xref: /linux/drivers/iommu/amd/iommu.c (revision 2f804aca48322f02a8f44cca540663845ee80fb1)
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 
41 #include "amd_iommu.h"
42 #include "../dma-iommu.h"
43 #include "../irq_remapping.h"
44 
45 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
46 
47 #define LOOP_TIMEOUT	100000
48 
49 /* IO virtual address start page frame number */
50 #define IOVA_START_PFN		(1)
51 #define IOVA_PFN(addr)		((addr) >> PAGE_SHIFT)
52 
53 /* Reserved IOVA ranges */
54 #define MSI_RANGE_START		(0xfee00000)
55 #define MSI_RANGE_END		(0xfeefffff)
56 #define HT_RANGE_START		(0xfd00000000ULL)
57 #define HT_RANGE_END		(0xffffffffffULL)
58 
59 #define DEFAULT_PGTABLE_LEVEL	PAGE_MODE_3_LEVEL
60 
61 static DEFINE_SPINLOCK(pd_bitmap_lock);
62 
63 LIST_HEAD(ioapic_map);
64 LIST_HEAD(hpet_map);
65 LIST_HEAD(acpihid_map);
66 
67 const struct iommu_ops amd_iommu_ops;
68 
69 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
70 int amd_iommu_max_glx_val = -1;
71 
72 /*
73  * general struct to manage commands send to an IOMMU
74  */
75 struct iommu_cmd {
76 	u32 data[4];
77 };
78 
79 struct kmem_cache *amd_iommu_irq_cache;
80 
81 static void detach_device(struct device *dev);
82 static int domain_enable_v2(struct protection_domain *domain, int pasids);
83 
84 /****************************************************************************
85  *
86  * Helper functions
87  *
88  ****************************************************************************/
89 
90 static inline int get_acpihid_device_id(struct device *dev,
91 					struct acpihid_map_entry **entry)
92 {
93 	struct acpi_device *adev = ACPI_COMPANION(dev);
94 	struct acpihid_map_entry *p;
95 
96 	if (!adev)
97 		return -ENODEV;
98 
99 	list_for_each_entry(p, &acpihid_map, list) {
100 		if (acpi_dev_hid_uid_match(adev, p->hid,
101 					   p->uid[0] ? p->uid : NULL)) {
102 			if (entry)
103 				*entry = p;
104 			return p->devid;
105 		}
106 	}
107 	return -EINVAL;
108 }
109 
110 static inline int get_device_sbdf_id(struct device *dev)
111 {
112 	int sbdf;
113 
114 	if (dev_is_pci(dev))
115 		sbdf = get_pci_sbdf_id(to_pci_dev(dev));
116 	else
117 		sbdf = get_acpihid_device_id(dev, NULL);
118 
119 	return sbdf;
120 }
121 
122 struct dev_table_entry *get_dev_table(struct amd_iommu *iommu)
123 {
124 	struct dev_table_entry *dev_table;
125 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
126 
127 	BUG_ON(pci_seg == NULL);
128 	dev_table = pci_seg->dev_table;
129 	BUG_ON(dev_table == NULL);
130 
131 	return dev_table;
132 }
133 
134 static inline u16 get_device_segment(struct device *dev)
135 {
136 	u16 seg;
137 
138 	if (dev_is_pci(dev)) {
139 		struct pci_dev *pdev = to_pci_dev(dev);
140 
141 		seg = pci_domain_nr(pdev->bus);
142 	} else {
143 		u32 devid = get_acpihid_device_id(dev, NULL);
144 
145 		seg = PCI_SBDF_TO_SEGID(devid);
146 	}
147 
148 	return seg;
149 }
150 
151 /* Writes the specific IOMMU for a device into the PCI segment rlookup table */
152 void amd_iommu_set_rlookup_table(struct amd_iommu *iommu, u16 devid)
153 {
154 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
155 
156 	pci_seg->rlookup_table[devid] = iommu;
157 }
158 
159 static struct amd_iommu *__rlookup_amd_iommu(u16 seg, u16 devid)
160 {
161 	struct amd_iommu_pci_seg *pci_seg;
162 
163 	for_each_pci_segment(pci_seg) {
164 		if (pci_seg->id == seg)
165 			return pci_seg->rlookup_table[devid];
166 	}
167 	return NULL;
168 }
169 
170 static struct amd_iommu *rlookup_amd_iommu(struct device *dev)
171 {
172 	u16 seg = get_device_segment(dev);
173 	int devid = get_device_sbdf_id(dev);
174 
175 	if (devid < 0)
176 		return NULL;
177 	return __rlookup_amd_iommu(seg, PCI_SBDF_TO_DEVID(devid));
178 }
179 
180 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
181 {
182 	return container_of(dom, struct protection_domain, domain);
183 }
184 
185 static struct iommu_dev_data *alloc_dev_data(struct amd_iommu *iommu, u16 devid)
186 {
187 	struct iommu_dev_data *dev_data;
188 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
189 
190 	dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
191 	if (!dev_data)
192 		return NULL;
193 
194 	spin_lock_init(&dev_data->lock);
195 	dev_data->devid = devid;
196 	ratelimit_default_init(&dev_data->rs);
197 
198 	llist_add(&dev_data->dev_data_list, &pci_seg->dev_data_list);
199 	return dev_data;
200 }
201 
202 static struct iommu_dev_data *search_dev_data(struct amd_iommu *iommu, u16 devid)
203 {
204 	struct iommu_dev_data *dev_data;
205 	struct llist_node *node;
206 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
207 
208 	if (llist_empty(&pci_seg->dev_data_list))
209 		return NULL;
210 
211 	node = pci_seg->dev_data_list.first;
212 	llist_for_each_entry(dev_data, node, dev_data_list) {
213 		if (dev_data->devid == devid)
214 			return dev_data;
215 	}
216 
217 	return NULL;
218 }
219 
220 static int clone_alias(struct pci_dev *pdev, u16 alias, void *data)
221 {
222 	struct amd_iommu *iommu;
223 	struct dev_table_entry *dev_table;
224 	u16 devid = pci_dev_id(pdev);
225 
226 	if (devid == alias)
227 		return 0;
228 
229 	iommu = rlookup_amd_iommu(&pdev->dev);
230 	if (!iommu)
231 		return 0;
232 
233 	amd_iommu_set_rlookup_table(iommu, alias);
234 	dev_table = get_dev_table(iommu);
235 	memcpy(dev_table[alias].data,
236 	       dev_table[devid].data,
237 	       sizeof(dev_table[alias].data));
238 
239 	return 0;
240 }
241 
242 static void clone_aliases(struct amd_iommu *iommu, struct device *dev)
243 {
244 	struct pci_dev *pdev;
245 
246 	if (!dev_is_pci(dev))
247 		return;
248 	pdev = to_pci_dev(dev);
249 
250 	/*
251 	 * The IVRS alias stored in the alias table may not be
252 	 * part of the PCI DMA aliases if it's bus differs
253 	 * from the original device.
254 	 */
255 	clone_alias(pdev, iommu->pci_seg->alias_table[pci_dev_id(pdev)], NULL);
256 
257 	pci_for_each_dma_alias(pdev, clone_alias, NULL);
258 }
259 
260 static void setup_aliases(struct amd_iommu *iommu, struct device *dev)
261 {
262 	struct pci_dev *pdev = to_pci_dev(dev);
263 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
264 	u16 ivrs_alias;
265 
266 	/* For ACPI HID devices, there are no aliases */
267 	if (!dev_is_pci(dev))
268 		return;
269 
270 	/*
271 	 * Add the IVRS alias to the pci aliases if it is on the same
272 	 * bus. The IVRS table may know about a quirk that we don't.
273 	 */
274 	ivrs_alias = pci_seg->alias_table[pci_dev_id(pdev)];
275 	if (ivrs_alias != pci_dev_id(pdev) &&
276 	    PCI_BUS_NUM(ivrs_alias) == pdev->bus->number)
277 		pci_add_dma_alias(pdev, ivrs_alias & 0xff, 1);
278 
279 	clone_aliases(iommu, dev);
280 }
281 
282 static struct iommu_dev_data *find_dev_data(struct amd_iommu *iommu, u16 devid)
283 {
284 	struct iommu_dev_data *dev_data;
285 
286 	dev_data = search_dev_data(iommu, devid);
287 
288 	if (dev_data == NULL) {
289 		dev_data = alloc_dev_data(iommu, devid);
290 		if (!dev_data)
291 			return NULL;
292 
293 		if (translation_pre_enabled(iommu))
294 			dev_data->defer_attach = true;
295 	}
296 
297 	return dev_data;
298 }
299 
300 /*
301 * Find or create an IOMMU group for a acpihid device.
302 */
303 static struct iommu_group *acpihid_device_group(struct device *dev)
304 {
305 	struct acpihid_map_entry *p, *entry = NULL;
306 	int devid;
307 
308 	devid = get_acpihid_device_id(dev, &entry);
309 	if (devid < 0)
310 		return ERR_PTR(devid);
311 
312 	list_for_each_entry(p, &acpihid_map, list) {
313 		if ((devid == p->devid) && p->group)
314 			entry->group = p->group;
315 	}
316 
317 	if (!entry->group)
318 		entry->group = generic_device_group(dev);
319 	else
320 		iommu_group_ref_get(entry->group);
321 
322 	return entry->group;
323 }
324 
325 static bool pci_iommuv2_capable(struct pci_dev *pdev)
326 {
327 	static const int caps[] = {
328 		PCI_EXT_CAP_ID_PRI,
329 		PCI_EXT_CAP_ID_PASID,
330 	};
331 	int i, pos;
332 
333 	if (!pci_ats_supported(pdev))
334 		return false;
335 
336 	for (i = 0; i < 2; ++i) {
337 		pos = pci_find_ext_capability(pdev, caps[i]);
338 		if (pos == 0)
339 			return false;
340 	}
341 
342 	return true;
343 }
344 
345 /*
346  * This function checks if the driver got a valid device from the caller to
347  * avoid dereferencing invalid pointers.
348  */
349 static bool check_device(struct device *dev)
350 {
351 	struct amd_iommu_pci_seg *pci_seg;
352 	struct amd_iommu *iommu;
353 	int devid, sbdf;
354 
355 	if (!dev)
356 		return false;
357 
358 	sbdf = get_device_sbdf_id(dev);
359 	if (sbdf < 0)
360 		return false;
361 	devid = PCI_SBDF_TO_DEVID(sbdf);
362 
363 	iommu = rlookup_amd_iommu(dev);
364 	if (!iommu)
365 		return false;
366 
367 	/* Out of our scope? */
368 	pci_seg = iommu->pci_seg;
369 	if (devid > pci_seg->last_bdf)
370 		return false;
371 
372 	return true;
373 }
374 
375 static int iommu_init_device(struct amd_iommu *iommu, struct device *dev)
376 {
377 	struct iommu_dev_data *dev_data;
378 	int devid, sbdf;
379 
380 	if (dev_iommu_priv_get(dev))
381 		return 0;
382 
383 	sbdf = get_device_sbdf_id(dev);
384 	if (sbdf < 0)
385 		return sbdf;
386 
387 	devid = PCI_SBDF_TO_DEVID(sbdf);
388 	dev_data = find_dev_data(iommu, devid);
389 	if (!dev_data)
390 		return -ENOMEM;
391 
392 	dev_data->dev = dev;
393 	setup_aliases(iommu, dev);
394 
395 	/*
396 	 * By default we use passthrough mode for IOMMUv2 capable device.
397 	 * But if amd_iommu=force_isolation is set (e.g. to debug DMA to
398 	 * invalid address), we ignore the capability for the device so
399 	 * it'll be forced to go into translation mode.
400 	 */
401 	if ((iommu_default_passthrough() || !amd_iommu_force_isolation) &&
402 	    dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
403 		dev_data->iommu_v2 = iommu->is_iommu_v2;
404 	}
405 
406 	dev_iommu_priv_set(dev, dev_data);
407 
408 	return 0;
409 }
410 
411 static void iommu_ignore_device(struct amd_iommu *iommu, struct device *dev)
412 {
413 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
414 	struct dev_table_entry *dev_table = get_dev_table(iommu);
415 	int devid, sbdf;
416 
417 	sbdf = get_device_sbdf_id(dev);
418 	if (sbdf < 0)
419 		return;
420 
421 	devid = PCI_SBDF_TO_DEVID(sbdf);
422 	pci_seg->rlookup_table[devid] = NULL;
423 	memset(&dev_table[devid], 0, sizeof(struct dev_table_entry));
424 
425 	setup_aliases(iommu, dev);
426 }
427 
428 static void amd_iommu_uninit_device(struct device *dev)
429 {
430 	struct iommu_dev_data *dev_data;
431 
432 	dev_data = dev_iommu_priv_get(dev);
433 	if (!dev_data)
434 		return;
435 
436 	if (dev_data->domain)
437 		detach_device(dev);
438 
439 	dev_iommu_priv_set(dev, NULL);
440 
441 	/*
442 	 * We keep dev_data around for unplugged devices and reuse it when the
443 	 * device is re-plugged - not doing so would introduce a ton of races.
444 	 */
445 }
446 
447 /****************************************************************************
448  *
449  * Interrupt handling functions
450  *
451  ****************************************************************************/
452 
453 static void dump_dte_entry(struct amd_iommu *iommu, u16 devid)
454 {
455 	int i;
456 	struct dev_table_entry *dev_table = get_dev_table(iommu);
457 
458 	for (i = 0; i < 4; ++i)
459 		pr_err("DTE[%d]: %016llx\n", i, dev_table[devid].data[i]);
460 }
461 
462 static void dump_command(unsigned long phys_addr)
463 {
464 	struct iommu_cmd *cmd = iommu_phys_to_virt(phys_addr);
465 	int i;
466 
467 	for (i = 0; i < 4; ++i)
468 		pr_err("CMD[%d]: %08x\n", i, cmd->data[i]);
469 }
470 
471 static void amd_iommu_report_rmp_hw_error(struct amd_iommu *iommu, volatile u32 *event)
472 {
473 	struct iommu_dev_data *dev_data = NULL;
474 	int devid, vmg_tag, flags;
475 	struct pci_dev *pdev;
476 	u64 spa;
477 
478 	devid   = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
479 	vmg_tag = (event[1]) & 0xFFFF;
480 	flags   = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
481 	spa     = ((u64)event[3] << 32) | (event[2] & 0xFFFFFFF8);
482 
483 	pdev = pci_get_domain_bus_and_slot(iommu->pci_seg->id, PCI_BUS_NUM(devid),
484 					   devid & 0xff);
485 	if (pdev)
486 		dev_data = dev_iommu_priv_get(&pdev->dev);
487 
488 	if (dev_data) {
489 		if (__ratelimit(&dev_data->rs)) {
490 			pci_err(pdev, "Event logged [RMP_HW_ERROR vmg_tag=0x%04x, spa=0x%llx, flags=0x%04x]\n",
491 				vmg_tag, spa, flags);
492 		}
493 	} else {
494 		pr_err_ratelimited("Event logged [RMP_HW_ERROR device=%04x:%02x:%02x.%x, vmg_tag=0x%04x, spa=0x%llx, flags=0x%04x]\n",
495 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
496 			vmg_tag, spa, flags);
497 	}
498 
499 	if (pdev)
500 		pci_dev_put(pdev);
501 }
502 
503 static void amd_iommu_report_rmp_fault(struct amd_iommu *iommu, volatile u32 *event)
504 {
505 	struct iommu_dev_data *dev_data = NULL;
506 	int devid, flags_rmp, vmg_tag, flags;
507 	struct pci_dev *pdev;
508 	u64 gpa;
509 
510 	devid     = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
511 	flags_rmp = (event[0] >> EVENT_FLAGS_SHIFT) & 0xFF;
512 	vmg_tag   = (event[1]) & 0xFFFF;
513 	flags     = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
514 	gpa       = ((u64)event[3] << 32) | event[2];
515 
516 	pdev = pci_get_domain_bus_and_slot(iommu->pci_seg->id, PCI_BUS_NUM(devid),
517 					   devid & 0xff);
518 	if (pdev)
519 		dev_data = dev_iommu_priv_get(&pdev->dev);
520 
521 	if (dev_data) {
522 		if (__ratelimit(&dev_data->rs)) {
523 			pci_err(pdev, "Event logged [RMP_PAGE_FAULT vmg_tag=0x%04x, gpa=0x%llx, flags_rmp=0x%04x, flags=0x%04x]\n",
524 				vmg_tag, gpa, flags_rmp, flags);
525 		}
526 	} else {
527 		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",
528 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
529 			vmg_tag, gpa, flags_rmp, flags);
530 	}
531 
532 	if (pdev)
533 		pci_dev_put(pdev);
534 }
535 
536 #define IS_IOMMU_MEM_TRANSACTION(flags)		\
537 	(((flags) & EVENT_FLAG_I) == 0)
538 
539 #define IS_WRITE_REQUEST(flags)			\
540 	((flags) & EVENT_FLAG_RW)
541 
542 static void amd_iommu_report_page_fault(struct amd_iommu *iommu,
543 					u16 devid, u16 domain_id,
544 					u64 address, int flags)
545 {
546 	struct iommu_dev_data *dev_data = NULL;
547 	struct pci_dev *pdev;
548 
549 	pdev = pci_get_domain_bus_and_slot(iommu->pci_seg->id, PCI_BUS_NUM(devid),
550 					   devid & 0xff);
551 	if (pdev)
552 		dev_data = dev_iommu_priv_get(&pdev->dev);
553 
554 	if (dev_data) {
555 		/*
556 		 * If this is a DMA fault (for which the I(nterrupt)
557 		 * bit will be unset), allow report_iommu_fault() to
558 		 * prevent logging it.
559 		 */
560 		if (IS_IOMMU_MEM_TRANSACTION(flags)) {
561 			/* Device not attached to domain properly */
562 			if (dev_data->domain == NULL) {
563 				pr_err_ratelimited("Event logged [Device not attached to domain properly]\n");
564 				pr_err_ratelimited("  device=%04x:%02x:%02x.%x domain=0x%04x\n",
565 						   iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid),
566 						   PCI_FUNC(devid), domain_id);
567 				goto out;
568 			}
569 
570 			if (!report_iommu_fault(&dev_data->domain->domain,
571 						&pdev->dev, address,
572 						IS_WRITE_REQUEST(flags) ?
573 							IOMMU_FAULT_WRITE :
574 							IOMMU_FAULT_READ))
575 				goto out;
576 		}
577 
578 		if (__ratelimit(&dev_data->rs)) {
579 			pci_err(pdev, "Event logged [IO_PAGE_FAULT domain=0x%04x address=0x%llx flags=0x%04x]\n",
580 				domain_id, address, flags);
581 		}
582 	} else {
583 		pr_err_ratelimited("Event logged [IO_PAGE_FAULT device=%04x:%02x:%02x.%x domain=0x%04x address=0x%llx flags=0x%04x]\n",
584 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
585 			domain_id, address, flags);
586 	}
587 
588 out:
589 	if (pdev)
590 		pci_dev_put(pdev);
591 }
592 
593 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
594 {
595 	struct device *dev = iommu->iommu.dev;
596 	int type, devid, flags, tag;
597 	volatile u32 *event = __evt;
598 	int count = 0;
599 	u64 address;
600 	u32 pasid;
601 
602 retry:
603 	type    = (event[1] >> EVENT_TYPE_SHIFT)  & EVENT_TYPE_MASK;
604 	devid   = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
605 	pasid   = (event[0] & EVENT_DOMID_MASK_HI) |
606 		  (event[1] & EVENT_DOMID_MASK_LO);
607 	flags   = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
608 	address = (u64)(((u64)event[3]) << 32) | event[2];
609 
610 	if (type == 0) {
611 		/* Did we hit the erratum? */
612 		if (++count == LOOP_TIMEOUT) {
613 			pr_err("No event written to event log\n");
614 			return;
615 		}
616 		udelay(1);
617 		goto retry;
618 	}
619 
620 	if (type == EVENT_TYPE_IO_FAULT) {
621 		amd_iommu_report_page_fault(iommu, devid, pasid, address, flags);
622 		return;
623 	}
624 
625 	switch (type) {
626 	case EVENT_TYPE_ILL_DEV:
627 		dev_err(dev, "Event logged [ILLEGAL_DEV_TABLE_ENTRY device=%04x:%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
628 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
629 			pasid, address, flags);
630 		dump_dte_entry(iommu, devid);
631 		break;
632 	case EVENT_TYPE_DEV_TAB_ERR:
633 		dev_err(dev, "Event logged [DEV_TAB_HARDWARE_ERROR device=%04x:%02x:%02x.%x "
634 			"address=0x%llx flags=0x%04x]\n",
635 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
636 			address, flags);
637 		break;
638 	case EVENT_TYPE_PAGE_TAB_ERR:
639 		dev_err(dev, "Event logged [PAGE_TAB_HARDWARE_ERROR device=%04x:%02x:%02x.%x pasid=0x%04x address=0x%llx flags=0x%04x]\n",
640 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
641 			pasid, address, flags);
642 		break;
643 	case EVENT_TYPE_ILL_CMD:
644 		dev_err(dev, "Event logged [ILLEGAL_COMMAND_ERROR address=0x%llx]\n", address);
645 		dump_command(address);
646 		break;
647 	case EVENT_TYPE_CMD_HARD_ERR:
648 		dev_err(dev, "Event logged [COMMAND_HARDWARE_ERROR address=0x%llx flags=0x%04x]\n",
649 			address, flags);
650 		break;
651 	case EVENT_TYPE_IOTLB_INV_TO:
652 		dev_err(dev, "Event logged [IOTLB_INV_TIMEOUT device=%04x:%02x:%02x.%x address=0x%llx]\n",
653 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
654 			address);
655 		break;
656 	case EVENT_TYPE_INV_DEV_REQ:
657 		dev_err(dev, "Event logged [INVALID_DEVICE_REQUEST device=%04x:%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
658 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
659 			pasid, address, flags);
660 		break;
661 	case EVENT_TYPE_RMP_FAULT:
662 		amd_iommu_report_rmp_fault(iommu, event);
663 		break;
664 	case EVENT_TYPE_RMP_HW_ERR:
665 		amd_iommu_report_rmp_hw_error(iommu, event);
666 		break;
667 	case EVENT_TYPE_INV_PPR_REQ:
668 		pasid = PPR_PASID(*((u64 *)__evt));
669 		tag = event[1] & 0x03FF;
670 		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",
671 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
672 			pasid, address, flags, tag);
673 		break;
674 	default:
675 		dev_err(dev, "Event logged [UNKNOWN event[0]=0x%08x event[1]=0x%08x event[2]=0x%08x event[3]=0x%08x\n",
676 			event[0], event[1], event[2], event[3]);
677 	}
678 
679 	/*
680 	 * To detect the hardware errata 732 we need to clear the
681 	 * entry back to zero. This issue does not exist on SNP
682 	 * enabled system. Also this buffer is not writeable on
683 	 * SNP enabled system.
684 	 */
685 	if (!amd_iommu_snp_en)
686 		memset(__evt, 0, 4 * sizeof(u32));
687 }
688 
689 static void iommu_poll_events(struct amd_iommu *iommu)
690 {
691 	u32 head, tail;
692 
693 	head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
694 	tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
695 
696 	while (head != tail) {
697 		iommu_print_event(iommu, iommu->evt_buf + head);
698 		head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
699 	}
700 
701 	writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
702 }
703 
704 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
705 {
706 	struct amd_iommu_fault fault;
707 
708 	if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
709 		pr_err_ratelimited("Unknown PPR request received\n");
710 		return;
711 	}
712 
713 	fault.address   = raw[1];
714 	fault.pasid     = PPR_PASID(raw[0]);
715 	fault.sbdf      = PCI_SEG_DEVID_TO_SBDF(iommu->pci_seg->id, PPR_DEVID(raw[0]));
716 	fault.tag       = PPR_TAG(raw[0]);
717 	fault.flags     = PPR_FLAGS(raw[0]);
718 
719 	atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
720 }
721 
722 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
723 {
724 	u32 head, tail;
725 
726 	if (iommu->ppr_log == NULL)
727 		return;
728 
729 	head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
730 	tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
731 
732 	while (head != tail) {
733 		volatile u64 *raw;
734 		u64 entry[2];
735 		int i;
736 
737 		raw = (u64 *)(iommu->ppr_log + head);
738 
739 		/*
740 		 * Hardware bug: Interrupt may arrive before the entry is
741 		 * written to memory. If this happens we need to wait for the
742 		 * entry to arrive.
743 		 */
744 		for (i = 0; i < LOOP_TIMEOUT; ++i) {
745 			if (PPR_REQ_TYPE(raw[0]) != 0)
746 				break;
747 			udelay(1);
748 		}
749 
750 		/* Avoid memcpy function-call overhead */
751 		entry[0] = raw[0];
752 		entry[1] = raw[1];
753 
754 		/*
755 		 * To detect the hardware errata 733 we need to clear the
756 		 * entry back to zero. This issue does not exist on SNP
757 		 * enabled system. Also this buffer is not writeable on
758 		 * SNP enabled system.
759 		 */
760 		if (!amd_iommu_snp_en)
761 			raw[0] = raw[1] = 0UL;
762 
763 		/* Update head pointer of hardware ring-buffer */
764 		head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
765 		writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
766 
767 		/* Handle PPR entry */
768 		iommu_handle_ppr_entry(iommu, entry);
769 
770 		/* Refresh ring-buffer information */
771 		head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
772 		tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
773 	}
774 }
775 
776 #ifdef CONFIG_IRQ_REMAP
777 static int (*iommu_ga_log_notifier)(u32);
778 
779 int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
780 {
781 	iommu_ga_log_notifier = notifier;
782 
783 	return 0;
784 }
785 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
786 
787 static void iommu_poll_ga_log(struct amd_iommu *iommu)
788 {
789 	u32 head, tail;
790 
791 	if (iommu->ga_log == NULL)
792 		return;
793 
794 	head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
795 	tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
796 
797 	while (head != tail) {
798 		volatile u64 *raw;
799 		u64 log_entry;
800 
801 		raw = (u64 *)(iommu->ga_log + head);
802 
803 		/* Avoid memcpy function-call overhead */
804 		log_entry = *raw;
805 
806 		/* Update head pointer of hardware ring-buffer */
807 		head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
808 		writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
809 
810 		/* Handle GA entry */
811 		switch (GA_REQ_TYPE(log_entry)) {
812 		case GA_GUEST_NR:
813 			if (!iommu_ga_log_notifier)
814 				break;
815 
816 			pr_debug("%s: devid=%#x, ga_tag=%#x\n",
817 				 __func__, GA_DEVID(log_entry),
818 				 GA_TAG(log_entry));
819 
820 			if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
821 				pr_err("GA log notifier failed.\n");
822 			break;
823 		default:
824 			break;
825 		}
826 	}
827 }
828 
829 static void
830 amd_iommu_set_pci_msi_domain(struct device *dev, struct amd_iommu *iommu)
831 {
832 	if (!irq_remapping_enabled || !dev_is_pci(dev) ||
833 	    !pci_dev_has_default_msi_parent_domain(to_pci_dev(dev)))
834 		return;
835 
836 	dev_set_msi_domain(dev, iommu->ir_domain);
837 }
838 
839 #else /* CONFIG_IRQ_REMAP */
840 static inline void
841 amd_iommu_set_pci_msi_domain(struct device *dev, struct amd_iommu *iommu) { }
842 #endif /* !CONFIG_IRQ_REMAP */
843 
844 #define AMD_IOMMU_INT_MASK	\
845 	(MMIO_STATUS_EVT_OVERFLOW_INT_MASK | \
846 	 MMIO_STATUS_EVT_INT_MASK | \
847 	 MMIO_STATUS_PPR_INT_MASK | \
848 	 MMIO_STATUS_GALOG_INT_MASK)
849 
850 irqreturn_t amd_iommu_int_thread(int irq, void *data)
851 {
852 	struct amd_iommu *iommu = (struct amd_iommu *) data;
853 	u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
854 
855 	while (status & AMD_IOMMU_INT_MASK) {
856 		/* Enable interrupt sources again */
857 		writel(AMD_IOMMU_INT_MASK,
858 			iommu->mmio_base + MMIO_STATUS_OFFSET);
859 
860 		if (status & MMIO_STATUS_EVT_INT_MASK) {
861 			pr_devel("Processing IOMMU Event Log\n");
862 			iommu_poll_events(iommu);
863 		}
864 
865 		if (status & MMIO_STATUS_PPR_INT_MASK) {
866 			pr_devel("Processing IOMMU PPR Log\n");
867 			iommu_poll_ppr_log(iommu);
868 		}
869 
870 #ifdef CONFIG_IRQ_REMAP
871 		if (status & MMIO_STATUS_GALOG_INT_MASK) {
872 			pr_devel("Processing IOMMU GA Log\n");
873 			iommu_poll_ga_log(iommu);
874 		}
875 #endif
876 
877 		if (status & MMIO_STATUS_EVT_OVERFLOW_INT_MASK) {
878 			pr_info_ratelimited("IOMMU event log overflow\n");
879 			amd_iommu_restart_event_logging(iommu);
880 		}
881 
882 		/*
883 		 * Hardware bug: ERBT1312
884 		 * When re-enabling interrupt (by writing 1
885 		 * to clear the bit), the hardware might also try to set
886 		 * the interrupt bit in the event status register.
887 		 * In this scenario, the bit will be set, and disable
888 		 * subsequent interrupts.
889 		 *
890 		 * Workaround: The IOMMU driver should read back the
891 		 * status register and check if the interrupt bits are cleared.
892 		 * If not, driver will need to go through the interrupt handler
893 		 * again and re-clear the bits
894 		 */
895 		status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
896 	}
897 	return IRQ_HANDLED;
898 }
899 
900 irqreturn_t amd_iommu_int_handler(int irq, void *data)
901 {
902 	return IRQ_WAKE_THREAD;
903 }
904 
905 /****************************************************************************
906  *
907  * IOMMU command queuing functions
908  *
909  ****************************************************************************/
910 
911 static int wait_on_sem(struct amd_iommu *iommu, u64 data)
912 {
913 	int i = 0;
914 
915 	while (*iommu->cmd_sem != data && i < LOOP_TIMEOUT) {
916 		udelay(1);
917 		i += 1;
918 	}
919 
920 	if (i == LOOP_TIMEOUT) {
921 		pr_alert("Completion-Wait loop timed out\n");
922 		return -EIO;
923 	}
924 
925 	return 0;
926 }
927 
928 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
929 			       struct iommu_cmd *cmd)
930 {
931 	u8 *target;
932 	u32 tail;
933 
934 	/* Copy command to buffer */
935 	tail = iommu->cmd_buf_tail;
936 	target = iommu->cmd_buf + tail;
937 	memcpy(target, cmd, sizeof(*cmd));
938 
939 	tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
940 	iommu->cmd_buf_tail = tail;
941 
942 	/* Tell the IOMMU about it */
943 	writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
944 }
945 
946 static void build_completion_wait(struct iommu_cmd *cmd,
947 				  struct amd_iommu *iommu,
948 				  u64 data)
949 {
950 	u64 paddr = iommu_virt_to_phys((void *)iommu->cmd_sem);
951 
952 	memset(cmd, 0, sizeof(*cmd));
953 	cmd->data[0] = lower_32_bits(paddr) | CMD_COMPL_WAIT_STORE_MASK;
954 	cmd->data[1] = upper_32_bits(paddr);
955 	cmd->data[2] = lower_32_bits(data);
956 	cmd->data[3] = upper_32_bits(data);
957 	CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
958 }
959 
960 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
961 {
962 	memset(cmd, 0, sizeof(*cmd));
963 	cmd->data[0] = devid;
964 	CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
965 }
966 
967 /*
968  * Builds an invalidation address which is suitable for one page or multiple
969  * pages. Sets the size bit (S) as needed is more than one page is flushed.
970  */
971 static inline u64 build_inv_address(u64 address, size_t size)
972 {
973 	u64 pages, end, msb_diff;
974 
975 	pages = iommu_num_pages(address, size, PAGE_SIZE);
976 
977 	if (pages == 1)
978 		return address & PAGE_MASK;
979 
980 	end = address + size - 1;
981 
982 	/*
983 	 * msb_diff would hold the index of the most significant bit that
984 	 * flipped between the start and end.
985 	 */
986 	msb_diff = fls64(end ^ address) - 1;
987 
988 	/*
989 	 * Bits 63:52 are sign extended. If for some reason bit 51 is different
990 	 * between the start and the end, invalidate everything.
991 	 */
992 	if (unlikely(msb_diff > 51)) {
993 		address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
994 	} else {
995 		/*
996 		 * The msb-bit must be clear on the address. Just set all the
997 		 * lower bits.
998 		 */
999 		address |= (1ull << msb_diff) - 1;
1000 	}
1001 
1002 	/* Clear bits 11:0 */
1003 	address &= PAGE_MASK;
1004 
1005 	/* Set the size bit - we flush more than one 4kb page */
1006 	return address | CMD_INV_IOMMU_PAGES_SIZE_MASK;
1007 }
1008 
1009 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
1010 				  size_t size, u16 domid, int pde)
1011 {
1012 	u64 inv_address = build_inv_address(address, size);
1013 
1014 	memset(cmd, 0, sizeof(*cmd));
1015 	cmd->data[1] |= domid;
1016 	cmd->data[2]  = lower_32_bits(inv_address);
1017 	cmd->data[3]  = upper_32_bits(inv_address);
1018 	CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
1019 	if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
1020 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
1021 }
1022 
1023 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
1024 				  u64 address, size_t size)
1025 {
1026 	u64 inv_address = build_inv_address(address, size);
1027 
1028 	memset(cmd, 0, sizeof(*cmd));
1029 	cmd->data[0]  = devid;
1030 	cmd->data[0] |= (qdep & 0xff) << 24;
1031 	cmd->data[1]  = devid;
1032 	cmd->data[2]  = lower_32_bits(inv_address);
1033 	cmd->data[3]  = upper_32_bits(inv_address);
1034 	CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
1035 }
1036 
1037 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, u32 pasid,
1038 				  u64 address, bool size)
1039 {
1040 	memset(cmd, 0, sizeof(*cmd));
1041 
1042 	address &= ~(0xfffULL);
1043 
1044 	cmd->data[0]  = pasid;
1045 	cmd->data[1]  = domid;
1046 	cmd->data[2]  = lower_32_bits(address);
1047 	cmd->data[3]  = upper_32_bits(address);
1048 	cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
1049 	cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
1050 	if (size)
1051 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
1052 	CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
1053 }
1054 
1055 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, u32 pasid,
1056 				  int qdep, u64 address, bool size)
1057 {
1058 	memset(cmd, 0, sizeof(*cmd));
1059 
1060 	address &= ~(0xfffULL);
1061 
1062 	cmd->data[0]  = devid;
1063 	cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
1064 	cmd->data[0] |= (qdep  & 0xff) << 24;
1065 	cmd->data[1]  = devid;
1066 	cmd->data[1] |= (pasid & 0xff) << 16;
1067 	cmd->data[2]  = lower_32_bits(address);
1068 	cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
1069 	cmd->data[3]  = upper_32_bits(address);
1070 	if (size)
1071 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
1072 	CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
1073 }
1074 
1075 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, u32 pasid,
1076 			       int status, int tag, bool gn)
1077 {
1078 	memset(cmd, 0, sizeof(*cmd));
1079 
1080 	cmd->data[0]  = devid;
1081 	if (gn) {
1082 		cmd->data[1]  = pasid;
1083 		cmd->data[2]  = CMD_INV_IOMMU_PAGES_GN_MASK;
1084 	}
1085 	cmd->data[3]  = tag & 0x1ff;
1086 	cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
1087 
1088 	CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
1089 }
1090 
1091 static void build_inv_all(struct iommu_cmd *cmd)
1092 {
1093 	memset(cmd, 0, sizeof(*cmd));
1094 	CMD_SET_TYPE(cmd, CMD_INV_ALL);
1095 }
1096 
1097 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1098 {
1099 	memset(cmd, 0, sizeof(*cmd));
1100 	cmd->data[0] = devid;
1101 	CMD_SET_TYPE(cmd, CMD_INV_IRT);
1102 }
1103 
1104 /*
1105  * Writes the command to the IOMMUs command buffer and informs the
1106  * hardware about the new command.
1107  */
1108 static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1109 				      struct iommu_cmd *cmd,
1110 				      bool sync)
1111 {
1112 	unsigned int count = 0;
1113 	u32 left, next_tail;
1114 
1115 	next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1116 again:
1117 	left      = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE;
1118 
1119 	if (left <= 0x20) {
1120 		/* Skip udelay() the first time around */
1121 		if (count++) {
1122 			if (count == LOOP_TIMEOUT) {
1123 				pr_err("Command buffer timeout\n");
1124 				return -EIO;
1125 			}
1126 
1127 			udelay(1);
1128 		}
1129 
1130 		/* Update head and recheck remaining space */
1131 		iommu->cmd_buf_head = readl(iommu->mmio_base +
1132 					    MMIO_CMD_HEAD_OFFSET);
1133 
1134 		goto again;
1135 	}
1136 
1137 	copy_cmd_to_buffer(iommu, cmd);
1138 
1139 	/* Do we need to make sure all commands are processed? */
1140 	iommu->need_sync = sync;
1141 
1142 	return 0;
1143 }
1144 
1145 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1146 				    struct iommu_cmd *cmd,
1147 				    bool sync)
1148 {
1149 	unsigned long flags;
1150 	int ret;
1151 
1152 	raw_spin_lock_irqsave(&iommu->lock, flags);
1153 	ret = __iommu_queue_command_sync(iommu, cmd, sync);
1154 	raw_spin_unlock_irqrestore(&iommu->lock, flags);
1155 
1156 	return ret;
1157 }
1158 
1159 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1160 {
1161 	return iommu_queue_command_sync(iommu, cmd, true);
1162 }
1163 
1164 /*
1165  * This function queues a completion wait command into the command
1166  * buffer of an IOMMU
1167  */
1168 static int iommu_completion_wait(struct amd_iommu *iommu)
1169 {
1170 	struct iommu_cmd cmd;
1171 	unsigned long flags;
1172 	int ret;
1173 	u64 data;
1174 
1175 	if (!iommu->need_sync)
1176 		return 0;
1177 
1178 	raw_spin_lock_irqsave(&iommu->lock, flags);
1179 
1180 	data = ++iommu->cmd_sem_val;
1181 	build_completion_wait(&cmd, iommu, data);
1182 
1183 	ret = __iommu_queue_command_sync(iommu, &cmd, false);
1184 	if (ret)
1185 		goto out_unlock;
1186 
1187 	ret = wait_on_sem(iommu, data);
1188 
1189 out_unlock:
1190 	raw_spin_unlock_irqrestore(&iommu->lock, flags);
1191 
1192 	return ret;
1193 }
1194 
1195 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1196 {
1197 	struct iommu_cmd cmd;
1198 
1199 	build_inv_dte(&cmd, devid);
1200 
1201 	return iommu_queue_command(iommu, &cmd);
1202 }
1203 
1204 static void amd_iommu_flush_dte_all(struct amd_iommu *iommu)
1205 {
1206 	u32 devid;
1207 	u16 last_bdf = iommu->pci_seg->last_bdf;
1208 
1209 	for (devid = 0; devid <= last_bdf; ++devid)
1210 		iommu_flush_dte(iommu, devid);
1211 
1212 	iommu_completion_wait(iommu);
1213 }
1214 
1215 /*
1216  * This function uses heavy locking and may disable irqs for some time. But
1217  * this is no issue because it is only called during resume.
1218  */
1219 static void amd_iommu_flush_tlb_all(struct amd_iommu *iommu)
1220 {
1221 	u32 dom_id;
1222 	u16 last_bdf = iommu->pci_seg->last_bdf;
1223 
1224 	for (dom_id = 0; dom_id <= last_bdf; ++dom_id) {
1225 		struct iommu_cmd cmd;
1226 		build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1227 				      dom_id, 1);
1228 		iommu_queue_command(iommu, &cmd);
1229 	}
1230 
1231 	iommu_completion_wait(iommu);
1232 }
1233 
1234 static void amd_iommu_flush_tlb_domid(struct amd_iommu *iommu, u32 dom_id)
1235 {
1236 	struct iommu_cmd cmd;
1237 
1238 	build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1239 			      dom_id, 1);
1240 	iommu_queue_command(iommu, &cmd);
1241 
1242 	iommu_completion_wait(iommu);
1243 }
1244 
1245 static void amd_iommu_flush_all(struct amd_iommu *iommu)
1246 {
1247 	struct iommu_cmd cmd;
1248 
1249 	build_inv_all(&cmd);
1250 
1251 	iommu_queue_command(iommu, &cmd);
1252 	iommu_completion_wait(iommu);
1253 }
1254 
1255 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1256 {
1257 	struct iommu_cmd cmd;
1258 
1259 	build_inv_irt(&cmd, devid);
1260 
1261 	iommu_queue_command(iommu, &cmd);
1262 }
1263 
1264 static void amd_iommu_flush_irt_all(struct amd_iommu *iommu)
1265 {
1266 	u32 devid;
1267 	u16 last_bdf = iommu->pci_seg->last_bdf;
1268 
1269 	for (devid = 0; devid <= last_bdf; devid++)
1270 		iommu_flush_irt(iommu, devid);
1271 
1272 	iommu_completion_wait(iommu);
1273 }
1274 
1275 void iommu_flush_all_caches(struct amd_iommu *iommu)
1276 {
1277 	if (iommu_feature(iommu, FEATURE_IA)) {
1278 		amd_iommu_flush_all(iommu);
1279 	} else {
1280 		amd_iommu_flush_dte_all(iommu);
1281 		amd_iommu_flush_irt_all(iommu);
1282 		amd_iommu_flush_tlb_all(iommu);
1283 	}
1284 }
1285 
1286 /*
1287  * Command send function for flushing on-device TLB
1288  */
1289 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1290 			      u64 address, size_t size)
1291 {
1292 	struct amd_iommu *iommu;
1293 	struct iommu_cmd cmd;
1294 	int qdep;
1295 
1296 	qdep     = dev_data->ats.qdep;
1297 	iommu    = rlookup_amd_iommu(dev_data->dev);
1298 	if (!iommu)
1299 		return -EINVAL;
1300 
1301 	build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1302 
1303 	return iommu_queue_command(iommu, &cmd);
1304 }
1305 
1306 static int device_flush_dte_alias(struct pci_dev *pdev, u16 alias, void *data)
1307 {
1308 	struct amd_iommu *iommu = data;
1309 
1310 	return iommu_flush_dte(iommu, alias);
1311 }
1312 
1313 /*
1314  * Command send function for invalidating a device table entry
1315  */
1316 static int device_flush_dte(struct iommu_dev_data *dev_data)
1317 {
1318 	struct amd_iommu *iommu;
1319 	struct pci_dev *pdev = NULL;
1320 	struct amd_iommu_pci_seg *pci_seg;
1321 	u16 alias;
1322 	int ret;
1323 
1324 	iommu = rlookup_amd_iommu(dev_data->dev);
1325 	if (!iommu)
1326 		return -EINVAL;
1327 
1328 	if (dev_is_pci(dev_data->dev))
1329 		pdev = to_pci_dev(dev_data->dev);
1330 
1331 	if (pdev)
1332 		ret = pci_for_each_dma_alias(pdev,
1333 					     device_flush_dte_alias, iommu);
1334 	else
1335 		ret = iommu_flush_dte(iommu, dev_data->devid);
1336 	if (ret)
1337 		return ret;
1338 
1339 	pci_seg = iommu->pci_seg;
1340 	alias = pci_seg->alias_table[dev_data->devid];
1341 	if (alias != dev_data->devid) {
1342 		ret = iommu_flush_dte(iommu, alias);
1343 		if (ret)
1344 			return ret;
1345 	}
1346 
1347 	if (dev_data->ats.enabled)
1348 		ret = device_flush_iotlb(dev_data, 0, ~0UL);
1349 
1350 	return ret;
1351 }
1352 
1353 /*
1354  * TLB invalidation function which is called from the mapping functions.
1355  * It invalidates a single PTE if the range to flush is within a single
1356  * page. Otherwise it flushes the whole TLB of the IOMMU.
1357  */
1358 static void __domain_flush_pages(struct protection_domain *domain,
1359 				 u64 address, size_t size, int pde)
1360 {
1361 	struct iommu_dev_data *dev_data;
1362 	struct iommu_cmd cmd;
1363 	int ret = 0, i;
1364 
1365 	build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1366 
1367 	for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1368 		if (!domain->dev_iommu[i])
1369 			continue;
1370 
1371 		/*
1372 		 * Devices of this domain are behind this IOMMU
1373 		 * We need a TLB flush
1374 		 */
1375 		ret |= iommu_queue_command(amd_iommus[i], &cmd);
1376 	}
1377 
1378 	list_for_each_entry(dev_data, &domain->dev_list, list) {
1379 
1380 		if (!dev_data->ats.enabled)
1381 			continue;
1382 
1383 		ret |= device_flush_iotlb(dev_data, address, size);
1384 	}
1385 
1386 	WARN_ON(ret);
1387 }
1388 
1389 static void domain_flush_pages(struct protection_domain *domain,
1390 			       u64 address, size_t size, int pde)
1391 {
1392 	if (likely(!amd_iommu_np_cache)) {
1393 		__domain_flush_pages(domain, address, size, pde);
1394 		return;
1395 	}
1396 
1397 	/*
1398 	 * When NpCache is on, we infer that we run in a VM and use a vIOMMU.
1399 	 * In such setups it is best to avoid flushes of ranges which are not
1400 	 * naturally aligned, since it would lead to flushes of unmodified
1401 	 * PTEs. Such flushes would require the hypervisor to do more work than
1402 	 * necessary. Therefore, perform repeated flushes of aligned ranges
1403 	 * until you cover the range. Each iteration flushes the smaller
1404 	 * between the natural alignment of the address that we flush and the
1405 	 * greatest naturally aligned region that fits in the range.
1406 	 */
1407 	while (size != 0) {
1408 		int addr_alignment = __ffs(address);
1409 		int size_alignment = __fls(size);
1410 		int min_alignment;
1411 		size_t flush_size;
1412 
1413 		/*
1414 		 * size is always non-zero, but address might be zero, causing
1415 		 * addr_alignment to be negative. As the casting of the
1416 		 * argument in __ffs(address) to long might trim the high bits
1417 		 * of the address on x86-32, cast to long when doing the check.
1418 		 */
1419 		if (likely((unsigned long)address != 0))
1420 			min_alignment = min(addr_alignment, size_alignment);
1421 		else
1422 			min_alignment = size_alignment;
1423 
1424 		flush_size = 1ul << min_alignment;
1425 
1426 		__domain_flush_pages(domain, address, flush_size, pde);
1427 		address += flush_size;
1428 		size -= flush_size;
1429 	}
1430 }
1431 
1432 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1433 void amd_iommu_domain_flush_tlb_pde(struct protection_domain *domain)
1434 {
1435 	domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1436 }
1437 
1438 void amd_iommu_domain_flush_complete(struct protection_domain *domain)
1439 {
1440 	int i;
1441 
1442 	for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1443 		if (domain && !domain->dev_iommu[i])
1444 			continue;
1445 
1446 		/*
1447 		 * Devices of this domain are behind this IOMMU
1448 		 * We need to wait for completion of all commands.
1449 		 */
1450 		iommu_completion_wait(amd_iommus[i]);
1451 	}
1452 }
1453 
1454 /* Flush the not present cache if it exists */
1455 static void domain_flush_np_cache(struct protection_domain *domain,
1456 		dma_addr_t iova, size_t size)
1457 {
1458 	if (unlikely(amd_iommu_np_cache)) {
1459 		unsigned long flags;
1460 
1461 		spin_lock_irqsave(&domain->lock, flags);
1462 		domain_flush_pages(domain, iova, size, 1);
1463 		amd_iommu_domain_flush_complete(domain);
1464 		spin_unlock_irqrestore(&domain->lock, flags);
1465 	}
1466 }
1467 
1468 
1469 /*
1470  * This function flushes the DTEs for all devices in domain
1471  */
1472 static void domain_flush_devices(struct protection_domain *domain)
1473 {
1474 	struct iommu_dev_data *dev_data;
1475 
1476 	list_for_each_entry(dev_data, &domain->dev_list, list)
1477 		device_flush_dte(dev_data);
1478 }
1479 
1480 /****************************************************************************
1481  *
1482  * The next functions belong to the domain allocation. A domain is
1483  * allocated for every IOMMU as the default domain. If device isolation
1484  * is enabled, every device get its own domain. The most important thing
1485  * about domains is the page table mapping the DMA address space they
1486  * contain.
1487  *
1488  ****************************************************************************/
1489 
1490 static u16 domain_id_alloc(void)
1491 {
1492 	int id;
1493 
1494 	spin_lock(&pd_bitmap_lock);
1495 	id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1496 	BUG_ON(id == 0);
1497 	if (id > 0 && id < MAX_DOMAIN_ID)
1498 		__set_bit(id, amd_iommu_pd_alloc_bitmap);
1499 	else
1500 		id = 0;
1501 	spin_unlock(&pd_bitmap_lock);
1502 
1503 	return id;
1504 }
1505 
1506 static void domain_id_free(int id)
1507 {
1508 	spin_lock(&pd_bitmap_lock);
1509 	if (id > 0 && id < MAX_DOMAIN_ID)
1510 		__clear_bit(id, amd_iommu_pd_alloc_bitmap);
1511 	spin_unlock(&pd_bitmap_lock);
1512 }
1513 
1514 static void free_gcr3_tbl_level1(u64 *tbl)
1515 {
1516 	u64 *ptr;
1517 	int i;
1518 
1519 	for (i = 0; i < 512; ++i) {
1520 		if (!(tbl[i] & GCR3_VALID))
1521 			continue;
1522 
1523 		ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1524 
1525 		free_page((unsigned long)ptr);
1526 	}
1527 }
1528 
1529 static void free_gcr3_tbl_level2(u64 *tbl)
1530 {
1531 	u64 *ptr;
1532 	int i;
1533 
1534 	for (i = 0; i < 512; ++i) {
1535 		if (!(tbl[i] & GCR3_VALID))
1536 			continue;
1537 
1538 		ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1539 
1540 		free_gcr3_tbl_level1(ptr);
1541 	}
1542 }
1543 
1544 static void free_gcr3_table(struct protection_domain *domain)
1545 {
1546 	if (domain->glx == 2)
1547 		free_gcr3_tbl_level2(domain->gcr3_tbl);
1548 	else if (domain->glx == 1)
1549 		free_gcr3_tbl_level1(domain->gcr3_tbl);
1550 	else
1551 		BUG_ON(domain->glx != 0);
1552 
1553 	free_page((unsigned long)domain->gcr3_tbl);
1554 }
1555 
1556 static void set_dte_entry(struct amd_iommu *iommu, u16 devid,
1557 			  struct protection_domain *domain, bool ats, bool ppr)
1558 {
1559 	u64 pte_root = 0;
1560 	u64 flags = 0;
1561 	u32 old_domid;
1562 	struct dev_table_entry *dev_table = get_dev_table(iommu);
1563 
1564 	if (domain->iop.mode != PAGE_MODE_NONE)
1565 		pte_root = iommu_virt_to_phys(domain->iop.root);
1566 
1567 	pte_root |= (domain->iop.mode & DEV_ENTRY_MODE_MASK)
1568 		    << DEV_ENTRY_MODE_SHIFT;
1569 
1570 	pte_root |= DTE_FLAG_IR | DTE_FLAG_IW | DTE_FLAG_V;
1571 
1572 	/*
1573 	 * When SNP is enabled, Only set TV bit when IOMMU
1574 	 * page translation is in use.
1575 	 */
1576 	if (!amd_iommu_snp_en || (domain->id != 0))
1577 		pte_root |= DTE_FLAG_TV;
1578 
1579 	flags = dev_table[devid].data[1];
1580 
1581 	if (ats)
1582 		flags |= DTE_FLAG_IOTLB;
1583 
1584 	if (ppr) {
1585 		if (iommu_feature(iommu, FEATURE_EPHSUP))
1586 			pte_root |= 1ULL << DEV_ENTRY_PPR;
1587 	}
1588 
1589 	if (domain->flags & PD_IOMMUV2_MASK) {
1590 		u64 gcr3 = iommu_virt_to_phys(domain->gcr3_tbl);
1591 		u64 glx  = domain->glx;
1592 		u64 tmp;
1593 
1594 		pte_root |= DTE_FLAG_GV;
1595 		pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1596 
1597 		/* First mask out possible old values for GCR3 table */
1598 		tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1599 		flags    &= ~tmp;
1600 
1601 		tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1602 		flags    &= ~tmp;
1603 
1604 		/* Encode GCR3 table into DTE */
1605 		tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1606 		pte_root |= tmp;
1607 
1608 		tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1609 		flags    |= tmp;
1610 
1611 		tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1612 		flags    |= tmp;
1613 
1614 		if (amd_iommu_gpt_level == PAGE_MODE_5_LEVEL) {
1615 			dev_table[devid].data[2] |=
1616 				((u64)GUEST_PGTABLE_5_LEVEL << DTE_GPT_LEVEL_SHIFT);
1617 		}
1618 
1619 		if (domain->flags & PD_GIOV_MASK)
1620 			pte_root |= DTE_FLAG_GIOV;
1621 	}
1622 
1623 	flags &= ~DEV_DOMID_MASK;
1624 	flags |= domain->id;
1625 
1626 	old_domid = dev_table[devid].data[1] & DEV_DOMID_MASK;
1627 	dev_table[devid].data[1]  = flags;
1628 	dev_table[devid].data[0]  = pte_root;
1629 
1630 	/*
1631 	 * A kdump kernel might be replacing a domain ID that was copied from
1632 	 * the previous kernel--if so, it needs to flush the translation cache
1633 	 * entries for the old domain ID that is being overwritten
1634 	 */
1635 	if (old_domid) {
1636 		amd_iommu_flush_tlb_domid(iommu, old_domid);
1637 	}
1638 }
1639 
1640 static void clear_dte_entry(struct amd_iommu *iommu, u16 devid)
1641 {
1642 	struct dev_table_entry *dev_table = get_dev_table(iommu);
1643 
1644 	/* remove entry from the device table seen by the hardware */
1645 	dev_table[devid].data[0]  = DTE_FLAG_V;
1646 
1647 	if (!amd_iommu_snp_en)
1648 		dev_table[devid].data[0] |= DTE_FLAG_TV;
1649 
1650 	dev_table[devid].data[1] &= DTE_FLAG_MASK;
1651 
1652 	amd_iommu_apply_erratum_63(iommu, devid);
1653 }
1654 
1655 static void do_attach(struct iommu_dev_data *dev_data,
1656 		      struct protection_domain *domain)
1657 {
1658 	struct amd_iommu *iommu;
1659 	bool ats;
1660 
1661 	iommu = rlookup_amd_iommu(dev_data->dev);
1662 	if (!iommu)
1663 		return;
1664 	ats   = dev_data->ats.enabled;
1665 
1666 	/* Update data structures */
1667 	dev_data->domain = domain;
1668 	list_add(&dev_data->list, &domain->dev_list);
1669 
1670 	/* Update NUMA Node ID */
1671 	if (domain->nid == NUMA_NO_NODE)
1672 		domain->nid = dev_to_node(dev_data->dev);
1673 
1674 	/* Do reference counting */
1675 	domain->dev_iommu[iommu->index] += 1;
1676 	domain->dev_cnt                 += 1;
1677 
1678 	/* Update device table */
1679 	set_dte_entry(iommu, dev_data->devid, domain,
1680 		      ats, dev_data->iommu_v2);
1681 	clone_aliases(iommu, dev_data->dev);
1682 
1683 	device_flush_dte(dev_data);
1684 }
1685 
1686 static void do_detach(struct iommu_dev_data *dev_data)
1687 {
1688 	struct protection_domain *domain = dev_data->domain;
1689 	struct amd_iommu *iommu;
1690 
1691 	iommu = rlookup_amd_iommu(dev_data->dev);
1692 	if (!iommu)
1693 		return;
1694 
1695 	/* Update data structures */
1696 	dev_data->domain = NULL;
1697 	list_del(&dev_data->list);
1698 	clear_dte_entry(iommu, dev_data->devid);
1699 	clone_aliases(iommu, dev_data->dev);
1700 
1701 	/* Flush the DTE entry */
1702 	device_flush_dte(dev_data);
1703 
1704 	/* Flush IOTLB */
1705 	amd_iommu_domain_flush_tlb_pde(domain);
1706 
1707 	/* Wait for the flushes to finish */
1708 	amd_iommu_domain_flush_complete(domain);
1709 
1710 	/* decrease reference counters - needs to happen after the flushes */
1711 	domain->dev_iommu[iommu->index] -= 1;
1712 	domain->dev_cnt                 -= 1;
1713 }
1714 
1715 static void pdev_iommuv2_disable(struct pci_dev *pdev)
1716 {
1717 	pci_disable_ats(pdev);
1718 	pci_disable_pri(pdev);
1719 	pci_disable_pasid(pdev);
1720 }
1721 
1722 static int pdev_pri_ats_enable(struct pci_dev *pdev)
1723 {
1724 	int ret;
1725 
1726 	/* Only allow access to user-accessible pages */
1727 	ret = pci_enable_pasid(pdev, 0);
1728 	if (ret)
1729 		return ret;
1730 
1731 	/* First reset the PRI state of the device */
1732 	ret = pci_reset_pri(pdev);
1733 	if (ret)
1734 		goto out_err_pasid;
1735 
1736 	/* Enable PRI */
1737 	/* FIXME: Hardcode number of outstanding requests for now */
1738 	ret = pci_enable_pri(pdev, 32);
1739 	if (ret)
1740 		goto out_err_pasid;
1741 
1742 	ret = pci_enable_ats(pdev, PAGE_SHIFT);
1743 	if (ret)
1744 		goto out_err_pri;
1745 
1746 	return 0;
1747 
1748 out_err_pri:
1749 	pci_disable_pri(pdev);
1750 
1751 out_err_pasid:
1752 	pci_disable_pasid(pdev);
1753 
1754 	return ret;
1755 }
1756 
1757 /*
1758  * If a device is not yet associated with a domain, this function makes the
1759  * device visible in the domain
1760  */
1761 static int attach_device(struct device *dev,
1762 			 struct protection_domain *domain)
1763 {
1764 	struct iommu_dev_data *dev_data;
1765 	struct pci_dev *pdev;
1766 	unsigned long flags;
1767 	int ret;
1768 
1769 	spin_lock_irqsave(&domain->lock, flags);
1770 
1771 	dev_data = dev_iommu_priv_get(dev);
1772 
1773 	spin_lock(&dev_data->lock);
1774 
1775 	ret = -EBUSY;
1776 	if (dev_data->domain != NULL)
1777 		goto out;
1778 
1779 	if (!dev_is_pci(dev))
1780 		goto skip_ats_check;
1781 
1782 	pdev = to_pci_dev(dev);
1783 	if (domain->flags & PD_IOMMUV2_MASK) {
1784 		struct iommu_domain *def_domain = iommu_get_dma_domain(dev);
1785 
1786 		ret = -EINVAL;
1787 
1788 		/*
1789 		 * In case of using AMD_IOMMU_V1 page table mode and the device
1790 		 * is enabling for PPR/ATS support (using v2 table),
1791 		 * we need to make sure that the domain type is identity map.
1792 		 */
1793 		if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
1794 		    def_domain->type != IOMMU_DOMAIN_IDENTITY) {
1795 			goto out;
1796 		}
1797 
1798 		if (dev_data->iommu_v2) {
1799 			if (pdev_pri_ats_enable(pdev) != 0)
1800 				goto out;
1801 
1802 			dev_data->ats.enabled = true;
1803 			dev_data->ats.qdep    = pci_ats_queue_depth(pdev);
1804 			dev_data->pri_tlp     = pci_prg_resp_pasid_required(pdev);
1805 		}
1806 	} else if (amd_iommu_iotlb_sup &&
1807 		   pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
1808 		dev_data->ats.enabled = true;
1809 		dev_data->ats.qdep    = pci_ats_queue_depth(pdev);
1810 	}
1811 
1812 skip_ats_check:
1813 	ret = 0;
1814 
1815 	do_attach(dev_data, domain);
1816 
1817 	/*
1818 	 * We might boot into a crash-kernel here. The crashed kernel
1819 	 * left the caches in the IOMMU dirty. So we have to flush
1820 	 * here to evict all dirty stuff.
1821 	 */
1822 	amd_iommu_domain_flush_tlb_pde(domain);
1823 
1824 	amd_iommu_domain_flush_complete(domain);
1825 
1826 out:
1827 	spin_unlock(&dev_data->lock);
1828 
1829 	spin_unlock_irqrestore(&domain->lock, flags);
1830 
1831 	return ret;
1832 }
1833 
1834 /*
1835  * Removes a device from a protection domain (with devtable_lock held)
1836  */
1837 static void detach_device(struct device *dev)
1838 {
1839 	struct protection_domain *domain;
1840 	struct iommu_dev_data *dev_data;
1841 	unsigned long flags;
1842 
1843 	dev_data = dev_iommu_priv_get(dev);
1844 	domain   = dev_data->domain;
1845 
1846 	spin_lock_irqsave(&domain->lock, flags);
1847 
1848 	spin_lock(&dev_data->lock);
1849 
1850 	/*
1851 	 * First check if the device is still attached. It might already
1852 	 * be detached from its domain because the generic
1853 	 * iommu_detach_group code detached it and we try again here in
1854 	 * our alias handling.
1855 	 */
1856 	if (WARN_ON(!dev_data->domain))
1857 		goto out;
1858 
1859 	do_detach(dev_data);
1860 
1861 	if (!dev_is_pci(dev))
1862 		goto out;
1863 
1864 	if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
1865 		pdev_iommuv2_disable(to_pci_dev(dev));
1866 	else if (dev_data->ats.enabled)
1867 		pci_disable_ats(to_pci_dev(dev));
1868 
1869 	dev_data->ats.enabled = false;
1870 
1871 out:
1872 	spin_unlock(&dev_data->lock);
1873 
1874 	spin_unlock_irqrestore(&domain->lock, flags);
1875 }
1876 
1877 static struct iommu_device *amd_iommu_probe_device(struct device *dev)
1878 {
1879 	struct iommu_device *iommu_dev;
1880 	struct amd_iommu *iommu;
1881 	int ret;
1882 
1883 	if (!check_device(dev))
1884 		return ERR_PTR(-ENODEV);
1885 
1886 	iommu = rlookup_amd_iommu(dev);
1887 	if (!iommu)
1888 		return ERR_PTR(-ENODEV);
1889 
1890 	/* Not registered yet? */
1891 	if (!iommu->iommu.ops)
1892 		return ERR_PTR(-ENODEV);
1893 
1894 	if (dev_iommu_priv_get(dev))
1895 		return &iommu->iommu;
1896 
1897 	ret = iommu_init_device(iommu, dev);
1898 	if (ret) {
1899 		if (ret != -ENOTSUPP)
1900 			dev_err(dev, "Failed to initialize - trying to proceed anyway\n");
1901 		iommu_dev = ERR_PTR(ret);
1902 		iommu_ignore_device(iommu, dev);
1903 	} else {
1904 		amd_iommu_set_pci_msi_domain(dev, iommu);
1905 		iommu_dev = &iommu->iommu;
1906 	}
1907 
1908 	iommu_completion_wait(iommu);
1909 
1910 	return iommu_dev;
1911 }
1912 
1913 static void amd_iommu_probe_finalize(struct device *dev)
1914 {
1915 	/* Domains are initialized for this device - have a look what we ended up with */
1916 	set_dma_ops(dev, NULL);
1917 	iommu_setup_dma_ops(dev, 0, U64_MAX);
1918 }
1919 
1920 static void amd_iommu_release_device(struct device *dev)
1921 {
1922 	struct amd_iommu *iommu;
1923 
1924 	if (!check_device(dev))
1925 		return;
1926 
1927 	iommu = rlookup_amd_iommu(dev);
1928 	if (!iommu)
1929 		return;
1930 
1931 	amd_iommu_uninit_device(dev);
1932 	iommu_completion_wait(iommu);
1933 }
1934 
1935 static struct iommu_group *amd_iommu_device_group(struct device *dev)
1936 {
1937 	if (dev_is_pci(dev))
1938 		return pci_device_group(dev);
1939 
1940 	return acpihid_device_group(dev);
1941 }
1942 
1943 /*****************************************************************************
1944  *
1945  * The next functions belong to the dma_ops mapping/unmapping code.
1946  *
1947  *****************************************************************************/
1948 
1949 static void update_device_table(struct protection_domain *domain)
1950 {
1951 	struct iommu_dev_data *dev_data;
1952 
1953 	list_for_each_entry(dev_data, &domain->dev_list, list) {
1954 		struct amd_iommu *iommu = rlookup_amd_iommu(dev_data->dev);
1955 
1956 		if (!iommu)
1957 			continue;
1958 		set_dte_entry(iommu, dev_data->devid, domain,
1959 			      dev_data->ats.enabled, dev_data->iommu_v2);
1960 		clone_aliases(iommu, dev_data->dev);
1961 	}
1962 }
1963 
1964 void amd_iommu_update_and_flush_device_table(struct protection_domain *domain)
1965 {
1966 	update_device_table(domain);
1967 	domain_flush_devices(domain);
1968 }
1969 
1970 void amd_iommu_domain_update(struct protection_domain *domain)
1971 {
1972 	/* Update device table */
1973 	amd_iommu_update_and_flush_device_table(domain);
1974 
1975 	/* Flush domain TLB(s) and wait for completion */
1976 	amd_iommu_domain_flush_tlb_pde(domain);
1977 	amd_iommu_domain_flush_complete(domain);
1978 }
1979 
1980 /*****************************************************************************
1981  *
1982  * The following functions belong to the exported interface of AMD IOMMU
1983  *
1984  * This interface allows access to lower level functions of the IOMMU
1985  * like protection domain handling and assignement of devices to domains
1986  * which is not possible with the dma_ops interface.
1987  *
1988  *****************************************************************************/
1989 
1990 static void cleanup_domain(struct protection_domain *domain)
1991 {
1992 	struct iommu_dev_data *entry;
1993 	unsigned long flags;
1994 
1995 	spin_lock_irqsave(&domain->lock, flags);
1996 
1997 	while (!list_empty(&domain->dev_list)) {
1998 		entry = list_first_entry(&domain->dev_list,
1999 					 struct iommu_dev_data, list);
2000 		BUG_ON(!entry->domain);
2001 		do_detach(entry);
2002 	}
2003 
2004 	spin_unlock_irqrestore(&domain->lock, flags);
2005 }
2006 
2007 static void protection_domain_free(struct protection_domain *domain)
2008 {
2009 	if (!domain)
2010 		return;
2011 
2012 	if (domain->iop.pgtbl_cfg.tlb)
2013 		free_io_pgtable_ops(&domain->iop.iop.ops);
2014 
2015 	if (domain->id)
2016 		domain_id_free(domain->id);
2017 
2018 	kfree(domain);
2019 }
2020 
2021 static int protection_domain_init_v1(struct protection_domain *domain, int mode)
2022 {
2023 	u64 *pt_root = NULL;
2024 
2025 	BUG_ON(mode < PAGE_MODE_NONE || mode > PAGE_MODE_6_LEVEL);
2026 
2027 	spin_lock_init(&domain->lock);
2028 	domain->id = domain_id_alloc();
2029 	if (!domain->id)
2030 		return -ENOMEM;
2031 	INIT_LIST_HEAD(&domain->dev_list);
2032 
2033 	if (mode != PAGE_MODE_NONE) {
2034 		pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2035 		if (!pt_root) {
2036 			domain_id_free(domain->id);
2037 			return -ENOMEM;
2038 		}
2039 	}
2040 
2041 	amd_iommu_domain_set_pgtable(domain, pt_root, mode);
2042 
2043 	return 0;
2044 }
2045 
2046 static int protection_domain_init_v2(struct protection_domain *domain)
2047 {
2048 	spin_lock_init(&domain->lock);
2049 	domain->id = domain_id_alloc();
2050 	if (!domain->id)
2051 		return -ENOMEM;
2052 	INIT_LIST_HEAD(&domain->dev_list);
2053 
2054 	domain->flags |= PD_GIOV_MASK;
2055 
2056 	domain->domain.pgsize_bitmap = AMD_IOMMU_PGSIZES_V2;
2057 
2058 	if (domain_enable_v2(domain, 1)) {
2059 		domain_id_free(domain->id);
2060 		return -ENOMEM;
2061 	}
2062 
2063 	return 0;
2064 }
2065 
2066 static struct protection_domain *protection_domain_alloc(unsigned int type)
2067 {
2068 	struct io_pgtable_ops *pgtbl_ops;
2069 	struct protection_domain *domain;
2070 	int pgtable = amd_iommu_pgtable;
2071 	int mode = DEFAULT_PGTABLE_LEVEL;
2072 	int ret;
2073 
2074 	domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2075 	if (!domain)
2076 		return NULL;
2077 
2078 	/*
2079 	 * Force IOMMU v1 page table when iommu=pt and
2080 	 * when allocating domain for pass-through devices.
2081 	 */
2082 	if (type == IOMMU_DOMAIN_IDENTITY) {
2083 		pgtable = AMD_IOMMU_V1;
2084 		mode = PAGE_MODE_NONE;
2085 	} else if (type == IOMMU_DOMAIN_UNMANAGED) {
2086 		pgtable = AMD_IOMMU_V1;
2087 	}
2088 
2089 	switch (pgtable) {
2090 	case AMD_IOMMU_V1:
2091 		ret = protection_domain_init_v1(domain, mode);
2092 		break;
2093 	case AMD_IOMMU_V2:
2094 		ret = protection_domain_init_v2(domain);
2095 		break;
2096 	default:
2097 		ret = -EINVAL;
2098 	}
2099 
2100 	if (ret)
2101 		goto out_err;
2102 
2103 	/* No need to allocate io pgtable ops in passthrough mode */
2104 	if (type == IOMMU_DOMAIN_IDENTITY)
2105 		return domain;
2106 
2107 	domain->nid = NUMA_NO_NODE;
2108 
2109 	pgtbl_ops = alloc_io_pgtable_ops(pgtable, &domain->iop.pgtbl_cfg, domain);
2110 	if (!pgtbl_ops) {
2111 		domain_id_free(domain->id);
2112 		goto out_err;
2113 	}
2114 
2115 	return domain;
2116 out_err:
2117 	kfree(domain);
2118 	return NULL;
2119 }
2120 
2121 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2122 {
2123 	struct protection_domain *domain;
2124 
2125 	/*
2126 	 * Since DTE[Mode]=0 is prohibited on SNP-enabled system,
2127 	 * default to use IOMMU_DOMAIN_DMA[_FQ].
2128 	 */
2129 	if (amd_iommu_snp_en && (type == IOMMU_DOMAIN_IDENTITY))
2130 		return NULL;
2131 
2132 	domain = protection_domain_alloc(type);
2133 	if (!domain)
2134 		return NULL;
2135 
2136 	domain->domain.geometry.aperture_start = 0;
2137 	domain->domain.geometry.aperture_end   = ~0ULL;
2138 	domain->domain.geometry.force_aperture = true;
2139 
2140 	return &domain->domain;
2141 }
2142 
2143 static void amd_iommu_domain_free(struct iommu_domain *dom)
2144 {
2145 	struct protection_domain *domain;
2146 
2147 	domain = to_pdomain(dom);
2148 
2149 	if (domain->dev_cnt > 0)
2150 		cleanup_domain(domain);
2151 
2152 	BUG_ON(domain->dev_cnt != 0);
2153 
2154 	if (!dom)
2155 		return;
2156 
2157 	if (domain->flags & PD_IOMMUV2_MASK)
2158 		free_gcr3_table(domain);
2159 
2160 	protection_domain_free(domain);
2161 }
2162 
2163 static int amd_iommu_attach_device(struct iommu_domain *dom,
2164 				   struct device *dev)
2165 {
2166 	struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2167 	struct protection_domain *domain = to_pdomain(dom);
2168 	struct amd_iommu *iommu = rlookup_amd_iommu(dev);
2169 	int ret;
2170 
2171 	/*
2172 	 * Skip attach device to domain if new domain is same as
2173 	 * devices current domain
2174 	 */
2175 	if (dev_data->domain == domain)
2176 		return 0;
2177 
2178 	dev_data->defer_attach = false;
2179 
2180 	if (dev_data->domain)
2181 		detach_device(dev);
2182 
2183 	ret = attach_device(dev, domain);
2184 
2185 #ifdef CONFIG_IRQ_REMAP
2186 	if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
2187 		if (dom->type == IOMMU_DOMAIN_UNMANAGED)
2188 			dev_data->use_vapic = 1;
2189 		else
2190 			dev_data->use_vapic = 0;
2191 	}
2192 #endif
2193 
2194 	iommu_completion_wait(iommu);
2195 
2196 	return ret;
2197 }
2198 
2199 static void amd_iommu_iotlb_sync_map(struct iommu_domain *dom,
2200 				     unsigned long iova, size_t size)
2201 {
2202 	struct protection_domain *domain = to_pdomain(dom);
2203 	struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2204 
2205 	if (ops->map_pages)
2206 		domain_flush_np_cache(domain, iova, size);
2207 }
2208 
2209 static int amd_iommu_map_pages(struct iommu_domain *dom, unsigned long iova,
2210 			       phys_addr_t paddr, size_t pgsize, size_t pgcount,
2211 			       int iommu_prot, gfp_t gfp, size_t *mapped)
2212 {
2213 	struct protection_domain *domain = to_pdomain(dom);
2214 	struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2215 	int prot = 0;
2216 	int ret = -EINVAL;
2217 
2218 	if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
2219 	    (domain->iop.mode == PAGE_MODE_NONE))
2220 		return -EINVAL;
2221 
2222 	if (iommu_prot & IOMMU_READ)
2223 		prot |= IOMMU_PROT_IR;
2224 	if (iommu_prot & IOMMU_WRITE)
2225 		prot |= IOMMU_PROT_IW;
2226 
2227 	if (ops->map_pages) {
2228 		ret = ops->map_pages(ops, iova, paddr, pgsize,
2229 				     pgcount, prot, gfp, mapped);
2230 	}
2231 
2232 	return ret;
2233 }
2234 
2235 static void amd_iommu_iotlb_gather_add_page(struct iommu_domain *domain,
2236 					    struct iommu_iotlb_gather *gather,
2237 					    unsigned long iova, size_t size)
2238 {
2239 	/*
2240 	 * AMD's IOMMU can flush as many pages as necessary in a single flush.
2241 	 * Unless we run in a virtual machine, which can be inferred according
2242 	 * to whether "non-present cache" is on, it is probably best to prefer
2243 	 * (potentially) too extensive TLB flushing (i.e., more misses) over
2244 	 * mutliple TLB flushes (i.e., more flushes). For virtual machines the
2245 	 * hypervisor needs to synchronize the host IOMMU PTEs with those of
2246 	 * the guest, and the trade-off is different: unnecessary TLB flushes
2247 	 * should be avoided.
2248 	 */
2249 	if (amd_iommu_np_cache &&
2250 	    iommu_iotlb_gather_is_disjoint(gather, iova, size))
2251 		iommu_iotlb_sync(domain, gather);
2252 
2253 	iommu_iotlb_gather_add_range(gather, iova, size);
2254 }
2255 
2256 static size_t amd_iommu_unmap_pages(struct iommu_domain *dom, unsigned long iova,
2257 				    size_t pgsize, size_t pgcount,
2258 				    struct iommu_iotlb_gather *gather)
2259 {
2260 	struct protection_domain *domain = to_pdomain(dom);
2261 	struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2262 	size_t r;
2263 
2264 	if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
2265 	    (domain->iop.mode == PAGE_MODE_NONE))
2266 		return 0;
2267 
2268 	r = (ops->unmap_pages) ? ops->unmap_pages(ops, iova, pgsize, pgcount, NULL) : 0;
2269 
2270 	if (r)
2271 		amd_iommu_iotlb_gather_add_page(dom, gather, iova, r);
2272 
2273 	return r;
2274 }
2275 
2276 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
2277 					  dma_addr_t iova)
2278 {
2279 	struct protection_domain *domain = to_pdomain(dom);
2280 	struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2281 
2282 	return ops->iova_to_phys(ops, iova);
2283 }
2284 
2285 static bool amd_iommu_capable(struct device *dev, enum iommu_cap cap)
2286 {
2287 	switch (cap) {
2288 	case IOMMU_CAP_CACHE_COHERENCY:
2289 		return true;
2290 	case IOMMU_CAP_NOEXEC:
2291 		return false;
2292 	case IOMMU_CAP_PRE_BOOT_PROTECTION:
2293 		return amdr_ivrs_remap_support;
2294 	case IOMMU_CAP_ENFORCE_CACHE_COHERENCY:
2295 		return true;
2296 	default:
2297 		break;
2298 	}
2299 
2300 	return false;
2301 }
2302 
2303 static void amd_iommu_get_resv_regions(struct device *dev,
2304 				       struct list_head *head)
2305 {
2306 	struct iommu_resv_region *region;
2307 	struct unity_map_entry *entry;
2308 	struct amd_iommu *iommu;
2309 	struct amd_iommu_pci_seg *pci_seg;
2310 	int devid, sbdf;
2311 
2312 	sbdf = get_device_sbdf_id(dev);
2313 	if (sbdf < 0)
2314 		return;
2315 
2316 	devid = PCI_SBDF_TO_DEVID(sbdf);
2317 	iommu = rlookup_amd_iommu(dev);
2318 	if (!iommu)
2319 		return;
2320 	pci_seg = iommu->pci_seg;
2321 
2322 	list_for_each_entry(entry, &pci_seg->unity_map, list) {
2323 		int type, prot = 0;
2324 		size_t length;
2325 
2326 		if (devid < entry->devid_start || devid > entry->devid_end)
2327 			continue;
2328 
2329 		type   = IOMMU_RESV_DIRECT;
2330 		length = entry->address_end - entry->address_start;
2331 		if (entry->prot & IOMMU_PROT_IR)
2332 			prot |= IOMMU_READ;
2333 		if (entry->prot & IOMMU_PROT_IW)
2334 			prot |= IOMMU_WRITE;
2335 		if (entry->prot & IOMMU_UNITY_MAP_FLAG_EXCL_RANGE)
2336 			/* Exclusion range */
2337 			type = IOMMU_RESV_RESERVED;
2338 
2339 		region = iommu_alloc_resv_region(entry->address_start,
2340 						 length, prot, type,
2341 						 GFP_KERNEL);
2342 		if (!region) {
2343 			dev_err(dev, "Out of memory allocating dm-regions\n");
2344 			return;
2345 		}
2346 		list_add_tail(&region->list, head);
2347 	}
2348 
2349 	region = iommu_alloc_resv_region(MSI_RANGE_START,
2350 					 MSI_RANGE_END - MSI_RANGE_START + 1,
2351 					 0, IOMMU_RESV_MSI, GFP_KERNEL);
2352 	if (!region)
2353 		return;
2354 	list_add_tail(&region->list, head);
2355 
2356 	region = iommu_alloc_resv_region(HT_RANGE_START,
2357 					 HT_RANGE_END - HT_RANGE_START + 1,
2358 					 0, IOMMU_RESV_RESERVED, GFP_KERNEL);
2359 	if (!region)
2360 		return;
2361 	list_add_tail(&region->list, head);
2362 }
2363 
2364 bool amd_iommu_is_attach_deferred(struct device *dev)
2365 {
2366 	struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2367 
2368 	return dev_data->defer_attach;
2369 }
2370 EXPORT_SYMBOL_GPL(amd_iommu_is_attach_deferred);
2371 
2372 static void amd_iommu_flush_iotlb_all(struct iommu_domain *domain)
2373 {
2374 	struct protection_domain *dom = to_pdomain(domain);
2375 	unsigned long flags;
2376 
2377 	spin_lock_irqsave(&dom->lock, flags);
2378 	amd_iommu_domain_flush_tlb_pde(dom);
2379 	amd_iommu_domain_flush_complete(dom);
2380 	spin_unlock_irqrestore(&dom->lock, flags);
2381 }
2382 
2383 static void amd_iommu_iotlb_sync(struct iommu_domain *domain,
2384 				 struct iommu_iotlb_gather *gather)
2385 {
2386 	struct protection_domain *dom = to_pdomain(domain);
2387 	unsigned long flags;
2388 
2389 	spin_lock_irqsave(&dom->lock, flags);
2390 	domain_flush_pages(dom, gather->start, gather->end - gather->start, 1);
2391 	amd_iommu_domain_flush_complete(dom);
2392 	spin_unlock_irqrestore(&dom->lock, flags);
2393 }
2394 
2395 static int amd_iommu_def_domain_type(struct device *dev)
2396 {
2397 	struct iommu_dev_data *dev_data;
2398 
2399 	dev_data = dev_iommu_priv_get(dev);
2400 	if (!dev_data)
2401 		return 0;
2402 
2403 	/*
2404 	 * Do not identity map IOMMUv2 capable devices when:
2405 	 *  - memory encryption is active, because some of those devices
2406 	 *    (AMD GPUs) don't have the encryption bit in their DMA-mask
2407 	 *    and require remapping.
2408 	 *  - SNP is enabled, because it prohibits DTE[Mode]=0.
2409 	 */
2410 	if (dev_data->iommu_v2 &&
2411 	    !cc_platform_has(CC_ATTR_MEM_ENCRYPT) &&
2412 	    !amd_iommu_snp_en) {
2413 		return IOMMU_DOMAIN_IDENTITY;
2414 	}
2415 
2416 	return 0;
2417 }
2418 
2419 static bool amd_iommu_enforce_cache_coherency(struct iommu_domain *domain)
2420 {
2421 	/* IOMMU_PTE_FC is always set */
2422 	return true;
2423 }
2424 
2425 const struct iommu_ops amd_iommu_ops = {
2426 	.capable = amd_iommu_capable,
2427 	.domain_alloc = amd_iommu_domain_alloc,
2428 	.probe_device = amd_iommu_probe_device,
2429 	.release_device = amd_iommu_release_device,
2430 	.probe_finalize = amd_iommu_probe_finalize,
2431 	.device_group = amd_iommu_device_group,
2432 	.get_resv_regions = amd_iommu_get_resv_regions,
2433 	.is_attach_deferred = amd_iommu_is_attach_deferred,
2434 	.pgsize_bitmap	= AMD_IOMMU_PGSIZES,
2435 	.def_domain_type = amd_iommu_def_domain_type,
2436 	.default_domain_ops = &(const struct iommu_domain_ops) {
2437 		.attach_dev	= amd_iommu_attach_device,
2438 		.map_pages	= amd_iommu_map_pages,
2439 		.unmap_pages	= amd_iommu_unmap_pages,
2440 		.iotlb_sync_map	= amd_iommu_iotlb_sync_map,
2441 		.iova_to_phys	= amd_iommu_iova_to_phys,
2442 		.flush_iotlb_all = amd_iommu_flush_iotlb_all,
2443 		.iotlb_sync	= amd_iommu_iotlb_sync,
2444 		.free		= amd_iommu_domain_free,
2445 		.enforce_cache_coherency = amd_iommu_enforce_cache_coherency,
2446 	}
2447 };
2448 
2449 /*****************************************************************************
2450  *
2451  * The next functions do a basic initialization of IOMMU for pass through
2452  * mode
2453  *
2454  * In passthrough mode the IOMMU is initialized and enabled but not used for
2455  * DMA-API translation.
2456  *
2457  *****************************************************************************/
2458 
2459 /* IOMMUv2 specific functions */
2460 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
2461 {
2462 	return atomic_notifier_chain_register(&ppr_notifier, nb);
2463 }
2464 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
2465 
2466 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
2467 {
2468 	return atomic_notifier_chain_unregister(&ppr_notifier, nb);
2469 }
2470 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
2471 
2472 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
2473 {
2474 	struct protection_domain *domain = to_pdomain(dom);
2475 	unsigned long flags;
2476 
2477 	spin_lock_irqsave(&domain->lock, flags);
2478 
2479 	if (domain->iop.pgtbl_cfg.tlb)
2480 		free_io_pgtable_ops(&domain->iop.iop.ops);
2481 
2482 	spin_unlock_irqrestore(&domain->lock, flags);
2483 }
2484 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
2485 
2486 /* Note: This function expects iommu_domain->lock to be held prior calling the function. */
2487 static int domain_enable_v2(struct protection_domain *domain, int pasids)
2488 {
2489 	int levels;
2490 
2491 	/* Number of GCR3 table levels required */
2492 	for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
2493 		levels += 1;
2494 
2495 	if (levels > amd_iommu_max_glx_val)
2496 		return -EINVAL;
2497 
2498 	domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
2499 	if (domain->gcr3_tbl == NULL)
2500 		return -ENOMEM;
2501 
2502 	domain->glx      = levels;
2503 	domain->flags   |= PD_IOMMUV2_MASK;
2504 
2505 	amd_iommu_domain_update(domain);
2506 
2507 	return 0;
2508 }
2509 
2510 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
2511 {
2512 	struct protection_domain *pdom = to_pdomain(dom);
2513 	unsigned long flags;
2514 	int ret;
2515 
2516 	spin_lock_irqsave(&pdom->lock, flags);
2517 
2518 	/*
2519 	 * Save us all sanity checks whether devices already in the
2520 	 * domain support IOMMUv2. Just force that the domain has no
2521 	 * devices attached when it is switched into IOMMUv2 mode.
2522 	 */
2523 	ret = -EBUSY;
2524 	if (pdom->dev_cnt > 0 || pdom->flags & PD_IOMMUV2_MASK)
2525 		goto out;
2526 
2527 	if (!pdom->gcr3_tbl)
2528 		ret = domain_enable_v2(pdom, pasids);
2529 
2530 out:
2531 	spin_unlock_irqrestore(&pdom->lock, flags);
2532 	return ret;
2533 }
2534 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
2535 
2536 static int __flush_pasid(struct protection_domain *domain, u32 pasid,
2537 			 u64 address, bool size)
2538 {
2539 	struct iommu_dev_data *dev_data;
2540 	struct iommu_cmd cmd;
2541 	int i, ret;
2542 
2543 	if (!(domain->flags & PD_IOMMUV2_MASK))
2544 		return -EINVAL;
2545 
2546 	build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
2547 
2548 	/*
2549 	 * IOMMU TLB needs to be flushed before Device TLB to
2550 	 * prevent device TLB refill from IOMMU TLB
2551 	 */
2552 	for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
2553 		if (domain->dev_iommu[i] == 0)
2554 			continue;
2555 
2556 		ret = iommu_queue_command(amd_iommus[i], &cmd);
2557 		if (ret != 0)
2558 			goto out;
2559 	}
2560 
2561 	/* Wait until IOMMU TLB flushes are complete */
2562 	amd_iommu_domain_flush_complete(domain);
2563 
2564 	/* Now flush device TLBs */
2565 	list_for_each_entry(dev_data, &domain->dev_list, list) {
2566 		struct amd_iommu *iommu;
2567 		int qdep;
2568 
2569 		/*
2570 		   There might be non-IOMMUv2 capable devices in an IOMMUv2
2571 		 * domain.
2572 		 */
2573 		if (!dev_data->ats.enabled)
2574 			continue;
2575 
2576 		qdep  = dev_data->ats.qdep;
2577 		iommu = rlookup_amd_iommu(dev_data->dev);
2578 		if (!iommu)
2579 			continue;
2580 		build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
2581 				      qdep, address, size);
2582 
2583 		ret = iommu_queue_command(iommu, &cmd);
2584 		if (ret != 0)
2585 			goto out;
2586 	}
2587 
2588 	/* Wait until all device TLBs are flushed */
2589 	amd_iommu_domain_flush_complete(domain);
2590 
2591 	ret = 0;
2592 
2593 out:
2594 
2595 	return ret;
2596 }
2597 
2598 static int __amd_iommu_flush_page(struct protection_domain *domain, u32 pasid,
2599 				  u64 address)
2600 {
2601 	return __flush_pasid(domain, pasid, address, false);
2602 }
2603 
2604 int amd_iommu_flush_page(struct iommu_domain *dom, u32 pasid,
2605 			 u64 address)
2606 {
2607 	struct protection_domain *domain = to_pdomain(dom);
2608 	unsigned long flags;
2609 	int ret;
2610 
2611 	spin_lock_irqsave(&domain->lock, flags);
2612 	ret = __amd_iommu_flush_page(domain, pasid, address);
2613 	spin_unlock_irqrestore(&domain->lock, flags);
2614 
2615 	return ret;
2616 }
2617 EXPORT_SYMBOL(amd_iommu_flush_page);
2618 
2619 static int __amd_iommu_flush_tlb(struct protection_domain *domain, u32 pasid)
2620 {
2621 	return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
2622 			     true);
2623 }
2624 
2625 int amd_iommu_flush_tlb(struct iommu_domain *dom, u32 pasid)
2626 {
2627 	struct protection_domain *domain = to_pdomain(dom);
2628 	unsigned long flags;
2629 	int ret;
2630 
2631 	spin_lock_irqsave(&domain->lock, flags);
2632 	ret = __amd_iommu_flush_tlb(domain, pasid);
2633 	spin_unlock_irqrestore(&domain->lock, flags);
2634 
2635 	return ret;
2636 }
2637 EXPORT_SYMBOL(amd_iommu_flush_tlb);
2638 
2639 static u64 *__get_gcr3_pte(u64 *root, int level, u32 pasid, bool alloc)
2640 {
2641 	int index;
2642 	u64 *pte;
2643 
2644 	while (true) {
2645 
2646 		index = (pasid >> (9 * level)) & 0x1ff;
2647 		pte   = &root[index];
2648 
2649 		if (level == 0)
2650 			break;
2651 
2652 		if (!(*pte & GCR3_VALID)) {
2653 			if (!alloc)
2654 				return NULL;
2655 
2656 			root = (void *)get_zeroed_page(GFP_ATOMIC);
2657 			if (root == NULL)
2658 				return NULL;
2659 
2660 			*pte = iommu_virt_to_phys(root) | GCR3_VALID;
2661 		}
2662 
2663 		root = iommu_phys_to_virt(*pte & PAGE_MASK);
2664 
2665 		level -= 1;
2666 	}
2667 
2668 	return pte;
2669 }
2670 
2671 static int __set_gcr3(struct protection_domain *domain, u32 pasid,
2672 		      unsigned long cr3)
2673 {
2674 	u64 *pte;
2675 
2676 	if (domain->iop.mode != PAGE_MODE_NONE)
2677 		return -EINVAL;
2678 
2679 	pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
2680 	if (pte == NULL)
2681 		return -ENOMEM;
2682 
2683 	*pte = (cr3 & PAGE_MASK) | GCR3_VALID;
2684 
2685 	return __amd_iommu_flush_tlb(domain, pasid);
2686 }
2687 
2688 static int __clear_gcr3(struct protection_domain *domain, u32 pasid)
2689 {
2690 	u64 *pte;
2691 
2692 	if (domain->iop.mode != PAGE_MODE_NONE)
2693 		return -EINVAL;
2694 
2695 	pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
2696 	if (pte == NULL)
2697 		return 0;
2698 
2699 	*pte = 0;
2700 
2701 	return __amd_iommu_flush_tlb(domain, pasid);
2702 }
2703 
2704 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, u32 pasid,
2705 			      unsigned long cr3)
2706 {
2707 	struct protection_domain *domain = to_pdomain(dom);
2708 	unsigned long flags;
2709 	int ret;
2710 
2711 	spin_lock_irqsave(&domain->lock, flags);
2712 	ret = __set_gcr3(domain, pasid, cr3);
2713 	spin_unlock_irqrestore(&domain->lock, flags);
2714 
2715 	return ret;
2716 }
2717 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
2718 
2719 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, u32 pasid)
2720 {
2721 	struct protection_domain *domain = to_pdomain(dom);
2722 	unsigned long flags;
2723 	int ret;
2724 
2725 	spin_lock_irqsave(&domain->lock, flags);
2726 	ret = __clear_gcr3(domain, pasid);
2727 	spin_unlock_irqrestore(&domain->lock, flags);
2728 
2729 	return ret;
2730 }
2731 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
2732 
2733 int amd_iommu_complete_ppr(struct pci_dev *pdev, u32 pasid,
2734 			   int status, int tag)
2735 {
2736 	struct iommu_dev_data *dev_data;
2737 	struct amd_iommu *iommu;
2738 	struct iommu_cmd cmd;
2739 
2740 	dev_data = dev_iommu_priv_get(&pdev->dev);
2741 	iommu    = rlookup_amd_iommu(&pdev->dev);
2742 	if (!iommu)
2743 		return -ENODEV;
2744 
2745 	build_complete_ppr(&cmd, dev_data->devid, pasid, status,
2746 			   tag, dev_data->pri_tlp);
2747 
2748 	return iommu_queue_command(iommu, &cmd);
2749 }
2750 EXPORT_SYMBOL(amd_iommu_complete_ppr);
2751 
2752 int amd_iommu_device_info(struct pci_dev *pdev,
2753                           struct amd_iommu_device_info *info)
2754 {
2755 	int max_pasids;
2756 	int pos;
2757 
2758 	if (pdev == NULL || info == NULL)
2759 		return -EINVAL;
2760 
2761 	if (!amd_iommu_v2_supported())
2762 		return -EINVAL;
2763 
2764 	memset(info, 0, sizeof(*info));
2765 
2766 	if (pci_ats_supported(pdev))
2767 		info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
2768 
2769 	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2770 	if (pos)
2771 		info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
2772 
2773 	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
2774 	if (pos) {
2775 		int features;
2776 
2777 		max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
2778 		max_pasids = min(max_pasids, (1 << 20));
2779 
2780 		info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
2781 		info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
2782 
2783 		features = pci_pasid_features(pdev);
2784 		if (features & PCI_PASID_CAP_EXEC)
2785 			info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
2786 		if (features & PCI_PASID_CAP_PRIV)
2787 			info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
2788 	}
2789 
2790 	return 0;
2791 }
2792 EXPORT_SYMBOL(amd_iommu_device_info);
2793 
2794 #ifdef CONFIG_IRQ_REMAP
2795 
2796 /*****************************************************************************
2797  *
2798  * Interrupt Remapping Implementation
2799  *
2800  *****************************************************************************/
2801 
2802 static struct irq_chip amd_ir_chip;
2803 static DEFINE_SPINLOCK(iommu_table_lock);
2804 
2805 static void set_dte_irq_entry(struct amd_iommu *iommu, u16 devid,
2806 			      struct irq_remap_table *table)
2807 {
2808 	u64 dte;
2809 	struct dev_table_entry *dev_table = get_dev_table(iommu);
2810 
2811 	dte	= dev_table[devid].data[2];
2812 	dte	&= ~DTE_IRQ_PHYS_ADDR_MASK;
2813 	dte	|= iommu_virt_to_phys(table->table);
2814 	dte	|= DTE_IRQ_REMAP_INTCTL;
2815 	dte	|= DTE_INTTABLEN;
2816 	dte	|= DTE_IRQ_REMAP_ENABLE;
2817 
2818 	dev_table[devid].data[2] = dte;
2819 }
2820 
2821 static struct irq_remap_table *get_irq_table(struct amd_iommu *iommu, u16 devid)
2822 {
2823 	struct irq_remap_table *table;
2824 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
2825 
2826 	if (WARN_ONCE(!pci_seg->rlookup_table[devid],
2827 		      "%s: no iommu for devid %x:%x\n",
2828 		      __func__, pci_seg->id, devid))
2829 		return NULL;
2830 
2831 	table = pci_seg->irq_lookup_table[devid];
2832 	if (WARN_ONCE(!table, "%s: no table for devid %x:%x\n",
2833 		      __func__, pci_seg->id, devid))
2834 		return NULL;
2835 
2836 	return table;
2837 }
2838 
2839 static struct irq_remap_table *__alloc_irq_table(void)
2840 {
2841 	struct irq_remap_table *table;
2842 
2843 	table = kzalloc(sizeof(*table), GFP_KERNEL);
2844 	if (!table)
2845 		return NULL;
2846 
2847 	table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_KERNEL);
2848 	if (!table->table) {
2849 		kfree(table);
2850 		return NULL;
2851 	}
2852 	raw_spin_lock_init(&table->lock);
2853 
2854 	if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
2855 		memset(table->table, 0,
2856 		       MAX_IRQS_PER_TABLE * sizeof(u32));
2857 	else
2858 		memset(table->table, 0,
2859 		       (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
2860 	return table;
2861 }
2862 
2863 static void set_remap_table_entry(struct amd_iommu *iommu, u16 devid,
2864 				  struct irq_remap_table *table)
2865 {
2866 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
2867 
2868 	pci_seg->irq_lookup_table[devid] = table;
2869 	set_dte_irq_entry(iommu, devid, table);
2870 	iommu_flush_dte(iommu, devid);
2871 }
2872 
2873 static int set_remap_table_entry_alias(struct pci_dev *pdev, u16 alias,
2874 				       void *data)
2875 {
2876 	struct irq_remap_table *table = data;
2877 	struct amd_iommu_pci_seg *pci_seg;
2878 	struct amd_iommu *iommu = rlookup_amd_iommu(&pdev->dev);
2879 
2880 	if (!iommu)
2881 		return -EINVAL;
2882 
2883 	pci_seg = iommu->pci_seg;
2884 	pci_seg->irq_lookup_table[alias] = table;
2885 	set_dte_irq_entry(iommu, alias, table);
2886 	iommu_flush_dte(pci_seg->rlookup_table[alias], alias);
2887 
2888 	return 0;
2889 }
2890 
2891 static struct irq_remap_table *alloc_irq_table(struct amd_iommu *iommu,
2892 					       u16 devid, struct pci_dev *pdev)
2893 {
2894 	struct irq_remap_table *table = NULL;
2895 	struct irq_remap_table *new_table = NULL;
2896 	struct amd_iommu_pci_seg *pci_seg;
2897 	unsigned long flags;
2898 	u16 alias;
2899 
2900 	spin_lock_irqsave(&iommu_table_lock, flags);
2901 
2902 	pci_seg = iommu->pci_seg;
2903 	table = pci_seg->irq_lookup_table[devid];
2904 	if (table)
2905 		goto out_unlock;
2906 
2907 	alias = pci_seg->alias_table[devid];
2908 	table = pci_seg->irq_lookup_table[alias];
2909 	if (table) {
2910 		set_remap_table_entry(iommu, devid, table);
2911 		goto out_wait;
2912 	}
2913 	spin_unlock_irqrestore(&iommu_table_lock, flags);
2914 
2915 	/* Nothing there yet, allocate new irq remapping table */
2916 	new_table = __alloc_irq_table();
2917 	if (!new_table)
2918 		return NULL;
2919 
2920 	spin_lock_irqsave(&iommu_table_lock, flags);
2921 
2922 	table = pci_seg->irq_lookup_table[devid];
2923 	if (table)
2924 		goto out_unlock;
2925 
2926 	table = pci_seg->irq_lookup_table[alias];
2927 	if (table) {
2928 		set_remap_table_entry(iommu, devid, table);
2929 		goto out_wait;
2930 	}
2931 
2932 	table = new_table;
2933 	new_table = NULL;
2934 
2935 	if (pdev)
2936 		pci_for_each_dma_alias(pdev, set_remap_table_entry_alias,
2937 				       table);
2938 	else
2939 		set_remap_table_entry(iommu, devid, table);
2940 
2941 	if (devid != alias)
2942 		set_remap_table_entry(iommu, alias, table);
2943 
2944 out_wait:
2945 	iommu_completion_wait(iommu);
2946 
2947 out_unlock:
2948 	spin_unlock_irqrestore(&iommu_table_lock, flags);
2949 
2950 	if (new_table) {
2951 		kmem_cache_free(amd_iommu_irq_cache, new_table->table);
2952 		kfree(new_table);
2953 	}
2954 	return table;
2955 }
2956 
2957 static int alloc_irq_index(struct amd_iommu *iommu, u16 devid, int count,
2958 			   bool align, struct pci_dev *pdev)
2959 {
2960 	struct irq_remap_table *table;
2961 	int index, c, alignment = 1;
2962 	unsigned long flags;
2963 
2964 	table = alloc_irq_table(iommu, devid, pdev);
2965 	if (!table)
2966 		return -ENODEV;
2967 
2968 	if (align)
2969 		alignment = roundup_pow_of_two(count);
2970 
2971 	raw_spin_lock_irqsave(&table->lock, flags);
2972 
2973 	/* Scan table for free entries */
2974 	for (index = ALIGN(table->min_index, alignment), c = 0;
2975 	     index < MAX_IRQS_PER_TABLE;) {
2976 		if (!iommu->irte_ops->is_allocated(table, index)) {
2977 			c += 1;
2978 		} else {
2979 			c     = 0;
2980 			index = ALIGN(index + 1, alignment);
2981 			continue;
2982 		}
2983 
2984 		if (c == count)	{
2985 			for (; c != 0; --c)
2986 				iommu->irte_ops->set_allocated(table, index - c + 1);
2987 
2988 			index -= count - 1;
2989 			goto out;
2990 		}
2991 
2992 		index++;
2993 	}
2994 
2995 	index = -ENOSPC;
2996 
2997 out:
2998 	raw_spin_unlock_irqrestore(&table->lock, flags);
2999 
3000 	return index;
3001 }
3002 
3003 static int modify_irte_ga(struct amd_iommu *iommu, u16 devid, int index,
3004 			  struct irte_ga *irte, struct amd_ir_data *data)
3005 {
3006 	bool ret;
3007 	struct irq_remap_table *table;
3008 	unsigned long flags;
3009 	struct irte_ga *entry;
3010 
3011 	table = get_irq_table(iommu, devid);
3012 	if (!table)
3013 		return -ENOMEM;
3014 
3015 	raw_spin_lock_irqsave(&table->lock, flags);
3016 
3017 	entry = (struct irte_ga *)table->table;
3018 	entry = &entry[index];
3019 
3020 	ret = cmpxchg_double(&entry->lo.val, &entry->hi.val,
3021 			     entry->lo.val, entry->hi.val,
3022 			     irte->lo.val, irte->hi.val);
3023 	/*
3024 	 * We use cmpxchg16 to atomically update the 128-bit IRTE,
3025 	 * and it cannot be updated by the hardware or other processors
3026 	 * behind us, so the return value of cmpxchg16 should be the
3027 	 * same as the old value.
3028 	 */
3029 	WARN_ON(!ret);
3030 
3031 	if (data)
3032 		data->ref = entry;
3033 
3034 	raw_spin_unlock_irqrestore(&table->lock, flags);
3035 
3036 	iommu_flush_irt(iommu, devid);
3037 	iommu_completion_wait(iommu);
3038 
3039 	return 0;
3040 }
3041 
3042 static int modify_irte(struct amd_iommu *iommu,
3043 		       u16 devid, int index, union irte *irte)
3044 {
3045 	struct irq_remap_table *table;
3046 	unsigned long flags;
3047 
3048 	table = get_irq_table(iommu, devid);
3049 	if (!table)
3050 		return -ENOMEM;
3051 
3052 	raw_spin_lock_irqsave(&table->lock, flags);
3053 	table->table[index] = irte->val;
3054 	raw_spin_unlock_irqrestore(&table->lock, flags);
3055 
3056 	iommu_flush_irt(iommu, devid);
3057 	iommu_completion_wait(iommu);
3058 
3059 	return 0;
3060 }
3061 
3062 static void free_irte(struct amd_iommu *iommu, u16 devid, int index)
3063 {
3064 	struct irq_remap_table *table;
3065 	unsigned long flags;
3066 
3067 	table = get_irq_table(iommu, devid);
3068 	if (!table)
3069 		return;
3070 
3071 	raw_spin_lock_irqsave(&table->lock, flags);
3072 	iommu->irte_ops->clear_allocated(table, index);
3073 	raw_spin_unlock_irqrestore(&table->lock, flags);
3074 
3075 	iommu_flush_irt(iommu, devid);
3076 	iommu_completion_wait(iommu);
3077 }
3078 
3079 static void irte_prepare(void *entry,
3080 			 u32 delivery_mode, bool dest_mode,
3081 			 u8 vector, u32 dest_apicid, int devid)
3082 {
3083 	union irte *irte = (union irte *) entry;
3084 
3085 	irte->val                = 0;
3086 	irte->fields.vector      = vector;
3087 	irte->fields.int_type    = delivery_mode;
3088 	irte->fields.destination = dest_apicid;
3089 	irte->fields.dm          = dest_mode;
3090 	irte->fields.valid       = 1;
3091 }
3092 
3093 static void irte_ga_prepare(void *entry,
3094 			    u32 delivery_mode, bool dest_mode,
3095 			    u8 vector, u32 dest_apicid, int devid)
3096 {
3097 	struct irte_ga *irte = (struct irte_ga *) entry;
3098 
3099 	irte->lo.val                      = 0;
3100 	irte->hi.val                      = 0;
3101 	irte->lo.fields_remap.int_type    = delivery_mode;
3102 	irte->lo.fields_remap.dm          = dest_mode;
3103 	irte->hi.fields.vector            = vector;
3104 	irte->lo.fields_remap.destination = APICID_TO_IRTE_DEST_LO(dest_apicid);
3105 	irte->hi.fields.destination       = APICID_TO_IRTE_DEST_HI(dest_apicid);
3106 	irte->lo.fields_remap.valid       = 1;
3107 }
3108 
3109 static void irte_activate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3110 {
3111 	union irte *irte = (union irte *) entry;
3112 
3113 	irte->fields.valid = 1;
3114 	modify_irte(iommu, devid, index, irte);
3115 }
3116 
3117 static void irte_ga_activate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3118 {
3119 	struct irte_ga *irte = (struct irte_ga *) entry;
3120 
3121 	irte->lo.fields_remap.valid = 1;
3122 	modify_irte_ga(iommu, devid, index, irte, NULL);
3123 }
3124 
3125 static void irte_deactivate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3126 {
3127 	union irte *irte = (union irte *) entry;
3128 
3129 	irte->fields.valid = 0;
3130 	modify_irte(iommu, devid, index, irte);
3131 }
3132 
3133 static void irte_ga_deactivate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3134 {
3135 	struct irte_ga *irte = (struct irte_ga *) entry;
3136 
3137 	irte->lo.fields_remap.valid = 0;
3138 	modify_irte_ga(iommu, devid, index, irte, NULL);
3139 }
3140 
3141 static void irte_set_affinity(struct amd_iommu *iommu, void *entry, u16 devid, u16 index,
3142 			      u8 vector, u32 dest_apicid)
3143 {
3144 	union irte *irte = (union irte *) entry;
3145 
3146 	irte->fields.vector = vector;
3147 	irte->fields.destination = dest_apicid;
3148 	modify_irte(iommu, devid, index, irte);
3149 }
3150 
3151 static void irte_ga_set_affinity(struct amd_iommu *iommu, void *entry, u16 devid, u16 index,
3152 				 u8 vector, u32 dest_apicid)
3153 {
3154 	struct irte_ga *irte = (struct irte_ga *) entry;
3155 
3156 	if (!irte->lo.fields_remap.guest_mode) {
3157 		irte->hi.fields.vector = vector;
3158 		irte->lo.fields_remap.destination =
3159 					APICID_TO_IRTE_DEST_LO(dest_apicid);
3160 		irte->hi.fields.destination =
3161 					APICID_TO_IRTE_DEST_HI(dest_apicid);
3162 		modify_irte_ga(iommu, devid, index, irte, NULL);
3163 	}
3164 }
3165 
3166 #define IRTE_ALLOCATED (~1U)
3167 static void irte_set_allocated(struct irq_remap_table *table, int index)
3168 {
3169 	table->table[index] = IRTE_ALLOCATED;
3170 }
3171 
3172 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
3173 {
3174 	struct irte_ga *ptr = (struct irte_ga *)table->table;
3175 	struct irte_ga *irte = &ptr[index];
3176 
3177 	memset(&irte->lo.val, 0, sizeof(u64));
3178 	memset(&irte->hi.val, 0, sizeof(u64));
3179 	irte->hi.fields.vector = 0xff;
3180 }
3181 
3182 static bool irte_is_allocated(struct irq_remap_table *table, int index)
3183 {
3184 	union irte *ptr = (union irte *)table->table;
3185 	union irte *irte = &ptr[index];
3186 
3187 	return irte->val != 0;
3188 }
3189 
3190 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
3191 {
3192 	struct irte_ga *ptr = (struct irte_ga *)table->table;
3193 	struct irte_ga *irte = &ptr[index];
3194 
3195 	return irte->hi.fields.vector != 0;
3196 }
3197 
3198 static void irte_clear_allocated(struct irq_remap_table *table, int index)
3199 {
3200 	table->table[index] = 0;
3201 }
3202 
3203 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
3204 {
3205 	struct irte_ga *ptr = (struct irte_ga *)table->table;
3206 	struct irte_ga *irte = &ptr[index];
3207 
3208 	memset(&irte->lo.val, 0, sizeof(u64));
3209 	memset(&irte->hi.val, 0, sizeof(u64));
3210 }
3211 
3212 static int get_devid(struct irq_alloc_info *info)
3213 {
3214 	switch (info->type) {
3215 	case X86_IRQ_ALLOC_TYPE_IOAPIC:
3216 		return get_ioapic_devid(info->devid);
3217 	case X86_IRQ_ALLOC_TYPE_HPET:
3218 		return get_hpet_devid(info->devid);
3219 	case X86_IRQ_ALLOC_TYPE_PCI_MSI:
3220 	case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
3221 		return get_device_sbdf_id(msi_desc_to_dev(info->desc));
3222 	default:
3223 		WARN_ON_ONCE(1);
3224 		return -1;
3225 	}
3226 }
3227 
3228 struct irq_remap_ops amd_iommu_irq_ops = {
3229 	.prepare		= amd_iommu_prepare,
3230 	.enable			= amd_iommu_enable,
3231 	.disable		= amd_iommu_disable,
3232 	.reenable		= amd_iommu_reenable,
3233 	.enable_faulting	= amd_iommu_enable_faulting,
3234 };
3235 
3236 static void fill_msi_msg(struct msi_msg *msg, u32 index)
3237 {
3238 	msg->data = index;
3239 	msg->address_lo = 0;
3240 	msg->arch_addr_lo.base_address = X86_MSI_BASE_ADDRESS_LOW;
3241 	msg->address_hi = X86_MSI_BASE_ADDRESS_HIGH;
3242 }
3243 
3244 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3245 				       struct irq_cfg *irq_cfg,
3246 				       struct irq_alloc_info *info,
3247 				       int devid, int index, int sub_handle)
3248 {
3249 	struct irq_2_irte *irte_info = &data->irq_2_irte;
3250 	struct amd_iommu *iommu = data->iommu;
3251 
3252 	if (!iommu)
3253 		return;
3254 
3255 	data->irq_2_irte.devid = devid;
3256 	data->irq_2_irte.index = index + sub_handle;
3257 	iommu->irte_ops->prepare(data->entry, apic->delivery_mode,
3258 				 apic->dest_mode_logical, irq_cfg->vector,
3259 				 irq_cfg->dest_apicid, devid);
3260 
3261 	switch (info->type) {
3262 	case X86_IRQ_ALLOC_TYPE_IOAPIC:
3263 	case X86_IRQ_ALLOC_TYPE_HPET:
3264 	case X86_IRQ_ALLOC_TYPE_PCI_MSI:
3265 	case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
3266 		fill_msi_msg(&data->msi_entry, irte_info->index);
3267 		break;
3268 
3269 	default:
3270 		BUG_ON(1);
3271 		break;
3272 	}
3273 }
3274 
3275 struct amd_irte_ops irte_32_ops = {
3276 	.prepare = irte_prepare,
3277 	.activate = irte_activate,
3278 	.deactivate = irte_deactivate,
3279 	.set_affinity = irte_set_affinity,
3280 	.set_allocated = irte_set_allocated,
3281 	.is_allocated = irte_is_allocated,
3282 	.clear_allocated = irte_clear_allocated,
3283 };
3284 
3285 struct amd_irte_ops irte_128_ops = {
3286 	.prepare = irte_ga_prepare,
3287 	.activate = irte_ga_activate,
3288 	.deactivate = irte_ga_deactivate,
3289 	.set_affinity = irte_ga_set_affinity,
3290 	.set_allocated = irte_ga_set_allocated,
3291 	.is_allocated = irte_ga_is_allocated,
3292 	.clear_allocated = irte_ga_clear_allocated,
3293 };
3294 
3295 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
3296 			       unsigned int nr_irqs, void *arg)
3297 {
3298 	struct irq_alloc_info *info = arg;
3299 	struct irq_data *irq_data;
3300 	struct amd_ir_data *data = NULL;
3301 	struct amd_iommu *iommu;
3302 	struct irq_cfg *cfg;
3303 	int i, ret, devid, seg, sbdf;
3304 	int index;
3305 
3306 	if (!info)
3307 		return -EINVAL;
3308 	if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_PCI_MSI)
3309 		return -EINVAL;
3310 
3311 	sbdf = get_devid(info);
3312 	if (sbdf < 0)
3313 		return -EINVAL;
3314 
3315 	seg = PCI_SBDF_TO_SEGID(sbdf);
3316 	devid = PCI_SBDF_TO_DEVID(sbdf);
3317 	iommu = __rlookup_amd_iommu(seg, devid);
3318 	if (!iommu)
3319 		return -EINVAL;
3320 
3321 	ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
3322 	if (ret < 0)
3323 		return ret;
3324 
3325 	if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
3326 		struct irq_remap_table *table;
3327 
3328 		table = alloc_irq_table(iommu, devid, NULL);
3329 		if (table) {
3330 			if (!table->min_index) {
3331 				/*
3332 				 * Keep the first 32 indexes free for IOAPIC
3333 				 * interrupts.
3334 				 */
3335 				table->min_index = 32;
3336 				for (i = 0; i < 32; ++i)
3337 					iommu->irte_ops->set_allocated(table, i);
3338 			}
3339 			WARN_ON(table->min_index != 32);
3340 			index = info->ioapic.pin;
3341 		} else {
3342 			index = -ENOMEM;
3343 		}
3344 	} else if (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI ||
3345 		   info->type == X86_IRQ_ALLOC_TYPE_PCI_MSIX) {
3346 		bool align = (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI);
3347 
3348 		index = alloc_irq_index(iommu, devid, nr_irqs, align,
3349 					msi_desc_to_pci_dev(info->desc));
3350 	} else {
3351 		index = alloc_irq_index(iommu, devid, nr_irqs, false, NULL);
3352 	}
3353 
3354 	if (index < 0) {
3355 		pr_warn("Failed to allocate IRTE\n");
3356 		ret = index;
3357 		goto out_free_parent;
3358 	}
3359 
3360 	for (i = 0; i < nr_irqs; i++) {
3361 		irq_data = irq_domain_get_irq_data(domain, virq + i);
3362 		cfg = irq_data ? irqd_cfg(irq_data) : NULL;
3363 		if (!cfg) {
3364 			ret = -EINVAL;
3365 			goto out_free_data;
3366 		}
3367 
3368 		ret = -ENOMEM;
3369 		data = kzalloc(sizeof(*data), GFP_KERNEL);
3370 		if (!data)
3371 			goto out_free_data;
3372 
3373 		if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3374 			data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
3375 		else
3376 			data->entry = kzalloc(sizeof(struct irte_ga),
3377 						     GFP_KERNEL);
3378 		if (!data->entry) {
3379 			kfree(data);
3380 			goto out_free_data;
3381 		}
3382 
3383 		data->iommu = iommu;
3384 		irq_data->hwirq = (devid << 16) + i;
3385 		irq_data->chip_data = data;
3386 		irq_data->chip = &amd_ir_chip;
3387 		irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
3388 		irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
3389 	}
3390 
3391 	return 0;
3392 
3393 out_free_data:
3394 	for (i--; i >= 0; i--) {
3395 		irq_data = irq_domain_get_irq_data(domain, virq + i);
3396 		if (irq_data)
3397 			kfree(irq_data->chip_data);
3398 	}
3399 	for (i = 0; i < nr_irqs; i++)
3400 		free_irte(iommu, devid, index + i);
3401 out_free_parent:
3402 	irq_domain_free_irqs_common(domain, virq, nr_irqs);
3403 	return ret;
3404 }
3405 
3406 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
3407 			       unsigned int nr_irqs)
3408 {
3409 	struct irq_2_irte *irte_info;
3410 	struct irq_data *irq_data;
3411 	struct amd_ir_data *data;
3412 	int i;
3413 
3414 	for (i = 0; i < nr_irqs; i++) {
3415 		irq_data = irq_domain_get_irq_data(domain, virq  + i);
3416 		if (irq_data && irq_data->chip_data) {
3417 			data = irq_data->chip_data;
3418 			irte_info = &data->irq_2_irte;
3419 			free_irte(data->iommu, irte_info->devid, irte_info->index);
3420 			kfree(data->entry);
3421 			kfree(data);
3422 		}
3423 	}
3424 	irq_domain_free_irqs_common(domain, virq, nr_irqs);
3425 }
3426 
3427 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
3428 			       struct amd_ir_data *ir_data,
3429 			       struct irq_2_irte *irte_info,
3430 			       struct irq_cfg *cfg);
3431 
3432 static int irq_remapping_activate(struct irq_domain *domain,
3433 				  struct irq_data *irq_data, bool reserve)
3434 {
3435 	struct amd_ir_data *data = irq_data->chip_data;
3436 	struct irq_2_irte *irte_info = &data->irq_2_irte;
3437 	struct amd_iommu *iommu = data->iommu;
3438 	struct irq_cfg *cfg = irqd_cfg(irq_data);
3439 
3440 	if (!iommu)
3441 		return 0;
3442 
3443 	iommu->irte_ops->activate(iommu, data->entry, irte_info->devid,
3444 				  irte_info->index);
3445 	amd_ir_update_irte(irq_data, iommu, data, irte_info, cfg);
3446 	return 0;
3447 }
3448 
3449 static void irq_remapping_deactivate(struct irq_domain *domain,
3450 				     struct irq_data *irq_data)
3451 {
3452 	struct amd_ir_data *data = irq_data->chip_data;
3453 	struct irq_2_irte *irte_info = &data->irq_2_irte;
3454 	struct amd_iommu *iommu = data->iommu;
3455 
3456 	if (iommu)
3457 		iommu->irte_ops->deactivate(iommu, data->entry, irte_info->devid,
3458 					    irte_info->index);
3459 }
3460 
3461 static int irq_remapping_select(struct irq_domain *d, struct irq_fwspec *fwspec,
3462 				enum irq_domain_bus_token bus_token)
3463 {
3464 	struct amd_iommu *iommu;
3465 	int devid = -1;
3466 
3467 	if (!amd_iommu_irq_remap)
3468 		return 0;
3469 
3470 	if (x86_fwspec_is_ioapic(fwspec))
3471 		devid = get_ioapic_devid(fwspec->param[0]);
3472 	else if (x86_fwspec_is_hpet(fwspec))
3473 		devid = get_hpet_devid(fwspec->param[0]);
3474 
3475 	if (devid < 0)
3476 		return 0;
3477 	iommu = __rlookup_amd_iommu((devid >> 16), (devid & 0xffff));
3478 
3479 	return iommu && iommu->ir_domain == d;
3480 }
3481 
3482 static const struct irq_domain_ops amd_ir_domain_ops = {
3483 	.select = irq_remapping_select,
3484 	.alloc = irq_remapping_alloc,
3485 	.free = irq_remapping_free,
3486 	.activate = irq_remapping_activate,
3487 	.deactivate = irq_remapping_deactivate,
3488 };
3489 
3490 int amd_iommu_activate_guest_mode(void *data)
3491 {
3492 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3493 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3494 	u64 valid;
3495 
3496 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3497 	    !entry || entry->lo.fields_vapic.guest_mode)
3498 		return 0;
3499 
3500 	valid = entry->lo.fields_vapic.valid;
3501 
3502 	entry->lo.val = 0;
3503 	entry->hi.val = 0;
3504 
3505 	entry->lo.fields_vapic.valid       = valid;
3506 	entry->lo.fields_vapic.guest_mode  = 1;
3507 	entry->lo.fields_vapic.ga_log_intr = 1;
3508 	entry->hi.fields.ga_root_ptr       = ir_data->ga_root_ptr;
3509 	entry->hi.fields.vector            = ir_data->ga_vector;
3510 	entry->lo.fields_vapic.ga_tag      = ir_data->ga_tag;
3511 
3512 	return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3513 			      ir_data->irq_2_irte.index, entry, ir_data);
3514 }
3515 EXPORT_SYMBOL(amd_iommu_activate_guest_mode);
3516 
3517 int amd_iommu_deactivate_guest_mode(void *data)
3518 {
3519 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3520 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3521 	struct irq_cfg *cfg = ir_data->cfg;
3522 	u64 valid;
3523 
3524 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3525 	    !entry || !entry->lo.fields_vapic.guest_mode)
3526 		return 0;
3527 
3528 	valid = entry->lo.fields_remap.valid;
3529 
3530 	entry->lo.val = 0;
3531 	entry->hi.val = 0;
3532 
3533 	entry->lo.fields_remap.valid       = valid;
3534 	entry->lo.fields_remap.dm          = apic->dest_mode_logical;
3535 	entry->lo.fields_remap.int_type    = apic->delivery_mode;
3536 	entry->hi.fields.vector            = cfg->vector;
3537 	entry->lo.fields_remap.destination =
3538 				APICID_TO_IRTE_DEST_LO(cfg->dest_apicid);
3539 	entry->hi.fields.destination =
3540 				APICID_TO_IRTE_DEST_HI(cfg->dest_apicid);
3541 
3542 	return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3543 			      ir_data->irq_2_irte.index, entry, ir_data);
3544 }
3545 EXPORT_SYMBOL(amd_iommu_deactivate_guest_mode);
3546 
3547 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
3548 {
3549 	int ret;
3550 	struct amd_iommu_pi_data *pi_data = vcpu_info;
3551 	struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
3552 	struct amd_ir_data *ir_data = data->chip_data;
3553 	struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3554 	struct iommu_dev_data *dev_data;
3555 
3556 	if (ir_data->iommu == NULL)
3557 		return -EINVAL;
3558 
3559 	dev_data = search_dev_data(ir_data->iommu, irte_info->devid);
3560 
3561 	/* Note:
3562 	 * This device has never been set up for guest mode.
3563 	 * we should not modify the IRTE
3564 	 */
3565 	if (!dev_data || !dev_data->use_vapic)
3566 		return 0;
3567 
3568 	ir_data->cfg = irqd_cfg(data);
3569 	pi_data->ir_data = ir_data;
3570 
3571 	/* Note:
3572 	 * SVM tries to set up for VAPIC mode, but we are in
3573 	 * legacy mode. So, we force legacy mode instead.
3574 	 */
3575 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3576 		pr_debug("%s: Fall back to using intr legacy remap\n",
3577 			 __func__);
3578 		pi_data->is_guest_mode = false;
3579 	}
3580 
3581 	pi_data->prev_ga_tag = ir_data->cached_ga_tag;
3582 	if (pi_data->is_guest_mode) {
3583 		ir_data->ga_root_ptr = (pi_data->base >> 12);
3584 		ir_data->ga_vector = vcpu_pi_info->vector;
3585 		ir_data->ga_tag = pi_data->ga_tag;
3586 		ret = amd_iommu_activate_guest_mode(ir_data);
3587 		if (!ret)
3588 			ir_data->cached_ga_tag = pi_data->ga_tag;
3589 	} else {
3590 		ret = amd_iommu_deactivate_guest_mode(ir_data);
3591 
3592 		/*
3593 		 * This communicates the ga_tag back to the caller
3594 		 * so that it can do all the necessary clean up.
3595 		 */
3596 		if (!ret)
3597 			ir_data->cached_ga_tag = 0;
3598 	}
3599 
3600 	return ret;
3601 }
3602 
3603 
3604 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
3605 			       struct amd_ir_data *ir_data,
3606 			       struct irq_2_irte *irte_info,
3607 			       struct irq_cfg *cfg)
3608 {
3609 
3610 	/*
3611 	 * Atomically updates the IRTE with the new destination, vector
3612 	 * and flushes the interrupt entry cache.
3613 	 */
3614 	iommu->irte_ops->set_affinity(iommu, ir_data->entry, irte_info->devid,
3615 				      irte_info->index, cfg->vector,
3616 				      cfg->dest_apicid);
3617 }
3618 
3619 static int amd_ir_set_affinity(struct irq_data *data,
3620 			       const struct cpumask *mask, bool force)
3621 {
3622 	struct amd_ir_data *ir_data = data->chip_data;
3623 	struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3624 	struct irq_cfg *cfg = irqd_cfg(data);
3625 	struct irq_data *parent = data->parent_data;
3626 	struct amd_iommu *iommu = ir_data->iommu;
3627 	int ret;
3628 
3629 	if (!iommu)
3630 		return -ENODEV;
3631 
3632 	ret = parent->chip->irq_set_affinity(parent, mask, force);
3633 	if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
3634 		return ret;
3635 
3636 	amd_ir_update_irte(data, iommu, ir_data, irte_info, cfg);
3637 	/*
3638 	 * After this point, all the interrupts will start arriving
3639 	 * at the new destination. So, time to cleanup the previous
3640 	 * vector allocation.
3641 	 */
3642 	send_cleanup_vector(cfg);
3643 
3644 	return IRQ_SET_MASK_OK_DONE;
3645 }
3646 
3647 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
3648 {
3649 	struct amd_ir_data *ir_data = irq_data->chip_data;
3650 
3651 	*msg = ir_data->msi_entry;
3652 }
3653 
3654 static struct irq_chip amd_ir_chip = {
3655 	.name			= "AMD-IR",
3656 	.irq_ack		= apic_ack_irq,
3657 	.irq_set_affinity	= amd_ir_set_affinity,
3658 	.irq_set_vcpu_affinity	= amd_ir_set_vcpu_affinity,
3659 	.irq_compose_msi_msg	= ir_compose_msi_msg,
3660 };
3661 
3662 static const struct msi_parent_ops amdvi_msi_parent_ops = {
3663 	.supported_flags	= X86_VECTOR_MSI_FLAGS_SUPPORTED |
3664 				  MSI_FLAG_MULTI_PCI_MSI |
3665 				  MSI_FLAG_PCI_IMS,
3666 	.prefix			= "IR-",
3667 	.init_dev_msi_info	= msi_parent_init_dev_msi_info,
3668 };
3669 
3670 static const struct msi_parent_ops virt_amdvi_msi_parent_ops = {
3671 	.supported_flags	= X86_VECTOR_MSI_FLAGS_SUPPORTED |
3672 				  MSI_FLAG_MULTI_PCI_MSI,
3673 	.prefix			= "vIR-",
3674 	.init_dev_msi_info	= msi_parent_init_dev_msi_info,
3675 };
3676 
3677 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
3678 {
3679 	struct fwnode_handle *fn;
3680 
3681 	fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index);
3682 	if (!fn)
3683 		return -ENOMEM;
3684 	iommu->ir_domain = irq_domain_create_hierarchy(arch_get_ir_parent_domain(), 0, 0,
3685 						       fn, &amd_ir_domain_ops, iommu);
3686 	if (!iommu->ir_domain) {
3687 		irq_domain_free_fwnode(fn);
3688 		return -ENOMEM;
3689 	}
3690 
3691 	irq_domain_update_bus_token(iommu->ir_domain,  DOMAIN_BUS_AMDVI);
3692 	iommu->ir_domain->flags |= IRQ_DOMAIN_FLAG_MSI_PARENT |
3693 				   IRQ_DOMAIN_FLAG_ISOLATED_MSI;
3694 
3695 	if (amd_iommu_np_cache)
3696 		iommu->ir_domain->msi_parent_ops = &virt_amdvi_msi_parent_ops;
3697 	else
3698 		iommu->ir_domain->msi_parent_ops = &amdvi_msi_parent_ops;
3699 
3700 	return 0;
3701 }
3702 
3703 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
3704 {
3705 	unsigned long flags;
3706 	struct amd_iommu *iommu;
3707 	struct irq_remap_table *table;
3708 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3709 	int devid = ir_data->irq_2_irte.devid;
3710 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3711 	struct irte_ga *ref = (struct irte_ga *) ir_data->ref;
3712 
3713 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3714 	    !ref || !entry || !entry->lo.fields_vapic.guest_mode)
3715 		return 0;
3716 
3717 	iommu = ir_data->iommu;
3718 	if (!iommu)
3719 		return -ENODEV;
3720 
3721 	table = get_irq_table(iommu, devid);
3722 	if (!table)
3723 		return -ENODEV;
3724 
3725 	raw_spin_lock_irqsave(&table->lock, flags);
3726 
3727 	if (ref->lo.fields_vapic.guest_mode) {
3728 		if (cpu >= 0) {
3729 			ref->lo.fields_vapic.destination =
3730 						APICID_TO_IRTE_DEST_LO(cpu);
3731 			ref->hi.fields.destination =
3732 						APICID_TO_IRTE_DEST_HI(cpu);
3733 		}
3734 		ref->lo.fields_vapic.is_run = is_run;
3735 		barrier();
3736 	}
3737 
3738 	raw_spin_unlock_irqrestore(&table->lock, flags);
3739 
3740 	iommu_flush_irt(iommu, devid);
3741 	iommu_completion_wait(iommu);
3742 	return 0;
3743 }
3744 EXPORT_SYMBOL(amd_iommu_update_ga);
3745 #endif
3746