xref: /openbmc/linux/drivers/dma/idxd/init.c (revision f0931824)
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 	},
63 };
64 
65 static struct pci_device_id idxd_pci_tbl[] = {
66 	/* DSA ver 1.0 platforms */
67 	{ PCI_DEVICE_DATA(INTEL, DSA_SPR0, &idxd_driver_data[IDXD_TYPE_DSA]) },
68 
69 	/* IAX ver 1.0 platforms */
70 	{ PCI_DEVICE_DATA(INTEL, IAX_SPR0, &idxd_driver_data[IDXD_TYPE_IAX]) },
71 	{ 0, }
72 };
73 MODULE_DEVICE_TABLE(pci, idxd_pci_tbl);
74 
75 static int idxd_setup_interrupts(struct idxd_device *idxd)
76 {
77 	struct pci_dev *pdev = idxd->pdev;
78 	struct device *dev = &pdev->dev;
79 	struct idxd_irq_entry *ie;
80 	int i, msixcnt;
81 	int rc = 0;
82 
83 	msixcnt = pci_msix_vec_count(pdev);
84 	if (msixcnt < 0) {
85 		dev_err(dev, "Not MSI-X interrupt capable.\n");
86 		return -ENOSPC;
87 	}
88 	idxd->irq_cnt = msixcnt;
89 
90 	rc = pci_alloc_irq_vectors(pdev, msixcnt, msixcnt, PCI_IRQ_MSIX);
91 	if (rc != msixcnt) {
92 		dev_err(dev, "Failed enabling %d MSIX entries: %d\n", msixcnt, rc);
93 		return -ENOSPC;
94 	}
95 	dev_dbg(dev, "Enabled %d msix vectors\n", msixcnt);
96 
97 
98 	ie = idxd_get_ie(idxd, 0);
99 	ie->vector = pci_irq_vector(pdev, 0);
100 	rc = request_threaded_irq(ie->vector, NULL, idxd_misc_thread, 0, "idxd-misc", ie);
101 	if (rc < 0) {
102 		dev_err(dev, "Failed to allocate misc interrupt.\n");
103 		goto err_misc_irq;
104 	}
105 	dev_dbg(dev, "Requested idxd-misc handler on msix vector %d\n", ie->vector);
106 
107 	for (i = 0; i < idxd->max_wqs; i++) {
108 		int msix_idx = i + 1;
109 
110 		ie = idxd_get_ie(idxd, msix_idx);
111 		ie->id = msix_idx;
112 		ie->int_handle = INVALID_INT_HANDLE;
113 		ie->pasid = IOMMU_PASID_INVALID;
114 
115 		spin_lock_init(&ie->list_lock);
116 		init_llist_head(&ie->pending_llist);
117 		INIT_LIST_HEAD(&ie->work_list);
118 	}
119 
120 	idxd_unmask_error_interrupts(idxd);
121 	return 0;
122 
123  err_misc_irq:
124 	idxd_mask_error_interrupts(idxd);
125 	pci_free_irq_vectors(pdev);
126 	dev_err(dev, "No usable interrupts\n");
127 	return rc;
128 }
129 
130 static void idxd_cleanup_interrupts(struct idxd_device *idxd)
131 {
132 	struct pci_dev *pdev = idxd->pdev;
133 	struct idxd_irq_entry *ie;
134 	int msixcnt;
135 
136 	msixcnt = pci_msix_vec_count(pdev);
137 	if (msixcnt <= 0)
138 		return;
139 
140 	ie = idxd_get_ie(idxd, 0);
141 	idxd_mask_error_interrupts(idxd);
142 	free_irq(ie->vector, ie);
143 	pci_free_irq_vectors(pdev);
144 }
145 
146 static int idxd_setup_wqs(struct idxd_device *idxd)
147 {
148 	struct device *dev = &idxd->pdev->dev;
149 	struct idxd_wq *wq;
150 	struct device *conf_dev;
151 	int i, rc;
152 
153 	idxd->wqs = kcalloc_node(idxd->max_wqs, sizeof(struct idxd_wq *),
154 				 GFP_KERNEL, dev_to_node(dev));
155 	if (!idxd->wqs)
156 		return -ENOMEM;
157 
158 	idxd->wq_enable_map = bitmap_zalloc_node(idxd->max_wqs, GFP_KERNEL, dev_to_node(dev));
159 	if (!idxd->wq_enable_map) {
160 		kfree(idxd->wqs);
161 		return -ENOMEM;
162 	}
163 
164 	for (i = 0; i < idxd->max_wqs; i++) {
165 		wq = kzalloc_node(sizeof(*wq), GFP_KERNEL, dev_to_node(dev));
166 		if (!wq) {
167 			rc = -ENOMEM;
168 			goto err;
169 		}
170 
171 		idxd_dev_set_type(&wq->idxd_dev, IDXD_DEV_WQ);
172 		conf_dev = wq_confdev(wq);
173 		wq->id = i;
174 		wq->idxd = idxd;
175 		device_initialize(wq_confdev(wq));
176 		conf_dev->parent = idxd_confdev(idxd);
177 		conf_dev->bus = &dsa_bus_type;
178 		conf_dev->type = &idxd_wq_device_type;
179 		rc = dev_set_name(conf_dev, "wq%d.%d", idxd->id, wq->id);
180 		if (rc < 0) {
181 			put_device(conf_dev);
182 			goto err;
183 		}
184 
185 		mutex_init(&wq->wq_lock);
186 		init_waitqueue_head(&wq->err_queue);
187 		init_completion(&wq->wq_dead);
188 		init_completion(&wq->wq_resurrect);
189 		wq->max_xfer_bytes = WQ_DEFAULT_MAX_XFER;
190 		idxd_wq_set_max_batch_size(idxd->data->type, wq, WQ_DEFAULT_MAX_BATCH);
191 		wq->enqcmds_retries = IDXD_ENQCMDS_RETRIES;
192 		wq->wqcfg = kzalloc_node(idxd->wqcfg_size, GFP_KERNEL, dev_to_node(dev));
193 		if (!wq->wqcfg) {
194 			put_device(conf_dev);
195 			rc = -ENOMEM;
196 			goto err;
197 		}
198 
199 		if (idxd->hw.wq_cap.op_config) {
200 			wq->opcap_bmap = bitmap_zalloc(IDXD_MAX_OPCAP_BITS, GFP_KERNEL);
201 			if (!wq->opcap_bmap) {
202 				put_device(conf_dev);
203 				rc = -ENOMEM;
204 				goto err;
205 			}
206 			bitmap_copy(wq->opcap_bmap, idxd->opcap_bmap, IDXD_MAX_OPCAP_BITS);
207 		}
208 		mutex_init(&wq->uc_lock);
209 		xa_init(&wq->upasid_xa);
210 		idxd->wqs[i] = wq;
211 	}
212 
213 	return 0;
214 
215  err:
216 	while (--i >= 0) {
217 		wq = idxd->wqs[i];
218 		conf_dev = wq_confdev(wq);
219 		put_device(conf_dev);
220 	}
221 	return rc;
222 }
223 
224 static int idxd_setup_engines(struct idxd_device *idxd)
225 {
226 	struct idxd_engine *engine;
227 	struct device *dev = &idxd->pdev->dev;
228 	struct device *conf_dev;
229 	int i, rc;
230 
231 	idxd->engines = kcalloc_node(idxd->max_engines, sizeof(struct idxd_engine *),
232 				     GFP_KERNEL, dev_to_node(dev));
233 	if (!idxd->engines)
234 		return -ENOMEM;
235 
236 	for (i = 0; i < idxd->max_engines; i++) {
237 		engine = kzalloc_node(sizeof(*engine), GFP_KERNEL, dev_to_node(dev));
238 		if (!engine) {
239 			rc = -ENOMEM;
240 			goto err;
241 		}
242 
243 		idxd_dev_set_type(&engine->idxd_dev, IDXD_DEV_ENGINE);
244 		conf_dev = engine_confdev(engine);
245 		engine->id = i;
246 		engine->idxd = idxd;
247 		device_initialize(conf_dev);
248 		conf_dev->parent = idxd_confdev(idxd);
249 		conf_dev->bus = &dsa_bus_type;
250 		conf_dev->type = &idxd_engine_device_type;
251 		rc = dev_set_name(conf_dev, "engine%d.%d", idxd->id, engine->id);
252 		if (rc < 0) {
253 			put_device(conf_dev);
254 			goto err;
255 		}
256 
257 		idxd->engines[i] = engine;
258 	}
259 
260 	return 0;
261 
262  err:
263 	while (--i >= 0) {
264 		engine = idxd->engines[i];
265 		conf_dev = engine_confdev(engine);
266 		put_device(conf_dev);
267 	}
268 	return rc;
269 }
270 
271 static int idxd_setup_groups(struct idxd_device *idxd)
272 {
273 	struct device *dev = &idxd->pdev->dev;
274 	struct device *conf_dev;
275 	struct idxd_group *group;
276 	int i, rc;
277 
278 	idxd->groups = kcalloc_node(idxd->max_groups, sizeof(struct idxd_group *),
279 				    GFP_KERNEL, dev_to_node(dev));
280 	if (!idxd->groups)
281 		return -ENOMEM;
282 
283 	for (i = 0; i < idxd->max_groups; i++) {
284 		group = kzalloc_node(sizeof(*group), GFP_KERNEL, dev_to_node(dev));
285 		if (!group) {
286 			rc = -ENOMEM;
287 			goto err;
288 		}
289 
290 		idxd_dev_set_type(&group->idxd_dev, IDXD_DEV_GROUP);
291 		conf_dev = group_confdev(group);
292 		group->id = i;
293 		group->idxd = idxd;
294 		device_initialize(conf_dev);
295 		conf_dev->parent = idxd_confdev(idxd);
296 		conf_dev->bus = &dsa_bus_type;
297 		conf_dev->type = &idxd_group_device_type;
298 		rc = dev_set_name(conf_dev, "group%d.%d", idxd->id, group->id);
299 		if (rc < 0) {
300 			put_device(conf_dev);
301 			goto err;
302 		}
303 
304 		idxd->groups[i] = group;
305 		if (idxd->hw.version <= DEVICE_VERSION_2 && !tc_override) {
306 			group->tc_a = 1;
307 			group->tc_b = 1;
308 		} else {
309 			group->tc_a = -1;
310 			group->tc_b = -1;
311 		}
312 		/*
313 		 * The default value is the same as the value of
314 		 * total read buffers in GRPCAP.
315 		 */
316 		group->rdbufs_allowed = idxd->max_rdbufs;
317 	}
318 
319 	return 0;
320 
321  err:
322 	while (--i >= 0) {
323 		group = idxd->groups[i];
324 		put_device(group_confdev(group));
325 	}
326 	return rc;
327 }
328 
329 static void idxd_cleanup_internals(struct idxd_device *idxd)
330 {
331 	int i;
332 
333 	for (i = 0; i < idxd->max_groups; i++)
334 		put_device(group_confdev(idxd->groups[i]));
335 	for (i = 0; i < idxd->max_engines; i++)
336 		put_device(engine_confdev(idxd->engines[i]));
337 	for (i = 0; i < idxd->max_wqs; i++)
338 		put_device(wq_confdev(idxd->wqs[i]));
339 	destroy_workqueue(idxd->wq);
340 }
341 
342 static int idxd_init_evl(struct idxd_device *idxd)
343 {
344 	struct device *dev = &idxd->pdev->dev;
345 	unsigned int evl_cache_size;
346 	struct idxd_evl *evl;
347 	const char *idxd_name;
348 
349 	if (idxd->hw.gen_cap.evl_support == 0)
350 		return 0;
351 
352 	evl = kzalloc_node(sizeof(*evl), GFP_KERNEL, dev_to_node(dev));
353 	if (!evl)
354 		return -ENOMEM;
355 
356 	mutex_init(&evl->lock);
357 	evl->size = IDXD_EVL_SIZE_MIN;
358 
359 	idxd_name = dev_name(idxd_confdev(idxd));
360 	evl_cache_size = sizeof(struct idxd_evl_fault) + evl_ent_size(idxd);
361 	/*
362 	 * Since completion record in evl_cache will be copied to user
363 	 * when handling completion record page fault, need to create
364 	 * the cache suitable for user copy.
365 	 */
366 	idxd->evl_cache = kmem_cache_create_usercopy(idxd_name, evl_cache_size,
367 						     0, 0, 0, evl_cache_size,
368 						     NULL);
369 	if (!idxd->evl_cache) {
370 		kfree(evl);
371 		return -ENOMEM;
372 	}
373 
374 	idxd->evl = evl;
375 	return 0;
376 }
377 
378 static int idxd_setup_internals(struct idxd_device *idxd)
379 {
380 	struct device *dev = &idxd->pdev->dev;
381 	int rc, i;
382 
383 	init_waitqueue_head(&idxd->cmd_waitq);
384 
385 	rc = idxd_setup_wqs(idxd);
386 	if (rc < 0)
387 		goto err_wqs;
388 
389 	rc = idxd_setup_engines(idxd);
390 	if (rc < 0)
391 		goto err_engine;
392 
393 	rc = idxd_setup_groups(idxd);
394 	if (rc < 0)
395 		goto err_group;
396 
397 	idxd->wq = create_workqueue(dev_name(dev));
398 	if (!idxd->wq) {
399 		rc = -ENOMEM;
400 		goto err_wkq_create;
401 	}
402 
403 	rc = idxd_init_evl(idxd);
404 	if (rc < 0)
405 		goto err_evl;
406 
407 	return 0;
408 
409  err_evl:
410 	destroy_workqueue(idxd->wq);
411  err_wkq_create:
412 	for (i = 0; i < idxd->max_groups; i++)
413 		put_device(group_confdev(idxd->groups[i]));
414  err_group:
415 	for (i = 0; i < idxd->max_engines; i++)
416 		put_device(engine_confdev(idxd->engines[i]));
417  err_engine:
418 	for (i = 0; i < idxd->max_wqs; i++)
419 		put_device(wq_confdev(idxd->wqs[i]));
420  err_wqs:
421 	return rc;
422 }
423 
424 static void idxd_read_table_offsets(struct idxd_device *idxd)
425 {
426 	union offsets_reg offsets;
427 	struct device *dev = &idxd->pdev->dev;
428 
429 	offsets.bits[0] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET);
430 	offsets.bits[1] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET + sizeof(u64));
431 	idxd->grpcfg_offset = offsets.grpcfg * IDXD_TABLE_MULT;
432 	dev_dbg(dev, "IDXD Group Config Offset: %#x\n", idxd->grpcfg_offset);
433 	idxd->wqcfg_offset = offsets.wqcfg * IDXD_TABLE_MULT;
434 	dev_dbg(dev, "IDXD Work Queue Config Offset: %#x\n", idxd->wqcfg_offset);
435 	idxd->msix_perm_offset = offsets.msix_perm * IDXD_TABLE_MULT;
436 	dev_dbg(dev, "IDXD MSIX Permission Offset: %#x\n", idxd->msix_perm_offset);
437 	idxd->perfmon_offset = offsets.perfmon * IDXD_TABLE_MULT;
438 	dev_dbg(dev, "IDXD Perfmon Offset: %#x\n", idxd->perfmon_offset);
439 }
440 
441 void multi_u64_to_bmap(unsigned long *bmap, u64 *val, int count)
442 {
443 	int i, j, nr;
444 
445 	for (i = 0, nr = 0; i < count; i++) {
446 		for (j = 0; j < BITS_PER_LONG_LONG; j++) {
447 			if (val[i] & BIT(j))
448 				set_bit(nr, bmap);
449 			nr++;
450 		}
451 	}
452 }
453 
454 static void idxd_read_caps(struct idxd_device *idxd)
455 {
456 	struct device *dev = &idxd->pdev->dev;
457 	int i;
458 
459 	/* reading generic capabilities */
460 	idxd->hw.gen_cap.bits = ioread64(idxd->reg_base + IDXD_GENCAP_OFFSET);
461 	dev_dbg(dev, "gen_cap: %#llx\n", idxd->hw.gen_cap.bits);
462 
463 	if (idxd->hw.gen_cap.cmd_cap) {
464 		idxd->hw.cmd_cap = ioread32(idxd->reg_base + IDXD_CMDCAP_OFFSET);
465 		dev_dbg(dev, "cmd_cap: %#x\n", idxd->hw.cmd_cap);
466 	}
467 
468 	/* reading command capabilities */
469 	if (idxd->hw.cmd_cap & BIT(IDXD_CMD_REQUEST_INT_HANDLE))
470 		idxd->request_int_handles = true;
471 
472 	idxd->max_xfer_bytes = 1ULL << idxd->hw.gen_cap.max_xfer_shift;
473 	dev_dbg(dev, "max xfer size: %llu bytes\n", idxd->max_xfer_bytes);
474 	idxd_set_max_batch_size(idxd->data->type, idxd, 1U << idxd->hw.gen_cap.max_batch_shift);
475 	dev_dbg(dev, "max batch size: %u\n", idxd->max_batch_size);
476 	if (idxd->hw.gen_cap.config_en)
477 		set_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags);
478 
479 	/* reading group capabilities */
480 	idxd->hw.group_cap.bits =
481 		ioread64(idxd->reg_base + IDXD_GRPCAP_OFFSET);
482 	dev_dbg(dev, "group_cap: %#llx\n", idxd->hw.group_cap.bits);
483 	idxd->max_groups = idxd->hw.group_cap.num_groups;
484 	dev_dbg(dev, "max groups: %u\n", idxd->max_groups);
485 	idxd->max_rdbufs = idxd->hw.group_cap.total_rdbufs;
486 	dev_dbg(dev, "max read buffers: %u\n", idxd->max_rdbufs);
487 	idxd->nr_rdbufs = idxd->max_rdbufs;
488 
489 	/* read engine capabilities */
490 	idxd->hw.engine_cap.bits =
491 		ioread64(idxd->reg_base + IDXD_ENGCAP_OFFSET);
492 	dev_dbg(dev, "engine_cap: %#llx\n", idxd->hw.engine_cap.bits);
493 	idxd->max_engines = idxd->hw.engine_cap.num_engines;
494 	dev_dbg(dev, "max engines: %u\n", idxd->max_engines);
495 
496 	/* read workqueue capabilities */
497 	idxd->hw.wq_cap.bits = ioread64(idxd->reg_base + IDXD_WQCAP_OFFSET);
498 	dev_dbg(dev, "wq_cap: %#llx\n", idxd->hw.wq_cap.bits);
499 	idxd->max_wq_size = idxd->hw.wq_cap.total_wq_size;
500 	dev_dbg(dev, "total workqueue size: %u\n", idxd->max_wq_size);
501 	idxd->max_wqs = idxd->hw.wq_cap.num_wqs;
502 	dev_dbg(dev, "max workqueues: %u\n", idxd->max_wqs);
503 	idxd->wqcfg_size = 1 << (idxd->hw.wq_cap.wqcfg_size + IDXD_WQCFG_MIN);
504 	dev_dbg(dev, "wqcfg size: %u\n", idxd->wqcfg_size);
505 
506 	/* reading operation capabilities */
507 	for (i = 0; i < 4; i++) {
508 		idxd->hw.opcap.bits[i] = ioread64(idxd->reg_base +
509 				IDXD_OPCAP_OFFSET + i * sizeof(u64));
510 		dev_dbg(dev, "opcap[%d]: %#llx\n", i, idxd->hw.opcap.bits[i]);
511 	}
512 	multi_u64_to_bmap(idxd->opcap_bmap, &idxd->hw.opcap.bits[0], 4);
513 
514 	/* read iaa cap */
515 	if (idxd->data->type == IDXD_TYPE_IAX && idxd->hw.version >= DEVICE_VERSION_2)
516 		idxd->hw.iaa_cap.bits = ioread64(idxd->reg_base + IDXD_IAACAP_OFFSET);
517 }
518 
519 static struct idxd_device *idxd_alloc(struct pci_dev *pdev, struct idxd_driver_data *data)
520 {
521 	struct device *dev = &pdev->dev;
522 	struct device *conf_dev;
523 	struct idxd_device *idxd;
524 	int rc;
525 
526 	idxd = kzalloc_node(sizeof(*idxd), GFP_KERNEL, dev_to_node(dev));
527 	if (!idxd)
528 		return NULL;
529 
530 	conf_dev = idxd_confdev(idxd);
531 	idxd->pdev = pdev;
532 	idxd->data = data;
533 	idxd_dev_set_type(&idxd->idxd_dev, idxd->data->type);
534 	idxd->id = ida_alloc(&idxd_ida, GFP_KERNEL);
535 	if (idxd->id < 0)
536 		return NULL;
537 
538 	idxd->opcap_bmap = bitmap_zalloc_node(IDXD_MAX_OPCAP_BITS, GFP_KERNEL, dev_to_node(dev));
539 	if (!idxd->opcap_bmap) {
540 		ida_free(&idxd_ida, idxd->id);
541 		return NULL;
542 	}
543 
544 	device_initialize(conf_dev);
545 	conf_dev->parent = dev;
546 	conf_dev->bus = &dsa_bus_type;
547 	conf_dev->type = idxd->data->dev_type;
548 	rc = dev_set_name(conf_dev, "%s%d", idxd->data->name_prefix, idxd->id);
549 	if (rc < 0) {
550 		put_device(conf_dev);
551 		return NULL;
552 	}
553 
554 	spin_lock_init(&idxd->dev_lock);
555 	spin_lock_init(&idxd->cmd_lock);
556 
557 	return idxd;
558 }
559 
560 static int idxd_enable_system_pasid(struct idxd_device *idxd)
561 {
562 	struct pci_dev *pdev = idxd->pdev;
563 	struct device *dev = &pdev->dev;
564 	struct iommu_domain *domain;
565 	ioasid_t pasid;
566 	int ret;
567 
568 	/*
569 	 * Attach a global PASID to the DMA domain so that we can use ENQCMDS
570 	 * to submit work on buffers mapped by DMA API.
571 	 */
572 	domain = iommu_get_domain_for_dev(dev);
573 	if (!domain)
574 		return -EPERM;
575 
576 	pasid = iommu_alloc_global_pasid(dev);
577 	if (pasid == IOMMU_PASID_INVALID)
578 		return -ENOSPC;
579 
580 	/*
581 	 * DMA domain is owned by the driver, it should support all valid
582 	 * types such as DMA-FQ, identity, etc.
583 	 */
584 	ret = iommu_attach_device_pasid(domain, dev, pasid);
585 	if (ret) {
586 		dev_err(dev, "failed to attach device pasid %d, domain type %d",
587 			pasid, domain->type);
588 		iommu_free_global_pasid(pasid);
589 		return ret;
590 	}
591 
592 	/* Since we set user privilege for kernel DMA, enable completion IRQ */
593 	idxd_set_user_intr(idxd, 1);
594 	idxd->pasid = pasid;
595 
596 	return ret;
597 }
598 
599 static void idxd_disable_system_pasid(struct idxd_device *idxd)
600 {
601 	struct pci_dev *pdev = idxd->pdev;
602 	struct device *dev = &pdev->dev;
603 	struct iommu_domain *domain;
604 
605 	domain = iommu_get_domain_for_dev(dev);
606 	if (!domain)
607 		return;
608 
609 	iommu_detach_device_pasid(domain, dev, idxd->pasid);
610 	iommu_free_global_pasid(idxd->pasid);
611 
612 	idxd_set_user_intr(idxd, 0);
613 	idxd->sva = NULL;
614 	idxd->pasid = IOMMU_PASID_INVALID;
615 }
616 
617 static int idxd_enable_sva(struct pci_dev *pdev)
618 {
619 	int ret;
620 
621 	ret = iommu_dev_enable_feature(&pdev->dev, IOMMU_DEV_FEAT_IOPF);
622 	if (ret)
623 		return ret;
624 
625 	ret = iommu_dev_enable_feature(&pdev->dev, IOMMU_DEV_FEAT_SVA);
626 	if (ret)
627 		iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_IOPF);
628 
629 	return ret;
630 }
631 
632 static void idxd_disable_sva(struct pci_dev *pdev)
633 {
634 	iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_SVA);
635 	iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_IOPF);
636 }
637 
638 static int idxd_probe(struct idxd_device *idxd)
639 {
640 	struct pci_dev *pdev = idxd->pdev;
641 	struct device *dev = &pdev->dev;
642 	int rc;
643 
644 	dev_dbg(dev, "%s entered and resetting device\n", __func__);
645 	rc = idxd_device_init_reset(idxd);
646 	if (rc < 0)
647 		return rc;
648 
649 	dev_dbg(dev, "IDXD reset complete\n");
650 
651 	if (IS_ENABLED(CONFIG_INTEL_IDXD_SVM) && sva) {
652 		if (idxd_enable_sva(pdev)) {
653 			dev_warn(dev, "Unable to turn on user SVA feature.\n");
654 		} else {
655 			set_bit(IDXD_FLAG_USER_PASID_ENABLED, &idxd->flags);
656 
657 			rc = idxd_enable_system_pasid(idxd);
658 			if (rc)
659 				dev_warn(dev, "No in-kernel DMA with PASID. %d\n", rc);
660 			else
661 				set_bit(IDXD_FLAG_PASID_ENABLED, &idxd->flags);
662 		}
663 	} else if (!sva) {
664 		dev_warn(dev, "User forced SVA off via module param.\n");
665 	}
666 
667 	idxd_read_caps(idxd);
668 	idxd_read_table_offsets(idxd);
669 
670 	rc = idxd_setup_internals(idxd);
671 	if (rc)
672 		goto err;
673 
674 	/* If the configs are readonly, then load them from device */
675 	if (!test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags)) {
676 		dev_dbg(dev, "Loading RO device config\n");
677 		rc = idxd_device_load_config(idxd);
678 		if (rc < 0)
679 			goto err_config;
680 	}
681 
682 	rc = idxd_setup_interrupts(idxd);
683 	if (rc)
684 		goto err_config;
685 
686 	idxd->major = idxd_cdev_get_major(idxd);
687 
688 	rc = perfmon_pmu_init(idxd);
689 	if (rc < 0)
690 		dev_warn(dev, "Failed to initialize perfmon. No PMU support: %d\n", rc);
691 
692 	dev_dbg(dev, "IDXD device %d probed successfully\n", idxd->id);
693 	return 0;
694 
695  err_config:
696 	idxd_cleanup_internals(idxd);
697  err:
698 	if (device_pasid_enabled(idxd))
699 		idxd_disable_system_pasid(idxd);
700 	if (device_user_pasid_enabled(idxd))
701 		idxd_disable_sva(pdev);
702 	return rc;
703 }
704 
705 static void idxd_cleanup(struct idxd_device *idxd)
706 {
707 	perfmon_pmu_remove(idxd);
708 	idxd_cleanup_interrupts(idxd);
709 	idxd_cleanup_internals(idxd);
710 	if (device_pasid_enabled(idxd))
711 		idxd_disable_system_pasid(idxd);
712 	if (device_user_pasid_enabled(idxd))
713 		idxd_disable_sva(idxd->pdev);
714 }
715 
716 static int idxd_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
717 {
718 	struct device *dev = &pdev->dev;
719 	struct idxd_device *idxd;
720 	struct idxd_driver_data *data = (struct idxd_driver_data *)id->driver_data;
721 	int rc;
722 
723 	rc = pci_enable_device(pdev);
724 	if (rc)
725 		return rc;
726 
727 	dev_dbg(dev, "Alloc IDXD context\n");
728 	idxd = idxd_alloc(pdev, data);
729 	if (!idxd) {
730 		rc = -ENOMEM;
731 		goto err_idxd_alloc;
732 	}
733 
734 	dev_dbg(dev, "Mapping BARs\n");
735 	idxd->reg_base = pci_iomap(pdev, IDXD_MMIO_BAR, 0);
736 	if (!idxd->reg_base) {
737 		rc = -ENOMEM;
738 		goto err_iomap;
739 	}
740 
741 	dev_dbg(dev, "Set DMA masks\n");
742 	rc = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
743 	if (rc)
744 		goto err;
745 
746 	dev_dbg(dev, "Set PCI master\n");
747 	pci_set_master(pdev);
748 	pci_set_drvdata(pdev, idxd);
749 
750 	idxd->hw.version = ioread32(idxd->reg_base + IDXD_VER_OFFSET);
751 	rc = idxd_probe(idxd);
752 	if (rc) {
753 		dev_err(dev, "Intel(R) IDXD DMA Engine init failed\n");
754 		goto err;
755 	}
756 
757 	rc = idxd_register_devices(idxd);
758 	if (rc) {
759 		dev_err(dev, "IDXD sysfs setup failed\n");
760 		goto err_dev_register;
761 	}
762 
763 	rc = idxd_device_init_debugfs(idxd);
764 	if (rc)
765 		dev_warn(dev, "IDXD debugfs failed to setup\n");
766 
767 	dev_info(&pdev->dev, "Intel(R) Accelerator Device (v%x)\n",
768 		 idxd->hw.version);
769 
770 	return 0;
771 
772  err_dev_register:
773 	idxd_cleanup(idxd);
774  err:
775 	pci_iounmap(pdev, idxd->reg_base);
776  err_iomap:
777 	put_device(idxd_confdev(idxd));
778  err_idxd_alloc:
779 	pci_disable_device(pdev);
780 	return rc;
781 }
782 
783 void idxd_wqs_quiesce(struct idxd_device *idxd)
784 {
785 	struct idxd_wq *wq;
786 	int i;
787 
788 	for (i = 0; i < idxd->max_wqs; i++) {
789 		wq = idxd->wqs[i];
790 		if (wq->state == IDXD_WQ_ENABLED && wq->type == IDXD_WQT_KERNEL)
791 			idxd_wq_quiesce(wq);
792 	}
793 }
794 
795 static void idxd_shutdown(struct pci_dev *pdev)
796 {
797 	struct idxd_device *idxd = pci_get_drvdata(pdev);
798 	struct idxd_irq_entry *irq_entry;
799 	int rc;
800 
801 	rc = idxd_device_disable(idxd);
802 	if (rc)
803 		dev_err(&pdev->dev, "Disabling device failed\n");
804 
805 	irq_entry = &idxd->ie;
806 	synchronize_irq(irq_entry->vector);
807 	idxd_mask_error_interrupts(idxd);
808 	flush_workqueue(idxd->wq);
809 }
810 
811 static void idxd_remove(struct pci_dev *pdev)
812 {
813 	struct idxd_device *idxd = pci_get_drvdata(pdev);
814 	struct idxd_irq_entry *irq_entry;
815 
816 	idxd_unregister_devices(idxd);
817 	/*
818 	 * When ->release() is called for the idxd->conf_dev, it frees all the memory related
819 	 * to the idxd context. The driver still needs those bits in order to do the rest of
820 	 * the cleanup. However, we do need to unbound the idxd sub-driver. So take a ref
821 	 * on the device here to hold off the freeing while allowing the idxd sub-driver
822 	 * to unbind.
823 	 */
824 	get_device(idxd_confdev(idxd));
825 	device_unregister(idxd_confdev(idxd));
826 	idxd_shutdown(pdev);
827 	if (device_pasid_enabled(idxd))
828 		idxd_disable_system_pasid(idxd);
829 	idxd_device_remove_debugfs(idxd);
830 
831 	irq_entry = idxd_get_ie(idxd, 0);
832 	free_irq(irq_entry->vector, irq_entry);
833 	pci_free_irq_vectors(pdev);
834 	pci_iounmap(pdev, idxd->reg_base);
835 	if (device_user_pasid_enabled(idxd))
836 		idxd_disable_sva(pdev);
837 	pci_disable_device(pdev);
838 	destroy_workqueue(idxd->wq);
839 	perfmon_pmu_remove(idxd);
840 	put_device(idxd_confdev(idxd));
841 }
842 
843 static struct pci_driver idxd_pci_driver = {
844 	.name		= DRV_NAME,
845 	.id_table	= idxd_pci_tbl,
846 	.probe		= idxd_pci_probe,
847 	.remove		= idxd_remove,
848 	.shutdown	= idxd_shutdown,
849 };
850 
851 static int __init idxd_init_module(void)
852 {
853 	int err;
854 
855 	/*
856 	 * If the CPU does not support MOVDIR64B or ENQCMDS, there's no point in
857 	 * enumerating the device. We can not utilize it.
858 	 */
859 	if (!cpu_feature_enabled(X86_FEATURE_MOVDIR64B)) {
860 		pr_warn("idxd driver failed to load without MOVDIR64B.\n");
861 		return -ENODEV;
862 	}
863 
864 	if (!cpu_feature_enabled(X86_FEATURE_ENQCMD))
865 		pr_warn("Platform does not have ENQCMD(S) support.\n");
866 	else
867 		support_enqcmd = true;
868 
869 	perfmon_init();
870 
871 	err = idxd_driver_register(&idxd_drv);
872 	if (err < 0)
873 		goto err_idxd_driver_register;
874 
875 	err = idxd_driver_register(&idxd_dmaengine_drv);
876 	if (err < 0)
877 		goto err_idxd_dmaengine_driver_register;
878 
879 	err = idxd_driver_register(&idxd_user_drv);
880 	if (err < 0)
881 		goto err_idxd_user_driver_register;
882 
883 	err = idxd_cdev_register();
884 	if (err)
885 		goto err_cdev_register;
886 
887 	err = idxd_init_debugfs();
888 	if (err)
889 		goto err_debugfs;
890 
891 	err = pci_register_driver(&idxd_pci_driver);
892 	if (err)
893 		goto err_pci_register;
894 
895 	return 0;
896 
897 err_pci_register:
898 	idxd_remove_debugfs();
899 err_debugfs:
900 	idxd_cdev_remove();
901 err_cdev_register:
902 	idxd_driver_unregister(&idxd_user_drv);
903 err_idxd_user_driver_register:
904 	idxd_driver_unregister(&idxd_dmaengine_drv);
905 err_idxd_dmaengine_driver_register:
906 	idxd_driver_unregister(&idxd_drv);
907 err_idxd_driver_register:
908 	return err;
909 }
910 module_init(idxd_init_module);
911 
912 static void __exit idxd_exit_module(void)
913 {
914 	idxd_driver_unregister(&idxd_user_drv);
915 	idxd_driver_unregister(&idxd_dmaengine_drv);
916 	idxd_driver_unregister(&idxd_drv);
917 	pci_unregister_driver(&idxd_pci_driver);
918 	idxd_cdev_remove();
919 	perfmon_exit();
920 	idxd_remove_debugfs();
921 }
922 module_exit(idxd_exit_module);
923