xref: /linux/drivers/dma/idxd/init.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
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_DESCRIPTION("Intel Data Streaming Accelerator and In-Memory Analytics Accelerator common driver");
26 MODULE_LICENSE("GPL v2");
27 MODULE_AUTHOR("Intel Corporation");
28 MODULE_IMPORT_NS(IDXD);
29 
30 static bool sva = true;
31 module_param(sva, bool, 0644);
32 MODULE_PARM_DESC(sva, "Toggle SVA support on/off");
33 
34 bool tc_override;
35 module_param(tc_override, bool, 0644);
36 MODULE_PARM_DESC(tc_override, "Override traffic class defaults");
37 
38 #define DRV_NAME "idxd"
39 
40 bool support_enqcmd;
41 DEFINE_IDA(idxd_ida);
42 
43 static struct idxd_driver_data idxd_driver_data[] = {
44 	[IDXD_TYPE_DSA] = {
45 		.name_prefix = "dsa",
46 		.type = IDXD_TYPE_DSA,
47 		.compl_size = sizeof(struct dsa_completion_record),
48 		.align = 32,
49 		.dev_type = &dsa_device_type,
50 		.evl_cr_off = offsetof(struct dsa_evl_entry, cr),
51 		.user_submission_safe = false, /* See INTEL-SA-01084 security advisory */
52 		.cr_status_off = offsetof(struct dsa_completion_record, status),
53 		.cr_result_off = offsetof(struct dsa_completion_record, result),
54 	},
55 	[IDXD_TYPE_IAX] = {
56 		.name_prefix = "iax",
57 		.type = IDXD_TYPE_IAX,
58 		.compl_size = sizeof(struct iax_completion_record),
59 		.align = 64,
60 		.dev_type = &iax_device_type,
61 		.evl_cr_off = offsetof(struct iax_evl_entry, cr),
62 		.user_submission_safe = false, /* See INTEL-SA-01084 security advisory */
63 		.cr_status_off = offsetof(struct iax_completion_record, status),
64 		.cr_result_off = offsetof(struct iax_completion_record, error_code),
65 		.load_device_defaults = idxd_load_iaa_device_defaults,
66 	},
67 };
68 
69 static struct pci_device_id idxd_pci_tbl[] = {
70 	/* DSA ver 1.0 platforms */
71 	{ PCI_DEVICE_DATA(INTEL, DSA_SPR0, &idxd_driver_data[IDXD_TYPE_DSA]) },
72 	/* DSA on GNR-D platforms */
73 	{ PCI_DEVICE_DATA(INTEL, DSA_GNRD, &idxd_driver_data[IDXD_TYPE_DSA]) },
74 	/* DSA on DMR platforms */
75 	{ PCI_DEVICE_DATA(INTEL, DSA_DMR, &idxd_driver_data[IDXD_TYPE_DSA]) },
76 
77 	/* IAX ver 1.0 platforms */
78 	{ PCI_DEVICE_DATA(INTEL, IAX_SPR0, &idxd_driver_data[IDXD_TYPE_IAX]) },
79 	/* IAA on DMR platforms */
80 	{ PCI_DEVICE_DATA(INTEL, IAA_DMR, &idxd_driver_data[IDXD_TYPE_IAX]) },
81 	{ 0, }
82 };
83 MODULE_DEVICE_TABLE(pci, idxd_pci_tbl);
84 
85 static int idxd_setup_interrupts(struct idxd_device *idxd)
86 {
87 	struct pci_dev *pdev = idxd->pdev;
88 	struct device *dev = &pdev->dev;
89 	struct idxd_irq_entry *ie;
90 	int i, msixcnt;
91 	int rc = 0;
92 
93 	msixcnt = pci_msix_vec_count(pdev);
94 	if (msixcnt < 0) {
95 		dev_err(dev, "Not MSI-X interrupt capable.\n");
96 		return -ENOSPC;
97 	}
98 	idxd->irq_cnt = msixcnt;
99 
100 	rc = pci_alloc_irq_vectors(pdev, msixcnt, msixcnt, PCI_IRQ_MSIX);
101 	if (rc != msixcnt) {
102 		dev_err(dev, "Failed enabling %d MSIX entries: %d\n", msixcnt, rc);
103 		return -ENOSPC;
104 	}
105 	dev_dbg(dev, "Enabled %d msix vectors\n", msixcnt);
106 
107 
108 	ie = idxd_get_ie(idxd, 0);
109 	ie->vector = pci_irq_vector(pdev, 0);
110 	rc = request_threaded_irq(ie->vector, NULL, idxd_misc_thread, 0, "idxd-misc", ie);
111 	if (rc < 0) {
112 		dev_err(dev, "Failed to allocate misc interrupt.\n");
113 		goto err_misc_irq;
114 	}
115 	dev_dbg(dev, "Requested idxd-misc handler on msix vector %d\n", ie->vector);
116 
117 	for (i = 0; i < idxd->max_wqs; i++) {
118 		int msix_idx = i + 1;
119 
120 		ie = idxd_get_ie(idxd, msix_idx);
121 		ie->id = msix_idx;
122 		ie->int_handle = INVALID_INT_HANDLE;
123 		ie->pasid = IOMMU_PASID_INVALID;
124 
125 		spin_lock_init(&ie->list_lock);
126 		init_llist_head(&ie->pending_llist);
127 		INIT_LIST_HEAD(&ie->work_list);
128 	}
129 
130 	idxd_unmask_error_interrupts(idxd);
131 	return 0;
132 
133  err_misc_irq:
134 	idxd_mask_error_interrupts(idxd);
135 	pci_free_irq_vectors(pdev);
136 	dev_err(dev, "No usable interrupts\n");
137 	return rc;
138 }
139 
140 static void idxd_cleanup_interrupts(struct idxd_device *idxd)
141 {
142 	struct pci_dev *pdev = idxd->pdev;
143 	struct idxd_irq_entry *ie;
144 	int msixcnt;
145 
146 	msixcnt = pci_msix_vec_count(pdev);
147 	if (msixcnt <= 0)
148 		return;
149 
150 	ie = idxd_get_ie(idxd, 0);
151 	idxd_mask_error_interrupts(idxd);
152 	free_irq(ie->vector, ie);
153 	pci_free_irq_vectors(pdev);
154 }
155 
156 static int idxd_setup_wqs(struct idxd_device *idxd)
157 {
158 	struct device *dev = &idxd->pdev->dev;
159 	struct idxd_wq *wq;
160 	struct device *conf_dev;
161 	int i, rc;
162 
163 	idxd->wqs = kcalloc_node(idxd->max_wqs, sizeof(struct idxd_wq *),
164 				 GFP_KERNEL, dev_to_node(dev));
165 	if (!idxd->wqs)
166 		return -ENOMEM;
167 
168 	idxd->wq_enable_map = bitmap_zalloc_node(idxd->max_wqs, GFP_KERNEL, dev_to_node(dev));
169 	if (!idxd->wq_enable_map) {
170 		kfree(idxd->wqs);
171 		return -ENOMEM;
172 	}
173 
174 	for (i = 0; i < idxd->max_wqs; i++) {
175 		wq = kzalloc_node(sizeof(*wq), GFP_KERNEL, dev_to_node(dev));
176 		if (!wq) {
177 			rc = -ENOMEM;
178 			goto err;
179 		}
180 
181 		idxd_dev_set_type(&wq->idxd_dev, IDXD_DEV_WQ);
182 		conf_dev = wq_confdev(wq);
183 		wq->id = i;
184 		wq->idxd = idxd;
185 		device_initialize(wq_confdev(wq));
186 		conf_dev->parent = idxd_confdev(idxd);
187 		conf_dev->bus = &dsa_bus_type;
188 		conf_dev->type = &idxd_wq_device_type;
189 		rc = dev_set_name(conf_dev, "wq%d.%d", idxd->id, wq->id);
190 		if (rc < 0) {
191 			put_device(conf_dev);
192 			goto err;
193 		}
194 
195 		mutex_init(&wq->wq_lock);
196 		init_waitqueue_head(&wq->err_queue);
197 		init_completion(&wq->wq_dead);
198 		init_completion(&wq->wq_resurrect);
199 		wq->max_xfer_bytes = WQ_DEFAULT_MAX_XFER;
200 		idxd_wq_set_max_batch_size(idxd->data->type, wq, WQ_DEFAULT_MAX_BATCH);
201 		wq->enqcmds_retries = IDXD_ENQCMDS_RETRIES;
202 		wq->wqcfg = kzalloc_node(idxd->wqcfg_size, GFP_KERNEL, dev_to_node(dev));
203 		if (!wq->wqcfg) {
204 			put_device(conf_dev);
205 			rc = -ENOMEM;
206 			goto err;
207 		}
208 
209 		if (idxd->hw.wq_cap.op_config) {
210 			wq->opcap_bmap = bitmap_zalloc(IDXD_MAX_OPCAP_BITS, GFP_KERNEL);
211 			if (!wq->opcap_bmap) {
212 				put_device(conf_dev);
213 				rc = -ENOMEM;
214 				goto err;
215 			}
216 			bitmap_copy(wq->opcap_bmap, idxd->opcap_bmap, IDXD_MAX_OPCAP_BITS);
217 		}
218 		mutex_init(&wq->uc_lock);
219 		xa_init(&wq->upasid_xa);
220 		idxd->wqs[i] = wq;
221 	}
222 
223 	return 0;
224 
225  err:
226 	while (--i >= 0) {
227 		wq = idxd->wqs[i];
228 		conf_dev = wq_confdev(wq);
229 		put_device(conf_dev);
230 	}
231 	return rc;
232 }
233 
234 static int idxd_setup_engines(struct idxd_device *idxd)
235 {
236 	struct idxd_engine *engine;
237 	struct device *dev = &idxd->pdev->dev;
238 	struct device *conf_dev;
239 	int i, rc;
240 
241 	idxd->engines = kcalloc_node(idxd->max_engines, sizeof(struct idxd_engine *),
242 				     GFP_KERNEL, dev_to_node(dev));
243 	if (!idxd->engines)
244 		return -ENOMEM;
245 
246 	for (i = 0; i < idxd->max_engines; i++) {
247 		engine = kzalloc_node(sizeof(*engine), GFP_KERNEL, dev_to_node(dev));
248 		if (!engine) {
249 			rc = -ENOMEM;
250 			goto err;
251 		}
252 
253 		idxd_dev_set_type(&engine->idxd_dev, IDXD_DEV_ENGINE);
254 		conf_dev = engine_confdev(engine);
255 		engine->id = i;
256 		engine->idxd = idxd;
257 		device_initialize(conf_dev);
258 		conf_dev->parent = idxd_confdev(idxd);
259 		conf_dev->bus = &dsa_bus_type;
260 		conf_dev->type = &idxd_engine_device_type;
261 		rc = dev_set_name(conf_dev, "engine%d.%d", idxd->id, engine->id);
262 		if (rc < 0) {
263 			put_device(conf_dev);
264 			goto err;
265 		}
266 
267 		idxd->engines[i] = engine;
268 	}
269 
270 	return 0;
271 
272  err:
273 	while (--i >= 0) {
274 		engine = idxd->engines[i];
275 		conf_dev = engine_confdev(engine);
276 		put_device(conf_dev);
277 	}
278 	return rc;
279 }
280 
281 static int idxd_setup_groups(struct idxd_device *idxd)
282 {
283 	struct device *dev = &idxd->pdev->dev;
284 	struct device *conf_dev;
285 	struct idxd_group *group;
286 	int i, rc;
287 
288 	idxd->groups = kcalloc_node(idxd->max_groups, sizeof(struct idxd_group *),
289 				    GFP_KERNEL, dev_to_node(dev));
290 	if (!idxd->groups)
291 		return -ENOMEM;
292 
293 	for (i = 0; i < idxd->max_groups; i++) {
294 		group = kzalloc_node(sizeof(*group), GFP_KERNEL, dev_to_node(dev));
295 		if (!group) {
296 			rc = -ENOMEM;
297 			goto err;
298 		}
299 
300 		idxd_dev_set_type(&group->idxd_dev, IDXD_DEV_GROUP);
301 		conf_dev = group_confdev(group);
302 		group->id = i;
303 		group->idxd = idxd;
304 		device_initialize(conf_dev);
305 		conf_dev->parent = idxd_confdev(idxd);
306 		conf_dev->bus = &dsa_bus_type;
307 		conf_dev->type = &idxd_group_device_type;
308 		rc = dev_set_name(conf_dev, "group%d.%d", idxd->id, group->id);
309 		if (rc < 0) {
310 			put_device(conf_dev);
311 			goto err;
312 		}
313 
314 		idxd->groups[i] = group;
315 		if (idxd->hw.version <= DEVICE_VERSION_2 && !tc_override) {
316 			group->tc_a = 1;
317 			group->tc_b = 1;
318 		} else {
319 			group->tc_a = -1;
320 			group->tc_b = -1;
321 		}
322 		/*
323 		 * The default value is the same as the value of
324 		 * total read buffers in GRPCAP.
325 		 */
326 		group->rdbufs_allowed = idxd->max_rdbufs;
327 	}
328 
329 	return 0;
330 
331  err:
332 	while (--i >= 0) {
333 		group = idxd->groups[i];
334 		put_device(group_confdev(group));
335 	}
336 	return rc;
337 }
338 
339 static void idxd_cleanup_internals(struct idxd_device *idxd)
340 {
341 	int i;
342 
343 	for (i = 0; i < idxd->max_groups; i++)
344 		put_device(group_confdev(idxd->groups[i]));
345 	for (i = 0; i < idxd->max_engines; i++)
346 		put_device(engine_confdev(idxd->engines[i]));
347 	for (i = 0; i < idxd->max_wqs; i++)
348 		put_device(wq_confdev(idxd->wqs[i]));
349 	destroy_workqueue(idxd->wq);
350 }
351 
352 static int idxd_init_evl(struct idxd_device *idxd)
353 {
354 	struct device *dev = &idxd->pdev->dev;
355 	unsigned int evl_cache_size;
356 	struct idxd_evl *evl;
357 	const char *idxd_name;
358 
359 	if (idxd->hw.gen_cap.evl_support == 0)
360 		return 0;
361 
362 	evl = kzalloc_node(sizeof(*evl), GFP_KERNEL, dev_to_node(dev));
363 	if (!evl)
364 		return -ENOMEM;
365 
366 	mutex_init(&evl->lock);
367 	evl->size = IDXD_EVL_SIZE_MIN;
368 
369 	idxd_name = dev_name(idxd_confdev(idxd));
370 	evl_cache_size = sizeof(struct idxd_evl_fault) + evl_ent_size(idxd);
371 	/*
372 	 * Since completion record in evl_cache will be copied to user
373 	 * when handling completion record page fault, need to create
374 	 * the cache suitable for user copy.
375 	 */
376 	idxd->evl_cache = kmem_cache_create_usercopy(idxd_name, evl_cache_size,
377 						     0, 0, 0, evl_cache_size,
378 						     NULL);
379 	if (!idxd->evl_cache) {
380 		kfree(evl);
381 		return -ENOMEM;
382 	}
383 
384 	idxd->evl = evl;
385 	return 0;
386 }
387 
388 static int idxd_setup_internals(struct idxd_device *idxd)
389 {
390 	struct device *dev = &idxd->pdev->dev;
391 	int rc, i;
392 
393 	init_waitqueue_head(&idxd->cmd_waitq);
394 
395 	rc = idxd_setup_wqs(idxd);
396 	if (rc < 0)
397 		goto err_wqs;
398 
399 	rc = idxd_setup_engines(idxd);
400 	if (rc < 0)
401 		goto err_engine;
402 
403 	rc = idxd_setup_groups(idxd);
404 	if (rc < 0)
405 		goto err_group;
406 
407 	idxd->wq = create_workqueue(dev_name(dev));
408 	if (!idxd->wq) {
409 		rc = -ENOMEM;
410 		goto err_wkq_create;
411 	}
412 
413 	rc = idxd_init_evl(idxd);
414 	if (rc < 0)
415 		goto err_evl;
416 
417 	return 0;
418 
419  err_evl:
420 	destroy_workqueue(idxd->wq);
421  err_wkq_create:
422 	for (i = 0; i < idxd->max_groups; i++)
423 		put_device(group_confdev(idxd->groups[i]));
424  err_group:
425 	for (i = 0; i < idxd->max_engines; i++)
426 		put_device(engine_confdev(idxd->engines[i]));
427  err_engine:
428 	for (i = 0; i < idxd->max_wqs; i++)
429 		put_device(wq_confdev(idxd->wqs[i]));
430  err_wqs:
431 	return rc;
432 }
433 
434 static void idxd_read_table_offsets(struct idxd_device *idxd)
435 {
436 	union offsets_reg offsets;
437 	struct device *dev = &idxd->pdev->dev;
438 
439 	offsets.bits[0] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET);
440 	offsets.bits[1] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET + sizeof(u64));
441 	idxd->grpcfg_offset = offsets.grpcfg * IDXD_TABLE_MULT;
442 	dev_dbg(dev, "IDXD Group Config Offset: %#x\n", idxd->grpcfg_offset);
443 	idxd->wqcfg_offset = offsets.wqcfg * IDXD_TABLE_MULT;
444 	dev_dbg(dev, "IDXD Work Queue Config Offset: %#x\n", idxd->wqcfg_offset);
445 	idxd->msix_perm_offset = offsets.msix_perm * IDXD_TABLE_MULT;
446 	dev_dbg(dev, "IDXD MSIX Permission Offset: %#x\n", idxd->msix_perm_offset);
447 	idxd->perfmon_offset = offsets.perfmon * IDXD_TABLE_MULT;
448 	dev_dbg(dev, "IDXD Perfmon Offset: %#x\n", idxd->perfmon_offset);
449 }
450 
451 void multi_u64_to_bmap(unsigned long *bmap, u64 *val, int count)
452 {
453 	int i, j, nr;
454 
455 	for (i = 0, nr = 0; i < count; i++) {
456 		for (j = 0; j < BITS_PER_LONG_LONG; j++) {
457 			if (val[i] & BIT(j))
458 				set_bit(nr, bmap);
459 			nr++;
460 		}
461 	}
462 }
463 
464 static void idxd_read_caps(struct idxd_device *idxd)
465 {
466 	struct device *dev = &idxd->pdev->dev;
467 	int i;
468 
469 	/* reading generic capabilities */
470 	idxd->hw.gen_cap.bits = ioread64(idxd->reg_base + IDXD_GENCAP_OFFSET);
471 	dev_dbg(dev, "gen_cap: %#llx\n", idxd->hw.gen_cap.bits);
472 
473 	if (idxd->hw.gen_cap.cmd_cap) {
474 		idxd->hw.cmd_cap = ioread32(idxd->reg_base + IDXD_CMDCAP_OFFSET);
475 		dev_dbg(dev, "cmd_cap: %#x\n", idxd->hw.cmd_cap);
476 	}
477 
478 	/* reading command capabilities */
479 	if (idxd->hw.cmd_cap & BIT(IDXD_CMD_REQUEST_INT_HANDLE))
480 		idxd->request_int_handles = true;
481 
482 	idxd->max_xfer_bytes = 1ULL << idxd->hw.gen_cap.max_xfer_shift;
483 	dev_dbg(dev, "max xfer size: %llu bytes\n", idxd->max_xfer_bytes);
484 	idxd_set_max_batch_size(idxd->data->type, idxd, 1U << idxd->hw.gen_cap.max_batch_shift);
485 	dev_dbg(dev, "max batch size: %u\n", idxd->max_batch_size);
486 	if (idxd->hw.gen_cap.config_en)
487 		set_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags);
488 
489 	/* reading group capabilities */
490 	idxd->hw.group_cap.bits =
491 		ioread64(idxd->reg_base + IDXD_GRPCAP_OFFSET);
492 	dev_dbg(dev, "group_cap: %#llx\n", idxd->hw.group_cap.bits);
493 	idxd->max_groups = idxd->hw.group_cap.num_groups;
494 	dev_dbg(dev, "max groups: %u\n", idxd->max_groups);
495 	idxd->max_rdbufs = idxd->hw.group_cap.total_rdbufs;
496 	dev_dbg(dev, "max read buffers: %u\n", idxd->max_rdbufs);
497 	idxd->nr_rdbufs = idxd->max_rdbufs;
498 
499 	/* read engine capabilities */
500 	idxd->hw.engine_cap.bits =
501 		ioread64(idxd->reg_base + IDXD_ENGCAP_OFFSET);
502 	dev_dbg(dev, "engine_cap: %#llx\n", idxd->hw.engine_cap.bits);
503 	idxd->max_engines = idxd->hw.engine_cap.num_engines;
504 	dev_dbg(dev, "max engines: %u\n", idxd->max_engines);
505 
506 	/* read workqueue capabilities */
507 	idxd->hw.wq_cap.bits = ioread64(idxd->reg_base + IDXD_WQCAP_OFFSET);
508 	dev_dbg(dev, "wq_cap: %#llx\n", idxd->hw.wq_cap.bits);
509 	idxd->max_wq_size = idxd->hw.wq_cap.total_wq_size;
510 	dev_dbg(dev, "total workqueue size: %u\n", idxd->max_wq_size);
511 	idxd->max_wqs = idxd->hw.wq_cap.num_wqs;
512 	dev_dbg(dev, "max workqueues: %u\n", idxd->max_wqs);
513 	idxd->wqcfg_size = 1 << (idxd->hw.wq_cap.wqcfg_size + IDXD_WQCFG_MIN);
514 	dev_dbg(dev, "wqcfg size: %u\n", idxd->wqcfg_size);
515 
516 	/* reading operation capabilities */
517 	for (i = 0; i < 4; i++) {
518 		idxd->hw.opcap.bits[i] = ioread64(idxd->reg_base +
519 				IDXD_OPCAP_OFFSET + i * sizeof(u64));
520 		dev_dbg(dev, "opcap[%d]: %#llx\n", i, idxd->hw.opcap.bits[i]);
521 	}
522 	multi_u64_to_bmap(idxd->opcap_bmap, &idxd->hw.opcap.bits[0], 4);
523 
524 	/* read iaa cap */
525 	if (idxd->data->type == IDXD_TYPE_IAX && idxd->hw.version >= DEVICE_VERSION_2)
526 		idxd->hw.iaa_cap.bits = ioread64(idxd->reg_base + IDXD_IAACAP_OFFSET);
527 }
528 
529 static struct idxd_device *idxd_alloc(struct pci_dev *pdev, struct idxd_driver_data *data)
530 {
531 	struct device *dev = &pdev->dev;
532 	struct device *conf_dev;
533 	struct idxd_device *idxd;
534 	int rc;
535 
536 	idxd = kzalloc_node(sizeof(*idxd), GFP_KERNEL, dev_to_node(dev));
537 	if (!idxd)
538 		return NULL;
539 
540 	conf_dev = idxd_confdev(idxd);
541 	idxd->pdev = pdev;
542 	idxd->data = data;
543 	idxd_dev_set_type(&idxd->idxd_dev, idxd->data->type);
544 	idxd->id = ida_alloc(&idxd_ida, GFP_KERNEL);
545 	if (idxd->id < 0)
546 		return NULL;
547 
548 	idxd->opcap_bmap = bitmap_zalloc_node(IDXD_MAX_OPCAP_BITS, GFP_KERNEL, dev_to_node(dev));
549 	if (!idxd->opcap_bmap) {
550 		ida_free(&idxd_ida, idxd->id);
551 		return NULL;
552 	}
553 
554 	device_initialize(conf_dev);
555 	conf_dev->parent = dev;
556 	conf_dev->bus = &dsa_bus_type;
557 	conf_dev->type = idxd->data->dev_type;
558 	rc = dev_set_name(conf_dev, "%s%d", idxd->data->name_prefix, idxd->id);
559 	if (rc < 0) {
560 		put_device(conf_dev);
561 		return NULL;
562 	}
563 
564 	spin_lock_init(&idxd->dev_lock);
565 	spin_lock_init(&idxd->cmd_lock);
566 
567 	return idxd;
568 }
569 
570 static int idxd_enable_system_pasid(struct idxd_device *idxd)
571 {
572 	struct pci_dev *pdev = idxd->pdev;
573 	struct device *dev = &pdev->dev;
574 	struct iommu_domain *domain;
575 	ioasid_t pasid;
576 	int ret;
577 
578 	/*
579 	 * Attach a global PASID to the DMA domain so that we can use ENQCMDS
580 	 * to submit work on buffers mapped by DMA API.
581 	 */
582 	domain = iommu_get_domain_for_dev(dev);
583 	if (!domain)
584 		return -EPERM;
585 
586 	pasid = iommu_alloc_global_pasid(dev);
587 	if (pasid == IOMMU_PASID_INVALID)
588 		return -ENOSPC;
589 
590 	/*
591 	 * DMA domain is owned by the driver, it should support all valid
592 	 * types such as DMA-FQ, identity, etc.
593 	 */
594 	ret = iommu_attach_device_pasid(domain, dev, pasid, NULL);
595 	if (ret) {
596 		dev_err(dev, "failed to attach device pasid %d, domain type %d",
597 			pasid, domain->type);
598 		iommu_free_global_pasid(pasid);
599 		return ret;
600 	}
601 
602 	/* Since we set user privilege for kernel DMA, enable completion IRQ */
603 	idxd_set_user_intr(idxd, 1);
604 	idxd->pasid = pasid;
605 
606 	return ret;
607 }
608 
609 static void idxd_disable_system_pasid(struct idxd_device *idxd)
610 {
611 	struct pci_dev *pdev = idxd->pdev;
612 	struct device *dev = &pdev->dev;
613 	struct iommu_domain *domain;
614 
615 	domain = iommu_get_domain_for_dev(dev);
616 	if (!domain)
617 		return;
618 
619 	iommu_detach_device_pasid(domain, dev, idxd->pasid);
620 	iommu_free_global_pasid(idxd->pasid);
621 
622 	idxd_set_user_intr(idxd, 0);
623 	idxd->sva = NULL;
624 	idxd->pasid = IOMMU_PASID_INVALID;
625 }
626 
627 static int idxd_enable_sva(struct pci_dev *pdev)
628 {
629 	int ret;
630 
631 	ret = iommu_dev_enable_feature(&pdev->dev, IOMMU_DEV_FEAT_IOPF);
632 	if (ret)
633 		return ret;
634 
635 	ret = iommu_dev_enable_feature(&pdev->dev, IOMMU_DEV_FEAT_SVA);
636 	if (ret)
637 		iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_IOPF);
638 
639 	return ret;
640 }
641 
642 static void idxd_disable_sva(struct pci_dev *pdev)
643 {
644 	iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_SVA);
645 	iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_IOPF);
646 }
647 
648 static int idxd_probe(struct idxd_device *idxd)
649 {
650 	struct pci_dev *pdev = idxd->pdev;
651 	struct device *dev = &pdev->dev;
652 	int rc;
653 
654 	dev_dbg(dev, "%s entered and resetting device\n", __func__);
655 	rc = idxd_device_init_reset(idxd);
656 	if (rc < 0)
657 		return rc;
658 
659 	dev_dbg(dev, "IDXD reset complete\n");
660 
661 	if (IS_ENABLED(CONFIG_INTEL_IDXD_SVM) && sva) {
662 		if (idxd_enable_sva(pdev)) {
663 			dev_warn(dev, "Unable to turn on user SVA feature.\n");
664 		} else {
665 			set_bit(IDXD_FLAG_USER_PASID_ENABLED, &idxd->flags);
666 
667 			rc = idxd_enable_system_pasid(idxd);
668 			if (rc)
669 				dev_warn(dev, "No in-kernel DMA with PASID. %d\n", rc);
670 			else
671 				set_bit(IDXD_FLAG_PASID_ENABLED, &idxd->flags);
672 		}
673 	} else if (!sva) {
674 		dev_warn(dev, "User forced SVA off via module param.\n");
675 	}
676 
677 	idxd_read_caps(idxd);
678 	idxd_read_table_offsets(idxd);
679 
680 	rc = idxd_setup_internals(idxd);
681 	if (rc)
682 		goto err;
683 
684 	/* If the configs are readonly, then load them from device */
685 	if (!test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags)) {
686 		dev_dbg(dev, "Loading RO device config\n");
687 		rc = idxd_device_load_config(idxd);
688 		if (rc < 0)
689 			goto err_config;
690 	}
691 
692 	rc = idxd_setup_interrupts(idxd);
693 	if (rc)
694 		goto err_config;
695 
696 	idxd->major = idxd_cdev_get_major(idxd);
697 
698 	rc = perfmon_pmu_init(idxd);
699 	if (rc < 0)
700 		dev_warn(dev, "Failed to initialize perfmon. No PMU support: %d\n", rc);
701 
702 	dev_dbg(dev, "IDXD device %d probed successfully\n", idxd->id);
703 	return 0;
704 
705  err_config:
706 	idxd_cleanup_internals(idxd);
707  err:
708 	if (device_pasid_enabled(idxd))
709 		idxd_disable_system_pasid(idxd);
710 	if (device_user_pasid_enabled(idxd))
711 		idxd_disable_sva(pdev);
712 	return rc;
713 }
714 
715 static void idxd_cleanup(struct idxd_device *idxd)
716 {
717 	perfmon_pmu_remove(idxd);
718 	idxd_cleanup_interrupts(idxd);
719 	idxd_cleanup_internals(idxd);
720 	if (device_pasid_enabled(idxd))
721 		idxd_disable_system_pasid(idxd);
722 	if (device_user_pasid_enabled(idxd))
723 		idxd_disable_sva(idxd->pdev);
724 }
725 
726 static int idxd_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
727 {
728 	struct device *dev = &pdev->dev;
729 	struct idxd_device *idxd;
730 	struct idxd_driver_data *data = (struct idxd_driver_data *)id->driver_data;
731 	int rc;
732 
733 	rc = pci_enable_device(pdev);
734 	if (rc)
735 		return rc;
736 
737 	dev_dbg(dev, "Alloc IDXD context\n");
738 	idxd = idxd_alloc(pdev, data);
739 	if (!idxd) {
740 		rc = -ENOMEM;
741 		goto err_idxd_alloc;
742 	}
743 
744 	dev_dbg(dev, "Mapping BARs\n");
745 	idxd->reg_base = pci_iomap(pdev, IDXD_MMIO_BAR, 0);
746 	if (!idxd->reg_base) {
747 		rc = -ENOMEM;
748 		goto err_iomap;
749 	}
750 
751 	dev_dbg(dev, "Set DMA masks\n");
752 	rc = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
753 	if (rc)
754 		goto err;
755 
756 	dev_dbg(dev, "Set PCI master\n");
757 	pci_set_master(pdev);
758 	pci_set_drvdata(pdev, idxd);
759 
760 	idxd->hw.version = ioread32(idxd->reg_base + IDXD_VER_OFFSET);
761 	rc = idxd_probe(idxd);
762 	if (rc) {
763 		dev_err(dev, "Intel(R) IDXD DMA Engine init failed\n");
764 		goto err;
765 	}
766 
767 	if (data->load_device_defaults) {
768 		rc = data->load_device_defaults(idxd);
769 		if (rc)
770 			dev_warn(dev, "IDXD loading device defaults failed\n");
771 	}
772 
773 	rc = idxd_register_devices(idxd);
774 	if (rc) {
775 		dev_err(dev, "IDXD sysfs setup failed\n");
776 		goto err_dev_register;
777 	}
778 
779 	rc = idxd_device_init_debugfs(idxd);
780 	if (rc)
781 		dev_warn(dev, "IDXD debugfs failed to setup\n");
782 
783 	dev_info(&pdev->dev, "Intel(R) Accelerator Device (v%x)\n",
784 		 idxd->hw.version);
785 
786 	idxd->user_submission_safe = data->user_submission_safe;
787 
788 	return 0;
789 
790  err_dev_register:
791 	idxd_cleanup(idxd);
792  err:
793 	pci_iounmap(pdev, idxd->reg_base);
794  err_iomap:
795 	put_device(idxd_confdev(idxd));
796  err_idxd_alloc:
797 	pci_disable_device(pdev);
798 	return rc;
799 }
800 
801 void idxd_wqs_quiesce(struct idxd_device *idxd)
802 {
803 	struct idxd_wq *wq;
804 	int i;
805 
806 	for (i = 0; i < idxd->max_wqs; i++) {
807 		wq = idxd->wqs[i];
808 		if (wq->state == IDXD_WQ_ENABLED && wq->type == IDXD_WQT_KERNEL)
809 			idxd_wq_quiesce(wq);
810 	}
811 }
812 
813 static void idxd_shutdown(struct pci_dev *pdev)
814 {
815 	struct idxd_device *idxd = pci_get_drvdata(pdev);
816 	struct idxd_irq_entry *irq_entry;
817 	int rc;
818 
819 	rc = idxd_device_disable(idxd);
820 	if (rc)
821 		dev_err(&pdev->dev, "Disabling device failed\n");
822 
823 	irq_entry = &idxd->ie;
824 	synchronize_irq(irq_entry->vector);
825 	idxd_mask_error_interrupts(idxd);
826 	flush_workqueue(idxd->wq);
827 }
828 
829 static void idxd_remove(struct pci_dev *pdev)
830 {
831 	struct idxd_device *idxd = pci_get_drvdata(pdev);
832 	struct idxd_irq_entry *irq_entry;
833 
834 	idxd_unregister_devices(idxd);
835 	/*
836 	 * When ->release() is called for the idxd->conf_dev, it frees all the memory related
837 	 * to the idxd context. The driver still needs those bits in order to do the rest of
838 	 * the cleanup. However, we do need to unbound the idxd sub-driver. So take a ref
839 	 * on the device here to hold off the freeing while allowing the idxd sub-driver
840 	 * to unbind.
841 	 */
842 	get_device(idxd_confdev(idxd));
843 	device_unregister(idxd_confdev(idxd));
844 	idxd_shutdown(pdev);
845 	if (device_pasid_enabled(idxd))
846 		idxd_disable_system_pasid(idxd);
847 	idxd_device_remove_debugfs(idxd);
848 
849 	irq_entry = idxd_get_ie(idxd, 0);
850 	free_irq(irq_entry->vector, irq_entry);
851 	pci_free_irq_vectors(pdev);
852 	pci_iounmap(pdev, idxd->reg_base);
853 	if (device_user_pasid_enabled(idxd))
854 		idxd_disable_sva(pdev);
855 	pci_disable_device(pdev);
856 	destroy_workqueue(idxd->wq);
857 	perfmon_pmu_remove(idxd);
858 	put_device(idxd_confdev(idxd));
859 }
860 
861 static struct pci_driver idxd_pci_driver = {
862 	.name		= DRV_NAME,
863 	.id_table	= idxd_pci_tbl,
864 	.probe		= idxd_pci_probe,
865 	.remove		= idxd_remove,
866 	.shutdown	= idxd_shutdown,
867 };
868 
869 static int __init idxd_init_module(void)
870 {
871 	int err;
872 
873 	/*
874 	 * If the CPU does not support MOVDIR64B or ENQCMDS, there's no point in
875 	 * enumerating the device. We can not utilize it.
876 	 */
877 	if (!cpu_feature_enabled(X86_FEATURE_MOVDIR64B)) {
878 		pr_warn("idxd driver failed to load without MOVDIR64B.\n");
879 		return -ENODEV;
880 	}
881 
882 	if (!cpu_feature_enabled(X86_FEATURE_ENQCMD))
883 		pr_warn("Platform does not have ENQCMD(S) support.\n");
884 	else
885 		support_enqcmd = true;
886 
887 	err = idxd_driver_register(&idxd_drv);
888 	if (err < 0)
889 		goto err_idxd_driver_register;
890 
891 	err = idxd_driver_register(&idxd_dmaengine_drv);
892 	if (err < 0)
893 		goto err_idxd_dmaengine_driver_register;
894 
895 	err = idxd_driver_register(&idxd_user_drv);
896 	if (err < 0)
897 		goto err_idxd_user_driver_register;
898 
899 	err = idxd_cdev_register();
900 	if (err)
901 		goto err_cdev_register;
902 
903 	err = idxd_init_debugfs();
904 	if (err)
905 		goto err_debugfs;
906 
907 	err = pci_register_driver(&idxd_pci_driver);
908 	if (err)
909 		goto err_pci_register;
910 
911 	return 0;
912 
913 err_pci_register:
914 	idxd_remove_debugfs();
915 err_debugfs:
916 	idxd_cdev_remove();
917 err_cdev_register:
918 	idxd_driver_unregister(&idxd_user_drv);
919 err_idxd_user_driver_register:
920 	idxd_driver_unregister(&idxd_dmaengine_drv);
921 err_idxd_dmaengine_driver_register:
922 	idxd_driver_unregister(&idxd_drv);
923 err_idxd_driver_register:
924 	return err;
925 }
926 module_init(idxd_init_module);
927 
928 static void __exit idxd_exit_module(void)
929 {
930 	idxd_driver_unregister(&idxd_user_drv);
931 	idxd_driver_unregister(&idxd_dmaengine_drv);
932 	idxd_driver_unregister(&idxd_drv);
933 	pci_unregister_driver(&idxd_pci_driver);
934 	idxd_cdev_remove();
935 	idxd_remove_debugfs();
936 }
937 module_exit(idxd_exit_module);
938