idxd->nr_rdbufs = idxd->max_rdbufs;
489 
490 	/* read engine capabilities */
491 	idxd->hw.engine_cap.bits =
492 		ioread64(idxd->reg_base + IDXD_ENGCAP_OFFSET);
493 	dev_dbg(dev, "engine_cap: %#llx\n", idxd->hw.engine_cap.bits);
494 	idxd->max_engines = idxd->hw.engine_cap.num_engines;
495 	dev_dbg(dev, "max engines: %u\n", idxd->max_engines);
496 
497 	/* read workqueue capabilities */
498 	idxd->hw.wq_cap.bits = ioread64(idxd->reg_base + IDXD_WQCAP_OFFSET);
499 	dev_dbg(dev, "wq_cap: %#llx\n", idxd->hw.wq_cap.bits);
500 	idxd->max_wq_size = idxd->hw.wq_cap.total_wq_size;
501 	dev_dbg(dev, "total workqueue size: %u\n", idxd->max_wq_size);
502 	idxd->max_wqs = idxd->hw.wq_cap.num_wqs;
503 	dev_dbg(dev, "max workqueues: %u\n", idxd->max_wqs);
504 	idxd->wqcfg_size = 1 << (idxd->hw.wq_cap.wqcfg_size + IDXD_WQCFG_MIN);
505 	dev_dbg(dev, "wqcfg size: %u\n", idxd->wqcfg_size);
506 
507 	/* reading operation capabilities */
508 	for (i = 0; i < 4; i++) {
509 		idxd->hw.opcap.bits[i] = ioread64(idxd->reg_base +
510 				IDXD_OPCAP_OFFSET + i * sizeof(u64));
511 		dev_dbg(dev, "opcap[%d]: %#llx\n", i, idxd->hw.opcap.bits[i]);
512 	}
513 	multi_u64_to_bmap(idxd->opcap_bmap, &idxd->hw.opcap.bits[0], 4);
514 
515 	/* read iaa cap */
516 	if (idxd->data->type == IDXD_TYPE_IAX && idxd->hw.version >= DEVICE_VERSION_2)
517 		idxd->hw.iaa_cap.bits = ioread64(idxd->reg_base + IDXD_IAACAP_OFFSET);
518 }
519 
520 static struct idxd_device *idxd_alloc(struct pci_dev *pdev, struct idxd_driver_data *data)
521 {
522 	struct device *dev = &pdev->dev;
523 	struct device *conf_dev;
524 	struct idxd_device *idxd;
525 	int rc;
526 
527 	idxd = kzalloc_node(sizeof(*idxd), GFP_KERNEL, dev_to_node(dev));
528 	if (!idxd)
529 		return NULL;
530 
531 	conf_dev = idxd_confdev(idxd);
532 	idxd->pdev = pdev;
533 	idxd->data = data;
534 	idxd_dev_set_type(&idxd->idxd_dev, idxd->data->type);
535 	idxd->id = ida_alloc(&idxd_ida, GFP_KERNEL);
536 	if (idxd->id < 0)
537 		return NULL;
538 
539 	idxd->opcap_bmap = bitmap_zalloc_node(IDXD_MAX_OPCAP_BITS, GFP_KERNEL, dev_to_node(dev));
540 	if (!idxd->opcap_bmap) {
541 		ida_free(&idxd_ida, idxd->id);
542 		return NULL;
543 	}
544 
545 	device_initialize(conf_dev);
546 	conf_dev->parent = dev;
547 	conf_dev->bus = &dsa_bus_type;
548 	conf_dev->type = idxd->data->dev_type;
549 	rc = dev_set_name(conf_dev, "%s%d", idxd->data->name_prefix, idxd->id);
550 	if (rc < 0) {
551 		put_device(conf_dev);
552 		return NULL;
553 	}
554 
555 	spin_lock_init(&idxd->dev_lock);
556 	spin_lock_init(&idxd->cmd_lock);
557 
558 	return idxd;
559 }
560 
561 static int idxd_enable_system_pasid(struct idxd_device *idxd)
562 {
563 	struct pci_dev *pdev = idxd->pdev;
564 	struct device *dev = &pdev->dev;
565 	struct iommu_domain *domain;
566 	ioasid_t pasid;
567 	int ret;
568 
569 	/*
570 	 * Attach a global PASID to the DMA domain so that we can use ENQCMDS
571 	 * to submit work on buffers mapped by DMA API.
572 	 */
573 	domain = iommu_get_domain_for_dev(dev);
574 	if (!domain)
575 		return -EPERM;
576 
577 	pasid = iommu_alloc_global_pasid(dev);
578 	if (pasid == IOMMU_PASID_INVALID)
579 		return -ENOSPC;
580 
581 	/*
582 	 * DMA domain is owned by the driver, it should support all valid
583 	 * types such as DMA-FQ, identity, etc.
584 	 */
585 	ret = iommu_attach_device_pasid(domain, dev, pasid);
586 	if (ret) {
587 		dev_err(dev, "failed to attach device pasid %d, domain type %d",
588 			pasid, domain->type);
589 		iommu_free_global_pasid(pasid);
590 		return ret;
591 	}
592 
593 	/* Since we set user privilege for kernel DMA, enable completion IRQ */
594 	idxd_set_user_intr(idxd, 1);
595 	idxd->pasid = pasid;
596 
597 	return ret;
598 }
599 
600 static void idxd_disable_system_pasid(struct idxd_device *idxd)
601 {
602 	struct pci_dev *pdev = idxd->pdev;
603 	struct device *dev = &pdev->dev;
604 	struct iommu_domain *domain;
605 
606 	domain = iommu_get_domain_for_dev(dev);
607 	if (!domain)
608 		return;
609 
610 	iommu_detach_device_pasid(domain, dev, idxd->pasid);
611 	iommu_free_global_pasid(idxd->pasid);
612 
613 	idxd_set_user_intr(idxd, 0);
614 	idxd->sva = NULL;
615 	idxd->pasid = IOMMU_PASID_INVALID;
616 }
617 
618 static int idxd_enable_sva(struct pci_dev *pdev)
619 {
620 	int ret;
621 
622 	ret = iommu_dev_enable_feature(&pdev->dev, IOMMU_DEV_FEAT_IOPF);
623 	if (ret)
624 		return ret;
625 
626 	ret = iommu_dev_enable_feature(&pdev->dev, IOMMU_DEV_FEAT_SVA);
627 	if (ret)
628 		iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_IOPF);
629 
630 	return ret;
631 }
632 
633 static void idxd_disable_sva(struct pci_dev *pdev)
634 {
635 	iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_SVA);
636 	iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_IOPF);
637 }
638 
639 static int idxd_probe(struct idxd_device *idxd)
640 {
641 	struct pci_dev *pdev = idxd->pdev;
642 	struct device *dev = &pdev->dev;
643 	int rc;
644 
645 	dev_dbg(dev, "%s entered and resetting device\n", __func__);
646 	rc = idxd_device_init_reset(idxd);
647 	if (rc < 0)
648 		return rc;
649 
650 	dev_dbg(dev, "IDXD reset complete\n");
651 
652 	if (IS_ENABLED(CONFIG_INTEL_IDXD_SVM) && sva) {
653 		if (idxd_enable_sva(pdev)) {
654 			dev_warn(dev, "Unable to turn on user SVA feature.\n");
655 		} else {
656 			set_bit(IDXD_FLAG_USER_PASID_ENABLED, &idxd->flags);
657 
658 			rc = idxd_enable_system_pasid(idxd);
659 			if (rc)
660 				dev_warn(dev, "No in-kernel DMA with PASID. %d\n", rc);
661 			else
662 				set_bit(IDXD_FLAG_PASID_ENABLED, &idxd->flags);
663 		}
664 	} else if (!sva) {
665 		dev_warn(dev, "User forced SVA off via module param.\n");
666 	}
667 
668 	idxd_read_caps(idxd);
669 	idxd_read_table_offsets(idxd);
670 
671 	rc = idxd_setup_internals(idxd);
672 	if (rc)
673 		goto err;
674 
675 	/* If the configs are readonly, then load them from device */
676 	if (!test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags)) {
677 		dev_dbg(dev, "Loading RO device config\n");
678 		rc = idxd_device_load_config(idxd);
679 		if (rc < 0)
680 			goto err_config;
681 	}
682 
683 	rc = idxd_setup_interrupts(idxd);
684 	if (rc)
685 		goto err_config;
686 
687 	idxd->major = idxd_cdev_get_major(idxd);
688 
689 	rc = perfmon_pmu_init(idxd);
690 	if (rc < 0)
691 		dev_warn(dev, "Failed to initialize perfmon. No PMU support: %d\n", rc);
692 
693 	dev_dbg(dev, "IDXD device %d probed successfully\n", idxd->id);
694 	return 0;
695 
696  err_config:
697 	idxd_cleanup_internals(idxd);
698  err:
699 	if (device_pasid_enabled(idxd))
700 		idxd_disable_system_pasid(idxd);
701 	if (device_user_pasid_enabled(idxd))
702 		idxd_disable_sva(pdev);
703 	return rc;
704 }
705 
706 static void idxd_cleanup(struct idxd_device *idxd)
707 {
708 	perfmon_pmu_remove(idxd);
709 	idxd_cleanup_interrupts(idxd);
710 	idxd_cleanup_internals(idxd);
711 	if (device_pasid_enabled(idxd))
712 		idxd_disable_system_pasid(idxd);
713 	if (device_user_pasid_enabled(idxd))
714 		idxd_disable_sva(idxd->pdev);
715 }
716 
717 static int idxd_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
718 {
719 	struct device *dev = &pdev->dev;
720 	struct idxd_device *idxd;
721 	struct idxd_driver_data *data = (struct idxd_driver_data *)id->driver_data;
722 	int rc;
723 
724 	rc = pci_enable_device(pdev);
725 	if (rc)
726 		return rc;
727 
728 	dev_dbg(dev, "Alloc IDXD context\n");
729 	idxd = idxd_alloc(pdev, data);
730 	if (!idxd) {
731 		rc = -ENOMEM;
732 		goto err_idxd_alloc;
733 	}
734 
735 	dev_dbg(dev, "Mapping BARs\n");
736 	idxd->reg_base = pci_iomap(pdev, IDXD_MMIO_BAR, 0);
737 	if (!idxd->reg_base) {
738 		rc = -ENOMEM;
739 		goto err_iomap;
740 	}
741 
742 	dev_dbg(dev, "Set DMA masks\n");
743 	rc = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
744 	if (rc)
745 		goto err;
746 
747 	dev_dbg(dev, "Set PCI master\n");
748 	pci_set_master(pdev);
749 	pci_set_drvdata(pdev, idxd);
750 
751 	idxd->hw.version = ioread32(idxd->reg_base + IDXD_VER_OFFSET);
752 	rc = idxd_probe(idxd);
753 	if (rc) {
754 		dev_err(dev, "Intel(R) IDXD DMA Engine init failed\n");
755 		goto err;
756 	}
757 
758 	if (data->load_device_defaults) {
759 		rc = data->load_device_defaults(idxd);
760 		if (rc)
761 			dev_warn(dev, "IDXD loading device defaults failed\n");
762 	}
763 
764 	rc = idxd_register_devices(idxd);
765 	if (rc) {
766 		dev_err(dev, "IDXD sysfs setup failed\n");
767 		goto err_dev_register;
768 	}
769 
770 	rc = idxd_device_init_debugfs(idxd);
771 	if (rc)
772 		dev_warn(dev, "IDXD debugfs failed to setup\n");
773 
774 	dev_info(&pdev->dev, "Intel(R) Accelerator Device (v%x)\n",
775 		 idxd->hw.version);
776 
777 	return 0;
778 
779  err_dev_register:
780 	idxd_cleanup(idxd);
781  err:
782 	pci_iounmap(pdev, idxd->reg_base);
783  err_iomap:
784 	put_device(idxd_confdev(idxd));
785  err_idxd_alloc:
786 	pci_disable_device(pdev);
787 	return rc;
788 }
789 
790 void idxd_wqs_quiesce(struct idxd_device *idxd)
791 {
792 	struct idxd_wq *wq;
793 	int i;
794 
795 	for (i = 0; i < idxd->max_wqs; i++) {
796 		wq = idxd->wqs[i];
797 		if (wq->state == IDXD_WQ_ENABLED && wq->type == IDXD_WQT_KERNEL)
798 			idxd_wq_quiesce(wq);
799 	}
800 }
801 
802 static void idxd_shutdown(struct pci_dev *pdev)
803 {
804 	struct idxd_device *idxd = pci_get_drvdata(pdev);
805 	struct idxd_irq_entry *irq_entry;
806 	int rc;
807 
808 	rc = idxd_device_disable(idxd);
809 	if (rc)
810 		dev_err(&pdev->dev, "Disabling device failed\n");
811 
812 	irq_entry = &idxd->ie;
813 	synchronize_irq(irq_entry->vector);
814 	idxd_mask_error_interrupts(idxd);
815 	flush_workqueue(idxd->wq);
816 }
817 
818 static void idxd_remove(struct pci_dev *pdev)
819 {
820 	struct idxd_device *idxd = pci_get_drvdata(pdev);
821 	struct idxd_irq_entry *irq_entry;
822 
823 	idxd_unregister_devices(idxd);
824 	/*
825 	 * When ->release() is called for the idxd->conf_dev, it frees all the memory related
826 	 * to the idxd context. The driver still needs those bits in order to do the rest of
827 	 * the cleanup. However, we do need to unbound the idxd sub-driver. So take a ref
828 	 * on the device here to hold off the freeing while allowing the idxd sub-driver
829 	 * to unbind.
830 	 */
831 	get_device(idxd_confdev(idxd));
832 	device_unregister(idxd_confdev(idxd));
833 	idxd_shutdown(pdev);
834 	if (device_pasid_enabled(idxd))
835 		idxd_disable_system_pasid(idxd);
836 	idxd_device_remove_debugfs(idxd);
837 
838 	irq_entry = idxd_get_ie(idxd, 0);
839 	free_irq(irq_entry->vector, irq_entry);
840 	pci_free_irq_vectors(pdev);
841 	pci_iounmap(pdev, idxd->reg_base);
842 	if (device_user_pasid_enabled(idxd))
843 		idxd_disable_sva(pdev);
844 	pci_disable_device(pdev);
845 	destroy_workqueue(idxd->wq);
846 	perfmon_pmu_remove(idxd);
847 	put_device(idxd_confdev(idxd));
848 }
849 
850 static struct pci_driver idxd_pci_driver = {
851 	.name		= DRV_NAME,
852 	.id_table	= idxd_pci_tbl,
853 	.probe		= idxd_pci_probe,
854 	.remove		= idxd_remove,
855 	.shutdown	= idxd_shutdown,
856 };
857 
858 static int __init idxd_init_module(void)
859 {
860 	int err;
861 
862 	/*
863 	 * If the CPU does not support MOVDIR64B or ENQCMDS, there's no point in
864 	 * enumerating the device. We can not utilize it.
865 	 */
866 	if (!cpu_feature_enabled(X86_FEATURE_MOVDIR64B)) {
867 		pr_warn("idxd driver failed to load without MOVDIR64B.\n");
868 		return -ENODEV;
869 	}
870 
871 	if (!cpu_feature_enabled(X86_FEATURE_ENQCMD))
872 		pr_warn("Platform does not have ENQCMD(S) support.\n");
873 	else
874 		support_enqcmd = true;
875 
876 	perfmon_init();
877 
878 	err = idxd_driver_register(&idxd_drv);
879 	if (err < 0)
880 		goto err_idxd_driver_register;
881 
882 	err = idxd_driver_register(&idxd_dmaengine_drv);
883 	if (err < 0)
884 		goto err_idxd_dmaengine_driver_register;
885 
886 	err = idxd_driver_register(&idxd_user_drv);
887 	if (err < 0)
888 		goto err_idxd_user_driver_register;
889 
890 	err = idxd_cdev_register();
891 	if (err)
892 		goto err_cdev_register;
893 
894 	err = idxd_init_debugfs();
895 	if (err)
896 		goto err_debugfs;
897 
898 	err = pci_register_driver(&idxd_pci_driver);
899 	if (err)
900 		goto err_pci_register;
901 
902 	return 0;
903 
904 err_pci_register:
905 	idxd_remove_debugfs();
906 err_debugfs:
907 	idxd_cdev_remove();
908 err_cdev_register:
909 	idxd_driver_unregister(&idxd_user_drv);
910 err_idxd_user_driver_register:
911 	idxd_driver_unregister(&idxd_dmaengine_drv);
912 err_idxd_dmaengine_driver_register:
913 	idxd_driver_unregister(&idxd_drv);
914 err_idxd_driver_register:
915 	return err;
916 }
917 module_init(idxd_init_module);
918 
919 static void __exit idxd_exit_module(void)
920 {
921 	idxd_driver_unregister(&idxd_user_drv);
922 	idxd_driver_unregister(&idxd_dmaengine_drv);
923 	idxd_driver_unregister(&idxd_drv);
924 	pci_unregister_driver(&idxd_pci_driver);
925 	idxd_cdev_remove();
926 	perfmon_exit();
927 	idxd_remove_debugfs();
928 }
929 module_exit(idxd_exit_module);
930