xref: /linux/drivers/iommu/amd/iommu.c (revision 314f14abdeca78de6b16f97d796a9966ce4b90ae)
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/amba/bus.h>
15 #include <linux/platform_device.h>
16 #include <linux/pci-ats.h>
17 #include <linux/bitmap.h>
18 #include <linux/slab.h>
19 #include <linux/debugfs.h>
20 #include <linux/scatterlist.h>
21 #include <linux/dma-map-ops.h>
22 #include <linux/dma-direct.h>
23 #include <linux/dma-iommu.h>
24 #include <linux/iommu-helper.h>
25 #include <linux/delay.h>
26 #include <linux/amd-iommu.h>
27 #include <linux/notifier.h>
28 #include <linux/export.h>
29 #include <linux/irq.h>
30 #include <linux/msi.h>
31 #include <linux/irqdomain.h>
32 #include <linux/percpu.h>
33 #include <linux/io-pgtable.h>
34 #include <linux/cc_platform.h>
35 #include <asm/irq_remapping.h>
36 #include <asm/io_apic.h>
37 #include <asm/apic.h>
38 #include <asm/hw_irq.h>
39 #include <asm/proto.h>
40 #include <asm/iommu.h>
41 #include <asm/gart.h>
42 #include <asm/dma.h>
43 
44 #include "amd_iommu.h"
45 #include "../irq_remapping.h"
46 
47 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
48 
49 #define LOOP_TIMEOUT	100000
50 
51 /* IO virtual address start page frame number */
52 #define IOVA_START_PFN		(1)
53 #define IOVA_PFN(addr)		((addr) >> PAGE_SHIFT)
54 
55 /* Reserved IOVA ranges */
56 #define MSI_RANGE_START		(0xfee00000)
57 #define MSI_RANGE_END		(0xfeefffff)
58 #define HT_RANGE_START		(0xfd00000000ULL)
59 #define HT_RANGE_END		(0xffffffffffULL)
60 
61 #define DEFAULT_PGTABLE_LEVEL	PAGE_MODE_3_LEVEL
62 
63 static DEFINE_SPINLOCK(pd_bitmap_lock);
64 
65 /* List of all available dev_data structures */
66 static LLIST_HEAD(dev_data_list);
67 
68 LIST_HEAD(ioapic_map);
69 LIST_HEAD(hpet_map);
70 LIST_HEAD(acpihid_map);
71 
72 /*
73  * Domain for untranslated devices - only allocated
74  * if iommu=pt passed on kernel cmd line.
75  */
76 const struct iommu_ops amd_iommu_ops;
77 
78 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
79 int amd_iommu_max_glx_val = -1;
80 
81 /*
82  * general struct to manage commands send to an IOMMU
83  */
84 struct iommu_cmd {
85 	u32 data[4];
86 };
87 
88 struct kmem_cache *amd_iommu_irq_cache;
89 
90 static void detach_device(struct device *dev);
91 
92 /****************************************************************************
93  *
94  * Helper functions
95  *
96  ****************************************************************************/
97 
98 static inline u16 get_pci_device_id(struct device *dev)
99 {
100 	struct pci_dev *pdev = to_pci_dev(dev);
101 
102 	return pci_dev_id(pdev);
103 }
104 
105 static inline int get_acpihid_device_id(struct device *dev,
106 					struct acpihid_map_entry **entry)
107 {
108 	struct acpi_device *adev = ACPI_COMPANION(dev);
109 	struct acpihid_map_entry *p;
110 
111 	if (!adev)
112 		return -ENODEV;
113 
114 	list_for_each_entry(p, &acpihid_map, list) {
115 		if (acpi_dev_hid_uid_match(adev, p->hid,
116 					   p->uid[0] ? p->uid : NULL)) {
117 			if (entry)
118 				*entry = p;
119 			return p->devid;
120 		}
121 	}
122 	return -EINVAL;
123 }
124 
125 static inline int get_device_id(struct device *dev)
126 {
127 	int devid;
128 
129 	if (dev_is_pci(dev))
130 		devid = get_pci_device_id(dev);
131 	else
132 		devid = get_acpihid_device_id(dev, NULL);
133 
134 	return devid;
135 }
136 
137 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
138 {
139 	return container_of(dom, struct protection_domain, domain);
140 }
141 
142 static struct iommu_dev_data *alloc_dev_data(u16 devid)
143 {
144 	struct iommu_dev_data *dev_data;
145 
146 	dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
147 	if (!dev_data)
148 		return NULL;
149 
150 	spin_lock_init(&dev_data->lock);
151 	dev_data->devid = devid;
152 	ratelimit_default_init(&dev_data->rs);
153 
154 	llist_add(&dev_data->dev_data_list, &dev_data_list);
155 	return dev_data;
156 }
157 
158 static struct iommu_dev_data *search_dev_data(u16 devid)
159 {
160 	struct iommu_dev_data *dev_data;
161 	struct llist_node *node;
162 
163 	if (llist_empty(&dev_data_list))
164 		return NULL;
165 
166 	node = dev_data_list.first;
167 	llist_for_each_entry(dev_data, node, dev_data_list) {
168 		if (dev_data->devid == devid)
169 			return dev_data;
170 	}
171 
172 	return NULL;
173 }
174 
175 static int clone_alias(struct pci_dev *pdev, u16 alias, void *data)
176 {
177 	u16 devid = pci_dev_id(pdev);
178 
179 	if (devid == alias)
180 		return 0;
181 
182 	amd_iommu_rlookup_table[alias] =
183 		amd_iommu_rlookup_table[devid];
184 	memcpy(amd_iommu_dev_table[alias].data,
185 	       amd_iommu_dev_table[devid].data,
186 	       sizeof(amd_iommu_dev_table[alias].data));
187 
188 	return 0;
189 }
190 
191 static void clone_aliases(struct pci_dev *pdev)
192 {
193 	if (!pdev)
194 		return;
195 
196 	/*
197 	 * The IVRS alias stored in the alias table may not be
198 	 * part of the PCI DMA aliases if it's bus differs
199 	 * from the original device.
200 	 */
201 	clone_alias(pdev, amd_iommu_alias_table[pci_dev_id(pdev)], NULL);
202 
203 	pci_for_each_dma_alias(pdev, clone_alias, NULL);
204 }
205 
206 static struct pci_dev *setup_aliases(struct device *dev)
207 {
208 	struct pci_dev *pdev = to_pci_dev(dev);
209 	u16 ivrs_alias;
210 
211 	/* For ACPI HID devices, there are no aliases */
212 	if (!dev_is_pci(dev))
213 		return NULL;
214 
215 	/*
216 	 * Add the IVRS alias to the pci aliases if it is on the same
217 	 * bus. The IVRS table may know about a quirk that we don't.
218 	 */
219 	ivrs_alias = amd_iommu_alias_table[pci_dev_id(pdev)];
220 	if (ivrs_alias != pci_dev_id(pdev) &&
221 	    PCI_BUS_NUM(ivrs_alias) == pdev->bus->number)
222 		pci_add_dma_alias(pdev, ivrs_alias & 0xff, 1);
223 
224 	clone_aliases(pdev);
225 
226 	return pdev;
227 }
228 
229 static struct iommu_dev_data *find_dev_data(u16 devid)
230 {
231 	struct iommu_dev_data *dev_data;
232 	struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
233 
234 	dev_data = search_dev_data(devid);
235 
236 	if (dev_data == NULL) {
237 		dev_data = alloc_dev_data(devid);
238 		if (!dev_data)
239 			return NULL;
240 
241 		if (translation_pre_enabled(iommu))
242 			dev_data->defer_attach = true;
243 	}
244 
245 	return dev_data;
246 }
247 
248 /*
249 * Find or create an IOMMU group for a acpihid device.
250 */
251 static struct iommu_group *acpihid_device_group(struct device *dev)
252 {
253 	struct acpihid_map_entry *p, *entry = NULL;
254 	int devid;
255 
256 	devid = get_acpihid_device_id(dev, &entry);
257 	if (devid < 0)
258 		return ERR_PTR(devid);
259 
260 	list_for_each_entry(p, &acpihid_map, list) {
261 		if ((devid == p->devid) && p->group)
262 			entry->group = p->group;
263 	}
264 
265 	if (!entry->group)
266 		entry->group = generic_device_group(dev);
267 	else
268 		iommu_group_ref_get(entry->group);
269 
270 	return entry->group;
271 }
272 
273 static bool pci_iommuv2_capable(struct pci_dev *pdev)
274 {
275 	static const int caps[] = {
276 		PCI_EXT_CAP_ID_PRI,
277 		PCI_EXT_CAP_ID_PASID,
278 	};
279 	int i, pos;
280 
281 	if (!pci_ats_supported(pdev))
282 		return false;
283 
284 	for (i = 0; i < 2; ++i) {
285 		pos = pci_find_ext_capability(pdev, caps[i]);
286 		if (pos == 0)
287 			return false;
288 	}
289 
290 	return true;
291 }
292 
293 /*
294  * This function checks if the driver got a valid device from the caller to
295  * avoid dereferencing invalid pointers.
296  */
297 static bool check_device(struct device *dev)
298 {
299 	int devid;
300 
301 	if (!dev)
302 		return false;
303 
304 	devid = get_device_id(dev);
305 	if (devid < 0)
306 		return false;
307 
308 	/* Out of our scope? */
309 	if (devid > amd_iommu_last_bdf)
310 		return false;
311 
312 	if (amd_iommu_rlookup_table[devid] == NULL)
313 		return false;
314 
315 	return true;
316 }
317 
318 static int iommu_init_device(struct device *dev)
319 {
320 	struct iommu_dev_data *dev_data;
321 	int devid;
322 
323 	if (dev_iommu_priv_get(dev))
324 		return 0;
325 
326 	devid = get_device_id(dev);
327 	if (devid < 0)
328 		return devid;
329 
330 	dev_data = find_dev_data(devid);
331 	if (!dev_data)
332 		return -ENOMEM;
333 
334 	dev_data->pdev = setup_aliases(dev);
335 
336 	/*
337 	 * By default we use passthrough mode for IOMMUv2 capable device.
338 	 * But if amd_iommu=force_isolation is set (e.g. to debug DMA to
339 	 * invalid address), we ignore the capability for the device so
340 	 * it'll be forced to go into translation mode.
341 	 */
342 	if ((iommu_default_passthrough() || !amd_iommu_force_isolation) &&
343 	    dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
344 		struct amd_iommu *iommu;
345 
346 		iommu = amd_iommu_rlookup_table[dev_data->devid];
347 		dev_data->iommu_v2 = iommu->is_iommu_v2;
348 	}
349 
350 	dev_iommu_priv_set(dev, dev_data);
351 
352 	return 0;
353 }
354 
355 static void iommu_ignore_device(struct device *dev)
356 {
357 	int devid;
358 
359 	devid = get_device_id(dev);
360 	if (devid < 0)
361 		return;
362 
363 	amd_iommu_rlookup_table[devid] = NULL;
364 	memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
365 
366 	setup_aliases(dev);
367 }
368 
369 static void amd_iommu_uninit_device(struct device *dev)
370 {
371 	struct iommu_dev_data *dev_data;
372 
373 	dev_data = dev_iommu_priv_get(dev);
374 	if (!dev_data)
375 		return;
376 
377 	if (dev_data->domain)
378 		detach_device(dev);
379 
380 	dev_iommu_priv_set(dev, NULL);
381 
382 	/*
383 	 * We keep dev_data around for unplugged devices and reuse it when the
384 	 * device is re-plugged - not doing so would introduce a ton of races.
385 	 */
386 }
387 
388 /****************************************************************************
389  *
390  * Interrupt handling functions
391  *
392  ****************************************************************************/
393 
394 static void dump_dte_entry(u16 devid)
395 {
396 	int i;
397 
398 	for (i = 0; i < 4; ++i)
399 		pr_err("DTE[%d]: %016llx\n", i,
400 			amd_iommu_dev_table[devid].data[i]);
401 }
402 
403 static void dump_command(unsigned long phys_addr)
404 {
405 	struct iommu_cmd *cmd = iommu_phys_to_virt(phys_addr);
406 	int i;
407 
408 	for (i = 0; i < 4; ++i)
409 		pr_err("CMD[%d]: %08x\n", i, cmd->data[i]);
410 }
411 
412 static void amd_iommu_report_rmp_hw_error(volatile u32 *event)
413 {
414 	struct iommu_dev_data *dev_data = NULL;
415 	int devid, vmg_tag, flags;
416 	struct pci_dev *pdev;
417 	u64 spa;
418 
419 	devid   = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
420 	vmg_tag = (event[1]) & 0xFFFF;
421 	flags   = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
422 	spa     = ((u64)event[3] << 32) | (event[2] & 0xFFFFFFF8);
423 
424 	pdev = pci_get_domain_bus_and_slot(0, PCI_BUS_NUM(devid),
425 					   devid & 0xff);
426 	if (pdev)
427 		dev_data = dev_iommu_priv_get(&pdev->dev);
428 
429 	if (dev_data) {
430 		if (__ratelimit(&dev_data->rs)) {
431 			pci_err(pdev, "Event logged [RMP_HW_ERROR vmg_tag=0x%04x, spa=0x%llx, flags=0x%04x]\n",
432 				vmg_tag, spa, flags);
433 		}
434 	} else {
435 		pr_err_ratelimited("Event logged [RMP_HW_ERROR device=%02x:%02x.%x, vmg_tag=0x%04x, spa=0x%llx, flags=0x%04x]\n",
436 			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
437 			vmg_tag, spa, flags);
438 	}
439 
440 	if (pdev)
441 		pci_dev_put(pdev);
442 }
443 
444 static void amd_iommu_report_rmp_fault(volatile u32 *event)
445 {
446 	struct iommu_dev_data *dev_data = NULL;
447 	int devid, flags_rmp, vmg_tag, flags;
448 	struct pci_dev *pdev;
449 	u64 gpa;
450 
451 	devid     = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
452 	flags_rmp = (event[0] >> EVENT_FLAGS_SHIFT) & 0xFF;
453 	vmg_tag   = (event[1]) & 0xFFFF;
454 	flags     = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
455 	gpa       = ((u64)event[3] << 32) | event[2];
456 
457 	pdev = pci_get_domain_bus_and_slot(0, PCI_BUS_NUM(devid),
458 					   devid & 0xff);
459 	if (pdev)
460 		dev_data = dev_iommu_priv_get(&pdev->dev);
461 
462 	if (dev_data) {
463 		if (__ratelimit(&dev_data->rs)) {
464 			pci_err(pdev, "Event logged [RMP_PAGE_FAULT vmg_tag=0x%04x, gpa=0x%llx, flags_rmp=0x%04x, flags=0x%04x]\n",
465 				vmg_tag, gpa, flags_rmp, flags);
466 		}
467 	} else {
468 		pr_err_ratelimited("Event logged [RMP_PAGE_FAULT device=%02x:%02x.%x, vmg_tag=0x%04x, gpa=0x%llx, flags_rmp=0x%04x, flags=0x%04x]\n",
469 			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
470 			vmg_tag, gpa, flags_rmp, flags);
471 	}
472 
473 	if (pdev)
474 		pci_dev_put(pdev);
475 }
476 
477 static void amd_iommu_report_page_fault(u16 devid, u16 domain_id,
478 					u64 address, int flags)
479 {
480 	struct iommu_dev_data *dev_data = NULL;
481 	struct pci_dev *pdev;
482 
483 	pdev = pci_get_domain_bus_and_slot(0, 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 [IO_PAGE_FAULT domain=0x%04x address=0x%llx flags=0x%04x]\n",
491 				domain_id, address, flags);
492 		}
493 	} else {
494 		pr_err_ratelimited("Event logged [IO_PAGE_FAULT device=%02x:%02x.%x domain=0x%04x address=0x%llx flags=0x%04x]\n",
495 			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
496 			domain_id, address, flags);
497 	}
498 
499 	if (pdev)
500 		pci_dev_put(pdev);
501 }
502 
503 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
504 {
505 	struct device *dev = iommu->iommu.dev;
506 	int type, devid, flags, tag;
507 	volatile u32 *event = __evt;
508 	int count = 0;
509 	u64 address;
510 	u32 pasid;
511 
512 retry:
513 	type    = (event[1] >> EVENT_TYPE_SHIFT)  & EVENT_TYPE_MASK;
514 	devid   = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
515 	pasid   = (event[0] & EVENT_DOMID_MASK_HI) |
516 		  (event[1] & EVENT_DOMID_MASK_LO);
517 	flags   = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
518 	address = (u64)(((u64)event[3]) << 32) | event[2];
519 
520 	if (type == 0) {
521 		/* Did we hit the erratum? */
522 		if (++count == LOOP_TIMEOUT) {
523 			pr_err("No event written to event log\n");
524 			return;
525 		}
526 		udelay(1);
527 		goto retry;
528 	}
529 
530 	if (type == EVENT_TYPE_IO_FAULT) {
531 		amd_iommu_report_page_fault(devid, pasid, address, flags);
532 		return;
533 	}
534 
535 	switch (type) {
536 	case EVENT_TYPE_ILL_DEV:
537 		dev_err(dev, "Event logged [ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
538 			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
539 			pasid, address, flags);
540 		dump_dte_entry(devid);
541 		break;
542 	case EVENT_TYPE_DEV_TAB_ERR:
543 		dev_err(dev, "Event logged [DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
544 			"address=0x%llx flags=0x%04x]\n",
545 			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
546 			address, flags);
547 		break;
548 	case EVENT_TYPE_PAGE_TAB_ERR:
549 		dev_err(dev, "Event logged [PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x pasid=0x%04x address=0x%llx flags=0x%04x]\n",
550 			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
551 			pasid, address, flags);
552 		break;
553 	case EVENT_TYPE_ILL_CMD:
554 		dev_err(dev, "Event logged [ILLEGAL_COMMAND_ERROR address=0x%llx]\n", address);
555 		dump_command(address);
556 		break;
557 	case EVENT_TYPE_CMD_HARD_ERR:
558 		dev_err(dev, "Event logged [COMMAND_HARDWARE_ERROR address=0x%llx flags=0x%04x]\n",
559 			address, flags);
560 		break;
561 	case EVENT_TYPE_IOTLB_INV_TO:
562 		dev_err(dev, "Event logged [IOTLB_INV_TIMEOUT device=%02x:%02x.%x address=0x%llx]\n",
563 			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
564 			address);
565 		break;
566 	case EVENT_TYPE_INV_DEV_REQ:
567 		dev_err(dev, "Event logged [INVALID_DEVICE_REQUEST device=%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
568 			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
569 			pasid, address, flags);
570 		break;
571 	case EVENT_TYPE_RMP_FAULT:
572 		amd_iommu_report_rmp_fault(event);
573 		break;
574 	case EVENT_TYPE_RMP_HW_ERR:
575 		amd_iommu_report_rmp_hw_error(event);
576 		break;
577 	case EVENT_TYPE_INV_PPR_REQ:
578 		pasid = PPR_PASID(*((u64 *)__evt));
579 		tag = event[1] & 0x03FF;
580 		dev_err(dev, "Event logged [INVALID_PPR_REQUEST device=%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x tag=0x%03x]\n",
581 			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
582 			pasid, address, flags, tag);
583 		break;
584 	default:
585 		dev_err(dev, "Event logged [UNKNOWN event[0]=0x%08x event[1]=0x%08x event[2]=0x%08x event[3]=0x%08x\n",
586 			event[0], event[1], event[2], event[3]);
587 	}
588 
589 	memset(__evt, 0, 4 * sizeof(u32));
590 }
591 
592 static void iommu_poll_events(struct amd_iommu *iommu)
593 {
594 	u32 head, tail;
595 
596 	head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
597 	tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
598 
599 	while (head != tail) {
600 		iommu_print_event(iommu, iommu->evt_buf + head);
601 		head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
602 	}
603 
604 	writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
605 }
606 
607 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
608 {
609 	struct amd_iommu_fault fault;
610 
611 	if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
612 		pr_err_ratelimited("Unknown PPR request received\n");
613 		return;
614 	}
615 
616 	fault.address   = raw[1];
617 	fault.pasid     = PPR_PASID(raw[0]);
618 	fault.device_id = PPR_DEVID(raw[0]);
619 	fault.tag       = PPR_TAG(raw[0]);
620 	fault.flags     = PPR_FLAGS(raw[0]);
621 
622 	atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
623 }
624 
625 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
626 {
627 	u32 head, tail;
628 
629 	if (iommu->ppr_log == NULL)
630 		return;
631 
632 	head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
633 	tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
634 
635 	while (head != tail) {
636 		volatile u64 *raw;
637 		u64 entry[2];
638 		int i;
639 
640 		raw = (u64 *)(iommu->ppr_log + head);
641 
642 		/*
643 		 * Hardware bug: Interrupt may arrive before the entry is
644 		 * written to memory. If this happens we need to wait for the
645 		 * entry to arrive.
646 		 */
647 		for (i = 0; i < LOOP_TIMEOUT; ++i) {
648 			if (PPR_REQ_TYPE(raw[0]) != 0)
649 				break;
650 			udelay(1);
651 		}
652 
653 		/* Avoid memcpy function-call overhead */
654 		entry[0] = raw[0];
655 		entry[1] = raw[1];
656 
657 		/*
658 		 * To detect the hardware bug we need to clear the entry
659 		 * back to zero.
660 		 */
661 		raw[0] = raw[1] = 0UL;
662 
663 		/* Update head pointer of hardware ring-buffer */
664 		head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
665 		writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
666 
667 		/* Handle PPR entry */
668 		iommu_handle_ppr_entry(iommu, entry);
669 
670 		/* Refresh ring-buffer information */
671 		head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
672 		tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
673 	}
674 }
675 
676 #ifdef CONFIG_IRQ_REMAP
677 static int (*iommu_ga_log_notifier)(u32);
678 
679 int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
680 {
681 	iommu_ga_log_notifier = notifier;
682 
683 	return 0;
684 }
685 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
686 
687 static void iommu_poll_ga_log(struct amd_iommu *iommu)
688 {
689 	u32 head, tail, cnt = 0;
690 
691 	if (iommu->ga_log == NULL)
692 		return;
693 
694 	head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
695 	tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
696 
697 	while (head != tail) {
698 		volatile u64 *raw;
699 		u64 log_entry;
700 
701 		raw = (u64 *)(iommu->ga_log + head);
702 		cnt++;
703 
704 		/* Avoid memcpy function-call overhead */
705 		log_entry = *raw;
706 
707 		/* Update head pointer of hardware ring-buffer */
708 		head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
709 		writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
710 
711 		/* Handle GA entry */
712 		switch (GA_REQ_TYPE(log_entry)) {
713 		case GA_GUEST_NR:
714 			if (!iommu_ga_log_notifier)
715 				break;
716 
717 			pr_debug("%s: devid=%#x, ga_tag=%#x\n",
718 				 __func__, GA_DEVID(log_entry),
719 				 GA_TAG(log_entry));
720 
721 			if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
722 				pr_err("GA log notifier failed.\n");
723 			break;
724 		default:
725 			break;
726 		}
727 	}
728 }
729 
730 static void
731 amd_iommu_set_pci_msi_domain(struct device *dev, struct amd_iommu *iommu)
732 {
733 	if (!irq_remapping_enabled || !dev_is_pci(dev) ||
734 	    pci_dev_has_special_msi_domain(to_pci_dev(dev)))
735 		return;
736 
737 	dev_set_msi_domain(dev, iommu->msi_domain);
738 }
739 
740 #else /* CONFIG_IRQ_REMAP */
741 static inline void
742 amd_iommu_set_pci_msi_domain(struct device *dev, struct amd_iommu *iommu) { }
743 #endif /* !CONFIG_IRQ_REMAP */
744 
745 #define AMD_IOMMU_INT_MASK	\
746 	(MMIO_STATUS_EVT_INT_MASK | \
747 	 MMIO_STATUS_PPR_INT_MASK | \
748 	 MMIO_STATUS_GALOG_INT_MASK)
749 
750 irqreturn_t amd_iommu_int_thread(int irq, void *data)
751 {
752 	struct amd_iommu *iommu = (struct amd_iommu *) data;
753 	u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
754 
755 	while (status & AMD_IOMMU_INT_MASK) {
756 		/* Enable EVT and PPR and GA interrupts again */
757 		writel(AMD_IOMMU_INT_MASK,
758 			iommu->mmio_base + MMIO_STATUS_OFFSET);
759 
760 		if (status & MMIO_STATUS_EVT_INT_MASK) {
761 			pr_devel("Processing IOMMU Event Log\n");
762 			iommu_poll_events(iommu);
763 		}
764 
765 		if (status & MMIO_STATUS_PPR_INT_MASK) {
766 			pr_devel("Processing IOMMU PPR Log\n");
767 			iommu_poll_ppr_log(iommu);
768 		}
769 
770 #ifdef CONFIG_IRQ_REMAP
771 		if (status & MMIO_STATUS_GALOG_INT_MASK) {
772 			pr_devel("Processing IOMMU GA Log\n");
773 			iommu_poll_ga_log(iommu);
774 		}
775 #endif
776 
777 		/*
778 		 * Hardware bug: ERBT1312
779 		 * When re-enabling interrupt (by writing 1
780 		 * to clear the bit), the hardware might also try to set
781 		 * the interrupt bit in the event status register.
782 		 * In this scenario, the bit will be set, and disable
783 		 * subsequent interrupts.
784 		 *
785 		 * Workaround: The IOMMU driver should read back the
786 		 * status register and check if the interrupt bits are cleared.
787 		 * If not, driver will need to go through the interrupt handler
788 		 * again and re-clear the bits
789 		 */
790 		status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
791 	}
792 	return IRQ_HANDLED;
793 }
794 
795 irqreturn_t amd_iommu_int_handler(int irq, void *data)
796 {
797 	return IRQ_WAKE_THREAD;
798 }
799 
800 /****************************************************************************
801  *
802  * IOMMU command queuing functions
803  *
804  ****************************************************************************/
805 
806 static int wait_on_sem(struct amd_iommu *iommu, u64 data)
807 {
808 	int i = 0;
809 
810 	while (*iommu->cmd_sem != data && i < LOOP_TIMEOUT) {
811 		udelay(1);
812 		i += 1;
813 	}
814 
815 	if (i == LOOP_TIMEOUT) {
816 		pr_alert("Completion-Wait loop timed out\n");
817 		return -EIO;
818 	}
819 
820 	return 0;
821 }
822 
823 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
824 			       struct iommu_cmd *cmd)
825 {
826 	u8 *target;
827 	u32 tail;
828 
829 	/* Copy command to buffer */
830 	tail = iommu->cmd_buf_tail;
831 	target = iommu->cmd_buf + tail;
832 	memcpy(target, cmd, sizeof(*cmd));
833 
834 	tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
835 	iommu->cmd_buf_tail = tail;
836 
837 	/* Tell the IOMMU about it */
838 	writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
839 }
840 
841 static void build_completion_wait(struct iommu_cmd *cmd,
842 				  struct amd_iommu *iommu,
843 				  u64 data)
844 {
845 	u64 paddr = iommu_virt_to_phys((void *)iommu->cmd_sem);
846 
847 	memset(cmd, 0, sizeof(*cmd));
848 	cmd->data[0] = lower_32_bits(paddr) | CMD_COMPL_WAIT_STORE_MASK;
849 	cmd->data[1] = upper_32_bits(paddr);
850 	cmd->data[2] = data;
851 	CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
852 }
853 
854 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
855 {
856 	memset(cmd, 0, sizeof(*cmd));
857 	cmd->data[0] = devid;
858 	CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
859 }
860 
861 /*
862  * Builds an invalidation address which is suitable for one page or multiple
863  * pages. Sets the size bit (S) as needed is more than one page is flushed.
864  */
865 static inline u64 build_inv_address(u64 address, size_t size)
866 {
867 	u64 pages, end, msb_diff;
868 
869 	pages = iommu_num_pages(address, size, PAGE_SIZE);
870 
871 	if (pages == 1)
872 		return address & PAGE_MASK;
873 
874 	end = address + size - 1;
875 
876 	/*
877 	 * msb_diff would hold the index of the most significant bit that
878 	 * flipped between the start and end.
879 	 */
880 	msb_diff = fls64(end ^ address) - 1;
881 
882 	/*
883 	 * Bits 63:52 are sign extended. If for some reason bit 51 is different
884 	 * between the start and the end, invalidate everything.
885 	 */
886 	if (unlikely(msb_diff > 51)) {
887 		address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
888 	} else {
889 		/*
890 		 * The msb-bit must be clear on the address. Just set all the
891 		 * lower bits.
892 		 */
893 		address |= (1ull << msb_diff) - 1;
894 	}
895 
896 	/* Clear bits 11:0 */
897 	address &= PAGE_MASK;
898 
899 	/* Set the size bit - we flush more than one 4kb page */
900 	return address | CMD_INV_IOMMU_PAGES_SIZE_MASK;
901 }
902 
903 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
904 				  size_t size, u16 domid, int pde)
905 {
906 	u64 inv_address = build_inv_address(address, size);
907 
908 	memset(cmd, 0, sizeof(*cmd));
909 	cmd->data[1] |= domid;
910 	cmd->data[2]  = lower_32_bits(inv_address);
911 	cmd->data[3]  = upper_32_bits(inv_address);
912 	CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
913 	if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
914 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
915 }
916 
917 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
918 				  u64 address, size_t size)
919 {
920 	u64 inv_address = build_inv_address(address, size);
921 
922 	memset(cmd, 0, sizeof(*cmd));
923 	cmd->data[0]  = devid;
924 	cmd->data[0] |= (qdep & 0xff) << 24;
925 	cmd->data[1]  = devid;
926 	cmd->data[2]  = lower_32_bits(inv_address);
927 	cmd->data[3]  = upper_32_bits(inv_address);
928 	CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
929 }
930 
931 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, u32 pasid,
932 				  u64 address, bool size)
933 {
934 	memset(cmd, 0, sizeof(*cmd));
935 
936 	address &= ~(0xfffULL);
937 
938 	cmd->data[0]  = pasid;
939 	cmd->data[1]  = domid;
940 	cmd->data[2]  = lower_32_bits(address);
941 	cmd->data[3]  = upper_32_bits(address);
942 	cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
943 	cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
944 	if (size)
945 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
946 	CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
947 }
948 
949 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, u32 pasid,
950 				  int qdep, u64 address, bool size)
951 {
952 	memset(cmd, 0, sizeof(*cmd));
953 
954 	address &= ~(0xfffULL);
955 
956 	cmd->data[0]  = devid;
957 	cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
958 	cmd->data[0] |= (qdep  & 0xff) << 24;
959 	cmd->data[1]  = devid;
960 	cmd->data[1] |= (pasid & 0xff) << 16;
961 	cmd->data[2]  = lower_32_bits(address);
962 	cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
963 	cmd->data[3]  = upper_32_bits(address);
964 	if (size)
965 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
966 	CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
967 }
968 
969 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, u32 pasid,
970 			       int status, int tag, bool gn)
971 {
972 	memset(cmd, 0, sizeof(*cmd));
973 
974 	cmd->data[0]  = devid;
975 	if (gn) {
976 		cmd->data[1]  = pasid;
977 		cmd->data[2]  = CMD_INV_IOMMU_PAGES_GN_MASK;
978 	}
979 	cmd->data[3]  = tag & 0x1ff;
980 	cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
981 
982 	CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
983 }
984 
985 static void build_inv_all(struct iommu_cmd *cmd)
986 {
987 	memset(cmd, 0, sizeof(*cmd));
988 	CMD_SET_TYPE(cmd, CMD_INV_ALL);
989 }
990 
991 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
992 {
993 	memset(cmd, 0, sizeof(*cmd));
994 	cmd->data[0] = devid;
995 	CMD_SET_TYPE(cmd, CMD_INV_IRT);
996 }
997 
998 /*
999  * Writes the command to the IOMMUs command buffer and informs the
1000  * hardware about the new command.
1001  */
1002 static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1003 				      struct iommu_cmd *cmd,
1004 				      bool sync)
1005 {
1006 	unsigned int count = 0;
1007 	u32 left, next_tail;
1008 
1009 	next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1010 again:
1011 	left      = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE;
1012 
1013 	if (left <= 0x20) {
1014 		/* Skip udelay() the first time around */
1015 		if (count++) {
1016 			if (count == LOOP_TIMEOUT) {
1017 				pr_err("Command buffer timeout\n");
1018 				return -EIO;
1019 			}
1020 
1021 			udelay(1);
1022 		}
1023 
1024 		/* Update head and recheck remaining space */
1025 		iommu->cmd_buf_head = readl(iommu->mmio_base +
1026 					    MMIO_CMD_HEAD_OFFSET);
1027 
1028 		goto again;
1029 	}
1030 
1031 	copy_cmd_to_buffer(iommu, cmd);
1032 
1033 	/* Do we need to make sure all commands are processed? */
1034 	iommu->need_sync = sync;
1035 
1036 	return 0;
1037 }
1038 
1039 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1040 				    struct iommu_cmd *cmd,
1041 				    bool sync)
1042 {
1043 	unsigned long flags;
1044 	int ret;
1045 
1046 	raw_spin_lock_irqsave(&iommu->lock, flags);
1047 	ret = __iommu_queue_command_sync(iommu, cmd, sync);
1048 	raw_spin_unlock_irqrestore(&iommu->lock, flags);
1049 
1050 	return ret;
1051 }
1052 
1053 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1054 {
1055 	return iommu_queue_command_sync(iommu, cmd, true);
1056 }
1057 
1058 /*
1059  * This function queues a completion wait command into the command
1060  * buffer of an IOMMU
1061  */
1062 static int iommu_completion_wait(struct amd_iommu *iommu)
1063 {
1064 	struct iommu_cmd cmd;
1065 	unsigned long flags;
1066 	int ret;
1067 	u64 data;
1068 
1069 	if (!iommu->need_sync)
1070 		return 0;
1071 
1072 	raw_spin_lock_irqsave(&iommu->lock, flags);
1073 
1074 	data = ++iommu->cmd_sem_val;
1075 	build_completion_wait(&cmd, iommu, data);
1076 
1077 	ret = __iommu_queue_command_sync(iommu, &cmd, false);
1078 	if (ret)
1079 		goto out_unlock;
1080 
1081 	ret = wait_on_sem(iommu, data);
1082 
1083 out_unlock:
1084 	raw_spin_unlock_irqrestore(&iommu->lock, flags);
1085 
1086 	return ret;
1087 }
1088 
1089 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1090 {
1091 	struct iommu_cmd cmd;
1092 
1093 	build_inv_dte(&cmd, devid);
1094 
1095 	return iommu_queue_command(iommu, &cmd);
1096 }
1097 
1098 static void amd_iommu_flush_dte_all(struct amd_iommu *iommu)
1099 {
1100 	u32 devid;
1101 
1102 	for (devid = 0; devid <= 0xffff; ++devid)
1103 		iommu_flush_dte(iommu, devid);
1104 
1105 	iommu_completion_wait(iommu);
1106 }
1107 
1108 /*
1109  * This function uses heavy locking and may disable irqs for some time. But
1110  * this is no issue because it is only called during resume.
1111  */
1112 static void amd_iommu_flush_tlb_all(struct amd_iommu *iommu)
1113 {
1114 	u32 dom_id;
1115 
1116 	for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1117 		struct iommu_cmd cmd;
1118 		build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1119 				      dom_id, 1);
1120 		iommu_queue_command(iommu, &cmd);
1121 	}
1122 
1123 	iommu_completion_wait(iommu);
1124 }
1125 
1126 static void amd_iommu_flush_tlb_domid(struct amd_iommu *iommu, u32 dom_id)
1127 {
1128 	struct iommu_cmd cmd;
1129 
1130 	build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1131 			      dom_id, 1);
1132 	iommu_queue_command(iommu, &cmd);
1133 
1134 	iommu_completion_wait(iommu);
1135 }
1136 
1137 static void amd_iommu_flush_all(struct amd_iommu *iommu)
1138 {
1139 	struct iommu_cmd cmd;
1140 
1141 	build_inv_all(&cmd);
1142 
1143 	iommu_queue_command(iommu, &cmd);
1144 	iommu_completion_wait(iommu);
1145 }
1146 
1147 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1148 {
1149 	struct iommu_cmd cmd;
1150 
1151 	build_inv_irt(&cmd, devid);
1152 
1153 	iommu_queue_command(iommu, &cmd);
1154 }
1155 
1156 static void amd_iommu_flush_irt_all(struct amd_iommu *iommu)
1157 {
1158 	u32 devid;
1159 
1160 	for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1161 		iommu_flush_irt(iommu, devid);
1162 
1163 	iommu_completion_wait(iommu);
1164 }
1165 
1166 void iommu_flush_all_caches(struct amd_iommu *iommu)
1167 {
1168 	if (iommu_feature(iommu, FEATURE_IA)) {
1169 		amd_iommu_flush_all(iommu);
1170 	} else {
1171 		amd_iommu_flush_dte_all(iommu);
1172 		amd_iommu_flush_irt_all(iommu);
1173 		amd_iommu_flush_tlb_all(iommu);
1174 	}
1175 }
1176 
1177 /*
1178  * Command send function for flushing on-device TLB
1179  */
1180 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1181 			      u64 address, size_t size)
1182 {
1183 	struct amd_iommu *iommu;
1184 	struct iommu_cmd cmd;
1185 	int qdep;
1186 
1187 	qdep     = dev_data->ats.qdep;
1188 	iommu    = amd_iommu_rlookup_table[dev_data->devid];
1189 
1190 	build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1191 
1192 	return iommu_queue_command(iommu, &cmd);
1193 }
1194 
1195 static int device_flush_dte_alias(struct pci_dev *pdev, u16 alias, void *data)
1196 {
1197 	struct amd_iommu *iommu = data;
1198 
1199 	return iommu_flush_dte(iommu, alias);
1200 }
1201 
1202 /*
1203  * Command send function for invalidating a device table entry
1204  */
1205 static int device_flush_dte(struct iommu_dev_data *dev_data)
1206 {
1207 	struct amd_iommu *iommu;
1208 	u16 alias;
1209 	int ret;
1210 
1211 	iommu = amd_iommu_rlookup_table[dev_data->devid];
1212 
1213 	if (dev_data->pdev)
1214 		ret = pci_for_each_dma_alias(dev_data->pdev,
1215 					     device_flush_dte_alias, iommu);
1216 	else
1217 		ret = iommu_flush_dte(iommu, dev_data->devid);
1218 	if (ret)
1219 		return ret;
1220 
1221 	alias = amd_iommu_alias_table[dev_data->devid];
1222 	if (alias != dev_data->devid) {
1223 		ret = iommu_flush_dte(iommu, alias);
1224 		if (ret)
1225 			return ret;
1226 	}
1227 
1228 	if (dev_data->ats.enabled)
1229 		ret = device_flush_iotlb(dev_data, 0, ~0UL);
1230 
1231 	return ret;
1232 }
1233 
1234 /*
1235  * TLB invalidation function which is called from the mapping functions.
1236  * It invalidates a single PTE if the range to flush is within a single
1237  * page. Otherwise it flushes the whole TLB of the IOMMU.
1238  */
1239 static void __domain_flush_pages(struct protection_domain *domain,
1240 				 u64 address, size_t size, int pde)
1241 {
1242 	struct iommu_dev_data *dev_data;
1243 	struct iommu_cmd cmd;
1244 	int ret = 0, i;
1245 
1246 	build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1247 
1248 	for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1249 		if (!domain->dev_iommu[i])
1250 			continue;
1251 
1252 		/*
1253 		 * Devices of this domain are behind this IOMMU
1254 		 * We need a TLB flush
1255 		 */
1256 		ret |= iommu_queue_command(amd_iommus[i], &cmd);
1257 	}
1258 
1259 	list_for_each_entry(dev_data, &domain->dev_list, list) {
1260 
1261 		if (!dev_data->ats.enabled)
1262 			continue;
1263 
1264 		ret |= device_flush_iotlb(dev_data, address, size);
1265 	}
1266 
1267 	WARN_ON(ret);
1268 }
1269 
1270 static void domain_flush_pages(struct protection_domain *domain,
1271 			       u64 address, size_t size, int pde)
1272 {
1273 	if (likely(!amd_iommu_np_cache)) {
1274 		__domain_flush_pages(domain, address, size, pde);
1275 		return;
1276 	}
1277 
1278 	/*
1279 	 * When NpCache is on, we infer that we run in a VM and use a vIOMMU.
1280 	 * In such setups it is best to avoid flushes of ranges which are not
1281 	 * naturally aligned, since it would lead to flushes of unmodified
1282 	 * PTEs. Such flushes would require the hypervisor to do more work than
1283 	 * necessary. Therefore, perform repeated flushes of aligned ranges
1284 	 * until you cover the range. Each iteration flushes the smaller
1285 	 * between the natural alignment of the address that we flush and the
1286 	 * greatest naturally aligned region that fits in the range.
1287 	 */
1288 	while (size != 0) {
1289 		int addr_alignment = __ffs(address);
1290 		int size_alignment = __fls(size);
1291 		int min_alignment;
1292 		size_t flush_size;
1293 
1294 		/*
1295 		 * size is always non-zero, but address might be zero, causing
1296 		 * addr_alignment to be negative. As the casting of the
1297 		 * argument in __ffs(address) to long might trim the high bits
1298 		 * of the address on x86-32, cast to long when doing the check.
1299 		 */
1300 		if (likely((unsigned long)address != 0))
1301 			min_alignment = min(addr_alignment, size_alignment);
1302 		else
1303 			min_alignment = size_alignment;
1304 
1305 		flush_size = 1ul << min_alignment;
1306 
1307 		__domain_flush_pages(domain, address, flush_size, pde);
1308 		address += flush_size;
1309 		size -= flush_size;
1310 	}
1311 }
1312 
1313 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1314 void amd_iommu_domain_flush_tlb_pde(struct protection_domain *domain)
1315 {
1316 	domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1317 }
1318 
1319 void amd_iommu_domain_flush_complete(struct protection_domain *domain)
1320 {
1321 	int i;
1322 
1323 	for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1324 		if (domain && !domain->dev_iommu[i])
1325 			continue;
1326 
1327 		/*
1328 		 * Devices of this domain are behind this IOMMU
1329 		 * We need to wait for completion of all commands.
1330 		 */
1331 		iommu_completion_wait(amd_iommus[i]);
1332 	}
1333 }
1334 
1335 /* Flush the not present cache if it exists */
1336 static void domain_flush_np_cache(struct protection_domain *domain,
1337 		dma_addr_t iova, size_t size)
1338 {
1339 	if (unlikely(amd_iommu_np_cache)) {
1340 		unsigned long flags;
1341 
1342 		spin_lock_irqsave(&domain->lock, flags);
1343 		domain_flush_pages(domain, iova, size, 1);
1344 		amd_iommu_domain_flush_complete(domain);
1345 		spin_unlock_irqrestore(&domain->lock, flags);
1346 	}
1347 }
1348 
1349 
1350 /*
1351  * This function flushes the DTEs for all devices in domain
1352  */
1353 static void domain_flush_devices(struct protection_domain *domain)
1354 {
1355 	struct iommu_dev_data *dev_data;
1356 
1357 	list_for_each_entry(dev_data, &domain->dev_list, list)
1358 		device_flush_dte(dev_data);
1359 }
1360 
1361 /****************************************************************************
1362  *
1363  * The next functions belong to the domain allocation. A domain is
1364  * allocated for every IOMMU as the default domain. If device isolation
1365  * is enabled, every device get its own domain. The most important thing
1366  * about domains is the page table mapping the DMA address space they
1367  * contain.
1368  *
1369  ****************************************************************************/
1370 
1371 static u16 domain_id_alloc(void)
1372 {
1373 	int id;
1374 
1375 	spin_lock(&pd_bitmap_lock);
1376 	id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1377 	BUG_ON(id == 0);
1378 	if (id > 0 && id < MAX_DOMAIN_ID)
1379 		__set_bit(id, amd_iommu_pd_alloc_bitmap);
1380 	else
1381 		id = 0;
1382 	spin_unlock(&pd_bitmap_lock);
1383 
1384 	return id;
1385 }
1386 
1387 static void domain_id_free(int id)
1388 {
1389 	spin_lock(&pd_bitmap_lock);
1390 	if (id > 0 && id < MAX_DOMAIN_ID)
1391 		__clear_bit(id, amd_iommu_pd_alloc_bitmap);
1392 	spin_unlock(&pd_bitmap_lock);
1393 }
1394 
1395 static void free_gcr3_tbl_level1(u64 *tbl)
1396 {
1397 	u64 *ptr;
1398 	int i;
1399 
1400 	for (i = 0; i < 512; ++i) {
1401 		if (!(tbl[i] & GCR3_VALID))
1402 			continue;
1403 
1404 		ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1405 
1406 		free_page((unsigned long)ptr);
1407 	}
1408 }
1409 
1410 static void free_gcr3_tbl_level2(u64 *tbl)
1411 {
1412 	u64 *ptr;
1413 	int i;
1414 
1415 	for (i = 0; i < 512; ++i) {
1416 		if (!(tbl[i] & GCR3_VALID))
1417 			continue;
1418 
1419 		ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1420 
1421 		free_gcr3_tbl_level1(ptr);
1422 	}
1423 }
1424 
1425 static void free_gcr3_table(struct protection_domain *domain)
1426 {
1427 	if (domain->glx == 2)
1428 		free_gcr3_tbl_level2(domain->gcr3_tbl);
1429 	else if (domain->glx == 1)
1430 		free_gcr3_tbl_level1(domain->gcr3_tbl);
1431 	else
1432 		BUG_ON(domain->glx != 0);
1433 
1434 	free_page((unsigned long)domain->gcr3_tbl);
1435 }
1436 
1437 static void set_dte_entry(u16 devid, struct protection_domain *domain,
1438 			  bool ats, bool ppr)
1439 {
1440 	u64 pte_root = 0;
1441 	u64 flags = 0;
1442 	u32 old_domid;
1443 
1444 	if (domain->iop.mode != PAGE_MODE_NONE)
1445 		pte_root = iommu_virt_to_phys(domain->iop.root);
1446 
1447 	pte_root |= (domain->iop.mode & DEV_ENTRY_MODE_MASK)
1448 		    << DEV_ENTRY_MODE_SHIFT;
1449 	pte_root |= DTE_FLAG_IR | DTE_FLAG_IW | DTE_FLAG_V | DTE_FLAG_TV;
1450 
1451 	flags = amd_iommu_dev_table[devid].data[1];
1452 
1453 	if (ats)
1454 		flags |= DTE_FLAG_IOTLB;
1455 
1456 	if (ppr) {
1457 		struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
1458 
1459 		if (iommu_feature(iommu, FEATURE_EPHSUP))
1460 			pte_root |= 1ULL << DEV_ENTRY_PPR;
1461 	}
1462 
1463 	if (domain->flags & PD_IOMMUV2_MASK) {
1464 		u64 gcr3 = iommu_virt_to_phys(domain->gcr3_tbl);
1465 		u64 glx  = domain->glx;
1466 		u64 tmp;
1467 
1468 		pte_root |= DTE_FLAG_GV;
1469 		pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1470 
1471 		/* First mask out possible old values for GCR3 table */
1472 		tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1473 		flags    &= ~tmp;
1474 
1475 		tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1476 		flags    &= ~tmp;
1477 
1478 		/* Encode GCR3 table into DTE */
1479 		tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1480 		pte_root |= tmp;
1481 
1482 		tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1483 		flags    |= tmp;
1484 
1485 		tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1486 		flags    |= tmp;
1487 	}
1488 
1489 	flags &= ~DEV_DOMID_MASK;
1490 	flags |= domain->id;
1491 
1492 	old_domid = amd_iommu_dev_table[devid].data[1] & DEV_DOMID_MASK;
1493 	amd_iommu_dev_table[devid].data[1]  = flags;
1494 	amd_iommu_dev_table[devid].data[0]  = pte_root;
1495 
1496 	/*
1497 	 * A kdump kernel might be replacing a domain ID that was copied from
1498 	 * the previous kernel--if so, it needs to flush the translation cache
1499 	 * entries for the old domain ID that is being overwritten
1500 	 */
1501 	if (old_domid) {
1502 		struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
1503 
1504 		amd_iommu_flush_tlb_domid(iommu, old_domid);
1505 	}
1506 }
1507 
1508 static void clear_dte_entry(u16 devid)
1509 {
1510 	/* remove entry from the device table seen by the hardware */
1511 	amd_iommu_dev_table[devid].data[0]  = DTE_FLAG_V | DTE_FLAG_TV;
1512 	amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
1513 
1514 	amd_iommu_apply_erratum_63(devid);
1515 }
1516 
1517 static void do_attach(struct iommu_dev_data *dev_data,
1518 		      struct protection_domain *domain)
1519 {
1520 	struct amd_iommu *iommu;
1521 	bool ats;
1522 
1523 	iommu = amd_iommu_rlookup_table[dev_data->devid];
1524 	ats   = dev_data->ats.enabled;
1525 
1526 	/* Update data structures */
1527 	dev_data->domain = domain;
1528 	list_add(&dev_data->list, &domain->dev_list);
1529 
1530 	/* Do reference counting */
1531 	domain->dev_iommu[iommu->index] += 1;
1532 	domain->dev_cnt                 += 1;
1533 
1534 	/* Update device table */
1535 	set_dte_entry(dev_data->devid, domain,
1536 		      ats, dev_data->iommu_v2);
1537 	clone_aliases(dev_data->pdev);
1538 
1539 	device_flush_dte(dev_data);
1540 }
1541 
1542 static void do_detach(struct iommu_dev_data *dev_data)
1543 {
1544 	struct protection_domain *domain = dev_data->domain;
1545 	struct amd_iommu *iommu;
1546 
1547 	iommu = amd_iommu_rlookup_table[dev_data->devid];
1548 
1549 	/* Update data structures */
1550 	dev_data->domain = NULL;
1551 	list_del(&dev_data->list);
1552 	clear_dte_entry(dev_data->devid);
1553 	clone_aliases(dev_data->pdev);
1554 
1555 	/* Flush the DTE entry */
1556 	device_flush_dte(dev_data);
1557 
1558 	/* Flush IOTLB */
1559 	amd_iommu_domain_flush_tlb_pde(domain);
1560 
1561 	/* Wait for the flushes to finish */
1562 	amd_iommu_domain_flush_complete(domain);
1563 
1564 	/* decrease reference counters - needs to happen after the flushes */
1565 	domain->dev_iommu[iommu->index] -= 1;
1566 	domain->dev_cnt                 -= 1;
1567 }
1568 
1569 static void pdev_iommuv2_disable(struct pci_dev *pdev)
1570 {
1571 	pci_disable_ats(pdev);
1572 	pci_disable_pri(pdev);
1573 	pci_disable_pasid(pdev);
1574 }
1575 
1576 static int pdev_iommuv2_enable(struct pci_dev *pdev)
1577 {
1578 	int ret;
1579 
1580 	/* Only allow access to user-accessible pages */
1581 	ret = pci_enable_pasid(pdev, 0);
1582 	if (ret)
1583 		goto out_err;
1584 
1585 	/* First reset the PRI state of the device */
1586 	ret = pci_reset_pri(pdev);
1587 	if (ret)
1588 		goto out_err;
1589 
1590 	/* Enable PRI */
1591 	/* FIXME: Hardcode number of outstanding requests for now */
1592 	ret = pci_enable_pri(pdev, 32);
1593 	if (ret)
1594 		goto out_err;
1595 
1596 	ret = pci_enable_ats(pdev, PAGE_SHIFT);
1597 	if (ret)
1598 		goto out_err;
1599 
1600 	return 0;
1601 
1602 out_err:
1603 	pci_disable_pri(pdev);
1604 	pci_disable_pasid(pdev);
1605 
1606 	return ret;
1607 }
1608 
1609 /*
1610  * If a device is not yet associated with a domain, this function makes the
1611  * device visible in the domain
1612  */
1613 static int attach_device(struct device *dev,
1614 			 struct protection_domain *domain)
1615 {
1616 	struct iommu_dev_data *dev_data;
1617 	struct pci_dev *pdev;
1618 	unsigned long flags;
1619 	int ret;
1620 
1621 	spin_lock_irqsave(&domain->lock, flags);
1622 
1623 	dev_data = dev_iommu_priv_get(dev);
1624 
1625 	spin_lock(&dev_data->lock);
1626 
1627 	ret = -EBUSY;
1628 	if (dev_data->domain != NULL)
1629 		goto out;
1630 
1631 	if (!dev_is_pci(dev))
1632 		goto skip_ats_check;
1633 
1634 	pdev = to_pci_dev(dev);
1635 	if (domain->flags & PD_IOMMUV2_MASK) {
1636 		struct iommu_domain *def_domain = iommu_get_dma_domain(dev);
1637 
1638 		ret = -EINVAL;
1639 		if (def_domain->type != IOMMU_DOMAIN_IDENTITY)
1640 			goto out;
1641 
1642 		if (dev_data->iommu_v2) {
1643 			if (pdev_iommuv2_enable(pdev) != 0)
1644 				goto out;
1645 
1646 			dev_data->ats.enabled = true;
1647 			dev_data->ats.qdep    = pci_ats_queue_depth(pdev);
1648 			dev_data->pri_tlp     = pci_prg_resp_pasid_required(pdev);
1649 		}
1650 	} else if (amd_iommu_iotlb_sup &&
1651 		   pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
1652 		dev_data->ats.enabled = true;
1653 		dev_data->ats.qdep    = pci_ats_queue_depth(pdev);
1654 	}
1655 
1656 skip_ats_check:
1657 	ret = 0;
1658 
1659 	do_attach(dev_data, domain);
1660 
1661 	/*
1662 	 * We might boot into a crash-kernel here. The crashed kernel
1663 	 * left the caches in the IOMMU dirty. So we have to flush
1664 	 * here to evict all dirty stuff.
1665 	 */
1666 	amd_iommu_domain_flush_tlb_pde(domain);
1667 
1668 	amd_iommu_domain_flush_complete(domain);
1669 
1670 out:
1671 	spin_unlock(&dev_data->lock);
1672 
1673 	spin_unlock_irqrestore(&domain->lock, flags);
1674 
1675 	return ret;
1676 }
1677 
1678 /*
1679  * Removes a device from a protection domain (with devtable_lock held)
1680  */
1681 static void detach_device(struct device *dev)
1682 {
1683 	struct protection_domain *domain;
1684 	struct iommu_dev_data *dev_data;
1685 	unsigned long flags;
1686 
1687 	dev_data = dev_iommu_priv_get(dev);
1688 	domain   = dev_data->domain;
1689 
1690 	spin_lock_irqsave(&domain->lock, flags);
1691 
1692 	spin_lock(&dev_data->lock);
1693 
1694 	/*
1695 	 * First check if the device is still attached. It might already
1696 	 * be detached from its domain because the generic
1697 	 * iommu_detach_group code detached it and we try again here in
1698 	 * our alias handling.
1699 	 */
1700 	if (WARN_ON(!dev_data->domain))
1701 		goto out;
1702 
1703 	do_detach(dev_data);
1704 
1705 	if (!dev_is_pci(dev))
1706 		goto out;
1707 
1708 	if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
1709 		pdev_iommuv2_disable(to_pci_dev(dev));
1710 	else if (dev_data->ats.enabled)
1711 		pci_disable_ats(to_pci_dev(dev));
1712 
1713 	dev_data->ats.enabled = false;
1714 
1715 out:
1716 	spin_unlock(&dev_data->lock);
1717 
1718 	spin_unlock_irqrestore(&domain->lock, flags);
1719 }
1720 
1721 static struct iommu_device *amd_iommu_probe_device(struct device *dev)
1722 {
1723 	struct iommu_device *iommu_dev;
1724 	struct amd_iommu *iommu;
1725 	int ret, devid;
1726 
1727 	if (!check_device(dev))
1728 		return ERR_PTR(-ENODEV);
1729 
1730 	devid = get_device_id(dev);
1731 	iommu = amd_iommu_rlookup_table[devid];
1732 
1733 	if (dev_iommu_priv_get(dev))
1734 		return &iommu->iommu;
1735 
1736 	ret = iommu_init_device(dev);
1737 	if (ret) {
1738 		if (ret != -ENOTSUPP)
1739 			dev_err(dev, "Failed to initialize - trying to proceed anyway\n");
1740 		iommu_dev = ERR_PTR(ret);
1741 		iommu_ignore_device(dev);
1742 	} else {
1743 		amd_iommu_set_pci_msi_domain(dev, iommu);
1744 		iommu_dev = &iommu->iommu;
1745 	}
1746 
1747 	iommu_completion_wait(iommu);
1748 
1749 	return iommu_dev;
1750 }
1751 
1752 static void amd_iommu_probe_finalize(struct device *dev)
1753 {
1754 	/* Domains are initialized for this device - have a look what we ended up with */
1755 	set_dma_ops(dev, NULL);
1756 	iommu_setup_dma_ops(dev, 0, U64_MAX);
1757 }
1758 
1759 static void amd_iommu_release_device(struct device *dev)
1760 {
1761 	int devid = get_device_id(dev);
1762 	struct amd_iommu *iommu;
1763 
1764 	if (!check_device(dev))
1765 		return;
1766 
1767 	iommu = amd_iommu_rlookup_table[devid];
1768 
1769 	amd_iommu_uninit_device(dev);
1770 	iommu_completion_wait(iommu);
1771 }
1772 
1773 static struct iommu_group *amd_iommu_device_group(struct device *dev)
1774 {
1775 	if (dev_is_pci(dev))
1776 		return pci_device_group(dev);
1777 
1778 	return acpihid_device_group(dev);
1779 }
1780 
1781 /*****************************************************************************
1782  *
1783  * The next functions belong to the dma_ops mapping/unmapping code.
1784  *
1785  *****************************************************************************/
1786 
1787 static void update_device_table(struct protection_domain *domain)
1788 {
1789 	struct iommu_dev_data *dev_data;
1790 
1791 	list_for_each_entry(dev_data, &domain->dev_list, list) {
1792 		set_dte_entry(dev_data->devid, domain,
1793 			      dev_data->ats.enabled, dev_data->iommu_v2);
1794 		clone_aliases(dev_data->pdev);
1795 	}
1796 }
1797 
1798 void amd_iommu_update_and_flush_device_table(struct protection_domain *domain)
1799 {
1800 	update_device_table(domain);
1801 	domain_flush_devices(domain);
1802 }
1803 
1804 void amd_iommu_domain_update(struct protection_domain *domain)
1805 {
1806 	/* Update device table */
1807 	amd_iommu_update_and_flush_device_table(domain);
1808 
1809 	/* Flush domain TLB(s) and wait for completion */
1810 	amd_iommu_domain_flush_tlb_pde(domain);
1811 	amd_iommu_domain_flush_complete(domain);
1812 }
1813 
1814 static void __init amd_iommu_init_dma_ops(void)
1815 {
1816 	swiotlb = (iommu_default_passthrough() || sme_me_mask) ? 1 : 0;
1817 }
1818 
1819 int __init amd_iommu_init_api(void)
1820 {
1821 	int err;
1822 
1823 	amd_iommu_init_dma_ops();
1824 
1825 	err = bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
1826 	if (err)
1827 		return err;
1828 #ifdef CONFIG_ARM_AMBA
1829 	err = bus_set_iommu(&amba_bustype, &amd_iommu_ops);
1830 	if (err)
1831 		return err;
1832 #endif
1833 	err = bus_set_iommu(&platform_bus_type, &amd_iommu_ops);
1834 	if (err)
1835 		return err;
1836 
1837 	return 0;
1838 }
1839 
1840 /*****************************************************************************
1841  *
1842  * The following functions belong to the exported interface of AMD IOMMU
1843  *
1844  * This interface allows access to lower level functions of the IOMMU
1845  * like protection domain handling and assignement of devices to domains
1846  * which is not possible with the dma_ops interface.
1847  *
1848  *****************************************************************************/
1849 
1850 static void cleanup_domain(struct protection_domain *domain)
1851 {
1852 	struct iommu_dev_data *entry;
1853 	unsigned long flags;
1854 
1855 	spin_lock_irqsave(&domain->lock, flags);
1856 
1857 	while (!list_empty(&domain->dev_list)) {
1858 		entry = list_first_entry(&domain->dev_list,
1859 					 struct iommu_dev_data, list);
1860 		BUG_ON(!entry->domain);
1861 		do_detach(entry);
1862 	}
1863 
1864 	spin_unlock_irqrestore(&domain->lock, flags);
1865 }
1866 
1867 static void protection_domain_free(struct protection_domain *domain)
1868 {
1869 	if (!domain)
1870 		return;
1871 
1872 	if (domain->id)
1873 		domain_id_free(domain->id);
1874 
1875 	if (domain->iop.pgtbl_cfg.tlb)
1876 		free_io_pgtable_ops(&domain->iop.iop.ops);
1877 
1878 	kfree(domain);
1879 }
1880 
1881 static int protection_domain_init_v1(struct protection_domain *domain, int mode)
1882 {
1883 	u64 *pt_root = NULL;
1884 
1885 	BUG_ON(mode < PAGE_MODE_NONE || mode > PAGE_MODE_6_LEVEL);
1886 
1887 	spin_lock_init(&domain->lock);
1888 	domain->id = domain_id_alloc();
1889 	if (!domain->id)
1890 		return -ENOMEM;
1891 	INIT_LIST_HEAD(&domain->dev_list);
1892 
1893 	if (mode != PAGE_MODE_NONE) {
1894 		pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1895 		if (!pt_root)
1896 			return -ENOMEM;
1897 	}
1898 
1899 	amd_iommu_domain_set_pgtable(domain, pt_root, mode);
1900 
1901 	return 0;
1902 }
1903 
1904 static struct protection_domain *protection_domain_alloc(unsigned int type)
1905 {
1906 	struct io_pgtable_ops *pgtbl_ops;
1907 	struct protection_domain *domain;
1908 	int pgtable = amd_iommu_pgtable;
1909 	int mode = DEFAULT_PGTABLE_LEVEL;
1910 	int ret;
1911 
1912 	domain = kzalloc(sizeof(*domain), GFP_KERNEL);
1913 	if (!domain)
1914 		return NULL;
1915 
1916 	/*
1917 	 * Force IOMMU v1 page table when iommu=pt and
1918 	 * when allocating domain for pass-through devices.
1919 	 */
1920 	if (type == IOMMU_DOMAIN_IDENTITY) {
1921 		pgtable = AMD_IOMMU_V1;
1922 		mode = PAGE_MODE_NONE;
1923 	} else if (type == IOMMU_DOMAIN_UNMANAGED) {
1924 		pgtable = AMD_IOMMU_V1;
1925 	}
1926 
1927 	switch (pgtable) {
1928 	case AMD_IOMMU_V1:
1929 		ret = protection_domain_init_v1(domain, mode);
1930 		break;
1931 	default:
1932 		ret = -EINVAL;
1933 	}
1934 
1935 	if (ret)
1936 		goto out_err;
1937 
1938 	pgtbl_ops = alloc_io_pgtable_ops(pgtable, &domain->iop.pgtbl_cfg, domain);
1939 	if (!pgtbl_ops)
1940 		goto out_err;
1941 
1942 	return domain;
1943 out_err:
1944 	kfree(domain);
1945 	return NULL;
1946 }
1947 
1948 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
1949 {
1950 	struct protection_domain *domain;
1951 
1952 	domain = protection_domain_alloc(type);
1953 	if (!domain)
1954 		return NULL;
1955 
1956 	domain->domain.geometry.aperture_start = 0;
1957 	domain->domain.geometry.aperture_end   = ~0ULL;
1958 	domain->domain.geometry.force_aperture = true;
1959 
1960 	return &domain->domain;
1961 }
1962 
1963 static void amd_iommu_domain_free(struct iommu_domain *dom)
1964 {
1965 	struct protection_domain *domain;
1966 
1967 	domain = to_pdomain(dom);
1968 
1969 	if (domain->dev_cnt > 0)
1970 		cleanup_domain(domain);
1971 
1972 	BUG_ON(domain->dev_cnt != 0);
1973 
1974 	if (!dom)
1975 		return;
1976 
1977 	if (domain->flags & PD_IOMMUV2_MASK)
1978 		free_gcr3_table(domain);
1979 
1980 	protection_domain_free(domain);
1981 }
1982 
1983 static void amd_iommu_detach_device(struct iommu_domain *dom,
1984 				    struct device *dev)
1985 {
1986 	struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
1987 	int devid = get_device_id(dev);
1988 	struct amd_iommu *iommu;
1989 
1990 	if (!check_device(dev))
1991 		return;
1992 
1993 	if (dev_data->domain != NULL)
1994 		detach_device(dev);
1995 
1996 	iommu = amd_iommu_rlookup_table[devid];
1997 	if (!iommu)
1998 		return;
1999 
2000 #ifdef CONFIG_IRQ_REMAP
2001 	if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) &&
2002 	    (dom->type == IOMMU_DOMAIN_UNMANAGED))
2003 		dev_data->use_vapic = 0;
2004 #endif
2005 
2006 	iommu_completion_wait(iommu);
2007 }
2008 
2009 static int amd_iommu_attach_device(struct iommu_domain *dom,
2010 				   struct device *dev)
2011 {
2012 	struct protection_domain *domain = to_pdomain(dom);
2013 	struct iommu_dev_data *dev_data;
2014 	struct amd_iommu *iommu;
2015 	int ret;
2016 
2017 	if (!check_device(dev))
2018 		return -EINVAL;
2019 
2020 	dev_data = dev_iommu_priv_get(dev);
2021 	dev_data->defer_attach = false;
2022 
2023 	iommu = amd_iommu_rlookup_table[dev_data->devid];
2024 	if (!iommu)
2025 		return -EINVAL;
2026 
2027 	if (dev_data->domain)
2028 		detach_device(dev);
2029 
2030 	ret = attach_device(dev, domain);
2031 
2032 #ifdef CONFIG_IRQ_REMAP
2033 	if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
2034 		if (dom->type == IOMMU_DOMAIN_UNMANAGED)
2035 			dev_data->use_vapic = 1;
2036 		else
2037 			dev_data->use_vapic = 0;
2038 	}
2039 #endif
2040 
2041 	iommu_completion_wait(iommu);
2042 
2043 	return ret;
2044 }
2045 
2046 static void amd_iommu_iotlb_sync_map(struct iommu_domain *dom,
2047 				     unsigned long iova, size_t size)
2048 {
2049 	struct protection_domain *domain = to_pdomain(dom);
2050 	struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2051 
2052 	if (ops->map)
2053 		domain_flush_np_cache(domain, iova, size);
2054 }
2055 
2056 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
2057 			 phys_addr_t paddr, size_t page_size, int iommu_prot,
2058 			 gfp_t gfp)
2059 {
2060 	struct protection_domain *domain = to_pdomain(dom);
2061 	struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2062 	int prot = 0;
2063 	int ret = -EINVAL;
2064 
2065 	if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
2066 	    (domain->iop.mode == PAGE_MODE_NONE))
2067 		return -EINVAL;
2068 
2069 	if (iommu_prot & IOMMU_READ)
2070 		prot |= IOMMU_PROT_IR;
2071 	if (iommu_prot & IOMMU_WRITE)
2072 		prot |= IOMMU_PROT_IW;
2073 
2074 	if (ops->map)
2075 		ret = ops->map(ops, iova, paddr, page_size, prot, gfp);
2076 
2077 	return ret;
2078 }
2079 
2080 static void amd_iommu_iotlb_gather_add_page(struct iommu_domain *domain,
2081 					    struct iommu_iotlb_gather *gather,
2082 					    unsigned long iova, size_t size)
2083 {
2084 	/*
2085 	 * AMD's IOMMU can flush as many pages as necessary in a single flush.
2086 	 * Unless we run in a virtual machine, which can be inferred according
2087 	 * to whether "non-present cache" is on, it is probably best to prefer
2088 	 * (potentially) too extensive TLB flushing (i.e., more misses) over
2089 	 * mutliple TLB flushes (i.e., more flushes). For virtual machines the
2090 	 * hypervisor needs to synchronize the host IOMMU PTEs with those of
2091 	 * the guest, and the trade-off is different: unnecessary TLB flushes
2092 	 * should be avoided.
2093 	 */
2094 	if (amd_iommu_np_cache &&
2095 	    iommu_iotlb_gather_is_disjoint(gather, iova, size))
2096 		iommu_iotlb_sync(domain, gather);
2097 
2098 	iommu_iotlb_gather_add_range(gather, iova, size);
2099 }
2100 
2101 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
2102 			      size_t page_size,
2103 			      struct iommu_iotlb_gather *gather)
2104 {
2105 	struct protection_domain *domain = to_pdomain(dom);
2106 	struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2107 	size_t r;
2108 
2109 	if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
2110 	    (domain->iop.mode == PAGE_MODE_NONE))
2111 		return 0;
2112 
2113 	r = (ops->unmap) ? ops->unmap(ops, iova, page_size, gather) : 0;
2114 
2115 	amd_iommu_iotlb_gather_add_page(dom, gather, iova, page_size);
2116 
2117 	return r;
2118 }
2119 
2120 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
2121 					  dma_addr_t iova)
2122 {
2123 	struct protection_domain *domain = to_pdomain(dom);
2124 	struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2125 
2126 	return ops->iova_to_phys(ops, iova);
2127 }
2128 
2129 static bool amd_iommu_capable(enum iommu_cap cap)
2130 {
2131 	switch (cap) {
2132 	case IOMMU_CAP_CACHE_COHERENCY:
2133 		return true;
2134 	case IOMMU_CAP_INTR_REMAP:
2135 		return (irq_remapping_enabled == 1);
2136 	case IOMMU_CAP_NOEXEC:
2137 		return false;
2138 	default:
2139 		break;
2140 	}
2141 
2142 	return false;
2143 }
2144 
2145 static void amd_iommu_get_resv_regions(struct device *dev,
2146 				       struct list_head *head)
2147 {
2148 	struct iommu_resv_region *region;
2149 	struct unity_map_entry *entry;
2150 	int devid;
2151 
2152 	devid = get_device_id(dev);
2153 	if (devid < 0)
2154 		return;
2155 
2156 	list_for_each_entry(entry, &amd_iommu_unity_map, list) {
2157 		int type, prot = 0;
2158 		size_t length;
2159 
2160 		if (devid < entry->devid_start || devid > entry->devid_end)
2161 			continue;
2162 
2163 		type   = IOMMU_RESV_DIRECT;
2164 		length = entry->address_end - entry->address_start;
2165 		if (entry->prot & IOMMU_PROT_IR)
2166 			prot |= IOMMU_READ;
2167 		if (entry->prot & IOMMU_PROT_IW)
2168 			prot |= IOMMU_WRITE;
2169 		if (entry->prot & IOMMU_UNITY_MAP_FLAG_EXCL_RANGE)
2170 			/* Exclusion range */
2171 			type = IOMMU_RESV_RESERVED;
2172 
2173 		region = iommu_alloc_resv_region(entry->address_start,
2174 						 length, prot, type);
2175 		if (!region) {
2176 			dev_err(dev, "Out of memory allocating dm-regions\n");
2177 			return;
2178 		}
2179 		list_add_tail(&region->list, head);
2180 	}
2181 
2182 	region = iommu_alloc_resv_region(MSI_RANGE_START,
2183 					 MSI_RANGE_END - MSI_RANGE_START + 1,
2184 					 0, IOMMU_RESV_MSI);
2185 	if (!region)
2186 		return;
2187 	list_add_tail(&region->list, head);
2188 
2189 	region = iommu_alloc_resv_region(HT_RANGE_START,
2190 					 HT_RANGE_END - HT_RANGE_START + 1,
2191 					 0, IOMMU_RESV_RESERVED);
2192 	if (!region)
2193 		return;
2194 	list_add_tail(&region->list, head);
2195 }
2196 
2197 bool amd_iommu_is_attach_deferred(struct iommu_domain *domain,
2198 				  struct device *dev)
2199 {
2200 	struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2201 
2202 	return dev_data->defer_attach;
2203 }
2204 EXPORT_SYMBOL_GPL(amd_iommu_is_attach_deferred);
2205 
2206 static void amd_iommu_flush_iotlb_all(struct iommu_domain *domain)
2207 {
2208 	struct protection_domain *dom = to_pdomain(domain);
2209 	unsigned long flags;
2210 
2211 	spin_lock_irqsave(&dom->lock, flags);
2212 	amd_iommu_domain_flush_tlb_pde(dom);
2213 	amd_iommu_domain_flush_complete(dom);
2214 	spin_unlock_irqrestore(&dom->lock, flags);
2215 }
2216 
2217 static void amd_iommu_iotlb_sync(struct iommu_domain *domain,
2218 				 struct iommu_iotlb_gather *gather)
2219 {
2220 	struct protection_domain *dom = to_pdomain(domain);
2221 	unsigned long flags;
2222 
2223 	spin_lock_irqsave(&dom->lock, flags);
2224 	domain_flush_pages(dom, gather->start, gather->end - gather->start, 1);
2225 	amd_iommu_domain_flush_complete(dom);
2226 	spin_unlock_irqrestore(&dom->lock, flags);
2227 }
2228 
2229 static int amd_iommu_def_domain_type(struct device *dev)
2230 {
2231 	struct iommu_dev_data *dev_data;
2232 
2233 	dev_data = dev_iommu_priv_get(dev);
2234 	if (!dev_data)
2235 		return 0;
2236 
2237 	/*
2238 	 * Do not identity map IOMMUv2 capable devices when memory encryption is
2239 	 * active, because some of those devices (AMD GPUs) don't have the
2240 	 * encryption bit in their DMA-mask and require remapping.
2241 	 */
2242 	if (!cc_platform_has(CC_ATTR_MEM_ENCRYPT) && dev_data->iommu_v2)
2243 		return IOMMU_DOMAIN_IDENTITY;
2244 
2245 	return 0;
2246 }
2247 
2248 const struct iommu_ops amd_iommu_ops = {
2249 	.capable = amd_iommu_capable,
2250 	.domain_alloc = amd_iommu_domain_alloc,
2251 	.domain_free  = amd_iommu_domain_free,
2252 	.attach_dev = amd_iommu_attach_device,
2253 	.detach_dev = amd_iommu_detach_device,
2254 	.map = amd_iommu_map,
2255 	.iotlb_sync_map	= amd_iommu_iotlb_sync_map,
2256 	.unmap = amd_iommu_unmap,
2257 	.iova_to_phys = amd_iommu_iova_to_phys,
2258 	.probe_device = amd_iommu_probe_device,
2259 	.release_device = amd_iommu_release_device,
2260 	.probe_finalize = amd_iommu_probe_finalize,
2261 	.device_group = amd_iommu_device_group,
2262 	.get_resv_regions = amd_iommu_get_resv_regions,
2263 	.put_resv_regions = generic_iommu_put_resv_regions,
2264 	.is_attach_deferred = amd_iommu_is_attach_deferred,
2265 	.pgsize_bitmap	= AMD_IOMMU_PGSIZES,
2266 	.flush_iotlb_all = amd_iommu_flush_iotlb_all,
2267 	.iotlb_sync = amd_iommu_iotlb_sync,
2268 	.def_domain_type = amd_iommu_def_domain_type,
2269 };
2270 
2271 /*****************************************************************************
2272  *
2273  * The next functions do a basic initialization of IOMMU for pass through
2274  * mode
2275  *
2276  * In passthrough mode the IOMMU is initialized and enabled but not used for
2277  * DMA-API translation.
2278  *
2279  *****************************************************************************/
2280 
2281 /* IOMMUv2 specific functions */
2282 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
2283 {
2284 	return atomic_notifier_chain_register(&ppr_notifier, nb);
2285 }
2286 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
2287 
2288 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
2289 {
2290 	return atomic_notifier_chain_unregister(&ppr_notifier, nb);
2291 }
2292 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
2293 
2294 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
2295 {
2296 	struct protection_domain *domain = to_pdomain(dom);
2297 	unsigned long flags;
2298 
2299 	spin_lock_irqsave(&domain->lock, flags);
2300 
2301 	if (domain->iop.pgtbl_cfg.tlb)
2302 		free_io_pgtable_ops(&domain->iop.iop.ops);
2303 
2304 	spin_unlock_irqrestore(&domain->lock, flags);
2305 }
2306 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
2307 
2308 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
2309 {
2310 	struct protection_domain *domain = to_pdomain(dom);
2311 	unsigned long flags;
2312 	int levels, ret;
2313 
2314 	/* Number of GCR3 table levels required */
2315 	for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
2316 		levels += 1;
2317 
2318 	if (levels > amd_iommu_max_glx_val)
2319 		return -EINVAL;
2320 
2321 	spin_lock_irqsave(&domain->lock, flags);
2322 
2323 	/*
2324 	 * Save us all sanity checks whether devices already in the
2325 	 * domain support IOMMUv2. Just force that the domain has no
2326 	 * devices attached when it is switched into IOMMUv2 mode.
2327 	 */
2328 	ret = -EBUSY;
2329 	if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
2330 		goto out;
2331 
2332 	ret = -ENOMEM;
2333 	domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
2334 	if (domain->gcr3_tbl == NULL)
2335 		goto out;
2336 
2337 	domain->glx      = levels;
2338 	domain->flags   |= PD_IOMMUV2_MASK;
2339 
2340 	amd_iommu_domain_update(domain);
2341 
2342 	ret = 0;
2343 
2344 out:
2345 	spin_unlock_irqrestore(&domain->lock, flags);
2346 
2347 	return ret;
2348 }
2349 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
2350 
2351 static int __flush_pasid(struct protection_domain *domain, u32 pasid,
2352 			 u64 address, bool size)
2353 {
2354 	struct iommu_dev_data *dev_data;
2355 	struct iommu_cmd cmd;
2356 	int i, ret;
2357 
2358 	if (!(domain->flags & PD_IOMMUV2_MASK))
2359 		return -EINVAL;
2360 
2361 	build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
2362 
2363 	/*
2364 	 * IOMMU TLB needs to be flushed before Device TLB to
2365 	 * prevent device TLB refill from IOMMU TLB
2366 	 */
2367 	for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
2368 		if (domain->dev_iommu[i] == 0)
2369 			continue;
2370 
2371 		ret = iommu_queue_command(amd_iommus[i], &cmd);
2372 		if (ret != 0)
2373 			goto out;
2374 	}
2375 
2376 	/* Wait until IOMMU TLB flushes are complete */
2377 	amd_iommu_domain_flush_complete(domain);
2378 
2379 	/* Now flush device TLBs */
2380 	list_for_each_entry(dev_data, &domain->dev_list, list) {
2381 		struct amd_iommu *iommu;
2382 		int qdep;
2383 
2384 		/*
2385 		   There might be non-IOMMUv2 capable devices in an IOMMUv2
2386 		 * domain.
2387 		 */
2388 		if (!dev_data->ats.enabled)
2389 			continue;
2390 
2391 		qdep  = dev_data->ats.qdep;
2392 		iommu = amd_iommu_rlookup_table[dev_data->devid];
2393 
2394 		build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
2395 				      qdep, address, size);
2396 
2397 		ret = iommu_queue_command(iommu, &cmd);
2398 		if (ret != 0)
2399 			goto out;
2400 	}
2401 
2402 	/* Wait until all device TLBs are flushed */
2403 	amd_iommu_domain_flush_complete(domain);
2404 
2405 	ret = 0;
2406 
2407 out:
2408 
2409 	return ret;
2410 }
2411 
2412 static int __amd_iommu_flush_page(struct protection_domain *domain, u32 pasid,
2413 				  u64 address)
2414 {
2415 	return __flush_pasid(domain, pasid, address, false);
2416 }
2417 
2418 int amd_iommu_flush_page(struct iommu_domain *dom, u32 pasid,
2419 			 u64 address)
2420 {
2421 	struct protection_domain *domain = to_pdomain(dom);
2422 	unsigned long flags;
2423 	int ret;
2424 
2425 	spin_lock_irqsave(&domain->lock, flags);
2426 	ret = __amd_iommu_flush_page(domain, pasid, address);
2427 	spin_unlock_irqrestore(&domain->lock, flags);
2428 
2429 	return ret;
2430 }
2431 EXPORT_SYMBOL(amd_iommu_flush_page);
2432 
2433 static int __amd_iommu_flush_tlb(struct protection_domain *domain, u32 pasid)
2434 {
2435 	return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
2436 			     true);
2437 }
2438 
2439 int amd_iommu_flush_tlb(struct iommu_domain *dom, u32 pasid)
2440 {
2441 	struct protection_domain *domain = to_pdomain(dom);
2442 	unsigned long flags;
2443 	int ret;
2444 
2445 	spin_lock_irqsave(&domain->lock, flags);
2446 	ret = __amd_iommu_flush_tlb(domain, pasid);
2447 	spin_unlock_irqrestore(&domain->lock, flags);
2448 
2449 	return ret;
2450 }
2451 EXPORT_SYMBOL(amd_iommu_flush_tlb);
2452 
2453 static u64 *__get_gcr3_pte(u64 *root, int level, u32 pasid, bool alloc)
2454 {
2455 	int index;
2456 	u64 *pte;
2457 
2458 	while (true) {
2459 
2460 		index = (pasid >> (9 * level)) & 0x1ff;
2461 		pte   = &root[index];
2462 
2463 		if (level == 0)
2464 			break;
2465 
2466 		if (!(*pte & GCR3_VALID)) {
2467 			if (!alloc)
2468 				return NULL;
2469 
2470 			root = (void *)get_zeroed_page(GFP_ATOMIC);
2471 			if (root == NULL)
2472 				return NULL;
2473 
2474 			*pte = iommu_virt_to_phys(root) | GCR3_VALID;
2475 		}
2476 
2477 		root = iommu_phys_to_virt(*pte & PAGE_MASK);
2478 
2479 		level -= 1;
2480 	}
2481 
2482 	return pte;
2483 }
2484 
2485 static int __set_gcr3(struct protection_domain *domain, u32 pasid,
2486 		      unsigned long cr3)
2487 {
2488 	u64 *pte;
2489 
2490 	if (domain->iop.mode != PAGE_MODE_NONE)
2491 		return -EINVAL;
2492 
2493 	pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
2494 	if (pte == NULL)
2495 		return -ENOMEM;
2496 
2497 	*pte = (cr3 & PAGE_MASK) | GCR3_VALID;
2498 
2499 	return __amd_iommu_flush_tlb(domain, pasid);
2500 }
2501 
2502 static int __clear_gcr3(struct protection_domain *domain, u32 pasid)
2503 {
2504 	u64 *pte;
2505 
2506 	if (domain->iop.mode != PAGE_MODE_NONE)
2507 		return -EINVAL;
2508 
2509 	pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
2510 	if (pte == NULL)
2511 		return 0;
2512 
2513 	*pte = 0;
2514 
2515 	return __amd_iommu_flush_tlb(domain, pasid);
2516 }
2517 
2518 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, u32 pasid,
2519 			      unsigned long cr3)
2520 {
2521 	struct protection_domain *domain = to_pdomain(dom);
2522 	unsigned long flags;
2523 	int ret;
2524 
2525 	spin_lock_irqsave(&domain->lock, flags);
2526 	ret = __set_gcr3(domain, pasid, cr3);
2527 	spin_unlock_irqrestore(&domain->lock, flags);
2528 
2529 	return ret;
2530 }
2531 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
2532 
2533 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, u32 pasid)
2534 {
2535 	struct protection_domain *domain = to_pdomain(dom);
2536 	unsigned long flags;
2537 	int ret;
2538 
2539 	spin_lock_irqsave(&domain->lock, flags);
2540 	ret = __clear_gcr3(domain, pasid);
2541 	spin_unlock_irqrestore(&domain->lock, flags);
2542 
2543 	return ret;
2544 }
2545 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
2546 
2547 int amd_iommu_complete_ppr(struct pci_dev *pdev, u32 pasid,
2548 			   int status, int tag)
2549 {
2550 	struct iommu_dev_data *dev_data;
2551 	struct amd_iommu *iommu;
2552 	struct iommu_cmd cmd;
2553 
2554 	dev_data = dev_iommu_priv_get(&pdev->dev);
2555 	iommu    = amd_iommu_rlookup_table[dev_data->devid];
2556 
2557 	build_complete_ppr(&cmd, dev_data->devid, pasid, status,
2558 			   tag, dev_data->pri_tlp);
2559 
2560 	return iommu_queue_command(iommu, &cmd);
2561 }
2562 EXPORT_SYMBOL(amd_iommu_complete_ppr);
2563 
2564 int amd_iommu_device_info(struct pci_dev *pdev,
2565                           struct amd_iommu_device_info *info)
2566 {
2567 	int max_pasids;
2568 	int pos;
2569 
2570 	if (pdev == NULL || info == NULL)
2571 		return -EINVAL;
2572 
2573 	if (!amd_iommu_v2_supported())
2574 		return -EINVAL;
2575 
2576 	memset(info, 0, sizeof(*info));
2577 
2578 	if (pci_ats_supported(pdev))
2579 		info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
2580 
2581 	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2582 	if (pos)
2583 		info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
2584 
2585 	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
2586 	if (pos) {
2587 		int features;
2588 
2589 		max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
2590 		max_pasids = min(max_pasids, (1 << 20));
2591 
2592 		info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
2593 		info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
2594 
2595 		features = pci_pasid_features(pdev);
2596 		if (features & PCI_PASID_CAP_EXEC)
2597 			info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
2598 		if (features & PCI_PASID_CAP_PRIV)
2599 			info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
2600 	}
2601 
2602 	return 0;
2603 }
2604 EXPORT_SYMBOL(amd_iommu_device_info);
2605 
2606 #ifdef CONFIG_IRQ_REMAP
2607 
2608 /*****************************************************************************
2609  *
2610  * Interrupt Remapping Implementation
2611  *
2612  *****************************************************************************/
2613 
2614 static struct irq_chip amd_ir_chip;
2615 static DEFINE_SPINLOCK(iommu_table_lock);
2616 
2617 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
2618 {
2619 	u64 dte;
2620 
2621 	dte	= amd_iommu_dev_table[devid].data[2];
2622 	dte	&= ~DTE_IRQ_PHYS_ADDR_MASK;
2623 	dte	|= iommu_virt_to_phys(table->table);
2624 	dte	|= DTE_IRQ_REMAP_INTCTL;
2625 	dte	|= DTE_INTTABLEN;
2626 	dte	|= DTE_IRQ_REMAP_ENABLE;
2627 
2628 	amd_iommu_dev_table[devid].data[2] = dte;
2629 }
2630 
2631 static struct irq_remap_table *get_irq_table(u16 devid)
2632 {
2633 	struct irq_remap_table *table;
2634 
2635 	if (WARN_ONCE(!amd_iommu_rlookup_table[devid],
2636 		      "%s: no iommu for devid %x\n", __func__, devid))
2637 		return NULL;
2638 
2639 	table = irq_lookup_table[devid];
2640 	if (WARN_ONCE(!table, "%s: no table for devid %x\n", __func__, devid))
2641 		return NULL;
2642 
2643 	return table;
2644 }
2645 
2646 static struct irq_remap_table *__alloc_irq_table(void)
2647 {
2648 	struct irq_remap_table *table;
2649 
2650 	table = kzalloc(sizeof(*table), GFP_KERNEL);
2651 	if (!table)
2652 		return NULL;
2653 
2654 	table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_KERNEL);
2655 	if (!table->table) {
2656 		kfree(table);
2657 		return NULL;
2658 	}
2659 	raw_spin_lock_init(&table->lock);
2660 
2661 	if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
2662 		memset(table->table, 0,
2663 		       MAX_IRQS_PER_TABLE * sizeof(u32));
2664 	else
2665 		memset(table->table, 0,
2666 		       (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
2667 	return table;
2668 }
2669 
2670 static void set_remap_table_entry(struct amd_iommu *iommu, u16 devid,
2671 				  struct irq_remap_table *table)
2672 {
2673 	irq_lookup_table[devid] = table;
2674 	set_dte_irq_entry(devid, table);
2675 	iommu_flush_dte(iommu, devid);
2676 }
2677 
2678 static int set_remap_table_entry_alias(struct pci_dev *pdev, u16 alias,
2679 				       void *data)
2680 {
2681 	struct irq_remap_table *table = data;
2682 
2683 	irq_lookup_table[alias] = table;
2684 	set_dte_irq_entry(alias, table);
2685 
2686 	iommu_flush_dte(amd_iommu_rlookup_table[alias], alias);
2687 
2688 	return 0;
2689 }
2690 
2691 static struct irq_remap_table *alloc_irq_table(u16 devid, struct pci_dev *pdev)
2692 {
2693 	struct irq_remap_table *table = NULL;
2694 	struct irq_remap_table *new_table = NULL;
2695 	struct amd_iommu *iommu;
2696 	unsigned long flags;
2697 	u16 alias;
2698 
2699 	spin_lock_irqsave(&iommu_table_lock, flags);
2700 
2701 	iommu = amd_iommu_rlookup_table[devid];
2702 	if (!iommu)
2703 		goto out_unlock;
2704 
2705 	table = irq_lookup_table[devid];
2706 	if (table)
2707 		goto out_unlock;
2708 
2709 	alias = amd_iommu_alias_table[devid];
2710 	table = irq_lookup_table[alias];
2711 	if (table) {
2712 		set_remap_table_entry(iommu, devid, table);
2713 		goto out_wait;
2714 	}
2715 	spin_unlock_irqrestore(&iommu_table_lock, flags);
2716 
2717 	/* Nothing there yet, allocate new irq remapping table */
2718 	new_table = __alloc_irq_table();
2719 	if (!new_table)
2720 		return NULL;
2721 
2722 	spin_lock_irqsave(&iommu_table_lock, flags);
2723 
2724 	table = irq_lookup_table[devid];
2725 	if (table)
2726 		goto out_unlock;
2727 
2728 	table = irq_lookup_table[alias];
2729 	if (table) {
2730 		set_remap_table_entry(iommu, devid, table);
2731 		goto out_wait;
2732 	}
2733 
2734 	table = new_table;
2735 	new_table = NULL;
2736 
2737 	if (pdev)
2738 		pci_for_each_dma_alias(pdev, set_remap_table_entry_alias,
2739 				       table);
2740 	else
2741 		set_remap_table_entry(iommu, devid, table);
2742 
2743 	if (devid != alias)
2744 		set_remap_table_entry(iommu, alias, table);
2745 
2746 out_wait:
2747 	iommu_completion_wait(iommu);
2748 
2749 out_unlock:
2750 	spin_unlock_irqrestore(&iommu_table_lock, flags);
2751 
2752 	if (new_table) {
2753 		kmem_cache_free(amd_iommu_irq_cache, new_table->table);
2754 		kfree(new_table);
2755 	}
2756 	return table;
2757 }
2758 
2759 static int alloc_irq_index(u16 devid, int count, bool align,
2760 			   struct pci_dev *pdev)
2761 {
2762 	struct irq_remap_table *table;
2763 	int index, c, alignment = 1;
2764 	unsigned long flags;
2765 	struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
2766 
2767 	if (!iommu)
2768 		return -ENODEV;
2769 
2770 	table = alloc_irq_table(devid, pdev);
2771 	if (!table)
2772 		return -ENODEV;
2773 
2774 	if (align)
2775 		alignment = roundup_pow_of_two(count);
2776 
2777 	raw_spin_lock_irqsave(&table->lock, flags);
2778 
2779 	/* Scan table for free entries */
2780 	for (index = ALIGN(table->min_index, alignment), c = 0;
2781 	     index < MAX_IRQS_PER_TABLE;) {
2782 		if (!iommu->irte_ops->is_allocated(table, index)) {
2783 			c += 1;
2784 		} else {
2785 			c     = 0;
2786 			index = ALIGN(index + 1, alignment);
2787 			continue;
2788 		}
2789 
2790 		if (c == count)	{
2791 			for (; c != 0; --c)
2792 				iommu->irte_ops->set_allocated(table, index - c + 1);
2793 
2794 			index -= count - 1;
2795 			goto out;
2796 		}
2797 
2798 		index++;
2799 	}
2800 
2801 	index = -ENOSPC;
2802 
2803 out:
2804 	raw_spin_unlock_irqrestore(&table->lock, flags);
2805 
2806 	return index;
2807 }
2808 
2809 static int modify_irte_ga(u16 devid, int index, struct irte_ga *irte,
2810 			  struct amd_ir_data *data)
2811 {
2812 	bool ret;
2813 	struct irq_remap_table *table;
2814 	struct amd_iommu *iommu;
2815 	unsigned long flags;
2816 	struct irte_ga *entry;
2817 
2818 	iommu = amd_iommu_rlookup_table[devid];
2819 	if (iommu == NULL)
2820 		return -EINVAL;
2821 
2822 	table = get_irq_table(devid);
2823 	if (!table)
2824 		return -ENOMEM;
2825 
2826 	raw_spin_lock_irqsave(&table->lock, flags);
2827 
2828 	entry = (struct irte_ga *)table->table;
2829 	entry = &entry[index];
2830 
2831 	ret = cmpxchg_double(&entry->lo.val, &entry->hi.val,
2832 			     entry->lo.val, entry->hi.val,
2833 			     irte->lo.val, irte->hi.val);
2834 	/*
2835 	 * We use cmpxchg16 to atomically update the 128-bit IRTE,
2836 	 * and it cannot be updated by the hardware or other processors
2837 	 * behind us, so the return value of cmpxchg16 should be the
2838 	 * same as the old value.
2839 	 */
2840 	WARN_ON(!ret);
2841 
2842 	if (data)
2843 		data->ref = entry;
2844 
2845 	raw_spin_unlock_irqrestore(&table->lock, flags);
2846 
2847 	iommu_flush_irt(iommu, devid);
2848 	iommu_completion_wait(iommu);
2849 
2850 	return 0;
2851 }
2852 
2853 static int modify_irte(u16 devid, int index, union irte *irte)
2854 {
2855 	struct irq_remap_table *table;
2856 	struct amd_iommu *iommu;
2857 	unsigned long flags;
2858 
2859 	iommu = amd_iommu_rlookup_table[devid];
2860 	if (iommu == NULL)
2861 		return -EINVAL;
2862 
2863 	table = get_irq_table(devid);
2864 	if (!table)
2865 		return -ENOMEM;
2866 
2867 	raw_spin_lock_irqsave(&table->lock, flags);
2868 	table->table[index] = irte->val;
2869 	raw_spin_unlock_irqrestore(&table->lock, flags);
2870 
2871 	iommu_flush_irt(iommu, devid);
2872 	iommu_completion_wait(iommu);
2873 
2874 	return 0;
2875 }
2876 
2877 static void free_irte(u16 devid, int index)
2878 {
2879 	struct irq_remap_table *table;
2880 	struct amd_iommu *iommu;
2881 	unsigned long flags;
2882 
2883 	iommu = amd_iommu_rlookup_table[devid];
2884 	if (iommu == NULL)
2885 		return;
2886 
2887 	table = get_irq_table(devid);
2888 	if (!table)
2889 		return;
2890 
2891 	raw_spin_lock_irqsave(&table->lock, flags);
2892 	iommu->irte_ops->clear_allocated(table, index);
2893 	raw_spin_unlock_irqrestore(&table->lock, flags);
2894 
2895 	iommu_flush_irt(iommu, devid);
2896 	iommu_completion_wait(iommu);
2897 }
2898 
2899 static void irte_prepare(void *entry,
2900 			 u32 delivery_mode, bool dest_mode,
2901 			 u8 vector, u32 dest_apicid, int devid)
2902 {
2903 	union irte *irte = (union irte *) entry;
2904 
2905 	irte->val                = 0;
2906 	irte->fields.vector      = vector;
2907 	irte->fields.int_type    = delivery_mode;
2908 	irte->fields.destination = dest_apicid;
2909 	irte->fields.dm          = dest_mode;
2910 	irte->fields.valid       = 1;
2911 }
2912 
2913 static void irte_ga_prepare(void *entry,
2914 			    u32 delivery_mode, bool dest_mode,
2915 			    u8 vector, u32 dest_apicid, int devid)
2916 {
2917 	struct irte_ga *irte = (struct irte_ga *) entry;
2918 
2919 	irte->lo.val                      = 0;
2920 	irte->hi.val                      = 0;
2921 	irte->lo.fields_remap.int_type    = delivery_mode;
2922 	irte->lo.fields_remap.dm          = dest_mode;
2923 	irte->hi.fields.vector            = vector;
2924 	irte->lo.fields_remap.destination = APICID_TO_IRTE_DEST_LO(dest_apicid);
2925 	irte->hi.fields.destination       = APICID_TO_IRTE_DEST_HI(dest_apicid);
2926 	irte->lo.fields_remap.valid       = 1;
2927 }
2928 
2929 static void irte_activate(void *entry, u16 devid, u16 index)
2930 {
2931 	union irte *irte = (union irte *) entry;
2932 
2933 	irte->fields.valid = 1;
2934 	modify_irte(devid, index, irte);
2935 }
2936 
2937 static void irte_ga_activate(void *entry, u16 devid, u16 index)
2938 {
2939 	struct irte_ga *irte = (struct irte_ga *) entry;
2940 
2941 	irte->lo.fields_remap.valid = 1;
2942 	modify_irte_ga(devid, index, irte, NULL);
2943 }
2944 
2945 static void irte_deactivate(void *entry, u16 devid, u16 index)
2946 {
2947 	union irte *irte = (union irte *) entry;
2948 
2949 	irte->fields.valid = 0;
2950 	modify_irte(devid, index, irte);
2951 }
2952 
2953 static void irte_ga_deactivate(void *entry, u16 devid, u16 index)
2954 {
2955 	struct irte_ga *irte = (struct irte_ga *) entry;
2956 
2957 	irte->lo.fields_remap.valid = 0;
2958 	modify_irte_ga(devid, index, irte, NULL);
2959 }
2960 
2961 static void irte_set_affinity(void *entry, u16 devid, u16 index,
2962 			      u8 vector, u32 dest_apicid)
2963 {
2964 	union irte *irte = (union irte *) entry;
2965 
2966 	irte->fields.vector = vector;
2967 	irte->fields.destination = dest_apicid;
2968 	modify_irte(devid, index, irte);
2969 }
2970 
2971 static void irte_ga_set_affinity(void *entry, u16 devid, u16 index,
2972 				 u8 vector, u32 dest_apicid)
2973 {
2974 	struct irte_ga *irte = (struct irte_ga *) entry;
2975 
2976 	if (!irte->lo.fields_remap.guest_mode) {
2977 		irte->hi.fields.vector = vector;
2978 		irte->lo.fields_remap.destination =
2979 					APICID_TO_IRTE_DEST_LO(dest_apicid);
2980 		irte->hi.fields.destination =
2981 					APICID_TO_IRTE_DEST_HI(dest_apicid);
2982 		modify_irte_ga(devid, index, irte, NULL);
2983 	}
2984 }
2985 
2986 #define IRTE_ALLOCATED (~1U)
2987 static void irte_set_allocated(struct irq_remap_table *table, int index)
2988 {
2989 	table->table[index] = IRTE_ALLOCATED;
2990 }
2991 
2992 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
2993 {
2994 	struct irte_ga *ptr = (struct irte_ga *)table->table;
2995 	struct irte_ga *irte = &ptr[index];
2996 
2997 	memset(&irte->lo.val, 0, sizeof(u64));
2998 	memset(&irte->hi.val, 0, sizeof(u64));
2999 	irte->hi.fields.vector = 0xff;
3000 }
3001 
3002 static bool irte_is_allocated(struct irq_remap_table *table, int index)
3003 {
3004 	union irte *ptr = (union irte *)table->table;
3005 	union irte *irte = &ptr[index];
3006 
3007 	return irte->val != 0;
3008 }
3009 
3010 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
3011 {
3012 	struct irte_ga *ptr = (struct irte_ga *)table->table;
3013 	struct irte_ga *irte = &ptr[index];
3014 
3015 	return irte->hi.fields.vector != 0;
3016 }
3017 
3018 static void irte_clear_allocated(struct irq_remap_table *table, int index)
3019 {
3020 	table->table[index] = 0;
3021 }
3022 
3023 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
3024 {
3025 	struct irte_ga *ptr = (struct irte_ga *)table->table;
3026 	struct irte_ga *irte = &ptr[index];
3027 
3028 	memset(&irte->lo.val, 0, sizeof(u64));
3029 	memset(&irte->hi.val, 0, sizeof(u64));
3030 }
3031 
3032 static int get_devid(struct irq_alloc_info *info)
3033 {
3034 	switch (info->type) {
3035 	case X86_IRQ_ALLOC_TYPE_IOAPIC:
3036 		return get_ioapic_devid(info->devid);
3037 	case X86_IRQ_ALLOC_TYPE_HPET:
3038 		return get_hpet_devid(info->devid);
3039 	case X86_IRQ_ALLOC_TYPE_PCI_MSI:
3040 	case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
3041 		return get_device_id(msi_desc_to_dev(info->desc));
3042 	default:
3043 		WARN_ON_ONCE(1);
3044 		return -1;
3045 	}
3046 }
3047 
3048 struct irq_remap_ops amd_iommu_irq_ops = {
3049 	.prepare		= amd_iommu_prepare,
3050 	.enable			= amd_iommu_enable,
3051 	.disable		= amd_iommu_disable,
3052 	.reenable		= amd_iommu_reenable,
3053 	.enable_faulting	= amd_iommu_enable_faulting,
3054 };
3055 
3056 static void fill_msi_msg(struct msi_msg *msg, u32 index)
3057 {
3058 	msg->data = index;
3059 	msg->address_lo = 0;
3060 	msg->arch_addr_lo.base_address = X86_MSI_BASE_ADDRESS_LOW;
3061 	msg->address_hi = X86_MSI_BASE_ADDRESS_HIGH;
3062 }
3063 
3064 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3065 				       struct irq_cfg *irq_cfg,
3066 				       struct irq_alloc_info *info,
3067 				       int devid, int index, int sub_handle)
3068 {
3069 	struct irq_2_irte *irte_info = &data->irq_2_irte;
3070 	struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
3071 
3072 	if (!iommu)
3073 		return;
3074 
3075 	data->irq_2_irte.devid = devid;
3076 	data->irq_2_irte.index = index + sub_handle;
3077 	iommu->irte_ops->prepare(data->entry, apic->delivery_mode,
3078 				 apic->dest_mode_logical, irq_cfg->vector,
3079 				 irq_cfg->dest_apicid, devid);
3080 
3081 	switch (info->type) {
3082 	case X86_IRQ_ALLOC_TYPE_IOAPIC:
3083 	case X86_IRQ_ALLOC_TYPE_HPET:
3084 	case X86_IRQ_ALLOC_TYPE_PCI_MSI:
3085 	case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
3086 		fill_msi_msg(&data->msi_entry, irte_info->index);
3087 		break;
3088 
3089 	default:
3090 		BUG_ON(1);
3091 		break;
3092 	}
3093 }
3094 
3095 struct amd_irte_ops irte_32_ops = {
3096 	.prepare = irte_prepare,
3097 	.activate = irte_activate,
3098 	.deactivate = irte_deactivate,
3099 	.set_affinity = irte_set_affinity,
3100 	.set_allocated = irte_set_allocated,
3101 	.is_allocated = irte_is_allocated,
3102 	.clear_allocated = irte_clear_allocated,
3103 };
3104 
3105 struct amd_irte_ops irte_128_ops = {
3106 	.prepare = irte_ga_prepare,
3107 	.activate = irte_ga_activate,
3108 	.deactivate = irte_ga_deactivate,
3109 	.set_affinity = irte_ga_set_affinity,
3110 	.set_allocated = irte_ga_set_allocated,
3111 	.is_allocated = irte_ga_is_allocated,
3112 	.clear_allocated = irte_ga_clear_allocated,
3113 };
3114 
3115 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
3116 			       unsigned int nr_irqs, void *arg)
3117 {
3118 	struct irq_alloc_info *info = arg;
3119 	struct irq_data *irq_data;
3120 	struct amd_ir_data *data = NULL;
3121 	struct irq_cfg *cfg;
3122 	int i, ret, devid;
3123 	int index;
3124 
3125 	if (!info)
3126 		return -EINVAL;
3127 	if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_PCI_MSI &&
3128 	    info->type != X86_IRQ_ALLOC_TYPE_PCI_MSIX)
3129 		return -EINVAL;
3130 
3131 	/*
3132 	 * With IRQ remapping enabled, don't need contiguous CPU vectors
3133 	 * to support multiple MSI interrupts.
3134 	 */
3135 	if (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI)
3136 		info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
3137 
3138 	devid = get_devid(info);
3139 	if (devid < 0)
3140 		return -EINVAL;
3141 
3142 	ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
3143 	if (ret < 0)
3144 		return ret;
3145 
3146 	if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
3147 		struct irq_remap_table *table;
3148 		struct amd_iommu *iommu;
3149 
3150 		table = alloc_irq_table(devid, NULL);
3151 		if (table) {
3152 			if (!table->min_index) {
3153 				/*
3154 				 * Keep the first 32 indexes free for IOAPIC
3155 				 * interrupts.
3156 				 */
3157 				table->min_index = 32;
3158 				iommu = amd_iommu_rlookup_table[devid];
3159 				for (i = 0; i < 32; ++i)
3160 					iommu->irte_ops->set_allocated(table, i);
3161 			}
3162 			WARN_ON(table->min_index != 32);
3163 			index = info->ioapic.pin;
3164 		} else {
3165 			index = -ENOMEM;
3166 		}
3167 	} else if (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI ||
3168 		   info->type == X86_IRQ_ALLOC_TYPE_PCI_MSIX) {
3169 		bool align = (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI);
3170 
3171 		index = alloc_irq_index(devid, nr_irqs, align,
3172 					msi_desc_to_pci_dev(info->desc));
3173 	} else {
3174 		index = alloc_irq_index(devid, nr_irqs, false, NULL);
3175 	}
3176 
3177 	if (index < 0) {
3178 		pr_warn("Failed to allocate IRTE\n");
3179 		ret = index;
3180 		goto out_free_parent;
3181 	}
3182 
3183 	for (i = 0; i < nr_irqs; i++) {
3184 		irq_data = irq_domain_get_irq_data(domain, virq + i);
3185 		cfg = irq_data ? irqd_cfg(irq_data) : NULL;
3186 		if (!cfg) {
3187 			ret = -EINVAL;
3188 			goto out_free_data;
3189 		}
3190 
3191 		ret = -ENOMEM;
3192 		data = kzalloc(sizeof(*data), GFP_KERNEL);
3193 		if (!data)
3194 			goto out_free_data;
3195 
3196 		if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3197 			data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
3198 		else
3199 			data->entry = kzalloc(sizeof(struct irte_ga),
3200 						     GFP_KERNEL);
3201 		if (!data->entry) {
3202 			kfree(data);
3203 			goto out_free_data;
3204 		}
3205 
3206 		irq_data->hwirq = (devid << 16) + i;
3207 		irq_data->chip_data = data;
3208 		irq_data->chip = &amd_ir_chip;
3209 		irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
3210 		irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
3211 	}
3212 
3213 	return 0;
3214 
3215 out_free_data:
3216 	for (i--; i >= 0; i--) {
3217 		irq_data = irq_domain_get_irq_data(domain, virq + i);
3218 		if (irq_data)
3219 			kfree(irq_data->chip_data);
3220 	}
3221 	for (i = 0; i < nr_irqs; i++)
3222 		free_irte(devid, index + i);
3223 out_free_parent:
3224 	irq_domain_free_irqs_common(domain, virq, nr_irqs);
3225 	return ret;
3226 }
3227 
3228 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
3229 			       unsigned int nr_irqs)
3230 {
3231 	struct irq_2_irte *irte_info;
3232 	struct irq_data *irq_data;
3233 	struct amd_ir_data *data;
3234 	int i;
3235 
3236 	for (i = 0; i < nr_irqs; i++) {
3237 		irq_data = irq_domain_get_irq_data(domain, virq  + i);
3238 		if (irq_data && irq_data->chip_data) {
3239 			data = irq_data->chip_data;
3240 			irte_info = &data->irq_2_irte;
3241 			free_irte(irte_info->devid, irte_info->index);
3242 			kfree(data->entry);
3243 			kfree(data);
3244 		}
3245 	}
3246 	irq_domain_free_irqs_common(domain, virq, nr_irqs);
3247 }
3248 
3249 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
3250 			       struct amd_ir_data *ir_data,
3251 			       struct irq_2_irte *irte_info,
3252 			       struct irq_cfg *cfg);
3253 
3254 static int irq_remapping_activate(struct irq_domain *domain,
3255 				  struct irq_data *irq_data, bool reserve)
3256 {
3257 	struct amd_ir_data *data = irq_data->chip_data;
3258 	struct irq_2_irte *irte_info = &data->irq_2_irte;
3259 	struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
3260 	struct irq_cfg *cfg = irqd_cfg(irq_data);
3261 
3262 	if (!iommu)
3263 		return 0;
3264 
3265 	iommu->irte_ops->activate(data->entry, irte_info->devid,
3266 				  irte_info->index);
3267 	amd_ir_update_irte(irq_data, iommu, data, irte_info, cfg);
3268 	return 0;
3269 }
3270 
3271 static void irq_remapping_deactivate(struct irq_domain *domain,
3272 				     struct irq_data *irq_data)
3273 {
3274 	struct amd_ir_data *data = irq_data->chip_data;
3275 	struct irq_2_irte *irte_info = &data->irq_2_irte;
3276 	struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
3277 
3278 	if (iommu)
3279 		iommu->irte_ops->deactivate(data->entry, irte_info->devid,
3280 					    irte_info->index);
3281 }
3282 
3283 static int irq_remapping_select(struct irq_domain *d, struct irq_fwspec *fwspec,
3284 				enum irq_domain_bus_token bus_token)
3285 {
3286 	struct amd_iommu *iommu;
3287 	int devid = -1;
3288 
3289 	if (!amd_iommu_irq_remap)
3290 		return 0;
3291 
3292 	if (x86_fwspec_is_ioapic(fwspec))
3293 		devid = get_ioapic_devid(fwspec->param[0]);
3294 	else if (x86_fwspec_is_hpet(fwspec))
3295 		devid = get_hpet_devid(fwspec->param[0]);
3296 
3297 	if (devid < 0)
3298 		return 0;
3299 
3300 	iommu = amd_iommu_rlookup_table[devid];
3301 	return iommu && iommu->ir_domain == d;
3302 }
3303 
3304 static const struct irq_domain_ops amd_ir_domain_ops = {
3305 	.select = irq_remapping_select,
3306 	.alloc = irq_remapping_alloc,
3307 	.free = irq_remapping_free,
3308 	.activate = irq_remapping_activate,
3309 	.deactivate = irq_remapping_deactivate,
3310 };
3311 
3312 int amd_iommu_activate_guest_mode(void *data)
3313 {
3314 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3315 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3316 	u64 valid;
3317 
3318 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3319 	    !entry || entry->lo.fields_vapic.guest_mode)
3320 		return 0;
3321 
3322 	valid = entry->lo.fields_vapic.valid;
3323 
3324 	entry->lo.val = 0;
3325 	entry->hi.val = 0;
3326 
3327 	entry->lo.fields_vapic.valid       = valid;
3328 	entry->lo.fields_vapic.guest_mode  = 1;
3329 	entry->lo.fields_vapic.ga_log_intr = 1;
3330 	entry->hi.fields.ga_root_ptr       = ir_data->ga_root_ptr;
3331 	entry->hi.fields.vector            = ir_data->ga_vector;
3332 	entry->lo.fields_vapic.ga_tag      = ir_data->ga_tag;
3333 
3334 	return modify_irte_ga(ir_data->irq_2_irte.devid,
3335 			      ir_data->irq_2_irte.index, entry, ir_data);
3336 }
3337 EXPORT_SYMBOL(amd_iommu_activate_guest_mode);
3338 
3339 int amd_iommu_deactivate_guest_mode(void *data)
3340 {
3341 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3342 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3343 	struct irq_cfg *cfg = ir_data->cfg;
3344 	u64 valid;
3345 
3346 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3347 	    !entry || !entry->lo.fields_vapic.guest_mode)
3348 		return 0;
3349 
3350 	valid = entry->lo.fields_remap.valid;
3351 
3352 	entry->lo.val = 0;
3353 	entry->hi.val = 0;
3354 
3355 	entry->lo.fields_remap.valid       = valid;
3356 	entry->lo.fields_remap.dm          = apic->dest_mode_logical;
3357 	entry->lo.fields_remap.int_type    = apic->delivery_mode;
3358 	entry->hi.fields.vector            = cfg->vector;
3359 	entry->lo.fields_remap.destination =
3360 				APICID_TO_IRTE_DEST_LO(cfg->dest_apicid);
3361 	entry->hi.fields.destination =
3362 				APICID_TO_IRTE_DEST_HI(cfg->dest_apicid);
3363 
3364 	return modify_irte_ga(ir_data->irq_2_irte.devid,
3365 			      ir_data->irq_2_irte.index, entry, ir_data);
3366 }
3367 EXPORT_SYMBOL(amd_iommu_deactivate_guest_mode);
3368 
3369 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
3370 {
3371 	int ret;
3372 	struct amd_iommu *iommu;
3373 	struct amd_iommu_pi_data *pi_data = vcpu_info;
3374 	struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
3375 	struct amd_ir_data *ir_data = data->chip_data;
3376 	struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3377 	struct iommu_dev_data *dev_data = search_dev_data(irte_info->devid);
3378 
3379 	/* Note:
3380 	 * This device has never been set up for guest mode.
3381 	 * we should not modify the IRTE
3382 	 */
3383 	if (!dev_data || !dev_data->use_vapic)
3384 		return 0;
3385 
3386 	ir_data->cfg = irqd_cfg(data);
3387 	pi_data->ir_data = ir_data;
3388 
3389 	/* Note:
3390 	 * SVM tries to set up for VAPIC mode, but we are in
3391 	 * legacy mode. So, we force legacy mode instead.
3392 	 */
3393 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3394 		pr_debug("%s: Fall back to using intr legacy remap\n",
3395 			 __func__);
3396 		pi_data->is_guest_mode = false;
3397 	}
3398 
3399 	iommu = amd_iommu_rlookup_table[irte_info->devid];
3400 	if (iommu == NULL)
3401 		return -EINVAL;
3402 
3403 	pi_data->prev_ga_tag = ir_data->cached_ga_tag;
3404 	if (pi_data->is_guest_mode) {
3405 		ir_data->ga_root_ptr = (pi_data->base >> 12);
3406 		ir_data->ga_vector = vcpu_pi_info->vector;
3407 		ir_data->ga_tag = pi_data->ga_tag;
3408 		ret = amd_iommu_activate_guest_mode(ir_data);
3409 		if (!ret)
3410 			ir_data->cached_ga_tag = pi_data->ga_tag;
3411 	} else {
3412 		ret = amd_iommu_deactivate_guest_mode(ir_data);
3413 
3414 		/*
3415 		 * This communicates the ga_tag back to the caller
3416 		 * so that it can do all the necessary clean up.
3417 		 */
3418 		if (!ret)
3419 			ir_data->cached_ga_tag = 0;
3420 	}
3421 
3422 	return ret;
3423 }
3424 
3425 
3426 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
3427 			       struct amd_ir_data *ir_data,
3428 			       struct irq_2_irte *irte_info,
3429 			       struct irq_cfg *cfg)
3430 {
3431 
3432 	/*
3433 	 * Atomically updates the IRTE with the new destination, vector
3434 	 * and flushes the interrupt entry cache.
3435 	 */
3436 	iommu->irte_ops->set_affinity(ir_data->entry, irte_info->devid,
3437 				      irte_info->index, cfg->vector,
3438 				      cfg->dest_apicid);
3439 }
3440 
3441 static int amd_ir_set_affinity(struct irq_data *data,
3442 			       const struct cpumask *mask, bool force)
3443 {
3444 	struct amd_ir_data *ir_data = data->chip_data;
3445 	struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3446 	struct irq_cfg *cfg = irqd_cfg(data);
3447 	struct irq_data *parent = data->parent_data;
3448 	struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
3449 	int ret;
3450 
3451 	if (!iommu)
3452 		return -ENODEV;
3453 
3454 	ret = parent->chip->irq_set_affinity(parent, mask, force);
3455 	if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
3456 		return ret;
3457 
3458 	amd_ir_update_irte(data, iommu, ir_data, irte_info, cfg);
3459 	/*
3460 	 * After this point, all the interrupts will start arriving
3461 	 * at the new destination. So, time to cleanup the previous
3462 	 * vector allocation.
3463 	 */
3464 	send_cleanup_vector(cfg);
3465 
3466 	return IRQ_SET_MASK_OK_DONE;
3467 }
3468 
3469 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
3470 {
3471 	struct amd_ir_data *ir_data = irq_data->chip_data;
3472 
3473 	*msg = ir_data->msi_entry;
3474 }
3475 
3476 static struct irq_chip amd_ir_chip = {
3477 	.name			= "AMD-IR",
3478 	.irq_ack		= apic_ack_irq,
3479 	.irq_set_affinity	= amd_ir_set_affinity,
3480 	.irq_set_vcpu_affinity	= amd_ir_set_vcpu_affinity,
3481 	.irq_compose_msi_msg	= ir_compose_msi_msg,
3482 };
3483 
3484 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
3485 {
3486 	struct fwnode_handle *fn;
3487 
3488 	fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index);
3489 	if (!fn)
3490 		return -ENOMEM;
3491 	iommu->ir_domain = irq_domain_create_tree(fn, &amd_ir_domain_ops, iommu);
3492 	if (!iommu->ir_domain) {
3493 		irq_domain_free_fwnode(fn);
3494 		return -ENOMEM;
3495 	}
3496 
3497 	iommu->ir_domain->parent = arch_get_ir_parent_domain();
3498 	iommu->msi_domain = arch_create_remap_msi_irq_domain(iommu->ir_domain,
3499 							     "AMD-IR-MSI",
3500 							     iommu->index);
3501 	return 0;
3502 }
3503 
3504 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
3505 {
3506 	unsigned long flags;
3507 	struct amd_iommu *iommu;
3508 	struct irq_remap_table *table;
3509 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3510 	int devid = ir_data->irq_2_irte.devid;
3511 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3512 	struct irte_ga *ref = (struct irte_ga *) ir_data->ref;
3513 
3514 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3515 	    !ref || !entry || !entry->lo.fields_vapic.guest_mode)
3516 		return 0;
3517 
3518 	iommu = amd_iommu_rlookup_table[devid];
3519 	if (!iommu)
3520 		return -ENODEV;
3521 
3522 	table = get_irq_table(devid);
3523 	if (!table)
3524 		return -ENODEV;
3525 
3526 	raw_spin_lock_irqsave(&table->lock, flags);
3527 
3528 	if (ref->lo.fields_vapic.guest_mode) {
3529 		if (cpu >= 0) {
3530 			ref->lo.fields_vapic.destination =
3531 						APICID_TO_IRTE_DEST_LO(cpu);
3532 			ref->hi.fields.destination =
3533 						APICID_TO_IRTE_DEST_HI(cpu);
3534 		}
3535 		ref->lo.fields_vapic.is_run = is_run;
3536 		barrier();
3537 	}
3538 
3539 	raw_spin_unlock_irqrestore(&table->lock, flags);
3540 
3541 	iommu_flush_irt(iommu, devid);
3542 	iommu_completion_wait(iommu);
3543 	return 0;
3544 }
3545 EXPORT_SYMBOL(amd_iommu_update_ga);
3546 #endif
3547