xref: /linux/drivers/dma/idxd/init.c (revision 955abe0a1b41de5ba61fe4cd614ebc123084d499)
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 
73 	/* IAX ver 1.0 platforms */
74 	{ PCI_DEVICE_DATA(INTEL, IAX_SPR0, &idxd_driver_data[IDXD_TYPE_IAX]) },
75 	{ 0, }
76 };
77 MODULE_DEVICE_TABLE(pci, idxd_pci_tbl);
78 
79 static int idxd_setup_interrupts(struct idxd_device *idxd)
80 {
81 	struct pci_dev *pdev = idxd->pdev;
82 	struct device *dev = &pdev->dev;
83 	struct idxd_irq_entry *ie;
84 	int i, msixcnt;
85 	int rc = 0;
86 
87 	msixcnt = pci_msix_vec_count(pdev);
88 	if (msixcnt < 0) {
89 		dev_err(dev, "Not MSI-X interrupt capable.\n");
90 		return -ENOSPC;
91 	}
92 	idxd->irq_cnt = msixcnt;
93 
94 	rc = pci_alloc_irq_vectors(pdev, msixcnt, msixcnt, PCI_IRQ_MSIX);
95 	if (rc != msixcnt) {
96 		dev_err(dev, "Failed enabling %d MSIX entries: %d\n", msixcnt, rc);
97 		return -ENOSPC;
98 	}
99 	dev_dbg(dev, "Enabled %d msix vectors\n", msixcnt);
100 
101 
102 	ie = idxd_get_ie(idxd, 0);
103 	ie->vector = pci_irq_vector(pdev, 0);
104 	rc = request_threaded_irq(ie->vector, NULL, idxd_misc_thread, 0, "idxd-misc", ie);
105 	if (rc < 0) {
106 		dev_err(dev, "Failed to allocate misc interrupt.\n");
107 		goto err_misc_irq;
108 	}
109 	dev_dbg(dev, "Requested idxd-misc handler on msix vector %d\n", ie->vector);
110 
111 	for (i = 0; i < idxd->max_wqs; i++) {
112 		int msix_idx = i + 1;
113 
114 		ie = idxd_get_ie(idxd, msix_idx);
115 		ie->id = msix_idx;
116 		ie->int_handle = INVALID_INT_HANDLE;
117 		ie->pasid = IOMMU_PASID_INVALID;
118 
119 		spin_lock_init(&ie->list_lock);
120 		init_llist_head(&ie->pending_llist);
121 		INIT_LIST_HEAD(&ie->work_list);
122 	}
123 
124 	idxd_unmask_error_interrupts(idxd);
125 	return 0;
126 
127  err_misc_irq:
128 	idxd_mask_error_interrupts(idxd);
129 	pci_free_irq_vectors(pdev);
130 	dev_err(dev, "No usable interrupts\n");
131 	return rc;
132 }
133 
134 static void idxd_cleanup_interrupts(struct idxd_device *idxd)
135 {
136 	struct pci_dev *pdev = idxd->pdev;
137 	struct idxd_irq_entry *ie;
138 	int msixcnt;
139 
140 	msixcnt = pci_msix_vec_count(pdev);
141 	if (msixcnt <= 0)
142 		return;
143 
144 	ie = idxd_get_ie(idxd, 0);
145 	idxd_mask_error_interrupts(idxd);
146 	free_irq(ie->vector, ie);
147 	pci_free_irq_vectors(pdev);
148 }
149 
150 static int idxd_setup_wqs(struct idxd_device *idxd)
151 {
152 	struct device *dev = &idxd->pdev->dev;
153 	struct idxd_wq *wq;
154 	struct device *conf_dev;
155 	int i, rc;
156 
157 	idxd->wqs = kcalloc_node(idxd->max_wqs, sizeof(struct idxd_wq *),
158 				 GFP_KERNEL, dev_to_node(dev));
159 	if (!idxd->wqs)
160 		return -ENOMEM;
161 
162 	idxd->wq_enable_map = bitmap_zalloc_node(idxd->max_wqs, GFP_KERNEL, dev_to_node(dev));
163 	if (!idxd->wq_enable_map) {
164 		kfree(idxd->wqs);
165 		return -ENOMEM;
166 	}
167 
168 	for (i = 0; i < idxd->max_wqs; i++) {
169 		wq = kzalloc_node(sizeof(*wq), GFP_KERNEL, dev_to_node(dev));
170 		if (!wq) {
171 			rc = -ENOMEM;
172 			goto err;
173 		}
174 
175 		idxd_dev_set_type(&wq->idxd_dev, IDXD_DEV_WQ);
176 		conf_dev = wq_confdev(wq);
177 		wq->id = i;
178 		wq->idxd = idxd;
179 		device_initialize(wq_confdev(wq));
180 		conf_dev->parent = idxd_confdev(idxd);
181 		conf_dev->bus = &dsa_bus_type;
182 		conf_dev->type = &idxd_wq_device_type;
183 		rc = dev_set_name(conf_dev, "wq%d.%d", idxd->id, wq->id);
184 		if (rc < 0) {
185 			put_device(conf_dev);
186 			goto err;
187 		}
188 
189 		mutex_init(&wq->wq_lock);
190 		init_waitqueue_head(&wq->err_queue);
191 		init_completion(&wq->wq_dead);
192 		init_completion(&wq->wq_resurrect);
193 		wq->max_xfer_bytes = WQ_DEFAULT_MAX_XFER;
194 		idxd_wq_set_max_batch_size(idxd->data->type, wq, WQ_DEFAULT_MAX_BATCH);
195 		wq->enqcmds_retries = IDXD_ENQCMDS_RETRIES;
196 		wq->wqcfg = kzalloc_node(idxd->wqcfg_size, GFP_KERNEL, dev_to_node(dev));
197 		if (!wq->wqcfg) {
198 			put_device(conf_dev);
199 			rc = -ENOMEM;
200 			goto err;
201 		}
202 
203 		if (idxd->hw.wq_cap.op_config) {
204 			wq->opcap_bmap = bitmap_zalloc(IDXD_MAX_OPCAP_BITS, GFP_KERNEL);
205 			if (!wq->opcap_bmap) {
206 				put_device(conf_dev);
207 				rc = -ENOMEM;
208 				goto err;
209 			}
210 			bitmap_copy(wq->opcap_bmap, idxd->opcap_bmap, IDXD_MAX_OPCAP_BITS);
211 		}
212 		mutex_init(&wq->uc_lock);
213 		xa_init(&wq->upasid_xa);
214 		idxd->wqs[i] = wq;
215 	}
216 
217 	return 0;
218 
219  err:
220 	while (--i >= 0) {
221 		wq = idxd->wqs[i];
222 		conf_dev = wq_confdev(wq);
223 		put_device(conf_dev);
224 	}
225 	return rc;
226 }
227 
228 static int idxd_setup_engines(struct idxd_device *idxd)
229 {
230 	struct idxd_engine *engine;
231 	struct device *dev = &idxd->pdev->dev;
232 	struct device *conf_dev;
233 	int i, rc;
234 
235 	idxd->engines = kcalloc_node(idxd->max_engines, sizeof(struct idxd_engine *),
236 				     GFP_KERNEL, dev_to_node(dev));
237 	if (!idxd->engines)
238 		return -ENOMEM;
239 
240 	for (i = 0; i < idxd->max_engines; i++) {
241 		engine = kzalloc_node(sizeof(*engine), GFP_KERNEL, dev_to_node(dev));
242 		if (!engine) {
243 			rc = -ENOMEM;
244 			goto err;
245 		}
246 
247 		idxd_dev_set_type(&engine->idxd_dev, IDXD_DEV_ENGINE);
248 		conf_dev = engine_confdev(engine);
249 		engine->id = i;
250 		engine->idxd = idxd;
251 		device_initialize(conf_dev);
252 		conf_dev->parent = idxd_confdev(idxd);
253 		conf_dev->bus = &dsa_bus_type;
254 		conf_dev->type = &idxd_engine_device_type;
255 		rc = dev_set_name(conf_dev, "engine%d.%d", idxd->id, engine->id);
256 		if (rc < 0) {
257 			put_device(conf_dev);
258 			goto err;
259 		}
260 
261 		idxd->engines[i] = engine;
262 	}
263 
264 	return 0;
265 
266  err:
267 	while (--i >= 0) {
268 		engine = idxd->engines[i];
269 		conf_dev = engine_confdev(engine);
270 		put_device(conf_dev);
271 	}
272 	return rc;
273 }
274 
275 static int idxd_setup_groups(struct idxd_device *idxd)
276 {
277 	struct device *dev = &idxd->pdev->dev;
278 	struct device *conf_dev;
279 	struct idxd_group *group;
280 	int i, rc;
281 
282 	idxd->groups = kcalloc_node(idxd->max_groups, sizeof(struct idxd_group *),
283 				    GFP_KERNEL, dev_to_node(dev));
284 	if (!idxd->groups)
285 		return -ENOMEM;
286 
287 	for (i = 0; i < idxd->max_groups; i++) {
288 		group = kzalloc_node(sizeof(*group), GFP_KERNEL, dev_to_node(dev));
289 		if (!group) {
290 			rc = -ENOMEM;
291 			goto err;
292 		}
293 
294 		idxd_dev_set_type(&group->idxd_dev, IDXD_DEV_GROUP);
295 		conf_dev = group_confdev(group);
296 		group->id = i;
297 		group->idxd = idxd;
298 		device_initialize(conf_dev);
299 		conf_dev->parent = idxd_confdev(idxd);
300 		conf_dev->bus = &dsa_bus_type;
301 		conf_dev->type = &idxd_group_device_type;
302 		rc = dev_set_name(conf_dev, "group%d.%d", idxd->id, group->id);
303 		if (rc < 0) {
304 			put_device(conf_dev);
305 			goto err;
306 		}
307 
308 		idxd->groups[i] = group;
309 		if (idxd->hw.version <= DEVICE_VERSION_2 && !tc_override) {
310 			group->tc_a = 1;
311 			group->tc_b = 1;
312 		} else {
313 			group->tc_a = -1;
314 			group->tc_b = -1;
315 		}
316 		/*
317 		 * The default value is the same as the value of
318 		 * total read buffers in GRPCAP.
319 		 */
320 		group->rdbufs_allowed = idxd->max_rdbufs;
321 	}
322 
323 	return 0;
324 
325  err:
326 	while (--i >= 0) {
327 		group = idxd->groups[i];
328 		put_device(group_confdev(group));
329 	}
330 	return rc;
331 }
332 
333 static void idxd_cleanup_internals(struct idxd_device *idxd)
334 {
335 	int i;
336 
337 	for (i = 0; i < idxd->max_groups; i++)
338 		put_device(group_confdev(idxd->groups[i]));
339 	for (i = 0; i < idxd->max_engines; i++)
340 		put_device(engine_confdev(idxd->engines[i]));
341 	for (i = 0; i < idxd->max_wqs; i++)
342 		put_device(wq_confdev(idxd->wqs[i]));
343 	destroy_workqueue(idxd->wq);
344 }
345 
346 static int idxd_init_evl(struct idxd_device *idxd)
347 {
348 	struct device *dev = &idxd->pdev->dev;
349 	unsigned int evl_cache_size;
350 	struct idxd_evl *evl;
351 	const char *idxd_name;
352 
353 	if (idxd->hw.gen_cap.evl_support == 0)
354 		return 0;
355 
356 	evl = kzalloc_node(sizeof(*evl), GFP_KERNEL, dev_to_node(dev));
357 	if (!evl)
358 		return -ENOMEM;
359 
360 	mutex_init(&evl->lock);
361 	evl->size = IDXD_EVL_SIZE_MIN;
362 
363 	idxd_name = dev_name(idxd_confdev(idxd));
364 	evl_cache_size = sizeof(struct idxd_evl_fault) + evl_ent_size(idxd);
365 	/*
366 	 * Since completion record in evl_cache will be copied to user
367 	 * when handling completion record page fault, need to create
368 	 * the cache suitable for user copy.
369 	 */
370 	idxd->evl_cache = kmem_cache_create_usercopy(idxd_name, evl_cache_size,
371 						     0, 0, 0, evl_cache_size,
372 						     NULL);
373 	if (!idxd->evl_cache) {
374 		kfree(evl);
375 		return -ENOMEM;
376 	}
377 
378 	idxd->evl = evl;
379 	return 0;
380 }
381 
382 static int idxd_setup_internals(struct idxd_device *idxd)
383 {
384 	struct device *dev = &idxd->pdev->dev;
385 	int rc, i;
386 
387 	init_waitqueue_head(&idxd->cmd_waitq);
388 
389 	rc = idxd_setup_wqs(idxd);
390 	if (rc < 0)
391 		goto err_wqs;
392 
393 	rc = idxd_setup_engines(idxd);
394 	if (rc < 0)
395 		goto err_engine;
396 
397 	rc = idxd_setup_groups(idxd);
398 	if (rc < 0)
399 		goto err_group;
400 
401 	idxd->wq = create_workqueue(dev_name(dev));
402 	if (!idxd->wq) {
403 		rc = -ENOMEM;
404 		goto err_wkq_create;
405 	}
406 
407 	rc = idxd_init_evl(idxd);
408 	if (rc < 0)
409 		goto err_evl;
410 
411 	return 0;
412 
413  err_evl:
414 	destroy_workqueue(idxd->wq);
415  err_wkq_create:
416 	for (i = 0; i < idxd->max_groups; i++)
417 		put_device(group_confdev(idxd->groups[i]));
418  err_group:
419 	for (i = 0; i < idxd->max_engines; i++)
420 		put_device(engine_confdev(idxd->engines[i]));
421  err_engine:
422 	for (i = 0; i < idxd->max_wqs; i++)
423 		put_device(wq_confdev(idxd->wqs[i]));
424  err_wqs:
425 	return rc;
426 }
427 
428 static void idxd_read_table_offsets(struct idxd_device *idxd)
429 {
430 	union offsets_reg offsets;
431 	struct device *dev = &idxd->pdev->dev;
432 
433 	offsets.bits[0] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET);
434 	offsets.bits[1] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET + sizeof(u64));
435 	idxd->grpcfg_offset = offsets.grpcfg * IDXD_TABLE_MULT;
436 	dev_dbg(dev, "IDXD Group Config Offset: %#x\n", idxd->grpcfg_offset);
437 	idxd->wqcfg_offset = offsets.wqcfg * IDXD_TABLE_MULT;
438 	dev_dbg(dev, "IDXD Work Queue Config Offset: %#x\n", idxd->wqcfg_offset);
439 	idxd->msix_perm_offset = offsets.msix_perm * IDXD_TABLE_MULT;
440 	dev_dbg(dev, "IDXD MSIX Permission Offset: %#x\n", idxd->msix_perm_offset);
441 	idxd->perfmon_offset = offsets.perfmon * IDXD_TABLE_MULT;
442 	dev_dbg(dev, "IDXD Perfmon Offset: %#x\n", idxd->perfmon_offset);
443 }
444 
445 void multi_u64_to_bmap(unsigned long *bmap, u64 *val, int count)
446 {
447 	int i, j, nr;
448 
449 	for (i = 0, nr = 0; i < count; i++) {
450 		for (j = 0; j < BITS_PER_LONG_LONG; j++) {
451 			if (val[i] & BIT(j))
452 				set_bit(nr, bmap);
453 			nr++;
454 		}
455 	}
456 }
457 
458 static void idxd_read_caps(struct idxd_device *idxd)
459 {
460 	struct device *dev = &idxd->pdev->dev;
461 	int i;
462 
463 	/* reading generic capabilities */
464 	idxd->hw.gen_cap.bits = ioread64(idxd->reg_base + IDXD_GENCAP_OFFSET);
465 	dev_dbg(dev, "gen_cap: %#llx\n", idxd->hw.gen_cap.bits);
466 
467 	if (idxd->hw.gen_cap.cmd_cap) {
468 		idxd->hw.cmd_cap = ioread32(idxd->reg_base + IDXD_CMDCAP_OFFSET);
469 		dev_dbg(dev, "cmd_cap: %#x\n", idxd->hw.cmd_cap);
470 	}
471 
472 	/* reading command capabilities */
473 	if (idxd->hw.cmd_cap & BIT(IDXD_CMD_REQUEST_INT_HANDLE))
474 		idxd->request_int_handles = true;
475 
476 	idxd->max_xfer_bytes = 1ULL << idxd->hw.gen_cap.max_xfer_shift;
477 	dev_dbg(dev, "max xfer size: %llu bytes\n", idxd->max_xfer_bytes);
478 	idxd_set_max_batch_size(idxd->data->type, idxd, 1U << idxd->hw.gen_cap.max_batch_shift);
479 	dev_dbg(dev, "max batch size: %u\n", idxd->max_batch_size);
480 	if (idxd->hw.gen_cap.config_en)
481 		set_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags);
482 
483 	/* reading group capabilities */
484 	idxd->hw.group_cap.bits =
485 		ioread64(idxd->reg_base + IDXD_GRPCAP_OFFSET);
486 	dev_dbg(dev, "group_cap: %#llx\n", idxd->hw.group_cap.bits);
487 	idxd->max_groups = idxd->hw.group_cap.num_groups;
488 	dev_dbg(dev, "max groups: %u\n", idxd->max_groups);
489 	idxd->max_rdbufs = idxd->hw.group_cap.total_rdbufs;
490 	dev_dbg(dev, "max read buffers: %u\n", idxd->max_rdbufs);
491 	idxd->nr_rdbufs = idxd->max_rdbufs;
492 
493 	/* read engine capabilities */
494 	idxd->hw.engine_cap.bits =
495 		ioread64(idxd->reg_base + IDXD_ENGCAP_OFFSET);
496 	dev_dbg(dev, "engine_cap: %#llx\n", idxd->hw.engine_cap.bits);
497 	idxd->max_engines = idxd->hw.engine_cap.num_engines;
498 	dev_dbg(dev, "max engines: %u\n", idxd->max_engines);
499 
500 	/* read workqueue capabilities */
501 	idxd->hw.wq_cap.bits = ioread64(idxd->reg_base + IDXD_WQCAP_OFFSET);
502 	dev_dbg(dev, "wq_cap: %#llx\n", idxd->hw.wq_cap.bits);
503 	idxd->max_wq_size = idxd->hw.wq_cap.total_wq_size;
504 	dev_dbg(dev, "total workqueue size: %u\n", idxd->max_wq_size);
505 	idxd->max_wqs = idxd->hw.wq_cap.num_wqs;
506 	dev_dbg(dev, "max workqueues: %u\n", idxd->max_wqs);
507 	idxd->wqcfg_size = 1 << (idxd->hw.wq_cap.wqcfg_size + IDXD_WQCFG_MIN);
508 	dev_dbg(dev, "wqcfg size: %u\n", idxd->wqcfg_size);
509 
510 	/* reading operation capabilities */
511 	for (i = 0; i < 4; i++) {
512 		idxd->hw.opcap.bits[i] = ioread64(idxd->reg_base +
513 				IDXD_OPCAP_OFFSET + i * sizeof(u64));
514 		dev_dbg(dev, "opcap[%d]: %#llx\n", i, idxd->hw.opcap.bits[i]);
515 	}
516 	multi_u64_to_bmap(idxd->opcap_bmap, &idxd->hw.opcap.bits[0], 4);
517 
518 	/* read iaa cap */
519 	if (idxd->data->type == IDXD_TYPE_IAX && idxd->hw.version >= DEVICE_VERSION_2)
520 		idxd->hw.iaa_cap.bits = ioread64(idxd->reg_base + IDXD_IAACAP_OFFSET);
521 }
522 
523 static struct idxd_device *idxd_alloc(struct pci_dev *pdev, struct idxd_driver_data *data)
524 {
525 	struct device *dev = &pdev->dev;
526 	struct device *conf_dev;
527 	struct idxd_device *idxd;
528 	int rc;
529 
530 	idxd = kzalloc_node(sizeof(*idxd), GFP_KERNEL, dev_to_node(dev));
531 	if (!idxd)
532 		return NULL;
533 
534 	conf_dev = idxd_confdev(idxd);
535 	idxd->pdev = pdev;
536 	idxd->data = data;
537 	idxd_dev_set_type(&idxd->idxd_dev, idxd->data->type);
538 	idxd->id = ida_alloc(&idxd_ida, GFP_KERNEL);
539 	if (idxd->id < 0)
540 		return NULL;
541 
542 	idxd->opcap_bmap = bitmap_zalloc_node(IDXD_MAX_OPCAP_BITS, GFP_KERNEL, dev_to_node(dev));
543 	if (!idxd->opcap_bmap) {
544 		ida_free(&idxd_ida, idxd->id);
545 		return NULL;
546 	}
547 
548 	device_initialize(conf_dev);
549 	conf_dev->parent = dev;
550 	conf_dev->bus = &dsa_bus_type;
551 	conf_dev->type = idxd->data->dev_type;
552 	rc = dev_set_name(conf_dev, "%s%d", idxd->data->name_prefix, idxd->id);
553 	if (rc < 0) {
554 		put_device(conf_dev);
555 		return NULL;
556 	}
557 
558 	spin_lock_init(&idxd->dev_lock);
559 	spin_lock_init(&idxd->cmd_lock);
560 
561 	return idxd;
562 }
563 
564 static int idxd_enable_system_pasid(struct idxd_device *idxd)
565 {
566 	struct pci_dev *pdev = idxd->pdev;
567 	struct device *dev = &pdev->dev;
568 	struct iommu_domain *domain;
569 	ioasid_t pasid;
570 	int ret;
571 
572 	/*
573 	 * Attach a global PASID to the DMA domain so that we can use ENQCMDS
574 	 * to submit work on buffers mapped by DMA API.
575 	 */
576 	domain = iommu_get_domain_for_dev(dev);
577 	if (!domain)
578 		return -EPERM;
579 
580 	pasid = iommu_alloc_global_pasid(dev);
581 	if (pasid == IOMMU_PASID_INVALID)
582 		return -ENOSPC;
583 
584 	/*
585 	 * DMA domain is owned by the driver, it should support all valid
586 	 * types such as DMA-FQ, identity, etc.
587 	 */
588 	ret = iommu_attach_device_pasid(domain, dev, pasid, NULL);
589 	if (ret) {
590 		dev_err(dev, "failed to attach device pasid %d, domain type %d",
591 			pasid, domain->type);
592 		iommu_free_global_pasid(pasid);
593 		return ret;
594 	}
595 
596 	/* Since we set user privilege for kernel DMA, enable completion IRQ */
597 	idxd_set_user_intr(idxd, 1);
598 	idxd->pasid = pasid;
599 
600 	return ret;
601 }
602 
603 static void idxd_disable_system_pasid(struct idxd_device *idxd)
604 {
605 	struct pci_dev *pdev = idxd->pdev;
606 	struct device *dev = &pdev->dev;
607 	struct iommu_domain *domain;
608 
609 	domain = iommu_get_domain_for_dev(dev);
610 	if (!domain)
611 		return;
612 
613 	iommu_detach_device_pasid(domain, dev, idxd->pasid);
614 	iommu_free_global_pasid(idxd->pasid);
615 
616 	idxd_set_user_intr(idxd, 0);
617 	idxd->sva = NULL;
618 	idxd->pasid = IOMMU_PASID_INVALID;
619 }
620 
621 static int idxd_enable_sva(struct pci_dev *pdev)
622 {
623 	int ret;
624 
625 	ret = iommu_dev_enable_feature(&pdev->dev, IOMMU_DEV_FEAT_IOPF);
626 	if (ret)
627 		return ret;
628 
629 	ret = iommu_dev_enable_feature(&pdev->dev, IOMMU_DEV_FEAT_SVA);
630 	if (ret)
631 		iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_IOPF);
632 
633 	return ret;
634 }
635 
636 static void idxd_disable_sva(struct pci_dev *pdev)
637 {
638 	iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_SVA);
639 	iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_IOPF);
640 }
641 
642 static int idxd_probe(struct idxd_device *idxd)
643 {
644 	struct pci_dev *pdev = idxd->pdev;
645 	struct device *dev = &pdev->dev;
646 	int rc;
647 
648 	dev_dbg(dev, "%s entered and resetting device\n", __func__);
649 	rc = idxd_device_init_reset(idxd);
650 	if (rc < 0)
651 		return rc;
652 
653 	dev_dbg(dev, "IDXD reset complete\n");
654 
655 	if (IS_ENABLED(CONFIG_INTEL_IDXD_SVM) && sva) {
656 		if (idxd_enable_sva(pdev)) {
657 			dev_warn(dev, "Unable to turn on user SVA feature.\n");
658 		} else {
659 			set_bit(IDXD_FLAG_USER_PASID_ENABLED, &idxd->flags);
660 
661 			rc = idxd_enable_system_pasid(idxd);
662 			if (rc)
663 				dev_warn(dev, "No in-kernel DMA with PASID. %d\n", rc);
664 			else
665 				set_bit(IDXD_FLAG_PASID_ENABLED, &idxd->flags);
666 		}
667 	} else if (!sva) {
668 		dev_warn(dev, "User forced SVA off via module param.\n");
669 	}
670 
671 	idxd_read_caps(idxd);
672 	idxd_read_table_offsets(idxd);
673 
674 	rc = idxd_setup_internals(idxd);
675 	if (rc)
676 		goto err;
677 
678 	/* If the configs are readonly, then load them from device */
679 	if (!test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags)) {
680 		dev_dbg(dev, "Loading RO device config\n");
681 		rc = idxd_device_load_config(idxd);
682 		if (rc < 0)
683 			goto err_config;
684 	}
685 
686 	rc = idxd_setup_interrupts(idxd);
687 	if (rc)
688 		goto err_config;
689 
690 	idxd->major = idxd_cdev_get_major(idxd);
691 
692 	rc = perfmon_pmu_init(idxd);
693 	if (rc < 0)
694 		dev_warn(dev, "Failed to initialize perfmon. No PMU support: %d\n", rc);
695 
696 	dev_dbg(dev, "IDXD device %d probed successfully\n", idxd->id);
697 	return 0;
698 
699  err_config:
700 	idxd_cleanup_internals(idxd);
701  err:
702 	if (device_pasid_enabled(idxd))
703 		idxd_disable_system_pasid(idxd);
704 	if (device_user_pasid_enabled(idxd))
705 		idxd_disable_sva(pdev);
706 	return rc;
707 }
708 
709 static void idxd_cleanup(struct idxd_device *idxd)
710 {
711 	perfmon_pmu_remove(idxd);
712 	idxd_cleanup_interrupts(idxd);
713 	idxd_cleanup_internals(idxd);
714 	if (device_pasid_enabled(idxd))
715 		idxd_disable_system_pasid(idxd);
716 	if (device_user_pasid_enabled(idxd))
717 		idxd_disable_sva(idxd->pdev);
718 }
719 
720 static int idxd_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
721 {
722 	struct device *dev = &pdev->dev;
723 	struct idxd_device *idxd;
724 	struct idxd_driver_data *data = (struct idxd_driver_data *)id->driver_data;
725 	int rc;
726 
727 	rc = pci_enable_device(pdev);
728 	if (rc)
729 		return rc;
730 
731 	dev_dbg(dev, "Alloc IDXD context\n");
732 	idxd = idxd_alloc(pdev, data);
733 	if (!idxd) {
734 		rc = -ENOMEM;
735 		goto err_idxd_alloc;
736 	}
737 
738 	dev_dbg(dev, "Mapping BARs\n");
739 	idxd->reg_base = pci_iomap(pdev, IDXD_MMIO_BAR, 0);
740 	if (!idxd->reg_base) {
741 		rc = -ENOMEM;
742 		goto err_iomap;
743 	}
744 
745 	dev_dbg(dev, "Set DMA masks\n");
746 	rc = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
747 	if (rc)
748 		goto err;
749 
750 	dev_dbg(dev, "Set PCI master\n");
751 	pci_set_master(pdev);
752 	pci_set_drvdata(pdev, idxd);
753 
754 	idxd->hw.version = ioread32(idxd->reg_base + IDXD_VER_OFFSET);
755 	rc = idxd_probe(idxd);
756 	if (rc) {
757 		dev_err(dev, "Intel(R) IDXD DMA Engine init failed\n");
758 		goto err;
759 	}
760 
761 	if (data->load_device_defaults) {
762 		rc = data->load_device_defaults(idxd);
763 		if (rc)
764 			dev_warn(dev, "IDXD loading device defaults failed\n");
765 	}
766 
767 	rc = idxd_register_devices(idxd);
768 	if (rc) {
769 		dev_err(dev, "IDXD sysfs setup failed\n");
770 		goto err_dev_register;
771 	}
772 
773 	rc = idxd_device_init_debugfs(idxd);
774 	if (rc)
775 		dev_warn(dev, "IDXD debugfs failed to setup\n");
776 
777 	dev_info(&pdev->dev, "Intel(R) Accelerator Device (v%x)\n",
778 		 idxd->hw.version);
779 
780 	idxd->user_submission_safe = data->user_submission_safe;
781 
782 	return 0;
783 
784  err_dev_register:
785 	idxd_cleanup(idxd);
786  err:
787 	pci_iounmap(pdev, idxd->reg_base);
788  err_iomap:
789 	put_device(idxd_confdev(idxd));
790  err_idxd_alloc:
791 	pci_disable_device(pdev);
792 	return rc;
793 }
794 
795 void idxd_wqs_quiesce(struct idxd_device *idxd)
796 {
797 	struct idxd_wq *wq;
798 	int i;
799 
800 	for (i = 0; i < idxd->max_wqs; i++) {
801 		wq = idxd->wqs[i];
802 		if (wq->state == IDXD_WQ_ENABLED && wq->type == IDXD_WQT_KERNEL)
803 			idxd_wq_quiesce(wq);
804 	}
805 }
806 
807 static void idxd_shutdown(struct pci_dev *pdev)
808 {
809 	struct idxd_device *idxd = pci_get_drvdata(pdev);
810 	struct idxd_irq_entry *irq_entry;
811 	int rc;
812 
813 	rc = idxd_device_disable(idxd);
814 	if (rc)
815 		dev_err(&pdev->dev, "Disabling device failed\n");
816 
817 	irq_entry = &idxd->ie;
818 	synchronize_irq(irq_entry->vector);
819 	idxd_mask_error_interrupts(idxd);
820 	flush_workqueue(idxd->wq);
821 }
822 
823 static void idxd_remove(struct pci_dev *pdev)
824 {
825 	struct idxd_device *idxd = pci_get_drvdata(pdev);
826 	struct idxd_irq_entry *irq_entry;
827 
828 	idxd_unregister_devices(idxd);
829 	/*
830 	 * When ->release() is called for the idxd->conf_dev, it frees all the memory related
831 	 * to the idxd context. The driver still needs those bits in order to do the rest of
832 	 * the cleanup. However, we do need to unbound the idxd sub-driver. So take a ref
833 	 * on the device here to hold off the freeing while allowing the idxd sub-driver
834 	 * to unbind.
835 	 */
836 	get_device(idxd_confdev(idxd));
837 	device_unregister(idxd_confdev(idxd));
838 	idxd_shutdown(pdev);
839 	if (device_pasid_enabled(idxd))
840 		idxd_disable_system_pasid(idxd);
841 	idxd_device_remove_debugfs(idxd);
842 
843 	irq_entry = idxd_get_ie(idxd, 0);
844 	free_irq(irq_entry->vector, irq_entry);
845 	pci_free_irq_vectors(pdev);
846 	pci_iounmap(pdev, idxd->reg_base);
847 	if (device_user_pasid_enabled(idxd))
848 		idxd_disable_sva(pdev);
849 	pci_disable_device(pdev);
850 	destroy_workqueue(idxd->wq);
851 	perfmon_pmu_remove(idxd);
852 	put_device(idxd_confdev(idxd));
853 }
854 
855 static struct pci_driver idxd_pci_driver = {
856 	.name		= DRV_NAME,
857 	.id_table	= idxd_pci_tbl,
858 	.probe		= idxd_pci_probe,
859 	.remove		= idxd_remove,
860 	.shutdown	= idxd_shutdown,
861 };
862 
863 static int __init idxd_init_module(void)
864 {
865 	int err;
866 
867 	/*
868 	 * If the CPU does not support MOVDIR64B or ENQCMDS, there's no point in
869 	 * enumerating the device. We can not utilize it.
870 	 */
871 	if (!cpu_feature_enabled(X86_FEATURE_MOVDIR64B)) {
872 		pr_warn("idxd driver failed to load without MOVDIR64B.\n");
873 		return -ENODEV;
874 	}
875 
876 	if (!cpu_feature_enabled(X86_FEATURE_ENQCMD))
877 		pr_warn("Platform does not have ENQCMD(S) support.\n");
878 	else
879 		support_enqcmd = true;
880 
881 	perfmon_init();
882 
883 	err = idxd_driver_register(&idxd_drv);
884 	if (err < 0)
885 		goto err_idxd_driver_register;
886 
887 	err = idxd_driver_register(&idxd_dmaengine_drv);
888 	if (err < 0)
889 		goto err_idxd_dmaengine_driver_register;
890 
891 	err = idxd_driver_register(&idxd_user_drv);
892 	if (err < 0)
893 		goto err_idxd_user_driver_register;
894 
895 	err = idxd_cdev_register();
896 	if (err)
897 		goto err_cdev_register;
898 
899 	err = idxd_init_debugfs();
900 	if (err)
901 		goto err_debugfs;
902 
903 	err = pci_register_driver(&idxd_pci_driver);
904 	if (err)
905 		goto err_pci_register;
906 
907 	return 0;
908 
909 err_pci_register:
910 	idxd_remove_debugfs();
911 err_debugfs:
912 	idxd_cdev_remove();
913 err_cdev_register:
914 	idxd_driver_unregister(&idxd_user_drv);
915 err_idxd_user_driver_register:
916 	idxd_driver_unregister(&idxd_dmaengine_drv);
917 err_idxd_dmaengine_driver_register:
918 	idxd_driver_unregister(&idxd_drv);
919 err_idxd_driver_register:
920 	return err;
921 }
922 module_init(idxd_init_module);
923 
924 static void __exit idxd_exit_module(void)
925 {
926 	idxd_driver_unregister(&idxd_user_drv);
927 	idxd_driver_unregister(&idxd_dmaengine_drv);
928 	idxd_driver_unregister(&idxd_drv);
929 	pci_unregister_driver(&idxd_pci_driver);
930 	idxd_cdev_remove();
931 	perfmon_exit();
932 	idxd_remove_debugfs();
933 }
934 module_exit(idxd_exit_module);
935