xref: /linux/drivers/dma/idxd/init.c (revision 019b383d1132e4051de0d2e43254454b86538cf4)
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 	struct idxd_evl *evl;
346 
347 	if (idxd->hw.gen_cap.evl_support == 0)
348 		return 0;
349 
350 	evl = kzalloc_node(sizeof(*evl), GFP_KERNEL, dev_to_node(dev));
351 	if (!evl)
352 		return -ENOMEM;
353 
354 	spin_lock_init(&evl->lock);
355 	evl->size = IDXD_EVL_SIZE_MIN;
356 
357 	idxd->evl_cache = kmem_cache_create(dev_name(idxd_confdev(idxd)),
358 					    sizeof(struct idxd_evl_fault) + evl_ent_size(idxd),
359 					    0, 0, NULL);
360 	if (!idxd->evl_cache) {
361 		kfree(evl);
362 		return -ENOMEM;
363 	}
364 
365 	idxd->evl = evl;
366 	return 0;
367 }
368 
369 static int idxd_setup_internals(struct idxd_device *idxd)
370 {
371 	struct device *dev = &idxd->pdev->dev;
372 	int rc, i;
373 
374 	init_waitqueue_head(&idxd->cmd_waitq);
375 
376 	rc = idxd_setup_wqs(idxd);
377 	if (rc < 0)
378 		goto err_wqs;
379 
380 	rc = idxd_setup_engines(idxd);
381 	if (rc < 0)
382 		goto err_engine;
383 
384 	rc = idxd_setup_groups(idxd);
385 	if (rc < 0)
386 		goto err_group;
387 
388 	idxd->wq = create_workqueue(dev_name(dev));
389 	if (!idxd->wq) {
390 		rc = -ENOMEM;
391 		goto err_wkq_create;
392 	}
393 
394 	rc = idxd_init_evl(idxd);
395 	if (rc < 0)
396 		goto err_evl;
397 
398 	return 0;
399 
400  err_evl:
401 	destroy_workqueue(idxd->wq);
402  err_wkq_create:
403 	for (i = 0; i < idxd->max_groups; i++)
404 		put_device(group_confdev(idxd->groups[i]));
405  err_group:
406 	for (i = 0; i < idxd->max_engines; i++)
407 		put_device(engine_confdev(idxd->engines[i]));
408  err_engine:
409 	for (i = 0; i < idxd->max_wqs; i++)
410 		put_device(wq_confdev(idxd->wqs[i]));
411  err_wqs:
412 	return rc;
413 }
414 
415 static void idxd_read_table_offsets(struct idxd_device *idxd)
416 {
417 	union offsets_reg offsets;
418 	struct device *dev = &idxd->pdev->dev;
419 
420 	offsets.bits[0] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET);
421 	offsets.bits[1] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET + sizeof(u64));
422 	idxd->grpcfg_offset = offsets.grpcfg * IDXD_TABLE_MULT;
423 	dev_dbg(dev, "IDXD Group Config Offset: %#x\n", idxd->grpcfg_offset);
424 	idxd->wqcfg_offset = offsets.wqcfg * IDXD_TABLE_MULT;
425 	dev_dbg(dev, "IDXD Work Queue Config Offset: %#x\n", idxd->wqcfg_offset);
426 	idxd->msix_perm_offset = offsets.msix_perm * IDXD_TABLE_MULT;
427 	dev_dbg(dev, "IDXD MSIX Permission Offset: %#x\n", idxd->msix_perm_offset);
428 	idxd->perfmon_offset = offsets.perfmon * IDXD_TABLE_MULT;
429 	dev_dbg(dev, "IDXD Perfmon Offset: %#x\n", idxd->perfmon_offset);
430 }
431 
432 void multi_u64_to_bmap(unsigned long *bmap, u64 *val, int count)
433 {
434 	int i, j, nr;
435 
436 	for (i = 0, nr = 0; i < count; i++) {
437 		for (j = 0; j < BITS_PER_LONG_LONG; j++) {
438 			if (val[i] & BIT(j))
439 				set_bit(nr, bmap);
440 			nr++;
441 		}
442 	}
443 }
444 
445 static void idxd_read_caps(struct idxd_device *idxd)
446 {
447 	struct device *dev = &idxd->pdev->dev;
448 	int i;
449 
450 	/* reading generic capabilities */
451 	idxd->hw.gen_cap.bits = ioread64(idxd->reg_base + IDXD_GENCAP_OFFSET);
452 	dev_dbg(dev, "gen_cap: %#llx\n", idxd->hw.gen_cap.bits);
453 
454 	if (idxd->hw.gen_cap.cmd_cap) {
455 		idxd->hw.cmd_cap = ioread32(idxd->reg_base + IDXD_CMDCAP_OFFSET);
456 		dev_dbg(dev, "cmd_cap: %#x\n", idxd->hw.cmd_cap);
457 	}
458 
459 	/* reading command capabilities */
460 	if (idxd->hw.cmd_cap & BIT(IDXD_CMD_REQUEST_INT_HANDLE))
461 		idxd->request_int_handles = true;
462 
463 	idxd->max_xfer_bytes = 1ULL << idxd->hw.gen_cap.max_xfer_shift;
464 	dev_dbg(dev, "max xfer size: %llu bytes\n", idxd->max_xfer_bytes);
465 	idxd_set_max_batch_size(idxd->data->type, idxd, 1U << idxd->hw.gen_cap.max_batch_shift);
466 	dev_dbg(dev, "max batch size: %u\n", idxd->max_batch_size);
467 	if (idxd->hw.gen_cap.config_en)
468 		set_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags);
469 
470 	/* reading group capabilities */
471 	idxd->hw.group_cap.bits =
472 		ioread64(idxd->reg_base + IDXD_GRPCAP_OFFSET);
473 	dev_dbg(dev, "group_cap: %#llx\n", idxd->hw.group_cap.bits);
474 	idxd->max_groups = idxd->hw.group_cap.num_groups;
475 	dev_dbg(dev, "max groups: %u\n", idxd->max_groups);
476 	idxd->max_rdbufs = idxd->hw.group_cap.total_rdbufs;
477 	dev_dbg(dev, "max read buffers: %u\n", idxd->max_rdbufs);
478 	idxd->nr_rdbufs = idxd->max_rdbufs;
479 
480 	/* read engine capabilities */
481 	idxd->hw.engine_cap.bits =
482 		ioread64(idxd->reg_base + IDXD_ENGCAP_OFFSET);
483 	dev_dbg(dev, "engine_cap: %#llx\n", idxd->hw.engine_cap.bits);
484 	idxd->max_engines = idxd->hw.engine_cap.num_engines;
485 	dev_dbg(dev, "max engines: %u\n", idxd->max_engines);
486 
487 	/* read workqueue capabilities */
488 	idxd->hw.wq_cap.bits = ioread64(idxd->reg_base + IDXD_WQCAP_OFFSET);
489 	dev_dbg(dev, "wq_cap: %#llx\n", idxd->hw.wq_cap.bits);
490 	idxd->max_wq_size = idxd->hw.wq_cap.total_wq_size;
491 	dev_dbg(dev, "total workqueue size: %u\n", idxd->max_wq_size);
492 	idxd->max_wqs = idxd->hw.wq_cap.num_wqs;
493 	dev_dbg(dev, "max workqueues: %u\n", idxd->max_wqs);
494 	idxd->wqcfg_size = 1 << (idxd->hw.wq_cap.wqcfg_size + IDXD_WQCFG_MIN);
495 	dev_dbg(dev, "wqcfg size: %u\n", idxd->wqcfg_size);
496 
497 	/* reading operation capabilities */
498 	for (i = 0; i < 4; i++) {
499 		idxd->hw.opcap.bits[i] = ioread64(idxd->reg_base +
500 				IDXD_OPCAP_OFFSET + i * sizeof(u64));
501 		dev_dbg(dev, "opcap[%d]: %#llx\n", i, idxd->hw.opcap.bits[i]);
502 	}
503 	multi_u64_to_bmap(idxd->opcap_bmap, &idxd->hw.opcap.bits[0], 4);
504 
505 	/* read iaa cap */
506 	if (idxd->data->type == IDXD_TYPE_IAX && idxd->hw.version >= DEVICE_VERSION_2)
507 		idxd->hw.iaa_cap.bits = ioread64(idxd->reg_base + IDXD_IAACAP_OFFSET);
508 }
509 
510 static struct idxd_device *idxd_alloc(struct pci_dev *pdev, struct idxd_driver_data *data)
511 {
512 	struct device *dev = &pdev->dev;
513 	struct device *conf_dev;
514 	struct idxd_device *idxd;
515 	int rc;
516 
517 	idxd = kzalloc_node(sizeof(*idxd), GFP_KERNEL, dev_to_node(dev));
518 	if (!idxd)
519 		return NULL;
520 
521 	conf_dev = idxd_confdev(idxd);
522 	idxd->pdev = pdev;
523 	idxd->data = data;
524 	idxd_dev_set_type(&idxd->idxd_dev, idxd->data->type);
525 	idxd->id = ida_alloc(&idxd_ida, GFP_KERNEL);
526 	if (idxd->id < 0)
527 		return NULL;
528 
529 	idxd->opcap_bmap = bitmap_zalloc_node(IDXD_MAX_OPCAP_BITS, GFP_KERNEL, dev_to_node(dev));
530 	if (!idxd->opcap_bmap) {
531 		ida_free(&idxd_ida, idxd->id);
532 		return NULL;
533 	}
534 
535 	device_initialize(conf_dev);
536 	conf_dev->parent = dev;
537 	conf_dev->bus = &dsa_bus_type;
538 	conf_dev->type = idxd->data->dev_type;
539 	rc = dev_set_name(conf_dev, "%s%d", idxd->data->name_prefix, idxd->id);
540 	if (rc < 0) {
541 		put_device(conf_dev);
542 		return NULL;
543 	}
544 
545 	spin_lock_init(&idxd->dev_lock);
546 	spin_lock_init(&idxd->cmd_lock);
547 
548 	return idxd;
549 }
550 
551 static int idxd_enable_system_pasid(struct idxd_device *idxd)
552 {
553 	struct pci_dev *pdev = idxd->pdev;
554 	struct device *dev = &pdev->dev;
555 	struct iommu_domain *domain;
556 	ioasid_t pasid;
557 	int ret;
558 
559 	/*
560 	 * Attach a global PASID to the DMA domain so that we can use ENQCMDS
561 	 * to submit work on buffers mapped by DMA API.
562 	 */
563 	domain = iommu_get_domain_for_dev(dev);
564 	if (!domain)
565 		return -EPERM;
566 
567 	pasid = iommu_alloc_global_pasid(dev);
568 	if (pasid == IOMMU_PASID_INVALID)
569 		return -ENOSPC;
570 
571 	/*
572 	 * DMA domain is owned by the driver, it should support all valid
573 	 * types such as DMA-FQ, identity, etc.
574 	 */
575 	ret = iommu_attach_device_pasid(domain, dev, pasid);
576 	if (ret) {
577 		dev_err(dev, "failed to attach device pasid %d, domain type %d",
578 			pasid, domain->type);
579 		iommu_free_global_pasid(pasid);
580 		return ret;
581 	}
582 
583 	/* Since we set user privilege for kernel DMA, enable completion IRQ */
584 	idxd_set_user_intr(idxd, 1);
585 	idxd->pasid = pasid;
586 
587 	return ret;
588 }
589 
590 static void idxd_disable_system_pasid(struct idxd_device *idxd)
591 {
592 	struct pci_dev *pdev = idxd->pdev;
593 	struct device *dev = &pdev->dev;
594 	struct iommu_domain *domain;
595 
596 	domain = iommu_get_domain_for_dev(dev);
597 	if (!domain)
598 		return;
599 
600 	iommu_detach_device_pasid(domain, dev, idxd->pasid);
601 	iommu_free_global_pasid(idxd->pasid);
602 
603 	idxd_set_user_intr(idxd, 0);
604 	idxd->sva = NULL;
605 	idxd->pasid = IOMMU_PASID_INVALID;
606 }
607 
608 static int idxd_enable_sva(struct pci_dev *pdev)
609 {
610 	int ret;
611 
612 	ret = iommu_dev_enable_feature(&pdev->dev, IOMMU_DEV_FEAT_IOPF);
613 	if (ret)
614 		return ret;
615 
616 	ret = iommu_dev_enable_feature(&pdev->dev, IOMMU_DEV_FEAT_SVA);
617 	if (ret)
618 		iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_IOPF);
619 
620 	return ret;
621 }
622 
623 static void idxd_disable_sva(struct pci_dev *pdev)
624 {
625 	iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_SVA);
626 	iommu_dev_disable_feature(&pdev->dev, IOMMU_DEV_FEAT_IOPF);
627 }
628 
629 static int idxd_probe(struct idxd_device *idxd)
630 {
631 	struct pci_dev *pdev = idxd->pdev;
632 	struct device *dev = &pdev->dev;
633 	int rc;
634 
635 	dev_dbg(dev, "%s entered and resetting device\n", __func__);
636 	rc = idxd_device_init_reset(idxd);
637 	if (rc < 0)
638 		return rc;
639 
640 	dev_dbg(dev, "IDXD reset complete\n");
641 
642 	if (IS_ENABLED(CONFIG_INTEL_IDXD_SVM) && sva) {
643 		if (idxd_enable_sva(pdev)) {
644 			dev_warn(dev, "Unable to turn on user SVA feature.\n");
645 		} else {
646 			set_bit(IDXD_FLAG_USER_PASID_ENABLED, &idxd->flags);
647 
648 			rc = idxd_enable_system_pasid(idxd);
649 			if (rc)
650 				dev_warn(dev, "No in-kernel DMA with PASID. %d\n", rc);
651 			else
652 				set_bit(IDXD_FLAG_PASID_ENABLED, &idxd->flags);
653 		}
654 	} else if (!sva) {
655 		dev_warn(dev, "User forced SVA off via module param.\n");
656 	}
657 
658 	idxd_read_caps(idxd);
659 	idxd_read_table_offsets(idxd);
660 
661 	rc = idxd_setup_internals(idxd);
662 	if (rc)
663 		goto err;
664 
665 	/* If the configs are readonly, then load them from device */
666 	if (!test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags)) {
667 		dev_dbg(dev, "Loading RO device config\n");
668 		rc = idxd_device_load_config(idxd);
669 		if (rc < 0)
670 			goto err_config;
671 	}
672 
673 	rc = idxd_setup_interrupts(idxd);
674 	if (rc)
675 		goto err_config;
676 
677 	idxd->major = idxd_cdev_get_major(idxd);
678 
679 	rc = perfmon_pmu_init(idxd);
680 	if (rc < 0)
681 		dev_warn(dev, "Failed to initialize perfmon. No PMU support: %d\n", rc);
682 
683 	dev_dbg(dev, "IDXD device %d probed successfully\n", idxd->id);
684 	return 0;
685 
686  err_config:
687 	idxd_cleanup_internals(idxd);
688  err:
689 	if (device_pasid_enabled(idxd))
690 		idxd_disable_system_pasid(idxd);
691 	if (device_user_pasid_enabled(idxd))
692 		idxd_disable_sva(pdev);
693 	return rc;
694 }
695 
696 static void idxd_cleanup(struct idxd_device *idxd)
697 {
698 	perfmon_pmu_remove(idxd);
699 	idxd_cleanup_interrupts(idxd);
700 	idxd_cleanup_internals(idxd);
701 	if (device_pasid_enabled(idxd))
702 		idxd_disable_system_pasid(idxd);
703 	if (device_user_pasid_enabled(idxd))
704 		idxd_disable_sva(idxd->pdev);
705 }
706 
707 static int idxd_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
708 {
709 	struct device *dev = &pdev->dev;
710 	struct idxd_device *idxd;
711 	struct idxd_driver_data *data = (struct idxd_driver_data *)id->driver_data;
712 	int rc;
713 
714 	rc = pci_enable_device(pdev);
715 	if (rc)
716 		return rc;
717 
718 	dev_dbg(dev, "Alloc IDXD context\n");
719 	idxd = idxd_alloc(pdev, data);
720 	if (!idxd) {
721 		rc = -ENOMEM;
722 		goto err_idxd_alloc;
723 	}
724 
725 	dev_dbg(dev, "Mapping BARs\n");
726 	idxd->reg_base = pci_iomap(pdev, IDXD_MMIO_BAR, 0);
727 	if (!idxd->reg_base) {
728 		rc = -ENOMEM;
729 		goto err_iomap;
730 	}
731 
732 	dev_dbg(dev, "Set DMA masks\n");
733 	rc = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
734 	if (rc)
735 		goto err;
736 
737 	dev_dbg(dev, "Set PCI master\n");
738 	pci_set_master(pdev);
739 	pci_set_drvdata(pdev, idxd);
740 
741 	idxd->hw.version = ioread32(idxd->reg_base + IDXD_VER_OFFSET);
742 	rc = idxd_probe(idxd);
743 	if (rc) {
744 		dev_err(dev, "Intel(R) IDXD DMA Engine init failed\n");
745 		goto err;
746 	}
747 
748 	rc = idxd_register_devices(idxd);
749 	if (rc) {
750 		dev_err(dev, "IDXD sysfs setup failed\n");
751 		goto err_dev_register;
752 	}
753 
754 	rc = idxd_device_init_debugfs(idxd);
755 	if (rc)
756 		dev_warn(dev, "IDXD debugfs failed to setup\n");
757 
758 	dev_info(&pdev->dev, "Intel(R) Accelerator Device (v%x)\n",
759 		 idxd->hw.version);
760 
761 	return 0;
762 
763  err_dev_register:
764 	idxd_cleanup(idxd);
765  err:
766 	pci_iounmap(pdev, idxd->reg_base);
767  err_iomap:
768 	put_device(idxd_confdev(idxd));
769  err_idxd_alloc:
770 	pci_disable_device(pdev);
771 	return rc;
772 }
773 
774 void idxd_wqs_quiesce(struct idxd_device *idxd)
775 {
776 	struct idxd_wq *wq;
777 	int i;
778 
779 	for (i = 0; i < idxd->max_wqs; i++) {
780 		wq = idxd->wqs[i];
781 		if (wq->state == IDXD_WQ_ENABLED && wq->type == IDXD_WQT_KERNEL)
782 			idxd_wq_quiesce(wq);
783 	}
784 }
785 
786 static void idxd_shutdown(struct pci_dev *pdev)
787 {
788 	struct idxd_device *idxd = pci_get_drvdata(pdev);
789 	struct idxd_irq_entry *irq_entry;
790 	int rc;
791 
792 	rc = idxd_device_disable(idxd);
793 	if (rc)
794 		dev_err(&pdev->dev, "Disabling device failed\n");
795 
796 	irq_entry = &idxd->ie;
797 	synchronize_irq(irq_entry->vector);
798 	idxd_mask_error_interrupts(idxd);
799 	flush_workqueue(idxd->wq);
800 }
801 
802 static void idxd_remove(struct pci_dev *pdev)
803 {
804 	struct idxd_device *idxd = pci_get_drvdata(pdev);
805 	struct idxd_irq_entry *irq_entry;
806 
807 	idxd_unregister_devices(idxd);
808 	/*
809 	 * When ->release() is called for the idxd->conf_dev, it frees all the memory related
810 	 * to the idxd context. The driver still needs those bits in order to do the rest of
811 	 * the cleanup. However, we do need to unbound the idxd sub-driver. So take a ref
812 	 * on the device here to hold off the freeing while allowing the idxd sub-driver
813 	 * to unbind.
814 	 */
815 	get_device(idxd_confdev(idxd));
816 	device_unregister(idxd_confdev(idxd));
817 	idxd_shutdown(pdev);
818 	if (device_pasid_enabled(idxd))
819 		idxd_disable_system_pasid(idxd);
820 	idxd_device_remove_debugfs(idxd);
821 
822 	irq_entry = idxd_get_ie(idxd, 0);
823 	free_irq(irq_entry->vector, irq_entry);
824 	pci_free_irq_vectors(pdev);
825 	pci_iounmap(pdev, idxd->reg_base);
826 	if (device_user_pasid_enabled(idxd))
827 		idxd_disable_sva(pdev);
828 	pci_disable_device(pdev);
829 	destroy_workqueue(idxd->wq);
830 	perfmon_pmu_remove(idxd);
831 	put_device(idxd_confdev(idxd));
832 }
833 
834 static struct pci_driver idxd_pci_driver = {
835 	.name		= DRV_NAME,
836 	.id_table	= idxd_pci_tbl,
837 	.probe		= idxd_pci_probe,
838 	.remove		= idxd_remove,
839 	.shutdown	= idxd_shutdown,
840 };
841 
842 static int __init idxd_init_module(void)
843 {
844 	int err;
845 
846 	/*
847 	 * If the CPU does not support MOVDIR64B or ENQCMDS, there's no point in
848 	 * enumerating the device. We can not utilize it.
849 	 */
850 	if (!cpu_feature_enabled(X86_FEATURE_MOVDIR64B)) {
851 		pr_warn("idxd driver failed to load without MOVDIR64B.\n");
852 		return -ENODEV;
853 	}
854 
855 	if (!cpu_feature_enabled(X86_FEATURE_ENQCMD))
856 		pr_warn("Platform does not have ENQCMD(S) support.\n");
857 	else
858 		support_enqcmd = true;
859 
860 	perfmon_init();
861 
862 	err = idxd_driver_register(&idxd_drv);
863 	if (err < 0)
864 		goto err_idxd_driver_register;
865 
866 	err = idxd_driver_register(&idxd_dmaengine_drv);
867 	if (err < 0)
868 		goto err_idxd_dmaengine_driver_register;
869 
870 	err = idxd_driver_register(&idxd_user_drv);
871 	if (err < 0)
872 		goto err_idxd_user_driver_register;
873 
874 	err = idxd_cdev_register();
875 	if (err)
876 		goto err_cdev_register;
877 
878 	err = idxd_init_debugfs();
879 	if (err)
880 		goto err_debugfs;
881 
882 	err = pci_register_driver(&idxd_pci_driver);
883 	if (err)
884 		goto err_pci_register;
885 
886 	return 0;
887 
888 err_pci_register:
889 	idxd_remove_debugfs();
890 err_debugfs:
891 	idxd_cdev_remove();
892 err_cdev_register:
893 	idxd_driver_unregister(&idxd_user_drv);
894 err_idxd_user_driver_register:
895 	idxd_driver_unregister(&idxd_dmaengine_drv);
896 err_idxd_dmaengine_driver_register:
897 	idxd_driver_unregister(&idxd_drv);
898 err_idxd_driver_register:
899 	return err;
900 }
901 module_init(idxd_init_module);
902 
903 static void __exit idxd_exit_module(void)
904 {
905 	idxd_driver_unregister(&idxd_user_drv);
906 	idxd_driver_unregister(&idxd_dmaengine_drv);
907 	idxd_driver_unregister(&idxd_drv);
908 	pci_unregister_driver(&idxd_pci_driver);
909 	idxd_cdev_remove();
910 	perfmon_exit();
911 	idxd_remove_debugfs();
912 }
913 module_exit(idxd_exit_module);
914