xref: /linux/drivers/iommu/amd/iommu.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
4  * Author: Joerg Roedel <jroedel@suse.de>
5  *         Leo Duran <leo.duran@amd.com>
6  */
7 
8 #define pr_fmt(fmt)     "AMD-Vi: " fmt
9 #define dev_fmt(fmt)    pr_fmt(fmt)
10 
11 #include <linux/ratelimit.h>
12 #include <linux/pci.h>
13 #include <linux/acpi.h>
14 #include <linux/pci-ats.h>
15 #include <linux/bitmap.h>
16 #include <linux/slab.h>
17 #include <linux/debugfs.h>
18 #include <linux/scatterlist.h>
19 #include <linux/dma-map-ops.h>
20 #include <linux/dma-direct.h>
21 #include <linux/iommu-helper.h>
22 #include <linux/delay.h>
23 #include <linux/amd-iommu.h>
24 #include <linux/notifier.h>
25 #include <linux/export.h>
26 #include <linux/irq.h>
27 #include <linux/msi.h>
28 #include <linux/irqdomain.h>
29 #include <linux/percpu.h>
30 #include <linux/io-pgtable.h>
31 #include <linux/cc_platform.h>
32 #include <asm/irq_remapping.h>
33 #include <asm/io_apic.h>
34 #include <asm/apic.h>
35 #include <asm/hw_irq.h>
36 #include <asm/proto.h>
37 #include <asm/iommu.h>
38 #include <asm/gart.h>
39 #include <asm/dma.h>
40 #include <uapi/linux/iommufd.h>
41 
42 #include "amd_iommu.h"
43 #include "../dma-iommu.h"
44 #include "../irq_remapping.h"
45 #include "../iommu-pages.h"
46 
47 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
48 
49 /* Reserved IOVA ranges */
50 #define MSI_RANGE_START		(0xfee00000)
51 #define MSI_RANGE_END		(0xfeefffff)
52 #define HT_RANGE_START		(0xfd00000000ULL)
53 #define HT_RANGE_END		(0xffffffffffULL)
54 
55 static DEFINE_SPINLOCK(pd_bitmap_lock);
56 
57 LIST_HEAD(ioapic_map);
58 LIST_HEAD(hpet_map);
59 LIST_HEAD(acpihid_map);
60 
61 const struct iommu_ops amd_iommu_ops;
62 static const struct iommu_dirty_ops amd_dirty_ops;
63 
64 int amd_iommu_max_glx_val = -1;
65 
66 /*
67  * general struct to manage commands send to an IOMMU
68  */
69 struct iommu_cmd {
70 	u32 data[4];
71 };
72 
73 struct kmem_cache *amd_iommu_irq_cache;
74 
75 static void detach_device(struct device *dev);
76 
77 static void set_dte_entry(struct amd_iommu *iommu,
78 			  struct iommu_dev_data *dev_data);
79 
80 /****************************************************************************
81  *
82  * Helper functions
83  *
84  ****************************************************************************/
85 
86 static inline bool pdom_is_v2_pgtbl_mode(struct protection_domain *pdom)
87 {
88 	return (pdom && (pdom->pd_mode == PD_MODE_V2));
89 }
90 
91 static inline bool pdom_is_in_pt_mode(struct protection_domain *pdom)
92 {
93 	return (pdom->domain.type == IOMMU_DOMAIN_IDENTITY);
94 }
95 
96 /*
97  * We cannot support PASID w/ existing v1 page table in the same domain
98  * since it will be nested. However, existing domain w/ v2 page table
99  * or passthrough mode can be used for PASID.
100  */
101 static inline bool pdom_is_sva_capable(struct protection_domain *pdom)
102 {
103 	return pdom_is_v2_pgtbl_mode(pdom) || pdom_is_in_pt_mode(pdom);
104 }
105 
106 static inline int get_acpihid_device_id(struct device *dev,
107 					struct acpihid_map_entry **entry)
108 {
109 	struct acpi_device *adev = ACPI_COMPANION(dev);
110 	struct acpihid_map_entry *p;
111 
112 	if (!adev)
113 		return -ENODEV;
114 
115 	list_for_each_entry(p, &acpihid_map, list) {
116 		if (acpi_dev_hid_uid_match(adev, p->hid,
117 					   p->uid[0] ? p->uid : NULL)) {
118 			if (entry)
119 				*entry = p;
120 			return p->devid;
121 		}
122 	}
123 	return -EINVAL;
124 }
125 
126 static inline int get_device_sbdf_id(struct device *dev)
127 {
128 	int sbdf;
129 
130 	if (dev_is_pci(dev))
131 		sbdf = get_pci_sbdf_id(to_pci_dev(dev));
132 	else
133 		sbdf = get_acpihid_device_id(dev, NULL);
134 
135 	return sbdf;
136 }
137 
138 struct dev_table_entry *get_dev_table(struct amd_iommu *iommu)
139 {
140 	struct dev_table_entry *dev_table;
141 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
142 
143 	BUG_ON(pci_seg == NULL);
144 	dev_table = pci_seg->dev_table;
145 	BUG_ON(dev_table == NULL);
146 
147 	return dev_table;
148 }
149 
150 static inline u16 get_device_segment(struct device *dev)
151 {
152 	u16 seg;
153 
154 	if (dev_is_pci(dev)) {
155 		struct pci_dev *pdev = to_pci_dev(dev);
156 
157 		seg = pci_domain_nr(pdev->bus);
158 	} else {
159 		u32 devid = get_acpihid_device_id(dev, NULL);
160 
161 		seg = PCI_SBDF_TO_SEGID(devid);
162 	}
163 
164 	return seg;
165 }
166 
167 /* Writes the specific IOMMU for a device into the PCI segment rlookup table */
168 void amd_iommu_set_rlookup_table(struct amd_iommu *iommu, u16 devid)
169 {
170 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
171 
172 	pci_seg->rlookup_table[devid] = iommu;
173 }
174 
175 static struct amd_iommu *__rlookup_amd_iommu(u16 seg, u16 devid)
176 {
177 	struct amd_iommu_pci_seg *pci_seg;
178 
179 	for_each_pci_segment(pci_seg) {
180 		if (pci_seg->id == seg)
181 			return pci_seg->rlookup_table[devid];
182 	}
183 	return NULL;
184 }
185 
186 static struct amd_iommu *rlookup_amd_iommu(struct device *dev)
187 {
188 	u16 seg = get_device_segment(dev);
189 	int devid = get_device_sbdf_id(dev);
190 
191 	if (devid < 0)
192 		return NULL;
193 	return __rlookup_amd_iommu(seg, PCI_SBDF_TO_DEVID(devid));
194 }
195 
196 static struct iommu_dev_data *alloc_dev_data(struct amd_iommu *iommu, u16 devid)
197 {
198 	struct iommu_dev_data *dev_data;
199 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
200 
201 	dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
202 	if (!dev_data)
203 		return NULL;
204 
205 	spin_lock_init(&dev_data->lock);
206 	dev_data->devid = devid;
207 	ratelimit_default_init(&dev_data->rs);
208 
209 	llist_add(&dev_data->dev_data_list, &pci_seg->dev_data_list);
210 	return dev_data;
211 }
212 
213 static struct iommu_dev_data *search_dev_data(struct amd_iommu *iommu, u16 devid)
214 {
215 	struct iommu_dev_data *dev_data;
216 	struct llist_node *node;
217 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
218 
219 	if (llist_empty(&pci_seg->dev_data_list))
220 		return NULL;
221 
222 	node = pci_seg->dev_data_list.first;
223 	llist_for_each_entry(dev_data, node, dev_data_list) {
224 		if (dev_data->devid == devid)
225 			return dev_data;
226 	}
227 
228 	return NULL;
229 }
230 
231 static int clone_alias(struct pci_dev *pdev, u16 alias, void *data)
232 {
233 	struct amd_iommu *iommu;
234 	struct dev_table_entry *dev_table;
235 	u16 devid = pci_dev_id(pdev);
236 
237 	if (devid == alias)
238 		return 0;
239 
240 	iommu = rlookup_amd_iommu(&pdev->dev);
241 	if (!iommu)
242 		return 0;
243 
244 	amd_iommu_set_rlookup_table(iommu, alias);
245 	dev_table = get_dev_table(iommu);
246 	memcpy(dev_table[alias].data,
247 	       dev_table[devid].data,
248 	       sizeof(dev_table[alias].data));
249 
250 	return 0;
251 }
252 
253 static void clone_aliases(struct amd_iommu *iommu, struct device *dev)
254 {
255 	struct pci_dev *pdev;
256 
257 	if (!dev_is_pci(dev))
258 		return;
259 	pdev = to_pci_dev(dev);
260 
261 	/*
262 	 * The IVRS alias stored in the alias table may not be
263 	 * part of the PCI DMA aliases if it's bus differs
264 	 * from the original device.
265 	 */
266 	clone_alias(pdev, iommu->pci_seg->alias_table[pci_dev_id(pdev)], NULL);
267 
268 	pci_for_each_dma_alias(pdev, clone_alias, NULL);
269 }
270 
271 static void setup_aliases(struct amd_iommu *iommu, struct device *dev)
272 {
273 	struct pci_dev *pdev = to_pci_dev(dev);
274 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
275 	u16 ivrs_alias;
276 
277 	/* For ACPI HID devices, there are no aliases */
278 	if (!dev_is_pci(dev))
279 		return;
280 
281 	/*
282 	 * Add the IVRS alias to the pci aliases if it is on the same
283 	 * bus. The IVRS table may know about a quirk that we don't.
284 	 */
285 	ivrs_alias = pci_seg->alias_table[pci_dev_id(pdev)];
286 	if (ivrs_alias != pci_dev_id(pdev) &&
287 	    PCI_BUS_NUM(ivrs_alias) == pdev->bus->number)
288 		pci_add_dma_alias(pdev, ivrs_alias & 0xff, 1);
289 
290 	clone_aliases(iommu, dev);
291 }
292 
293 static struct iommu_dev_data *find_dev_data(struct amd_iommu *iommu, u16 devid)
294 {
295 	struct iommu_dev_data *dev_data;
296 
297 	dev_data = search_dev_data(iommu, devid);
298 
299 	if (dev_data == NULL) {
300 		dev_data = alloc_dev_data(iommu, devid);
301 		if (!dev_data)
302 			return NULL;
303 
304 		if (translation_pre_enabled(iommu))
305 			dev_data->defer_attach = true;
306 	}
307 
308 	return dev_data;
309 }
310 
311 /*
312 * Find or create an IOMMU group for a acpihid device.
313 */
314 static struct iommu_group *acpihid_device_group(struct device *dev)
315 {
316 	struct acpihid_map_entry *p, *entry = NULL;
317 	int devid;
318 
319 	devid = get_acpihid_device_id(dev, &entry);
320 	if (devid < 0)
321 		return ERR_PTR(devid);
322 
323 	list_for_each_entry(p, &acpihid_map, list) {
324 		if ((devid == p->devid) && p->group)
325 			entry->group = p->group;
326 	}
327 
328 	if (!entry->group)
329 		entry->group = generic_device_group(dev);
330 	else
331 		iommu_group_ref_get(entry->group);
332 
333 	return entry->group;
334 }
335 
336 static inline bool pdev_pasid_supported(struct iommu_dev_data *dev_data)
337 {
338 	return (dev_data->flags & AMD_IOMMU_DEVICE_FLAG_PASID_SUP);
339 }
340 
341 static u32 pdev_get_caps(struct pci_dev *pdev)
342 {
343 	int features;
344 	u32 flags = 0;
345 
346 	if (pci_ats_supported(pdev))
347 		flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
348 
349 	if (pci_pri_supported(pdev))
350 		flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
351 
352 	features = pci_pasid_features(pdev);
353 	if (features >= 0) {
354 		flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
355 
356 		if (features & PCI_PASID_CAP_EXEC)
357 			flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
358 
359 		if (features & PCI_PASID_CAP_PRIV)
360 			flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
361 	}
362 
363 	return flags;
364 }
365 
366 static inline int pdev_enable_cap_ats(struct pci_dev *pdev)
367 {
368 	struct iommu_dev_data *dev_data = dev_iommu_priv_get(&pdev->dev);
369 	int ret = -EINVAL;
370 
371 	if (dev_data->ats_enabled)
372 		return 0;
373 
374 	if (amd_iommu_iotlb_sup &&
375 	    (dev_data->flags & AMD_IOMMU_DEVICE_FLAG_ATS_SUP)) {
376 		ret = pci_enable_ats(pdev, PAGE_SHIFT);
377 		if (!ret) {
378 			dev_data->ats_enabled = 1;
379 			dev_data->ats_qdep    = pci_ats_queue_depth(pdev);
380 		}
381 	}
382 
383 	return ret;
384 }
385 
386 static inline void pdev_disable_cap_ats(struct pci_dev *pdev)
387 {
388 	struct iommu_dev_data *dev_data = dev_iommu_priv_get(&pdev->dev);
389 
390 	if (dev_data->ats_enabled) {
391 		pci_disable_ats(pdev);
392 		dev_data->ats_enabled = 0;
393 	}
394 }
395 
396 static inline int pdev_enable_cap_pri(struct pci_dev *pdev)
397 {
398 	struct iommu_dev_data *dev_data = dev_iommu_priv_get(&pdev->dev);
399 	int ret = -EINVAL;
400 
401 	if (dev_data->pri_enabled)
402 		return 0;
403 
404 	if (!dev_data->ats_enabled)
405 		return 0;
406 
407 	if (dev_data->flags & AMD_IOMMU_DEVICE_FLAG_PRI_SUP) {
408 		/*
409 		 * First reset the PRI state of the device.
410 		 * FIXME: Hardcode number of outstanding requests for now
411 		 */
412 		if (!pci_reset_pri(pdev) && !pci_enable_pri(pdev, 32)) {
413 			dev_data->pri_enabled = 1;
414 			dev_data->pri_tlp     = pci_prg_resp_pasid_required(pdev);
415 
416 			ret = 0;
417 		}
418 	}
419 
420 	return ret;
421 }
422 
423 static inline void pdev_disable_cap_pri(struct pci_dev *pdev)
424 {
425 	struct iommu_dev_data *dev_data = dev_iommu_priv_get(&pdev->dev);
426 
427 	if (dev_data->pri_enabled) {
428 		pci_disable_pri(pdev);
429 		dev_data->pri_enabled = 0;
430 	}
431 }
432 
433 static inline int pdev_enable_cap_pasid(struct pci_dev *pdev)
434 {
435 	struct iommu_dev_data *dev_data = dev_iommu_priv_get(&pdev->dev);
436 	int ret = -EINVAL;
437 
438 	if (dev_data->pasid_enabled)
439 		return 0;
440 
441 	if (dev_data->flags & AMD_IOMMU_DEVICE_FLAG_PASID_SUP) {
442 		/* Only allow access to user-accessible pages */
443 		ret = pci_enable_pasid(pdev, 0);
444 		if (!ret)
445 			dev_data->pasid_enabled = 1;
446 	}
447 
448 	return ret;
449 }
450 
451 static inline void pdev_disable_cap_pasid(struct pci_dev *pdev)
452 {
453 	struct iommu_dev_data *dev_data = dev_iommu_priv_get(&pdev->dev);
454 
455 	if (dev_data->pasid_enabled) {
456 		pci_disable_pasid(pdev);
457 		dev_data->pasid_enabled = 0;
458 	}
459 }
460 
461 static void pdev_enable_caps(struct pci_dev *pdev)
462 {
463 	pdev_enable_cap_ats(pdev);
464 	pdev_enable_cap_pasid(pdev);
465 	pdev_enable_cap_pri(pdev);
466 }
467 
468 static void pdev_disable_caps(struct pci_dev *pdev)
469 {
470 	pdev_disable_cap_ats(pdev);
471 	pdev_disable_cap_pasid(pdev);
472 	pdev_disable_cap_pri(pdev);
473 }
474 
475 /*
476  * This function checks if the driver got a valid device from the caller to
477  * avoid dereferencing invalid pointers.
478  */
479 static bool check_device(struct device *dev)
480 {
481 	struct amd_iommu_pci_seg *pci_seg;
482 	struct amd_iommu *iommu;
483 	int devid, sbdf;
484 
485 	if (!dev)
486 		return false;
487 
488 	sbdf = get_device_sbdf_id(dev);
489 	if (sbdf < 0)
490 		return false;
491 	devid = PCI_SBDF_TO_DEVID(sbdf);
492 
493 	iommu = rlookup_amd_iommu(dev);
494 	if (!iommu)
495 		return false;
496 
497 	/* Out of our scope? */
498 	pci_seg = iommu->pci_seg;
499 	if (devid > pci_seg->last_bdf)
500 		return false;
501 
502 	return true;
503 }
504 
505 static int iommu_init_device(struct amd_iommu *iommu, struct device *dev)
506 {
507 	struct iommu_dev_data *dev_data;
508 	int devid, sbdf;
509 
510 	if (dev_iommu_priv_get(dev))
511 		return 0;
512 
513 	sbdf = get_device_sbdf_id(dev);
514 	if (sbdf < 0)
515 		return sbdf;
516 
517 	devid = PCI_SBDF_TO_DEVID(sbdf);
518 	dev_data = find_dev_data(iommu, devid);
519 	if (!dev_data)
520 		return -ENOMEM;
521 
522 	dev_data->dev = dev;
523 	setup_aliases(iommu, dev);
524 
525 	/*
526 	 * By default we use passthrough mode for IOMMUv2 capable device.
527 	 * But if amd_iommu=force_isolation is set (e.g. to debug DMA to
528 	 * invalid address), we ignore the capability for the device so
529 	 * it'll be forced to go into translation mode.
530 	 */
531 	if ((iommu_default_passthrough() || !amd_iommu_force_isolation) &&
532 	    dev_is_pci(dev) && amd_iommu_gt_ppr_supported()) {
533 		dev_data->flags = pdev_get_caps(to_pci_dev(dev));
534 	}
535 
536 	dev_iommu_priv_set(dev, dev_data);
537 
538 	return 0;
539 }
540 
541 static void iommu_ignore_device(struct amd_iommu *iommu, struct device *dev)
542 {
543 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
544 	struct dev_table_entry *dev_table = get_dev_table(iommu);
545 	int devid, sbdf;
546 
547 	sbdf = get_device_sbdf_id(dev);
548 	if (sbdf < 0)
549 		return;
550 
551 	devid = PCI_SBDF_TO_DEVID(sbdf);
552 	pci_seg->rlookup_table[devid] = NULL;
553 	memset(&dev_table[devid], 0, sizeof(struct dev_table_entry));
554 
555 	setup_aliases(iommu, dev);
556 }
557 
558 static void amd_iommu_uninit_device(struct device *dev)
559 {
560 	struct iommu_dev_data *dev_data;
561 
562 	dev_data = dev_iommu_priv_get(dev);
563 	if (!dev_data)
564 		return;
565 
566 	if (dev_data->domain)
567 		detach_device(dev);
568 
569 	/*
570 	 * We keep dev_data around for unplugged devices and reuse it when the
571 	 * device is re-plugged - not doing so would introduce a ton of races.
572 	 */
573 }
574 
575 /****************************************************************************
576  *
577  * Interrupt handling functions
578  *
579  ****************************************************************************/
580 
581 static void dump_dte_entry(struct amd_iommu *iommu, u16 devid)
582 {
583 	int i;
584 	struct dev_table_entry *dev_table = get_dev_table(iommu);
585 
586 	for (i = 0; i < 4; ++i)
587 		pr_err("DTE[%d]: %016llx\n", i, dev_table[devid].data[i]);
588 }
589 
590 static void dump_command(unsigned long phys_addr)
591 {
592 	struct iommu_cmd *cmd = iommu_phys_to_virt(phys_addr);
593 	int i;
594 
595 	for (i = 0; i < 4; ++i)
596 		pr_err("CMD[%d]: %08x\n", i, cmd->data[i]);
597 }
598 
599 static void amd_iommu_report_rmp_hw_error(struct amd_iommu *iommu, volatile u32 *event)
600 {
601 	struct iommu_dev_data *dev_data = NULL;
602 	int devid, vmg_tag, flags;
603 	struct pci_dev *pdev;
604 	u64 spa;
605 
606 	devid   = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
607 	vmg_tag = (event[1]) & 0xFFFF;
608 	flags   = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
609 	spa     = ((u64)event[3] << 32) | (event[2] & 0xFFFFFFF8);
610 
611 	pdev = pci_get_domain_bus_and_slot(iommu->pci_seg->id, PCI_BUS_NUM(devid),
612 					   devid & 0xff);
613 	if (pdev)
614 		dev_data = dev_iommu_priv_get(&pdev->dev);
615 
616 	if (dev_data) {
617 		if (__ratelimit(&dev_data->rs)) {
618 			pci_err(pdev, "Event logged [RMP_HW_ERROR vmg_tag=0x%04x, spa=0x%llx, flags=0x%04x]\n",
619 				vmg_tag, spa, flags);
620 		}
621 	} else {
622 		pr_err_ratelimited("Event logged [RMP_HW_ERROR device=%04x:%02x:%02x.%x, vmg_tag=0x%04x, spa=0x%llx, flags=0x%04x]\n",
623 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
624 			vmg_tag, spa, flags);
625 	}
626 
627 	if (pdev)
628 		pci_dev_put(pdev);
629 }
630 
631 static void amd_iommu_report_rmp_fault(struct amd_iommu *iommu, volatile u32 *event)
632 {
633 	struct iommu_dev_data *dev_data = NULL;
634 	int devid, flags_rmp, vmg_tag, flags;
635 	struct pci_dev *pdev;
636 	u64 gpa;
637 
638 	devid     = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
639 	flags_rmp = (event[0] >> EVENT_FLAGS_SHIFT) & 0xFF;
640 	vmg_tag   = (event[1]) & 0xFFFF;
641 	flags     = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
642 	gpa       = ((u64)event[3] << 32) | event[2];
643 
644 	pdev = pci_get_domain_bus_and_slot(iommu->pci_seg->id, PCI_BUS_NUM(devid),
645 					   devid & 0xff);
646 	if (pdev)
647 		dev_data = dev_iommu_priv_get(&pdev->dev);
648 
649 	if (dev_data) {
650 		if (__ratelimit(&dev_data->rs)) {
651 			pci_err(pdev, "Event logged [RMP_PAGE_FAULT vmg_tag=0x%04x, gpa=0x%llx, flags_rmp=0x%04x, flags=0x%04x]\n",
652 				vmg_tag, gpa, flags_rmp, flags);
653 		}
654 	} else {
655 		pr_err_ratelimited("Event logged [RMP_PAGE_FAULT device=%04x:%02x:%02x.%x, vmg_tag=0x%04x, gpa=0x%llx, flags_rmp=0x%04x, flags=0x%04x]\n",
656 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
657 			vmg_tag, gpa, flags_rmp, flags);
658 	}
659 
660 	if (pdev)
661 		pci_dev_put(pdev);
662 }
663 
664 #define IS_IOMMU_MEM_TRANSACTION(flags)		\
665 	(((flags) & EVENT_FLAG_I) == 0)
666 
667 #define IS_WRITE_REQUEST(flags)			\
668 	((flags) & EVENT_FLAG_RW)
669 
670 static void amd_iommu_report_page_fault(struct amd_iommu *iommu,
671 					u16 devid, u16 domain_id,
672 					u64 address, int flags)
673 {
674 	struct iommu_dev_data *dev_data = NULL;
675 	struct pci_dev *pdev;
676 
677 	pdev = pci_get_domain_bus_and_slot(iommu->pci_seg->id, PCI_BUS_NUM(devid),
678 					   devid & 0xff);
679 	if (pdev)
680 		dev_data = dev_iommu_priv_get(&pdev->dev);
681 
682 	if (dev_data) {
683 		/*
684 		 * If this is a DMA fault (for which the I(nterrupt)
685 		 * bit will be unset), allow report_iommu_fault() to
686 		 * prevent logging it.
687 		 */
688 		if (IS_IOMMU_MEM_TRANSACTION(flags)) {
689 			/* Device not attached to domain properly */
690 			if (dev_data->domain == NULL) {
691 				pr_err_ratelimited("Event logged [Device not attached to domain properly]\n");
692 				pr_err_ratelimited("  device=%04x:%02x:%02x.%x domain=0x%04x\n",
693 						   iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid),
694 						   PCI_FUNC(devid), domain_id);
695 				goto out;
696 			}
697 
698 			if (!report_iommu_fault(&dev_data->domain->domain,
699 						&pdev->dev, address,
700 						IS_WRITE_REQUEST(flags) ?
701 							IOMMU_FAULT_WRITE :
702 							IOMMU_FAULT_READ))
703 				goto out;
704 		}
705 
706 		if (__ratelimit(&dev_data->rs)) {
707 			pci_err(pdev, "Event logged [IO_PAGE_FAULT domain=0x%04x address=0x%llx flags=0x%04x]\n",
708 				domain_id, address, flags);
709 		}
710 	} else {
711 		pr_err_ratelimited("Event logged [IO_PAGE_FAULT device=%04x:%02x:%02x.%x domain=0x%04x address=0x%llx flags=0x%04x]\n",
712 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
713 			domain_id, address, flags);
714 	}
715 
716 out:
717 	if (pdev)
718 		pci_dev_put(pdev);
719 }
720 
721 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
722 {
723 	struct device *dev = iommu->iommu.dev;
724 	int type, devid, flags, tag;
725 	volatile u32 *event = __evt;
726 	int count = 0;
727 	u64 address;
728 	u32 pasid;
729 
730 retry:
731 	type    = (event[1] >> EVENT_TYPE_SHIFT)  & EVENT_TYPE_MASK;
732 	devid   = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
733 	pasid   = (event[0] & EVENT_DOMID_MASK_HI) |
734 		  (event[1] & EVENT_DOMID_MASK_LO);
735 	flags   = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
736 	address = (u64)(((u64)event[3]) << 32) | event[2];
737 
738 	if (type == 0) {
739 		/* Did we hit the erratum? */
740 		if (++count == LOOP_TIMEOUT) {
741 			pr_err("No event written to event log\n");
742 			return;
743 		}
744 		udelay(1);
745 		goto retry;
746 	}
747 
748 	if (type == EVENT_TYPE_IO_FAULT) {
749 		amd_iommu_report_page_fault(iommu, devid, pasid, address, flags);
750 		return;
751 	}
752 
753 	switch (type) {
754 	case EVENT_TYPE_ILL_DEV:
755 		dev_err(dev, "Event logged [ILLEGAL_DEV_TABLE_ENTRY device=%04x:%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
756 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
757 			pasid, address, flags);
758 		dump_dte_entry(iommu, devid);
759 		break;
760 	case EVENT_TYPE_DEV_TAB_ERR:
761 		dev_err(dev, "Event logged [DEV_TAB_HARDWARE_ERROR device=%04x:%02x:%02x.%x "
762 			"address=0x%llx flags=0x%04x]\n",
763 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
764 			address, flags);
765 		break;
766 	case EVENT_TYPE_PAGE_TAB_ERR:
767 		dev_err(dev, "Event logged [PAGE_TAB_HARDWARE_ERROR device=%04x:%02x:%02x.%x pasid=0x%04x address=0x%llx flags=0x%04x]\n",
768 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
769 			pasid, address, flags);
770 		break;
771 	case EVENT_TYPE_ILL_CMD:
772 		dev_err(dev, "Event logged [ILLEGAL_COMMAND_ERROR address=0x%llx]\n", address);
773 		dump_command(address);
774 		break;
775 	case EVENT_TYPE_CMD_HARD_ERR:
776 		dev_err(dev, "Event logged [COMMAND_HARDWARE_ERROR address=0x%llx flags=0x%04x]\n",
777 			address, flags);
778 		break;
779 	case EVENT_TYPE_IOTLB_INV_TO:
780 		dev_err(dev, "Event logged [IOTLB_INV_TIMEOUT device=%04x:%02x:%02x.%x address=0x%llx]\n",
781 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
782 			address);
783 		break;
784 	case EVENT_TYPE_INV_DEV_REQ:
785 		dev_err(dev, "Event logged [INVALID_DEVICE_REQUEST device=%04x:%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
786 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
787 			pasid, address, flags);
788 		break;
789 	case EVENT_TYPE_RMP_FAULT:
790 		amd_iommu_report_rmp_fault(iommu, event);
791 		break;
792 	case EVENT_TYPE_RMP_HW_ERR:
793 		amd_iommu_report_rmp_hw_error(iommu, event);
794 		break;
795 	case EVENT_TYPE_INV_PPR_REQ:
796 		pasid = PPR_PASID(*((u64 *)__evt));
797 		tag = event[1] & 0x03FF;
798 		dev_err(dev, "Event logged [INVALID_PPR_REQUEST device=%04x:%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x tag=0x%03x]\n",
799 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
800 			pasid, address, flags, tag);
801 		break;
802 	default:
803 		dev_err(dev, "Event logged [UNKNOWN event[0]=0x%08x event[1]=0x%08x event[2]=0x%08x event[3]=0x%08x\n",
804 			event[0], event[1], event[2], event[3]);
805 	}
806 
807 	/*
808 	 * To detect the hardware errata 732 we need to clear the
809 	 * entry back to zero. This issue does not exist on SNP
810 	 * enabled system. Also this buffer is not writeable on
811 	 * SNP enabled system.
812 	 */
813 	if (!amd_iommu_snp_en)
814 		memset(__evt, 0, 4 * sizeof(u32));
815 }
816 
817 static void iommu_poll_events(struct amd_iommu *iommu)
818 {
819 	u32 head, tail;
820 
821 	head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
822 	tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
823 
824 	while (head != tail) {
825 		iommu_print_event(iommu, iommu->evt_buf + head);
826 
827 		/* Update head pointer of hardware ring-buffer */
828 		head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
829 		writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
830 	}
831 
832 }
833 
834 #ifdef CONFIG_IRQ_REMAP
835 static int (*iommu_ga_log_notifier)(u32);
836 
837 int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
838 {
839 	iommu_ga_log_notifier = notifier;
840 
841 	return 0;
842 }
843 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
844 
845 static void iommu_poll_ga_log(struct amd_iommu *iommu)
846 {
847 	u32 head, tail;
848 
849 	if (iommu->ga_log == NULL)
850 		return;
851 
852 	head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
853 	tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
854 
855 	while (head != tail) {
856 		volatile u64 *raw;
857 		u64 log_entry;
858 
859 		raw = (u64 *)(iommu->ga_log + head);
860 
861 		/* Avoid memcpy function-call overhead */
862 		log_entry = *raw;
863 
864 		/* Update head pointer of hardware ring-buffer */
865 		head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
866 		writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
867 
868 		/* Handle GA entry */
869 		switch (GA_REQ_TYPE(log_entry)) {
870 		case GA_GUEST_NR:
871 			if (!iommu_ga_log_notifier)
872 				break;
873 
874 			pr_debug("%s: devid=%#x, ga_tag=%#x\n",
875 				 __func__, GA_DEVID(log_entry),
876 				 GA_TAG(log_entry));
877 
878 			if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
879 				pr_err("GA log notifier failed.\n");
880 			break;
881 		default:
882 			break;
883 		}
884 	}
885 }
886 
887 static void
888 amd_iommu_set_pci_msi_domain(struct device *dev, struct amd_iommu *iommu)
889 {
890 	if (!irq_remapping_enabled || !dev_is_pci(dev) ||
891 	    !pci_dev_has_default_msi_parent_domain(to_pci_dev(dev)))
892 		return;
893 
894 	dev_set_msi_domain(dev, iommu->ir_domain);
895 }
896 
897 #else /* CONFIG_IRQ_REMAP */
898 static inline void
899 amd_iommu_set_pci_msi_domain(struct device *dev, struct amd_iommu *iommu) { }
900 #endif /* !CONFIG_IRQ_REMAP */
901 
902 static void amd_iommu_handle_irq(void *data, const char *evt_type,
903 				 u32 int_mask, u32 overflow_mask,
904 				 void (*int_handler)(struct amd_iommu *),
905 				 void (*overflow_handler)(struct amd_iommu *))
906 {
907 	struct amd_iommu *iommu = (struct amd_iommu *) data;
908 	u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
909 	u32 mask = int_mask | overflow_mask;
910 
911 	while (status & mask) {
912 		/* Enable interrupt sources again */
913 		writel(mask, iommu->mmio_base + MMIO_STATUS_OFFSET);
914 
915 		if (int_handler) {
916 			pr_devel("Processing IOMMU (ivhd%d) %s Log\n",
917 				 iommu->index, evt_type);
918 			int_handler(iommu);
919 		}
920 
921 		if ((status & overflow_mask) && overflow_handler)
922 			overflow_handler(iommu);
923 
924 		/*
925 		 * Hardware bug: ERBT1312
926 		 * When re-enabling interrupt (by writing 1
927 		 * to clear the bit), the hardware might also try to set
928 		 * the interrupt bit in the event status register.
929 		 * In this scenario, the bit will be set, and disable
930 		 * subsequent interrupts.
931 		 *
932 		 * Workaround: The IOMMU driver should read back the
933 		 * status register and check if the interrupt bits are cleared.
934 		 * If not, driver will need to go through the interrupt handler
935 		 * again and re-clear the bits
936 		 */
937 		status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
938 	}
939 }
940 
941 irqreturn_t amd_iommu_int_thread_evtlog(int irq, void *data)
942 {
943 	amd_iommu_handle_irq(data, "Evt", MMIO_STATUS_EVT_INT_MASK,
944 			     MMIO_STATUS_EVT_OVERFLOW_MASK,
945 			     iommu_poll_events, amd_iommu_restart_event_logging);
946 
947 	return IRQ_HANDLED;
948 }
949 
950 irqreturn_t amd_iommu_int_thread_pprlog(int irq, void *data)
951 {
952 	amd_iommu_handle_irq(data, "PPR", MMIO_STATUS_PPR_INT_MASK,
953 			     MMIO_STATUS_PPR_OVERFLOW_MASK,
954 			     amd_iommu_poll_ppr_log, amd_iommu_restart_ppr_log);
955 
956 	return IRQ_HANDLED;
957 }
958 
959 irqreturn_t amd_iommu_int_thread_galog(int irq, void *data)
960 {
961 #ifdef CONFIG_IRQ_REMAP
962 	amd_iommu_handle_irq(data, "GA", MMIO_STATUS_GALOG_INT_MASK,
963 			     MMIO_STATUS_GALOG_OVERFLOW_MASK,
964 			     iommu_poll_ga_log, amd_iommu_restart_ga_log);
965 #endif
966 
967 	return IRQ_HANDLED;
968 }
969 
970 irqreturn_t amd_iommu_int_thread(int irq, void *data)
971 {
972 	amd_iommu_int_thread_evtlog(irq, data);
973 	amd_iommu_int_thread_pprlog(irq, data);
974 	amd_iommu_int_thread_galog(irq, data);
975 
976 	return IRQ_HANDLED;
977 }
978 
979 irqreturn_t amd_iommu_int_handler(int irq, void *data)
980 {
981 	return IRQ_WAKE_THREAD;
982 }
983 
984 /****************************************************************************
985  *
986  * IOMMU command queuing functions
987  *
988  ****************************************************************************/
989 
990 static int wait_on_sem(struct amd_iommu *iommu, u64 data)
991 {
992 	int i = 0;
993 
994 	while (*iommu->cmd_sem != data && i < LOOP_TIMEOUT) {
995 		udelay(1);
996 		i += 1;
997 	}
998 
999 	if (i == LOOP_TIMEOUT) {
1000 		pr_alert("Completion-Wait loop timed out\n");
1001 		return -EIO;
1002 	}
1003 
1004 	return 0;
1005 }
1006 
1007 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
1008 			       struct iommu_cmd *cmd)
1009 {
1010 	u8 *target;
1011 	u32 tail;
1012 
1013 	/* Copy command to buffer */
1014 	tail = iommu->cmd_buf_tail;
1015 	target = iommu->cmd_buf + tail;
1016 	memcpy(target, cmd, sizeof(*cmd));
1017 
1018 	tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1019 	iommu->cmd_buf_tail = tail;
1020 
1021 	/* Tell the IOMMU about it */
1022 	writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
1023 }
1024 
1025 static void build_completion_wait(struct iommu_cmd *cmd,
1026 				  struct amd_iommu *iommu,
1027 				  u64 data)
1028 {
1029 	u64 paddr = iommu_virt_to_phys((void *)iommu->cmd_sem);
1030 
1031 	memset(cmd, 0, sizeof(*cmd));
1032 	cmd->data[0] = lower_32_bits(paddr) | CMD_COMPL_WAIT_STORE_MASK;
1033 	cmd->data[1] = upper_32_bits(paddr);
1034 	cmd->data[2] = lower_32_bits(data);
1035 	cmd->data[3] = upper_32_bits(data);
1036 	CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
1037 }
1038 
1039 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
1040 {
1041 	memset(cmd, 0, sizeof(*cmd));
1042 	cmd->data[0] = devid;
1043 	CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
1044 }
1045 
1046 /*
1047  * Builds an invalidation address which is suitable for one page or multiple
1048  * pages. Sets the size bit (S) as needed is more than one page is flushed.
1049  */
1050 static inline u64 build_inv_address(u64 address, size_t size)
1051 {
1052 	u64 pages, end, msb_diff;
1053 
1054 	pages = iommu_num_pages(address, size, PAGE_SIZE);
1055 
1056 	if (pages == 1)
1057 		return address & PAGE_MASK;
1058 
1059 	end = address + size - 1;
1060 
1061 	/*
1062 	 * msb_diff would hold the index of the most significant bit that
1063 	 * flipped between the start and end.
1064 	 */
1065 	msb_diff = fls64(end ^ address) - 1;
1066 
1067 	/*
1068 	 * Bits 63:52 are sign extended. If for some reason bit 51 is different
1069 	 * between the start and the end, invalidate everything.
1070 	 */
1071 	if (unlikely(msb_diff > 51)) {
1072 		address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
1073 	} else {
1074 		/*
1075 		 * The msb-bit must be clear on the address. Just set all the
1076 		 * lower bits.
1077 		 */
1078 		address |= (1ull << msb_diff) - 1;
1079 	}
1080 
1081 	/* Clear bits 11:0 */
1082 	address &= PAGE_MASK;
1083 
1084 	/* Set the size bit - we flush more than one 4kb page */
1085 	return address | CMD_INV_IOMMU_PAGES_SIZE_MASK;
1086 }
1087 
1088 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
1089 				  size_t size, u16 domid,
1090 				  ioasid_t pasid, bool gn)
1091 {
1092 	u64 inv_address = build_inv_address(address, size);
1093 
1094 	memset(cmd, 0, sizeof(*cmd));
1095 
1096 	cmd->data[1] |= domid;
1097 	cmd->data[2]  = lower_32_bits(inv_address);
1098 	cmd->data[3]  = upper_32_bits(inv_address);
1099 	/* PDE bit - we want to flush everything, not only the PTEs */
1100 	cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
1101 	if (gn) {
1102 		cmd->data[0] |= pasid;
1103 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
1104 	}
1105 	CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
1106 }
1107 
1108 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
1109 				  u64 address, size_t size,
1110 				  ioasid_t pasid, bool gn)
1111 {
1112 	u64 inv_address = build_inv_address(address, size);
1113 
1114 	memset(cmd, 0, sizeof(*cmd));
1115 
1116 	cmd->data[0]  = devid;
1117 	cmd->data[0] |= (qdep & 0xff) << 24;
1118 	cmd->data[1]  = devid;
1119 	cmd->data[2]  = lower_32_bits(inv_address);
1120 	cmd->data[3]  = upper_32_bits(inv_address);
1121 	if (gn) {
1122 		cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
1123 		cmd->data[1] |= (pasid & 0xff) << 16;
1124 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
1125 	}
1126 
1127 	CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
1128 }
1129 
1130 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, u32 pasid,
1131 			       int status, int tag, u8 gn)
1132 {
1133 	memset(cmd, 0, sizeof(*cmd));
1134 
1135 	cmd->data[0]  = devid;
1136 	if (gn) {
1137 		cmd->data[1]  = pasid;
1138 		cmd->data[2]  = CMD_INV_IOMMU_PAGES_GN_MASK;
1139 	}
1140 	cmd->data[3]  = tag & 0x1ff;
1141 	cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
1142 
1143 	CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
1144 }
1145 
1146 static void build_inv_all(struct iommu_cmd *cmd)
1147 {
1148 	memset(cmd, 0, sizeof(*cmd));
1149 	CMD_SET_TYPE(cmd, CMD_INV_ALL);
1150 }
1151 
1152 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1153 {
1154 	memset(cmd, 0, sizeof(*cmd));
1155 	cmd->data[0] = devid;
1156 	CMD_SET_TYPE(cmd, CMD_INV_IRT);
1157 }
1158 
1159 /*
1160  * Writes the command to the IOMMUs command buffer and informs the
1161  * hardware about the new command.
1162  */
1163 static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1164 				      struct iommu_cmd *cmd,
1165 				      bool sync)
1166 {
1167 	unsigned int count = 0;
1168 	u32 left, next_tail;
1169 
1170 	next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1171 again:
1172 	left      = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE;
1173 
1174 	if (left <= 0x20) {
1175 		/* Skip udelay() the first time around */
1176 		if (count++) {
1177 			if (count == LOOP_TIMEOUT) {
1178 				pr_err("Command buffer timeout\n");
1179 				return -EIO;
1180 			}
1181 
1182 			udelay(1);
1183 		}
1184 
1185 		/* Update head and recheck remaining space */
1186 		iommu->cmd_buf_head = readl(iommu->mmio_base +
1187 					    MMIO_CMD_HEAD_OFFSET);
1188 
1189 		goto again;
1190 	}
1191 
1192 	copy_cmd_to_buffer(iommu, cmd);
1193 
1194 	/* Do we need to make sure all commands are processed? */
1195 	iommu->need_sync = sync;
1196 
1197 	return 0;
1198 }
1199 
1200 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1201 				    struct iommu_cmd *cmd,
1202 				    bool sync)
1203 {
1204 	unsigned long flags;
1205 	int ret;
1206 
1207 	raw_spin_lock_irqsave(&iommu->lock, flags);
1208 	ret = __iommu_queue_command_sync(iommu, cmd, sync);
1209 	raw_spin_unlock_irqrestore(&iommu->lock, flags);
1210 
1211 	return ret;
1212 }
1213 
1214 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1215 {
1216 	return iommu_queue_command_sync(iommu, cmd, true);
1217 }
1218 
1219 /*
1220  * This function queues a completion wait command into the command
1221  * buffer of an IOMMU
1222  */
1223 static int iommu_completion_wait(struct amd_iommu *iommu)
1224 {
1225 	struct iommu_cmd cmd;
1226 	unsigned long flags;
1227 	int ret;
1228 	u64 data;
1229 
1230 	if (!iommu->need_sync)
1231 		return 0;
1232 
1233 	data = atomic64_add_return(1, &iommu->cmd_sem_val);
1234 	build_completion_wait(&cmd, iommu, data);
1235 
1236 	raw_spin_lock_irqsave(&iommu->lock, flags);
1237 
1238 	ret = __iommu_queue_command_sync(iommu, &cmd, false);
1239 	if (ret)
1240 		goto out_unlock;
1241 
1242 	ret = wait_on_sem(iommu, data);
1243 
1244 out_unlock:
1245 	raw_spin_unlock_irqrestore(&iommu->lock, flags);
1246 
1247 	return ret;
1248 }
1249 
1250 static void domain_flush_complete(struct protection_domain *domain)
1251 {
1252 	int i;
1253 
1254 	for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1255 		if (domain && !domain->dev_iommu[i])
1256 			continue;
1257 
1258 		/*
1259 		 * Devices of this domain are behind this IOMMU
1260 		 * We need to wait for completion of all commands.
1261 		 */
1262 		iommu_completion_wait(amd_iommus[i]);
1263 	}
1264 }
1265 
1266 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1267 {
1268 	struct iommu_cmd cmd;
1269 
1270 	build_inv_dte(&cmd, devid);
1271 
1272 	return iommu_queue_command(iommu, &cmd);
1273 }
1274 
1275 static void amd_iommu_flush_dte_all(struct amd_iommu *iommu)
1276 {
1277 	u32 devid;
1278 	u16 last_bdf = iommu->pci_seg->last_bdf;
1279 
1280 	for (devid = 0; devid <= last_bdf; ++devid)
1281 		iommu_flush_dte(iommu, devid);
1282 
1283 	iommu_completion_wait(iommu);
1284 }
1285 
1286 /*
1287  * This function uses heavy locking and may disable irqs for some time. But
1288  * this is no issue because it is only called during resume.
1289  */
1290 static void amd_iommu_flush_tlb_all(struct amd_iommu *iommu)
1291 {
1292 	u32 dom_id;
1293 	u16 last_bdf = iommu->pci_seg->last_bdf;
1294 
1295 	for (dom_id = 0; dom_id <= last_bdf; ++dom_id) {
1296 		struct iommu_cmd cmd;
1297 		build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1298 				      dom_id, IOMMU_NO_PASID, false);
1299 		iommu_queue_command(iommu, &cmd);
1300 	}
1301 
1302 	iommu_completion_wait(iommu);
1303 }
1304 
1305 static void amd_iommu_flush_tlb_domid(struct amd_iommu *iommu, u32 dom_id)
1306 {
1307 	struct iommu_cmd cmd;
1308 
1309 	build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1310 			      dom_id, IOMMU_NO_PASID, false);
1311 	iommu_queue_command(iommu, &cmd);
1312 
1313 	iommu_completion_wait(iommu);
1314 }
1315 
1316 static void amd_iommu_flush_all(struct amd_iommu *iommu)
1317 {
1318 	struct iommu_cmd cmd;
1319 
1320 	build_inv_all(&cmd);
1321 
1322 	iommu_queue_command(iommu, &cmd);
1323 	iommu_completion_wait(iommu);
1324 }
1325 
1326 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1327 {
1328 	struct iommu_cmd cmd;
1329 
1330 	build_inv_irt(&cmd, devid);
1331 
1332 	iommu_queue_command(iommu, &cmd);
1333 }
1334 
1335 static void amd_iommu_flush_irt_all(struct amd_iommu *iommu)
1336 {
1337 	u32 devid;
1338 	u16 last_bdf = iommu->pci_seg->last_bdf;
1339 
1340 	if (iommu->irtcachedis_enabled)
1341 		return;
1342 
1343 	for (devid = 0; devid <= last_bdf; devid++)
1344 		iommu_flush_irt(iommu, devid);
1345 
1346 	iommu_completion_wait(iommu);
1347 }
1348 
1349 void amd_iommu_flush_all_caches(struct amd_iommu *iommu)
1350 {
1351 	if (check_feature(FEATURE_IA)) {
1352 		amd_iommu_flush_all(iommu);
1353 	} else {
1354 		amd_iommu_flush_dte_all(iommu);
1355 		amd_iommu_flush_irt_all(iommu);
1356 		amd_iommu_flush_tlb_all(iommu);
1357 	}
1358 }
1359 
1360 /*
1361  * Command send function for flushing on-device TLB
1362  */
1363 static int device_flush_iotlb(struct iommu_dev_data *dev_data, u64 address,
1364 			      size_t size, ioasid_t pasid, bool gn)
1365 {
1366 	struct amd_iommu *iommu = get_amd_iommu_from_dev_data(dev_data);
1367 	struct iommu_cmd cmd;
1368 	int qdep = dev_data->ats_qdep;
1369 
1370 	build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address,
1371 			      size, pasid, gn);
1372 
1373 	return iommu_queue_command(iommu, &cmd);
1374 }
1375 
1376 static int device_flush_dte_alias(struct pci_dev *pdev, u16 alias, void *data)
1377 {
1378 	struct amd_iommu *iommu = data;
1379 
1380 	return iommu_flush_dte(iommu, alias);
1381 }
1382 
1383 /*
1384  * Command send function for invalidating a device table entry
1385  */
1386 static int device_flush_dte(struct iommu_dev_data *dev_data)
1387 {
1388 	struct amd_iommu *iommu = get_amd_iommu_from_dev_data(dev_data);
1389 	struct pci_dev *pdev = NULL;
1390 	struct amd_iommu_pci_seg *pci_seg;
1391 	u16 alias;
1392 	int ret;
1393 
1394 	if (dev_is_pci(dev_data->dev))
1395 		pdev = to_pci_dev(dev_data->dev);
1396 
1397 	if (pdev)
1398 		ret = pci_for_each_dma_alias(pdev,
1399 					     device_flush_dte_alias, iommu);
1400 	else
1401 		ret = iommu_flush_dte(iommu, dev_data->devid);
1402 	if (ret)
1403 		return ret;
1404 
1405 	pci_seg = iommu->pci_seg;
1406 	alias = pci_seg->alias_table[dev_data->devid];
1407 	if (alias != dev_data->devid) {
1408 		ret = iommu_flush_dte(iommu, alias);
1409 		if (ret)
1410 			return ret;
1411 	}
1412 
1413 	if (dev_data->ats_enabled) {
1414 		/* Invalidate the entire contents of an IOTLB */
1415 		ret = device_flush_iotlb(dev_data, 0, ~0UL,
1416 					 IOMMU_NO_PASID, false);
1417 	}
1418 
1419 	return ret;
1420 }
1421 
1422 static int domain_flush_pages_v2(struct protection_domain *pdom,
1423 				 u64 address, size_t size)
1424 {
1425 	struct iommu_dev_data *dev_data;
1426 	struct iommu_cmd cmd;
1427 	int ret = 0;
1428 
1429 	list_for_each_entry(dev_data, &pdom->dev_list, list) {
1430 		struct amd_iommu *iommu = get_amd_iommu_from_dev(dev_data->dev);
1431 		u16 domid = dev_data->gcr3_info.domid;
1432 
1433 		build_inv_iommu_pages(&cmd, address, size,
1434 				      domid, IOMMU_NO_PASID, true);
1435 
1436 		ret |= iommu_queue_command(iommu, &cmd);
1437 	}
1438 
1439 	return ret;
1440 }
1441 
1442 static int domain_flush_pages_v1(struct protection_domain *pdom,
1443 				 u64 address, size_t size)
1444 {
1445 	struct iommu_cmd cmd;
1446 	int ret = 0, i;
1447 
1448 	build_inv_iommu_pages(&cmd, address, size,
1449 			      pdom->id, IOMMU_NO_PASID, false);
1450 
1451 	for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1452 		if (!pdom->dev_iommu[i])
1453 			continue;
1454 
1455 		/*
1456 		 * Devices of this domain are behind this IOMMU
1457 		 * We need a TLB flush
1458 		 */
1459 		ret |= iommu_queue_command(amd_iommus[i], &cmd);
1460 	}
1461 
1462 	return ret;
1463 }
1464 
1465 /*
1466  * TLB invalidation function which is called from the mapping functions.
1467  * It flushes range of PTEs of the domain.
1468  */
1469 static void __domain_flush_pages(struct protection_domain *domain,
1470 				 u64 address, size_t size)
1471 {
1472 	struct iommu_dev_data *dev_data;
1473 	int ret = 0;
1474 	ioasid_t pasid = IOMMU_NO_PASID;
1475 	bool gn = false;
1476 
1477 	if (pdom_is_v2_pgtbl_mode(domain)) {
1478 		gn = true;
1479 		ret = domain_flush_pages_v2(domain, address, size);
1480 	} else {
1481 		ret = domain_flush_pages_v1(domain, address, size);
1482 	}
1483 
1484 	list_for_each_entry(dev_data, &domain->dev_list, list) {
1485 
1486 		if (!dev_data->ats_enabled)
1487 			continue;
1488 
1489 		ret |= device_flush_iotlb(dev_data, address, size, pasid, gn);
1490 	}
1491 
1492 	WARN_ON(ret);
1493 }
1494 
1495 void amd_iommu_domain_flush_pages(struct protection_domain *domain,
1496 				  u64 address, size_t size)
1497 {
1498 	if (likely(!amd_iommu_np_cache)) {
1499 		__domain_flush_pages(domain, address, size);
1500 
1501 		/* Wait until IOMMU TLB and all device IOTLB flushes are complete */
1502 		domain_flush_complete(domain);
1503 
1504 		return;
1505 	}
1506 
1507 	/*
1508 	 * When NpCache is on, we infer that we run in a VM and use a vIOMMU.
1509 	 * In such setups it is best to avoid flushes of ranges which are not
1510 	 * naturally aligned, since it would lead to flushes of unmodified
1511 	 * PTEs. Such flushes would require the hypervisor to do more work than
1512 	 * necessary. Therefore, perform repeated flushes of aligned ranges
1513 	 * until you cover the range. Each iteration flushes the smaller
1514 	 * between the natural alignment of the address that we flush and the
1515 	 * greatest naturally aligned region that fits in the range.
1516 	 */
1517 	while (size != 0) {
1518 		int addr_alignment = __ffs(address);
1519 		int size_alignment = __fls(size);
1520 		int min_alignment;
1521 		size_t flush_size;
1522 
1523 		/*
1524 		 * size is always non-zero, but address might be zero, causing
1525 		 * addr_alignment to be negative. As the casting of the
1526 		 * argument in __ffs(address) to long might trim the high bits
1527 		 * of the address on x86-32, cast to long when doing the check.
1528 		 */
1529 		if (likely((unsigned long)address != 0))
1530 			min_alignment = min(addr_alignment, size_alignment);
1531 		else
1532 			min_alignment = size_alignment;
1533 
1534 		flush_size = 1ul << min_alignment;
1535 
1536 		__domain_flush_pages(domain, address, flush_size);
1537 		address += flush_size;
1538 		size -= flush_size;
1539 	}
1540 
1541 	/* Wait until IOMMU TLB and all device IOTLB flushes are complete */
1542 	domain_flush_complete(domain);
1543 }
1544 
1545 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1546 static void amd_iommu_domain_flush_all(struct protection_domain *domain)
1547 {
1548 	amd_iommu_domain_flush_pages(domain, 0,
1549 				     CMD_INV_IOMMU_ALL_PAGES_ADDRESS);
1550 }
1551 
1552 void amd_iommu_dev_flush_pasid_pages(struct iommu_dev_data *dev_data,
1553 				     ioasid_t pasid, u64 address, size_t size)
1554 {
1555 	struct iommu_cmd cmd;
1556 	struct amd_iommu *iommu = get_amd_iommu_from_dev(dev_data->dev);
1557 
1558 	build_inv_iommu_pages(&cmd, address, size,
1559 			      dev_data->gcr3_info.domid, pasid, true);
1560 	iommu_queue_command(iommu, &cmd);
1561 
1562 	if (dev_data->ats_enabled)
1563 		device_flush_iotlb(dev_data, address, size, pasid, true);
1564 
1565 	iommu_completion_wait(iommu);
1566 }
1567 
1568 static void dev_flush_pasid_all(struct iommu_dev_data *dev_data,
1569 				ioasid_t pasid)
1570 {
1571 	amd_iommu_dev_flush_pasid_pages(dev_data, pasid, 0,
1572 					CMD_INV_IOMMU_ALL_PAGES_ADDRESS);
1573 }
1574 
1575 /* Flush the not present cache if it exists */
1576 static void domain_flush_np_cache(struct protection_domain *domain,
1577 		dma_addr_t iova, size_t size)
1578 {
1579 	if (unlikely(amd_iommu_np_cache)) {
1580 		unsigned long flags;
1581 
1582 		spin_lock_irqsave(&domain->lock, flags);
1583 		amd_iommu_domain_flush_pages(domain, iova, size);
1584 		spin_unlock_irqrestore(&domain->lock, flags);
1585 	}
1586 }
1587 
1588 
1589 /*
1590  * This function flushes the DTEs for all devices in domain
1591  */
1592 void amd_iommu_update_and_flush_device_table(struct protection_domain *domain)
1593 {
1594 	struct iommu_dev_data *dev_data;
1595 
1596 	list_for_each_entry(dev_data, &domain->dev_list, list) {
1597 		struct amd_iommu *iommu = rlookup_amd_iommu(dev_data->dev);
1598 
1599 		set_dte_entry(iommu, dev_data);
1600 		clone_aliases(iommu, dev_data->dev);
1601 	}
1602 
1603 	list_for_each_entry(dev_data, &domain->dev_list, list)
1604 		device_flush_dte(dev_data);
1605 
1606 	domain_flush_complete(domain);
1607 }
1608 
1609 void amd_iommu_domain_update(struct protection_domain *domain)
1610 {
1611 	/* Update device table */
1612 	amd_iommu_update_and_flush_device_table(domain);
1613 
1614 	/* Flush domain TLB(s) and wait for completion */
1615 	amd_iommu_domain_flush_all(domain);
1616 }
1617 
1618 int amd_iommu_complete_ppr(struct device *dev, u32 pasid, int status, int tag)
1619 {
1620 	struct iommu_dev_data *dev_data;
1621 	struct amd_iommu *iommu;
1622 	struct iommu_cmd cmd;
1623 
1624 	dev_data = dev_iommu_priv_get(dev);
1625 	iommu    = get_amd_iommu_from_dev(dev);
1626 
1627 	build_complete_ppr(&cmd, dev_data->devid, pasid, status,
1628 			   tag, dev_data->pri_tlp);
1629 
1630 	return iommu_queue_command(iommu, &cmd);
1631 }
1632 
1633 /****************************************************************************
1634  *
1635  * The next functions belong to the domain allocation. A domain is
1636  * allocated for every IOMMU as the default domain. If device isolation
1637  * is enabled, every device get its own domain. The most important thing
1638  * about domains is the page table mapping the DMA address space they
1639  * contain.
1640  *
1641  ****************************************************************************/
1642 
1643 static u16 domain_id_alloc(void)
1644 {
1645 	unsigned long flags;
1646 	int id;
1647 
1648 	spin_lock_irqsave(&pd_bitmap_lock, flags);
1649 	id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1650 	BUG_ON(id == 0);
1651 	if (id > 0 && id < MAX_DOMAIN_ID)
1652 		__set_bit(id, amd_iommu_pd_alloc_bitmap);
1653 	else
1654 		id = 0;
1655 	spin_unlock_irqrestore(&pd_bitmap_lock, flags);
1656 
1657 	return id;
1658 }
1659 
1660 static void domain_id_free(int id)
1661 {
1662 	unsigned long flags;
1663 
1664 	spin_lock_irqsave(&pd_bitmap_lock, flags);
1665 	if (id > 0 && id < MAX_DOMAIN_ID)
1666 		__clear_bit(id, amd_iommu_pd_alloc_bitmap);
1667 	spin_unlock_irqrestore(&pd_bitmap_lock, flags);
1668 }
1669 
1670 static void free_gcr3_tbl_level1(u64 *tbl)
1671 {
1672 	u64 *ptr;
1673 	int i;
1674 
1675 	for (i = 0; i < 512; ++i) {
1676 		if (!(tbl[i] & GCR3_VALID))
1677 			continue;
1678 
1679 		ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1680 
1681 		iommu_free_page(ptr);
1682 	}
1683 }
1684 
1685 static void free_gcr3_tbl_level2(u64 *tbl)
1686 {
1687 	u64 *ptr;
1688 	int i;
1689 
1690 	for (i = 0; i < 512; ++i) {
1691 		if (!(tbl[i] & GCR3_VALID))
1692 			continue;
1693 
1694 		ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1695 
1696 		free_gcr3_tbl_level1(ptr);
1697 	}
1698 }
1699 
1700 static void free_gcr3_table(struct gcr3_tbl_info *gcr3_info)
1701 {
1702 	if (gcr3_info->glx == 2)
1703 		free_gcr3_tbl_level2(gcr3_info->gcr3_tbl);
1704 	else if (gcr3_info->glx == 1)
1705 		free_gcr3_tbl_level1(gcr3_info->gcr3_tbl);
1706 	else
1707 		WARN_ON_ONCE(gcr3_info->glx != 0);
1708 
1709 	gcr3_info->glx = 0;
1710 
1711 	/* Free per device domain ID */
1712 	domain_id_free(gcr3_info->domid);
1713 
1714 	iommu_free_page(gcr3_info->gcr3_tbl);
1715 	gcr3_info->gcr3_tbl = NULL;
1716 }
1717 
1718 /*
1719  * Number of GCR3 table levels required. Level must be 4-Kbyte
1720  * page and can contain up to 512 entries.
1721  */
1722 static int get_gcr3_levels(int pasids)
1723 {
1724 	int levels;
1725 
1726 	if (pasids == -1)
1727 		return amd_iommu_max_glx_val;
1728 
1729 	levels = get_count_order(pasids);
1730 
1731 	return levels ? (DIV_ROUND_UP(levels, 9) - 1) : levels;
1732 }
1733 
1734 static int setup_gcr3_table(struct gcr3_tbl_info *gcr3_info,
1735 			    struct amd_iommu *iommu, int pasids)
1736 {
1737 	int levels = get_gcr3_levels(pasids);
1738 	int nid = iommu ? dev_to_node(&iommu->dev->dev) : NUMA_NO_NODE;
1739 
1740 	if (levels > amd_iommu_max_glx_val)
1741 		return -EINVAL;
1742 
1743 	if (gcr3_info->gcr3_tbl)
1744 		return -EBUSY;
1745 
1746 	/* Allocate per device domain ID */
1747 	gcr3_info->domid = domain_id_alloc();
1748 
1749 	gcr3_info->gcr3_tbl = iommu_alloc_page_node(nid, GFP_ATOMIC);
1750 	if (gcr3_info->gcr3_tbl == NULL) {
1751 		domain_id_free(gcr3_info->domid);
1752 		return -ENOMEM;
1753 	}
1754 
1755 	gcr3_info->glx = levels;
1756 
1757 	return 0;
1758 }
1759 
1760 static u64 *__get_gcr3_pte(struct gcr3_tbl_info *gcr3_info,
1761 			   ioasid_t pasid, bool alloc)
1762 {
1763 	int index;
1764 	u64 *pte;
1765 	u64 *root = gcr3_info->gcr3_tbl;
1766 	int level = gcr3_info->glx;
1767 
1768 	while (true) {
1769 
1770 		index = (pasid >> (9 * level)) & 0x1ff;
1771 		pte   = &root[index];
1772 
1773 		if (level == 0)
1774 			break;
1775 
1776 		if (!(*pte & GCR3_VALID)) {
1777 			if (!alloc)
1778 				return NULL;
1779 
1780 			root = (void *)get_zeroed_page(GFP_ATOMIC);
1781 			if (root == NULL)
1782 				return NULL;
1783 
1784 			*pte = iommu_virt_to_phys(root) | GCR3_VALID;
1785 		}
1786 
1787 		root = iommu_phys_to_virt(*pte & PAGE_MASK);
1788 
1789 		level -= 1;
1790 	}
1791 
1792 	return pte;
1793 }
1794 
1795 static int update_gcr3(struct iommu_dev_data *dev_data,
1796 		       ioasid_t pasid, unsigned long gcr3, bool set)
1797 {
1798 	struct gcr3_tbl_info *gcr3_info = &dev_data->gcr3_info;
1799 	u64 *pte;
1800 
1801 	pte = __get_gcr3_pte(gcr3_info, pasid, true);
1802 	if (pte == NULL)
1803 		return -ENOMEM;
1804 
1805 	if (set)
1806 		*pte = (gcr3 & PAGE_MASK) | GCR3_VALID;
1807 	else
1808 		*pte = 0;
1809 
1810 	dev_flush_pasid_all(dev_data, pasid);
1811 	return 0;
1812 }
1813 
1814 int amd_iommu_set_gcr3(struct iommu_dev_data *dev_data, ioasid_t pasid,
1815 		       unsigned long gcr3)
1816 {
1817 	struct gcr3_tbl_info *gcr3_info = &dev_data->gcr3_info;
1818 	int ret;
1819 
1820 	iommu_group_mutex_assert(dev_data->dev);
1821 
1822 	ret = update_gcr3(dev_data, pasid, gcr3, true);
1823 	if (ret)
1824 		return ret;
1825 
1826 	gcr3_info->pasid_cnt++;
1827 	return ret;
1828 }
1829 
1830 int amd_iommu_clear_gcr3(struct iommu_dev_data *dev_data, ioasid_t pasid)
1831 {
1832 	struct gcr3_tbl_info *gcr3_info = &dev_data->gcr3_info;
1833 	int ret;
1834 
1835 	iommu_group_mutex_assert(dev_data->dev);
1836 
1837 	ret = update_gcr3(dev_data, pasid, 0, false);
1838 	if (ret)
1839 		return ret;
1840 
1841 	gcr3_info->pasid_cnt--;
1842 	return ret;
1843 }
1844 
1845 static void set_dte_entry(struct amd_iommu *iommu,
1846 			  struct iommu_dev_data *dev_data)
1847 {
1848 	u64 pte_root = 0;
1849 	u64 flags = 0;
1850 	u32 old_domid;
1851 	u16 devid = dev_data->devid;
1852 	u16 domid;
1853 	struct protection_domain *domain = dev_data->domain;
1854 	struct dev_table_entry *dev_table = get_dev_table(iommu);
1855 	struct gcr3_tbl_info *gcr3_info = &dev_data->gcr3_info;
1856 
1857 	if (gcr3_info && gcr3_info->gcr3_tbl)
1858 		domid = dev_data->gcr3_info.domid;
1859 	else
1860 		domid = domain->id;
1861 
1862 	if (domain->iop.mode != PAGE_MODE_NONE)
1863 		pte_root = iommu_virt_to_phys(domain->iop.root);
1864 
1865 	pte_root |= (domain->iop.mode & DEV_ENTRY_MODE_MASK)
1866 		    << DEV_ENTRY_MODE_SHIFT;
1867 
1868 	pte_root |= DTE_FLAG_IR | DTE_FLAG_IW | DTE_FLAG_V;
1869 
1870 	/*
1871 	 * When SNP is enabled, Only set TV bit when IOMMU
1872 	 * page translation is in use.
1873 	 */
1874 	if (!amd_iommu_snp_en || (domid != 0))
1875 		pte_root |= DTE_FLAG_TV;
1876 
1877 	flags = dev_table[devid].data[1];
1878 
1879 	if (dev_data->ats_enabled)
1880 		flags |= DTE_FLAG_IOTLB;
1881 
1882 	if (dev_data->ppr)
1883 		pte_root |= 1ULL << DEV_ENTRY_PPR;
1884 
1885 	if (domain->dirty_tracking)
1886 		pte_root |= DTE_FLAG_HAD;
1887 
1888 	if (gcr3_info && gcr3_info->gcr3_tbl) {
1889 		u64 gcr3 = iommu_virt_to_phys(gcr3_info->gcr3_tbl);
1890 		u64 glx  = gcr3_info->glx;
1891 		u64 tmp;
1892 
1893 		pte_root |= DTE_FLAG_GV;
1894 		pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1895 
1896 		/* First mask out possible old values for GCR3 table */
1897 		tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1898 		flags    &= ~tmp;
1899 
1900 		tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1901 		flags    &= ~tmp;
1902 
1903 		/* Encode GCR3 table into DTE */
1904 		tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1905 		pte_root |= tmp;
1906 
1907 		tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1908 		flags    |= tmp;
1909 
1910 		tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1911 		flags    |= tmp;
1912 
1913 		if (amd_iommu_gpt_level == PAGE_MODE_5_LEVEL) {
1914 			dev_table[devid].data[2] |=
1915 				((u64)GUEST_PGTABLE_5_LEVEL << DTE_GPT_LEVEL_SHIFT);
1916 		}
1917 
1918 		/* GIOV is supported with V2 page table mode only */
1919 		if (pdom_is_v2_pgtbl_mode(domain))
1920 			pte_root |= DTE_FLAG_GIOV;
1921 	}
1922 
1923 	flags &= ~DEV_DOMID_MASK;
1924 	flags |= domid;
1925 
1926 	old_domid = dev_table[devid].data[1] & DEV_DOMID_MASK;
1927 	dev_table[devid].data[1]  = flags;
1928 	dev_table[devid].data[0]  = pte_root;
1929 
1930 	/*
1931 	 * A kdump kernel might be replacing a domain ID that was copied from
1932 	 * the previous kernel--if so, it needs to flush the translation cache
1933 	 * entries for the old domain ID that is being overwritten
1934 	 */
1935 	if (old_domid) {
1936 		amd_iommu_flush_tlb_domid(iommu, old_domid);
1937 	}
1938 }
1939 
1940 static void clear_dte_entry(struct amd_iommu *iommu, u16 devid)
1941 {
1942 	struct dev_table_entry *dev_table = get_dev_table(iommu);
1943 
1944 	/* remove entry from the device table seen by the hardware */
1945 	dev_table[devid].data[0]  = DTE_FLAG_V;
1946 
1947 	if (!amd_iommu_snp_en)
1948 		dev_table[devid].data[0] |= DTE_FLAG_TV;
1949 
1950 	dev_table[devid].data[1] &= DTE_FLAG_MASK;
1951 
1952 	amd_iommu_apply_erratum_63(iommu, devid);
1953 }
1954 
1955 /* Update and flush DTE for the given device */
1956 static void dev_update_dte(struct iommu_dev_data *dev_data, bool set)
1957 {
1958 	struct amd_iommu *iommu = get_amd_iommu_from_dev(dev_data->dev);
1959 
1960 	if (set)
1961 		set_dte_entry(iommu, dev_data);
1962 	else
1963 		clear_dte_entry(iommu, dev_data->devid);
1964 
1965 	clone_aliases(iommu, dev_data->dev);
1966 	device_flush_dte(dev_data);
1967 	iommu_completion_wait(iommu);
1968 }
1969 
1970 /*
1971  * If domain is SVA capable then initialize GCR3 table. Also if domain is
1972  * in v2 page table mode then update GCR3[0].
1973  */
1974 static int init_gcr3_table(struct iommu_dev_data *dev_data,
1975 			   struct protection_domain *pdom)
1976 {
1977 	struct amd_iommu *iommu = get_amd_iommu_from_dev_data(dev_data);
1978 	int max_pasids = dev_data->max_pasids;
1979 	int ret = 0;
1980 
1981 	 /*
1982 	  * If domain is in pt mode then setup GCR3 table only if device
1983 	  * is PASID capable
1984 	  */
1985 	if (pdom_is_in_pt_mode(pdom) && !pdev_pasid_supported(dev_data))
1986 		return ret;
1987 
1988 	/*
1989 	 * By default, setup GCR3 table to support MAX PASIDs
1990 	 * supported by the device/IOMMU.
1991 	 */
1992 	ret = setup_gcr3_table(&dev_data->gcr3_info, iommu,
1993 			       max_pasids > 0 ?  max_pasids : 1);
1994 	if (ret)
1995 		return ret;
1996 
1997 	/* Setup GCR3[0] only if domain is setup with v2 page table mode */
1998 	if (!pdom_is_v2_pgtbl_mode(pdom))
1999 		return ret;
2000 
2001 	ret = update_gcr3(dev_data, 0, iommu_virt_to_phys(pdom->iop.pgd), true);
2002 	if (ret)
2003 		free_gcr3_table(&dev_data->gcr3_info);
2004 
2005 	return ret;
2006 }
2007 
2008 static void destroy_gcr3_table(struct iommu_dev_data *dev_data,
2009 			       struct protection_domain *pdom)
2010 {
2011 	struct gcr3_tbl_info *gcr3_info = &dev_data->gcr3_info;
2012 
2013 	if (pdom_is_v2_pgtbl_mode(pdom))
2014 		update_gcr3(dev_data, 0, 0, false);
2015 
2016 	if (gcr3_info->gcr3_tbl == NULL)
2017 		return;
2018 
2019 	free_gcr3_table(gcr3_info);
2020 }
2021 
2022 static int do_attach(struct iommu_dev_data *dev_data,
2023 		     struct protection_domain *domain)
2024 {
2025 	struct amd_iommu *iommu = get_amd_iommu_from_dev_data(dev_data);
2026 	struct io_pgtable_cfg *cfg = &domain->iop.pgtbl.cfg;
2027 	int ret = 0;
2028 
2029 	/* Update data structures */
2030 	dev_data->domain = domain;
2031 	list_add(&dev_data->list, &domain->dev_list);
2032 
2033 	/* Update NUMA Node ID */
2034 	if (cfg->amd.nid == NUMA_NO_NODE)
2035 		cfg->amd.nid = dev_to_node(dev_data->dev);
2036 
2037 	/* Do reference counting */
2038 	domain->dev_iommu[iommu->index] += 1;
2039 	domain->dev_cnt                 += 1;
2040 
2041 	/* Setup GCR3 table */
2042 	if (pdom_is_sva_capable(domain)) {
2043 		ret = init_gcr3_table(dev_data, domain);
2044 		if (ret)
2045 			return ret;
2046 	}
2047 
2048 	return ret;
2049 }
2050 
2051 static void do_detach(struct iommu_dev_data *dev_data)
2052 {
2053 	struct protection_domain *domain = dev_data->domain;
2054 	struct amd_iommu *iommu = get_amd_iommu_from_dev_data(dev_data);
2055 
2056 	/* Clear DTE and flush the entry */
2057 	dev_update_dte(dev_data, false);
2058 
2059 	/* Flush IOTLB and wait for the flushes to finish */
2060 	amd_iommu_domain_flush_all(domain);
2061 
2062 	/* Clear GCR3 table */
2063 	if (pdom_is_sva_capable(domain))
2064 		destroy_gcr3_table(dev_data, domain);
2065 
2066 	/* Update data structures */
2067 	dev_data->domain = NULL;
2068 	list_del(&dev_data->list);
2069 
2070 	/* decrease reference counters - needs to happen after the flushes */
2071 	domain->dev_iommu[iommu->index] -= 1;
2072 	domain->dev_cnt                 -= 1;
2073 }
2074 
2075 /*
2076  * If a device is not yet associated with a domain, this function makes the
2077  * device visible in the domain
2078  */
2079 static int attach_device(struct device *dev,
2080 			 struct protection_domain *domain)
2081 {
2082 	struct iommu_dev_data *dev_data;
2083 	unsigned long flags;
2084 	int ret = 0;
2085 
2086 	spin_lock_irqsave(&domain->lock, flags);
2087 
2088 	dev_data = dev_iommu_priv_get(dev);
2089 
2090 	spin_lock(&dev_data->lock);
2091 
2092 	if (dev_data->domain != NULL) {
2093 		ret = -EBUSY;
2094 		goto out;
2095 	}
2096 
2097 	ret = do_attach(dev_data, domain);
2098 
2099 out:
2100 	spin_unlock(&dev_data->lock);
2101 
2102 	spin_unlock_irqrestore(&domain->lock, flags);
2103 
2104 	return ret;
2105 }
2106 
2107 /*
2108  * Removes a device from a protection domain (with devtable_lock held)
2109  */
2110 static void detach_device(struct device *dev)
2111 {
2112 	struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2113 	struct protection_domain *domain = dev_data->domain;
2114 	struct amd_iommu *iommu = get_amd_iommu_from_dev_data(dev_data);
2115 	unsigned long flags;
2116 	bool ppr = dev_data->ppr;
2117 
2118 	spin_lock_irqsave(&domain->lock, flags);
2119 
2120 	spin_lock(&dev_data->lock);
2121 
2122 	/*
2123 	 * First check if the device is still attached. It might already
2124 	 * be detached from its domain because the generic
2125 	 * iommu_detach_group code detached it and we try again here in
2126 	 * our alias handling.
2127 	 */
2128 	if (WARN_ON(!dev_data->domain))
2129 		goto out;
2130 
2131 	if (ppr) {
2132 		iopf_queue_flush_dev(dev);
2133 
2134 		/* Updated here so that it gets reflected in DTE */
2135 		dev_data->ppr = false;
2136 	}
2137 
2138 	do_detach(dev_data);
2139 
2140 out:
2141 	spin_unlock(&dev_data->lock);
2142 
2143 	spin_unlock_irqrestore(&domain->lock, flags);
2144 
2145 	/* Remove IOPF handler */
2146 	if (ppr)
2147 		amd_iommu_iopf_remove_device(iommu, dev_data);
2148 
2149 	if (dev_is_pci(dev))
2150 		pdev_disable_caps(to_pci_dev(dev));
2151 
2152 }
2153 
2154 static struct iommu_device *amd_iommu_probe_device(struct device *dev)
2155 {
2156 	struct iommu_device *iommu_dev;
2157 	struct amd_iommu *iommu;
2158 	struct iommu_dev_data *dev_data;
2159 	int ret;
2160 
2161 	if (!check_device(dev))
2162 		return ERR_PTR(-ENODEV);
2163 
2164 	iommu = rlookup_amd_iommu(dev);
2165 	if (!iommu)
2166 		return ERR_PTR(-ENODEV);
2167 
2168 	/* Not registered yet? */
2169 	if (!iommu->iommu.ops)
2170 		return ERR_PTR(-ENODEV);
2171 
2172 	if (dev_iommu_priv_get(dev))
2173 		return &iommu->iommu;
2174 
2175 	ret = iommu_init_device(iommu, dev);
2176 	if (ret) {
2177 		dev_err(dev, "Failed to initialize - trying to proceed anyway\n");
2178 		iommu_dev = ERR_PTR(ret);
2179 		iommu_ignore_device(iommu, dev);
2180 		goto out_err;
2181 	}
2182 
2183 	amd_iommu_set_pci_msi_domain(dev, iommu);
2184 	iommu_dev = &iommu->iommu;
2185 
2186 	/*
2187 	 * If IOMMU and device supports PASID then it will contain max
2188 	 * supported PASIDs, else it will be zero.
2189 	 */
2190 	dev_data = dev_iommu_priv_get(dev);
2191 	if (amd_iommu_pasid_supported() && dev_is_pci(dev) &&
2192 	    pdev_pasid_supported(dev_data)) {
2193 		dev_data->max_pasids = min_t(u32, iommu->iommu.max_pasids,
2194 					     pci_max_pasids(to_pci_dev(dev)));
2195 	}
2196 
2197 out_err:
2198 	iommu_completion_wait(iommu);
2199 
2200 	if (dev_is_pci(dev))
2201 		pci_prepare_ats(to_pci_dev(dev), PAGE_SHIFT);
2202 
2203 	return iommu_dev;
2204 }
2205 
2206 static void amd_iommu_release_device(struct device *dev)
2207 {
2208 	struct amd_iommu *iommu;
2209 
2210 	if (!check_device(dev))
2211 		return;
2212 
2213 	iommu = rlookup_amd_iommu(dev);
2214 	if (!iommu)
2215 		return;
2216 
2217 	amd_iommu_uninit_device(dev);
2218 	iommu_completion_wait(iommu);
2219 }
2220 
2221 static struct iommu_group *amd_iommu_device_group(struct device *dev)
2222 {
2223 	if (dev_is_pci(dev))
2224 		return pci_device_group(dev);
2225 
2226 	return acpihid_device_group(dev);
2227 }
2228 
2229 /*****************************************************************************
2230  *
2231  * The following functions belong to the exported interface of AMD IOMMU
2232  *
2233  * This interface allows access to lower level functions of the IOMMU
2234  * like protection domain handling and assignement of devices to domains
2235  * which is not possible with the dma_ops interface.
2236  *
2237  *****************************************************************************/
2238 
2239 static void cleanup_domain(struct protection_domain *domain)
2240 {
2241 	struct iommu_dev_data *entry;
2242 
2243 	lockdep_assert_held(&domain->lock);
2244 
2245 	if (!domain->dev_cnt)
2246 		return;
2247 
2248 	while (!list_empty(&domain->dev_list)) {
2249 		entry = list_first_entry(&domain->dev_list,
2250 					 struct iommu_dev_data, list);
2251 		BUG_ON(!entry->domain);
2252 		do_detach(entry);
2253 	}
2254 	WARN_ON(domain->dev_cnt != 0);
2255 }
2256 
2257 void protection_domain_free(struct protection_domain *domain)
2258 {
2259 	WARN_ON(!list_empty(&domain->dev_list));
2260 	if (domain->domain.type & __IOMMU_DOMAIN_PAGING)
2261 		free_io_pgtable_ops(&domain->iop.pgtbl.ops);
2262 	domain_id_free(domain->id);
2263 	kfree(domain);
2264 }
2265 
2266 struct protection_domain *protection_domain_alloc(unsigned int type, int nid)
2267 {
2268 	struct io_pgtable_ops *pgtbl_ops;
2269 	struct protection_domain *domain;
2270 	int pgtable;
2271 
2272 	domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2273 	if (!domain)
2274 		return NULL;
2275 
2276 	domain->id = domain_id_alloc();
2277 	if (!domain->id)
2278 		goto err_free;
2279 
2280 	spin_lock_init(&domain->lock);
2281 	INIT_LIST_HEAD(&domain->dev_list);
2282 	INIT_LIST_HEAD(&domain->dev_data_list);
2283 	domain->iop.pgtbl.cfg.amd.nid = nid;
2284 
2285 	switch (type) {
2286 	/* No need to allocate io pgtable ops in passthrough mode */
2287 	case IOMMU_DOMAIN_IDENTITY:
2288 	case IOMMU_DOMAIN_SVA:
2289 		return domain;
2290 	case IOMMU_DOMAIN_DMA:
2291 		pgtable = amd_iommu_pgtable;
2292 		break;
2293 	/*
2294 	 * Force IOMMU v1 page table when allocating
2295 	 * domain for pass-through devices.
2296 	 */
2297 	case IOMMU_DOMAIN_UNMANAGED:
2298 		pgtable = AMD_IOMMU_V1;
2299 		break;
2300 	default:
2301 		goto err_id;
2302 	}
2303 
2304 	switch (pgtable) {
2305 	case AMD_IOMMU_V1:
2306 		domain->pd_mode = PD_MODE_V1;
2307 		break;
2308 	case AMD_IOMMU_V2:
2309 		domain->pd_mode = PD_MODE_V2;
2310 		break;
2311 	default:
2312 		goto err_id;
2313 	}
2314 
2315 	pgtbl_ops =
2316 		alloc_io_pgtable_ops(pgtable, &domain->iop.pgtbl.cfg, domain);
2317 	if (!pgtbl_ops)
2318 		goto err_id;
2319 
2320 	return domain;
2321 err_id:
2322 	domain_id_free(domain->id);
2323 err_free:
2324 	kfree(domain);
2325 	return NULL;
2326 }
2327 
2328 static inline u64 dma_max_address(void)
2329 {
2330 	if (amd_iommu_pgtable == AMD_IOMMU_V1)
2331 		return ~0ULL;
2332 
2333 	/* V2 with 4/5 level page table */
2334 	return ((1ULL << PM_LEVEL_SHIFT(amd_iommu_gpt_level)) - 1);
2335 }
2336 
2337 static bool amd_iommu_hd_support(struct amd_iommu *iommu)
2338 {
2339 	return iommu && (iommu->features & FEATURE_HDSUP);
2340 }
2341 
2342 static struct iommu_domain *do_iommu_domain_alloc(unsigned int type,
2343 						  struct device *dev, u32 flags)
2344 {
2345 	bool dirty_tracking = flags & IOMMU_HWPT_ALLOC_DIRTY_TRACKING;
2346 	struct protection_domain *domain;
2347 	struct amd_iommu *iommu = NULL;
2348 
2349 	if (dev)
2350 		iommu = get_amd_iommu_from_dev(dev);
2351 
2352 	/*
2353 	 * Since DTE[Mode]=0 is prohibited on SNP-enabled system,
2354 	 * default to use IOMMU_DOMAIN_DMA[_FQ].
2355 	 */
2356 	if (amd_iommu_snp_en && (type == IOMMU_DOMAIN_IDENTITY))
2357 		return ERR_PTR(-EINVAL);
2358 
2359 	if (dirty_tracking && !amd_iommu_hd_support(iommu))
2360 		return ERR_PTR(-EOPNOTSUPP);
2361 
2362 	domain = protection_domain_alloc(type,
2363 					 dev ? dev_to_node(dev) : NUMA_NO_NODE);
2364 	if (!domain)
2365 		return ERR_PTR(-ENOMEM);
2366 
2367 	domain->domain.geometry.aperture_start = 0;
2368 	domain->domain.geometry.aperture_end   = dma_max_address();
2369 	domain->domain.geometry.force_aperture = true;
2370 	domain->domain.pgsize_bitmap = domain->iop.pgtbl.cfg.pgsize_bitmap;
2371 
2372 	if (iommu) {
2373 		domain->domain.type = type;
2374 		domain->domain.ops = iommu->iommu.ops->default_domain_ops;
2375 
2376 		if (dirty_tracking)
2377 			domain->domain.dirty_ops = &amd_dirty_ops;
2378 	}
2379 
2380 	return &domain->domain;
2381 }
2382 
2383 static struct iommu_domain *amd_iommu_domain_alloc(unsigned int type)
2384 {
2385 	struct iommu_domain *domain;
2386 
2387 	domain = do_iommu_domain_alloc(type, NULL, 0);
2388 	if (IS_ERR(domain))
2389 		return NULL;
2390 
2391 	return domain;
2392 }
2393 
2394 static struct iommu_domain *
2395 amd_iommu_domain_alloc_user(struct device *dev, u32 flags,
2396 			    struct iommu_domain *parent,
2397 			    const struct iommu_user_data *user_data)
2398 
2399 {
2400 	unsigned int type = IOMMU_DOMAIN_UNMANAGED;
2401 
2402 	if ((flags & ~IOMMU_HWPT_ALLOC_DIRTY_TRACKING) || parent || user_data)
2403 		return ERR_PTR(-EOPNOTSUPP);
2404 
2405 	return do_iommu_domain_alloc(type, dev, flags);
2406 }
2407 
2408 void amd_iommu_domain_free(struct iommu_domain *dom)
2409 {
2410 	struct protection_domain *domain;
2411 	unsigned long flags;
2412 
2413 	domain = to_pdomain(dom);
2414 
2415 	spin_lock_irqsave(&domain->lock, flags);
2416 
2417 	cleanup_domain(domain);
2418 
2419 	spin_unlock_irqrestore(&domain->lock, flags);
2420 
2421 	protection_domain_free(domain);
2422 }
2423 
2424 static int blocked_domain_attach_device(struct iommu_domain *domain,
2425 					struct device *dev)
2426 {
2427 	struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2428 
2429 	if (dev_data->domain)
2430 		detach_device(dev);
2431 
2432 	/* Clear DTE and flush the entry */
2433 	spin_lock(&dev_data->lock);
2434 	dev_update_dte(dev_data, false);
2435 	spin_unlock(&dev_data->lock);
2436 
2437 	return 0;
2438 }
2439 
2440 static struct iommu_domain blocked_domain = {
2441 	.type = IOMMU_DOMAIN_BLOCKED,
2442 	.ops = &(const struct iommu_domain_ops) {
2443 		.attach_dev     = blocked_domain_attach_device,
2444 	}
2445 };
2446 
2447 static int amd_iommu_attach_device(struct iommu_domain *dom,
2448 				   struct device *dev)
2449 {
2450 	struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2451 	struct protection_domain *domain = to_pdomain(dom);
2452 	struct amd_iommu *iommu = get_amd_iommu_from_dev(dev);
2453 	struct pci_dev *pdev;
2454 	int ret;
2455 
2456 	/*
2457 	 * Skip attach device to domain if new domain is same as
2458 	 * devices current domain
2459 	 */
2460 	if (dev_data->domain == domain)
2461 		return 0;
2462 
2463 	dev_data->defer_attach = false;
2464 
2465 	/*
2466 	 * Restrict to devices with compatible IOMMU hardware support
2467 	 * when enforcement of dirty tracking is enabled.
2468 	 */
2469 	if (dom->dirty_ops && !amd_iommu_hd_support(iommu))
2470 		return -EINVAL;
2471 
2472 	if (dev_data->domain)
2473 		detach_device(dev);
2474 
2475 	ret = attach_device(dev, domain);
2476 
2477 #ifdef CONFIG_IRQ_REMAP
2478 	if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
2479 		if (dom->type == IOMMU_DOMAIN_UNMANAGED)
2480 			dev_data->use_vapic = 1;
2481 		else
2482 			dev_data->use_vapic = 0;
2483 	}
2484 #endif
2485 
2486 	pdev = dev_is_pci(dev_data->dev) ? to_pci_dev(dev_data->dev) : NULL;
2487 	if (pdev && pdom_is_sva_capable(domain)) {
2488 		pdev_enable_caps(pdev);
2489 
2490 		/*
2491 		 * Device can continue to function even if IOPF
2492 		 * enablement failed. Hence in error path just
2493 		 * disable device PRI support.
2494 		 */
2495 		if (amd_iommu_iopf_add_device(iommu, dev_data))
2496 			pdev_disable_cap_pri(pdev);
2497 	} else if (pdev) {
2498 		pdev_enable_cap_ats(pdev);
2499 	}
2500 
2501 	/* Update device table */
2502 	dev_update_dte(dev_data, true);
2503 
2504 	return ret;
2505 }
2506 
2507 static int amd_iommu_iotlb_sync_map(struct iommu_domain *dom,
2508 				    unsigned long iova, size_t size)
2509 {
2510 	struct protection_domain *domain = to_pdomain(dom);
2511 	struct io_pgtable_ops *ops = &domain->iop.pgtbl.ops;
2512 
2513 	if (ops->map_pages)
2514 		domain_flush_np_cache(domain, iova, size);
2515 	return 0;
2516 }
2517 
2518 static int amd_iommu_map_pages(struct iommu_domain *dom, unsigned long iova,
2519 			       phys_addr_t paddr, size_t pgsize, size_t pgcount,
2520 			       int iommu_prot, gfp_t gfp, size_t *mapped)
2521 {
2522 	struct protection_domain *domain = to_pdomain(dom);
2523 	struct io_pgtable_ops *ops = &domain->iop.pgtbl.ops;
2524 	int prot = 0;
2525 	int ret = -EINVAL;
2526 
2527 	if ((domain->pd_mode == PD_MODE_V1) &&
2528 	    (domain->iop.mode == PAGE_MODE_NONE))
2529 		return -EINVAL;
2530 
2531 	if (iommu_prot & IOMMU_READ)
2532 		prot |= IOMMU_PROT_IR;
2533 	if (iommu_prot & IOMMU_WRITE)
2534 		prot |= IOMMU_PROT_IW;
2535 
2536 	if (ops->map_pages) {
2537 		ret = ops->map_pages(ops, iova, paddr, pgsize,
2538 				     pgcount, prot, gfp, mapped);
2539 	}
2540 
2541 	return ret;
2542 }
2543 
2544 static void amd_iommu_iotlb_gather_add_page(struct iommu_domain *domain,
2545 					    struct iommu_iotlb_gather *gather,
2546 					    unsigned long iova, size_t size)
2547 {
2548 	/*
2549 	 * AMD's IOMMU can flush as many pages as necessary in a single flush.
2550 	 * Unless we run in a virtual machine, which can be inferred according
2551 	 * to whether "non-present cache" is on, it is probably best to prefer
2552 	 * (potentially) too extensive TLB flushing (i.e., more misses) over
2553 	 * mutliple TLB flushes (i.e., more flushes). For virtual machines the
2554 	 * hypervisor needs to synchronize the host IOMMU PTEs with those of
2555 	 * the guest, and the trade-off is different: unnecessary TLB flushes
2556 	 * should be avoided.
2557 	 */
2558 	if (amd_iommu_np_cache &&
2559 	    iommu_iotlb_gather_is_disjoint(gather, iova, size))
2560 		iommu_iotlb_sync(domain, gather);
2561 
2562 	iommu_iotlb_gather_add_range(gather, iova, size);
2563 }
2564 
2565 static size_t amd_iommu_unmap_pages(struct iommu_domain *dom, unsigned long iova,
2566 				    size_t pgsize, size_t pgcount,
2567 				    struct iommu_iotlb_gather *gather)
2568 {
2569 	struct protection_domain *domain = to_pdomain(dom);
2570 	struct io_pgtable_ops *ops = &domain->iop.pgtbl.ops;
2571 	size_t r;
2572 
2573 	if ((domain->pd_mode == PD_MODE_V1) &&
2574 	    (domain->iop.mode == PAGE_MODE_NONE))
2575 		return 0;
2576 
2577 	r = (ops->unmap_pages) ? ops->unmap_pages(ops, iova, pgsize, pgcount, NULL) : 0;
2578 
2579 	if (r)
2580 		amd_iommu_iotlb_gather_add_page(dom, gather, iova, r);
2581 
2582 	return r;
2583 }
2584 
2585 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
2586 					  dma_addr_t iova)
2587 {
2588 	struct protection_domain *domain = to_pdomain(dom);
2589 	struct io_pgtable_ops *ops = &domain->iop.pgtbl.ops;
2590 
2591 	return ops->iova_to_phys(ops, iova);
2592 }
2593 
2594 static bool amd_iommu_capable(struct device *dev, enum iommu_cap cap)
2595 {
2596 	switch (cap) {
2597 	case IOMMU_CAP_CACHE_COHERENCY:
2598 		return true;
2599 	case IOMMU_CAP_NOEXEC:
2600 		return false;
2601 	case IOMMU_CAP_PRE_BOOT_PROTECTION:
2602 		return amdr_ivrs_remap_support;
2603 	case IOMMU_CAP_ENFORCE_CACHE_COHERENCY:
2604 		return true;
2605 	case IOMMU_CAP_DEFERRED_FLUSH:
2606 		return true;
2607 	case IOMMU_CAP_DIRTY_TRACKING: {
2608 		struct amd_iommu *iommu = get_amd_iommu_from_dev(dev);
2609 
2610 		return amd_iommu_hd_support(iommu);
2611 	}
2612 	default:
2613 		break;
2614 	}
2615 
2616 	return false;
2617 }
2618 
2619 static int amd_iommu_set_dirty_tracking(struct iommu_domain *domain,
2620 					bool enable)
2621 {
2622 	struct protection_domain *pdomain = to_pdomain(domain);
2623 	struct dev_table_entry *dev_table;
2624 	struct iommu_dev_data *dev_data;
2625 	bool domain_flush = false;
2626 	struct amd_iommu *iommu;
2627 	unsigned long flags;
2628 	u64 pte_root;
2629 
2630 	spin_lock_irqsave(&pdomain->lock, flags);
2631 	if (!(pdomain->dirty_tracking ^ enable)) {
2632 		spin_unlock_irqrestore(&pdomain->lock, flags);
2633 		return 0;
2634 	}
2635 
2636 	list_for_each_entry(dev_data, &pdomain->dev_list, list) {
2637 		iommu = get_amd_iommu_from_dev_data(dev_data);
2638 
2639 		dev_table = get_dev_table(iommu);
2640 		pte_root = dev_table[dev_data->devid].data[0];
2641 
2642 		pte_root = (enable ? pte_root | DTE_FLAG_HAD :
2643 				     pte_root & ~DTE_FLAG_HAD);
2644 
2645 		/* Flush device DTE */
2646 		dev_table[dev_data->devid].data[0] = pte_root;
2647 		device_flush_dte(dev_data);
2648 		domain_flush = true;
2649 	}
2650 
2651 	/* Flush IOTLB to mark IOPTE dirty on the next translation(s) */
2652 	if (domain_flush)
2653 		amd_iommu_domain_flush_all(pdomain);
2654 
2655 	pdomain->dirty_tracking = enable;
2656 	spin_unlock_irqrestore(&pdomain->lock, flags);
2657 
2658 	return 0;
2659 }
2660 
2661 static int amd_iommu_read_and_clear_dirty(struct iommu_domain *domain,
2662 					  unsigned long iova, size_t size,
2663 					  unsigned long flags,
2664 					  struct iommu_dirty_bitmap *dirty)
2665 {
2666 	struct protection_domain *pdomain = to_pdomain(domain);
2667 	struct io_pgtable_ops *ops = &pdomain->iop.pgtbl.ops;
2668 	unsigned long lflags;
2669 
2670 	if (!ops || !ops->read_and_clear_dirty)
2671 		return -EOPNOTSUPP;
2672 
2673 	spin_lock_irqsave(&pdomain->lock, lflags);
2674 	if (!pdomain->dirty_tracking && dirty->bitmap) {
2675 		spin_unlock_irqrestore(&pdomain->lock, lflags);
2676 		return -EINVAL;
2677 	}
2678 	spin_unlock_irqrestore(&pdomain->lock, lflags);
2679 
2680 	return ops->read_and_clear_dirty(ops, iova, size, flags, dirty);
2681 }
2682 
2683 static void amd_iommu_get_resv_regions(struct device *dev,
2684 				       struct list_head *head)
2685 {
2686 	struct iommu_resv_region *region;
2687 	struct unity_map_entry *entry;
2688 	struct amd_iommu *iommu;
2689 	struct amd_iommu_pci_seg *pci_seg;
2690 	int devid, sbdf;
2691 
2692 	sbdf = get_device_sbdf_id(dev);
2693 	if (sbdf < 0)
2694 		return;
2695 
2696 	devid = PCI_SBDF_TO_DEVID(sbdf);
2697 	iommu = get_amd_iommu_from_dev(dev);
2698 	pci_seg = iommu->pci_seg;
2699 
2700 	list_for_each_entry(entry, &pci_seg->unity_map, list) {
2701 		int type, prot = 0;
2702 		size_t length;
2703 
2704 		if (devid < entry->devid_start || devid > entry->devid_end)
2705 			continue;
2706 
2707 		type   = IOMMU_RESV_DIRECT;
2708 		length = entry->address_end - entry->address_start;
2709 		if (entry->prot & IOMMU_PROT_IR)
2710 			prot |= IOMMU_READ;
2711 		if (entry->prot & IOMMU_PROT_IW)
2712 			prot |= IOMMU_WRITE;
2713 		if (entry->prot & IOMMU_UNITY_MAP_FLAG_EXCL_RANGE)
2714 			/* Exclusion range */
2715 			type = IOMMU_RESV_RESERVED;
2716 
2717 		region = iommu_alloc_resv_region(entry->address_start,
2718 						 length, prot, type,
2719 						 GFP_KERNEL);
2720 		if (!region) {
2721 			dev_err(dev, "Out of memory allocating dm-regions\n");
2722 			return;
2723 		}
2724 		list_add_tail(&region->list, head);
2725 	}
2726 
2727 	region = iommu_alloc_resv_region(MSI_RANGE_START,
2728 					 MSI_RANGE_END - MSI_RANGE_START + 1,
2729 					 0, IOMMU_RESV_MSI, GFP_KERNEL);
2730 	if (!region)
2731 		return;
2732 	list_add_tail(&region->list, head);
2733 
2734 	region = iommu_alloc_resv_region(HT_RANGE_START,
2735 					 HT_RANGE_END - HT_RANGE_START + 1,
2736 					 0, IOMMU_RESV_RESERVED, GFP_KERNEL);
2737 	if (!region)
2738 		return;
2739 	list_add_tail(&region->list, head);
2740 }
2741 
2742 static bool amd_iommu_is_attach_deferred(struct device *dev)
2743 {
2744 	struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2745 
2746 	return dev_data->defer_attach;
2747 }
2748 
2749 static void amd_iommu_flush_iotlb_all(struct iommu_domain *domain)
2750 {
2751 	struct protection_domain *dom = to_pdomain(domain);
2752 	unsigned long flags;
2753 
2754 	spin_lock_irqsave(&dom->lock, flags);
2755 	amd_iommu_domain_flush_all(dom);
2756 	spin_unlock_irqrestore(&dom->lock, flags);
2757 }
2758 
2759 static void amd_iommu_iotlb_sync(struct iommu_domain *domain,
2760 				 struct iommu_iotlb_gather *gather)
2761 {
2762 	struct protection_domain *dom = to_pdomain(domain);
2763 	unsigned long flags;
2764 
2765 	spin_lock_irqsave(&dom->lock, flags);
2766 	amd_iommu_domain_flush_pages(dom, gather->start,
2767 				     gather->end - gather->start + 1);
2768 	spin_unlock_irqrestore(&dom->lock, flags);
2769 }
2770 
2771 static int amd_iommu_def_domain_type(struct device *dev)
2772 {
2773 	struct iommu_dev_data *dev_data;
2774 
2775 	dev_data = dev_iommu_priv_get(dev);
2776 	if (!dev_data)
2777 		return 0;
2778 
2779 	/* Always use DMA domain for untrusted device */
2780 	if (dev_is_pci(dev) && to_pci_dev(dev)->untrusted)
2781 		return IOMMU_DOMAIN_DMA;
2782 
2783 	/*
2784 	 * Do not identity map IOMMUv2 capable devices when:
2785 	 *  - memory encryption is active, because some of those devices
2786 	 *    (AMD GPUs) don't have the encryption bit in their DMA-mask
2787 	 *    and require remapping.
2788 	 *  - SNP is enabled, because it prohibits DTE[Mode]=0.
2789 	 */
2790 	if (pdev_pasid_supported(dev_data) &&
2791 	    !cc_platform_has(CC_ATTR_MEM_ENCRYPT) &&
2792 	    !amd_iommu_snp_en) {
2793 		return IOMMU_DOMAIN_IDENTITY;
2794 	}
2795 
2796 	return 0;
2797 }
2798 
2799 static bool amd_iommu_enforce_cache_coherency(struct iommu_domain *domain)
2800 {
2801 	/* IOMMU_PTE_FC is always set */
2802 	return true;
2803 }
2804 
2805 static const struct iommu_dirty_ops amd_dirty_ops = {
2806 	.set_dirty_tracking = amd_iommu_set_dirty_tracking,
2807 	.read_and_clear_dirty = amd_iommu_read_and_clear_dirty,
2808 };
2809 
2810 static int amd_iommu_dev_enable_feature(struct device *dev,
2811 					enum iommu_dev_features feat)
2812 {
2813 	int ret = 0;
2814 
2815 	switch (feat) {
2816 	case IOMMU_DEV_FEAT_IOPF:
2817 	case IOMMU_DEV_FEAT_SVA:
2818 		break;
2819 	default:
2820 		ret = -EINVAL;
2821 		break;
2822 	}
2823 	return ret;
2824 }
2825 
2826 static int amd_iommu_dev_disable_feature(struct device *dev,
2827 					 enum iommu_dev_features feat)
2828 {
2829 	int ret = 0;
2830 
2831 	switch (feat) {
2832 	case IOMMU_DEV_FEAT_IOPF:
2833 	case IOMMU_DEV_FEAT_SVA:
2834 		break;
2835 	default:
2836 		ret = -EINVAL;
2837 		break;
2838 	}
2839 	return ret;
2840 }
2841 
2842 const struct iommu_ops amd_iommu_ops = {
2843 	.capable = amd_iommu_capable,
2844 	.blocked_domain = &blocked_domain,
2845 	.domain_alloc = amd_iommu_domain_alloc,
2846 	.domain_alloc_user = amd_iommu_domain_alloc_user,
2847 	.domain_alloc_sva = amd_iommu_domain_alloc_sva,
2848 	.probe_device = amd_iommu_probe_device,
2849 	.release_device = amd_iommu_release_device,
2850 	.device_group = amd_iommu_device_group,
2851 	.get_resv_regions = amd_iommu_get_resv_regions,
2852 	.is_attach_deferred = amd_iommu_is_attach_deferred,
2853 	.def_domain_type = amd_iommu_def_domain_type,
2854 	.dev_enable_feat = amd_iommu_dev_enable_feature,
2855 	.dev_disable_feat = amd_iommu_dev_disable_feature,
2856 	.remove_dev_pasid = amd_iommu_remove_dev_pasid,
2857 	.page_response = amd_iommu_page_response,
2858 	.default_domain_ops = &(const struct iommu_domain_ops) {
2859 		.attach_dev	= amd_iommu_attach_device,
2860 		.map_pages	= amd_iommu_map_pages,
2861 		.unmap_pages	= amd_iommu_unmap_pages,
2862 		.iotlb_sync_map	= amd_iommu_iotlb_sync_map,
2863 		.iova_to_phys	= amd_iommu_iova_to_phys,
2864 		.flush_iotlb_all = amd_iommu_flush_iotlb_all,
2865 		.iotlb_sync	= amd_iommu_iotlb_sync,
2866 		.free		= amd_iommu_domain_free,
2867 		.enforce_cache_coherency = amd_iommu_enforce_cache_coherency,
2868 	}
2869 };
2870 
2871 #ifdef CONFIG_IRQ_REMAP
2872 
2873 /*****************************************************************************
2874  *
2875  * Interrupt Remapping Implementation
2876  *
2877  *****************************************************************************/
2878 
2879 static struct irq_chip amd_ir_chip;
2880 static DEFINE_SPINLOCK(iommu_table_lock);
2881 
2882 static void iommu_flush_irt_and_complete(struct amd_iommu *iommu, u16 devid)
2883 {
2884 	int ret;
2885 	u64 data;
2886 	unsigned long flags;
2887 	struct iommu_cmd cmd, cmd2;
2888 
2889 	if (iommu->irtcachedis_enabled)
2890 		return;
2891 
2892 	build_inv_irt(&cmd, devid);
2893 	data = atomic64_add_return(1, &iommu->cmd_sem_val);
2894 	build_completion_wait(&cmd2, iommu, data);
2895 
2896 	raw_spin_lock_irqsave(&iommu->lock, flags);
2897 	ret = __iommu_queue_command_sync(iommu, &cmd, true);
2898 	if (ret)
2899 		goto out;
2900 	ret = __iommu_queue_command_sync(iommu, &cmd2, false);
2901 	if (ret)
2902 		goto out;
2903 	wait_on_sem(iommu, data);
2904 out:
2905 	raw_spin_unlock_irqrestore(&iommu->lock, flags);
2906 }
2907 
2908 static void set_dte_irq_entry(struct amd_iommu *iommu, u16 devid,
2909 			      struct irq_remap_table *table)
2910 {
2911 	u64 dte;
2912 	struct dev_table_entry *dev_table = get_dev_table(iommu);
2913 
2914 	dte	= dev_table[devid].data[2];
2915 	dte	&= ~DTE_IRQ_PHYS_ADDR_MASK;
2916 	dte	|= iommu_virt_to_phys(table->table);
2917 	dte	|= DTE_IRQ_REMAP_INTCTL;
2918 	dte	|= DTE_INTTABLEN;
2919 	dte	|= DTE_IRQ_REMAP_ENABLE;
2920 
2921 	dev_table[devid].data[2] = dte;
2922 }
2923 
2924 static struct irq_remap_table *get_irq_table(struct amd_iommu *iommu, u16 devid)
2925 {
2926 	struct irq_remap_table *table;
2927 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
2928 
2929 	if (WARN_ONCE(!pci_seg->rlookup_table[devid],
2930 		      "%s: no iommu for devid %x:%x\n",
2931 		      __func__, pci_seg->id, devid))
2932 		return NULL;
2933 
2934 	table = pci_seg->irq_lookup_table[devid];
2935 	if (WARN_ONCE(!table, "%s: no table for devid %x:%x\n",
2936 		      __func__, pci_seg->id, devid))
2937 		return NULL;
2938 
2939 	return table;
2940 }
2941 
2942 static struct irq_remap_table *__alloc_irq_table(void)
2943 {
2944 	struct irq_remap_table *table;
2945 
2946 	table = kzalloc(sizeof(*table), GFP_KERNEL);
2947 	if (!table)
2948 		return NULL;
2949 
2950 	table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_KERNEL);
2951 	if (!table->table) {
2952 		kfree(table);
2953 		return NULL;
2954 	}
2955 	raw_spin_lock_init(&table->lock);
2956 
2957 	if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
2958 		memset(table->table, 0,
2959 		       MAX_IRQS_PER_TABLE * sizeof(u32));
2960 	else
2961 		memset(table->table, 0,
2962 		       (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
2963 	return table;
2964 }
2965 
2966 static void set_remap_table_entry(struct amd_iommu *iommu, u16 devid,
2967 				  struct irq_remap_table *table)
2968 {
2969 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
2970 
2971 	pci_seg->irq_lookup_table[devid] = table;
2972 	set_dte_irq_entry(iommu, devid, table);
2973 	iommu_flush_dte(iommu, devid);
2974 }
2975 
2976 static int set_remap_table_entry_alias(struct pci_dev *pdev, u16 alias,
2977 				       void *data)
2978 {
2979 	struct irq_remap_table *table = data;
2980 	struct amd_iommu_pci_seg *pci_seg;
2981 	struct amd_iommu *iommu = rlookup_amd_iommu(&pdev->dev);
2982 
2983 	if (!iommu)
2984 		return -EINVAL;
2985 
2986 	pci_seg = iommu->pci_seg;
2987 	pci_seg->irq_lookup_table[alias] = table;
2988 	set_dte_irq_entry(iommu, alias, table);
2989 	iommu_flush_dte(pci_seg->rlookup_table[alias], alias);
2990 
2991 	return 0;
2992 }
2993 
2994 static struct irq_remap_table *alloc_irq_table(struct amd_iommu *iommu,
2995 					       u16 devid, struct pci_dev *pdev)
2996 {
2997 	struct irq_remap_table *table = NULL;
2998 	struct irq_remap_table *new_table = NULL;
2999 	struct amd_iommu_pci_seg *pci_seg;
3000 	unsigned long flags;
3001 	u16 alias;
3002 
3003 	spin_lock_irqsave(&iommu_table_lock, flags);
3004 
3005 	pci_seg = iommu->pci_seg;
3006 	table = pci_seg->irq_lookup_table[devid];
3007 	if (table)
3008 		goto out_unlock;
3009 
3010 	alias = pci_seg->alias_table[devid];
3011 	table = pci_seg->irq_lookup_table[alias];
3012 	if (table) {
3013 		set_remap_table_entry(iommu, devid, table);
3014 		goto out_wait;
3015 	}
3016 	spin_unlock_irqrestore(&iommu_table_lock, flags);
3017 
3018 	/* Nothing there yet, allocate new irq remapping table */
3019 	new_table = __alloc_irq_table();
3020 	if (!new_table)
3021 		return NULL;
3022 
3023 	spin_lock_irqsave(&iommu_table_lock, flags);
3024 
3025 	table = pci_seg->irq_lookup_table[devid];
3026 	if (table)
3027 		goto out_unlock;
3028 
3029 	table = pci_seg->irq_lookup_table[alias];
3030 	if (table) {
3031 		set_remap_table_entry(iommu, devid, table);
3032 		goto out_wait;
3033 	}
3034 
3035 	table = new_table;
3036 	new_table = NULL;
3037 
3038 	if (pdev)
3039 		pci_for_each_dma_alias(pdev, set_remap_table_entry_alias,
3040 				       table);
3041 	else
3042 		set_remap_table_entry(iommu, devid, table);
3043 
3044 	if (devid != alias)
3045 		set_remap_table_entry(iommu, alias, table);
3046 
3047 out_wait:
3048 	iommu_completion_wait(iommu);
3049 
3050 out_unlock:
3051 	spin_unlock_irqrestore(&iommu_table_lock, flags);
3052 
3053 	if (new_table) {
3054 		kmem_cache_free(amd_iommu_irq_cache, new_table->table);
3055 		kfree(new_table);
3056 	}
3057 	return table;
3058 }
3059 
3060 static int alloc_irq_index(struct amd_iommu *iommu, u16 devid, int count,
3061 			   bool align, struct pci_dev *pdev)
3062 {
3063 	struct irq_remap_table *table;
3064 	int index, c, alignment = 1;
3065 	unsigned long flags;
3066 
3067 	table = alloc_irq_table(iommu, devid, pdev);
3068 	if (!table)
3069 		return -ENODEV;
3070 
3071 	if (align)
3072 		alignment = roundup_pow_of_two(count);
3073 
3074 	raw_spin_lock_irqsave(&table->lock, flags);
3075 
3076 	/* Scan table for free entries */
3077 	for (index = ALIGN(table->min_index, alignment), c = 0;
3078 	     index < MAX_IRQS_PER_TABLE;) {
3079 		if (!iommu->irte_ops->is_allocated(table, index)) {
3080 			c += 1;
3081 		} else {
3082 			c     = 0;
3083 			index = ALIGN(index + 1, alignment);
3084 			continue;
3085 		}
3086 
3087 		if (c == count)	{
3088 			for (; c != 0; --c)
3089 				iommu->irte_ops->set_allocated(table, index - c + 1);
3090 
3091 			index -= count - 1;
3092 			goto out;
3093 		}
3094 
3095 		index++;
3096 	}
3097 
3098 	index = -ENOSPC;
3099 
3100 out:
3101 	raw_spin_unlock_irqrestore(&table->lock, flags);
3102 
3103 	return index;
3104 }
3105 
3106 static int __modify_irte_ga(struct amd_iommu *iommu, u16 devid, int index,
3107 			    struct irte_ga *irte)
3108 {
3109 	struct irq_remap_table *table;
3110 	struct irte_ga *entry;
3111 	unsigned long flags;
3112 	u128 old;
3113 
3114 	table = get_irq_table(iommu, devid);
3115 	if (!table)
3116 		return -ENOMEM;
3117 
3118 	raw_spin_lock_irqsave(&table->lock, flags);
3119 
3120 	entry = (struct irte_ga *)table->table;
3121 	entry = &entry[index];
3122 
3123 	/*
3124 	 * We use cmpxchg16 to atomically update the 128-bit IRTE,
3125 	 * and it cannot be updated by the hardware or other processors
3126 	 * behind us, so the return value of cmpxchg16 should be the
3127 	 * same as the old value.
3128 	 */
3129 	old = entry->irte;
3130 	WARN_ON(!try_cmpxchg128(&entry->irte, &old, irte->irte));
3131 
3132 	raw_spin_unlock_irqrestore(&table->lock, flags);
3133 
3134 	return 0;
3135 }
3136 
3137 static int modify_irte_ga(struct amd_iommu *iommu, u16 devid, int index,
3138 			  struct irte_ga *irte)
3139 {
3140 	bool ret;
3141 
3142 	ret = __modify_irte_ga(iommu, devid, index, irte);
3143 	if (ret)
3144 		return ret;
3145 
3146 	iommu_flush_irt_and_complete(iommu, devid);
3147 
3148 	return 0;
3149 }
3150 
3151 static int modify_irte(struct amd_iommu *iommu,
3152 		       u16 devid, int index, union irte *irte)
3153 {
3154 	struct irq_remap_table *table;
3155 	unsigned long flags;
3156 
3157 	table = get_irq_table(iommu, devid);
3158 	if (!table)
3159 		return -ENOMEM;
3160 
3161 	raw_spin_lock_irqsave(&table->lock, flags);
3162 	table->table[index] = irte->val;
3163 	raw_spin_unlock_irqrestore(&table->lock, flags);
3164 
3165 	iommu_flush_irt_and_complete(iommu, devid);
3166 
3167 	return 0;
3168 }
3169 
3170 static void free_irte(struct amd_iommu *iommu, u16 devid, int index)
3171 {
3172 	struct irq_remap_table *table;
3173 	unsigned long flags;
3174 
3175 	table = get_irq_table(iommu, devid);
3176 	if (!table)
3177 		return;
3178 
3179 	raw_spin_lock_irqsave(&table->lock, flags);
3180 	iommu->irte_ops->clear_allocated(table, index);
3181 	raw_spin_unlock_irqrestore(&table->lock, flags);
3182 
3183 	iommu_flush_irt_and_complete(iommu, devid);
3184 }
3185 
3186 static void irte_prepare(void *entry,
3187 			 u32 delivery_mode, bool dest_mode,
3188 			 u8 vector, u32 dest_apicid, int devid)
3189 {
3190 	union irte *irte = (union irte *) entry;
3191 
3192 	irte->val                = 0;
3193 	irte->fields.vector      = vector;
3194 	irte->fields.int_type    = delivery_mode;
3195 	irte->fields.destination = dest_apicid;
3196 	irte->fields.dm          = dest_mode;
3197 	irte->fields.valid       = 1;
3198 }
3199 
3200 static void irte_ga_prepare(void *entry,
3201 			    u32 delivery_mode, bool dest_mode,
3202 			    u8 vector, u32 dest_apicid, int devid)
3203 {
3204 	struct irte_ga *irte = (struct irte_ga *) entry;
3205 
3206 	irte->lo.val                      = 0;
3207 	irte->hi.val                      = 0;
3208 	irte->lo.fields_remap.int_type    = delivery_mode;
3209 	irte->lo.fields_remap.dm          = dest_mode;
3210 	irte->hi.fields.vector            = vector;
3211 	irte->lo.fields_remap.destination = APICID_TO_IRTE_DEST_LO(dest_apicid);
3212 	irte->hi.fields.destination       = APICID_TO_IRTE_DEST_HI(dest_apicid);
3213 	irte->lo.fields_remap.valid       = 1;
3214 }
3215 
3216 static void irte_activate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3217 {
3218 	union irte *irte = (union irte *) entry;
3219 
3220 	irte->fields.valid = 1;
3221 	modify_irte(iommu, devid, index, irte);
3222 }
3223 
3224 static void irte_ga_activate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3225 {
3226 	struct irte_ga *irte = (struct irte_ga *) entry;
3227 
3228 	irte->lo.fields_remap.valid = 1;
3229 	modify_irte_ga(iommu, devid, index, irte);
3230 }
3231 
3232 static void irte_deactivate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3233 {
3234 	union irte *irte = (union irte *) entry;
3235 
3236 	irte->fields.valid = 0;
3237 	modify_irte(iommu, devid, index, irte);
3238 }
3239 
3240 static void irte_ga_deactivate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3241 {
3242 	struct irte_ga *irte = (struct irte_ga *) entry;
3243 
3244 	irte->lo.fields_remap.valid = 0;
3245 	modify_irte_ga(iommu, devid, index, irte);
3246 }
3247 
3248 static void irte_set_affinity(struct amd_iommu *iommu, void *entry, u16 devid, u16 index,
3249 			      u8 vector, u32 dest_apicid)
3250 {
3251 	union irte *irte = (union irte *) entry;
3252 
3253 	irte->fields.vector = vector;
3254 	irte->fields.destination = dest_apicid;
3255 	modify_irte(iommu, devid, index, irte);
3256 }
3257 
3258 static void irte_ga_set_affinity(struct amd_iommu *iommu, void *entry, u16 devid, u16 index,
3259 				 u8 vector, u32 dest_apicid)
3260 {
3261 	struct irte_ga *irte = (struct irte_ga *) entry;
3262 
3263 	if (!irte->lo.fields_remap.guest_mode) {
3264 		irte->hi.fields.vector = vector;
3265 		irte->lo.fields_remap.destination =
3266 					APICID_TO_IRTE_DEST_LO(dest_apicid);
3267 		irte->hi.fields.destination =
3268 					APICID_TO_IRTE_DEST_HI(dest_apicid);
3269 		modify_irte_ga(iommu, devid, index, irte);
3270 	}
3271 }
3272 
3273 #define IRTE_ALLOCATED (~1U)
3274 static void irte_set_allocated(struct irq_remap_table *table, int index)
3275 {
3276 	table->table[index] = IRTE_ALLOCATED;
3277 }
3278 
3279 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
3280 {
3281 	struct irte_ga *ptr = (struct irte_ga *)table->table;
3282 	struct irte_ga *irte = &ptr[index];
3283 
3284 	memset(&irte->lo.val, 0, sizeof(u64));
3285 	memset(&irte->hi.val, 0, sizeof(u64));
3286 	irte->hi.fields.vector = 0xff;
3287 }
3288 
3289 static bool irte_is_allocated(struct irq_remap_table *table, int index)
3290 {
3291 	union irte *ptr = (union irte *)table->table;
3292 	union irte *irte = &ptr[index];
3293 
3294 	return irte->val != 0;
3295 }
3296 
3297 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
3298 {
3299 	struct irte_ga *ptr = (struct irte_ga *)table->table;
3300 	struct irte_ga *irte = &ptr[index];
3301 
3302 	return irte->hi.fields.vector != 0;
3303 }
3304 
3305 static void irte_clear_allocated(struct irq_remap_table *table, int index)
3306 {
3307 	table->table[index] = 0;
3308 }
3309 
3310 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
3311 {
3312 	struct irte_ga *ptr = (struct irte_ga *)table->table;
3313 	struct irte_ga *irte = &ptr[index];
3314 
3315 	memset(&irte->lo.val, 0, sizeof(u64));
3316 	memset(&irte->hi.val, 0, sizeof(u64));
3317 }
3318 
3319 static int get_devid(struct irq_alloc_info *info)
3320 {
3321 	switch (info->type) {
3322 	case X86_IRQ_ALLOC_TYPE_IOAPIC:
3323 		return get_ioapic_devid(info->devid);
3324 	case X86_IRQ_ALLOC_TYPE_HPET:
3325 		return get_hpet_devid(info->devid);
3326 	case X86_IRQ_ALLOC_TYPE_PCI_MSI:
3327 	case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
3328 		return get_device_sbdf_id(msi_desc_to_dev(info->desc));
3329 	default:
3330 		WARN_ON_ONCE(1);
3331 		return -1;
3332 	}
3333 }
3334 
3335 struct irq_remap_ops amd_iommu_irq_ops = {
3336 	.prepare		= amd_iommu_prepare,
3337 	.enable			= amd_iommu_enable,
3338 	.disable		= amd_iommu_disable,
3339 	.reenable		= amd_iommu_reenable,
3340 	.enable_faulting	= amd_iommu_enable_faulting,
3341 };
3342 
3343 static void fill_msi_msg(struct msi_msg *msg, u32 index)
3344 {
3345 	msg->data = index;
3346 	msg->address_lo = 0;
3347 	msg->arch_addr_lo.base_address = X86_MSI_BASE_ADDRESS_LOW;
3348 	msg->address_hi = X86_MSI_BASE_ADDRESS_HIGH;
3349 }
3350 
3351 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3352 				       struct irq_cfg *irq_cfg,
3353 				       struct irq_alloc_info *info,
3354 				       int devid, int index, int sub_handle)
3355 {
3356 	struct irq_2_irte *irte_info = &data->irq_2_irte;
3357 	struct amd_iommu *iommu = data->iommu;
3358 
3359 	if (!iommu)
3360 		return;
3361 
3362 	data->irq_2_irte.devid = devid;
3363 	data->irq_2_irte.index = index + sub_handle;
3364 	iommu->irte_ops->prepare(data->entry, APIC_DELIVERY_MODE_FIXED,
3365 				 apic->dest_mode_logical, irq_cfg->vector,
3366 				 irq_cfg->dest_apicid, devid);
3367 
3368 	switch (info->type) {
3369 	case X86_IRQ_ALLOC_TYPE_IOAPIC:
3370 	case X86_IRQ_ALLOC_TYPE_HPET:
3371 	case X86_IRQ_ALLOC_TYPE_PCI_MSI:
3372 	case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
3373 		fill_msi_msg(&data->msi_entry, irte_info->index);
3374 		break;
3375 
3376 	default:
3377 		BUG_ON(1);
3378 		break;
3379 	}
3380 }
3381 
3382 struct amd_irte_ops irte_32_ops = {
3383 	.prepare = irte_prepare,
3384 	.activate = irte_activate,
3385 	.deactivate = irte_deactivate,
3386 	.set_affinity = irte_set_affinity,
3387 	.set_allocated = irte_set_allocated,
3388 	.is_allocated = irte_is_allocated,
3389 	.clear_allocated = irte_clear_allocated,
3390 };
3391 
3392 struct amd_irte_ops irte_128_ops = {
3393 	.prepare = irte_ga_prepare,
3394 	.activate = irte_ga_activate,
3395 	.deactivate = irte_ga_deactivate,
3396 	.set_affinity = irte_ga_set_affinity,
3397 	.set_allocated = irte_ga_set_allocated,
3398 	.is_allocated = irte_ga_is_allocated,
3399 	.clear_allocated = irte_ga_clear_allocated,
3400 };
3401 
3402 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
3403 			       unsigned int nr_irqs, void *arg)
3404 {
3405 	struct irq_alloc_info *info = arg;
3406 	struct irq_data *irq_data;
3407 	struct amd_ir_data *data = NULL;
3408 	struct amd_iommu *iommu;
3409 	struct irq_cfg *cfg;
3410 	int i, ret, devid, seg, sbdf;
3411 	int index;
3412 
3413 	if (!info)
3414 		return -EINVAL;
3415 	if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_PCI_MSI)
3416 		return -EINVAL;
3417 
3418 	sbdf = get_devid(info);
3419 	if (sbdf < 0)
3420 		return -EINVAL;
3421 
3422 	seg = PCI_SBDF_TO_SEGID(sbdf);
3423 	devid = PCI_SBDF_TO_DEVID(sbdf);
3424 	iommu = __rlookup_amd_iommu(seg, devid);
3425 	if (!iommu)
3426 		return -EINVAL;
3427 
3428 	ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
3429 	if (ret < 0)
3430 		return ret;
3431 
3432 	if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
3433 		struct irq_remap_table *table;
3434 
3435 		table = alloc_irq_table(iommu, devid, NULL);
3436 		if (table) {
3437 			if (!table->min_index) {
3438 				/*
3439 				 * Keep the first 32 indexes free for IOAPIC
3440 				 * interrupts.
3441 				 */
3442 				table->min_index = 32;
3443 				for (i = 0; i < 32; ++i)
3444 					iommu->irte_ops->set_allocated(table, i);
3445 			}
3446 			WARN_ON(table->min_index != 32);
3447 			index = info->ioapic.pin;
3448 		} else {
3449 			index = -ENOMEM;
3450 		}
3451 	} else if (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI ||
3452 		   info->type == X86_IRQ_ALLOC_TYPE_PCI_MSIX) {
3453 		bool align = (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI);
3454 
3455 		index = alloc_irq_index(iommu, devid, nr_irqs, align,
3456 					msi_desc_to_pci_dev(info->desc));
3457 	} else {
3458 		index = alloc_irq_index(iommu, devid, nr_irqs, false, NULL);
3459 	}
3460 
3461 	if (index < 0) {
3462 		pr_warn("Failed to allocate IRTE\n");
3463 		ret = index;
3464 		goto out_free_parent;
3465 	}
3466 
3467 	for (i = 0; i < nr_irqs; i++) {
3468 		irq_data = irq_domain_get_irq_data(domain, virq + i);
3469 		cfg = irq_data ? irqd_cfg(irq_data) : NULL;
3470 		if (!cfg) {
3471 			ret = -EINVAL;
3472 			goto out_free_data;
3473 		}
3474 
3475 		ret = -ENOMEM;
3476 		data = kzalloc(sizeof(*data), GFP_KERNEL);
3477 		if (!data)
3478 			goto out_free_data;
3479 
3480 		if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3481 			data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
3482 		else
3483 			data->entry = kzalloc(sizeof(struct irte_ga),
3484 						     GFP_KERNEL);
3485 		if (!data->entry) {
3486 			kfree(data);
3487 			goto out_free_data;
3488 		}
3489 
3490 		data->iommu = iommu;
3491 		irq_data->hwirq = (devid << 16) + i;
3492 		irq_data->chip_data = data;
3493 		irq_data->chip = &amd_ir_chip;
3494 		irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
3495 		irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
3496 	}
3497 
3498 	return 0;
3499 
3500 out_free_data:
3501 	for (i--; i >= 0; i--) {
3502 		irq_data = irq_domain_get_irq_data(domain, virq + i);
3503 		if (irq_data)
3504 			kfree(irq_data->chip_data);
3505 	}
3506 	for (i = 0; i < nr_irqs; i++)
3507 		free_irte(iommu, devid, index + i);
3508 out_free_parent:
3509 	irq_domain_free_irqs_common(domain, virq, nr_irqs);
3510 	return ret;
3511 }
3512 
3513 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
3514 			       unsigned int nr_irqs)
3515 {
3516 	struct irq_2_irte *irte_info;
3517 	struct irq_data *irq_data;
3518 	struct amd_ir_data *data;
3519 	int i;
3520 
3521 	for (i = 0; i < nr_irqs; i++) {
3522 		irq_data = irq_domain_get_irq_data(domain, virq  + i);
3523 		if (irq_data && irq_data->chip_data) {
3524 			data = irq_data->chip_data;
3525 			irte_info = &data->irq_2_irte;
3526 			free_irte(data->iommu, irte_info->devid, irte_info->index);
3527 			kfree(data->entry);
3528 			kfree(data);
3529 		}
3530 	}
3531 	irq_domain_free_irqs_common(domain, virq, nr_irqs);
3532 }
3533 
3534 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
3535 			       struct amd_ir_data *ir_data,
3536 			       struct irq_2_irte *irte_info,
3537 			       struct irq_cfg *cfg);
3538 
3539 static int irq_remapping_activate(struct irq_domain *domain,
3540 				  struct irq_data *irq_data, bool reserve)
3541 {
3542 	struct amd_ir_data *data = irq_data->chip_data;
3543 	struct irq_2_irte *irte_info = &data->irq_2_irte;
3544 	struct amd_iommu *iommu = data->iommu;
3545 	struct irq_cfg *cfg = irqd_cfg(irq_data);
3546 
3547 	if (!iommu)
3548 		return 0;
3549 
3550 	iommu->irte_ops->activate(iommu, data->entry, irte_info->devid,
3551 				  irte_info->index);
3552 	amd_ir_update_irte(irq_data, iommu, data, irte_info, cfg);
3553 	return 0;
3554 }
3555 
3556 static void irq_remapping_deactivate(struct irq_domain *domain,
3557 				     struct irq_data *irq_data)
3558 {
3559 	struct amd_ir_data *data = irq_data->chip_data;
3560 	struct irq_2_irte *irte_info = &data->irq_2_irte;
3561 	struct amd_iommu *iommu = data->iommu;
3562 
3563 	if (iommu)
3564 		iommu->irte_ops->deactivate(iommu, data->entry, irte_info->devid,
3565 					    irte_info->index);
3566 }
3567 
3568 static int irq_remapping_select(struct irq_domain *d, struct irq_fwspec *fwspec,
3569 				enum irq_domain_bus_token bus_token)
3570 {
3571 	struct amd_iommu *iommu;
3572 	int devid = -1;
3573 
3574 	if (!amd_iommu_irq_remap)
3575 		return 0;
3576 
3577 	if (x86_fwspec_is_ioapic(fwspec))
3578 		devid = get_ioapic_devid(fwspec->param[0]);
3579 	else if (x86_fwspec_is_hpet(fwspec))
3580 		devid = get_hpet_devid(fwspec->param[0]);
3581 
3582 	if (devid < 0)
3583 		return 0;
3584 	iommu = __rlookup_amd_iommu((devid >> 16), (devid & 0xffff));
3585 
3586 	return iommu && iommu->ir_domain == d;
3587 }
3588 
3589 static const struct irq_domain_ops amd_ir_domain_ops = {
3590 	.select = irq_remapping_select,
3591 	.alloc = irq_remapping_alloc,
3592 	.free = irq_remapping_free,
3593 	.activate = irq_remapping_activate,
3594 	.deactivate = irq_remapping_deactivate,
3595 };
3596 
3597 int amd_iommu_activate_guest_mode(void *data)
3598 {
3599 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3600 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3601 	u64 valid;
3602 
3603 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) || !entry)
3604 		return 0;
3605 
3606 	valid = entry->lo.fields_vapic.valid;
3607 
3608 	entry->lo.val = 0;
3609 	entry->hi.val = 0;
3610 
3611 	entry->lo.fields_vapic.valid       = valid;
3612 	entry->lo.fields_vapic.guest_mode  = 1;
3613 	entry->lo.fields_vapic.ga_log_intr = 1;
3614 	entry->hi.fields.ga_root_ptr       = ir_data->ga_root_ptr;
3615 	entry->hi.fields.vector            = ir_data->ga_vector;
3616 	entry->lo.fields_vapic.ga_tag      = ir_data->ga_tag;
3617 
3618 	return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3619 			      ir_data->irq_2_irte.index, entry);
3620 }
3621 EXPORT_SYMBOL(amd_iommu_activate_guest_mode);
3622 
3623 int amd_iommu_deactivate_guest_mode(void *data)
3624 {
3625 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3626 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3627 	struct irq_cfg *cfg = ir_data->cfg;
3628 	u64 valid;
3629 
3630 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3631 	    !entry || !entry->lo.fields_vapic.guest_mode)
3632 		return 0;
3633 
3634 	valid = entry->lo.fields_remap.valid;
3635 
3636 	entry->lo.val = 0;
3637 	entry->hi.val = 0;
3638 
3639 	entry->lo.fields_remap.valid       = valid;
3640 	entry->lo.fields_remap.dm          = apic->dest_mode_logical;
3641 	entry->lo.fields_remap.int_type    = APIC_DELIVERY_MODE_FIXED;
3642 	entry->hi.fields.vector            = cfg->vector;
3643 	entry->lo.fields_remap.destination =
3644 				APICID_TO_IRTE_DEST_LO(cfg->dest_apicid);
3645 	entry->hi.fields.destination =
3646 				APICID_TO_IRTE_DEST_HI(cfg->dest_apicid);
3647 
3648 	return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3649 			      ir_data->irq_2_irte.index, entry);
3650 }
3651 EXPORT_SYMBOL(amd_iommu_deactivate_guest_mode);
3652 
3653 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
3654 {
3655 	int ret;
3656 	struct amd_iommu_pi_data *pi_data = vcpu_info;
3657 	struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
3658 	struct amd_ir_data *ir_data = data->chip_data;
3659 	struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3660 	struct iommu_dev_data *dev_data;
3661 
3662 	if (ir_data->iommu == NULL)
3663 		return -EINVAL;
3664 
3665 	dev_data = search_dev_data(ir_data->iommu, irte_info->devid);
3666 
3667 	/* Note:
3668 	 * This device has never been set up for guest mode.
3669 	 * we should not modify the IRTE
3670 	 */
3671 	if (!dev_data || !dev_data->use_vapic)
3672 		return 0;
3673 
3674 	ir_data->cfg = irqd_cfg(data);
3675 	pi_data->ir_data = ir_data;
3676 
3677 	/* Note:
3678 	 * SVM tries to set up for VAPIC mode, but we are in
3679 	 * legacy mode. So, we force legacy mode instead.
3680 	 */
3681 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3682 		pr_debug("%s: Fall back to using intr legacy remap\n",
3683 			 __func__);
3684 		pi_data->is_guest_mode = false;
3685 	}
3686 
3687 	pi_data->prev_ga_tag = ir_data->cached_ga_tag;
3688 	if (pi_data->is_guest_mode) {
3689 		ir_data->ga_root_ptr = (pi_data->base >> 12);
3690 		ir_data->ga_vector = vcpu_pi_info->vector;
3691 		ir_data->ga_tag = pi_data->ga_tag;
3692 		ret = amd_iommu_activate_guest_mode(ir_data);
3693 		if (!ret)
3694 			ir_data->cached_ga_tag = pi_data->ga_tag;
3695 	} else {
3696 		ret = amd_iommu_deactivate_guest_mode(ir_data);
3697 
3698 		/*
3699 		 * This communicates the ga_tag back to the caller
3700 		 * so that it can do all the necessary clean up.
3701 		 */
3702 		if (!ret)
3703 			ir_data->cached_ga_tag = 0;
3704 	}
3705 
3706 	return ret;
3707 }
3708 
3709 
3710 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
3711 			       struct amd_ir_data *ir_data,
3712 			       struct irq_2_irte *irte_info,
3713 			       struct irq_cfg *cfg)
3714 {
3715 
3716 	/*
3717 	 * Atomically updates the IRTE with the new destination, vector
3718 	 * and flushes the interrupt entry cache.
3719 	 */
3720 	iommu->irte_ops->set_affinity(iommu, ir_data->entry, irte_info->devid,
3721 				      irte_info->index, cfg->vector,
3722 				      cfg->dest_apicid);
3723 }
3724 
3725 static int amd_ir_set_affinity(struct irq_data *data,
3726 			       const struct cpumask *mask, bool force)
3727 {
3728 	struct amd_ir_data *ir_data = data->chip_data;
3729 	struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3730 	struct irq_cfg *cfg = irqd_cfg(data);
3731 	struct irq_data *parent = data->parent_data;
3732 	struct amd_iommu *iommu = ir_data->iommu;
3733 	int ret;
3734 
3735 	if (!iommu)
3736 		return -ENODEV;
3737 
3738 	ret = parent->chip->irq_set_affinity(parent, mask, force);
3739 	if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
3740 		return ret;
3741 
3742 	amd_ir_update_irte(data, iommu, ir_data, irte_info, cfg);
3743 	/*
3744 	 * After this point, all the interrupts will start arriving
3745 	 * at the new destination. So, time to cleanup the previous
3746 	 * vector allocation.
3747 	 */
3748 	vector_schedule_cleanup(cfg);
3749 
3750 	return IRQ_SET_MASK_OK_DONE;
3751 }
3752 
3753 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
3754 {
3755 	struct amd_ir_data *ir_data = irq_data->chip_data;
3756 
3757 	*msg = ir_data->msi_entry;
3758 }
3759 
3760 static struct irq_chip amd_ir_chip = {
3761 	.name			= "AMD-IR",
3762 	.irq_ack		= apic_ack_irq,
3763 	.irq_set_affinity	= amd_ir_set_affinity,
3764 	.irq_set_vcpu_affinity	= amd_ir_set_vcpu_affinity,
3765 	.irq_compose_msi_msg	= ir_compose_msi_msg,
3766 };
3767 
3768 static const struct msi_parent_ops amdvi_msi_parent_ops = {
3769 	.supported_flags	= X86_VECTOR_MSI_FLAGS_SUPPORTED | MSI_FLAG_MULTI_PCI_MSI,
3770 	.prefix			= "IR-",
3771 	.init_dev_msi_info	= msi_parent_init_dev_msi_info,
3772 };
3773 
3774 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
3775 {
3776 	struct fwnode_handle *fn;
3777 
3778 	fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index);
3779 	if (!fn)
3780 		return -ENOMEM;
3781 	iommu->ir_domain = irq_domain_create_hierarchy(arch_get_ir_parent_domain(), 0, 0,
3782 						       fn, &amd_ir_domain_ops, iommu);
3783 	if (!iommu->ir_domain) {
3784 		irq_domain_free_fwnode(fn);
3785 		return -ENOMEM;
3786 	}
3787 
3788 	irq_domain_update_bus_token(iommu->ir_domain,  DOMAIN_BUS_AMDVI);
3789 	iommu->ir_domain->flags |= IRQ_DOMAIN_FLAG_MSI_PARENT |
3790 				   IRQ_DOMAIN_FLAG_ISOLATED_MSI;
3791 	iommu->ir_domain->msi_parent_ops = &amdvi_msi_parent_ops;
3792 
3793 	return 0;
3794 }
3795 
3796 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
3797 {
3798 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3799 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3800 
3801 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3802 	    !entry || !entry->lo.fields_vapic.guest_mode)
3803 		return 0;
3804 
3805 	if (!ir_data->iommu)
3806 		return -ENODEV;
3807 
3808 	if (cpu >= 0) {
3809 		entry->lo.fields_vapic.destination =
3810 					APICID_TO_IRTE_DEST_LO(cpu);
3811 		entry->hi.fields.destination =
3812 					APICID_TO_IRTE_DEST_HI(cpu);
3813 	}
3814 	entry->lo.fields_vapic.is_run = is_run;
3815 
3816 	return __modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3817 				ir_data->irq_2_irte.index, entry);
3818 }
3819 EXPORT_SYMBOL(amd_iommu_update_ga);
3820 #endif
3821