xref: /openbmc/linux/drivers/iommu/amd/iommu.c (revision c29b9772)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
4  * Author: Joerg Roedel <jroedel@suse.de>
5  *         Leo Duran <leo.duran@amd.com>
6  */
7 
8 #define pr_fmt(fmt)     "AMD-Vi: " fmt
9 #define dev_fmt(fmt)    pr_fmt(fmt)
10 
11 #include <linux/ratelimit.h>
12 #include <linux/pci.h>
13 #include <linux/acpi.h>
14 #include <linux/pci-ats.h>
15 #include <linux/bitmap.h>
16 #include <linux/slab.h>
17 #include <linux/debugfs.h>
18 #include <linux/scatterlist.h>
19 #include <linux/dma-map-ops.h>
20 #include <linux/dma-direct.h>
21 #include <linux/iommu-helper.h>
22 #include <linux/delay.h>
23 #include <linux/amd-iommu.h>
24 #include <linux/notifier.h>
25 #include <linux/export.h>
26 #include <linux/irq.h>
27 #include <linux/msi.h>
28 #include <linux/irqdomain.h>
29 #include <linux/percpu.h>
30 #include <linux/io-pgtable.h>
31 #include <linux/cc_platform.h>
32 #include <asm/irq_remapping.h>
33 #include <asm/io_apic.h>
34 #include <asm/apic.h>
35 #include <asm/hw_irq.h>
36 #include <asm/proto.h>
37 #include <asm/iommu.h>
38 #include <asm/gart.h>
39 #include <asm/dma.h>
40 
41 #include "amd_iommu.h"
42 #include "../dma-iommu.h"
43 #include "../irq_remapping.h"
44 
45 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
46 
47 #define LOOP_TIMEOUT	100000
48 
49 /* IO virtual address start page frame number */
50 #define IOVA_START_PFN		(1)
51 #define IOVA_PFN(addr)		((addr) >> PAGE_SHIFT)
52 
53 /* Reserved IOVA ranges */
54 #define MSI_RANGE_START		(0xfee00000)
55 #define MSI_RANGE_END		(0xfeefffff)
56 #define HT_RANGE_START		(0xfd00000000ULL)
57 #define HT_RANGE_END		(0xffffffffffULL)
58 
59 #define DEFAULT_PGTABLE_LEVEL	PAGE_MODE_3_LEVEL
60 
61 static DEFINE_SPINLOCK(pd_bitmap_lock);
62 
63 LIST_HEAD(ioapic_map);
64 LIST_HEAD(hpet_map);
65 LIST_HEAD(acpihid_map);
66 
67 const struct iommu_ops amd_iommu_ops;
68 
69 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
70 int amd_iommu_max_glx_val = -1;
71 
72 /*
73  * general struct to manage commands send to an IOMMU
74  */
75 struct iommu_cmd {
76 	u32 data[4];
77 };
78 
79 struct kmem_cache *amd_iommu_irq_cache;
80 
81 static void detach_device(struct device *dev);
82 static int domain_enable_v2(struct protection_domain *domain, int pasids);
83 
84 /****************************************************************************
85  *
86  * Helper functions
87  *
88  ****************************************************************************/
89 
90 static inline int get_acpihid_device_id(struct device *dev,
91 					struct acpihid_map_entry **entry)
92 {
93 	struct acpi_device *adev = ACPI_COMPANION(dev);
94 	struct acpihid_map_entry *p;
95 
96 	if (!adev)
97 		return -ENODEV;
98 
99 	list_for_each_entry(p, &acpihid_map, list) {
100 		if (acpi_dev_hid_uid_match(adev, p->hid,
101 					   p->uid[0] ? p->uid : NULL)) {
102 			if (entry)
103 				*entry = p;
104 			return p->devid;
105 		}
106 	}
107 	return -EINVAL;
108 }
109 
110 static inline int get_device_sbdf_id(struct device *dev)
111 {
112 	int sbdf;
113 
114 	if (dev_is_pci(dev))
115 		sbdf = get_pci_sbdf_id(to_pci_dev(dev));
116 	else
117 		sbdf = get_acpihid_device_id(dev, NULL);
118 
119 	return sbdf;
120 }
121 
122 struct dev_table_entry *get_dev_table(struct amd_iommu *iommu)
123 {
124 	struct dev_table_entry *dev_table;
125 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
126 
127 	BUG_ON(pci_seg == NULL);
128 	dev_table = pci_seg->dev_table;
129 	BUG_ON(dev_table == NULL);
130 
131 	return dev_table;
132 }
133 
134 static inline u16 get_device_segment(struct device *dev)
135 {
136 	u16 seg;
137 
138 	if (dev_is_pci(dev)) {
139 		struct pci_dev *pdev = to_pci_dev(dev);
140 
141 		seg = pci_domain_nr(pdev->bus);
142 	} else {
143 		u32 devid = get_acpihid_device_id(dev, NULL);
144 
145 		seg = PCI_SBDF_TO_SEGID(devid);
146 	}
147 
148 	return seg;
149 }
150 
151 /* Writes the specific IOMMU for a device into the PCI segment rlookup table */
152 void amd_iommu_set_rlookup_table(struct amd_iommu *iommu, u16 devid)
153 {
154 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
155 
156 	pci_seg->rlookup_table[devid] = iommu;
157 }
158 
159 static struct amd_iommu *__rlookup_amd_iommu(u16 seg, u16 devid)
160 {
161 	struct amd_iommu_pci_seg *pci_seg;
162 
163 	for_each_pci_segment(pci_seg) {
164 		if (pci_seg->id == seg)
165 			return pci_seg->rlookup_table[devid];
166 	}
167 	return NULL;
168 }
169 
170 static struct amd_iommu *rlookup_amd_iommu(struct device *dev)
171 {
172 	u16 seg = get_device_segment(dev);
173 	int devid = get_device_sbdf_id(dev);
174 
175 	if (devid < 0)
176 		return NULL;
177 	return __rlookup_amd_iommu(seg, PCI_SBDF_TO_DEVID(devid));
178 }
179 
180 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
181 {
182 	return container_of(dom, struct protection_domain, domain);
183 }
184 
185 static struct iommu_dev_data *alloc_dev_data(struct amd_iommu *iommu, u16 devid)
186 {
187 	struct iommu_dev_data *dev_data;
188 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
189 
190 	dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
191 	if (!dev_data)
192 		return NULL;
193 
194 	spin_lock_init(&dev_data->lock);
195 	dev_data->devid = devid;
196 	ratelimit_default_init(&dev_data->rs);
197 
198 	llist_add(&dev_data->dev_data_list, &pci_seg->dev_data_list);
199 	return dev_data;
200 }
201 
202 static struct iommu_dev_data *search_dev_data(struct amd_iommu *iommu, u16 devid)
203 {
204 	struct iommu_dev_data *dev_data;
205 	struct llist_node *node;
206 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
207 
208 	if (llist_empty(&pci_seg->dev_data_list))
209 		return NULL;
210 
211 	node = pci_seg->dev_data_list.first;
212 	llist_for_each_entry(dev_data, node, dev_data_list) {
213 		if (dev_data->devid == devid)
214 			return dev_data;
215 	}
216 
217 	return NULL;
218 }
219 
220 static int clone_alias(struct pci_dev *pdev, u16 alias, void *data)
221 {
222 	struct amd_iommu *iommu;
223 	struct dev_table_entry *dev_table;
224 	u16 devid = pci_dev_id(pdev);
225 
226 	if (devid == alias)
227 		return 0;
228 
229 	iommu = rlookup_amd_iommu(&pdev->dev);
230 	if (!iommu)
231 		return 0;
232 
233 	amd_iommu_set_rlookup_table(iommu, alias);
234 	dev_table = get_dev_table(iommu);
235 	memcpy(dev_table[alias].data,
236 	       dev_table[devid].data,
237 	       sizeof(dev_table[alias].data));
238 
239 	return 0;
240 }
241 
242 static void clone_aliases(struct amd_iommu *iommu, struct device *dev)
243 {
244 	struct pci_dev *pdev;
245 
246 	if (!dev_is_pci(dev))
247 		return;
248 	pdev = to_pci_dev(dev);
249 
250 	/*
251 	 * The IVRS alias stored in the alias table may not be
252 	 * part of the PCI DMA aliases if it's bus differs
253 	 * from the original device.
254 	 */
255 	clone_alias(pdev, iommu->pci_seg->alias_table[pci_dev_id(pdev)], NULL);
256 
257 	pci_for_each_dma_alias(pdev, clone_alias, NULL);
258 }
259 
260 static void setup_aliases(struct amd_iommu *iommu, struct device *dev)
261 {
262 	struct pci_dev *pdev = to_pci_dev(dev);
263 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
264 	u16 ivrs_alias;
265 
266 	/* For ACPI HID devices, there are no aliases */
267 	if (!dev_is_pci(dev))
268 		return;
269 
270 	/*
271 	 * Add the IVRS alias to the pci aliases if it is on the same
272 	 * bus. The IVRS table may know about a quirk that we don't.
273 	 */
274 	ivrs_alias = pci_seg->alias_table[pci_dev_id(pdev)];
275 	if (ivrs_alias != pci_dev_id(pdev) &&
276 	    PCI_BUS_NUM(ivrs_alias) == pdev->bus->number)
277 		pci_add_dma_alias(pdev, ivrs_alias & 0xff, 1);
278 
279 	clone_aliases(iommu, dev);
280 }
281 
282 static struct iommu_dev_data *find_dev_data(struct amd_iommu *iommu, u16 devid)
283 {
284 	struct iommu_dev_data *dev_data;
285 
286 	dev_data = search_dev_data(iommu, devid);
287 
288 	if (dev_data == NULL) {
289 		dev_data = alloc_dev_data(iommu, devid);
290 		if (!dev_data)
291 			return NULL;
292 
293 		if (translation_pre_enabled(iommu))
294 			dev_data->defer_attach = true;
295 	}
296 
297 	return dev_data;
298 }
299 
300 /*
301 * Find or create an IOMMU group for a acpihid device.
302 */
303 static struct iommu_group *acpihid_device_group(struct device *dev)
304 {
305 	struct acpihid_map_entry *p, *entry = NULL;
306 	int devid;
307 
308 	devid = get_acpihid_device_id(dev, &entry);
309 	if (devid < 0)
310 		return ERR_PTR(devid);
311 
312 	list_for_each_entry(p, &acpihid_map, list) {
313 		if ((devid == p->devid) && p->group)
314 			entry->group = p->group;
315 	}
316 
317 	if (!entry->group)
318 		entry->group = generic_device_group(dev);
319 	else
320 		iommu_group_ref_get(entry->group);
321 
322 	return entry->group;
323 }
324 
325 static bool pci_iommuv2_capable(struct pci_dev *pdev)
326 {
327 	static const int caps[] = {
328 		PCI_EXT_CAP_ID_PRI,
329 		PCI_EXT_CAP_ID_PASID,
330 	};
331 	int i, pos;
332 
333 	if (!pci_ats_supported(pdev))
334 		return false;
335 
336 	for (i = 0; i < 2; ++i) {
337 		pos = pci_find_ext_capability(pdev, caps[i]);
338 		if (pos == 0)
339 			return false;
340 	}
341 
342 	return true;
343 }
344 
345 /*
346  * This function checks if the driver got a valid device from the caller to
347  * avoid dereferencing invalid pointers.
348  */
349 static bool check_device(struct device *dev)
350 {
351 	struct amd_iommu_pci_seg *pci_seg;
352 	struct amd_iommu *iommu;
353 	int devid, sbdf;
354 
355 	if (!dev)
356 		return false;
357 
358 	sbdf = get_device_sbdf_id(dev);
359 	if (sbdf < 0)
360 		return false;
361 	devid = PCI_SBDF_TO_DEVID(sbdf);
362 
363 	iommu = rlookup_amd_iommu(dev);
364 	if (!iommu)
365 		return false;
366 
367 	/* Out of our scope? */
368 	pci_seg = iommu->pci_seg;
369 	if (devid > pci_seg->last_bdf)
370 		return false;
371 
372 	return true;
373 }
374 
375 static int iommu_init_device(struct amd_iommu *iommu, struct device *dev)
376 {
377 	struct iommu_dev_data *dev_data;
378 	int devid, sbdf;
379 
380 	if (dev_iommu_priv_get(dev))
381 		return 0;
382 
383 	sbdf = get_device_sbdf_id(dev);
384 	if (sbdf < 0)
385 		return sbdf;
386 
387 	devid = PCI_SBDF_TO_DEVID(sbdf);
388 	dev_data = find_dev_data(iommu, devid);
389 	if (!dev_data)
390 		return -ENOMEM;
391 
392 	dev_data->dev = dev;
393 	setup_aliases(iommu, dev);
394 
395 	/*
396 	 * By default we use passthrough mode for IOMMUv2 capable device.
397 	 * But if amd_iommu=force_isolation is set (e.g. to debug DMA to
398 	 * invalid address), we ignore the capability for the device so
399 	 * it'll be forced to go into translation mode.
400 	 */
401 	if ((iommu_default_passthrough() || !amd_iommu_force_isolation) &&
402 	    dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
403 		dev_data->iommu_v2 = iommu->is_iommu_v2;
404 	}
405 
406 	dev_iommu_priv_set(dev, dev_data);
407 
408 	return 0;
409 }
410 
411 static void iommu_ignore_device(struct amd_iommu *iommu, struct device *dev)
412 {
413 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
414 	struct dev_table_entry *dev_table = get_dev_table(iommu);
415 	int devid, sbdf;
416 
417 	sbdf = get_device_sbdf_id(dev);
418 	if (sbdf < 0)
419 		return;
420 
421 	devid = PCI_SBDF_TO_DEVID(sbdf);
422 	pci_seg->rlookup_table[devid] = NULL;
423 	memset(&dev_table[devid], 0, sizeof(struct dev_table_entry));
424 
425 	setup_aliases(iommu, dev);
426 }
427 
428 static void amd_iommu_uninit_device(struct device *dev)
429 {
430 	struct iommu_dev_data *dev_data;
431 
432 	dev_data = dev_iommu_priv_get(dev);
433 	if (!dev_data)
434 		return;
435 
436 	if (dev_data->domain)
437 		detach_device(dev);
438 
439 	dev_iommu_priv_set(dev, NULL);
440 
441 	/*
442 	 * We keep dev_data around for unplugged devices and reuse it when the
443 	 * device is re-plugged - not doing so would introduce a ton of races.
444 	 */
445 }
446 
447 /****************************************************************************
448  *
449  * Interrupt handling functions
450  *
451  ****************************************************************************/
452 
453 static void dump_dte_entry(struct amd_iommu *iommu, u16 devid)
454 {
455 	int i;
456 	struct dev_table_entry *dev_table = get_dev_table(iommu);
457 
458 	for (i = 0; i < 4; ++i)
459 		pr_err("DTE[%d]: %016llx\n", i, dev_table[devid].data[i]);
460 }
461 
462 static void dump_command(unsigned long phys_addr)
463 {
464 	struct iommu_cmd *cmd = iommu_phys_to_virt(phys_addr);
465 	int i;
466 
467 	for (i = 0; i < 4; ++i)
468 		pr_err("CMD[%d]: %08x\n", i, cmd->data[i]);
469 }
470 
471 static void amd_iommu_report_rmp_hw_error(struct amd_iommu *iommu, volatile u32 *event)
472 {
473 	struct iommu_dev_data *dev_data = NULL;
474 	int devid, vmg_tag, flags;
475 	struct pci_dev *pdev;
476 	u64 spa;
477 
478 	devid   = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
479 	vmg_tag = (event[1]) & 0xFFFF;
480 	flags   = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
481 	spa     = ((u64)event[3] << 32) | (event[2] & 0xFFFFFFF8);
482 
483 	pdev = pci_get_domain_bus_and_slot(iommu->pci_seg->id, PCI_BUS_NUM(devid),
484 					   devid & 0xff);
485 	if (pdev)
486 		dev_data = dev_iommu_priv_get(&pdev->dev);
487 
488 	if (dev_data) {
489 		if (__ratelimit(&dev_data->rs)) {
490 			pci_err(pdev, "Event logged [RMP_HW_ERROR vmg_tag=0x%04x, spa=0x%llx, flags=0x%04x]\n",
491 				vmg_tag, spa, flags);
492 		}
493 	} else {
494 		pr_err_ratelimited("Event logged [RMP_HW_ERROR device=%04x:%02x:%02x.%x, vmg_tag=0x%04x, spa=0x%llx, flags=0x%04x]\n",
495 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
496 			vmg_tag, spa, flags);
497 	}
498 
499 	if (pdev)
500 		pci_dev_put(pdev);
501 }
502 
503 static void amd_iommu_report_rmp_fault(struct amd_iommu *iommu, volatile u32 *event)
504 {
505 	struct iommu_dev_data *dev_data = NULL;
506 	int devid, flags_rmp, vmg_tag, flags;
507 	struct pci_dev *pdev;
508 	u64 gpa;
509 
510 	devid     = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
511 	flags_rmp = (event[0] >> EVENT_FLAGS_SHIFT) & 0xFF;
512 	vmg_tag   = (event[1]) & 0xFFFF;
513 	flags     = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
514 	gpa       = ((u64)event[3] << 32) | event[2];
515 
516 	pdev = pci_get_domain_bus_and_slot(iommu->pci_seg->id, PCI_BUS_NUM(devid),
517 					   devid & 0xff);
518 	if (pdev)
519 		dev_data = dev_iommu_priv_get(&pdev->dev);
520 
521 	if (dev_data) {
522 		if (__ratelimit(&dev_data->rs)) {
523 			pci_err(pdev, "Event logged [RMP_PAGE_FAULT vmg_tag=0x%04x, gpa=0x%llx, flags_rmp=0x%04x, flags=0x%04x]\n",
524 				vmg_tag, gpa, flags_rmp, flags);
525 		}
526 	} else {
527 		pr_err_ratelimited("Event logged [RMP_PAGE_FAULT device=%04x:%02x:%02x.%x, vmg_tag=0x%04x, gpa=0x%llx, flags_rmp=0x%04x, flags=0x%04x]\n",
528 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
529 			vmg_tag, gpa, flags_rmp, flags);
530 	}
531 
532 	if (pdev)
533 		pci_dev_put(pdev);
534 }
535 
536 #define IS_IOMMU_MEM_TRANSACTION(flags)		\
537 	(((flags) & EVENT_FLAG_I) == 0)
538 
539 #define IS_WRITE_REQUEST(flags)			\
540 	((flags) & EVENT_FLAG_RW)
541 
542 static void amd_iommu_report_page_fault(struct amd_iommu *iommu,
543 					u16 devid, u16 domain_id,
544 					u64 address, int flags)
545 {
546 	struct iommu_dev_data *dev_data = NULL;
547 	struct pci_dev *pdev;
548 
549 	pdev = pci_get_domain_bus_and_slot(iommu->pci_seg->id, PCI_BUS_NUM(devid),
550 					   devid & 0xff);
551 	if (pdev)
552 		dev_data = dev_iommu_priv_get(&pdev->dev);
553 
554 	if (dev_data) {
555 		/*
556 		 * If this is a DMA fault (for which the I(nterrupt)
557 		 * bit will be unset), allow report_iommu_fault() to
558 		 * prevent logging it.
559 		 */
560 		if (IS_IOMMU_MEM_TRANSACTION(flags)) {
561 			/* Device not attached to domain properly */
562 			if (dev_data->domain == NULL) {
563 				pr_err_ratelimited("Event logged [Device not attached to domain properly]\n");
564 				pr_err_ratelimited("  device=%04x:%02x:%02x.%x domain=0x%04x\n",
565 						   iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid),
566 						   PCI_FUNC(devid), domain_id);
567 				goto out;
568 			}
569 
570 			if (!report_iommu_fault(&dev_data->domain->domain,
571 						&pdev->dev, address,
572 						IS_WRITE_REQUEST(flags) ?
573 							IOMMU_FAULT_WRITE :
574 							IOMMU_FAULT_READ))
575 				goto out;
576 		}
577 
578 		if (__ratelimit(&dev_data->rs)) {
579 			pci_err(pdev, "Event logged [IO_PAGE_FAULT domain=0x%04x address=0x%llx flags=0x%04x]\n",
580 				domain_id, address, flags);
581 		}
582 	} else {
583 		pr_err_ratelimited("Event logged [IO_PAGE_FAULT device=%04x:%02x:%02x.%x domain=0x%04x address=0x%llx flags=0x%04x]\n",
584 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
585 			domain_id, address, flags);
586 	}
587 
588 out:
589 	if (pdev)
590 		pci_dev_put(pdev);
591 }
592 
593 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
594 {
595 	struct device *dev = iommu->iommu.dev;
596 	int type, devid, flags, tag;
597 	volatile u32 *event = __evt;
598 	int count = 0;
599 	u64 address;
600 	u32 pasid;
601 
602 retry:
603 	type    = (event[1] >> EVENT_TYPE_SHIFT)  & EVENT_TYPE_MASK;
604 	devid   = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
605 	pasid   = (event[0] & EVENT_DOMID_MASK_HI) |
606 		  (event[1] & EVENT_DOMID_MASK_LO);
607 	flags   = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
608 	address = (u64)(((u64)event[3]) << 32) | event[2];
609 
610 	if (type == 0) {
611 		/* Did we hit the erratum? */
612 		if (++count == LOOP_TIMEOUT) {
613 			pr_err("No event written to event log\n");
614 			return;
615 		}
616 		udelay(1);
617 		goto retry;
618 	}
619 
620 	if (type == EVENT_TYPE_IO_FAULT) {
621 		amd_iommu_report_page_fault(iommu, devid, pasid, address, flags);
622 		return;
623 	}
624 
625 	switch (type) {
626 	case EVENT_TYPE_ILL_DEV:
627 		dev_err(dev, "Event logged [ILLEGAL_DEV_TABLE_ENTRY device=%04x:%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
628 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
629 			pasid, address, flags);
630 		dump_dte_entry(iommu, devid);
631 		break;
632 	case EVENT_TYPE_DEV_TAB_ERR:
633 		dev_err(dev, "Event logged [DEV_TAB_HARDWARE_ERROR device=%04x:%02x:%02x.%x "
634 			"address=0x%llx flags=0x%04x]\n",
635 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
636 			address, flags);
637 		break;
638 	case EVENT_TYPE_PAGE_TAB_ERR:
639 		dev_err(dev, "Event logged [PAGE_TAB_HARDWARE_ERROR device=%04x:%02x:%02x.%x pasid=0x%04x address=0x%llx flags=0x%04x]\n",
640 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
641 			pasid, address, flags);
642 		break;
643 	case EVENT_TYPE_ILL_CMD:
644 		dev_err(dev, "Event logged [ILLEGAL_COMMAND_ERROR address=0x%llx]\n", address);
645 		dump_command(address);
646 		break;
647 	case EVENT_TYPE_CMD_HARD_ERR:
648 		dev_err(dev, "Event logged [COMMAND_HARDWARE_ERROR address=0x%llx flags=0x%04x]\n",
649 			address, flags);
650 		break;
651 	case EVENT_TYPE_IOTLB_INV_TO:
652 		dev_err(dev, "Event logged [IOTLB_INV_TIMEOUT device=%04x:%02x:%02x.%x address=0x%llx]\n",
653 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
654 			address);
655 		break;
656 	case EVENT_TYPE_INV_DEV_REQ:
657 		dev_err(dev, "Event logged [INVALID_DEVICE_REQUEST device=%04x:%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
658 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
659 			pasid, address, flags);
660 		break;
661 	case EVENT_TYPE_RMP_FAULT:
662 		amd_iommu_report_rmp_fault(iommu, event);
663 		break;
664 	case EVENT_TYPE_RMP_HW_ERR:
665 		amd_iommu_report_rmp_hw_error(iommu, event);
666 		break;
667 	case EVENT_TYPE_INV_PPR_REQ:
668 		pasid = PPR_PASID(*((u64 *)__evt));
669 		tag = event[1] & 0x03FF;
670 		dev_err(dev, "Event logged [INVALID_PPR_REQUEST device=%04x:%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x tag=0x%03x]\n",
671 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
672 			pasid, address, flags, tag);
673 		break;
674 	default:
675 		dev_err(dev, "Event logged [UNKNOWN event[0]=0x%08x event[1]=0x%08x event[2]=0x%08x event[3]=0x%08x\n",
676 			event[0], event[1], event[2], event[3]);
677 	}
678 
679 	/*
680 	 * To detect the hardware errata 732 we need to clear the
681 	 * entry back to zero. This issue does not exist on SNP
682 	 * enabled system. Also this buffer is not writeable on
683 	 * SNP enabled system.
684 	 */
685 	if (!amd_iommu_snp_en)
686 		memset(__evt, 0, 4 * sizeof(u32));
687 }
688 
689 static void iommu_poll_events(struct amd_iommu *iommu)
690 {
691 	u32 head, tail;
692 
693 	head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
694 	tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
695 
696 	while (head != tail) {
697 		iommu_print_event(iommu, iommu->evt_buf + head);
698 		head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
699 	}
700 
701 	writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
702 }
703 
704 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
705 {
706 	struct amd_iommu_fault fault;
707 
708 	if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
709 		pr_err_ratelimited("Unknown PPR request received\n");
710 		return;
711 	}
712 
713 	fault.address   = raw[1];
714 	fault.pasid     = PPR_PASID(raw[0]);
715 	fault.sbdf      = PCI_SEG_DEVID_TO_SBDF(iommu->pci_seg->id, PPR_DEVID(raw[0]));
716 	fault.tag       = PPR_TAG(raw[0]);
717 	fault.flags     = PPR_FLAGS(raw[0]);
718 
719 	atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
720 }
721 
722 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
723 {
724 	u32 head, tail;
725 
726 	if (iommu->ppr_log == NULL)
727 		return;
728 
729 	head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
730 	tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
731 
732 	while (head != tail) {
733 		volatile u64 *raw;
734 		u64 entry[2];
735 		int i;
736 
737 		raw = (u64 *)(iommu->ppr_log + head);
738 
739 		/*
740 		 * Hardware bug: Interrupt may arrive before the entry is
741 		 * written to memory. If this happens we need to wait for the
742 		 * entry to arrive.
743 		 */
744 		for (i = 0; i < LOOP_TIMEOUT; ++i) {
745 			if (PPR_REQ_TYPE(raw[0]) != 0)
746 				break;
747 			udelay(1);
748 		}
749 
750 		/* Avoid memcpy function-call overhead */
751 		entry[0] = raw[0];
752 		entry[1] = raw[1];
753 
754 		/*
755 		 * To detect the hardware errata 733 we need to clear the
756 		 * entry back to zero. This issue does not exist on SNP
757 		 * enabled system. Also this buffer is not writeable on
758 		 * SNP enabled system.
759 		 */
760 		if (!amd_iommu_snp_en)
761 			raw[0] = raw[1] = 0UL;
762 
763 		/* Update head pointer of hardware ring-buffer */
764 		head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
765 		writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
766 
767 		/* Handle PPR entry */
768 		iommu_handle_ppr_entry(iommu, entry);
769 
770 		/* Refresh ring-buffer information */
771 		head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
772 		tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
773 	}
774 }
775 
776 #ifdef CONFIG_IRQ_REMAP
777 static int (*iommu_ga_log_notifier)(u32);
778 
779 int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
780 {
781 	iommu_ga_log_notifier = notifier;
782 
783 	return 0;
784 }
785 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
786 
787 static void iommu_poll_ga_log(struct amd_iommu *iommu)
788 {
789 	u32 head, tail;
790 
791 	if (iommu->ga_log == NULL)
792 		return;
793 
794 	head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
795 	tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
796 
797 	while (head != tail) {
798 		volatile u64 *raw;
799 		u64 log_entry;
800 
801 		raw = (u64 *)(iommu->ga_log + head);
802 
803 		/* Avoid memcpy function-call overhead */
804 		log_entry = *raw;
805 
806 		/* Update head pointer of hardware ring-buffer */
807 		head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
808 		writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
809 
810 		/* Handle GA entry */
811 		switch (GA_REQ_TYPE(log_entry)) {
812 		case GA_GUEST_NR:
813 			if (!iommu_ga_log_notifier)
814 				break;
815 
816 			pr_debug("%s: devid=%#x, ga_tag=%#x\n",
817 				 __func__, GA_DEVID(log_entry),
818 				 GA_TAG(log_entry));
819 
820 			if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
821 				pr_err("GA log notifier failed.\n");
822 			break;
823 		default:
824 			break;
825 		}
826 	}
827 }
828 
829 static void
830 amd_iommu_set_pci_msi_domain(struct device *dev, struct amd_iommu *iommu)
831 {
832 	if (!irq_remapping_enabled || !dev_is_pci(dev) ||
833 	    !pci_dev_has_default_msi_parent_domain(to_pci_dev(dev)))
834 		return;
835 
836 	dev_set_msi_domain(dev, iommu->ir_domain);
837 }
838 
839 #else /* CONFIG_IRQ_REMAP */
840 static inline void
841 amd_iommu_set_pci_msi_domain(struct device *dev, struct amd_iommu *iommu) { }
842 #endif /* !CONFIG_IRQ_REMAP */
843 
844 #define AMD_IOMMU_INT_MASK	\
845 	(MMIO_STATUS_EVT_OVERFLOW_INT_MASK | \
846 	 MMIO_STATUS_EVT_INT_MASK | \
847 	 MMIO_STATUS_PPR_INT_MASK | \
848 	 MMIO_STATUS_GALOG_INT_MASK)
849 
850 irqreturn_t amd_iommu_int_thread(int irq, void *data)
851 {
852 	struct amd_iommu *iommu = (struct amd_iommu *) data;
853 	u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
854 
855 	while (status & AMD_IOMMU_INT_MASK) {
856 		/* Enable interrupt sources again */
857 		writel(AMD_IOMMU_INT_MASK,
858 			iommu->mmio_base + MMIO_STATUS_OFFSET);
859 
860 		if (status & MMIO_STATUS_EVT_INT_MASK) {
861 			pr_devel("Processing IOMMU Event Log\n");
862 			iommu_poll_events(iommu);
863 		}
864 
865 		if (status & MMIO_STATUS_PPR_INT_MASK) {
866 			pr_devel("Processing IOMMU PPR Log\n");
867 			iommu_poll_ppr_log(iommu);
868 		}
869 
870 #ifdef CONFIG_IRQ_REMAP
871 		if (status & MMIO_STATUS_GALOG_INT_MASK) {
872 			pr_devel("Processing IOMMU GA Log\n");
873 			iommu_poll_ga_log(iommu);
874 		}
875 #endif
876 
877 		if (status & MMIO_STATUS_EVT_OVERFLOW_INT_MASK) {
878 			pr_info_ratelimited("IOMMU event log overflow\n");
879 			amd_iommu_restart_event_logging(iommu);
880 		}
881 
882 		/*
883 		 * Hardware bug: ERBT1312
884 		 * When re-enabling interrupt (by writing 1
885 		 * to clear the bit), the hardware might also try to set
886 		 * the interrupt bit in the event status register.
887 		 * In this scenario, the bit will be set, and disable
888 		 * subsequent interrupts.
889 		 *
890 		 * Workaround: The IOMMU driver should read back the
891 		 * status register and check if the interrupt bits are cleared.
892 		 * If not, driver will need to go through the interrupt handler
893 		 * again and re-clear the bits
894 		 */
895 		status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
896 	}
897 	return IRQ_HANDLED;
898 }
899 
900 irqreturn_t amd_iommu_int_handler(int irq, void *data)
901 {
902 	return IRQ_WAKE_THREAD;
903 }
904 
905 /****************************************************************************
906  *
907  * IOMMU command queuing functions
908  *
909  ****************************************************************************/
910 
911 static int wait_on_sem(struct amd_iommu *iommu, u64 data)
912 {
913 	int i = 0;
914 
915 	while (*iommu->cmd_sem != data && i < LOOP_TIMEOUT) {
916 		udelay(1);
917 		i += 1;
918 	}
919 
920 	if (i == LOOP_TIMEOUT) {
921 		pr_alert("Completion-Wait loop timed out\n");
922 		return -EIO;
923 	}
924 
925 	return 0;
926 }
927 
928 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
929 			       struct iommu_cmd *cmd)
930 {
931 	u8 *target;
932 	u32 tail;
933 
934 	/* Copy command to buffer */
935 	tail = iommu->cmd_buf_tail;
936 	target = iommu->cmd_buf + tail;
937 	memcpy(target, cmd, sizeof(*cmd));
938 
939 	tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
940 	iommu->cmd_buf_tail = tail;
941 
942 	/* Tell the IOMMU about it */
943 	writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
944 }
945 
946 static void build_completion_wait(struct iommu_cmd *cmd,
947 				  struct amd_iommu *iommu,
948 				  u64 data)
949 {
950 	u64 paddr = iommu_virt_to_phys((void *)iommu->cmd_sem);
951 
952 	memset(cmd, 0, sizeof(*cmd));
953 	cmd->data[0] = lower_32_bits(paddr) | CMD_COMPL_WAIT_STORE_MASK;
954 	cmd->data[1] = upper_32_bits(paddr);
955 	cmd->data[2] = lower_32_bits(data);
956 	cmd->data[3] = upper_32_bits(data);
957 	CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
958 }
959 
960 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
961 {
962 	memset(cmd, 0, sizeof(*cmd));
963 	cmd->data[0] = devid;
964 	CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
965 }
966 
967 /*
968  * Builds an invalidation address which is suitable for one page or multiple
969  * pages. Sets the size bit (S) as needed is more than one page is flushed.
970  */
971 static inline u64 build_inv_address(u64 address, size_t size)
972 {
973 	u64 pages, end, msb_diff;
974 
975 	pages = iommu_num_pages(address, size, PAGE_SIZE);
976 
977 	if (pages == 1)
978 		return address & PAGE_MASK;
979 
980 	end = address + size - 1;
981 
982 	/*
983 	 * msb_diff would hold the index of the most significant bit that
984 	 * flipped between the start and end.
985 	 */
986 	msb_diff = fls64(end ^ address) - 1;
987 
988 	/*
989 	 * Bits 63:52 are sign extended. If for some reason bit 51 is different
990 	 * between the start and the end, invalidate everything.
991 	 */
992 	if (unlikely(msb_diff > 51)) {
993 		address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
994 	} else {
995 		/*
996 		 * The msb-bit must be clear on the address. Just set all the
997 		 * lower bits.
998 		 */
999 		address |= (1ull << msb_diff) - 1;
1000 	}
1001 
1002 	/* Clear bits 11:0 */
1003 	address &= PAGE_MASK;
1004 
1005 	/* Set the size bit - we flush more than one 4kb page */
1006 	return address | CMD_INV_IOMMU_PAGES_SIZE_MASK;
1007 }
1008 
1009 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
1010 				  size_t size, u16 domid, int pde)
1011 {
1012 	u64 inv_address = build_inv_address(address, size);
1013 
1014 	memset(cmd, 0, sizeof(*cmd));
1015 	cmd->data[1] |= domid;
1016 	cmd->data[2]  = lower_32_bits(inv_address);
1017 	cmd->data[3]  = upper_32_bits(inv_address);
1018 	CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
1019 	if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
1020 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
1021 }
1022 
1023 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
1024 				  u64 address, size_t size)
1025 {
1026 	u64 inv_address = build_inv_address(address, size);
1027 
1028 	memset(cmd, 0, sizeof(*cmd));
1029 	cmd->data[0]  = devid;
1030 	cmd->data[0] |= (qdep & 0xff) << 24;
1031 	cmd->data[1]  = devid;
1032 	cmd->data[2]  = lower_32_bits(inv_address);
1033 	cmd->data[3]  = upper_32_bits(inv_address);
1034 	CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
1035 }
1036 
1037 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, u32 pasid,
1038 				  u64 address, bool size)
1039 {
1040 	memset(cmd, 0, sizeof(*cmd));
1041 
1042 	address &= ~(0xfffULL);
1043 
1044 	cmd->data[0]  = pasid;
1045 	cmd->data[1]  = domid;
1046 	cmd->data[2]  = lower_32_bits(address);
1047 	cmd->data[3]  = upper_32_bits(address);
1048 	cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
1049 	cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
1050 	if (size)
1051 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
1052 	CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
1053 }
1054 
1055 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, u32 pasid,
1056 				  int qdep, u64 address, bool size)
1057 {
1058 	memset(cmd, 0, sizeof(*cmd));
1059 
1060 	address &= ~(0xfffULL);
1061 
1062 	cmd->data[0]  = devid;
1063 	cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
1064 	cmd->data[0] |= (qdep  & 0xff) << 24;
1065 	cmd->data[1]  = devid;
1066 	cmd->data[1] |= (pasid & 0xff) << 16;
1067 	cmd->data[2]  = lower_32_bits(address);
1068 	cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
1069 	cmd->data[3]  = upper_32_bits(address);
1070 	if (size)
1071 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
1072 	CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
1073 }
1074 
1075 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, u32 pasid,
1076 			       int status, int tag, bool gn)
1077 {
1078 	memset(cmd, 0, sizeof(*cmd));
1079 
1080 	cmd->data[0]  = devid;
1081 	if (gn) {
1082 		cmd->data[1]  = pasid;
1083 		cmd->data[2]  = CMD_INV_IOMMU_PAGES_GN_MASK;
1084 	}
1085 	cmd->data[3]  = tag & 0x1ff;
1086 	cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
1087 
1088 	CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
1089 }
1090 
1091 static void build_inv_all(struct iommu_cmd *cmd)
1092 {
1093 	memset(cmd, 0, sizeof(*cmd));
1094 	CMD_SET_TYPE(cmd, CMD_INV_ALL);
1095 }
1096 
1097 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1098 {
1099 	memset(cmd, 0, sizeof(*cmd));
1100 	cmd->data[0] = devid;
1101 	CMD_SET_TYPE(cmd, CMD_INV_IRT);
1102 }
1103 
1104 /*
1105  * Writes the command to the IOMMUs command buffer and informs the
1106  * hardware about the new command.
1107  */
1108 static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1109 				      struct iommu_cmd *cmd,
1110 				      bool sync)
1111 {
1112 	unsigned int count = 0;
1113 	u32 left, next_tail;
1114 
1115 	next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1116 again:
1117 	left      = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE;
1118 
1119 	if (left <= 0x20) {
1120 		/* Skip udelay() the first time around */
1121 		if (count++) {
1122 			if (count == LOOP_TIMEOUT) {
1123 				pr_err("Command buffer timeout\n");
1124 				return -EIO;
1125 			}
1126 
1127 			udelay(1);
1128 		}
1129 
1130 		/* Update head and recheck remaining space */
1131 		iommu->cmd_buf_head = readl(iommu->mmio_base +
1132 					    MMIO_CMD_HEAD_OFFSET);
1133 
1134 		goto again;
1135 	}
1136 
1137 	copy_cmd_to_buffer(iommu, cmd);
1138 
1139 	/* Do we need to make sure all commands are processed? */
1140 	iommu->need_sync = sync;
1141 
1142 	return 0;
1143 }
1144 
1145 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1146 				    struct iommu_cmd *cmd,
1147 				    bool sync)
1148 {
1149 	unsigned long flags;
1150 	int ret;
1151 
1152 	raw_spin_lock_irqsave(&iommu->lock, flags);
1153 	ret = __iommu_queue_command_sync(iommu, cmd, sync);
1154 	raw_spin_unlock_irqrestore(&iommu->lock, flags);
1155 
1156 	return ret;
1157 }
1158 
1159 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1160 {
1161 	return iommu_queue_command_sync(iommu, cmd, true);
1162 }
1163 
1164 /*
1165  * This function queues a completion wait command into the command
1166  * buffer of an IOMMU
1167  */
1168 static int iommu_completion_wait(struct amd_iommu *iommu)
1169 {
1170 	struct iommu_cmd cmd;
1171 	unsigned long flags;
1172 	int ret;
1173 	u64 data;
1174 
1175 	if (!iommu->need_sync)
1176 		return 0;
1177 
1178 	raw_spin_lock_irqsave(&iommu->lock, flags);
1179 
1180 	data = ++iommu->cmd_sem_val;
1181 	build_completion_wait(&cmd, iommu, data);
1182 
1183 	ret = __iommu_queue_command_sync(iommu, &cmd, false);
1184 	if (ret)
1185 		goto out_unlock;
1186 
1187 	ret = wait_on_sem(iommu, data);
1188 
1189 out_unlock:
1190 	raw_spin_unlock_irqrestore(&iommu->lock, flags);
1191 
1192 	return ret;
1193 }
1194 
1195 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1196 {
1197 	struct iommu_cmd cmd;
1198 
1199 	build_inv_dte(&cmd, devid);
1200 
1201 	return iommu_queue_command(iommu, &cmd);
1202 }
1203 
1204 static void amd_iommu_flush_dte_all(struct amd_iommu *iommu)
1205 {
1206 	u32 devid;
1207 	u16 last_bdf = iommu->pci_seg->last_bdf;
1208 
1209 	for (devid = 0; devid <= last_bdf; ++devid)
1210 		iommu_flush_dte(iommu, devid);
1211 
1212 	iommu_completion_wait(iommu);
1213 }
1214 
1215 /*
1216  * This function uses heavy locking and may disable irqs for some time. But
1217  * this is no issue because it is only called during resume.
1218  */
1219 static void amd_iommu_flush_tlb_all(struct amd_iommu *iommu)
1220 {
1221 	u32 dom_id;
1222 	u16 last_bdf = iommu->pci_seg->last_bdf;
1223 
1224 	for (dom_id = 0; dom_id <= last_bdf; ++dom_id) {
1225 		struct iommu_cmd cmd;
1226 		build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1227 				      dom_id, 1);
1228 		iommu_queue_command(iommu, &cmd);
1229 	}
1230 
1231 	iommu_completion_wait(iommu);
1232 }
1233 
1234 static void amd_iommu_flush_tlb_domid(struct amd_iommu *iommu, u32 dom_id)
1235 {
1236 	struct iommu_cmd cmd;
1237 
1238 	build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1239 			      dom_id, 1);
1240 	iommu_queue_command(iommu, &cmd);
1241 
1242 	iommu_completion_wait(iommu);
1243 }
1244 
1245 static void amd_iommu_flush_all(struct amd_iommu *iommu)
1246 {
1247 	struct iommu_cmd cmd;
1248 
1249 	build_inv_all(&cmd);
1250 
1251 	iommu_queue_command(iommu, &cmd);
1252 	iommu_completion_wait(iommu);
1253 }
1254 
1255 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1256 {
1257 	struct iommu_cmd cmd;
1258 
1259 	build_inv_irt(&cmd, devid);
1260 
1261 	iommu_queue_command(iommu, &cmd);
1262 }
1263 
1264 static void amd_iommu_flush_irt_all(struct amd_iommu *iommu)
1265 {
1266 	u32 devid;
1267 	u16 last_bdf = iommu->pci_seg->last_bdf;
1268 
1269 	for (devid = 0; devid <= last_bdf; devid++)
1270 		iommu_flush_irt(iommu, devid);
1271 
1272 	iommu_completion_wait(iommu);
1273 }
1274 
1275 void iommu_flush_all_caches(struct amd_iommu *iommu)
1276 {
1277 	if (iommu_feature(iommu, FEATURE_IA)) {
1278 		amd_iommu_flush_all(iommu);
1279 	} else {
1280 		amd_iommu_flush_dte_all(iommu);
1281 		amd_iommu_flush_irt_all(iommu);
1282 		amd_iommu_flush_tlb_all(iommu);
1283 	}
1284 }
1285 
1286 /*
1287  * Command send function for flushing on-device TLB
1288  */
1289 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1290 			      u64 address, size_t size)
1291 {
1292 	struct amd_iommu *iommu;
1293 	struct iommu_cmd cmd;
1294 	int qdep;
1295 
1296 	qdep     = dev_data->ats.qdep;
1297 	iommu    = rlookup_amd_iommu(dev_data->dev);
1298 	if (!iommu)
1299 		return -EINVAL;
1300 
1301 	build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1302 
1303 	return iommu_queue_command(iommu, &cmd);
1304 }
1305 
1306 static int device_flush_dte_alias(struct pci_dev *pdev, u16 alias, void *data)
1307 {
1308 	struct amd_iommu *iommu = data;
1309 
1310 	return iommu_flush_dte(iommu, alias);
1311 }
1312 
1313 /*
1314  * Command send function for invalidating a device table entry
1315  */
1316 static int device_flush_dte(struct iommu_dev_data *dev_data)
1317 {
1318 	struct amd_iommu *iommu;
1319 	struct pci_dev *pdev = NULL;
1320 	struct amd_iommu_pci_seg *pci_seg;
1321 	u16 alias;
1322 	int ret;
1323 
1324 	iommu = rlookup_amd_iommu(dev_data->dev);
1325 	if (!iommu)
1326 		return -EINVAL;
1327 
1328 	if (dev_is_pci(dev_data->dev))
1329 		pdev = to_pci_dev(dev_data->dev);
1330 
1331 	if (pdev)
1332 		ret = pci_for_each_dma_alias(pdev,
1333 					     device_flush_dte_alias, iommu);
1334 	else
1335 		ret = iommu_flush_dte(iommu, dev_data->devid);
1336 	if (ret)
1337 		return ret;
1338 
1339 	pci_seg = iommu->pci_seg;
1340 	alias = pci_seg->alias_table[dev_data->devid];
1341 	if (alias != dev_data->devid) {
1342 		ret = iommu_flush_dte(iommu, alias);
1343 		if (ret)
1344 			return ret;
1345 	}
1346 
1347 	if (dev_data->ats.enabled)
1348 		ret = device_flush_iotlb(dev_data, 0, ~0UL);
1349 
1350 	return ret;
1351 }
1352 
1353 /*
1354  * TLB invalidation function which is called from the mapping functions.
1355  * It invalidates a single PTE if the range to flush is within a single
1356  * page. Otherwise it flushes the whole TLB of the IOMMU.
1357  */
1358 static void __domain_flush_pages(struct protection_domain *domain,
1359 				 u64 address, size_t size, int pde)
1360 {
1361 	struct iommu_dev_data *dev_data;
1362 	struct iommu_cmd cmd;
1363 	int ret = 0, i;
1364 
1365 	build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1366 
1367 	for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1368 		if (!domain->dev_iommu[i])
1369 			continue;
1370 
1371 		/*
1372 		 * Devices of this domain are behind this IOMMU
1373 		 * We need a TLB flush
1374 		 */
1375 		ret |= iommu_queue_command(amd_iommus[i], &cmd);
1376 	}
1377 
1378 	list_for_each_entry(dev_data, &domain->dev_list, list) {
1379 
1380 		if (!dev_data->ats.enabled)
1381 			continue;
1382 
1383 		ret |= device_flush_iotlb(dev_data, address, size);
1384 	}
1385 
1386 	WARN_ON(ret);
1387 }
1388 
1389 static void domain_flush_pages(struct protection_domain *domain,
1390 			       u64 address, size_t size, int pde)
1391 {
1392 	if (likely(!amd_iommu_np_cache)) {
1393 		__domain_flush_pages(domain, address, size, pde);
1394 		return;
1395 	}
1396 
1397 	/*
1398 	 * When NpCache is on, we infer that we run in a VM and use a vIOMMU.
1399 	 * In such setups it is best to avoid flushes of ranges which are not
1400 	 * naturally aligned, since it would lead to flushes of unmodified
1401 	 * PTEs. Such flushes would require the hypervisor to do more work than
1402 	 * necessary. Therefore, perform repeated flushes of aligned ranges
1403 	 * until you cover the range. Each iteration flushes the smaller
1404 	 * between the natural alignment of the address that we flush and the
1405 	 * greatest naturally aligned region that fits in the range.
1406 	 */
1407 	while (size != 0) {
1408 		int addr_alignment = __ffs(address);
1409 		int size_alignment = __fls(size);
1410 		int min_alignment;
1411 		size_t flush_size;
1412 
1413 		/*
1414 		 * size is always non-zero, but address might be zero, causing
1415 		 * addr_alignment to be negative. As the casting of the
1416 		 * argument in __ffs(address) to long might trim the high bits
1417 		 * of the address on x86-32, cast to long when doing the check.
1418 		 */
1419 		if (likely((unsigned long)address != 0))
1420 			min_alignment = min(addr_alignment, size_alignment);
1421 		else
1422 			min_alignment = size_alignment;
1423 
1424 		flush_size = 1ul << min_alignment;
1425 
1426 		__domain_flush_pages(domain, address, flush_size, pde);
1427 		address += flush_size;
1428 		size -= flush_size;
1429 	}
1430 }
1431 
1432 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1433 void amd_iommu_domain_flush_tlb_pde(struct protection_domain *domain)
1434 {
1435 	domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1436 }
1437 
1438 void amd_iommu_domain_flush_complete(struct protection_domain *domain)
1439 {
1440 	int i;
1441 
1442 	for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1443 		if (domain && !domain->dev_iommu[i])
1444 			continue;
1445 
1446 		/*
1447 		 * Devices of this domain are behind this IOMMU
1448 		 * We need to wait for completion of all commands.
1449 		 */
1450 		iommu_completion_wait(amd_iommus[i]);
1451 	}
1452 }
1453 
1454 /* Flush the not present cache if it exists */
1455 static void domain_flush_np_cache(struct protection_domain *domain,
1456 		dma_addr_t iova, size_t size)
1457 {
1458 	if (unlikely(amd_iommu_np_cache)) {
1459 		unsigned long flags;
1460 
1461 		spin_lock_irqsave(&domain->lock, flags);
1462 		domain_flush_pages(domain, iova, size, 1);
1463 		amd_iommu_domain_flush_complete(domain);
1464 		spin_unlock_irqrestore(&domain->lock, flags);
1465 	}
1466 }
1467 
1468 
1469 /*
1470  * This function flushes the DTEs for all devices in domain
1471  */
1472 static void domain_flush_devices(struct protection_domain *domain)
1473 {
1474 	struct iommu_dev_data *dev_data;
1475 
1476 	list_for_each_entry(dev_data, &domain->dev_list, list)
1477 		device_flush_dte(dev_data);
1478 }
1479 
1480 /****************************************************************************
1481  *
1482  * The next functions belong to the domain allocation. A domain is
1483  * allocated for every IOMMU as the default domain. If device isolation
1484  * is enabled, every device get its own domain. The most important thing
1485  * about domains is the page table mapping the DMA address space they
1486  * contain.
1487  *
1488  ****************************************************************************/
1489 
1490 static u16 domain_id_alloc(void)
1491 {
1492 	int id;
1493 
1494 	spin_lock(&pd_bitmap_lock);
1495 	id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1496 	BUG_ON(id == 0);
1497 	if (id > 0 && id < MAX_DOMAIN_ID)
1498 		__set_bit(id, amd_iommu_pd_alloc_bitmap);
1499 	else
1500 		id = 0;
1501 	spin_unlock(&pd_bitmap_lock);
1502 
1503 	return id;
1504 }
1505 
1506 static void domain_id_free(int id)
1507 {
1508 	spin_lock(&pd_bitmap_lock);
1509 	if (id > 0 && id < MAX_DOMAIN_ID)
1510 		__clear_bit(id, amd_iommu_pd_alloc_bitmap);
1511 	spin_unlock(&pd_bitmap_lock);
1512 }
1513 
1514 static void free_gcr3_tbl_level1(u64 *tbl)
1515 {
1516 	u64 *ptr;
1517 	int i;
1518 
1519 	for (i = 0; i < 512; ++i) {
1520 		if (!(tbl[i] & GCR3_VALID))
1521 			continue;
1522 
1523 		ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1524 
1525 		free_page((unsigned long)ptr);
1526 	}
1527 }
1528 
1529 static void free_gcr3_tbl_level2(u64 *tbl)
1530 {
1531 	u64 *ptr;
1532 	int i;
1533 
1534 	for (i = 0; i < 512; ++i) {
1535 		if (!(tbl[i] & GCR3_VALID))
1536 			continue;
1537 
1538 		ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1539 
1540 		free_gcr3_tbl_level1(ptr);
1541 	}
1542 }
1543 
1544 static void free_gcr3_table(struct protection_domain *domain)
1545 {
1546 	if (domain->glx == 2)
1547 		free_gcr3_tbl_level2(domain->gcr3_tbl);
1548 	else if (domain->glx == 1)
1549 		free_gcr3_tbl_level1(domain->gcr3_tbl);
1550 	else
1551 		BUG_ON(domain->glx != 0);
1552 
1553 	free_page((unsigned long)domain->gcr3_tbl);
1554 }
1555 
1556 static void set_dte_entry(struct amd_iommu *iommu, u16 devid,
1557 			  struct protection_domain *domain, bool ats, bool ppr)
1558 {
1559 	u64 pte_root = 0;
1560 	u64 flags = 0;
1561 	u32 old_domid;
1562 	struct dev_table_entry *dev_table = get_dev_table(iommu);
1563 
1564 	if (domain->iop.mode != PAGE_MODE_NONE)
1565 		pte_root = iommu_virt_to_phys(domain->iop.root);
1566 
1567 	pte_root |= (domain->iop.mode & DEV_ENTRY_MODE_MASK)
1568 		    << DEV_ENTRY_MODE_SHIFT;
1569 
1570 	pte_root |= DTE_FLAG_IR | DTE_FLAG_IW | DTE_FLAG_V;
1571 
1572 	/*
1573 	 * When SNP is enabled, Only set TV bit when IOMMU
1574 	 * page translation is in use.
1575 	 */
1576 	if (!amd_iommu_snp_en || (domain->id != 0))
1577 		pte_root |= DTE_FLAG_TV;
1578 
1579 	flags = dev_table[devid].data[1];
1580 
1581 	if (ats)
1582 		flags |= DTE_FLAG_IOTLB;
1583 
1584 	if (ppr) {
1585 		if (iommu_feature(iommu, FEATURE_EPHSUP))
1586 			pte_root |= 1ULL << DEV_ENTRY_PPR;
1587 	}
1588 
1589 	if (domain->flags & PD_IOMMUV2_MASK) {
1590 		u64 gcr3 = iommu_virt_to_phys(domain->gcr3_tbl);
1591 		u64 glx  = domain->glx;
1592 		u64 tmp;
1593 
1594 		pte_root |= DTE_FLAG_GV;
1595 		pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1596 
1597 		/* First mask out possible old values for GCR3 table */
1598 		tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1599 		flags    &= ~tmp;
1600 
1601 		tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1602 		flags    &= ~tmp;
1603 
1604 		/* Encode GCR3 table into DTE */
1605 		tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1606 		pte_root |= tmp;
1607 
1608 		tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1609 		flags    |= tmp;
1610 
1611 		tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1612 		flags    |= tmp;
1613 
1614 		if (domain->flags & PD_GIOV_MASK)
1615 			pte_root |= DTE_FLAG_GIOV;
1616 	}
1617 
1618 	flags &= ~DEV_DOMID_MASK;
1619 	flags |= domain->id;
1620 
1621 	old_domid = dev_table[devid].data[1] & DEV_DOMID_MASK;
1622 	dev_table[devid].data[1]  = flags;
1623 	dev_table[devid].data[0]  = pte_root;
1624 
1625 	/*
1626 	 * A kdump kernel might be replacing a domain ID that was copied from
1627 	 * the previous kernel--if so, it needs to flush the translation cache
1628 	 * entries for the old domain ID that is being overwritten
1629 	 */
1630 	if (old_domid) {
1631 		amd_iommu_flush_tlb_domid(iommu, old_domid);
1632 	}
1633 }
1634 
1635 static void clear_dte_entry(struct amd_iommu *iommu, u16 devid)
1636 {
1637 	struct dev_table_entry *dev_table = get_dev_table(iommu);
1638 
1639 	/* remove entry from the device table seen by the hardware */
1640 	dev_table[devid].data[0]  = DTE_FLAG_V;
1641 
1642 	if (!amd_iommu_snp_en)
1643 		dev_table[devid].data[0] |= DTE_FLAG_TV;
1644 
1645 	dev_table[devid].data[1] &= DTE_FLAG_MASK;
1646 
1647 	amd_iommu_apply_erratum_63(iommu, devid);
1648 }
1649 
1650 static void do_attach(struct iommu_dev_data *dev_data,
1651 		      struct protection_domain *domain)
1652 {
1653 	struct amd_iommu *iommu;
1654 	bool ats;
1655 
1656 	iommu = rlookup_amd_iommu(dev_data->dev);
1657 	if (!iommu)
1658 		return;
1659 	ats   = dev_data->ats.enabled;
1660 
1661 	/* Update data structures */
1662 	dev_data->domain = domain;
1663 	list_add(&dev_data->list, &domain->dev_list);
1664 
1665 	/* Do reference counting */
1666 	domain->dev_iommu[iommu->index] += 1;
1667 	domain->dev_cnt                 += 1;
1668 
1669 	/* Override supported page sizes */
1670 	if (domain->flags & PD_GIOV_MASK)
1671 		domain->domain.pgsize_bitmap = AMD_IOMMU_PGSIZES_V2;
1672 
1673 	/* Update device table */
1674 	set_dte_entry(iommu, dev_data->devid, domain,
1675 		      ats, dev_data->iommu_v2);
1676 	clone_aliases(iommu, dev_data->dev);
1677 
1678 	device_flush_dte(dev_data);
1679 }
1680 
1681 static void do_detach(struct iommu_dev_data *dev_data)
1682 {
1683 	struct protection_domain *domain = dev_data->domain;
1684 	struct amd_iommu *iommu;
1685 
1686 	iommu = rlookup_amd_iommu(dev_data->dev);
1687 	if (!iommu)
1688 		return;
1689 
1690 	/* Update data structures */
1691 	dev_data->domain = NULL;
1692 	list_del(&dev_data->list);
1693 	clear_dte_entry(iommu, dev_data->devid);
1694 	clone_aliases(iommu, dev_data->dev);
1695 
1696 	/* Flush the DTE entry */
1697 	device_flush_dte(dev_data);
1698 
1699 	/* Flush IOTLB */
1700 	amd_iommu_domain_flush_tlb_pde(domain);
1701 
1702 	/* Wait for the flushes to finish */
1703 	amd_iommu_domain_flush_complete(domain);
1704 
1705 	/* decrease reference counters - needs to happen after the flushes */
1706 	domain->dev_iommu[iommu->index] -= 1;
1707 	domain->dev_cnt                 -= 1;
1708 }
1709 
1710 static void pdev_iommuv2_disable(struct pci_dev *pdev)
1711 {
1712 	pci_disable_ats(pdev);
1713 	pci_disable_pri(pdev);
1714 	pci_disable_pasid(pdev);
1715 }
1716 
1717 static int pdev_pri_ats_enable(struct pci_dev *pdev)
1718 {
1719 	int ret;
1720 
1721 	/* Only allow access to user-accessible pages */
1722 	ret = pci_enable_pasid(pdev, 0);
1723 	if (ret)
1724 		return ret;
1725 
1726 	/* First reset the PRI state of the device */
1727 	ret = pci_reset_pri(pdev);
1728 	if (ret)
1729 		goto out_err_pasid;
1730 
1731 	/* Enable PRI */
1732 	/* FIXME: Hardcode number of outstanding requests for now */
1733 	ret = pci_enable_pri(pdev, 32);
1734 	if (ret)
1735 		goto out_err_pasid;
1736 
1737 	ret = pci_enable_ats(pdev, PAGE_SHIFT);
1738 	if (ret)
1739 		goto out_err_pri;
1740 
1741 	return 0;
1742 
1743 out_err_pri:
1744 	pci_disable_pri(pdev);
1745 
1746 out_err_pasid:
1747 	pci_disable_pasid(pdev);
1748 
1749 	return ret;
1750 }
1751 
1752 /*
1753  * If a device is not yet associated with a domain, this function makes the
1754  * device visible in the domain
1755  */
1756 static int attach_device(struct device *dev,
1757 			 struct protection_domain *domain)
1758 {
1759 	struct iommu_dev_data *dev_data;
1760 	struct pci_dev *pdev;
1761 	unsigned long flags;
1762 	int ret;
1763 
1764 	spin_lock_irqsave(&domain->lock, flags);
1765 
1766 	dev_data = dev_iommu_priv_get(dev);
1767 
1768 	spin_lock(&dev_data->lock);
1769 
1770 	ret = -EBUSY;
1771 	if (dev_data->domain != NULL)
1772 		goto out;
1773 
1774 	if (!dev_is_pci(dev))
1775 		goto skip_ats_check;
1776 
1777 	pdev = to_pci_dev(dev);
1778 	if (domain->flags & PD_IOMMUV2_MASK) {
1779 		struct iommu_domain *def_domain = iommu_get_dma_domain(dev);
1780 
1781 		ret = -EINVAL;
1782 
1783 		/*
1784 		 * In case of using AMD_IOMMU_V1 page table mode and the device
1785 		 * is enabling for PPR/ATS support (using v2 table),
1786 		 * we need to make sure that the domain type is identity map.
1787 		 */
1788 		if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
1789 		    def_domain->type != IOMMU_DOMAIN_IDENTITY) {
1790 			goto out;
1791 		}
1792 
1793 		if (dev_data->iommu_v2) {
1794 			if (pdev_pri_ats_enable(pdev) != 0)
1795 				goto out;
1796 
1797 			dev_data->ats.enabled = true;
1798 			dev_data->ats.qdep    = pci_ats_queue_depth(pdev);
1799 			dev_data->pri_tlp     = pci_prg_resp_pasid_required(pdev);
1800 		}
1801 	} else if (amd_iommu_iotlb_sup &&
1802 		   pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
1803 		dev_data->ats.enabled = true;
1804 		dev_data->ats.qdep    = pci_ats_queue_depth(pdev);
1805 	}
1806 
1807 skip_ats_check:
1808 	ret = 0;
1809 
1810 	do_attach(dev_data, domain);
1811 
1812 	/*
1813 	 * We might boot into a crash-kernel here. The crashed kernel
1814 	 * left the caches in the IOMMU dirty. So we have to flush
1815 	 * here to evict all dirty stuff.
1816 	 */
1817 	amd_iommu_domain_flush_tlb_pde(domain);
1818 
1819 	amd_iommu_domain_flush_complete(domain);
1820 
1821 out:
1822 	spin_unlock(&dev_data->lock);
1823 
1824 	spin_unlock_irqrestore(&domain->lock, flags);
1825 
1826 	return ret;
1827 }
1828 
1829 /*
1830  * Removes a device from a protection domain (with devtable_lock held)
1831  */
1832 static void detach_device(struct device *dev)
1833 {
1834 	struct protection_domain *domain;
1835 	struct iommu_dev_data *dev_data;
1836 	unsigned long flags;
1837 
1838 	dev_data = dev_iommu_priv_get(dev);
1839 	domain   = dev_data->domain;
1840 
1841 	spin_lock_irqsave(&domain->lock, flags);
1842 
1843 	spin_lock(&dev_data->lock);
1844 
1845 	/*
1846 	 * First check if the device is still attached. It might already
1847 	 * be detached from its domain because the generic
1848 	 * iommu_detach_group code detached it and we try again here in
1849 	 * our alias handling.
1850 	 */
1851 	if (WARN_ON(!dev_data->domain))
1852 		goto out;
1853 
1854 	do_detach(dev_data);
1855 
1856 	if (!dev_is_pci(dev))
1857 		goto out;
1858 
1859 	if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
1860 		pdev_iommuv2_disable(to_pci_dev(dev));
1861 	else if (dev_data->ats.enabled)
1862 		pci_disable_ats(to_pci_dev(dev));
1863 
1864 	dev_data->ats.enabled = false;
1865 
1866 out:
1867 	spin_unlock(&dev_data->lock);
1868 
1869 	spin_unlock_irqrestore(&domain->lock, flags);
1870 }
1871 
1872 static struct iommu_device *amd_iommu_probe_device(struct device *dev)
1873 {
1874 	struct iommu_device *iommu_dev;
1875 	struct amd_iommu *iommu;
1876 	int ret;
1877 
1878 	if (!check_device(dev))
1879 		return ERR_PTR(-ENODEV);
1880 
1881 	iommu = rlookup_amd_iommu(dev);
1882 	if (!iommu)
1883 		return ERR_PTR(-ENODEV);
1884 
1885 	/* Not registered yet? */
1886 	if (!iommu->iommu.ops)
1887 		return ERR_PTR(-ENODEV);
1888 
1889 	if (dev_iommu_priv_get(dev))
1890 		return &iommu->iommu;
1891 
1892 	ret = iommu_init_device(iommu, dev);
1893 	if (ret) {
1894 		if (ret != -ENOTSUPP)
1895 			dev_err(dev, "Failed to initialize - trying to proceed anyway\n");
1896 		iommu_dev = ERR_PTR(ret);
1897 		iommu_ignore_device(iommu, dev);
1898 	} else {
1899 		amd_iommu_set_pci_msi_domain(dev, iommu);
1900 		iommu_dev = &iommu->iommu;
1901 	}
1902 
1903 	iommu_completion_wait(iommu);
1904 
1905 	return iommu_dev;
1906 }
1907 
1908 static void amd_iommu_probe_finalize(struct device *dev)
1909 {
1910 	/* Domains are initialized for this device - have a look what we ended up with */
1911 	set_dma_ops(dev, NULL);
1912 	iommu_setup_dma_ops(dev, 0, U64_MAX);
1913 }
1914 
1915 static void amd_iommu_release_device(struct device *dev)
1916 {
1917 	struct amd_iommu *iommu;
1918 
1919 	if (!check_device(dev))
1920 		return;
1921 
1922 	iommu = rlookup_amd_iommu(dev);
1923 	if (!iommu)
1924 		return;
1925 
1926 	amd_iommu_uninit_device(dev);
1927 	iommu_completion_wait(iommu);
1928 }
1929 
1930 static struct iommu_group *amd_iommu_device_group(struct device *dev)
1931 {
1932 	if (dev_is_pci(dev))
1933 		return pci_device_group(dev);
1934 
1935 	return acpihid_device_group(dev);
1936 }
1937 
1938 /*****************************************************************************
1939  *
1940  * The next functions belong to the dma_ops mapping/unmapping code.
1941  *
1942  *****************************************************************************/
1943 
1944 static void update_device_table(struct protection_domain *domain)
1945 {
1946 	struct iommu_dev_data *dev_data;
1947 
1948 	list_for_each_entry(dev_data, &domain->dev_list, list) {
1949 		struct amd_iommu *iommu = rlookup_amd_iommu(dev_data->dev);
1950 
1951 		if (!iommu)
1952 			continue;
1953 		set_dte_entry(iommu, dev_data->devid, domain,
1954 			      dev_data->ats.enabled, dev_data->iommu_v2);
1955 		clone_aliases(iommu, dev_data->dev);
1956 	}
1957 }
1958 
1959 void amd_iommu_update_and_flush_device_table(struct protection_domain *domain)
1960 {
1961 	update_device_table(domain);
1962 	domain_flush_devices(domain);
1963 }
1964 
1965 void amd_iommu_domain_update(struct protection_domain *domain)
1966 {
1967 	/* Update device table */
1968 	amd_iommu_update_and_flush_device_table(domain);
1969 
1970 	/* Flush domain TLB(s) and wait for completion */
1971 	amd_iommu_domain_flush_tlb_pde(domain);
1972 	amd_iommu_domain_flush_complete(domain);
1973 }
1974 
1975 /*****************************************************************************
1976  *
1977  * The following functions belong to the exported interface of AMD IOMMU
1978  *
1979  * This interface allows access to lower level functions of the IOMMU
1980  * like protection domain handling and assignement of devices to domains
1981  * which is not possible with the dma_ops interface.
1982  *
1983  *****************************************************************************/
1984 
1985 static void cleanup_domain(struct protection_domain *domain)
1986 {
1987 	struct iommu_dev_data *entry;
1988 	unsigned long flags;
1989 
1990 	spin_lock_irqsave(&domain->lock, flags);
1991 
1992 	while (!list_empty(&domain->dev_list)) {
1993 		entry = list_first_entry(&domain->dev_list,
1994 					 struct iommu_dev_data, list);
1995 		BUG_ON(!entry->domain);
1996 		do_detach(entry);
1997 	}
1998 
1999 	spin_unlock_irqrestore(&domain->lock, flags);
2000 }
2001 
2002 static void protection_domain_free(struct protection_domain *domain)
2003 {
2004 	if (!domain)
2005 		return;
2006 
2007 	if (domain->iop.pgtbl_cfg.tlb)
2008 		free_io_pgtable_ops(&domain->iop.iop.ops);
2009 
2010 	if (domain->id)
2011 		domain_id_free(domain->id);
2012 
2013 	kfree(domain);
2014 }
2015 
2016 static int protection_domain_init_v1(struct protection_domain *domain, int mode)
2017 {
2018 	u64 *pt_root = NULL;
2019 
2020 	BUG_ON(mode < PAGE_MODE_NONE || mode > PAGE_MODE_6_LEVEL);
2021 
2022 	spin_lock_init(&domain->lock);
2023 	domain->id = domain_id_alloc();
2024 	if (!domain->id)
2025 		return -ENOMEM;
2026 	INIT_LIST_HEAD(&domain->dev_list);
2027 
2028 	if (mode != PAGE_MODE_NONE) {
2029 		pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2030 		if (!pt_root) {
2031 			domain_id_free(domain->id);
2032 			return -ENOMEM;
2033 		}
2034 	}
2035 
2036 	amd_iommu_domain_set_pgtable(domain, pt_root, mode);
2037 
2038 	return 0;
2039 }
2040 
2041 static int protection_domain_init_v2(struct protection_domain *domain)
2042 {
2043 	spin_lock_init(&domain->lock);
2044 	domain->id = domain_id_alloc();
2045 	if (!domain->id)
2046 		return -ENOMEM;
2047 	INIT_LIST_HEAD(&domain->dev_list);
2048 
2049 	domain->flags |= PD_GIOV_MASK;
2050 
2051 	if (domain_enable_v2(domain, 1)) {
2052 		domain_id_free(domain->id);
2053 		return -ENOMEM;
2054 	}
2055 
2056 	return 0;
2057 }
2058 
2059 static struct protection_domain *protection_domain_alloc(unsigned int type)
2060 {
2061 	struct io_pgtable_ops *pgtbl_ops;
2062 	struct protection_domain *domain;
2063 	int pgtable = amd_iommu_pgtable;
2064 	int mode = DEFAULT_PGTABLE_LEVEL;
2065 	int ret;
2066 
2067 	domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2068 	if (!domain)
2069 		return NULL;
2070 
2071 	/*
2072 	 * Force IOMMU v1 page table when iommu=pt and
2073 	 * when allocating domain for pass-through devices.
2074 	 */
2075 	if (type == IOMMU_DOMAIN_IDENTITY) {
2076 		pgtable = AMD_IOMMU_V1;
2077 		mode = PAGE_MODE_NONE;
2078 	} else if (type == IOMMU_DOMAIN_UNMANAGED) {
2079 		pgtable = AMD_IOMMU_V1;
2080 	}
2081 
2082 	switch (pgtable) {
2083 	case AMD_IOMMU_V1:
2084 		ret = protection_domain_init_v1(domain, mode);
2085 		break;
2086 	case AMD_IOMMU_V2:
2087 		ret = protection_domain_init_v2(domain);
2088 		break;
2089 	default:
2090 		ret = -EINVAL;
2091 	}
2092 
2093 	if (ret)
2094 		goto out_err;
2095 
2096 	/* No need to allocate io pgtable ops in passthrough mode */
2097 	if (type == IOMMU_DOMAIN_IDENTITY)
2098 		return domain;
2099 
2100 	pgtbl_ops = alloc_io_pgtable_ops(pgtable, &domain->iop.pgtbl_cfg, domain);
2101 	if (!pgtbl_ops) {
2102 		domain_id_free(domain->id);
2103 		goto out_err;
2104 	}
2105 
2106 	return domain;
2107 out_err:
2108 	kfree(domain);
2109 	return NULL;
2110 }
2111 
2112 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2113 {
2114 	struct protection_domain *domain;
2115 
2116 	/*
2117 	 * Since DTE[Mode]=0 is prohibited on SNP-enabled system,
2118 	 * default to use IOMMU_DOMAIN_DMA[_FQ].
2119 	 */
2120 	if (amd_iommu_snp_en && (type == IOMMU_DOMAIN_IDENTITY))
2121 		return NULL;
2122 
2123 	domain = protection_domain_alloc(type);
2124 	if (!domain)
2125 		return NULL;
2126 
2127 	domain->domain.geometry.aperture_start = 0;
2128 	domain->domain.geometry.aperture_end   = ~0ULL;
2129 	domain->domain.geometry.force_aperture = true;
2130 
2131 	return &domain->domain;
2132 }
2133 
2134 static void amd_iommu_domain_free(struct iommu_domain *dom)
2135 {
2136 	struct protection_domain *domain;
2137 
2138 	domain = to_pdomain(dom);
2139 
2140 	if (domain->dev_cnt > 0)
2141 		cleanup_domain(domain);
2142 
2143 	BUG_ON(domain->dev_cnt != 0);
2144 
2145 	if (!dom)
2146 		return;
2147 
2148 	if (domain->flags & PD_IOMMUV2_MASK)
2149 		free_gcr3_table(domain);
2150 
2151 	protection_domain_free(domain);
2152 }
2153 
2154 static int amd_iommu_attach_device(struct iommu_domain *dom,
2155 				   struct device *dev)
2156 {
2157 	struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2158 	struct protection_domain *domain = to_pdomain(dom);
2159 	struct amd_iommu *iommu = rlookup_amd_iommu(dev);
2160 	int ret;
2161 
2162 	/*
2163 	 * Skip attach device to domain if new domain is same as
2164 	 * devices current domain
2165 	 */
2166 	if (dev_data->domain == domain)
2167 		return 0;
2168 
2169 	dev_data->defer_attach = false;
2170 
2171 	if (dev_data->domain)
2172 		detach_device(dev);
2173 
2174 	ret = attach_device(dev, domain);
2175 
2176 #ifdef CONFIG_IRQ_REMAP
2177 	if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
2178 		if (dom->type == IOMMU_DOMAIN_UNMANAGED)
2179 			dev_data->use_vapic = 1;
2180 		else
2181 			dev_data->use_vapic = 0;
2182 	}
2183 #endif
2184 
2185 	iommu_completion_wait(iommu);
2186 
2187 	return ret;
2188 }
2189 
2190 static void amd_iommu_iotlb_sync_map(struct iommu_domain *dom,
2191 				     unsigned long iova, size_t size)
2192 {
2193 	struct protection_domain *domain = to_pdomain(dom);
2194 	struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2195 
2196 	if (ops->map_pages)
2197 		domain_flush_np_cache(domain, iova, size);
2198 }
2199 
2200 static int amd_iommu_map_pages(struct iommu_domain *dom, unsigned long iova,
2201 			       phys_addr_t paddr, size_t pgsize, size_t pgcount,
2202 			       int iommu_prot, gfp_t gfp, size_t *mapped)
2203 {
2204 	struct protection_domain *domain = to_pdomain(dom);
2205 	struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2206 	int prot = 0;
2207 	int ret = -EINVAL;
2208 
2209 	if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
2210 	    (domain->iop.mode == PAGE_MODE_NONE))
2211 		return -EINVAL;
2212 
2213 	if (iommu_prot & IOMMU_READ)
2214 		prot |= IOMMU_PROT_IR;
2215 	if (iommu_prot & IOMMU_WRITE)
2216 		prot |= IOMMU_PROT_IW;
2217 
2218 	if (ops->map_pages) {
2219 		ret = ops->map_pages(ops, iova, paddr, pgsize,
2220 				     pgcount, prot, gfp, mapped);
2221 	}
2222 
2223 	return ret;
2224 }
2225 
2226 static void amd_iommu_iotlb_gather_add_page(struct iommu_domain *domain,
2227 					    struct iommu_iotlb_gather *gather,
2228 					    unsigned long iova, size_t size)
2229 {
2230 	/*
2231 	 * AMD's IOMMU can flush as many pages as necessary in a single flush.
2232 	 * Unless we run in a virtual machine, which can be inferred according
2233 	 * to whether "non-present cache" is on, it is probably best to prefer
2234 	 * (potentially) too extensive TLB flushing (i.e., more misses) over
2235 	 * mutliple TLB flushes (i.e., more flushes). For virtual machines the
2236 	 * hypervisor needs to synchronize the host IOMMU PTEs with those of
2237 	 * the guest, and the trade-off is different: unnecessary TLB flushes
2238 	 * should be avoided.
2239 	 */
2240 	if (amd_iommu_np_cache &&
2241 	    iommu_iotlb_gather_is_disjoint(gather, iova, size))
2242 		iommu_iotlb_sync(domain, gather);
2243 
2244 	iommu_iotlb_gather_add_range(gather, iova, size);
2245 }
2246 
2247 static size_t amd_iommu_unmap_pages(struct iommu_domain *dom, unsigned long iova,
2248 				    size_t pgsize, size_t pgcount,
2249 				    struct iommu_iotlb_gather *gather)
2250 {
2251 	struct protection_domain *domain = to_pdomain(dom);
2252 	struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2253 	size_t r;
2254 
2255 	if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
2256 	    (domain->iop.mode == PAGE_MODE_NONE))
2257 		return 0;
2258 
2259 	r = (ops->unmap_pages) ? ops->unmap_pages(ops, iova, pgsize, pgcount, NULL) : 0;
2260 
2261 	if (r)
2262 		amd_iommu_iotlb_gather_add_page(dom, gather, iova, r);
2263 
2264 	return r;
2265 }
2266 
2267 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
2268 					  dma_addr_t iova)
2269 {
2270 	struct protection_domain *domain = to_pdomain(dom);
2271 	struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2272 
2273 	return ops->iova_to_phys(ops, iova);
2274 }
2275 
2276 static bool amd_iommu_capable(struct device *dev, enum iommu_cap cap)
2277 {
2278 	switch (cap) {
2279 	case IOMMU_CAP_CACHE_COHERENCY:
2280 		return true;
2281 	case IOMMU_CAP_NOEXEC:
2282 		return false;
2283 	case IOMMU_CAP_PRE_BOOT_PROTECTION:
2284 		return amdr_ivrs_remap_support;
2285 	case IOMMU_CAP_ENFORCE_CACHE_COHERENCY:
2286 		return true;
2287 	default:
2288 		break;
2289 	}
2290 
2291 	return false;
2292 }
2293 
2294 static void amd_iommu_get_resv_regions(struct device *dev,
2295 				       struct list_head *head)
2296 {
2297 	struct iommu_resv_region *region;
2298 	struct unity_map_entry *entry;
2299 	struct amd_iommu *iommu;
2300 	struct amd_iommu_pci_seg *pci_seg;
2301 	int devid, sbdf;
2302 
2303 	sbdf = get_device_sbdf_id(dev);
2304 	if (sbdf < 0)
2305 		return;
2306 
2307 	devid = PCI_SBDF_TO_DEVID(sbdf);
2308 	iommu = rlookup_amd_iommu(dev);
2309 	if (!iommu)
2310 		return;
2311 	pci_seg = iommu->pci_seg;
2312 
2313 	list_for_each_entry(entry, &pci_seg->unity_map, list) {
2314 		int type, prot = 0;
2315 		size_t length;
2316 
2317 		if (devid < entry->devid_start || devid > entry->devid_end)
2318 			continue;
2319 
2320 		type   = IOMMU_RESV_DIRECT;
2321 		length = entry->address_end - entry->address_start;
2322 		if (entry->prot & IOMMU_PROT_IR)
2323 			prot |= IOMMU_READ;
2324 		if (entry->prot & IOMMU_PROT_IW)
2325 			prot |= IOMMU_WRITE;
2326 		if (entry->prot & IOMMU_UNITY_MAP_FLAG_EXCL_RANGE)
2327 			/* Exclusion range */
2328 			type = IOMMU_RESV_RESERVED;
2329 
2330 		region = iommu_alloc_resv_region(entry->address_start,
2331 						 length, prot, type,
2332 						 GFP_KERNEL);
2333 		if (!region) {
2334 			dev_err(dev, "Out of memory allocating dm-regions\n");
2335 			return;
2336 		}
2337 		list_add_tail(&region->list, head);
2338 	}
2339 
2340 	region = iommu_alloc_resv_region(MSI_RANGE_START,
2341 					 MSI_RANGE_END - MSI_RANGE_START + 1,
2342 					 0, IOMMU_RESV_MSI, GFP_KERNEL);
2343 	if (!region)
2344 		return;
2345 	list_add_tail(&region->list, head);
2346 
2347 	region = iommu_alloc_resv_region(HT_RANGE_START,
2348 					 HT_RANGE_END - HT_RANGE_START + 1,
2349 					 0, IOMMU_RESV_RESERVED, GFP_KERNEL);
2350 	if (!region)
2351 		return;
2352 	list_add_tail(&region->list, head);
2353 }
2354 
2355 bool amd_iommu_is_attach_deferred(struct device *dev)
2356 {
2357 	struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2358 
2359 	return dev_data->defer_attach;
2360 }
2361 EXPORT_SYMBOL_GPL(amd_iommu_is_attach_deferred);
2362 
2363 static void amd_iommu_flush_iotlb_all(struct iommu_domain *domain)
2364 {
2365 	struct protection_domain *dom = to_pdomain(domain);
2366 	unsigned long flags;
2367 
2368 	spin_lock_irqsave(&dom->lock, flags);
2369 	amd_iommu_domain_flush_tlb_pde(dom);
2370 	amd_iommu_domain_flush_complete(dom);
2371 	spin_unlock_irqrestore(&dom->lock, flags);
2372 }
2373 
2374 static void amd_iommu_iotlb_sync(struct iommu_domain *domain,
2375 				 struct iommu_iotlb_gather *gather)
2376 {
2377 	struct protection_domain *dom = to_pdomain(domain);
2378 	unsigned long flags;
2379 
2380 	spin_lock_irqsave(&dom->lock, flags);
2381 	domain_flush_pages(dom, gather->start, gather->end - gather->start, 1);
2382 	amd_iommu_domain_flush_complete(dom);
2383 	spin_unlock_irqrestore(&dom->lock, flags);
2384 }
2385 
2386 static int amd_iommu_def_domain_type(struct device *dev)
2387 {
2388 	struct iommu_dev_data *dev_data;
2389 
2390 	dev_data = dev_iommu_priv_get(dev);
2391 	if (!dev_data)
2392 		return 0;
2393 
2394 	/*
2395 	 * Do not identity map IOMMUv2 capable devices when:
2396 	 *  - memory encryption is active, because some of those devices
2397 	 *    (AMD GPUs) don't have the encryption bit in their DMA-mask
2398 	 *    and require remapping.
2399 	 *  - SNP is enabled, because it prohibits DTE[Mode]=0.
2400 	 */
2401 	if (dev_data->iommu_v2 &&
2402 	    !cc_platform_has(CC_ATTR_MEM_ENCRYPT) &&
2403 	    !amd_iommu_snp_en) {
2404 		return IOMMU_DOMAIN_IDENTITY;
2405 	}
2406 
2407 	return 0;
2408 }
2409 
2410 static bool amd_iommu_enforce_cache_coherency(struct iommu_domain *domain)
2411 {
2412 	/* IOMMU_PTE_FC is always set */
2413 	return true;
2414 }
2415 
2416 const struct iommu_ops amd_iommu_ops = {
2417 	.capable = amd_iommu_capable,
2418 	.domain_alloc = amd_iommu_domain_alloc,
2419 	.probe_device = amd_iommu_probe_device,
2420 	.release_device = amd_iommu_release_device,
2421 	.probe_finalize = amd_iommu_probe_finalize,
2422 	.device_group = amd_iommu_device_group,
2423 	.get_resv_regions = amd_iommu_get_resv_regions,
2424 	.is_attach_deferred = amd_iommu_is_attach_deferred,
2425 	.pgsize_bitmap	= AMD_IOMMU_PGSIZES,
2426 	.def_domain_type = amd_iommu_def_domain_type,
2427 	.default_domain_ops = &(const struct iommu_domain_ops) {
2428 		.attach_dev	= amd_iommu_attach_device,
2429 		.map_pages	= amd_iommu_map_pages,
2430 		.unmap_pages	= amd_iommu_unmap_pages,
2431 		.iotlb_sync_map	= amd_iommu_iotlb_sync_map,
2432 		.iova_to_phys	= amd_iommu_iova_to_phys,
2433 		.flush_iotlb_all = amd_iommu_flush_iotlb_all,
2434 		.iotlb_sync	= amd_iommu_iotlb_sync,
2435 		.free		= amd_iommu_domain_free,
2436 		.enforce_cache_coherency = amd_iommu_enforce_cache_coherency,
2437 	}
2438 };
2439 
2440 /*****************************************************************************
2441  *
2442  * The next functions do a basic initialization of IOMMU for pass through
2443  * mode
2444  *
2445  * In passthrough mode the IOMMU is initialized and enabled but not used for
2446  * DMA-API translation.
2447  *
2448  *****************************************************************************/
2449 
2450 /* IOMMUv2 specific functions */
2451 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
2452 {
2453 	return atomic_notifier_chain_register(&ppr_notifier, nb);
2454 }
2455 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
2456 
2457 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
2458 {
2459 	return atomic_notifier_chain_unregister(&ppr_notifier, nb);
2460 }
2461 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
2462 
2463 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
2464 {
2465 	struct protection_domain *domain = to_pdomain(dom);
2466 	unsigned long flags;
2467 
2468 	spin_lock_irqsave(&domain->lock, flags);
2469 
2470 	if (domain->iop.pgtbl_cfg.tlb)
2471 		free_io_pgtable_ops(&domain->iop.iop.ops);
2472 
2473 	spin_unlock_irqrestore(&domain->lock, flags);
2474 }
2475 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
2476 
2477 /* Note: This function expects iommu_domain->lock to be held prior calling the function. */
2478 static int domain_enable_v2(struct protection_domain *domain, int pasids)
2479 {
2480 	int levels;
2481 
2482 	/* Number of GCR3 table levels required */
2483 	for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
2484 		levels += 1;
2485 
2486 	if (levels > amd_iommu_max_glx_val)
2487 		return -EINVAL;
2488 
2489 	domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
2490 	if (domain->gcr3_tbl == NULL)
2491 		return -ENOMEM;
2492 
2493 	domain->glx      = levels;
2494 	domain->flags   |= PD_IOMMUV2_MASK;
2495 
2496 	amd_iommu_domain_update(domain);
2497 
2498 	return 0;
2499 }
2500 
2501 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
2502 {
2503 	struct protection_domain *pdom = to_pdomain(dom);
2504 	unsigned long flags;
2505 	int ret;
2506 
2507 	spin_lock_irqsave(&pdom->lock, flags);
2508 
2509 	/*
2510 	 * Save us all sanity checks whether devices already in the
2511 	 * domain support IOMMUv2. Just force that the domain has no
2512 	 * devices attached when it is switched into IOMMUv2 mode.
2513 	 */
2514 	ret = -EBUSY;
2515 	if (pdom->dev_cnt > 0 || pdom->flags & PD_IOMMUV2_MASK)
2516 		goto out;
2517 
2518 	if (!pdom->gcr3_tbl)
2519 		ret = domain_enable_v2(pdom, pasids);
2520 
2521 out:
2522 	spin_unlock_irqrestore(&pdom->lock, flags);
2523 	return ret;
2524 }
2525 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
2526 
2527 static int __flush_pasid(struct protection_domain *domain, u32 pasid,
2528 			 u64 address, bool size)
2529 {
2530 	struct iommu_dev_data *dev_data;
2531 	struct iommu_cmd cmd;
2532 	int i, ret;
2533 
2534 	if (!(domain->flags & PD_IOMMUV2_MASK))
2535 		return -EINVAL;
2536 
2537 	build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
2538 
2539 	/*
2540 	 * IOMMU TLB needs to be flushed before Device TLB to
2541 	 * prevent device TLB refill from IOMMU TLB
2542 	 */
2543 	for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
2544 		if (domain->dev_iommu[i] == 0)
2545 			continue;
2546 
2547 		ret = iommu_queue_command(amd_iommus[i], &cmd);
2548 		if (ret != 0)
2549 			goto out;
2550 	}
2551 
2552 	/* Wait until IOMMU TLB flushes are complete */
2553 	amd_iommu_domain_flush_complete(domain);
2554 
2555 	/* Now flush device TLBs */
2556 	list_for_each_entry(dev_data, &domain->dev_list, list) {
2557 		struct amd_iommu *iommu;
2558 		int qdep;
2559 
2560 		/*
2561 		   There might be non-IOMMUv2 capable devices in an IOMMUv2
2562 		 * domain.
2563 		 */
2564 		if (!dev_data->ats.enabled)
2565 			continue;
2566 
2567 		qdep  = dev_data->ats.qdep;
2568 		iommu = rlookup_amd_iommu(dev_data->dev);
2569 		if (!iommu)
2570 			continue;
2571 		build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
2572 				      qdep, address, size);
2573 
2574 		ret = iommu_queue_command(iommu, &cmd);
2575 		if (ret != 0)
2576 			goto out;
2577 	}
2578 
2579 	/* Wait until all device TLBs are flushed */
2580 	amd_iommu_domain_flush_complete(domain);
2581 
2582 	ret = 0;
2583 
2584 out:
2585 
2586 	return ret;
2587 }
2588 
2589 static int __amd_iommu_flush_page(struct protection_domain *domain, u32 pasid,
2590 				  u64 address)
2591 {
2592 	return __flush_pasid(domain, pasid, address, false);
2593 }
2594 
2595 int amd_iommu_flush_page(struct iommu_domain *dom, u32 pasid,
2596 			 u64 address)
2597 {
2598 	struct protection_domain *domain = to_pdomain(dom);
2599 	unsigned long flags;
2600 	int ret;
2601 
2602 	spin_lock_irqsave(&domain->lock, flags);
2603 	ret = __amd_iommu_flush_page(domain, pasid, address);
2604 	spin_unlock_irqrestore(&domain->lock, flags);
2605 
2606 	return ret;
2607 }
2608 EXPORT_SYMBOL(amd_iommu_flush_page);
2609 
2610 static int __amd_iommu_flush_tlb(struct protection_domain *domain, u32 pasid)
2611 {
2612 	return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
2613 			     true);
2614 }
2615 
2616 int amd_iommu_flush_tlb(struct iommu_domain *dom, u32 pasid)
2617 {
2618 	struct protection_domain *domain = to_pdomain(dom);
2619 	unsigned long flags;
2620 	int ret;
2621 
2622 	spin_lock_irqsave(&domain->lock, flags);
2623 	ret = __amd_iommu_flush_tlb(domain, pasid);
2624 	spin_unlock_irqrestore(&domain->lock, flags);
2625 
2626 	return ret;
2627 }
2628 EXPORT_SYMBOL(amd_iommu_flush_tlb);
2629 
2630 static u64 *__get_gcr3_pte(u64 *root, int level, u32 pasid, bool alloc)
2631 {
2632 	int index;
2633 	u64 *pte;
2634 
2635 	while (true) {
2636 
2637 		index = (pasid >> (9 * level)) & 0x1ff;
2638 		pte   = &root[index];
2639 
2640 		if (level == 0)
2641 			break;
2642 
2643 		if (!(*pte & GCR3_VALID)) {
2644 			if (!alloc)
2645 				return NULL;
2646 
2647 			root = (void *)get_zeroed_page(GFP_ATOMIC);
2648 			if (root == NULL)
2649 				return NULL;
2650 
2651 			*pte = iommu_virt_to_phys(root) | GCR3_VALID;
2652 		}
2653 
2654 		root = iommu_phys_to_virt(*pte & PAGE_MASK);
2655 
2656 		level -= 1;
2657 	}
2658 
2659 	return pte;
2660 }
2661 
2662 static int __set_gcr3(struct protection_domain *domain, u32 pasid,
2663 		      unsigned long cr3)
2664 {
2665 	u64 *pte;
2666 
2667 	if (domain->iop.mode != PAGE_MODE_NONE)
2668 		return -EINVAL;
2669 
2670 	pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
2671 	if (pte == NULL)
2672 		return -ENOMEM;
2673 
2674 	*pte = (cr3 & PAGE_MASK) | GCR3_VALID;
2675 
2676 	return __amd_iommu_flush_tlb(domain, pasid);
2677 }
2678 
2679 static int __clear_gcr3(struct protection_domain *domain, u32 pasid)
2680 {
2681 	u64 *pte;
2682 
2683 	if (domain->iop.mode != PAGE_MODE_NONE)
2684 		return -EINVAL;
2685 
2686 	pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
2687 	if (pte == NULL)
2688 		return 0;
2689 
2690 	*pte = 0;
2691 
2692 	return __amd_iommu_flush_tlb(domain, pasid);
2693 }
2694 
2695 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, u32 pasid,
2696 			      unsigned long cr3)
2697 {
2698 	struct protection_domain *domain = to_pdomain(dom);
2699 	unsigned long flags;
2700 	int ret;
2701 
2702 	spin_lock_irqsave(&domain->lock, flags);
2703 	ret = __set_gcr3(domain, pasid, cr3);
2704 	spin_unlock_irqrestore(&domain->lock, flags);
2705 
2706 	return ret;
2707 }
2708 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
2709 
2710 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, u32 pasid)
2711 {
2712 	struct protection_domain *domain = to_pdomain(dom);
2713 	unsigned long flags;
2714 	int ret;
2715 
2716 	spin_lock_irqsave(&domain->lock, flags);
2717 	ret = __clear_gcr3(domain, pasid);
2718 	spin_unlock_irqrestore(&domain->lock, flags);
2719 
2720 	return ret;
2721 }
2722 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
2723 
2724 int amd_iommu_complete_ppr(struct pci_dev *pdev, u32 pasid,
2725 			   int status, int tag)
2726 {
2727 	struct iommu_dev_data *dev_data;
2728 	struct amd_iommu *iommu;
2729 	struct iommu_cmd cmd;
2730 
2731 	dev_data = dev_iommu_priv_get(&pdev->dev);
2732 	iommu    = rlookup_amd_iommu(&pdev->dev);
2733 	if (!iommu)
2734 		return -ENODEV;
2735 
2736 	build_complete_ppr(&cmd, dev_data->devid, pasid, status,
2737 			   tag, dev_data->pri_tlp);
2738 
2739 	return iommu_queue_command(iommu, &cmd);
2740 }
2741 EXPORT_SYMBOL(amd_iommu_complete_ppr);
2742 
2743 int amd_iommu_device_info(struct pci_dev *pdev,
2744                           struct amd_iommu_device_info *info)
2745 {
2746 	int max_pasids;
2747 	int pos;
2748 
2749 	if (pdev == NULL || info == NULL)
2750 		return -EINVAL;
2751 
2752 	if (!amd_iommu_v2_supported())
2753 		return -EINVAL;
2754 
2755 	memset(info, 0, sizeof(*info));
2756 
2757 	if (pci_ats_supported(pdev))
2758 		info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
2759 
2760 	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2761 	if (pos)
2762 		info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
2763 
2764 	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
2765 	if (pos) {
2766 		int features;
2767 
2768 		max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
2769 		max_pasids = min(max_pasids, (1 << 20));
2770 
2771 		info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
2772 		info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
2773 
2774 		features = pci_pasid_features(pdev);
2775 		if (features & PCI_PASID_CAP_EXEC)
2776 			info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
2777 		if (features & PCI_PASID_CAP_PRIV)
2778 			info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
2779 	}
2780 
2781 	return 0;
2782 }
2783 EXPORT_SYMBOL(amd_iommu_device_info);
2784 
2785 #ifdef CONFIG_IRQ_REMAP
2786 
2787 /*****************************************************************************
2788  *
2789  * Interrupt Remapping Implementation
2790  *
2791  *****************************************************************************/
2792 
2793 static struct irq_chip amd_ir_chip;
2794 static DEFINE_SPINLOCK(iommu_table_lock);
2795 
2796 static void set_dte_irq_entry(struct amd_iommu *iommu, u16 devid,
2797 			      struct irq_remap_table *table)
2798 {
2799 	u64 dte;
2800 	struct dev_table_entry *dev_table = get_dev_table(iommu);
2801 
2802 	dte	= dev_table[devid].data[2];
2803 	dte	&= ~DTE_IRQ_PHYS_ADDR_MASK;
2804 	dte	|= iommu_virt_to_phys(table->table);
2805 	dte	|= DTE_IRQ_REMAP_INTCTL;
2806 	dte	|= DTE_INTTABLEN;
2807 	dte	|= DTE_IRQ_REMAP_ENABLE;
2808 
2809 	dev_table[devid].data[2] = dte;
2810 }
2811 
2812 static struct irq_remap_table *get_irq_table(struct amd_iommu *iommu, u16 devid)
2813 {
2814 	struct irq_remap_table *table;
2815 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
2816 
2817 	if (WARN_ONCE(!pci_seg->rlookup_table[devid],
2818 		      "%s: no iommu for devid %x:%x\n",
2819 		      __func__, pci_seg->id, devid))
2820 		return NULL;
2821 
2822 	table = pci_seg->irq_lookup_table[devid];
2823 	if (WARN_ONCE(!table, "%s: no table for devid %x:%x\n",
2824 		      __func__, pci_seg->id, devid))
2825 		return NULL;
2826 
2827 	return table;
2828 }
2829 
2830 static struct irq_remap_table *__alloc_irq_table(void)
2831 {
2832 	struct irq_remap_table *table;
2833 
2834 	table = kzalloc(sizeof(*table), GFP_KERNEL);
2835 	if (!table)
2836 		return NULL;
2837 
2838 	table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_KERNEL);
2839 	if (!table->table) {
2840 		kfree(table);
2841 		return NULL;
2842 	}
2843 	raw_spin_lock_init(&table->lock);
2844 
2845 	if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
2846 		memset(table->table, 0,
2847 		       MAX_IRQS_PER_TABLE * sizeof(u32));
2848 	else
2849 		memset(table->table, 0,
2850 		       (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
2851 	return table;
2852 }
2853 
2854 static void set_remap_table_entry(struct amd_iommu *iommu, u16 devid,
2855 				  struct irq_remap_table *table)
2856 {
2857 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
2858 
2859 	pci_seg->irq_lookup_table[devid] = table;
2860 	set_dte_irq_entry(iommu, devid, table);
2861 	iommu_flush_dte(iommu, devid);
2862 }
2863 
2864 static int set_remap_table_entry_alias(struct pci_dev *pdev, u16 alias,
2865 				       void *data)
2866 {
2867 	struct irq_remap_table *table = data;
2868 	struct amd_iommu_pci_seg *pci_seg;
2869 	struct amd_iommu *iommu = rlookup_amd_iommu(&pdev->dev);
2870 
2871 	if (!iommu)
2872 		return -EINVAL;
2873 
2874 	pci_seg = iommu->pci_seg;
2875 	pci_seg->irq_lookup_table[alias] = table;
2876 	set_dte_irq_entry(iommu, alias, table);
2877 	iommu_flush_dte(pci_seg->rlookup_table[alias], alias);
2878 
2879 	return 0;
2880 }
2881 
2882 static struct irq_remap_table *alloc_irq_table(struct amd_iommu *iommu,
2883 					       u16 devid, struct pci_dev *pdev)
2884 {
2885 	struct irq_remap_table *table = NULL;
2886 	struct irq_remap_table *new_table = NULL;
2887 	struct amd_iommu_pci_seg *pci_seg;
2888 	unsigned long flags;
2889 	u16 alias;
2890 
2891 	spin_lock_irqsave(&iommu_table_lock, flags);
2892 
2893 	pci_seg = iommu->pci_seg;
2894 	table = pci_seg->irq_lookup_table[devid];
2895 	if (table)
2896 		goto out_unlock;
2897 
2898 	alias = pci_seg->alias_table[devid];
2899 	table = pci_seg->irq_lookup_table[alias];
2900 	if (table) {
2901 		set_remap_table_entry(iommu, devid, table);
2902 		goto out_wait;
2903 	}
2904 	spin_unlock_irqrestore(&iommu_table_lock, flags);
2905 
2906 	/* Nothing there yet, allocate new irq remapping table */
2907 	new_table = __alloc_irq_table();
2908 	if (!new_table)
2909 		return NULL;
2910 
2911 	spin_lock_irqsave(&iommu_table_lock, flags);
2912 
2913 	table = pci_seg->irq_lookup_table[devid];
2914 	if (table)
2915 		goto out_unlock;
2916 
2917 	table = pci_seg->irq_lookup_table[alias];
2918 	if (table) {
2919 		set_remap_table_entry(iommu, devid, table);
2920 		goto out_wait;
2921 	}
2922 
2923 	table = new_table;
2924 	new_table = NULL;
2925 
2926 	if (pdev)
2927 		pci_for_each_dma_alias(pdev, set_remap_table_entry_alias,
2928 				       table);
2929 	else
2930 		set_remap_table_entry(iommu, devid, table);
2931 
2932 	if (devid != alias)
2933 		set_remap_table_entry(iommu, alias, table);
2934 
2935 out_wait:
2936 	iommu_completion_wait(iommu);
2937 
2938 out_unlock:
2939 	spin_unlock_irqrestore(&iommu_table_lock, flags);
2940 
2941 	if (new_table) {
2942 		kmem_cache_free(amd_iommu_irq_cache, new_table->table);
2943 		kfree(new_table);
2944 	}
2945 	return table;
2946 }
2947 
2948 static int alloc_irq_index(struct amd_iommu *iommu, u16 devid, int count,
2949 			   bool align, struct pci_dev *pdev)
2950 {
2951 	struct irq_remap_table *table;
2952 	int index, c, alignment = 1;
2953 	unsigned long flags;
2954 
2955 	table = alloc_irq_table(iommu, devid, pdev);
2956 	if (!table)
2957 		return -ENODEV;
2958 
2959 	if (align)
2960 		alignment = roundup_pow_of_two(count);
2961 
2962 	raw_spin_lock_irqsave(&table->lock, flags);
2963 
2964 	/* Scan table for free entries */
2965 	for (index = ALIGN(table->min_index, alignment), c = 0;
2966 	     index < MAX_IRQS_PER_TABLE;) {
2967 		if (!iommu->irte_ops->is_allocated(table, index)) {
2968 			c += 1;
2969 		} else {
2970 			c     = 0;
2971 			index = ALIGN(index + 1, alignment);
2972 			continue;
2973 		}
2974 
2975 		if (c == count)	{
2976 			for (; c != 0; --c)
2977 				iommu->irte_ops->set_allocated(table, index - c + 1);
2978 
2979 			index -= count - 1;
2980 			goto out;
2981 		}
2982 
2983 		index++;
2984 	}
2985 
2986 	index = -ENOSPC;
2987 
2988 out:
2989 	raw_spin_unlock_irqrestore(&table->lock, flags);
2990 
2991 	return index;
2992 }
2993 
2994 static int modify_irte_ga(struct amd_iommu *iommu, u16 devid, int index,
2995 			  struct irte_ga *irte, struct amd_ir_data *data)
2996 {
2997 	bool ret;
2998 	struct irq_remap_table *table;
2999 	unsigned long flags;
3000 	struct irte_ga *entry;
3001 
3002 	table = get_irq_table(iommu, devid);
3003 	if (!table)
3004 		return -ENOMEM;
3005 
3006 	raw_spin_lock_irqsave(&table->lock, flags);
3007 
3008 	entry = (struct irte_ga *)table->table;
3009 	entry = &entry[index];
3010 
3011 	ret = cmpxchg_double(&entry->lo.val, &entry->hi.val,
3012 			     entry->lo.val, entry->hi.val,
3013 			     irte->lo.val, irte->hi.val);
3014 	/*
3015 	 * We use cmpxchg16 to atomically update the 128-bit IRTE,
3016 	 * and it cannot be updated by the hardware or other processors
3017 	 * behind us, so the return value of cmpxchg16 should be the
3018 	 * same as the old value.
3019 	 */
3020 	WARN_ON(!ret);
3021 
3022 	if (data)
3023 		data->ref = entry;
3024 
3025 	raw_spin_unlock_irqrestore(&table->lock, flags);
3026 
3027 	iommu_flush_irt(iommu, devid);
3028 	iommu_completion_wait(iommu);
3029 
3030 	return 0;
3031 }
3032 
3033 static int modify_irte(struct amd_iommu *iommu,
3034 		       u16 devid, int index, union irte *irte)
3035 {
3036 	struct irq_remap_table *table;
3037 	unsigned long flags;
3038 
3039 	table = get_irq_table(iommu, devid);
3040 	if (!table)
3041 		return -ENOMEM;
3042 
3043 	raw_spin_lock_irqsave(&table->lock, flags);
3044 	table->table[index] = irte->val;
3045 	raw_spin_unlock_irqrestore(&table->lock, flags);
3046 
3047 	iommu_flush_irt(iommu, devid);
3048 	iommu_completion_wait(iommu);
3049 
3050 	return 0;
3051 }
3052 
3053 static void free_irte(struct amd_iommu *iommu, u16 devid, int index)
3054 {
3055 	struct irq_remap_table *table;
3056 	unsigned long flags;
3057 
3058 	table = get_irq_table(iommu, devid);
3059 	if (!table)
3060 		return;
3061 
3062 	raw_spin_lock_irqsave(&table->lock, flags);
3063 	iommu->irte_ops->clear_allocated(table, index);
3064 	raw_spin_unlock_irqrestore(&table->lock, flags);
3065 
3066 	iommu_flush_irt(iommu, devid);
3067 	iommu_completion_wait(iommu);
3068 }
3069 
3070 static void irte_prepare(void *entry,
3071 			 u32 delivery_mode, bool dest_mode,
3072 			 u8 vector, u32 dest_apicid, int devid)
3073 {
3074 	union irte *irte = (union irte *) entry;
3075 
3076 	irte->val                = 0;
3077 	irte->fields.vector      = vector;
3078 	irte->fields.int_type    = delivery_mode;
3079 	irte->fields.destination = dest_apicid;
3080 	irte->fields.dm          = dest_mode;
3081 	irte->fields.valid       = 1;
3082 }
3083 
3084 static void irte_ga_prepare(void *entry,
3085 			    u32 delivery_mode, bool dest_mode,
3086 			    u8 vector, u32 dest_apicid, int devid)
3087 {
3088 	struct irte_ga *irte = (struct irte_ga *) entry;
3089 
3090 	irte->lo.val                      = 0;
3091 	irte->hi.val                      = 0;
3092 	irte->lo.fields_remap.int_type    = delivery_mode;
3093 	irte->lo.fields_remap.dm          = dest_mode;
3094 	irte->hi.fields.vector            = vector;
3095 	irte->lo.fields_remap.destination = APICID_TO_IRTE_DEST_LO(dest_apicid);
3096 	irte->hi.fields.destination       = APICID_TO_IRTE_DEST_HI(dest_apicid);
3097 	irte->lo.fields_remap.valid       = 1;
3098 }
3099 
3100 static void irte_activate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3101 {
3102 	union irte *irte = (union irte *) entry;
3103 
3104 	irte->fields.valid = 1;
3105 	modify_irte(iommu, devid, index, irte);
3106 }
3107 
3108 static void irte_ga_activate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3109 {
3110 	struct irte_ga *irte = (struct irte_ga *) entry;
3111 
3112 	irte->lo.fields_remap.valid = 1;
3113 	modify_irte_ga(iommu, devid, index, irte, NULL);
3114 }
3115 
3116 static void irte_deactivate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3117 {
3118 	union irte *irte = (union irte *) entry;
3119 
3120 	irte->fields.valid = 0;
3121 	modify_irte(iommu, devid, index, irte);
3122 }
3123 
3124 static void irte_ga_deactivate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3125 {
3126 	struct irte_ga *irte = (struct irte_ga *) entry;
3127 
3128 	irte->lo.fields_remap.valid = 0;
3129 	modify_irte_ga(iommu, devid, index, irte, NULL);
3130 }
3131 
3132 static void irte_set_affinity(struct amd_iommu *iommu, void *entry, u16 devid, u16 index,
3133 			      u8 vector, u32 dest_apicid)
3134 {
3135 	union irte *irte = (union irte *) entry;
3136 
3137 	irte->fields.vector = vector;
3138 	irte->fields.destination = dest_apicid;
3139 	modify_irte(iommu, devid, index, irte);
3140 }
3141 
3142 static void irte_ga_set_affinity(struct amd_iommu *iommu, void *entry, u16 devid, u16 index,
3143 				 u8 vector, u32 dest_apicid)
3144 {
3145 	struct irte_ga *irte = (struct irte_ga *) entry;
3146 
3147 	if (!irte->lo.fields_remap.guest_mode) {
3148 		irte->hi.fields.vector = vector;
3149 		irte->lo.fields_remap.destination =
3150 					APICID_TO_IRTE_DEST_LO(dest_apicid);
3151 		irte->hi.fields.destination =
3152 					APICID_TO_IRTE_DEST_HI(dest_apicid);
3153 		modify_irte_ga(iommu, devid, index, irte, NULL);
3154 	}
3155 }
3156 
3157 #define IRTE_ALLOCATED (~1U)
3158 static void irte_set_allocated(struct irq_remap_table *table, int index)
3159 {
3160 	table->table[index] = IRTE_ALLOCATED;
3161 }
3162 
3163 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
3164 {
3165 	struct irte_ga *ptr = (struct irte_ga *)table->table;
3166 	struct irte_ga *irte = &ptr[index];
3167 
3168 	memset(&irte->lo.val, 0, sizeof(u64));
3169 	memset(&irte->hi.val, 0, sizeof(u64));
3170 	irte->hi.fields.vector = 0xff;
3171 }
3172 
3173 static bool irte_is_allocated(struct irq_remap_table *table, int index)
3174 {
3175 	union irte *ptr = (union irte *)table->table;
3176 	union irte *irte = &ptr[index];
3177 
3178 	return irte->val != 0;
3179 }
3180 
3181 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
3182 {
3183 	struct irte_ga *ptr = (struct irte_ga *)table->table;
3184 	struct irte_ga *irte = &ptr[index];
3185 
3186 	return irte->hi.fields.vector != 0;
3187 }
3188 
3189 static void irte_clear_allocated(struct irq_remap_table *table, int index)
3190 {
3191 	table->table[index] = 0;
3192 }
3193 
3194 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
3195 {
3196 	struct irte_ga *ptr = (struct irte_ga *)table->table;
3197 	struct irte_ga *irte = &ptr[index];
3198 
3199 	memset(&irte->lo.val, 0, sizeof(u64));
3200 	memset(&irte->hi.val, 0, sizeof(u64));
3201 }
3202 
3203 static int get_devid(struct irq_alloc_info *info)
3204 {
3205 	switch (info->type) {
3206 	case X86_IRQ_ALLOC_TYPE_IOAPIC:
3207 		return get_ioapic_devid(info->devid);
3208 	case X86_IRQ_ALLOC_TYPE_HPET:
3209 		return get_hpet_devid(info->devid);
3210 	case X86_IRQ_ALLOC_TYPE_PCI_MSI:
3211 	case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
3212 		return get_device_sbdf_id(msi_desc_to_dev(info->desc));
3213 	default:
3214 		WARN_ON_ONCE(1);
3215 		return -1;
3216 	}
3217 }
3218 
3219 struct irq_remap_ops amd_iommu_irq_ops = {
3220 	.prepare		= amd_iommu_prepare,
3221 	.enable			= amd_iommu_enable,
3222 	.disable		= amd_iommu_disable,
3223 	.reenable		= amd_iommu_reenable,
3224 	.enable_faulting	= amd_iommu_enable_faulting,
3225 };
3226 
3227 static void fill_msi_msg(struct msi_msg *msg, u32 index)
3228 {
3229 	msg->data = index;
3230 	msg->address_lo = 0;
3231 	msg->arch_addr_lo.base_address = X86_MSI_BASE_ADDRESS_LOW;
3232 	msg->address_hi = X86_MSI_BASE_ADDRESS_HIGH;
3233 }
3234 
3235 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3236 				       struct irq_cfg *irq_cfg,
3237 				       struct irq_alloc_info *info,
3238 				       int devid, int index, int sub_handle)
3239 {
3240 	struct irq_2_irte *irte_info = &data->irq_2_irte;
3241 	struct amd_iommu *iommu = data->iommu;
3242 
3243 	if (!iommu)
3244 		return;
3245 
3246 	data->irq_2_irte.devid = devid;
3247 	data->irq_2_irte.index = index + sub_handle;
3248 	iommu->irte_ops->prepare(data->entry, apic->delivery_mode,
3249 				 apic->dest_mode_logical, irq_cfg->vector,
3250 				 irq_cfg->dest_apicid, devid);
3251 
3252 	switch (info->type) {
3253 	case X86_IRQ_ALLOC_TYPE_IOAPIC:
3254 	case X86_IRQ_ALLOC_TYPE_HPET:
3255 	case X86_IRQ_ALLOC_TYPE_PCI_MSI:
3256 	case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
3257 		fill_msi_msg(&data->msi_entry, irte_info->index);
3258 		break;
3259 
3260 	default:
3261 		BUG_ON(1);
3262 		break;
3263 	}
3264 }
3265 
3266 struct amd_irte_ops irte_32_ops = {
3267 	.prepare = irte_prepare,
3268 	.activate = irte_activate,
3269 	.deactivate = irte_deactivate,
3270 	.set_affinity = irte_set_affinity,
3271 	.set_allocated = irte_set_allocated,
3272 	.is_allocated = irte_is_allocated,
3273 	.clear_allocated = irte_clear_allocated,
3274 };
3275 
3276 struct amd_irte_ops irte_128_ops = {
3277 	.prepare = irte_ga_prepare,
3278 	.activate = irte_ga_activate,
3279 	.deactivate = irte_ga_deactivate,
3280 	.set_affinity = irte_ga_set_affinity,
3281 	.set_allocated = irte_ga_set_allocated,
3282 	.is_allocated = irte_ga_is_allocated,
3283 	.clear_allocated = irte_ga_clear_allocated,
3284 };
3285 
3286 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
3287 			       unsigned int nr_irqs, void *arg)
3288 {
3289 	struct irq_alloc_info *info = arg;
3290 	struct irq_data *irq_data;
3291 	struct amd_ir_data *data = NULL;
3292 	struct amd_iommu *iommu;
3293 	struct irq_cfg *cfg;
3294 	int i, ret, devid, seg, sbdf;
3295 	int index;
3296 
3297 	if (!info)
3298 		return -EINVAL;
3299 	if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_PCI_MSI)
3300 		return -EINVAL;
3301 
3302 	sbdf = get_devid(info);
3303 	if (sbdf < 0)
3304 		return -EINVAL;
3305 
3306 	seg = PCI_SBDF_TO_SEGID(sbdf);
3307 	devid = PCI_SBDF_TO_DEVID(sbdf);
3308 	iommu = __rlookup_amd_iommu(seg, devid);
3309 	if (!iommu)
3310 		return -EINVAL;
3311 
3312 	ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
3313 	if (ret < 0)
3314 		return ret;
3315 
3316 	if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
3317 		struct irq_remap_table *table;
3318 
3319 		table = alloc_irq_table(iommu, devid, NULL);
3320 		if (table) {
3321 			if (!table->min_index) {
3322 				/*
3323 				 * Keep the first 32 indexes free for IOAPIC
3324 				 * interrupts.
3325 				 */
3326 				table->min_index = 32;
3327 				for (i = 0; i < 32; ++i)
3328 					iommu->irte_ops->set_allocated(table, i);
3329 			}
3330 			WARN_ON(table->min_index != 32);
3331 			index = info->ioapic.pin;
3332 		} else {
3333 			index = -ENOMEM;
3334 		}
3335 	} else if (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI ||
3336 		   info->type == X86_IRQ_ALLOC_TYPE_PCI_MSIX) {
3337 		bool align = (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI);
3338 
3339 		index = alloc_irq_index(iommu, devid, nr_irqs, align,
3340 					msi_desc_to_pci_dev(info->desc));
3341 	} else {
3342 		index = alloc_irq_index(iommu, devid, nr_irqs, false, NULL);
3343 	}
3344 
3345 	if (index < 0) {
3346 		pr_warn("Failed to allocate IRTE\n");
3347 		ret = index;
3348 		goto out_free_parent;
3349 	}
3350 
3351 	for (i = 0; i < nr_irqs; i++) {
3352 		irq_data = irq_domain_get_irq_data(domain, virq + i);
3353 		cfg = irq_data ? irqd_cfg(irq_data) : NULL;
3354 		if (!cfg) {
3355 			ret = -EINVAL;
3356 			goto out_free_data;
3357 		}
3358 
3359 		ret = -ENOMEM;
3360 		data = kzalloc(sizeof(*data), GFP_KERNEL);
3361 		if (!data)
3362 			goto out_free_data;
3363 
3364 		if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3365 			data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
3366 		else
3367 			data->entry = kzalloc(sizeof(struct irte_ga),
3368 						     GFP_KERNEL);
3369 		if (!data->entry) {
3370 			kfree(data);
3371 			goto out_free_data;
3372 		}
3373 
3374 		data->iommu = iommu;
3375 		irq_data->hwirq = (devid << 16) + i;
3376 		irq_data->chip_data = data;
3377 		irq_data->chip = &amd_ir_chip;
3378 		irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
3379 		irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
3380 	}
3381 
3382 	return 0;
3383 
3384 out_free_data:
3385 	for (i--; i >= 0; i--) {
3386 		irq_data = irq_domain_get_irq_data(domain, virq + i);
3387 		if (irq_data)
3388 			kfree(irq_data->chip_data);
3389 	}
3390 	for (i = 0; i < nr_irqs; i++)
3391 		free_irte(iommu, devid, index + i);
3392 out_free_parent:
3393 	irq_domain_free_irqs_common(domain, virq, nr_irqs);
3394 	return ret;
3395 }
3396 
3397 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
3398 			       unsigned int nr_irqs)
3399 {
3400 	struct irq_2_irte *irte_info;
3401 	struct irq_data *irq_data;
3402 	struct amd_ir_data *data;
3403 	int i;
3404 
3405 	for (i = 0; i < nr_irqs; i++) {
3406 		irq_data = irq_domain_get_irq_data(domain, virq  + i);
3407 		if (irq_data && irq_data->chip_data) {
3408 			data = irq_data->chip_data;
3409 			irte_info = &data->irq_2_irte;
3410 			free_irte(data->iommu, irte_info->devid, irte_info->index);
3411 			kfree(data->entry);
3412 			kfree(data);
3413 		}
3414 	}
3415 	irq_domain_free_irqs_common(domain, virq, nr_irqs);
3416 }
3417 
3418 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
3419 			       struct amd_ir_data *ir_data,
3420 			       struct irq_2_irte *irte_info,
3421 			       struct irq_cfg *cfg);
3422 
3423 static int irq_remapping_activate(struct irq_domain *domain,
3424 				  struct irq_data *irq_data, bool reserve)
3425 {
3426 	struct amd_ir_data *data = irq_data->chip_data;
3427 	struct irq_2_irte *irte_info = &data->irq_2_irte;
3428 	struct amd_iommu *iommu = data->iommu;
3429 	struct irq_cfg *cfg = irqd_cfg(irq_data);
3430 
3431 	if (!iommu)
3432 		return 0;
3433 
3434 	iommu->irte_ops->activate(iommu, data->entry, irte_info->devid,
3435 				  irte_info->index);
3436 	amd_ir_update_irte(irq_data, iommu, data, irte_info, cfg);
3437 	return 0;
3438 }
3439 
3440 static void irq_remapping_deactivate(struct irq_domain *domain,
3441 				     struct irq_data *irq_data)
3442 {
3443 	struct amd_ir_data *data = irq_data->chip_data;
3444 	struct irq_2_irte *irte_info = &data->irq_2_irte;
3445 	struct amd_iommu *iommu = data->iommu;
3446 
3447 	if (iommu)
3448 		iommu->irte_ops->deactivate(iommu, data->entry, irte_info->devid,
3449 					    irte_info->index);
3450 }
3451 
3452 static int irq_remapping_select(struct irq_domain *d, struct irq_fwspec *fwspec,
3453 				enum irq_domain_bus_token bus_token)
3454 {
3455 	struct amd_iommu *iommu;
3456 	int devid = -1;
3457 
3458 	if (!amd_iommu_irq_remap)
3459 		return 0;
3460 
3461 	if (x86_fwspec_is_ioapic(fwspec))
3462 		devid = get_ioapic_devid(fwspec->param[0]);
3463 	else if (x86_fwspec_is_hpet(fwspec))
3464 		devid = get_hpet_devid(fwspec->param[0]);
3465 
3466 	if (devid < 0)
3467 		return 0;
3468 	iommu = __rlookup_amd_iommu((devid >> 16), (devid & 0xffff));
3469 
3470 	return iommu && iommu->ir_domain == d;
3471 }
3472 
3473 static const struct irq_domain_ops amd_ir_domain_ops = {
3474 	.select = irq_remapping_select,
3475 	.alloc = irq_remapping_alloc,
3476 	.free = irq_remapping_free,
3477 	.activate = irq_remapping_activate,
3478 	.deactivate = irq_remapping_deactivate,
3479 };
3480 
3481 int amd_iommu_activate_guest_mode(void *data)
3482 {
3483 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3484 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3485 	u64 valid;
3486 
3487 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3488 	    !entry || entry->lo.fields_vapic.guest_mode)
3489 		return 0;
3490 
3491 	valid = entry->lo.fields_vapic.valid;
3492 
3493 	entry->lo.val = 0;
3494 	entry->hi.val = 0;
3495 
3496 	entry->lo.fields_vapic.valid       = valid;
3497 	entry->lo.fields_vapic.guest_mode  = 1;
3498 	entry->lo.fields_vapic.ga_log_intr = 1;
3499 	entry->hi.fields.ga_root_ptr       = ir_data->ga_root_ptr;
3500 	entry->hi.fields.vector            = ir_data->ga_vector;
3501 	entry->lo.fields_vapic.ga_tag      = ir_data->ga_tag;
3502 
3503 	return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3504 			      ir_data->irq_2_irte.index, entry, ir_data);
3505 }
3506 EXPORT_SYMBOL(amd_iommu_activate_guest_mode);
3507 
3508 int amd_iommu_deactivate_guest_mode(void *data)
3509 {
3510 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3511 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3512 	struct irq_cfg *cfg = ir_data->cfg;
3513 	u64 valid;
3514 
3515 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3516 	    !entry || !entry->lo.fields_vapic.guest_mode)
3517 		return 0;
3518 
3519 	valid = entry->lo.fields_remap.valid;
3520 
3521 	entry->lo.val = 0;
3522 	entry->hi.val = 0;
3523 
3524 	entry->lo.fields_remap.valid       = valid;
3525 	entry->lo.fields_remap.dm          = apic->dest_mode_logical;
3526 	entry->lo.fields_remap.int_type    = apic->delivery_mode;
3527 	entry->hi.fields.vector            = cfg->vector;
3528 	entry->lo.fields_remap.destination =
3529 				APICID_TO_IRTE_DEST_LO(cfg->dest_apicid);
3530 	entry->hi.fields.destination =
3531 				APICID_TO_IRTE_DEST_HI(cfg->dest_apicid);
3532 
3533 	return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3534 			      ir_data->irq_2_irte.index, entry, ir_data);
3535 }
3536 EXPORT_SYMBOL(amd_iommu_deactivate_guest_mode);
3537 
3538 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
3539 {
3540 	int ret;
3541 	struct amd_iommu_pi_data *pi_data = vcpu_info;
3542 	struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
3543 	struct amd_ir_data *ir_data = data->chip_data;
3544 	struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3545 	struct iommu_dev_data *dev_data;
3546 
3547 	if (ir_data->iommu == NULL)
3548 		return -EINVAL;
3549 
3550 	dev_data = search_dev_data(ir_data->iommu, irte_info->devid);
3551 
3552 	/* Note:
3553 	 * This device has never been set up for guest mode.
3554 	 * we should not modify the IRTE
3555 	 */
3556 	if (!dev_data || !dev_data->use_vapic)
3557 		return 0;
3558 
3559 	ir_data->cfg = irqd_cfg(data);
3560 	pi_data->ir_data = ir_data;
3561 
3562 	/* Note:
3563 	 * SVM tries to set up for VAPIC mode, but we are in
3564 	 * legacy mode. So, we force legacy mode instead.
3565 	 */
3566 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3567 		pr_debug("%s: Fall back to using intr legacy remap\n",
3568 			 __func__);
3569 		pi_data->is_guest_mode = false;
3570 	}
3571 
3572 	pi_data->prev_ga_tag = ir_data->cached_ga_tag;
3573 	if (pi_data->is_guest_mode) {
3574 		ir_data->ga_root_ptr = (pi_data->base >> 12);
3575 		ir_data->ga_vector = vcpu_pi_info->vector;
3576 		ir_data->ga_tag = pi_data->ga_tag;
3577 		ret = amd_iommu_activate_guest_mode(ir_data);
3578 		if (!ret)
3579 			ir_data->cached_ga_tag = pi_data->ga_tag;
3580 	} else {
3581 		ret = amd_iommu_deactivate_guest_mode(ir_data);
3582 
3583 		/*
3584 		 * This communicates the ga_tag back to the caller
3585 		 * so that it can do all the necessary clean up.
3586 		 */
3587 		if (!ret)
3588 			ir_data->cached_ga_tag = 0;
3589 	}
3590 
3591 	return ret;
3592 }
3593 
3594 
3595 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
3596 			       struct amd_ir_data *ir_data,
3597 			       struct irq_2_irte *irte_info,
3598 			       struct irq_cfg *cfg)
3599 {
3600 
3601 	/*
3602 	 * Atomically updates the IRTE with the new destination, vector
3603 	 * and flushes the interrupt entry cache.
3604 	 */
3605 	iommu->irte_ops->set_affinity(iommu, ir_data->entry, irte_info->devid,
3606 				      irte_info->index, cfg->vector,
3607 				      cfg->dest_apicid);
3608 }
3609 
3610 static int amd_ir_set_affinity(struct irq_data *data,
3611 			       const struct cpumask *mask, bool force)
3612 {
3613 	struct amd_ir_data *ir_data = data->chip_data;
3614 	struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3615 	struct irq_cfg *cfg = irqd_cfg(data);
3616 	struct irq_data *parent = data->parent_data;
3617 	struct amd_iommu *iommu = ir_data->iommu;
3618 	int ret;
3619 
3620 	if (!iommu)
3621 		return -ENODEV;
3622 
3623 	ret = parent->chip->irq_set_affinity(parent, mask, force);
3624 	if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
3625 		return ret;
3626 
3627 	amd_ir_update_irte(data, iommu, ir_data, irte_info, cfg);
3628 	/*
3629 	 * After this point, all the interrupts will start arriving
3630 	 * at the new destination. So, time to cleanup the previous
3631 	 * vector allocation.
3632 	 */
3633 	send_cleanup_vector(cfg);
3634 
3635 	return IRQ_SET_MASK_OK_DONE;
3636 }
3637 
3638 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
3639 {
3640 	struct amd_ir_data *ir_data = irq_data->chip_data;
3641 
3642 	*msg = ir_data->msi_entry;
3643 }
3644 
3645 static struct irq_chip amd_ir_chip = {
3646 	.name			= "AMD-IR",
3647 	.irq_ack		= apic_ack_irq,
3648 	.irq_set_affinity	= amd_ir_set_affinity,
3649 	.irq_set_vcpu_affinity	= amd_ir_set_vcpu_affinity,
3650 	.irq_compose_msi_msg	= ir_compose_msi_msg,
3651 };
3652 
3653 static const struct msi_parent_ops amdvi_msi_parent_ops = {
3654 	.supported_flags	= X86_VECTOR_MSI_FLAGS_SUPPORTED |
3655 				  MSI_FLAG_MULTI_PCI_MSI |
3656 				  MSI_FLAG_PCI_IMS,
3657 	.prefix			= "IR-",
3658 	.init_dev_msi_info	= msi_parent_init_dev_msi_info,
3659 };
3660 
3661 static const struct msi_parent_ops virt_amdvi_msi_parent_ops = {
3662 	.supported_flags	= X86_VECTOR_MSI_FLAGS_SUPPORTED |
3663 				  MSI_FLAG_MULTI_PCI_MSI,
3664 	.prefix			= "vIR-",
3665 	.init_dev_msi_info	= msi_parent_init_dev_msi_info,
3666 };
3667 
3668 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
3669 {
3670 	struct fwnode_handle *fn;
3671 
3672 	fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index);
3673 	if (!fn)
3674 		return -ENOMEM;
3675 	iommu->ir_domain = irq_domain_create_hierarchy(arch_get_ir_parent_domain(), 0, 0,
3676 						       fn, &amd_ir_domain_ops, iommu);
3677 	if (!iommu->ir_domain) {
3678 		irq_domain_free_fwnode(fn);
3679 		return -ENOMEM;
3680 	}
3681 
3682 	irq_domain_update_bus_token(iommu->ir_domain,  DOMAIN_BUS_AMDVI);
3683 	iommu->ir_domain->flags |= IRQ_DOMAIN_FLAG_MSI_PARENT |
3684 				   IRQ_DOMAIN_FLAG_ISOLATED_MSI;
3685 
3686 	if (amd_iommu_np_cache)
3687 		iommu->ir_domain->msi_parent_ops = &virt_amdvi_msi_parent_ops;
3688 	else
3689 		iommu->ir_domain->msi_parent_ops = &amdvi_msi_parent_ops;
3690 
3691 	return 0;
3692 }
3693 
3694 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
3695 {
3696 	unsigned long flags;
3697 	struct amd_iommu *iommu;
3698 	struct irq_remap_table *table;
3699 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3700 	int devid = ir_data->irq_2_irte.devid;
3701 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3702 	struct irte_ga *ref = (struct irte_ga *) ir_data->ref;
3703 
3704 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3705 	    !ref || !entry || !entry->lo.fields_vapic.guest_mode)
3706 		return 0;
3707 
3708 	iommu = ir_data->iommu;
3709 	if (!iommu)
3710 		return -ENODEV;
3711 
3712 	table = get_irq_table(iommu, devid);
3713 	if (!table)
3714 		return -ENODEV;
3715 
3716 	raw_spin_lock_irqsave(&table->lock, flags);
3717 
3718 	if (ref->lo.fields_vapic.guest_mode) {
3719 		if (cpu >= 0) {
3720 			ref->lo.fields_vapic.destination =
3721 						APICID_TO_IRTE_DEST_LO(cpu);
3722 			ref->hi.fields.destination =
3723 						APICID_TO_IRTE_DEST_HI(cpu);
3724 		}
3725 		ref->lo.fields_vapic.is_run = is_run;
3726 		barrier();
3727 	}
3728 
3729 	raw_spin_unlock_irqrestore(&table->lock, flags);
3730 
3731 	iommu_flush_irt(iommu, devid);
3732 	iommu_completion_wait(iommu);
3733 	return 0;
3734 }
3735 EXPORT_SYMBOL(amd_iommu_update_ga);
3736 #endif
3737