xref: /linux/drivers/dma/idxd/init.c (revision 0676181a87540fa2aabc8920dc30bca5071b8f1a)
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 	/* IAA PTL platforms */
82 	{ PCI_DEVICE_DATA(INTEL, IAA_PTL, &idxd_driver_data[IDXD_TYPE_IAX]) },
83 	{ 0, }
84 };
85 MODULE_DEVICE_TABLE(pci, idxd_pci_tbl);
86 
idxd_setup_interrupts(struct idxd_device * idxd)87 static int idxd_setup_interrupts(struct idxd_device *idxd)
88 {
89 	struct pci_dev *pdev = idxd->pdev;
90 	struct device *dev = &pdev->dev;
91 	struct idxd_irq_entry *ie;
92 	int i, msixcnt;
93 	int rc = 0;
94 
95 	msixcnt = pci_msix_vec_count(pdev);
96 	if (msixcnt < 0) {
97 		dev_err(dev, "Not MSI-X interrupt capable.\n");
98 		return -ENOSPC;
99 	}
100 	idxd->irq_cnt = msixcnt;
101 
102 	rc = pci_alloc_irq_vectors(pdev, msixcnt, msixcnt, PCI_IRQ_MSIX);
103 	if (rc != msixcnt) {
104 		dev_err(dev, "Failed enabling %d MSIX entries: %d\n", msixcnt, rc);
105 		return -ENOSPC;
106 	}
107 	dev_dbg(dev, "Enabled %d msix vectors\n", msixcnt);
108 
109 
110 	ie = idxd_get_ie(idxd, 0);
111 	ie->vector = pci_irq_vector(pdev, 0);
112 	rc = request_threaded_irq(ie->vector, NULL, idxd_misc_thread, 0, "idxd-misc", ie);
113 	if (rc < 0) {
114 		dev_err(dev, "Failed to allocate misc interrupt.\n");
115 		goto err_misc_irq;
116 	}
117 	dev_dbg(dev, "Requested idxd-misc handler on msix vector %d\n", ie->vector);
118 
119 	for (i = 0; i < idxd->max_wqs; i++) {
120 		int msix_idx = i + 1;
121 
122 		ie = idxd_get_ie(idxd, msix_idx);
123 		ie->id = msix_idx;
124 		ie->int_handle = INVALID_INT_HANDLE;
125 		ie->pasid = IOMMU_PASID_INVALID;
126 
127 		spin_lock_init(&ie->list_lock);
128 		init_llist_head(&ie->pending_llist);
129 		INIT_LIST_HEAD(&ie->work_list);
130 	}
131 
132 	idxd_unmask_error_interrupts(idxd);
133 	return 0;
134 
135  err_misc_irq:
136 	idxd_mask_error_interrupts(idxd);
137 	pci_free_irq_vectors(pdev);
138 	dev_err(dev, "No usable interrupts\n");
139 	return rc;
140 }
141 
idxd_cleanup_interrupts(struct idxd_device * idxd)142 static void idxd_cleanup_interrupts(struct idxd_device *idxd)
143 {
144 	struct pci_dev *pdev = idxd->pdev;
145 	struct idxd_irq_entry *ie;
146 	int msixcnt;
147 
148 	msixcnt = pci_msix_vec_count(pdev);
149 	if (msixcnt <= 0)
150 		return;
151 
152 	ie = idxd_get_ie(idxd, 0);
153 	idxd_mask_error_interrupts(idxd);
154 	free_irq(ie->vector, ie);
155 	pci_free_irq_vectors(pdev);
156 }
157 
idxd_clean_wqs(struct idxd_device * idxd)158 static void idxd_clean_wqs(struct idxd_device *idxd)
159 {
160 	struct idxd_wq *wq;
161 	struct device *conf_dev;
162 	int i;
163 
164 	for (i = 0; i < idxd->max_wqs; i++) {
165 		wq = idxd->wqs[i];
166 		if (idxd->hw.wq_cap.op_config)
167 			bitmap_free(wq->opcap_bmap);
168 		kfree(wq->wqcfg);
169 		conf_dev = wq_confdev(wq);
170 		put_device(conf_dev);
171 		kfree(wq);
172 	}
173 	bitmap_free(idxd->wq_enable_map);
174 	kfree(idxd->wqs);
175 }
176 
idxd_setup_wqs(struct idxd_device * idxd)177 static int idxd_setup_wqs(struct idxd_device *idxd)
178 {
179 	struct device *dev = &idxd->pdev->dev;
180 	struct idxd_wq *wq;
181 	struct device *conf_dev;
182 	int i, rc;
183 
184 	idxd->wqs = kcalloc_node(idxd->max_wqs, sizeof(struct idxd_wq *),
185 				 GFP_KERNEL, dev_to_node(dev));
186 	if (!idxd->wqs)
187 		return -ENOMEM;
188 
189 	idxd->wq_enable_map = bitmap_zalloc_node(idxd->max_wqs, GFP_KERNEL, dev_to_node(dev));
190 	if (!idxd->wq_enable_map) {
191 		rc = -ENOMEM;
192 		goto err_free_wqs;
193 	}
194 
195 	for (i = 0; i < idxd->max_wqs; i++) {
196 		wq = kzalloc_node(sizeof(*wq), GFP_KERNEL, dev_to_node(dev));
197 		if (!wq) {
198 			rc = -ENOMEM;
199 			goto err_unwind;
200 		}
201 
202 		idxd_dev_set_type(&wq->idxd_dev, IDXD_DEV_WQ);
203 		conf_dev = wq_confdev(wq);
204 		wq->id = i;
205 		wq->idxd = idxd;
206 		device_initialize(conf_dev);
207 		conf_dev->parent = idxd_confdev(idxd);
208 		conf_dev->bus = &dsa_bus_type;
209 		conf_dev->type = &idxd_wq_device_type;
210 		rc = dev_set_name(conf_dev, "wq%d.%d", idxd->id, wq->id);
211 		if (rc < 0) {
212 			put_device(conf_dev);
213 			kfree(wq);
214 			goto err_unwind;
215 		}
216 
217 		mutex_init(&wq->wq_lock);
218 		init_waitqueue_head(&wq->err_queue);
219 		init_completion(&wq->wq_dead);
220 		init_completion(&wq->wq_resurrect);
221 		wq->max_xfer_bytes = WQ_DEFAULT_MAX_XFER;
222 		idxd_wq_set_max_batch_size(idxd->data->type, wq, WQ_DEFAULT_MAX_BATCH);
223 		wq->enqcmds_retries = IDXD_ENQCMDS_RETRIES;
224 		wq->wqcfg = kzalloc_node(idxd->wqcfg_size, GFP_KERNEL, dev_to_node(dev));
225 		if (!wq->wqcfg) {
226 			put_device(conf_dev);
227 			kfree(wq);
228 			rc = -ENOMEM;
229 			goto err_unwind;
230 		}
231 
232 		if (idxd->hw.wq_cap.op_config) {
233 			wq->opcap_bmap = bitmap_zalloc(IDXD_MAX_OPCAP_BITS, GFP_KERNEL);
234 			if (!wq->opcap_bmap) {
235 				kfree(wq->wqcfg);
236 				put_device(conf_dev);
237 				kfree(wq);
238 				rc = -ENOMEM;
239 				goto err_unwind;
240 			}
241 			bitmap_copy(wq->opcap_bmap, idxd->opcap_bmap, IDXD_MAX_OPCAP_BITS);
242 		}
243 		mutex_init(&wq->uc_lock);
244 		xa_init(&wq->upasid_xa);
245 		idxd->wqs[i] = wq;
246 	}
247 
248 	return 0;
249 
250 err_unwind:
251 	while (--i >= 0) {
252 		wq = idxd->wqs[i];
253 		if (idxd->hw.wq_cap.op_config)
254 			bitmap_free(wq->opcap_bmap);
255 		kfree(wq->wqcfg);
256 		conf_dev = wq_confdev(wq);
257 		put_device(conf_dev);
258 		kfree(wq);
259 	}
260 	bitmap_free(idxd->wq_enable_map);
261 
262 err_free_wqs:
263 	kfree(idxd->wqs);
264 
265 	return rc;
266 }
267 
idxd_clean_engines(struct idxd_device * idxd)268 static void idxd_clean_engines(struct idxd_device *idxd)
269 {
270 	struct idxd_engine *engine;
271 	struct device *conf_dev;
272 	int i;
273 
274 	for (i = 0; i < idxd->max_engines; i++) {
275 		engine = idxd->engines[i];
276 		conf_dev = engine_confdev(engine);
277 		put_device(conf_dev);
278 		kfree(engine);
279 	}
280 	kfree(idxd->engines);
281 }
282 
idxd_setup_engines(struct idxd_device * idxd)283 static int idxd_setup_engines(struct idxd_device *idxd)
284 {
285 	struct idxd_engine *engine;
286 	struct device *dev = &idxd->pdev->dev;
287 	struct device *conf_dev;
288 	int i, rc;
289 
290 	idxd->engines = kcalloc_node(idxd->max_engines, sizeof(struct idxd_engine *),
291 				     GFP_KERNEL, dev_to_node(dev));
292 	if (!idxd->engines)
293 		return -ENOMEM;
294 
295 	for (i = 0; i < idxd->max_engines; i++) {
296 		engine = kzalloc_node(sizeof(*engine), GFP_KERNEL, dev_to_node(dev));
297 		if (!engine) {
298 			rc = -ENOMEM;
299 			goto err;
300 		}
301 
302 		idxd_dev_set_type(&engine->idxd_dev, IDXD_DEV_ENGINE);
303 		conf_dev = engine_confdev(engine);
304 		engine->id = i;
305 		engine->idxd = idxd;
306 		device_initialize(conf_dev);
307 		conf_dev->parent = idxd_confdev(idxd);
308 		conf_dev->bus = &dsa_bus_type;
309 		conf_dev->type = &idxd_engine_device_type;
310 		rc = dev_set_name(conf_dev, "engine%d.%d", idxd->id, engine->id);
311 		if (rc < 0) {
312 			put_device(conf_dev);
313 			kfree(engine);
314 			goto err;
315 		}
316 
317 		idxd->engines[i] = engine;
318 	}
319 
320 	return 0;
321 
322  err:
323 	while (--i >= 0) {
324 		engine = idxd->engines[i];
325 		conf_dev = engine_confdev(engine);
326 		put_device(conf_dev);
327 		kfree(engine);
328 	}
329 	kfree(idxd->engines);
330 
331 	return rc;
332 }
333 
idxd_clean_groups(struct idxd_device * idxd)334 static void idxd_clean_groups(struct idxd_device *idxd)
335 {
336 	struct idxd_group *group;
337 	int i;
338 
339 	for (i = 0; i < idxd->max_groups; i++) {
340 		group = idxd->groups[i];
341 		put_device(group_confdev(group));
342 		kfree(group);
343 	}
344 	kfree(idxd->groups);
345 }
346 
idxd_setup_groups(struct idxd_device * idxd)347 static int idxd_setup_groups(struct idxd_device *idxd)
348 {
349 	struct device *dev = &idxd->pdev->dev;
350 	struct device *conf_dev;
351 	struct idxd_group *group;
352 	int i, rc;
353 
354 	idxd->groups = kcalloc_node(idxd->max_groups, sizeof(struct idxd_group *),
355 				    GFP_KERNEL, dev_to_node(dev));
356 	if (!idxd->groups)
357 		return -ENOMEM;
358 
359 	for (i = 0; i < idxd->max_groups; i++) {
360 		group = kzalloc_node(sizeof(*group), GFP_KERNEL, dev_to_node(dev));
361 		if (!group) {
362 			rc = -ENOMEM;
363 			goto err;
364 		}
365 
366 		idxd_dev_set_type(&group->idxd_dev, IDXD_DEV_GROUP);
367 		conf_dev = group_confdev(group);
368 		group->id = i;
369 		group->idxd = idxd;
370 		device_initialize(conf_dev);
371 		conf_dev->parent = idxd_confdev(idxd);
372 		conf_dev->bus = &dsa_bus_type;
373 		conf_dev->type = &idxd_group_device_type;
374 		rc = dev_set_name(conf_dev, "group%d.%d", idxd->id, group->id);
375 		if (rc < 0) {
376 			put_device(conf_dev);
377 			kfree(group);
378 			goto err;
379 		}
380 
381 		idxd->groups[i] = group;
382 		if (idxd->hw.version <= DEVICE_VERSION_2 && !tc_override) {
383 			group->tc_a = 1;
384 			group->tc_b = 1;
385 		} else {
386 			group->tc_a = -1;
387 			group->tc_b = -1;
388 		}
389 		/*
390 		 * The default value is the same as the value of
391 		 * total read buffers in GRPCAP.
392 		 */
393 		group->rdbufs_allowed = idxd->max_rdbufs;
394 	}
395 
396 	return 0;
397 
398  err:
399 	while (--i >= 0) {
400 		group = idxd->groups[i];
401 		put_device(group_confdev(group));
402 		kfree(group);
403 	}
404 	kfree(idxd->groups);
405 
406 	return rc;
407 }
408 
idxd_cleanup_internals(struct idxd_device * idxd)409 static void idxd_cleanup_internals(struct idxd_device *idxd)
410 {
411 	idxd_clean_groups(idxd);
412 	idxd_clean_engines(idxd);
413 	idxd_clean_wqs(idxd);
414 	destroy_workqueue(idxd->wq);
415 }
416 
idxd_init_evl(struct idxd_device * idxd)417 static int idxd_init_evl(struct idxd_device *idxd)
418 {
419 	struct device *dev = &idxd->pdev->dev;
420 	unsigned int evl_cache_size;
421 	struct idxd_evl *evl;
422 	const char *idxd_name;
423 
424 	if (idxd->hw.gen_cap.evl_support == 0)
425 		return 0;
426 
427 	evl = kzalloc_node(sizeof(*evl), GFP_KERNEL, dev_to_node(dev));
428 	if (!evl)
429 		return -ENOMEM;
430 
431 	mutex_init(&evl->lock);
432 	evl->size = IDXD_EVL_SIZE_MIN;
433 
434 	idxd_name = dev_name(idxd_confdev(idxd));
435 	evl_cache_size = sizeof(struct idxd_evl_fault) + evl_ent_size(idxd);
436 	/*
437 	 * Since completion record in evl_cache will be copied to user
438 	 * when handling completion record page fault, need to create
439 	 * the cache suitable for user copy.
440 	 */
441 	idxd->evl_cache = kmem_cache_create_usercopy(idxd_name, evl_cache_size,
442 						     0, 0, 0, evl_cache_size,
443 						     NULL);
444 	if (!idxd->evl_cache) {
445 		kfree(evl);
446 		return -ENOMEM;
447 	}
448 
449 	idxd->evl = evl;
450 	return 0;
451 }
452 
idxd_setup_internals(struct idxd_device * idxd)453 static int idxd_setup_internals(struct idxd_device *idxd)
454 {
455 	struct device *dev = &idxd->pdev->dev;
456 	int rc;
457 
458 	init_waitqueue_head(&idxd->cmd_waitq);
459 
460 	rc = idxd_setup_wqs(idxd);
461 	if (rc < 0)
462 		goto err_wqs;
463 
464 	rc = idxd_setup_engines(idxd);
465 	if (rc < 0)
466 		goto err_engine;
467 
468 	rc = idxd_setup_groups(idxd);
469 	if (rc < 0)
470 		goto err_group;
471 
472 	idxd->wq = create_workqueue(dev_name(dev));
473 	if (!idxd->wq) {
474 		rc = -ENOMEM;
475 		goto err_wkq_create;
476 	}
477 
478 	rc = idxd_init_evl(idxd);
479 	if (rc < 0)
480 		goto err_evl;
481 
482 	return 0;
483 
484  err_evl:
485 	destroy_workqueue(idxd->wq);
486  err_wkq_create:
487 	idxd_clean_groups(idxd);
488  err_group:
489 	idxd_clean_engines(idxd);
490  err_engine:
491 	idxd_clean_wqs(idxd);
492  err_wqs:
493 	return rc;
494 }
495 
idxd_read_table_offsets(struct idxd_device * idxd)496 static void idxd_read_table_offsets(struct idxd_device *idxd)
497 {
498 	union offsets_reg offsets;
499 	struct device *dev = &idxd->pdev->dev;
500 
501 	offsets.bits[0] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET);
502 	offsets.bits[1] = ioread64(idxd->reg_base + IDXD_TABLE_OFFSET + sizeof(u64));
503 	idxd->grpcfg_offset = offsets.grpcfg * IDXD_TABLE_MULT;
504 	dev_dbg(dev, "IDXD Group Config Offset: %#x\n", idxd->grpcfg_offset);
505 	idxd->wqcfg_offset = offsets.wqcfg * IDXD_TABLE_MULT;
506 	dev_dbg(dev, "IDXD Work Queue Config Offset: %#x\n", idxd->wqcfg_offset);
507 	idxd->msix_perm_offset = offsets.msix_perm * IDXD_TABLE_MULT;
508 	dev_dbg(dev, "IDXD MSIX Permission Offset: %#x\n", idxd->msix_perm_offset);
509 	idxd->perfmon_offset = offsets.perfmon * IDXD_TABLE_MULT;
510 	dev_dbg(dev, "IDXD Perfmon Offset: %#x\n", idxd->perfmon_offset);
511 }
512 
multi_u64_to_bmap(unsigned long * bmap,u64 * val,int count)513 void multi_u64_to_bmap(unsigned long *bmap, u64 *val, int count)
514 {
515 	int i, j, nr;
516 
517 	for (i = 0, nr = 0; i < count; i++) {
518 		for (j = 0; j < BITS_PER_LONG_LONG; j++) {
519 			if (val[i] & BIT(j))
520 				set_bit(nr, bmap);
521 			nr++;
522 		}
523 	}
524 }
525 
idxd_read_caps(struct idxd_device * idxd)526 static void idxd_read_caps(struct idxd_device *idxd)
527 {
528 	struct device *dev = &idxd->pdev->dev;
529 	int i;
530 
531 	/* reading generic capabilities */
532 	idxd->hw.gen_cap.bits = ioread64(idxd->reg_base + IDXD_GENCAP_OFFSET);
533 	dev_dbg(dev, "gen_cap: %#llx\n", idxd->hw.gen_cap.bits);
534 
535 	if (idxd->hw.gen_cap.cmd_cap) {
536 		idxd->hw.cmd_cap = ioread32(idxd->reg_base + IDXD_CMDCAP_OFFSET);
537 		dev_dbg(dev, "cmd_cap: %#x\n", idxd->hw.cmd_cap);
538 	}
539 
540 	/* reading command capabilities */
541 	if (idxd->hw.cmd_cap & BIT(IDXD_CMD_REQUEST_INT_HANDLE))
542 		idxd->request_int_handles = true;
543 
544 	idxd->max_xfer_bytes = 1ULL << idxd->hw.gen_cap.max_xfer_shift;
545 	dev_dbg(dev, "max xfer size: %llu bytes\n", idxd->max_xfer_bytes);
546 	idxd_set_max_batch_size(idxd->data->type, idxd, 1U << idxd->hw.gen_cap.max_batch_shift);
547 	dev_dbg(dev, "max batch size: %u\n", idxd->max_batch_size);
548 	if (idxd->hw.gen_cap.config_en)
549 		set_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags);
550 
551 	/* reading group capabilities */
552 	idxd->hw.group_cap.bits =
553 		ioread64(idxd->reg_base + IDXD_GRPCAP_OFFSET);
554 	dev_dbg(dev, "group_cap: %#llx\n", idxd->hw.group_cap.bits);
555 	idxd->max_groups = idxd->hw.group_cap.num_groups;
556 	dev_dbg(dev, "max groups: %u\n", idxd->max_groups);
557 	idxd->max_rdbufs = idxd->hw.group_cap.total_rdbufs;
558 	dev_dbg(dev, "max read buffers: %u\n", idxd->max_rdbufs);
559 	idxd->nr_rdbufs = idxd->max_rdbufs;
560 
561 	/* read engine capabilities */
562 	idxd->hw.engine_cap.bits =
563 		ioread64(idxd->reg_base + IDXD_ENGCAP_OFFSET);
564 	dev_dbg(dev, "engine_cap: %#llx\n", idxd->hw.engine_cap.bits);
565 	idxd->max_engines = idxd->hw.engine_cap.num_engines;
566 	dev_dbg(dev, "max engines: %u\n", idxd->max_engines);
567 
568 	/* read workqueue capabilities */
569 	idxd->hw.wq_cap.bits = ioread64(idxd->reg_base + IDXD_WQCAP_OFFSET);
570 	dev_dbg(dev, "wq_cap: %#llx\n", idxd->hw.wq_cap.bits);
571 	idxd->max_wq_size = idxd->hw.wq_cap.total_wq_size;
572 	dev_dbg(dev, "total workqueue size: %u\n", idxd->max_wq_size);
573 	idxd->max_wqs = idxd->hw.wq_cap.num_wqs;
574 	dev_dbg(dev, "max workqueues: %u\n", idxd->max_wqs);
575 	idxd->wqcfg_size = 1 << (idxd->hw.wq_cap.wqcfg_size + IDXD_WQCFG_MIN);
576 	dev_dbg(dev, "wqcfg size: %u\n", idxd->wqcfg_size);
577 
578 	/* reading operation capabilities */
579 	for (i = 0; i < 4; i++) {
580 		idxd->hw.opcap.bits[i] = ioread64(idxd->reg_base +
581 				IDXD_OPCAP_OFFSET + i * sizeof(u64));
582 		dev_dbg(dev, "opcap[%d]: %#llx\n", i, idxd->hw.opcap.bits[i]);
583 	}
584 	multi_u64_to_bmap(idxd->opcap_bmap, &idxd->hw.opcap.bits[0], 4);
585 
586 	/* read iaa cap */
587 	if (idxd->data->type == IDXD_TYPE_IAX && idxd->hw.version >= DEVICE_VERSION_2)
588 		idxd->hw.iaa_cap.bits = ioread64(idxd->reg_base + IDXD_IAACAP_OFFSET);
589 }
590 
idxd_free(struct idxd_device * idxd)591 static void idxd_free(struct idxd_device *idxd)
592 {
593 	if (!idxd)
594 		return;
595 
596 	put_device(idxd_confdev(idxd));
597 	bitmap_free(idxd->opcap_bmap);
598 	ida_free(&idxd_ida, idxd->id);
599 	kfree(idxd);
600 }
601 
idxd_alloc(struct pci_dev * pdev,struct idxd_driver_data * data)602 static struct idxd_device *idxd_alloc(struct pci_dev *pdev, struct idxd_driver_data *data)
603 {
604 	struct device *dev = &pdev->dev;
605 	struct device *conf_dev;
606 	struct idxd_device *idxd;
607 	int rc;
608 
609 	idxd = kzalloc_node(sizeof(*idxd), GFP_KERNEL, dev_to_node(dev));
610 	if (!idxd)
611 		return NULL;
612 
613 	conf_dev = idxd_confdev(idxd);
614 	idxd->pdev = pdev;
615 	idxd->data = data;
616 	idxd_dev_set_type(&idxd->idxd_dev, idxd->data->type);
617 	idxd->id = ida_alloc(&idxd_ida, GFP_KERNEL);
618 	if (idxd->id < 0)
619 		goto err_ida;
620 
621 	idxd->opcap_bmap = bitmap_zalloc_node(IDXD_MAX_OPCAP_BITS, GFP_KERNEL, dev_to_node(dev));
622 	if (!idxd->opcap_bmap)
623 		goto err_opcap;
624 
625 	device_initialize(conf_dev);
626 	conf_dev->parent = dev;
627 	conf_dev->bus = &dsa_bus_type;
628 	conf_dev->type = idxd->data->dev_type;
629 	rc = dev_set_name(conf_dev, "%s%d", idxd->data->name_prefix, idxd->id);
630 	if (rc < 0)
631 		goto err_name;
632 
633 	spin_lock_init(&idxd->dev_lock);
634 	spin_lock_init(&idxd->cmd_lock);
635 
636 	return idxd;
637 
638 err_name:
639 	put_device(conf_dev);
640 	bitmap_free(idxd->opcap_bmap);
641 err_opcap:
642 	ida_free(&idxd_ida, idxd->id);
643 err_ida:
644 	kfree(idxd);
645 
646 	return NULL;
647 }
648 
idxd_enable_system_pasid(struct idxd_device * idxd)649 static int idxd_enable_system_pasid(struct idxd_device *idxd)
650 {
651 	struct pci_dev *pdev = idxd->pdev;
652 	struct device *dev = &pdev->dev;
653 	struct iommu_domain *domain;
654 	ioasid_t pasid;
655 	int ret;
656 
657 	/*
658 	 * Attach a global PASID to the DMA domain so that we can use ENQCMDS
659 	 * to submit work on buffers mapped by DMA API.
660 	 */
661 	domain = iommu_get_domain_for_dev(dev);
662 	if (!domain)
663 		return -EPERM;
664 
665 	pasid = iommu_alloc_global_pasid(dev);
666 	if (pasid == IOMMU_PASID_INVALID)
667 		return -ENOSPC;
668 
669 	/*
670 	 * DMA domain is owned by the driver, it should support all valid
671 	 * types such as DMA-FQ, identity, etc.
672 	 */
673 	ret = iommu_attach_device_pasid(domain, dev, pasid, NULL);
674 	if (ret) {
675 		dev_err(dev, "failed to attach device pasid %d, domain type %d",
676 			pasid, domain->type);
677 		iommu_free_global_pasid(pasid);
678 		return ret;
679 	}
680 
681 	/* Since we set user privilege for kernel DMA, enable completion IRQ */
682 	idxd_set_user_intr(idxd, 1);
683 	idxd->pasid = pasid;
684 
685 	return ret;
686 }
687 
idxd_disable_system_pasid(struct idxd_device * idxd)688 static void idxd_disable_system_pasid(struct idxd_device *idxd)
689 {
690 	struct pci_dev *pdev = idxd->pdev;
691 	struct device *dev = &pdev->dev;
692 	struct iommu_domain *domain;
693 
694 	domain = iommu_get_domain_for_dev(dev);
695 	if (!domain)
696 		return;
697 
698 	iommu_detach_device_pasid(domain, dev, idxd->pasid);
699 	iommu_free_global_pasid(idxd->pasid);
700 
701 	idxd_set_user_intr(idxd, 0);
702 	idxd->sva = NULL;
703 	idxd->pasid = IOMMU_PASID_INVALID;
704 }
705 
idxd_probe(struct idxd_device * idxd)706 static int idxd_probe(struct idxd_device *idxd)
707 {
708 	struct pci_dev *pdev = idxd->pdev;
709 	struct device *dev = &pdev->dev;
710 	int rc;
711 
712 	dev_dbg(dev, "%s entered and resetting device\n", __func__);
713 	rc = idxd_device_init_reset(idxd);
714 	if (rc < 0)
715 		return rc;
716 
717 	dev_dbg(dev, "IDXD reset complete\n");
718 
719 	if (IS_ENABLED(CONFIG_INTEL_IDXD_SVM) && sva) {
720 		set_bit(IDXD_FLAG_USER_PASID_ENABLED, &idxd->flags);
721 
722 		rc = idxd_enable_system_pasid(idxd);
723 		if (rc)
724 			dev_warn(dev, "No in-kernel DMA with PASID. %d\n", rc);
725 		else
726 			set_bit(IDXD_FLAG_PASID_ENABLED, &idxd->flags);
727 	} else if (!sva) {
728 		dev_warn(dev, "User forced SVA off via module param.\n");
729 	}
730 
731 	idxd_read_caps(idxd);
732 	idxd_read_table_offsets(idxd);
733 
734 	rc = idxd_setup_internals(idxd);
735 	if (rc)
736 		goto err;
737 
738 	/* If the configs are readonly, then load them from device */
739 	if (!test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags)) {
740 		dev_dbg(dev, "Loading RO device config\n");
741 		rc = idxd_device_load_config(idxd);
742 		if (rc < 0)
743 			goto err_config;
744 	}
745 
746 	rc = idxd_setup_interrupts(idxd);
747 	if (rc)
748 		goto err_config;
749 
750 	idxd->major = idxd_cdev_get_major(idxd);
751 
752 	rc = perfmon_pmu_init(idxd);
753 	if (rc < 0)
754 		dev_warn(dev, "Failed to initialize perfmon. No PMU support: %d\n", rc);
755 
756 	dev_dbg(dev, "IDXD device %d probed successfully\n", idxd->id);
757 	return 0;
758 
759  err_config:
760 	idxd_cleanup_internals(idxd);
761  err:
762 	if (device_pasid_enabled(idxd))
763 		idxd_disable_system_pasid(idxd);
764 	return rc;
765 }
766 
idxd_cleanup(struct idxd_device * idxd)767 static void idxd_cleanup(struct idxd_device *idxd)
768 {
769 	perfmon_pmu_remove(idxd);
770 	idxd_cleanup_interrupts(idxd);
771 	idxd_cleanup_internals(idxd);
772 	if (device_pasid_enabled(idxd))
773 		idxd_disable_system_pasid(idxd);
774 }
775 
776 /*
777  * Attach IDXD device to IDXD driver.
778  */
idxd_bind(struct device_driver * drv,const char * buf)779 static int idxd_bind(struct device_driver *drv, const char *buf)
780 {
781 	const struct bus_type *bus = drv->bus;
782 	struct device *dev;
783 	int err = -ENODEV;
784 
785 	dev = bus_find_device_by_name(bus, NULL, buf);
786 	if (dev)
787 		err = device_driver_attach(drv, dev);
788 
789 	put_device(dev);
790 
791 	return err;
792 }
793 
794 /*
795  * Detach IDXD device from driver.
796  */
idxd_unbind(struct device_driver * drv,const char * buf)797 static void idxd_unbind(struct device_driver *drv, const char *buf)
798 {
799 	const struct bus_type *bus = drv->bus;
800 	struct device *dev;
801 
802 	dev = bus_find_device_by_name(bus, NULL, buf);
803 	if (dev && dev->driver == drv)
804 		device_release_driver(dev);
805 
806 	put_device(dev);
807 }
808 
809 #define idxd_free_saved_configs(saved_configs, count)	\
810 	do {						\
811 		int i;					\
812 							\
813 		for (i = 0; i < (count); i++)		\
814 			kfree(saved_configs[i]);	\
815 	} while (0)
816 
idxd_free_saved(struct idxd_group ** saved_groups,struct idxd_engine ** saved_engines,struct idxd_wq ** saved_wqs,struct idxd_device * idxd)817 static void idxd_free_saved(struct idxd_group **saved_groups,
818 			    struct idxd_engine **saved_engines,
819 			    struct idxd_wq **saved_wqs,
820 			    struct idxd_device *idxd)
821 {
822 	if (saved_groups)
823 		idxd_free_saved_configs(saved_groups, idxd->max_groups);
824 	if (saved_engines)
825 		idxd_free_saved_configs(saved_engines, idxd->max_engines);
826 	if (saved_wqs)
827 		idxd_free_saved_configs(saved_wqs, idxd->max_wqs);
828 }
829 
830 /*
831  * Save IDXD device configurations including engines, groups, wqs etc.
832  * The saved configurations can be restored when needed.
833  */
idxd_device_config_save(struct idxd_device * idxd,struct idxd_saved_states * idxd_saved)834 static int idxd_device_config_save(struct idxd_device *idxd,
835 				   struct idxd_saved_states *idxd_saved)
836 {
837 	struct device *dev = &idxd->pdev->dev;
838 	int i;
839 
840 	memcpy(&idxd_saved->saved_idxd, idxd, sizeof(*idxd));
841 
842 	if (idxd->evl) {
843 		memcpy(&idxd_saved->saved_evl, idxd->evl,
844 		       sizeof(struct idxd_evl));
845 	}
846 
847 	struct idxd_group **saved_groups __free(kfree) =
848 			kcalloc_node(idxd->max_groups,
849 				     sizeof(struct idxd_group *),
850 				     GFP_KERNEL, dev_to_node(dev));
851 	if (!saved_groups)
852 		return -ENOMEM;
853 
854 	for (i = 0; i < idxd->max_groups; i++) {
855 		struct idxd_group *saved_group __free(kfree) =
856 			kzalloc_node(sizeof(*saved_group), GFP_KERNEL,
857 				     dev_to_node(dev));
858 
859 		if (!saved_group) {
860 			/* Free saved groups */
861 			idxd_free_saved(saved_groups, NULL, NULL, idxd);
862 
863 			return -ENOMEM;
864 		}
865 
866 		memcpy(saved_group, idxd->groups[i], sizeof(*saved_group));
867 		saved_groups[i] = no_free_ptr(saved_group);
868 	}
869 
870 	struct idxd_engine **saved_engines =
871 			kcalloc_node(idxd->max_engines,
872 				     sizeof(struct idxd_engine *),
873 				     GFP_KERNEL, dev_to_node(dev));
874 	if (!saved_engines) {
875 		/* Free saved groups */
876 		idxd_free_saved(saved_groups, NULL, NULL, idxd);
877 
878 		return -ENOMEM;
879 	}
880 	for (i = 0; i < idxd->max_engines; i++) {
881 		struct idxd_engine *saved_engine __free(kfree) =
882 				kzalloc_node(sizeof(*saved_engine), GFP_KERNEL,
883 					     dev_to_node(dev));
884 		if (!saved_engine) {
885 			/* Free saved groups and engines */
886 			idxd_free_saved(saved_groups, saved_engines, NULL,
887 					idxd);
888 
889 			return -ENOMEM;
890 		}
891 
892 		memcpy(saved_engine, idxd->engines[i], sizeof(*saved_engine));
893 		saved_engines[i] = no_free_ptr(saved_engine);
894 	}
895 
896 	unsigned long *saved_wq_enable_map __free(bitmap) =
897 			bitmap_zalloc_node(idxd->max_wqs, GFP_KERNEL,
898 					   dev_to_node(dev));
899 	if (!saved_wq_enable_map) {
900 		/* Free saved groups and engines */
901 		idxd_free_saved(saved_groups, saved_engines, NULL, idxd);
902 
903 		return -ENOMEM;
904 	}
905 
906 	bitmap_copy(saved_wq_enable_map, idxd->wq_enable_map, idxd->max_wqs);
907 
908 	struct idxd_wq **saved_wqs __free(kfree) =
909 			kcalloc_node(idxd->max_wqs, sizeof(struct idxd_wq *),
910 				     GFP_KERNEL, dev_to_node(dev));
911 	if (!saved_wqs) {
912 		/* Free saved groups and engines */
913 		idxd_free_saved(saved_groups, saved_engines, NULL, idxd);
914 
915 		return -ENOMEM;
916 	}
917 
918 	for (i = 0; i < idxd->max_wqs; i++) {
919 		struct idxd_wq *saved_wq __free(kfree) =
920 			kzalloc_node(sizeof(*saved_wq), GFP_KERNEL,
921 				     dev_to_node(dev));
922 		struct idxd_wq *wq;
923 
924 		if (!saved_wq) {
925 			/* Free saved groups, engines, and wqs */
926 			idxd_free_saved(saved_groups, saved_engines, saved_wqs,
927 					idxd);
928 
929 			return -ENOMEM;
930 		}
931 
932 		if (!test_bit(i, saved_wq_enable_map))
933 			continue;
934 
935 		wq = idxd->wqs[i];
936 		mutex_lock(&wq->wq_lock);
937 		memcpy(saved_wq, wq, sizeof(*saved_wq));
938 		saved_wqs[i] = no_free_ptr(saved_wq);
939 		mutex_unlock(&wq->wq_lock);
940 	}
941 
942 	/* Save configurations */
943 	idxd_saved->saved_groups = no_free_ptr(saved_groups);
944 	idxd_saved->saved_engines = no_free_ptr(saved_engines);
945 	idxd_saved->saved_wq_enable_map = no_free_ptr(saved_wq_enable_map);
946 	idxd_saved->saved_wqs = no_free_ptr(saved_wqs);
947 
948 	return 0;
949 }
950 
951 /*
952  * Restore IDXD device configurations including engines, groups, wqs etc
953  * that were saved before.
954  */
idxd_device_config_restore(struct idxd_device * idxd,struct idxd_saved_states * idxd_saved)955 static void idxd_device_config_restore(struct idxd_device *idxd,
956 				       struct idxd_saved_states *idxd_saved)
957 {
958 	struct idxd_evl *saved_evl = &idxd_saved->saved_evl;
959 	int i;
960 
961 	idxd->rdbuf_limit = idxd_saved->saved_idxd.rdbuf_limit;
962 
963 	idxd->evl->size = saved_evl->size;
964 
965 	for (i = 0; i < idxd->max_groups; i++) {
966 		struct idxd_group *saved_group, *group;
967 
968 		saved_group = idxd_saved->saved_groups[i];
969 		group = idxd->groups[i];
970 
971 		group->rdbufs_allowed = saved_group->rdbufs_allowed;
972 		group->rdbufs_reserved = saved_group->rdbufs_reserved;
973 		group->tc_a = saved_group->tc_a;
974 		group->tc_b = saved_group->tc_b;
975 		group->use_rdbuf_limit = saved_group->use_rdbuf_limit;
976 
977 		kfree(saved_group);
978 	}
979 	kfree(idxd_saved->saved_groups);
980 
981 	for (i = 0; i < idxd->max_engines; i++) {
982 		struct idxd_engine *saved_engine, *engine;
983 
984 		saved_engine = idxd_saved->saved_engines[i];
985 		engine = idxd->engines[i];
986 
987 		engine->group = saved_engine->group;
988 
989 		kfree(saved_engine);
990 	}
991 	kfree(idxd_saved->saved_engines);
992 
993 	bitmap_copy(idxd->wq_enable_map, idxd_saved->saved_wq_enable_map,
994 		    idxd->max_wqs);
995 	bitmap_free(idxd_saved->saved_wq_enable_map);
996 
997 	for (i = 0; i < idxd->max_wqs; i++) {
998 		struct idxd_wq *saved_wq, *wq;
999 		size_t len;
1000 
1001 		if (!test_bit(i, idxd->wq_enable_map))
1002 			continue;
1003 
1004 		saved_wq = idxd_saved->saved_wqs[i];
1005 		wq = idxd->wqs[i];
1006 
1007 		mutex_lock(&wq->wq_lock);
1008 
1009 		wq->group = saved_wq->group;
1010 		wq->flags = saved_wq->flags;
1011 		wq->threshold = saved_wq->threshold;
1012 		wq->size = saved_wq->size;
1013 		wq->priority = saved_wq->priority;
1014 		wq->type = saved_wq->type;
1015 		len = strlen(saved_wq->name) + 1;
1016 		strscpy(wq->name, saved_wq->name, len);
1017 		wq->max_xfer_bytes = saved_wq->max_xfer_bytes;
1018 		wq->max_batch_size = saved_wq->max_batch_size;
1019 		wq->enqcmds_retries = saved_wq->enqcmds_retries;
1020 		wq->descs = saved_wq->descs;
1021 		wq->idxd_chan = saved_wq->idxd_chan;
1022 		len = strlen(saved_wq->driver_name) + 1;
1023 		strscpy(wq->driver_name, saved_wq->driver_name, len);
1024 
1025 		mutex_unlock(&wq->wq_lock);
1026 
1027 		kfree(saved_wq);
1028 	}
1029 
1030 	kfree(idxd_saved->saved_wqs);
1031 }
1032 
idxd_reset_prepare(struct pci_dev * pdev)1033 static void idxd_reset_prepare(struct pci_dev *pdev)
1034 {
1035 	struct idxd_device *idxd = pci_get_drvdata(pdev);
1036 	struct device *dev = &idxd->pdev->dev;
1037 	const char *idxd_name;
1038 	int rc;
1039 
1040 	idxd_name = dev_name(idxd_confdev(idxd));
1041 
1042 	struct idxd_saved_states *idxd_saved __free(kfree) =
1043 			kzalloc_node(sizeof(*idxd_saved), GFP_KERNEL,
1044 				     dev_to_node(&pdev->dev));
1045 	if (!idxd_saved) {
1046 		dev_err(dev, "HALT: no memory\n");
1047 
1048 		return;
1049 	}
1050 
1051 	/* Save IDXD configurations. */
1052 	rc = idxd_device_config_save(idxd, idxd_saved);
1053 	if (rc < 0) {
1054 		dev_err(dev, "HALT: cannot save %s configs\n", idxd_name);
1055 
1056 		return;
1057 	}
1058 
1059 	idxd->idxd_saved = no_free_ptr(idxd_saved);
1060 
1061 	/* Save PCI device state. */
1062 	pci_save_state(idxd->pdev);
1063 }
1064 
idxd_reset_done(struct pci_dev * pdev)1065 static void idxd_reset_done(struct pci_dev *pdev)
1066 {
1067 	struct idxd_device *idxd = pci_get_drvdata(pdev);
1068 	const char *idxd_name;
1069 	struct device *dev;
1070 	int rc, i;
1071 
1072 	if (!idxd->idxd_saved)
1073 		return;
1074 
1075 	dev = &idxd->pdev->dev;
1076 	idxd_name = dev_name(idxd_confdev(idxd));
1077 
1078 	/* Restore PCI device state. */
1079 	pci_restore_state(idxd->pdev);
1080 
1081 	/* Unbind idxd device from driver. */
1082 	idxd_unbind(&idxd_drv.drv, idxd_name);
1083 
1084 	/*
1085 	 * Probe PCI device without allocating or changing
1086 	 * idxd software data which keeps the same as before FLR.
1087 	 */
1088 	idxd_pci_probe_alloc(idxd, NULL, NULL);
1089 
1090 	/* Restore IDXD configurations. */
1091 	idxd_device_config_restore(idxd, idxd->idxd_saved);
1092 
1093 	/* Re-configure IDXD device if allowed. */
1094 	if (test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags)) {
1095 		rc = idxd_device_config(idxd);
1096 		if (rc < 0) {
1097 			dev_err(dev, "HALT: %s config fails\n", idxd_name);
1098 			goto out;
1099 		}
1100 	}
1101 
1102 	/* Bind IDXD device to driver. */
1103 	rc = idxd_bind(&idxd_drv.drv, idxd_name);
1104 	if (rc < 0) {
1105 		dev_err(dev, "HALT: binding %s to driver fails\n", idxd_name);
1106 		goto out;
1107 	}
1108 
1109 	/* Bind enabled wq in the IDXD device to driver. */
1110 	for (i = 0; i < idxd->max_wqs; i++) {
1111 		if (test_bit(i, idxd->wq_enable_map)) {
1112 			struct idxd_wq *wq = idxd->wqs[i];
1113 			char wq_name[32];
1114 
1115 			wq->state = IDXD_WQ_DISABLED;
1116 			sprintf(wq_name, "wq%d.%d", idxd->id, wq->id);
1117 			/*
1118 			 * Bind to user driver depending on wq type.
1119 			 *
1120 			 * Currently only support user type WQ. Will support
1121 			 * kernel type WQ in the future.
1122 			 */
1123 			if (wq->type == IDXD_WQT_USER)
1124 				rc = idxd_bind(&idxd_user_drv.drv, wq_name);
1125 			else
1126 				rc = -EINVAL;
1127 			if (rc < 0) {
1128 				clear_bit(i, idxd->wq_enable_map);
1129 				dev_err(dev,
1130 					"HALT: unable to re-enable wq %s\n",
1131 					dev_name(wq_confdev(wq)));
1132 			}
1133 		}
1134 	}
1135 out:
1136 	kfree(idxd->idxd_saved);
1137 }
1138 
1139 static const struct pci_error_handlers idxd_error_handler = {
1140 	.reset_prepare	= idxd_reset_prepare,
1141 	.reset_done	= idxd_reset_done,
1142 };
1143 
1144 /*
1145  * Probe idxd PCI device.
1146  * If idxd is not given, need to allocate idxd and set up its data.
1147  *
1148  * If idxd is given, idxd was allocated and setup already. Just need to
1149  * configure device without re-allocating and re-configuring idxd data.
1150  * This is useful for recovering from FLR.
1151  */
idxd_pci_probe_alloc(struct idxd_device * idxd,struct pci_dev * pdev,const struct pci_device_id * id)1152 int idxd_pci_probe_alloc(struct idxd_device *idxd, struct pci_dev *pdev,
1153 			 const struct pci_device_id *id)
1154 {
1155 	bool alloc_idxd = idxd ? false : true;
1156 	struct idxd_driver_data *data;
1157 	struct device *dev;
1158 	int rc;
1159 
1160 	pdev = idxd ? idxd->pdev : pdev;
1161 	dev = &pdev->dev;
1162 	data = id ? (struct idxd_driver_data *)id->driver_data : NULL;
1163 	rc = pci_enable_device(pdev);
1164 	if (rc)
1165 		return rc;
1166 
1167 	if (alloc_idxd) {
1168 		dev_dbg(dev, "Alloc IDXD context\n");
1169 		idxd = idxd_alloc(pdev, data);
1170 		if (!idxd) {
1171 			rc = -ENOMEM;
1172 			goto err_idxd_alloc;
1173 		}
1174 
1175 		dev_dbg(dev, "Mapping BARs\n");
1176 		idxd->reg_base = pci_iomap(pdev, IDXD_MMIO_BAR, 0);
1177 		if (!idxd->reg_base) {
1178 			rc = -ENOMEM;
1179 			goto err_iomap;
1180 		}
1181 
1182 		dev_dbg(dev, "Set DMA masks\n");
1183 		rc = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
1184 		if (rc)
1185 			goto err;
1186 	}
1187 
1188 	dev_dbg(dev, "Set PCI master\n");
1189 	pci_set_master(pdev);
1190 	pci_set_drvdata(pdev, idxd);
1191 
1192 	if (alloc_idxd) {
1193 		idxd->hw.version = ioread32(idxd->reg_base + IDXD_VER_OFFSET);
1194 		rc = idxd_probe(idxd);
1195 		if (rc) {
1196 			dev_err(dev, "Intel(R) IDXD DMA Engine init failed\n");
1197 			goto err;
1198 		}
1199 
1200 		if (data->load_device_defaults) {
1201 			rc = data->load_device_defaults(idxd);
1202 			if (rc)
1203 				dev_warn(dev, "IDXD loading device defaults failed\n");
1204 		}
1205 
1206 		rc = idxd_register_devices(idxd);
1207 		if (rc) {
1208 			dev_err(dev, "IDXD sysfs setup failed\n");
1209 			goto err_dev_register;
1210 		}
1211 
1212 		rc = idxd_device_init_debugfs(idxd);
1213 		if (rc)
1214 			dev_warn(dev, "IDXD debugfs failed to setup\n");
1215 	}
1216 
1217 	if (!alloc_idxd) {
1218 		/* Release interrupts in the IDXD device. */
1219 		idxd_cleanup_interrupts(idxd);
1220 
1221 		/* Re-enable interrupts in the IDXD device. */
1222 		rc = idxd_setup_interrupts(idxd);
1223 		if (rc)
1224 			dev_warn(dev, "IDXD interrupts failed to setup\n");
1225 	}
1226 
1227 	dev_info(&pdev->dev, "Intel(R) Accelerator Device (v%x)\n",
1228 		 idxd->hw.version);
1229 
1230 	if (data)
1231 		idxd->user_submission_safe = data->user_submission_safe;
1232 
1233 	return 0;
1234 
1235  err_dev_register:
1236 	idxd_cleanup(idxd);
1237  err:
1238 	pci_iounmap(pdev, idxd->reg_base);
1239  err_iomap:
1240 	idxd_free(idxd);
1241  err_idxd_alloc:
1242 	pci_disable_device(pdev);
1243 	return rc;
1244 }
1245 
idxd_pci_probe(struct pci_dev * pdev,const struct pci_device_id * id)1246 static int idxd_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1247 {
1248 	return idxd_pci_probe_alloc(NULL, pdev, id);
1249 }
1250 
idxd_wqs_quiesce(struct idxd_device * idxd)1251 void idxd_wqs_quiesce(struct idxd_device *idxd)
1252 {
1253 	struct idxd_wq *wq;
1254 	int i;
1255 
1256 	for (i = 0; i < idxd->max_wqs; i++) {
1257 		wq = idxd->wqs[i];
1258 		if (wq->state == IDXD_WQ_ENABLED && wq->type == IDXD_WQT_KERNEL)
1259 			idxd_wq_quiesce(wq);
1260 	}
1261 }
1262 
idxd_shutdown(struct pci_dev * pdev)1263 static void idxd_shutdown(struct pci_dev *pdev)
1264 {
1265 	struct idxd_device *idxd = pci_get_drvdata(pdev);
1266 	struct idxd_irq_entry *irq_entry;
1267 	int rc;
1268 
1269 	rc = idxd_device_disable(idxd);
1270 	if (rc)
1271 		dev_err(&pdev->dev, "Disabling device failed\n");
1272 
1273 	irq_entry = &idxd->ie;
1274 	synchronize_irq(irq_entry->vector);
1275 	idxd_mask_error_interrupts(idxd);
1276 	flush_workqueue(idxd->wq);
1277 }
1278 
idxd_remove(struct pci_dev * pdev)1279 static void idxd_remove(struct pci_dev *pdev)
1280 {
1281 	struct idxd_device *idxd = pci_get_drvdata(pdev);
1282 
1283 	idxd_unregister_devices(idxd);
1284 	/*
1285 	 * When ->release() is called for the idxd->conf_dev, it frees all the memory related
1286 	 * to the idxd context. The driver still needs those bits in order to do the rest of
1287 	 * the cleanup. However, we do need to unbound the idxd sub-driver. So take a ref
1288 	 * on the device here to hold off the freeing while allowing the idxd sub-driver
1289 	 * to unbind.
1290 	 */
1291 	get_device(idxd_confdev(idxd));
1292 	device_unregister(idxd_confdev(idxd));
1293 	idxd_shutdown(pdev);
1294 	idxd_device_remove_debugfs(idxd);
1295 	perfmon_pmu_remove(idxd);
1296 	idxd_cleanup_interrupts(idxd);
1297 	if (device_pasid_enabled(idxd))
1298 		idxd_disable_system_pasid(idxd);
1299 	pci_iounmap(pdev, idxd->reg_base);
1300 	put_device(idxd_confdev(idxd));
1301 	pci_disable_device(pdev);
1302 }
1303 
1304 static struct pci_driver idxd_pci_driver = {
1305 	.name		= DRV_NAME,
1306 	.id_table	= idxd_pci_tbl,
1307 	.probe		= idxd_pci_probe,
1308 	.remove		= idxd_remove,
1309 	.shutdown	= idxd_shutdown,
1310 	.err_handler	= &idxd_error_handler,
1311 };
1312 
idxd_init_module(void)1313 static int __init idxd_init_module(void)
1314 {
1315 	int err;
1316 
1317 	/*
1318 	 * If the CPU does not support MOVDIR64B or ENQCMDS, there's no point in
1319 	 * enumerating the device. We can not utilize it.
1320 	 */
1321 	if (!cpu_feature_enabled(X86_FEATURE_MOVDIR64B)) {
1322 		pr_warn("idxd driver failed to load without MOVDIR64B.\n");
1323 		return -ENODEV;
1324 	}
1325 
1326 	if (!cpu_feature_enabled(X86_FEATURE_ENQCMD))
1327 		pr_warn("Platform does not have ENQCMD(S) support.\n");
1328 	else
1329 		support_enqcmd = true;
1330 
1331 	err = idxd_driver_register(&idxd_drv);
1332 	if (err < 0)
1333 		goto err_idxd_driver_register;
1334 
1335 	err = idxd_driver_register(&idxd_dmaengine_drv);
1336 	if (err < 0)
1337 		goto err_idxd_dmaengine_driver_register;
1338 
1339 	err = idxd_driver_register(&idxd_user_drv);
1340 	if (err < 0)
1341 		goto err_idxd_user_driver_register;
1342 
1343 	err = idxd_cdev_register();
1344 	if (err)
1345 		goto err_cdev_register;
1346 
1347 	err = idxd_init_debugfs();
1348 	if (err)
1349 		goto err_debugfs;
1350 
1351 	err = pci_register_driver(&idxd_pci_driver);
1352 	if (err)
1353 		goto err_pci_register;
1354 
1355 	return 0;
1356 
1357 err_pci_register:
1358 	idxd_remove_debugfs();
1359 err_debugfs:
1360 	idxd_cdev_remove();
1361 err_cdev_register:
1362 	idxd_driver_unregister(&idxd_user_drv);
1363 err_idxd_user_driver_register:
1364 	idxd_driver_unregister(&idxd_dmaengine_drv);
1365 err_idxd_dmaengine_driver_register:
1366 	idxd_driver_unregister(&idxd_drv);
1367 err_idxd_driver_register:
1368 	return err;
1369 }
1370 module_init(idxd_init_module);
1371 
idxd_exit_module(void)1372 static void __exit idxd_exit_module(void)
1373 {
1374 	idxd_driver_unregister(&idxd_user_drv);
1375 	idxd_driver_unregister(&idxd_dmaengine_drv);
1376 	idxd_driver_unregister(&idxd_drv);
1377 	pci_unregister_driver(&idxd_pci_driver);
1378 	idxd_cdev_remove();
1379 	idxd_remove_debugfs();
1380 }
1381 module_exit(idxd_exit_module);
1382