• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
4  * Author: Joerg Roedel <jroedel@suse.de>
5  *         Leo Duran <leo.duran@amd.com>
6  */
7 
8 #define pr_fmt(fmt)     "AMD-Vi: " fmt
9 #define dev_fmt(fmt)    pr_fmt(fmt)
10 
11 #include <linux/ratelimit.h>
12 #include <linux/pci.h>
13 #include <linux/acpi.h>
14 #include <linux/amba/bus.h>
15 #include <linux/platform_device.h>
16 #include <linux/pci-ats.h>
17 #include <linux/bitmap.h>
18 #include <linux/slab.h>
19 #include <linux/debugfs.h>
20 #include <linux/scatterlist.h>
21 #include <linux/dma-map-ops.h>
22 #include <linux/dma-direct.h>
23 #include <linux/dma-iommu.h>
24 #include <linux/iommu-helper.h>
25 #include <linux/delay.h>
26 #include <linux/amd-iommu.h>
27 #include <linux/notifier.h>
28 #include <linux/export.h>
29 #include <linux/irq.h>
30 #include <linux/msi.h>
31 #include <linux/irqdomain.h>
32 #include <linux/percpu.h>
33 #include <linux/iova.h>
34 #include <asm/irq_remapping.h>
35 #include <asm/io_apic.h>
36 #include <asm/apic.h>
37 #include <asm/hw_irq.h>
38 #include <asm/msidef.h>
39 #include <asm/proto.h>
40 #include <asm/iommu.h>
41 #include <asm/gart.h>
42 #include <asm/dma.h>
43 
44 #include "amd_iommu.h"
45 #include "../irq_remapping.h"
46 
47 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
48 
49 #define LOOP_TIMEOUT	100000
50 
51 /* IO virtual address start page frame number */
52 #define IOVA_START_PFN		(1)
53 #define IOVA_PFN(addr)		((addr) >> PAGE_SHIFT)
54 
55 /* Reserved IOVA ranges */
56 #define MSI_RANGE_START		(0xfee00000)
57 #define MSI_RANGE_END		(0xfeefffff)
58 #define HT_RANGE_START		(0xfd00000000ULL)
59 #define HT_RANGE_END		(0xffffffffffULL)
60 
61 /*
62  * This bitmap is used to advertise the page sizes our hardware support
63  * to the IOMMU core, which will then use this information to split
64  * physically contiguous memory regions it is mapping into page sizes
65  * that we support.
66  *
67  * 512GB Pages are not supported due to a hardware bug
68  */
69 #define AMD_IOMMU_PGSIZES	((~0xFFFUL) & ~(2ULL << 38))
70 
71 #define DEFAULT_PGTABLE_LEVEL	PAGE_MODE_3_LEVEL
72 
73 static DEFINE_SPINLOCK(pd_bitmap_lock);
74 
75 /* List of all available dev_data structures */
76 static LLIST_HEAD(dev_data_list);
77 
78 LIST_HEAD(ioapic_map);
79 LIST_HEAD(hpet_map);
80 LIST_HEAD(acpihid_map);
81 
82 /*
83  * Domain for untranslated devices - only allocated
84  * if iommu=pt passed on kernel cmd line.
85  */
86 const struct iommu_ops amd_iommu_ops;
87 
88 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
89 int amd_iommu_max_glx_val = -1;
90 
91 /*
92  * general struct to manage commands send to an IOMMU
93  */
94 struct iommu_cmd {
95 	u32 data[4];
96 };
97 
98 struct kmem_cache *amd_iommu_irq_cache;
99 
100 static void update_domain(struct protection_domain *domain);
101 static void detach_device(struct device *dev);
102 static void update_and_flush_device_table(struct protection_domain *domain,
103 					  struct domain_pgtable *pgtable);
104 
105 /****************************************************************************
106  *
107  * Helper functions
108  *
109  ****************************************************************************/
110 
get_pci_device_id(struct device * dev)111 static inline u16 get_pci_device_id(struct device *dev)
112 {
113 	struct pci_dev *pdev = to_pci_dev(dev);
114 
115 	return pci_dev_id(pdev);
116 }
117 
get_acpihid_device_id(struct device * dev,struct acpihid_map_entry ** entry)118 static inline int get_acpihid_device_id(struct device *dev,
119 					struct acpihid_map_entry **entry)
120 {
121 	struct acpi_device *adev = ACPI_COMPANION(dev);
122 	struct acpihid_map_entry *p;
123 
124 	if (!adev)
125 		return -ENODEV;
126 
127 	list_for_each_entry(p, &acpihid_map, list) {
128 		if (acpi_dev_hid_uid_match(adev, p->hid,
129 					   p->uid[0] ? p->uid : NULL)) {
130 			if (entry)
131 				*entry = p;
132 			return p->devid;
133 		}
134 	}
135 	return -EINVAL;
136 }
137 
get_device_id(struct device * dev)138 static inline int get_device_id(struct device *dev)
139 {
140 	int devid;
141 
142 	if (dev_is_pci(dev))
143 		devid = get_pci_device_id(dev);
144 	else
145 		devid = get_acpihid_device_id(dev, NULL);
146 
147 	return devid;
148 }
149 
to_pdomain(struct iommu_domain * dom)150 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
151 {
152 	return container_of(dom, struct protection_domain, domain);
153 }
154 
amd_iommu_domain_get_pgtable(struct protection_domain * domain,struct domain_pgtable * pgtable)155 static void amd_iommu_domain_get_pgtable(struct protection_domain *domain,
156 					 struct domain_pgtable *pgtable)
157 {
158 	u64 pt_root = atomic64_read(&domain->pt_root);
159 
160 	pgtable->root = (u64 *)(pt_root & PAGE_MASK);
161 	pgtable->mode = pt_root & 7; /* lowest 3 bits encode pgtable mode */
162 }
163 
amd_iommu_domain_set_pt_root(struct protection_domain * domain,u64 root)164 static void amd_iommu_domain_set_pt_root(struct protection_domain *domain, u64 root)
165 {
166 	atomic64_set(&domain->pt_root, root);
167 }
168 
amd_iommu_domain_clr_pt_root(struct protection_domain * domain)169 static void amd_iommu_domain_clr_pt_root(struct protection_domain *domain)
170 {
171 	amd_iommu_domain_set_pt_root(domain, 0);
172 }
173 
amd_iommu_domain_set_pgtable(struct protection_domain * domain,u64 * root,int mode)174 static void amd_iommu_domain_set_pgtable(struct protection_domain *domain,
175 					 u64 *root, int mode)
176 {
177 	u64 pt_root;
178 
179 	/* lowest 3 bits encode pgtable mode */
180 	pt_root = mode & 7;
181 	pt_root |= (u64)root;
182 
183 	amd_iommu_domain_set_pt_root(domain, pt_root);
184 }
185 
alloc_dev_data(u16 devid)186 static struct iommu_dev_data *alloc_dev_data(u16 devid)
187 {
188 	struct iommu_dev_data *dev_data;
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, &dev_data_list);
199 	return dev_data;
200 }
201 
search_dev_data(u16 devid)202 static struct iommu_dev_data *search_dev_data(u16 devid)
203 {
204 	struct iommu_dev_data *dev_data;
205 	struct llist_node *node;
206 
207 	if (llist_empty(&dev_data_list))
208 		return NULL;
209 
210 	node = dev_data_list.first;
211 	llist_for_each_entry(dev_data, node, dev_data_list) {
212 		if (dev_data->devid == devid)
213 			return dev_data;
214 	}
215 
216 	return NULL;
217 }
218 
clone_alias(struct pci_dev * pdev,u16 alias,void * data)219 static int clone_alias(struct pci_dev *pdev, u16 alias, void *data)
220 {
221 	u16 devid = pci_dev_id(pdev);
222 
223 	if (devid == alias)
224 		return 0;
225 
226 	amd_iommu_rlookup_table[alias] =
227 		amd_iommu_rlookup_table[devid];
228 	memcpy(amd_iommu_dev_table[alias].data,
229 	       amd_iommu_dev_table[devid].data,
230 	       sizeof(amd_iommu_dev_table[alias].data));
231 
232 	return 0;
233 }
234 
clone_aliases(struct pci_dev * pdev)235 static void clone_aliases(struct pci_dev *pdev)
236 {
237 	if (!pdev)
238 		return;
239 
240 	/*
241 	 * The IVRS alias stored in the alias table may not be
242 	 * part of the PCI DMA aliases if it's bus differs
243 	 * from the original device.
244 	 */
245 	clone_alias(pdev, amd_iommu_alias_table[pci_dev_id(pdev)], NULL);
246 
247 	pci_for_each_dma_alias(pdev, clone_alias, NULL);
248 }
249 
setup_aliases(struct device * dev)250 static struct pci_dev *setup_aliases(struct device *dev)
251 {
252 	struct pci_dev *pdev = to_pci_dev(dev);
253 	u16 ivrs_alias;
254 
255 	/* For ACPI HID devices, there are no aliases */
256 	if (!dev_is_pci(dev))
257 		return NULL;
258 
259 	/*
260 	 * Add the IVRS alias to the pci aliases if it is on the same
261 	 * bus. The IVRS table may know about a quirk that we don't.
262 	 */
263 	ivrs_alias = amd_iommu_alias_table[pci_dev_id(pdev)];
264 	if (ivrs_alias != pci_dev_id(pdev) &&
265 	    PCI_BUS_NUM(ivrs_alias) == pdev->bus->number)
266 		pci_add_dma_alias(pdev, ivrs_alias & 0xff, 1);
267 
268 	clone_aliases(pdev);
269 
270 	return pdev;
271 }
272 
find_dev_data(u16 devid)273 static struct iommu_dev_data *find_dev_data(u16 devid)
274 {
275 	struct iommu_dev_data *dev_data;
276 	struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
277 
278 	dev_data = search_dev_data(devid);
279 
280 	if (dev_data == NULL) {
281 		dev_data = alloc_dev_data(devid);
282 		if (!dev_data)
283 			return NULL;
284 
285 		if (translation_pre_enabled(iommu))
286 			dev_data->defer_attach = true;
287 	}
288 
289 	return dev_data;
290 }
291 
292 /*
293 * Find or create an IOMMU group for a acpihid device.
294 */
acpihid_device_group(struct device * dev)295 static struct iommu_group *acpihid_device_group(struct device *dev)
296 {
297 	struct acpihid_map_entry *p, *entry = NULL;
298 	int devid;
299 
300 	devid = get_acpihid_device_id(dev, &entry);
301 	if (devid < 0)
302 		return ERR_PTR(devid);
303 
304 	list_for_each_entry(p, &acpihid_map, list) {
305 		if ((devid == p->devid) && p->group)
306 			entry->group = p->group;
307 	}
308 
309 	if (!entry->group)
310 		entry->group = generic_device_group(dev);
311 	else
312 		iommu_group_ref_get(entry->group);
313 
314 	return entry->group;
315 }
316 
pci_iommuv2_capable(struct pci_dev * pdev)317 static bool pci_iommuv2_capable(struct pci_dev *pdev)
318 {
319 	static const int caps[] = {
320 		PCI_EXT_CAP_ID_PRI,
321 		PCI_EXT_CAP_ID_PASID,
322 	};
323 	int i, pos;
324 
325 	if (!pci_ats_supported(pdev))
326 		return false;
327 
328 	for (i = 0; i < 2; ++i) {
329 		pos = pci_find_ext_capability(pdev, caps[i]);
330 		if (pos == 0)
331 			return false;
332 	}
333 
334 	return true;
335 }
336 
pdev_pri_erratum(struct pci_dev * pdev,u32 erratum)337 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
338 {
339 	struct iommu_dev_data *dev_data;
340 
341 	dev_data = dev_iommu_priv_get(&pdev->dev);
342 
343 	return dev_data->errata & (1 << erratum) ? true : false;
344 }
345 
346 /*
347  * This function checks if the driver got a valid device from the caller to
348  * avoid dereferencing invalid pointers.
349  */
check_device(struct device * dev)350 static bool check_device(struct device *dev)
351 {
352 	int devid;
353 
354 	if (!dev)
355 		return false;
356 
357 	devid = get_device_id(dev);
358 	if (devid < 0)
359 		return false;
360 
361 	/* Out of our scope? */
362 	if (devid > amd_iommu_last_bdf)
363 		return false;
364 
365 	if (amd_iommu_rlookup_table[devid] == NULL)
366 		return false;
367 
368 	return true;
369 }
370 
iommu_init_device(struct device * dev)371 static int iommu_init_device(struct device *dev)
372 {
373 	struct iommu_dev_data *dev_data;
374 	int devid;
375 
376 	if (dev_iommu_priv_get(dev))
377 		return 0;
378 
379 	devid = get_device_id(dev);
380 	if (devid < 0)
381 		return devid;
382 
383 	dev_data = find_dev_data(devid);
384 	if (!dev_data)
385 		return -ENOMEM;
386 
387 	dev_data->pdev = setup_aliases(dev);
388 
389 	/*
390 	 * By default we use passthrough mode for IOMMUv2 capable device.
391 	 * But if amd_iommu=force_isolation is set (e.g. to debug DMA to
392 	 * invalid address), we ignore the capability for the device so
393 	 * it'll be forced to go into translation mode.
394 	 */
395 	if ((iommu_default_passthrough() || !amd_iommu_force_isolation) &&
396 	    dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
397 		struct amd_iommu *iommu;
398 
399 		iommu = amd_iommu_rlookup_table[dev_data->devid];
400 		dev_data->iommu_v2 = iommu->is_iommu_v2;
401 	}
402 
403 	dev_iommu_priv_set(dev, dev_data);
404 
405 	return 0;
406 }
407 
iommu_ignore_device(struct device * dev)408 static void iommu_ignore_device(struct device *dev)
409 {
410 	int devid;
411 
412 	devid = get_device_id(dev);
413 	if (devid < 0)
414 		return;
415 
416 	amd_iommu_rlookup_table[devid] = NULL;
417 	memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
418 
419 	setup_aliases(dev);
420 }
421 
amd_iommu_uninit_device(struct device * dev)422 static void amd_iommu_uninit_device(struct device *dev)
423 {
424 	struct iommu_dev_data *dev_data;
425 
426 	dev_data = dev_iommu_priv_get(dev);
427 	if (!dev_data)
428 		return;
429 
430 	if (dev_data->domain)
431 		detach_device(dev);
432 
433 	dev_iommu_priv_set(dev, NULL);
434 
435 	/*
436 	 * We keep dev_data around for unplugged devices and reuse it when the
437 	 * device is re-plugged - not doing so would introduce a ton of races.
438 	 */
439 }
440 
441 /*
442  * Helper function to get the first pte of a large mapping
443  */
first_pte_l7(u64 * pte,unsigned long * page_size,unsigned long * count)444 static u64 *first_pte_l7(u64 *pte, unsigned long *page_size,
445 			 unsigned long *count)
446 {
447 	unsigned long pte_mask, pg_size, cnt;
448 	u64 *fpte;
449 
450 	pg_size  = PTE_PAGE_SIZE(*pte);
451 	cnt      = PAGE_SIZE_PTE_COUNT(pg_size);
452 	pte_mask = ~((cnt << 3) - 1);
453 	fpte     = (u64 *)(((unsigned long)pte) & pte_mask);
454 
455 	if (page_size)
456 		*page_size = pg_size;
457 
458 	if (count)
459 		*count = cnt;
460 
461 	return fpte;
462 }
463 
464 /****************************************************************************
465  *
466  * Interrupt handling functions
467  *
468  ****************************************************************************/
469 
dump_dte_entry(u16 devid)470 static void dump_dte_entry(u16 devid)
471 {
472 	int i;
473 
474 	for (i = 0; i < 4; ++i)
475 		pr_err("DTE[%d]: %016llx\n", i,
476 			amd_iommu_dev_table[devid].data[i]);
477 }
478 
dump_command(unsigned long phys_addr)479 static void dump_command(unsigned long phys_addr)
480 {
481 	struct iommu_cmd *cmd = iommu_phys_to_virt(phys_addr);
482 	int i;
483 
484 	for (i = 0; i < 4; ++i)
485 		pr_err("CMD[%d]: %08x\n", i, cmd->data[i]);
486 }
487 
amd_iommu_report_rmp_hw_error(volatile u32 * event)488 static void amd_iommu_report_rmp_hw_error(volatile u32 *event)
489 {
490 	struct iommu_dev_data *dev_data = NULL;
491 	int devid, vmg_tag, flags;
492 	struct pci_dev *pdev;
493 	u64 spa;
494 
495 	devid   = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
496 	vmg_tag = (event[1]) & 0xFFFF;
497 	flags   = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
498 	spa     = ((u64)event[3] << 32) | (event[2] & 0xFFFFFFF8);
499 
500 	pdev = pci_get_domain_bus_and_slot(0, PCI_BUS_NUM(devid),
501 					   devid & 0xff);
502 	if (pdev)
503 		dev_data = dev_iommu_priv_get(&pdev->dev);
504 
505 	if (dev_data && __ratelimit(&dev_data->rs)) {
506 		pci_err(pdev, "Event logged [RMP_HW_ERROR vmg_tag=0x%04x, spa=0x%llx, flags=0x%04x]\n",
507 			vmg_tag, spa, flags);
508 	} else {
509 		pr_err_ratelimited("Event logged [RMP_HW_ERROR device=%02x:%02x.%x, vmg_tag=0x%04x, spa=0x%llx, flags=0x%04x]\n",
510 			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
511 			vmg_tag, spa, flags);
512 	}
513 
514 	if (pdev)
515 		pci_dev_put(pdev);
516 }
517 
amd_iommu_report_rmp_fault(volatile u32 * event)518 static void amd_iommu_report_rmp_fault(volatile u32 *event)
519 {
520 	struct iommu_dev_data *dev_data = NULL;
521 	int devid, flags_rmp, vmg_tag, flags;
522 	struct pci_dev *pdev;
523 	u64 gpa;
524 
525 	devid     = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
526 	flags_rmp = (event[0] >> EVENT_FLAGS_SHIFT) & 0xFF;
527 	vmg_tag   = (event[1]) & 0xFFFF;
528 	flags     = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
529 	gpa       = ((u64)event[3] << 32) | event[2];
530 
531 	pdev = pci_get_domain_bus_and_slot(0, PCI_BUS_NUM(devid),
532 					   devid & 0xff);
533 	if (pdev)
534 		dev_data = dev_iommu_priv_get(&pdev->dev);
535 
536 	if (dev_data && __ratelimit(&dev_data->rs)) {
537 		pci_err(pdev, "Event logged [RMP_PAGE_FAULT vmg_tag=0x%04x, gpa=0x%llx, flags_rmp=0x%04x, flags=0x%04x]\n",
538 			vmg_tag, gpa, flags_rmp, flags);
539 	} else {
540 		pr_err_ratelimited("Event logged [RMP_PAGE_FAULT device=%02x:%02x.%x, vmg_tag=0x%04x, gpa=0x%llx, flags_rmp=0x%04x, flags=0x%04x]\n",
541 			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
542 			vmg_tag, gpa, flags_rmp, flags);
543 	}
544 
545 	if (pdev)
546 		pci_dev_put(pdev);
547 }
548 
amd_iommu_report_page_fault(u16 devid,u16 domain_id,u64 address,int flags)549 static void amd_iommu_report_page_fault(u16 devid, u16 domain_id,
550 					u64 address, int flags)
551 {
552 	struct iommu_dev_data *dev_data = NULL;
553 	struct pci_dev *pdev;
554 
555 	pdev = pci_get_domain_bus_and_slot(0, PCI_BUS_NUM(devid),
556 					   devid & 0xff);
557 	if (pdev)
558 		dev_data = dev_iommu_priv_get(&pdev->dev);
559 
560 	if (dev_data && __ratelimit(&dev_data->rs)) {
561 		pci_err(pdev, "Event logged [IO_PAGE_FAULT domain=0x%04x address=0x%llx flags=0x%04x]\n",
562 			domain_id, address, flags);
563 	} else if (printk_ratelimit()) {
564 		pr_err("Event logged [IO_PAGE_FAULT device=%02x:%02x.%x domain=0x%04x address=0x%llx flags=0x%04x]\n",
565 			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
566 			domain_id, address, flags);
567 	}
568 
569 	if (pdev)
570 		pci_dev_put(pdev);
571 }
572 
iommu_print_event(struct amd_iommu * iommu,void * __evt)573 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
574 {
575 	struct device *dev = iommu->iommu.dev;
576 	int type, devid, flags, tag;
577 	volatile u32 *event = __evt;
578 	int count = 0;
579 	u64 address;
580 	u32 pasid;
581 
582 retry:
583 	type    = (event[1] >> EVENT_TYPE_SHIFT)  & EVENT_TYPE_MASK;
584 	devid   = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
585 	pasid   = (event[0] & EVENT_DOMID_MASK_HI) |
586 		  (event[1] & EVENT_DOMID_MASK_LO);
587 	flags   = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
588 	address = (u64)(((u64)event[3]) << 32) | event[2];
589 
590 	if (type == 0) {
591 		/* Did we hit the erratum? */
592 		if (++count == LOOP_TIMEOUT) {
593 			pr_err("No event written to event log\n");
594 			return;
595 		}
596 		udelay(1);
597 		goto retry;
598 	}
599 
600 	if (type == EVENT_TYPE_IO_FAULT) {
601 		amd_iommu_report_page_fault(devid, pasid, address, flags);
602 		return;
603 	}
604 
605 	switch (type) {
606 	case EVENT_TYPE_ILL_DEV:
607 		dev_err(dev, "Event logged [ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
608 			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
609 			pasid, address, flags);
610 		dump_dte_entry(devid);
611 		break;
612 	case EVENT_TYPE_DEV_TAB_ERR:
613 		dev_err(dev, "Event logged [DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
614 			"address=0x%llx flags=0x%04x]\n",
615 			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
616 			address, flags);
617 		break;
618 	case EVENT_TYPE_PAGE_TAB_ERR:
619 		dev_err(dev, "Event logged [PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x pasid=0x%04x address=0x%llx flags=0x%04x]\n",
620 			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
621 			pasid, address, flags);
622 		break;
623 	case EVENT_TYPE_ILL_CMD:
624 		dev_err(dev, "Event logged [ILLEGAL_COMMAND_ERROR address=0x%llx]\n", address);
625 		dump_command(address);
626 		break;
627 	case EVENT_TYPE_CMD_HARD_ERR:
628 		dev_err(dev, "Event logged [COMMAND_HARDWARE_ERROR address=0x%llx flags=0x%04x]\n",
629 			address, flags);
630 		break;
631 	case EVENT_TYPE_IOTLB_INV_TO:
632 		dev_err(dev, "Event logged [IOTLB_INV_TIMEOUT device=%02x:%02x.%x address=0x%llx]\n",
633 			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
634 			address);
635 		break;
636 	case EVENT_TYPE_INV_DEV_REQ:
637 		dev_err(dev, "Event logged [INVALID_DEVICE_REQUEST device=%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
638 			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
639 			pasid, address, flags);
640 		break;
641 	case EVENT_TYPE_RMP_FAULT:
642 		amd_iommu_report_rmp_fault(event);
643 		break;
644 	case EVENT_TYPE_RMP_HW_ERR:
645 		amd_iommu_report_rmp_hw_error(event);
646 		break;
647 	case EVENT_TYPE_INV_PPR_REQ:
648 		pasid = PPR_PASID(*((u64 *)__evt));
649 		tag = event[1] & 0x03FF;
650 		dev_err(dev, "Event logged [INVALID_PPR_REQUEST device=%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x tag=0x%03x]\n",
651 			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
652 			pasid, address, flags, tag);
653 		break;
654 	default:
655 		dev_err(dev, "Event logged [UNKNOWN event[0]=0x%08x event[1]=0x%08x event[2]=0x%08x event[3]=0x%08x\n",
656 			event[0], event[1], event[2], event[3]);
657 	}
658 
659 	memset(__evt, 0, 4 * sizeof(u32));
660 }
661 
iommu_poll_events(struct amd_iommu * iommu)662 static void iommu_poll_events(struct amd_iommu *iommu)
663 {
664 	u32 head, tail;
665 
666 	head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
667 	tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
668 
669 	while (head != tail) {
670 		iommu_print_event(iommu, iommu->evt_buf + head);
671 		head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
672 	}
673 
674 	writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
675 }
676 
iommu_handle_ppr_entry(struct amd_iommu * iommu,u64 * raw)677 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
678 {
679 	struct amd_iommu_fault fault;
680 
681 	if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
682 		pr_err_ratelimited("Unknown PPR request received\n");
683 		return;
684 	}
685 
686 	fault.address   = raw[1];
687 	fault.pasid     = PPR_PASID(raw[0]);
688 	fault.device_id = PPR_DEVID(raw[0]);
689 	fault.tag       = PPR_TAG(raw[0]);
690 	fault.flags     = PPR_FLAGS(raw[0]);
691 
692 	atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
693 }
694 
iommu_poll_ppr_log(struct amd_iommu * iommu)695 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
696 {
697 	u32 head, tail;
698 
699 	if (iommu->ppr_log == NULL)
700 		return;
701 
702 	head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
703 	tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
704 
705 	while (head != tail) {
706 		volatile u64 *raw;
707 		u64 entry[2];
708 		int i;
709 
710 		raw = (u64 *)(iommu->ppr_log + head);
711 
712 		/*
713 		 * Hardware bug: Interrupt may arrive before the entry is
714 		 * written to memory. If this happens we need to wait for the
715 		 * entry to arrive.
716 		 */
717 		for (i = 0; i < LOOP_TIMEOUT; ++i) {
718 			if (PPR_REQ_TYPE(raw[0]) != 0)
719 				break;
720 			udelay(1);
721 		}
722 
723 		/* Avoid memcpy function-call overhead */
724 		entry[0] = raw[0];
725 		entry[1] = raw[1];
726 
727 		/*
728 		 * To detect the hardware bug we need to clear the entry
729 		 * back to zero.
730 		 */
731 		raw[0] = raw[1] = 0UL;
732 
733 		/* Update head pointer of hardware ring-buffer */
734 		head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
735 		writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
736 
737 		/* Handle PPR entry */
738 		iommu_handle_ppr_entry(iommu, entry);
739 
740 		/* Refresh ring-buffer information */
741 		head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
742 		tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
743 	}
744 }
745 
746 #ifdef CONFIG_IRQ_REMAP
747 static int (*iommu_ga_log_notifier)(u32);
748 
amd_iommu_register_ga_log_notifier(int (* notifier)(u32))749 int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
750 {
751 	iommu_ga_log_notifier = notifier;
752 
753 	return 0;
754 }
755 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
756 
iommu_poll_ga_log(struct amd_iommu * iommu)757 static void iommu_poll_ga_log(struct amd_iommu *iommu)
758 {
759 	u32 head, tail, cnt = 0;
760 
761 	if (iommu->ga_log == NULL)
762 		return;
763 
764 	head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
765 	tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
766 
767 	while (head != tail) {
768 		volatile u64 *raw;
769 		u64 log_entry;
770 
771 		raw = (u64 *)(iommu->ga_log + head);
772 		cnt++;
773 
774 		/* Avoid memcpy function-call overhead */
775 		log_entry = *raw;
776 
777 		/* Update head pointer of hardware ring-buffer */
778 		head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
779 		writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
780 
781 		/* Handle GA entry */
782 		switch (GA_REQ_TYPE(log_entry)) {
783 		case GA_GUEST_NR:
784 			if (!iommu_ga_log_notifier)
785 				break;
786 
787 			pr_debug("%s: devid=%#x, ga_tag=%#x\n",
788 				 __func__, GA_DEVID(log_entry),
789 				 GA_TAG(log_entry));
790 
791 			if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
792 				pr_err("GA log notifier failed.\n");
793 			break;
794 		default:
795 			break;
796 		}
797 	}
798 }
799 
800 static void
amd_iommu_set_pci_msi_domain(struct device * dev,struct amd_iommu * iommu)801 amd_iommu_set_pci_msi_domain(struct device *dev, struct amd_iommu *iommu)
802 {
803 	if (!irq_remapping_enabled || !dev_is_pci(dev) ||
804 	    pci_dev_has_special_msi_domain(to_pci_dev(dev)))
805 		return;
806 
807 	dev_set_msi_domain(dev, iommu->msi_domain);
808 }
809 
810 #else /* CONFIG_IRQ_REMAP */
811 static inline void
amd_iommu_set_pci_msi_domain(struct device * dev,struct amd_iommu * iommu)812 amd_iommu_set_pci_msi_domain(struct device *dev, struct amd_iommu *iommu) { }
813 #endif /* !CONFIG_IRQ_REMAP */
814 
815 #define AMD_IOMMU_INT_MASK	\
816 	(MMIO_STATUS_EVT_INT_MASK | \
817 	 MMIO_STATUS_PPR_INT_MASK | \
818 	 MMIO_STATUS_GALOG_INT_MASK)
819 
amd_iommu_int_thread(int irq,void * data)820 irqreturn_t amd_iommu_int_thread(int irq, void *data)
821 {
822 	struct amd_iommu *iommu = (struct amd_iommu *) data;
823 	u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
824 
825 	while (status & AMD_IOMMU_INT_MASK) {
826 		/* Enable EVT and PPR and GA interrupts again */
827 		writel(AMD_IOMMU_INT_MASK,
828 			iommu->mmio_base + MMIO_STATUS_OFFSET);
829 
830 		if (status & MMIO_STATUS_EVT_INT_MASK) {
831 			pr_devel("Processing IOMMU Event Log\n");
832 			iommu_poll_events(iommu);
833 		}
834 
835 		if (status & MMIO_STATUS_PPR_INT_MASK) {
836 			pr_devel("Processing IOMMU PPR Log\n");
837 			iommu_poll_ppr_log(iommu);
838 		}
839 
840 #ifdef CONFIG_IRQ_REMAP
841 		if (status & MMIO_STATUS_GALOG_INT_MASK) {
842 			pr_devel("Processing IOMMU GA Log\n");
843 			iommu_poll_ga_log(iommu);
844 		}
845 #endif
846 
847 		/*
848 		 * Hardware bug: ERBT1312
849 		 * When re-enabling interrupt (by writing 1
850 		 * to clear the bit), the hardware might also try to set
851 		 * the interrupt bit in the event status register.
852 		 * In this scenario, the bit will be set, and disable
853 		 * subsequent interrupts.
854 		 *
855 		 * Workaround: The IOMMU driver should read back the
856 		 * status register and check if the interrupt bits are cleared.
857 		 * If not, driver will need to go through the interrupt handler
858 		 * again and re-clear the bits
859 		 */
860 		status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
861 	}
862 	return IRQ_HANDLED;
863 }
864 
amd_iommu_int_handler(int irq,void * data)865 irqreturn_t amd_iommu_int_handler(int irq, void *data)
866 {
867 	return IRQ_WAKE_THREAD;
868 }
869 
870 /****************************************************************************
871  *
872  * IOMMU command queuing functions
873  *
874  ****************************************************************************/
875 
wait_on_sem(struct amd_iommu * iommu,u64 data)876 static int wait_on_sem(struct amd_iommu *iommu, u64 data)
877 {
878 	int i = 0;
879 
880 	while (*iommu->cmd_sem != data && i < LOOP_TIMEOUT) {
881 		udelay(1);
882 		i += 1;
883 	}
884 
885 	if (i == LOOP_TIMEOUT) {
886 		pr_alert("Completion-Wait loop timed out\n");
887 		return -EIO;
888 	}
889 
890 	return 0;
891 }
892 
copy_cmd_to_buffer(struct amd_iommu * iommu,struct iommu_cmd * cmd)893 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
894 			       struct iommu_cmd *cmd)
895 {
896 	u8 *target;
897 	u32 tail;
898 
899 	/* Copy command to buffer */
900 	tail = iommu->cmd_buf_tail;
901 	target = iommu->cmd_buf + tail;
902 	memcpy(target, cmd, sizeof(*cmd));
903 
904 	tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
905 	iommu->cmd_buf_tail = tail;
906 
907 	/* Tell the IOMMU about it */
908 	writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
909 }
910 
build_completion_wait(struct iommu_cmd * cmd,struct amd_iommu * iommu,u64 data)911 static void build_completion_wait(struct iommu_cmd *cmd,
912 				  struct amd_iommu *iommu,
913 				  u64 data)
914 {
915 	u64 paddr = iommu_virt_to_phys((void *)iommu->cmd_sem);
916 
917 	memset(cmd, 0, sizeof(*cmd));
918 	cmd->data[0] = lower_32_bits(paddr) | CMD_COMPL_WAIT_STORE_MASK;
919 	cmd->data[1] = upper_32_bits(paddr);
920 	cmd->data[2] = data;
921 	CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
922 }
923 
build_inv_dte(struct iommu_cmd * cmd,u16 devid)924 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
925 {
926 	memset(cmd, 0, sizeof(*cmd));
927 	cmd->data[0] = devid;
928 	CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
929 }
930 
build_inv_iommu_pages(struct iommu_cmd * cmd,u64 address,size_t size,u16 domid,int pde)931 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
932 				  size_t size, u16 domid, int pde)
933 {
934 	u64 pages;
935 	bool s;
936 
937 	pages = iommu_num_pages(address, size, PAGE_SIZE);
938 	s     = false;
939 
940 	if (pages > 1) {
941 		/*
942 		 * If we have to flush more than one page, flush all
943 		 * TLB entries for this domain
944 		 */
945 		address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
946 		s = true;
947 	}
948 
949 	address &= PAGE_MASK;
950 
951 	memset(cmd, 0, sizeof(*cmd));
952 	cmd->data[1] |= domid;
953 	cmd->data[2]  = lower_32_bits(address);
954 	cmd->data[3]  = upper_32_bits(address);
955 	CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
956 	if (s) /* size bit - we flush more than one 4kb page */
957 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
958 	if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
959 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
960 }
961 
build_inv_iotlb_pages(struct iommu_cmd * cmd,u16 devid,int qdep,u64 address,size_t size)962 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
963 				  u64 address, size_t size)
964 {
965 	u64 pages;
966 	bool s;
967 
968 	pages = iommu_num_pages(address, size, PAGE_SIZE);
969 	s     = false;
970 
971 	if (pages > 1) {
972 		/*
973 		 * If we have to flush more than one page, flush all
974 		 * TLB entries for this domain
975 		 */
976 		address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
977 		s = true;
978 	}
979 
980 	address &= PAGE_MASK;
981 
982 	memset(cmd, 0, sizeof(*cmd));
983 	cmd->data[0]  = devid;
984 	cmd->data[0] |= (qdep & 0xff) << 24;
985 	cmd->data[1]  = devid;
986 	cmd->data[2]  = lower_32_bits(address);
987 	cmd->data[3]  = upper_32_bits(address);
988 	CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
989 	if (s)
990 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
991 }
992 
build_inv_iommu_pasid(struct iommu_cmd * cmd,u16 domid,u32 pasid,u64 address,bool size)993 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, u32 pasid,
994 				  u64 address, bool size)
995 {
996 	memset(cmd, 0, sizeof(*cmd));
997 
998 	address &= ~(0xfffULL);
999 
1000 	cmd->data[0]  = pasid;
1001 	cmd->data[1]  = domid;
1002 	cmd->data[2]  = lower_32_bits(address);
1003 	cmd->data[3]  = upper_32_bits(address);
1004 	cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
1005 	cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
1006 	if (size)
1007 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
1008 	CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
1009 }
1010 
build_inv_iotlb_pasid(struct iommu_cmd * cmd,u16 devid,u32 pasid,int qdep,u64 address,bool size)1011 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, u32 pasid,
1012 				  int qdep, u64 address, bool size)
1013 {
1014 	memset(cmd, 0, sizeof(*cmd));
1015 
1016 	address &= ~(0xfffULL);
1017 
1018 	cmd->data[0]  = devid;
1019 	cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
1020 	cmd->data[0] |= (qdep  & 0xff) << 24;
1021 	cmd->data[1]  = devid;
1022 	cmd->data[1] |= (pasid & 0xff) << 16;
1023 	cmd->data[2]  = lower_32_bits(address);
1024 	cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
1025 	cmd->data[3]  = upper_32_bits(address);
1026 	if (size)
1027 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
1028 	CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
1029 }
1030 
build_complete_ppr(struct iommu_cmd * cmd,u16 devid,u32 pasid,int status,int tag,bool gn)1031 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, u32 pasid,
1032 			       int status, int tag, bool gn)
1033 {
1034 	memset(cmd, 0, sizeof(*cmd));
1035 
1036 	cmd->data[0]  = devid;
1037 	if (gn) {
1038 		cmd->data[1]  = pasid;
1039 		cmd->data[2]  = CMD_INV_IOMMU_PAGES_GN_MASK;
1040 	}
1041 	cmd->data[3]  = tag & 0x1ff;
1042 	cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
1043 
1044 	CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
1045 }
1046 
build_inv_all(struct iommu_cmd * cmd)1047 static void build_inv_all(struct iommu_cmd *cmd)
1048 {
1049 	memset(cmd, 0, sizeof(*cmd));
1050 	CMD_SET_TYPE(cmd, CMD_INV_ALL);
1051 }
1052 
build_inv_irt(struct iommu_cmd * cmd,u16 devid)1053 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1054 {
1055 	memset(cmd, 0, sizeof(*cmd));
1056 	cmd->data[0] = devid;
1057 	CMD_SET_TYPE(cmd, CMD_INV_IRT);
1058 }
1059 
1060 /*
1061  * Writes the command to the IOMMUs command buffer and informs the
1062  * hardware about the new command.
1063  */
__iommu_queue_command_sync(struct amd_iommu * iommu,struct iommu_cmd * cmd,bool sync)1064 static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1065 				      struct iommu_cmd *cmd,
1066 				      bool sync)
1067 {
1068 	unsigned int count = 0;
1069 	u32 left, next_tail;
1070 
1071 	next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1072 again:
1073 	left      = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE;
1074 
1075 	if (left <= 0x20) {
1076 		/* Skip udelay() the first time around */
1077 		if (count++) {
1078 			if (count == LOOP_TIMEOUT) {
1079 				pr_err("Command buffer timeout\n");
1080 				return -EIO;
1081 			}
1082 
1083 			udelay(1);
1084 		}
1085 
1086 		/* Update head and recheck remaining space */
1087 		iommu->cmd_buf_head = readl(iommu->mmio_base +
1088 					    MMIO_CMD_HEAD_OFFSET);
1089 
1090 		goto again;
1091 	}
1092 
1093 	copy_cmd_to_buffer(iommu, cmd);
1094 
1095 	/* Do we need to make sure all commands are processed? */
1096 	iommu->need_sync = sync;
1097 
1098 	return 0;
1099 }
1100 
iommu_queue_command_sync(struct amd_iommu * iommu,struct iommu_cmd * cmd,bool sync)1101 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1102 				    struct iommu_cmd *cmd,
1103 				    bool sync)
1104 {
1105 	unsigned long flags;
1106 	int ret;
1107 
1108 	raw_spin_lock_irqsave(&iommu->lock, flags);
1109 	ret = __iommu_queue_command_sync(iommu, cmd, sync);
1110 	raw_spin_unlock_irqrestore(&iommu->lock, flags);
1111 
1112 	return ret;
1113 }
1114 
iommu_queue_command(struct amd_iommu * iommu,struct iommu_cmd * cmd)1115 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1116 {
1117 	return iommu_queue_command_sync(iommu, cmd, true);
1118 }
1119 
1120 /*
1121  * This function queues a completion wait command into the command
1122  * buffer of an IOMMU
1123  */
iommu_completion_wait(struct amd_iommu * iommu)1124 static int iommu_completion_wait(struct amd_iommu *iommu)
1125 {
1126 	struct iommu_cmd cmd;
1127 	unsigned long flags;
1128 	int ret;
1129 	u64 data;
1130 
1131 	if (!iommu->need_sync)
1132 		return 0;
1133 
1134 	raw_spin_lock_irqsave(&iommu->lock, flags);
1135 
1136 	data = ++iommu->cmd_sem_val;
1137 	build_completion_wait(&cmd, iommu, data);
1138 
1139 	ret = __iommu_queue_command_sync(iommu, &cmd, false);
1140 	if (ret)
1141 		goto out_unlock;
1142 
1143 	ret = wait_on_sem(iommu, data);
1144 
1145 out_unlock:
1146 	raw_spin_unlock_irqrestore(&iommu->lock, flags);
1147 
1148 	return ret;
1149 }
1150 
iommu_flush_dte(struct amd_iommu * iommu,u16 devid)1151 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1152 {
1153 	struct iommu_cmd cmd;
1154 
1155 	build_inv_dte(&cmd, devid);
1156 
1157 	return iommu_queue_command(iommu, &cmd);
1158 }
1159 
amd_iommu_flush_dte_all(struct amd_iommu * iommu)1160 static void amd_iommu_flush_dte_all(struct amd_iommu *iommu)
1161 {
1162 	u32 devid;
1163 
1164 	for (devid = 0; devid <= 0xffff; ++devid)
1165 		iommu_flush_dte(iommu, devid);
1166 
1167 	iommu_completion_wait(iommu);
1168 }
1169 
1170 /*
1171  * This function uses heavy locking and may disable irqs for some time. But
1172  * this is no issue because it is only called during resume.
1173  */
amd_iommu_flush_tlb_all(struct amd_iommu * iommu)1174 static void amd_iommu_flush_tlb_all(struct amd_iommu *iommu)
1175 {
1176 	u32 dom_id;
1177 
1178 	for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1179 		struct iommu_cmd cmd;
1180 		build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1181 				      dom_id, 1);
1182 		iommu_queue_command(iommu, &cmd);
1183 	}
1184 
1185 	iommu_completion_wait(iommu);
1186 }
1187 
amd_iommu_flush_tlb_domid(struct amd_iommu * iommu,u32 dom_id)1188 static void amd_iommu_flush_tlb_domid(struct amd_iommu *iommu, u32 dom_id)
1189 {
1190 	struct iommu_cmd cmd;
1191 
1192 	build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1193 			      dom_id, 1);
1194 	iommu_queue_command(iommu, &cmd);
1195 
1196 	iommu_completion_wait(iommu);
1197 }
1198 
amd_iommu_flush_all(struct amd_iommu * iommu)1199 static void amd_iommu_flush_all(struct amd_iommu *iommu)
1200 {
1201 	struct iommu_cmd cmd;
1202 
1203 	build_inv_all(&cmd);
1204 
1205 	iommu_queue_command(iommu, &cmd);
1206 	iommu_completion_wait(iommu);
1207 }
1208 
iommu_flush_irt(struct amd_iommu * iommu,u16 devid)1209 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1210 {
1211 	struct iommu_cmd cmd;
1212 
1213 	build_inv_irt(&cmd, devid);
1214 
1215 	iommu_queue_command(iommu, &cmd);
1216 }
1217 
amd_iommu_flush_irt_all(struct amd_iommu * iommu)1218 static void amd_iommu_flush_irt_all(struct amd_iommu *iommu)
1219 {
1220 	u32 devid;
1221 
1222 	for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1223 		iommu_flush_irt(iommu, devid);
1224 
1225 	iommu_completion_wait(iommu);
1226 }
1227 
iommu_flush_all_caches(struct amd_iommu * iommu)1228 void iommu_flush_all_caches(struct amd_iommu *iommu)
1229 {
1230 	if (iommu_feature(iommu, FEATURE_IA)) {
1231 		amd_iommu_flush_all(iommu);
1232 	} else {
1233 		amd_iommu_flush_dte_all(iommu);
1234 		amd_iommu_flush_irt_all(iommu);
1235 		amd_iommu_flush_tlb_all(iommu);
1236 	}
1237 }
1238 
1239 /*
1240  * Command send function for flushing on-device TLB
1241  */
device_flush_iotlb(struct iommu_dev_data * dev_data,u64 address,size_t size)1242 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1243 			      u64 address, size_t size)
1244 {
1245 	struct amd_iommu *iommu;
1246 	struct iommu_cmd cmd;
1247 	int qdep;
1248 
1249 	qdep     = dev_data->ats.qdep;
1250 	iommu    = amd_iommu_rlookup_table[dev_data->devid];
1251 
1252 	build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1253 
1254 	return iommu_queue_command(iommu, &cmd);
1255 }
1256 
device_flush_dte_alias(struct pci_dev * pdev,u16 alias,void * data)1257 static int device_flush_dte_alias(struct pci_dev *pdev, u16 alias, void *data)
1258 {
1259 	struct amd_iommu *iommu = data;
1260 
1261 	return iommu_flush_dte(iommu, alias);
1262 }
1263 
1264 /*
1265  * Command send function for invalidating a device table entry
1266  */
device_flush_dte(struct iommu_dev_data * dev_data)1267 static int device_flush_dte(struct iommu_dev_data *dev_data)
1268 {
1269 	struct amd_iommu *iommu;
1270 	u16 alias;
1271 	int ret;
1272 
1273 	iommu = amd_iommu_rlookup_table[dev_data->devid];
1274 
1275 	if (dev_data->pdev)
1276 		ret = pci_for_each_dma_alias(dev_data->pdev,
1277 					     device_flush_dte_alias, iommu);
1278 	else
1279 		ret = iommu_flush_dte(iommu, dev_data->devid);
1280 	if (ret)
1281 		return ret;
1282 
1283 	alias = amd_iommu_alias_table[dev_data->devid];
1284 	if (alias != dev_data->devid) {
1285 		ret = iommu_flush_dte(iommu, alias);
1286 		if (ret)
1287 			return ret;
1288 	}
1289 
1290 	if (dev_data->ats.enabled)
1291 		ret = device_flush_iotlb(dev_data, 0, ~0UL);
1292 
1293 	return ret;
1294 }
1295 
1296 /*
1297  * TLB invalidation function which is called from the mapping functions.
1298  * It invalidates a single PTE if the range to flush is within a single
1299  * page. Otherwise it flushes the whole TLB of the IOMMU.
1300  */
__domain_flush_pages(struct protection_domain * domain,u64 address,size_t size,int pde)1301 static void __domain_flush_pages(struct protection_domain *domain,
1302 				 u64 address, size_t size, int pde)
1303 {
1304 	struct iommu_dev_data *dev_data;
1305 	struct iommu_cmd cmd;
1306 	int ret = 0, i;
1307 
1308 	build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1309 
1310 	for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1311 		if (!domain->dev_iommu[i])
1312 			continue;
1313 
1314 		/*
1315 		 * Devices of this domain are behind this IOMMU
1316 		 * We need a TLB flush
1317 		 */
1318 		ret |= iommu_queue_command(amd_iommus[i], &cmd);
1319 	}
1320 
1321 	list_for_each_entry(dev_data, &domain->dev_list, list) {
1322 
1323 		if (!dev_data->ats.enabled)
1324 			continue;
1325 
1326 		ret |= device_flush_iotlb(dev_data, address, size);
1327 	}
1328 
1329 	WARN_ON(ret);
1330 }
1331 
domain_flush_pages(struct protection_domain * domain,u64 address,size_t size)1332 static void domain_flush_pages(struct protection_domain *domain,
1333 			       u64 address, size_t size)
1334 {
1335 	__domain_flush_pages(domain, address, size, 0);
1336 }
1337 
1338 /* Flush the whole IO/TLB for a given protection domain - including PDE */
domain_flush_tlb_pde(struct protection_domain * domain)1339 static void domain_flush_tlb_pde(struct protection_domain *domain)
1340 {
1341 	__domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1342 }
1343 
domain_flush_complete(struct protection_domain * domain)1344 static void domain_flush_complete(struct protection_domain *domain)
1345 {
1346 	int i;
1347 
1348 	for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1349 		if (domain && !domain->dev_iommu[i])
1350 			continue;
1351 
1352 		/*
1353 		 * Devices of this domain are behind this IOMMU
1354 		 * We need to wait for completion of all commands.
1355 		 */
1356 		iommu_completion_wait(amd_iommus[i]);
1357 	}
1358 }
1359 
1360 /* Flush the not present cache if it exists */
domain_flush_np_cache(struct protection_domain * domain,dma_addr_t iova,size_t size)1361 static void domain_flush_np_cache(struct protection_domain *domain,
1362 		dma_addr_t iova, size_t size)
1363 {
1364 	if (unlikely(amd_iommu_np_cache)) {
1365 		unsigned long flags;
1366 
1367 		spin_lock_irqsave(&domain->lock, flags);
1368 		domain_flush_pages(domain, iova, size);
1369 		domain_flush_complete(domain);
1370 		spin_unlock_irqrestore(&domain->lock, flags);
1371 	}
1372 }
1373 
1374 
1375 /*
1376  * This function flushes the DTEs for all devices in domain
1377  */
domain_flush_devices(struct protection_domain * domain)1378 static void domain_flush_devices(struct protection_domain *domain)
1379 {
1380 	struct iommu_dev_data *dev_data;
1381 
1382 	list_for_each_entry(dev_data, &domain->dev_list, list)
1383 		device_flush_dte(dev_data);
1384 }
1385 
1386 /****************************************************************************
1387  *
1388  * The functions below are used the create the page table mappings for
1389  * unity mapped regions.
1390  *
1391  ****************************************************************************/
1392 
free_page_list(struct page * freelist)1393 static void free_page_list(struct page *freelist)
1394 {
1395 	while (freelist != NULL) {
1396 		unsigned long p = (unsigned long)page_address(freelist);
1397 		freelist = freelist->freelist;
1398 		free_page(p);
1399 	}
1400 }
1401 
free_pt_page(unsigned long pt,struct page * freelist)1402 static struct page *free_pt_page(unsigned long pt, struct page *freelist)
1403 {
1404 	struct page *p = virt_to_page((void *)pt);
1405 
1406 	p->freelist = freelist;
1407 
1408 	return p;
1409 }
1410 
1411 #define DEFINE_FREE_PT_FN(LVL, FN)						\
1412 static struct page *free_pt_##LVL (unsigned long __pt, struct page *freelist)	\
1413 {										\
1414 	unsigned long p;							\
1415 	u64 *pt;								\
1416 	int i;									\
1417 										\
1418 	pt = (u64 *)__pt;							\
1419 										\
1420 	for (i = 0; i < 512; ++i) {						\
1421 		/* PTE present? */						\
1422 		if (!IOMMU_PTE_PRESENT(pt[i]))					\
1423 			continue;						\
1424 										\
1425 		/* Large PTE? */						\
1426 		if (PM_PTE_LEVEL(pt[i]) == 0 ||					\
1427 		    PM_PTE_LEVEL(pt[i]) == 7)					\
1428 			continue;						\
1429 										\
1430 		p = (unsigned long)IOMMU_PTE_PAGE(pt[i]);			\
1431 		freelist = FN(p, freelist);					\
1432 	}									\
1433 										\
1434 	return free_pt_page((unsigned long)pt, freelist);			\
1435 }
1436 
DEFINE_FREE_PT_FN(l2,free_pt_page)1437 DEFINE_FREE_PT_FN(l2, free_pt_page)
1438 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1439 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1440 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1441 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1442 
1443 static struct page *free_sub_pt(unsigned long root, int mode,
1444 				struct page *freelist)
1445 {
1446 	switch (mode) {
1447 	case PAGE_MODE_NONE:
1448 	case PAGE_MODE_7_LEVEL:
1449 		break;
1450 	case PAGE_MODE_1_LEVEL:
1451 		freelist = free_pt_page(root, freelist);
1452 		break;
1453 	case PAGE_MODE_2_LEVEL:
1454 		freelist = free_pt_l2(root, freelist);
1455 		break;
1456 	case PAGE_MODE_3_LEVEL:
1457 		freelist = free_pt_l3(root, freelist);
1458 		break;
1459 	case PAGE_MODE_4_LEVEL:
1460 		freelist = free_pt_l4(root, freelist);
1461 		break;
1462 	case PAGE_MODE_5_LEVEL:
1463 		freelist = free_pt_l5(root, freelist);
1464 		break;
1465 	case PAGE_MODE_6_LEVEL:
1466 		freelist = free_pt_l6(root, freelist);
1467 		break;
1468 	default:
1469 		BUG();
1470 	}
1471 
1472 	return freelist;
1473 }
1474 
free_pagetable(struct domain_pgtable * pgtable)1475 static void free_pagetable(struct domain_pgtable *pgtable)
1476 {
1477 	struct page *freelist = NULL;
1478 	unsigned long root;
1479 
1480 	if (pgtable->mode == PAGE_MODE_NONE)
1481 		return;
1482 
1483 	BUG_ON(pgtable->mode < PAGE_MODE_NONE ||
1484 	       pgtable->mode > PAGE_MODE_6_LEVEL);
1485 
1486 	root = (unsigned long)pgtable->root;
1487 	freelist = free_sub_pt(root, pgtable->mode, freelist);
1488 
1489 	free_page_list(freelist);
1490 }
1491 
1492 /*
1493  * This function is used to add another level to an IO page table. Adding
1494  * another level increases the size of the address space by 9 bits to a size up
1495  * to 64 bits.
1496  */
increase_address_space(struct protection_domain * domain,unsigned long address,gfp_t gfp)1497 static bool increase_address_space(struct protection_domain *domain,
1498 				   unsigned long address,
1499 				   gfp_t gfp)
1500 {
1501 	struct domain_pgtable pgtable;
1502 	unsigned long flags;
1503 	bool ret = true;
1504 	u64 *pte;
1505 
1506 	pte = (void *)get_zeroed_page(gfp);
1507 	if (!pte)
1508 		return false;
1509 
1510 	spin_lock_irqsave(&domain->lock, flags);
1511 
1512 	amd_iommu_domain_get_pgtable(domain, &pgtable);
1513 
1514 	if (address <= PM_LEVEL_SIZE(pgtable.mode))
1515 		goto out;
1516 
1517 	ret = false;
1518 	if (WARN_ON_ONCE(pgtable.mode == PAGE_MODE_6_LEVEL))
1519 		goto out;
1520 
1521 	*pte = PM_LEVEL_PDE(pgtable.mode, iommu_virt_to_phys(pgtable.root));
1522 
1523 	pgtable.root  = pte;
1524 	pgtable.mode += 1;
1525 	update_and_flush_device_table(domain, &pgtable);
1526 	domain_flush_complete(domain);
1527 
1528 	/*
1529 	 * Device Table needs to be updated and flushed before the new root can
1530 	 * be published.
1531 	 */
1532 	amd_iommu_domain_set_pgtable(domain, pte, pgtable.mode);
1533 
1534 	pte = NULL;
1535 	ret = true;
1536 
1537 out:
1538 	spin_unlock_irqrestore(&domain->lock, flags);
1539 	free_page((unsigned long)pte);
1540 
1541 	return ret;
1542 }
1543 
alloc_pte(struct protection_domain * domain,unsigned long address,unsigned long page_size,u64 ** pte_page,gfp_t gfp,bool * updated)1544 static u64 *alloc_pte(struct protection_domain *domain,
1545 		      unsigned long address,
1546 		      unsigned long page_size,
1547 		      u64 **pte_page,
1548 		      gfp_t gfp,
1549 		      bool *updated)
1550 {
1551 	struct domain_pgtable pgtable;
1552 	int level, end_lvl;
1553 	u64 *pte, *page;
1554 
1555 	BUG_ON(!is_power_of_2(page_size));
1556 
1557 	amd_iommu_domain_get_pgtable(domain, &pgtable);
1558 
1559 	while (address > PM_LEVEL_SIZE(pgtable.mode)) {
1560 		/*
1561 		 * Return an error if there is no memory to update the
1562 		 * page-table.
1563 		 */
1564 		if (!increase_address_space(domain, address, gfp))
1565 			return NULL;
1566 
1567 		/* Read new values to check if update was successful */
1568 		amd_iommu_domain_get_pgtable(domain, &pgtable);
1569 	}
1570 
1571 
1572 	level   = pgtable.mode - 1;
1573 	pte     = &pgtable.root[PM_LEVEL_INDEX(level, address)];
1574 	address = PAGE_SIZE_ALIGN(address, page_size);
1575 	end_lvl = PAGE_SIZE_LEVEL(page_size);
1576 
1577 	while (level > end_lvl) {
1578 		u64 __pte, __npte;
1579 		int pte_level;
1580 
1581 		__pte     = *pte;
1582 		pte_level = PM_PTE_LEVEL(__pte);
1583 
1584 		/*
1585 		 * If we replace a series of large PTEs, we need
1586 		 * to tear down all of them.
1587 		 */
1588 		if (IOMMU_PTE_PRESENT(__pte) &&
1589 		    pte_level == PAGE_MODE_7_LEVEL) {
1590 			unsigned long count, i;
1591 			u64 *lpte;
1592 
1593 			lpte = first_pte_l7(pte, NULL, &count);
1594 
1595 			/*
1596 			 * Unmap the replicated PTEs that still match the
1597 			 * original large mapping
1598 			 */
1599 			for (i = 0; i < count; ++i)
1600 				cmpxchg64(&lpte[i], __pte, 0ULL);
1601 
1602 			*updated = true;
1603 			continue;
1604 		}
1605 
1606 		if (!IOMMU_PTE_PRESENT(__pte) ||
1607 		    pte_level == PAGE_MODE_NONE) {
1608 			page = (u64 *)get_zeroed_page(gfp);
1609 
1610 			if (!page)
1611 				return NULL;
1612 
1613 			__npte = PM_LEVEL_PDE(level, iommu_virt_to_phys(page));
1614 
1615 			/* pte could have been changed somewhere. */
1616 			if (cmpxchg64(pte, __pte, __npte) != __pte)
1617 				free_page((unsigned long)page);
1618 			else if (IOMMU_PTE_PRESENT(__pte))
1619 				*updated = true;
1620 
1621 			continue;
1622 		}
1623 
1624 		/* No level skipping support yet */
1625 		if (pte_level != level)
1626 			return NULL;
1627 
1628 		level -= 1;
1629 
1630 		pte = IOMMU_PTE_PAGE(__pte);
1631 
1632 		if (pte_page && level == end_lvl)
1633 			*pte_page = pte;
1634 
1635 		pte = &pte[PM_LEVEL_INDEX(level, address)];
1636 	}
1637 
1638 	return pte;
1639 }
1640 
1641 /*
1642  * This function checks if there is a PTE for a given dma address. If
1643  * there is one, it returns the pointer to it.
1644  */
fetch_pte(struct protection_domain * domain,unsigned long address,unsigned long * page_size)1645 static u64 *fetch_pte(struct protection_domain *domain,
1646 		      unsigned long address,
1647 		      unsigned long *page_size)
1648 {
1649 	struct domain_pgtable pgtable;
1650 	int level;
1651 	u64 *pte;
1652 
1653 	*page_size = 0;
1654 
1655 	amd_iommu_domain_get_pgtable(domain, &pgtable);
1656 
1657 	if (address > PM_LEVEL_SIZE(pgtable.mode))
1658 		return NULL;
1659 
1660 	level	   =  pgtable.mode - 1;
1661 	pte	   = &pgtable.root[PM_LEVEL_INDEX(level, address)];
1662 	*page_size =  PTE_LEVEL_PAGE_SIZE(level);
1663 
1664 	while (level > 0) {
1665 
1666 		/* Not Present */
1667 		if (!IOMMU_PTE_PRESENT(*pte))
1668 			return NULL;
1669 
1670 		/* Large PTE */
1671 		if (PM_PTE_LEVEL(*pte) == 7 ||
1672 		    PM_PTE_LEVEL(*pte) == 0)
1673 			break;
1674 
1675 		/* No level skipping support yet */
1676 		if (PM_PTE_LEVEL(*pte) != level)
1677 			return NULL;
1678 
1679 		level -= 1;
1680 
1681 		/* Walk to the next level */
1682 		pte	   = IOMMU_PTE_PAGE(*pte);
1683 		pte	   = &pte[PM_LEVEL_INDEX(level, address)];
1684 		*page_size = PTE_LEVEL_PAGE_SIZE(level);
1685 	}
1686 
1687 	/*
1688 	 * If we have a series of large PTEs, make
1689 	 * sure to return a pointer to the first one.
1690 	 */
1691 	if (PM_PTE_LEVEL(*pte) == PAGE_MODE_7_LEVEL)
1692 		pte = first_pte_l7(pte, page_size, NULL);
1693 
1694 	return pte;
1695 }
1696 
free_clear_pte(u64 * pte,u64 pteval,struct page * freelist)1697 static struct page *free_clear_pte(u64 *pte, u64 pteval, struct page *freelist)
1698 {
1699 	unsigned long pt;
1700 	int mode;
1701 
1702 	while (cmpxchg64(pte, pteval, 0) != pteval) {
1703 		pr_warn("AMD-Vi: IOMMU pte changed since we read it\n");
1704 		pteval = *pte;
1705 	}
1706 
1707 	if (!IOMMU_PTE_PRESENT(pteval))
1708 		return freelist;
1709 
1710 	pt   = (unsigned long)IOMMU_PTE_PAGE(pteval);
1711 	mode = IOMMU_PTE_MODE(pteval);
1712 
1713 	return free_sub_pt(pt, mode, freelist);
1714 }
1715 
1716 /*
1717  * Generic mapping functions. It maps a physical address into a DMA
1718  * address space. It allocates the page table pages if necessary.
1719  * In the future it can be extended to a generic mapping function
1720  * supporting all features of AMD IOMMU page tables like level skipping
1721  * and full 64 bit address spaces.
1722  */
iommu_map_page(struct protection_domain * dom,unsigned long bus_addr,unsigned long phys_addr,unsigned long page_size,int prot,gfp_t gfp)1723 static int iommu_map_page(struct protection_domain *dom,
1724 			  unsigned long bus_addr,
1725 			  unsigned long phys_addr,
1726 			  unsigned long page_size,
1727 			  int prot,
1728 			  gfp_t gfp)
1729 {
1730 	struct page *freelist = NULL;
1731 	bool updated = false;
1732 	u64 __pte, *pte;
1733 	int ret, i, count;
1734 
1735 	BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1736 	BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1737 
1738 	ret = -EINVAL;
1739 	if (!(prot & IOMMU_PROT_MASK))
1740 		goto out;
1741 
1742 	count = PAGE_SIZE_PTE_COUNT(page_size);
1743 	pte   = alloc_pte(dom, bus_addr, page_size, NULL, gfp, &updated);
1744 
1745 	ret = -ENOMEM;
1746 	if (!pte)
1747 		goto out;
1748 
1749 	for (i = 0; i < count; ++i)
1750 		freelist = free_clear_pte(&pte[i], pte[i], freelist);
1751 
1752 	if (freelist != NULL)
1753 		updated = true;
1754 
1755 	if (count > 1) {
1756 		__pte = PAGE_SIZE_PTE(__sme_set(phys_addr), page_size);
1757 		__pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_PR | IOMMU_PTE_FC;
1758 	} else
1759 		__pte = __sme_set(phys_addr) | IOMMU_PTE_PR | IOMMU_PTE_FC;
1760 
1761 	if (prot & IOMMU_PROT_IR)
1762 		__pte |= IOMMU_PTE_IR;
1763 	if (prot & IOMMU_PROT_IW)
1764 		__pte |= IOMMU_PTE_IW;
1765 
1766 	for (i = 0; i < count; ++i)
1767 		pte[i] = __pte;
1768 
1769 	ret = 0;
1770 
1771 out:
1772 	if (updated) {
1773 		unsigned long flags;
1774 
1775 		spin_lock_irqsave(&dom->lock, flags);
1776 		/*
1777 		 * Flush domain TLB(s) and wait for completion. Any Device-Table
1778 		 * Updates and flushing already happened in
1779 		 * increase_address_space().
1780 		 */
1781 		domain_flush_tlb_pde(dom);
1782 		domain_flush_complete(dom);
1783 		spin_unlock_irqrestore(&dom->lock, flags);
1784 	}
1785 
1786 	/* Everything flushed out, free pages now */
1787 	free_page_list(freelist);
1788 
1789 	return ret;
1790 }
1791 
iommu_unmap_page(struct protection_domain * dom,unsigned long bus_addr,unsigned long page_size)1792 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1793 				      unsigned long bus_addr,
1794 				      unsigned long page_size)
1795 {
1796 	unsigned long long unmapped;
1797 	unsigned long unmap_size;
1798 	u64 *pte;
1799 
1800 	BUG_ON(!is_power_of_2(page_size));
1801 
1802 	unmapped = 0;
1803 
1804 	while (unmapped < page_size) {
1805 
1806 		pte = fetch_pte(dom, bus_addr, &unmap_size);
1807 
1808 		if (pte) {
1809 			int i, count;
1810 
1811 			count = PAGE_SIZE_PTE_COUNT(unmap_size);
1812 			for (i = 0; i < count; i++)
1813 				pte[i] = 0ULL;
1814 		}
1815 
1816 		bus_addr  = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1817 		unmapped += unmap_size;
1818 	}
1819 
1820 	BUG_ON(unmapped && !is_power_of_2(unmapped));
1821 
1822 	return unmapped;
1823 }
1824 
1825 /****************************************************************************
1826  *
1827  * The next functions belong to the domain allocation. A domain is
1828  * allocated for every IOMMU as the default domain. If device isolation
1829  * is enabled, every device get its own domain. The most important thing
1830  * about domains is the page table mapping the DMA address space they
1831  * contain.
1832  *
1833  ****************************************************************************/
1834 
domain_id_alloc(void)1835 static u16 domain_id_alloc(void)
1836 {
1837 	int id;
1838 
1839 	spin_lock(&pd_bitmap_lock);
1840 	id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1841 	BUG_ON(id == 0);
1842 	if (id > 0 && id < MAX_DOMAIN_ID)
1843 		__set_bit(id, amd_iommu_pd_alloc_bitmap);
1844 	else
1845 		id = 0;
1846 	spin_unlock(&pd_bitmap_lock);
1847 
1848 	return id;
1849 }
1850 
domain_id_free(int id)1851 static void domain_id_free(int id)
1852 {
1853 	spin_lock(&pd_bitmap_lock);
1854 	if (id > 0 && id < MAX_DOMAIN_ID)
1855 		__clear_bit(id, amd_iommu_pd_alloc_bitmap);
1856 	spin_unlock(&pd_bitmap_lock);
1857 }
1858 
free_gcr3_tbl_level1(u64 * tbl)1859 static void free_gcr3_tbl_level1(u64 *tbl)
1860 {
1861 	u64 *ptr;
1862 	int i;
1863 
1864 	for (i = 0; i < 512; ++i) {
1865 		if (!(tbl[i] & GCR3_VALID))
1866 			continue;
1867 
1868 		ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1869 
1870 		free_page((unsigned long)ptr);
1871 	}
1872 }
1873 
free_gcr3_tbl_level2(u64 * tbl)1874 static void free_gcr3_tbl_level2(u64 *tbl)
1875 {
1876 	u64 *ptr;
1877 	int i;
1878 
1879 	for (i = 0; i < 512; ++i) {
1880 		if (!(tbl[i] & GCR3_VALID))
1881 			continue;
1882 
1883 		ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1884 
1885 		free_gcr3_tbl_level1(ptr);
1886 	}
1887 }
1888 
free_gcr3_table(struct protection_domain * domain)1889 static void free_gcr3_table(struct protection_domain *domain)
1890 {
1891 	if (domain->glx == 2)
1892 		free_gcr3_tbl_level2(domain->gcr3_tbl);
1893 	else if (domain->glx == 1)
1894 		free_gcr3_tbl_level1(domain->gcr3_tbl);
1895 	else
1896 		BUG_ON(domain->glx != 0);
1897 
1898 	free_page((unsigned long)domain->gcr3_tbl);
1899 }
1900 
set_dte_entry(u16 devid,struct protection_domain * domain,struct domain_pgtable * pgtable,bool ats,bool ppr)1901 static void set_dte_entry(u16 devid, struct protection_domain *domain,
1902 			  struct domain_pgtable *pgtable,
1903 			  bool ats, bool ppr)
1904 {
1905 	u64 pte_root = 0;
1906 	u64 flags = 0;
1907 	u32 old_domid;
1908 
1909 	if (pgtable->mode != PAGE_MODE_NONE)
1910 		pte_root = iommu_virt_to_phys(pgtable->root);
1911 
1912 	pte_root |= (pgtable->mode & DEV_ENTRY_MODE_MASK)
1913 		    << DEV_ENTRY_MODE_SHIFT;
1914 	pte_root |= DTE_FLAG_IR | DTE_FLAG_IW | DTE_FLAG_V | DTE_FLAG_TV;
1915 
1916 	flags = amd_iommu_dev_table[devid].data[1];
1917 
1918 	if (ats)
1919 		flags |= DTE_FLAG_IOTLB;
1920 
1921 	if (ppr) {
1922 		struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
1923 
1924 		if (iommu_feature(iommu, FEATURE_EPHSUP))
1925 			pte_root |= 1ULL << DEV_ENTRY_PPR;
1926 	}
1927 
1928 	if (domain->flags & PD_IOMMUV2_MASK) {
1929 		u64 gcr3 = iommu_virt_to_phys(domain->gcr3_tbl);
1930 		u64 glx  = domain->glx;
1931 		u64 tmp;
1932 
1933 		pte_root |= DTE_FLAG_GV;
1934 		pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1935 
1936 		/* First mask out possible old values for GCR3 table */
1937 		tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1938 		flags    &= ~tmp;
1939 
1940 		tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1941 		flags    &= ~tmp;
1942 
1943 		/* Encode GCR3 table into DTE */
1944 		tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1945 		pte_root |= tmp;
1946 
1947 		tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1948 		flags    |= tmp;
1949 
1950 		tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1951 		flags    |= tmp;
1952 	}
1953 
1954 	flags &= ~DEV_DOMID_MASK;
1955 	flags |= domain->id;
1956 
1957 	old_domid = amd_iommu_dev_table[devid].data[1] & DEV_DOMID_MASK;
1958 	amd_iommu_dev_table[devid].data[1]  = flags;
1959 	amd_iommu_dev_table[devid].data[0]  = pte_root;
1960 
1961 	/*
1962 	 * A kdump kernel might be replacing a domain ID that was copied from
1963 	 * the previous kernel--if so, it needs to flush the translation cache
1964 	 * entries for the old domain ID that is being overwritten
1965 	 */
1966 	if (old_domid) {
1967 		struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
1968 
1969 		amd_iommu_flush_tlb_domid(iommu, old_domid);
1970 	}
1971 }
1972 
clear_dte_entry(u16 devid)1973 static void clear_dte_entry(u16 devid)
1974 {
1975 	/* remove entry from the device table seen by the hardware */
1976 	amd_iommu_dev_table[devid].data[0]  = DTE_FLAG_V | DTE_FLAG_TV;
1977 	amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
1978 
1979 	amd_iommu_apply_erratum_63(devid);
1980 }
1981 
do_attach(struct iommu_dev_data * dev_data,struct protection_domain * domain)1982 static void do_attach(struct iommu_dev_data *dev_data,
1983 		      struct protection_domain *domain)
1984 {
1985 	struct domain_pgtable pgtable;
1986 	struct amd_iommu *iommu;
1987 	bool ats;
1988 
1989 	iommu = amd_iommu_rlookup_table[dev_data->devid];
1990 	ats   = dev_data->ats.enabled;
1991 
1992 	/* Update data structures */
1993 	dev_data->domain = domain;
1994 	list_add(&dev_data->list, &domain->dev_list);
1995 
1996 	/* Do reference counting */
1997 	domain->dev_iommu[iommu->index] += 1;
1998 	domain->dev_cnt                 += 1;
1999 
2000 	/* Update device table */
2001 	amd_iommu_domain_get_pgtable(domain, &pgtable);
2002 	set_dte_entry(dev_data->devid, domain, &pgtable,
2003 		      ats, dev_data->iommu_v2);
2004 	clone_aliases(dev_data->pdev);
2005 
2006 	device_flush_dte(dev_data);
2007 }
2008 
do_detach(struct iommu_dev_data * dev_data)2009 static void do_detach(struct iommu_dev_data *dev_data)
2010 {
2011 	struct protection_domain *domain = dev_data->domain;
2012 	struct amd_iommu *iommu;
2013 
2014 	iommu = amd_iommu_rlookup_table[dev_data->devid];
2015 
2016 	/* Update data structures */
2017 	dev_data->domain = NULL;
2018 	list_del(&dev_data->list);
2019 	clear_dte_entry(dev_data->devid);
2020 	clone_aliases(dev_data->pdev);
2021 
2022 	/* Flush the DTE entry */
2023 	device_flush_dte(dev_data);
2024 
2025 	/* Flush IOTLB */
2026 	domain_flush_tlb_pde(domain);
2027 
2028 	/* Wait for the flushes to finish */
2029 	domain_flush_complete(domain);
2030 
2031 	/* decrease reference counters - needs to happen after the flushes */
2032 	domain->dev_iommu[iommu->index] -= 1;
2033 	domain->dev_cnt                 -= 1;
2034 }
2035 
pdev_iommuv2_disable(struct pci_dev * pdev)2036 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2037 {
2038 	pci_disable_ats(pdev);
2039 	pci_disable_pri(pdev);
2040 	pci_disable_pasid(pdev);
2041 }
2042 
2043 /* FIXME: Change generic reset-function to do the same */
pri_reset_while_enabled(struct pci_dev * pdev)2044 static int pri_reset_while_enabled(struct pci_dev *pdev)
2045 {
2046 	u16 control;
2047 	int pos;
2048 
2049 	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2050 	if (!pos)
2051 		return -EINVAL;
2052 
2053 	pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2054 	control |= PCI_PRI_CTRL_RESET;
2055 	pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2056 
2057 	return 0;
2058 }
2059 
pdev_iommuv2_enable(struct pci_dev * pdev)2060 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2061 {
2062 	bool reset_enable;
2063 	int reqs, ret;
2064 
2065 	/* FIXME: Hardcode number of outstanding requests for now */
2066 	reqs = 32;
2067 	if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2068 		reqs = 1;
2069 	reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2070 
2071 	/* Only allow access to user-accessible pages */
2072 	ret = pci_enable_pasid(pdev, 0);
2073 	if (ret)
2074 		goto out_err;
2075 
2076 	/* First reset the PRI state of the device */
2077 	ret = pci_reset_pri(pdev);
2078 	if (ret)
2079 		goto out_err;
2080 
2081 	/* Enable PRI */
2082 	ret = pci_enable_pri(pdev, reqs);
2083 	if (ret)
2084 		goto out_err;
2085 
2086 	if (reset_enable) {
2087 		ret = pri_reset_while_enabled(pdev);
2088 		if (ret)
2089 			goto out_err;
2090 	}
2091 
2092 	ret = pci_enable_ats(pdev, PAGE_SHIFT);
2093 	if (ret)
2094 		goto out_err;
2095 
2096 	return 0;
2097 
2098 out_err:
2099 	pci_disable_pri(pdev);
2100 	pci_disable_pasid(pdev);
2101 
2102 	return ret;
2103 }
2104 
2105 /*
2106  * If a device is not yet associated with a domain, this function makes the
2107  * device visible in the domain
2108  */
attach_device(struct device * dev,struct protection_domain * domain)2109 static int attach_device(struct device *dev,
2110 			 struct protection_domain *domain)
2111 {
2112 	struct iommu_dev_data *dev_data;
2113 	struct pci_dev *pdev;
2114 	unsigned long flags;
2115 	int ret;
2116 
2117 	spin_lock_irqsave(&domain->lock, flags);
2118 
2119 	dev_data = dev_iommu_priv_get(dev);
2120 
2121 	spin_lock(&dev_data->lock);
2122 
2123 	ret = -EBUSY;
2124 	if (dev_data->domain != NULL)
2125 		goto out;
2126 
2127 	if (!dev_is_pci(dev))
2128 		goto skip_ats_check;
2129 
2130 	pdev = to_pci_dev(dev);
2131 	if (domain->flags & PD_IOMMUV2_MASK) {
2132 		struct iommu_domain *def_domain = iommu_get_dma_domain(dev);
2133 
2134 		ret = -EINVAL;
2135 		if (def_domain->type != IOMMU_DOMAIN_IDENTITY)
2136 			goto out;
2137 
2138 		if (dev_data->iommu_v2) {
2139 			if (pdev_iommuv2_enable(pdev) != 0)
2140 				goto out;
2141 
2142 			dev_data->ats.enabled = true;
2143 			dev_data->ats.qdep    = pci_ats_queue_depth(pdev);
2144 			dev_data->pri_tlp     = pci_prg_resp_pasid_required(pdev);
2145 		}
2146 	} else if (amd_iommu_iotlb_sup &&
2147 		   pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2148 		dev_data->ats.enabled = true;
2149 		dev_data->ats.qdep    = pci_ats_queue_depth(pdev);
2150 	}
2151 
2152 skip_ats_check:
2153 	ret = 0;
2154 
2155 	do_attach(dev_data, domain);
2156 
2157 	/*
2158 	 * We might boot into a crash-kernel here. The crashed kernel
2159 	 * left the caches in the IOMMU dirty. So we have to flush
2160 	 * here to evict all dirty stuff.
2161 	 */
2162 	domain_flush_tlb_pde(domain);
2163 
2164 	domain_flush_complete(domain);
2165 
2166 out:
2167 	spin_unlock(&dev_data->lock);
2168 
2169 	spin_unlock_irqrestore(&domain->lock, flags);
2170 
2171 	return ret;
2172 }
2173 
2174 /*
2175  * Removes a device from a protection domain (with devtable_lock held)
2176  */
detach_device(struct device * dev)2177 static void detach_device(struct device *dev)
2178 {
2179 	struct protection_domain *domain;
2180 	struct iommu_dev_data *dev_data;
2181 	unsigned long flags;
2182 
2183 	dev_data = dev_iommu_priv_get(dev);
2184 	domain   = dev_data->domain;
2185 
2186 	spin_lock_irqsave(&domain->lock, flags);
2187 
2188 	spin_lock(&dev_data->lock);
2189 
2190 	/*
2191 	 * First check if the device is still attached. It might already
2192 	 * be detached from its domain because the generic
2193 	 * iommu_detach_group code detached it and we try again here in
2194 	 * our alias handling.
2195 	 */
2196 	if (WARN_ON(!dev_data->domain))
2197 		goto out;
2198 
2199 	do_detach(dev_data);
2200 
2201 	if (!dev_is_pci(dev))
2202 		goto out;
2203 
2204 	if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2205 		pdev_iommuv2_disable(to_pci_dev(dev));
2206 	else if (dev_data->ats.enabled)
2207 		pci_disable_ats(to_pci_dev(dev));
2208 
2209 	dev_data->ats.enabled = false;
2210 
2211 out:
2212 	spin_unlock(&dev_data->lock);
2213 
2214 	spin_unlock_irqrestore(&domain->lock, flags);
2215 }
2216 
amd_iommu_probe_device(struct device * dev)2217 static struct iommu_device *amd_iommu_probe_device(struct device *dev)
2218 {
2219 	struct iommu_device *iommu_dev;
2220 	struct amd_iommu *iommu;
2221 	int ret, devid;
2222 
2223 	if (!check_device(dev))
2224 		return ERR_PTR(-ENODEV);
2225 
2226 	devid = get_device_id(dev);
2227 	if (devid < 0)
2228 		return ERR_PTR(devid);
2229 
2230 	iommu = amd_iommu_rlookup_table[devid];
2231 
2232 	if (dev_iommu_priv_get(dev))
2233 		return &iommu->iommu;
2234 
2235 	ret = iommu_init_device(dev);
2236 	if (ret) {
2237 		if (ret != -ENOTSUPP)
2238 			dev_err(dev, "Failed to initialize - trying to proceed anyway\n");
2239 		iommu_dev = ERR_PTR(ret);
2240 		iommu_ignore_device(dev);
2241 	} else {
2242 		amd_iommu_set_pci_msi_domain(dev, iommu);
2243 		iommu_dev = &iommu->iommu;
2244 	}
2245 
2246 	iommu_completion_wait(iommu);
2247 
2248 	return iommu_dev;
2249 }
2250 
amd_iommu_probe_finalize(struct device * dev)2251 static void amd_iommu_probe_finalize(struct device *dev)
2252 {
2253 	struct iommu_domain *domain;
2254 
2255 	/* Domains are initialized for this device - have a look what we ended up with */
2256 	domain = iommu_get_domain_for_dev(dev);
2257 	if (domain->type == IOMMU_DOMAIN_DMA)
2258 		iommu_setup_dma_ops(dev, IOVA_START_PFN << PAGE_SHIFT, 0);
2259 }
2260 
amd_iommu_release_device(struct device * dev)2261 static void amd_iommu_release_device(struct device *dev)
2262 {
2263 	int devid = get_device_id(dev);
2264 	struct amd_iommu *iommu;
2265 
2266 	if (!check_device(dev))
2267 		return;
2268 
2269 	iommu = amd_iommu_rlookup_table[devid];
2270 
2271 	amd_iommu_uninit_device(dev);
2272 	iommu_completion_wait(iommu);
2273 }
2274 
amd_iommu_device_group(struct device * dev)2275 static struct iommu_group *amd_iommu_device_group(struct device *dev)
2276 {
2277 	if (dev_is_pci(dev))
2278 		return pci_device_group(dev);
2279 
2280 	return acpihid_device_group(dev);
2281 }
2282 
amd_iommu_domain_get_attr(struct iommu_domain * domain,enum iommu_attr attr,void * data)2283 static int amd_iommu_domain_get_attr(struct iommu_domain *domain,
2284 		enum iommu_attr attr, void *data)
2285 {
2286 	switch (domain->type) {
2287 	case IOMMU_DOMAIN_UNMANAGED:
2288 		return -ENODEV;
2289 	case IOMMU_DOMAIN_DMA:
2290 		switch (attr) {
2291 		case DOMAIN_ATTR_DMA_USE_FLUSH_QUEUE:
2292 			*(int *)data = !amd_iommu_unmap_flush;
2293 			return 0;
2294 		default:
2295 			return -ENODEV;
2296 		}
2297 		break;
2298 	default:
2299 		return -EINVAL;
2300 	}
2301 }
2302 
2303 /*****************************************************************************
2304  *
2305  * The next functions belong to the dma_ops mapping/unmapping code.
2306  *
2307  *****************************************************************************/
2308 
update_device_table(struct protection_domain * domain,struct domain_pgtable * pgtable)2309 static void update_device_table(struct protection_domain *domain,
2310 				struct domain_pgtable *pgtable)
2311 {
2312 	struct iommu_dev_data *dev_data;
2313 
2314 	list_for_each_entry(dev_data, &domain->dev_list, list) {
2315 		set_dte_entry(dev_data->devid, domain, pgtable,
2316 			      dev_data->ats.enabled, dev_data->iommu_v2);
2317 		clone_aliases(dev_data->pdev);
2318 	}
2319 }
2320 
update_and_flush_device_table(struct protection_domain * domain,struct domain_pgtable * pgtable)2321 static void update_and_flush_device_table(struct protection_domain *domain,
2322 					  struct domain_pgtable *pgtable)
2323 {
2324 	update_device_table(domain, pgtable);
2325 	domain_flush_devices(domain);
2326 }
2327 
update_domain(struct protection_domain * domain)2328 static void update_domain(struct protection_domain *domain)
2329 {
2330 	struct domain_pgtable pgtable;
2331 
2332 	/* Update device table */
2333 	amd_iommu_domain_get_pgtable(domain, &pgtable);
2334 	update_and_flush_device_table(domain, &pgtable);
2335 
2336 	/* Flush domain TLB(s) and wait for completion */
2337 	domain_flush_tlb_pde(domain);
2338 	domain_flush_complete(domain);
2339 }
2340 
amd_iommu_init_api(void)2341 int __init amd_iommu_init_api(void)
2342 {
2343 	int ret, err = 0;
2344 
2345 	ret = iova_cache_get();
2346 	if (ret)
2347 		return ret;
2348 
2349 	err = bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2350 	if (err)
2351 		return err;
2352 #ifdef CONFIG_ARM_AMBA
2353 	err = bus_set_iommu(&amba_bustype, &amd_iommu_ops);
2354 	if (err)
2355 		return err;
2356 #endif
2357 	err = bus_set_iommu(&platform_bus_type, &amd_iommu_ops);
2358 	if (err)
2359 		return err;
2360 
2361 	return 0;
2362 }
2363 
amd_iommu_init_dma_ops(void)2364 int __init amd_iommu_init_dma_ops(void)
2365 {
2366 	swiotlb        = (iommu_default_passthrough() || sme_me_mask) ? 1 : 0;
2367 
2368 	if (amd_iommu_unmap_flush)
2369 		pr_info("IO/TLB flush on unmap enabled\n");
2370 	else
2371 		pr_info("Lazy IO/TLB flushing enabled\n");
2372 
2373 	return 0;
2374 
2375 }
2376 
2377 /*****************************************************************************
2378  *
2379  * The following functions belong to the exported interface of AMD IOMMU
2380  *
2381  * This interface allows access to lower level functions of the IOMMU
2382  * like protection domain handling and assignement of devices to domains
2383  * which is not possible with the dma_ops interface.
2384  *
2385  *****************************************************************************/
2386 
cleanup_domain(struct protection_domain * domain)2387 static void cleanup_domain(struct protection_domain *domain)
2388 {
2389 	struct iommu_dev_data *entry;
2390 	unsigned long flags;
2391 
2392 	spin_lock_irqsave(&domain->lock, flags);
2393 
2394 	while (!list_empty(&domain->dev_list)) {
2395 		entry = list_first_entry(&domain->dev_list,
2396 					 struct iommu_dev_data, list);
2397 		BUG_ON(!entry->domain);
2398 		do_detach(entry);
2399 	}
2400 
2401 	spin_unlock_irqrestore(&domain->lock, flags);
2402 }
2403 
protection_domain_free(struct protection_domain * domain)2404 static void protection_domain_free(struct protection_domain *domain)
2405 {
2406 	struct domain_pgtable pgtable;
2407 
2408 	if (!domain)
2409 		return;
2410 
2411 	if (domain->id)
2412 		domain_id_free(domain->id);
2413 
2414 	amd_iommu_domain_get_pgtable(domain, &pgtable);
2415 	amd_iommu_domain_clr_pt_root(domain);
2416 	free_pagetable(&pgtable);
2417 
2418 	kfree(domain);
2419 }
2420 
protection_domain_init(struct protection_domain * domain,int mode)2421 static int protection_domain_init(struct protection_domain *domain, int mode)
2422 {
2423 	u64 *pt_root = NULL;
2424 
2425 	BUG_ON(mode < PAGE_MODE_NONE || mode > PAGE_MODE_6_LEVEL);
2426 
2427 	spin_lock_init(&domain->lock);
2428 	domain->id = domain_id_alloc();
2429 	if (!domain->id)
2430 		return -ENOMEM;
2431 	INIT_LIST_HEAD(&domain->dev_list);
2432 
2433 	if (mode != PAGE_MODE_NONE) {
2434 		pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2435 		if (!pt_root)
2436 			return -ENOMEM;
2437 	}
2438 
2439 	amd_iommu_domain_set_pgtable(domain, pt_root, mode);
2440 
2441 	return 0;
2442 }
2443 
protection_domain_alloc(int mode)2444 static struct protection_domain *protection_domain_alloc(int mode)
2445 {
2446 	struct protection_domain *domain;
2447 
2448 	domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2449 	if (!domain)
2450 		return NULL;
2451 
2452 	if (protection_domain_init(domain, mode))
2453 		goto out_err;
2454 
2455 	return domain;
2456 
2457 out_err:
2458 	kfree(domain);
2459 
2460 	return NULL;
2461 }
2462 
amd_iommu_domain_alloc(unsigned type)2463 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2464 {
2465 	struct protection_domain *domain;
2466 	int mode = DEFAULT_PGTABLE_LEVEL;
2467 
2468 	if (type == IOMMU_DOMAIN_IDENTITY)
2469 		mode = PAGE_MODE_NONE;
2470 
2471 	domain = protection_domain_alloc(mode);
2472 	if (!domain)
2473 		return NULL;
2474 
2475 	domain->domain.geometry.aperture_start = 0;
2476 	domain->domain.geometry.aperture_end   = ~0ULL;
2477 	domain->domain.geometry.force_aperture = true;
2478 
2479 	if (type == IOMMU_DOMAIN_DMA &&
2480 	    iommu_get_dma_cookie(&domain->domain) == -ENOMEM)
2481 		goto free_domain;
2482 
2483 	return &domain->domain;
2484 
2485 free_domain:
2486 	protection_domain_free(domain);
2487 
2488 	return NULL;
2489 }
2490 
amd_iommu_domain_free(struct iommu_domain * dom)2491 static void amd_iommu_domain_free(struct iommu_domain *dom)
2492 {
2493 	struct protection_domain *domain;
2494 
2495 	domain = to_pdomain(dom);
2496 
2497 	if (domain->dev_cnt > 0)
2498 		cleanup_domain(domain);
2499 
2500 	BUG_ON(domain->dev_cnt != 0);
2501 
2502 	if (!dom)
2503 		return;
2504 
2505 	if (dom->type == IOMMU_DOMAIN_DMA)
2506 		iommu_put_dma_cookie(&domain->domain);
2507 
2508 	if (domain->flags & PD_IOMMUV2_MASK)
2509 		free_gcr3_table(domain);
2510 
2511 	protection_domain_free(domain);
2512 }
2513 
amd_iommu_detach_device(struct iommu_domain * dom,struct device * dev)2514 static void amd_iommu_detach_device(struct iommu_domain *dom,
2515 				    struct device *dev)
2516 {
2517 	struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2518 	struct amd_iommu *iommu;
2519 	int devid;
2520 
2521 	if (!check_device(dev))
2522 		return;
2523 
2524 	devid = get_device_id(dev);
2525 	if (devid < 0)
2526 		return;
2527 
2528 	if (dev_data->domain != NULL)
2529 		detach_device(dev);
2530 
2531 	iommu = amd_iommu_rlookup_table[devid];
2532 	if (!iommu)
2533 		return;
2534 
2535 #ifdef CONFIG_IRQ_REMAP
2536 	if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) &&
2537 	    (dom->type == IOMMU_DOMAIN_UNMANAGED))
2538 		dev_data->use_vapic = 0;
2539 #endif
2540 
2541 	iommu_completion_wait(iommu);
2542 }
2543 
amd_iommu_attach_device(struct iommu_domain * dom,struct device * dev)2544 static int amd_iommu_attach_device(struct iommu_domain *dom,
2545 				   struct device *dev)
2546 {
2547 	struct protection_domain *domain = to_pdomain(dom);
2548 	struct iommu_dev_data *dev_data;
2549 	struct amd_iommu *iommu;
2550 	int ret;
2551 
2552 	if (!check_device(dev))
2553 		return -EINVAL;
2554 
2555 	dev_data = dev_iommu_priv_get(dev);
2556 	dev_data->defer_attach = false;
2557 
2558 	iommu = amd_iommu_rlookup_table[dev_data->devid];
2559 	if (!iommu)
2560 		return -EINVAL;
2561 
2562 	if (dev_data->domain)
2563 		detach_device(dev);
2564 
2565 	ret = attach_device(dev, domain);
2566 
2567 #ifdef CONFIG_IRQ_REMAP
2568 	if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
2569 		if (dom->type == IOMMU_DOMAIN_UNMANAGED)
2570 			dev_data->use_vapic = 1;
2571 		else
2572 			dev_data->use_vapic = 0;
2573 	}
2574 #endif
2575 
2576 	iommu_completion_wait(iommu);
2577 
2578 	return ret;
2579 }
2580 
amd_iommu_map(struct iommu_domain * dom,unsigned long iova,phys_addr_t paddr,size_t page_size,int iommu_prot,gfp_t gfp)2581 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
2582 			 phys_addr_t paddr, size_t page_size, int iommu_prot,
2583 			 gfp_t gfp)
2584 {
2585 	struct protection_domain *domain = to_pdomain(dom);
2586 	struct domain_pgtable pgtable;
2587 	int prot = 0;
2588 	int ret;
2589 
2590 	amd_iommu_domain_get_pgtable(domain, &pgtable);
2591 	if (pgtable.mode == PAGE_MODE_NONE)
2592 		return -EINVAL;
2593 
2594 	if (iommu_prot & IOMMU_READ)
2595 		prot |= IOMMU_PROT_IR;
2596 	if (iommu_prot & IOMMU_WRITE)
2597 		prot |= IOMMU_PROT_IW;
2598 
2599 	ret = iommu_map_page(domain, iova, paddr, page_size, prot, gfp);
2600 
2601 	domain_flush_np_cache(domain, iova, page_size);
2602 
2603 	return ret;
2604 }
2605 
amd_iommu_unmap(struct iommu_domain * dom,unsigned long iova,size_t page_size,struct iommu_iotlb_gather * gather)2606 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
2607 			      size_t page_size,
2608 			      struct iommu_iotlb_gather *gather)
2609 {
2610 	struct protection_domain *domain = to_pdomain(dom);
2611 	struct domain_pgtable pgtable;
2612 
2613 	amd_iommu_domain_get_pgtable(domain, &pgtable);
2614 	if (pgtable.mode == PAGE_MODE_NONE)
2615 		return 0;
2616 
2617 	return iommu_unmap_page(domain, iova, page_size);
2618 }
2619 
amd_iommu_iova_to_phys(struct iommu_domain * dom,dma_addr_t iova)2620 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
2621 					  dma_addr_t iova)
2622 {
2623 	struct protection_domain *domain = to_pdomain(dom);
2624 	unsigned long offset_mask, pte_pgsize;
2625 	struct domain_pgtable pgtable;
2626 	u64 *pte, __pte;
2627 
2628 	amd_iommu_domain_get_pgtable(domain, &pgtable);
2629 	if (pgtable.mode == PAGE_MODE_NONE)
2630 		return iova;
2631 
2632 	pte = fetch_pte(domain, iova, &pte_pgsize);
2633 
2634 	if (!pte || !IOMMU_PTE_PRESENT(*pte))
2635 		return 0;
2636 
2637 	offset_mask = pte_pgsize - 1;
2638 	__pte	    = __sme_clr(*pte & PM_ADDR_MASK);
2639 
2640 	return (__pte & ~offset_mask) | (iova & offset_mask);
2641 }
2642 
amd_iommu_capable(enum iommu_cap cap)2643 static bool amd_iommu_capable(enum iommu_cap cap)
2644 {
2645 	switch (cap) {
2646 	case IOMMU_CAP_CACHE_COHERENCY:
2647 		return true;
2648 	case IOMMU_CAP_INTR_REMAP:
2649 		return (irq_remapping_enabled == 1);
2650 	case IOMMU_CAP_NOEXEC:
2651 		return false;
2652 	default:
2653 		break;
2654 	}
2655 
2656 	return false;
2657 }
2658 
amd_iommu_get_resv_regions(struct device * dev,struct list_head * head)2659 static void amd_iommu_get_resv_regions(struct device *dev,
2660 				       struct list_head *head)
2661 {
2662 	struct iommu_resv_region *region;
2663 	struct unity_map_entry *entry;
2664 	int devid;
2665 
2666 	devid = get_device_id(dev);
2667 	if (devid < 0)
2668 		return;
2669 
2670 	list_for_each_entry(entry, &amd_iommu_unity_map, list) {
2671 		int type, prot = 0;
2672 		size_t length;
2673 
2674 		if (devid < entry->devid_start || devid > entry->devid_end)
2675 			continue;
2676 
2677 		type   = IOMMU_RESV_DIRECT;
2678 		length = entry->address_end - entry->address_start;
2679 		if (entry->prot & IOMMU_PROT_IR)
2680 			prot |= IOMMU_READ;
2681 		if (entry->prot & IOMMU_PROT_IW)
2682 			prot |= IOMMU_WRITE;
2683 		if (entry->prot & IOMMU_UNITY_MAP_FLAG_EXCL_RANGE)
2684 			/* Exclusion range */
2685 			type = IOMMU_RESV_RESERVED;
2686 
2687 		region = iommu_alloc_resv_region(entry->address_start,
2688 						 length, prot, type);
2689 		if (!region) {
2690 			dev_err(dev, "Out of memory allocating dm-regions\n");
2691 			return;
2692 		}
2693 		list_add_tail(&region->list, head);
2694 	}
2695 
2696 	region = iommu_alloc_resv_region(MSI_RANGE_START,
2697 					 MSI_RANGE_END - MSI_RANGE_START + 1,
2698 					 0, IOMMU_RESV_MSI);
2699 	if (!region)
2700 		return;
2701 	list_add_tail(&region->list, head);
2702 
2703 	region = iommu_alloc_resv_region(HT_RANGE_START,
2704 					 HT_RANGE_END - HT_RANGE_START + 1,
2705 					 0, IOMMU_RESV_RESERVED);
2706 	if (!region)
2707 		return;
2708 	list_add_tail(&region->list, head);
2709 }
2710 
amd_iommu_is_attach_deferred(struct iommu_domain * domain,struct device * dev)2711 bool amd_iommu_is_attach_deferred(struct iommu_domain *domain,
2712 				  struct device *dev)
2713 {
2714 	struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2715 
2716 	return dev_data->defer_attach;
2717 }
2718 EXPORT_SYMBOL_GPL(amd_iommu_is_attach_deferred);
2719 
amd_iommu_flush_iotlb_all(struct iommu_domain * domain)2720 static void amd_iommu_flush_iotlb_all(struct iommu_domain *domain)
2721 {
2722 	struct protection_domain *dom = to_pdomain(domain);
2723 	unsigned long flags;
2724 
2725 	spin_lock_irqsave(&dom->lock, flags);
2726 	domain_flush_tlb_pde(dom);
2727 	domain_flush_complete(dom);
2728 	spin_unlock_irqrestore(&dom->lock, flags);
2729 }
2730 
amd_iommu_iotlb_sync(struct iommu_domain * domain,struct iommu_iotlb_gather * gather)2731 static void amd_iommu_iotlb_sync(struct iommu_domain *domain,
2732 				 struct iommu_iotlb_gather *gather)
2733 {
2734 	amd_iommu_flush_iotlb_all(domain);
2735 }
2736 
amd_iommu_def_domain_type(struct device * dev)2737 static int amd_iommu_def_domain_type(struct device *dev)
2738 {
2739 	struct iommu_dev_data *dev_data;
2740 
2741 	dev_data = dev_iommu_priv_get(dev);
2742 	if (!dev_data)
2743 		return 0;
2744 
2745 	/*
2746 	 * Do not identity map IOMMUv2 capable devices when memory encryption is
2747 	 * active, because some of those devices (AMD GPUs) don't have the
2748 	 * encryption bit in their DMA-mask and require remapping.
2749 	 */
2750 	if (!mem_encrypt_active() && dev_data->iommu_v2)
2751 		return IOMMU_DOMAIN_IDENTITY;
2752 
2753 	return 0;
2754 }
2755 
2756 const struct iommu_ops amd_iommu_ops = {
2757 	.capable = amd_iommu_capable,
2758 	.domain_alloc = amd_iommu_domain_alloc,
2759 	.domain_free  = amd_iommu_domain_free,
2760 	.attach_dev = amd_iommu_attach_device,
2761 	.detach_dev = amd_iommu_detach_device,
2762 	.map = amd_iommu_map,
2763 	.unmap = amd_iommu_unmap,
2764 	.iova_to_phys = amd_iommu_iova_to_phys,
2765 	.probe_device = amd_iommu_probe_device,
2766 	.release_device = amd_iommu_release_device,
2767 	.probe_finalize = amd_iommu_probe_finalize,
2768 	.device_group = amd_iommu_device_group,
2769 	.domain_get_attr = amd_iommu_domain_get_attr,
2770 	.get_resv_regions = amd_iommu_get_resv_regions,
2771 	.put_resv_regions = generic_iommu_put_resv_regions,
2772 	.is_attach_deferred = amd_iommu_is_attach_deferred,
2773 	.pgsize_bitmap	= AMD_IOMMU_PGSIZES,
2774 	.flush_iotlb_all = amd_iommu_flush_iotlb_all,
2775 	.iotlb_sync = amd_iommu_iotlb_sync,
2776 	.def_domain_type = amd_iommu_def_domain_type,
2777 };
2778 
2779 /*****************************************************************************
2780  *
2781  * The next functions do a basic initialization of IOMMU for pass through
2782  * mode
2783  *
2784  * In passthrough mode the IOMMU is initialized and enabled but not used for
2785  * DMA-API translation.
2786  *
2787  *****************************************************************************/
2788 
2789 /* IOMMUv2 specific functions */
amd_iommu_register_ppr_notifier(struct notifier_block * nb)2790 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
2791 {
2792 	return atomic_notifier_chain_register(&ppr_notifier, nb);
2793 }
2794 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
2795 
amd_iommu_unregister_ppr_notifier(struct notifier_block * nb)2796 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
2797 {
2798 	return atomic_notifier_chain_unregister(&ppr_notifier, nb);
2799 }
2800 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
2801 
amd_iommu_domain_direct_map(struct iommu_domain * dom)2802 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
2803 {
2804 	struct protection_domain *domain = to_pdomain(dom);
2805 	struct domain_pgtable pgtable;
2806 	unsigned long flags;
2807 
2808 	spin_lock_irqsave(&domain->lock, flags);
2809 
2810 	/* First save pgtable configuration*/
2811 	amd_iommu_domain_get_pgtable(domain, &pgtable);
2812 
2813 	/* Remove page-table from domain */
2814 	amd_iommu_domain_clr_pt_root(domain);
2815 
2816 	/* Make changes visible to IOMMUs */
2817 	update_domain(domain);
2818 
2819 	/* Page-table is not visible to IOMMU anymore, so free it */
2820 	free_pagetable(&pgtable);
2821 
2822 	spin_unlock_irqrestore(&domain->lock, flags);
2823 }
2824 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
2825 
amd_iommu_domain_enable_v2(struct iommu_domain * dom,int pasids)2826 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
2827 {
2828 	struct protection_domain *domain = to_pdomain(dom);
2829 	unsigned long flags;
2830 	int levels, ret;
2831 
2832 	if (pasids <= 0 || pasids > (PASID_MASK + 1))
2833 		return -EINVAL;
2834 
2835 	/* Number of GCR3 table levels required */
2836 	for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
2837 		levels += 1;
2838 
2839 	if (levels > amd_iommu_max_glx_val)
2840 		return -EINVAL;
2841 
2842 	spin_lock_irqsave(&domain->lock, flags);
2843 
2844 	/*
2845 	 * Save us all sanity checks whether devices already in the
2846 	 * domain support IOMMUv2. Just force that the domain has no
2847 	 * devices attached when it is switched into IOMMUv2 mode.
2848 	 */
2849 	ret = -EBUSY;
2850 	if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
2851 		goto out;
2852 
2853 	ret = -ENOMEM;
2854 	domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
2855 	if (domain->gcr3_tbl == NULL)
2856 		goto out;
2857 
2858 	domain->glx      = levels;
2859 	domain->flags   |= PD_IOMMUV2_MASK;
2860 
2861 	update_domain(domain);
2862 
2863 	ret = 0;
2864 
2865 out:
2866 	spin_unlock_irqrestore(&domain->lock, flags);
2867 
2868 	return ret;
2869 }
2870 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
2871 
__flush_pasid(struct protection_domain * domain,u32 pasid,u64 address,bool size)2872 static int __flush_pasid(struct protection_domain *domain, u32 pasid,
2873 			 u64 address, bool size)
2874 {
2875 	struct iommu_dev_data *dev_data;
2876 	struct iommu_cmd cmd;
2877 	int i, ret;
2878 
2879 	if (!(domain->flags & PD_IOMMUV2_MASK))
2880 		return -EINVAL;
2881 
2882 	build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
2883 
2884 	/*
2885 	 * IOMMU TLB needs to be flushed before Device TLB to
2886 	 * prevent device TLB refill from IOMMU TLB
2887 	 */
2888 	for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
2889 		if (domain->dev_iommu[i] == 0)
2890 			continue;
2891 
2892 		ret = iommu_queue_command(amd_iommus[i], &cmd);
2893 		if (ret != 0)
2894 			goto out;
2895 	}
2896 
2897 	/* Wait until IOMMU TLB flushes are complete */
2898 	domain_flush_complete(domain);
2899 
2900 	/* Now flush device TLBs */
2901 	list_for_each_entry(dev_data, &domain->dev_list, list) {
2902 		struct amd_iommu *iommu;
2903 		int qdep;
2904 
2905 		/*
2906 		   There might be non-IOMMUv2 capable devices in an IOMMUv2
2907 		 * domain.
2908 		 */
2909 		if (!dev_data->ats.enabled)
2910 			continue;
2911 
2912 		qdep  = dev_data->ats.qdep;
2913 		iommu = amd_iommu_rlookup_table[dev_data->devid];
2914 
2915 		build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
2916 				      qdep, address, size);
2917 
2918 		ret = iommu_queue_command(iommu, &cmd);
2919 		if (ret != 0)
2920 			goto out;
2921 	}
2922 
2923 	/* Wait until all device TLBs are flushed */
2924 	domain_flush_complete(domain);
2925 
2926 	ret = 0;
2927 
2928 out:
2929 
2930 	return ret;
2931 }
2932 
__amd_iommu_flush_page(struct protection_domain * domain,u32 pasid,u64 address)2933 static int __amd_iommu_flush_page(struct protection_domain *domain, u32 pasid,
2934 				  u64 address)
2935 {
2936 	return __flush_pasid(domain, pasid, address, false);
2937 }
2938 
amd_iommu_flush_page(struct iommu_domain * dom,u32 pasid,u64 address)2939 int amd_iommu_flush_page(struct iommu_domain *dom, u32 pasid,
2940 			 u64 address)
2941 {
2942 	struct protection_domain *domain = to_pdomain(dom);
2943 	unsigned long flags;
2944 	int ret;
2945 
2946 	spin_lock_irqsave(&domain->lock, flags);
2947 	ret = __amd_iommu_flush_page(domain, pasid, address);
2948 	spin_unlock_irqrestore(&domain->lock, flags);
2949 
2950 	return ret;
2951 }
2952 EXPORT_SYMBOL(amd_iommu_flush_page);
2953 
__amd_iommu_flush_tlb(struct protection_domain * domain,u32 pasid)2954 static int __amd_iommu_flush_tlb(struct protection_domain *domain, u32 pasid)
2955 {
2956 	return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
2957 			     true);
2958 }
2959 
amd_iommu_flush_tlb(struct iommu_domain * dom,u32 pasid)2960 int amd_iommu_flush_tlb(struct iommu_domain *dom, u32 pasid)
2961 {
2962 	struct protection_domain *domain = to_pdomain(dom);
2963 	unsigned long flags;
2964 	int ret;
2965 
2966 	spin_lock_irqsave(&domain->lock, flags);
2967 	ret = __amd_iommu_flush_tlb(domain, pasid);
2968 	spin_unlock_irqrestore(&domain->lock, flags);
2969 
2970 	return ret;
2971 }
2972 EXPORT_SYMBOL(amd_iommu_flush_tlb);
2973 
__get_gcr3_pte(u64 * root,int level,u32 pasid,bool alloc)2974 static u64 *__get_gcr3_pte(u64 *root, int level, u32 pasid, bool alloc)
2975 {
2976 	int index;
2977 	u64 *pte;
2978 
2979 	while (true) {
2980 
2981 		index = (pasid >> (9 * level)) & 0x1ff;
2982 		pte   = &root[index];
2983 
2984 		if (level == 0)
2985 			break;
2986 
2987 		if (!(*pte & GCR3_VALID)) {
2988 			if (!alloc)
2989 				return NULL;
2990 
2991 			root = (void *)get_zeroed_page(GFP_ATOMIC);
2992 			if (root == NULL)
2993 				return NULL;
2994 
2995 			*pte = iommu_virt_to_phys(root) | GCR3_VALID;
2996 		}
2997 
2998 		root = iommu_phys_to_virt(*pte & PAGE_MASK);
2999 
3000 		level -= 1;
3001 	}
3002 
3003 	return pte;
3004 }
3005 
__set_gcr3(struct protection_domain * domain,u32 pasid,unsigned long cr3)3006 static int __set_gcr3(struct protection_domain *domain, u32 pasid,
3007 		      unsigned long cr3)
3008 {
3009 	struct domain_pgtable pgtable;
3010 	u64 *pte;
3011 
3012 	amd_iommu_domain_get_pgtable(domain, &pgtable);
3013 	if (pgtable.mode != PAGE_MODE_NONE)
3014 		return -EINVAL;
3015 
3016 	pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3017 	if (pte == NULL)
3018 		return -ENOMEM;
3019 
3020 	*pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3021 
3022 	return __amd_iommu_flush_tlb(domain, pasid);
3023 }
3024 
__clear_gcr3(struct protection_domain * domain,u32 pasid)3025 static int __clear_gcr3(struct protection_domain *domain, u32 pasid)
3026 {
3027 	struct domain_pgtable pgtable;
3028 	u64 *pte;
3029 
3030 	amd_iommu_domain_get_pgtable(domain, &pgtable);
3031 	if (pgtable.mode != PAGE_MODE_NONE)
3032 		return -EINVAL;
3033 
3034 	pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3035 	if (pte == NULL)
3036 		return 0;
3037 
3038 	*pte = 0;
3039 
3040 	return __amd_iommu_flush_tlb(domain, pasid);
3041 }
3042 
amd_iommu_domain_set_gcr3(struct iommu_domain * dom,u32 pasid,unsigned long cr3)3043 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, u32 pasid,
3044 			      unsigned long cr3)
3045 {
3046 	struct protection_domain *domain = to_pdomain(dom);
3047 	unsigned long flags;
3048 	int ret;
3049 
3050 	spin_lock_irqsave(&domain->lock, flags);
3051 	ret = __set_gcr3(domain, pasid, cr3);
3052 	spin_unlock_irqrestore(&domain->lock, flags);
3053 
3054 	return ret;
3055 }
3056 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3057 
amd_iommu_domain_clear_gcr3(struct iommu_domain * dom,u32 pasid)3058 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, u32 pasid)
3059 {
3060 	struct protection_domain *domain = to_pdomain(dom);
3061 	unsigned long flags;
3062 	int ret;
3063 
3064 	spin_lock_irqsave(&domain->lock, flags);
3065 	ret = __clear_gcr3(domain, pasid);
3066 	spin_unlock_irqrestore(&domain->lock, flags);
3067 
3068 	return ret;
3069 }
3070 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3071 
amd_iommu_complete_ppr(struct pci_dev * pdev,u32 pasid,int status,int tag)3072 int amd_iommu_complete_ppr(struct pci_dev *pdev, u32 pasid,
3073 			   int status, int tag)
3074 {
3075 	struct iommu_dev_data *dev_data;
3076 	struct amd_iommu *iommu;
3077 	struct iommu_cmd cmd;
3078 
3079 	dev_data = dev_iommu_priv_get(&pdev->dev);
3080 	iommu    = amd_iommu_rlookup_table[dev_data->devid];
3081 
3082 	build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3083 			   tag, dev_data->pri_tlp);
3084 
3085 	return iommu_queue_command(iommu, &cmd);
3086 }
3087 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3088 
amd_iommu_get_v2_domain(struct pci_dev * pdev)3089 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3090 {
3091 	struct protection_domain *pdomain;
3092 	struct iommu_dev_data *dev_data;
3093 	struct device *dev = &pdev->dev;
3094 	struct iommu_domain *io_domain;
3095 
3096 	if (!check_device(dev))
3097 		return NULL;
3098 
3099 	dev_data  = dev_iommu_priv_get(&pdev->dev);
3100 	pdomain   = dev_data->domain;
3101 	io_domain = iommu_get_domain_for_dev(dev);
3102 
3103 	if (pdomain == NULL && dev_data->defer_attach) {
3104 		dev_data->defer_attach = false;
3105 		pdomain = to_pdomain(io_domain);
3106 		attach_device(dev, pdomain);
3107 	}
3108 
3109 	if (pdomain == NULL)
3110 		return NULL;
3111 
3112 	if (io_domain->type != IOMMU_DOMAIN_DMA)
3113 		return NULL;
3114 
3115 	/* Only return IOMMUv2 domains */
3116 	if (!(pdomain->flags & PD_IOMMUV2_MASK))
3117 		return NULL;
3118 
3119 	return &pdomain->domain;
3120 }
3121 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3122 
amd_iommu_enable_device_erratum(struct pci_dev * pdev,u32 erratum)3123 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3124 {
3125 	struct iommu_dev_data *dev_data;
3126 
3127 	if (!amd_iommu_v2_supported())
3128 		return;
3129 
3130 	dev_data = dev_iommu_priv_get(&pdev->dev);
3131 	dev_data->errata |= (1 << erratum);
3132 }
3133 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3134 
amd_iommu_device_info(struct pci_dev * pdev,struct amd_iommu_device_info * info)3135 int amd_iommu_device_info(struct pci_dev *pdev,
3136                           struct amd_iommu_device_info *info)
3137 {
3138 	int max_pasids;
3139 	int pos;
3140 
3141 	if (pdev == NULL || info == NULL)
3142 		return -EINVAL;
3143 
3144 	if (!amd_iommu_v2_supported())
3145 		return -EINVAL;
3146 
3147 	memset(info, 0, sizeof(*info));
3148 
3149 	if (pci_ats_supported(pdev))
3150 		info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3151 
3152 	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3153 	if (pos)
3154 		info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3155 
3156 	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3157 	if (pos) {
3158 		int features;
3159 
3160 		max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3161 		max_pasids = min(max_pasids, (1 << 20));
3162 
3163 		info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3164 		info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3165 
3166 		features = pci_pasid_features(pdev);
3167 		if (features & PCI_PASID_CAP_EXEC)
3168 			info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3169 		if (features & PCI_PASID_CAP_PRIV)
3170 			info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3171 	}
3172 
3173 	return 0;
3174 }
3175 EXPORT_SYMBOL(amd_iommu_device_info);
3176 
3177 #ifdef CONFIG_IRQ_REMAP
3178 
3179 /*****************************************************************************
3180  *
3181  * Interrupt Remapping Implementation
3182  *
3183  *****************************************************************************/
3184 
3185 static struct irq_chip amd_ir_chip;
3186 static DEFINE_SPINLOCK(iommu_table_lock);
3187 
set_dte_irq_entry(u16 devid,struct irq_remap_table * table)3188 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3189 {
3190 	u64 dte;
3191 
3192 	dte	= amd_iommu_dev_table[devid].data[2];
3193 	dte	&= ~DTE_IRQ_PHYS_ADDR_MASK;
3194 	dte	|= iommu_virt_to_phys(table->table);
3195 	dte	|= DTE_IRQ_REMAP_INTCTL;
3196 	dte	|= DTE_IRQ_TABLE_LEN;
3197 	dte	|= DTE_IRQ_REMAP_ENABLE;
3198 
3199 	amd_iommu_dev_table[devid].data[2] = dte;
3200 }
3201 
get_irq_table(u16 devid)3202 static struct irq_remap_table *get_irq_table(u16 devid)
3203 {
3204 	struct irq_remap_table *table;
3205 
3206 	if (WARN_ONCE(!amd_iommu_rlookup_table[devid],
3207 		      "%s: no iommu for devid %x\n", __func__, devid))
3208 		return NULL;
3209 
3210 	table = irq_lookup_table[devid];
3211 	if (WARN_ONCE(!table, "%s: no table for devid %x\n", __func__, devid))
3212 		return NULL;
3213 
3214 	return table;
3215 }
3216 
__alloc_irq_table(void)3217 static struct irq_remap_table *__alloc_irq_table(void)
3218 {
3219 	struct irq_remap_table *table;
3220 
3221 	table = kzalloc(sizeof(*table), GFP_KERNEL);
3222 	if (!table)
3223 		return NULL;
3224 
3225 	table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_KERNEL);
3226 	if (!table->table) {
3227 		kfree(table);
3228 		return NULL;
3229 	}
3230 	raw_spin_lock_init(&table->lock);
3231 
3232 	if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3233 		memset(table->table, 0,
3234 		       MAX_IRQS_PER_TABLE * sizeof(u32));
3235 	else
3236 		memset(table->table, 0,
3237 		       (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
3238 	return table;
3239 }
3240 
set_remap_table_entry(struct amd_iommu * iommu,u16 devid,struct irq_remap_table * table)3241 static void set_remap_table_entry(struct amd_iommu *iommu, u16 devid,
3242 				  struct irq_remap_table *table)
3243 {
3244 	irq_lookup_table[devid] = table;
3245 	set_dte_irq_entry(devid, table);
3246 	iommu_flush_dte(iommu, devid);
3247 }
3248 
set_remap_table_entry_alias(struct pci_dev * pdev,u16 alias,void * data)3249 static int set_remap_table_entry_alias(struct pci_dev *pdev, u16 alias,
3250 				       void *data)
3251 {
3252 	struct irq_remap_table *table = data;
3253 
3254 	irq_lookup_table[alias] = table;
3255 	set_dte_irq_entry(alias, table);
3256 
3257 	iommu_flush_dte(amd_iommu_rlookup_table[alias], alias);
3258 
3259 	return 0;
3260 }
3261 
alloc_irq_table(u16 devid,struct pci_dev * pdev)3262 static struct irq_remap_table *alloc_irq_table(u16 devid, struct pci_dev *pdev)
3263 {
3264 	struct irq_remap_table *table = NULL;
3265 	struct irq_remap_table *new_table = NULL;
3266 	struct amd_iommu *iommu;
3267 	unsigned long flags;
3268 	u16 alias;
3269 
3270 	spin_lock_irqsave(&iommu_table_lock, flags);
3271 
3272 	iommu = amd_iommu_rlookup_table[devid];
3273 	if (!iommu)
3274 		goto out_unlock;
3275 
3276 	table = irq_lookup_table[devid];
3277 	if (table)
3278 		goto out_unlock;
3279 
3280 	alias = amd_iommu_alias_table[devid];
3281 	table = irq_lookup_table[alias];
3282 	if (table) {
3283 		set_remap_table_entry(iommu, devid, table);
3284 		goto out_wait;
3285 	}
3286 	spin_unlock_irqrestore(&iommu_table_lock, flags);
3287 
3288 	/* Nothing there yet, allocate new irq remapping table */
3289 	new_table = __alloc_irq_table();
3290 	if (!new_table)
3291 		return NULL;
3292 
3293 	spin_lock_irqsave(&iommu_table_lock, flags);
3294 
3295 	table = irq_lookup_table[devid];
3296 	if (table)
3297 		goto out_unlock;
3298 
3299 	table = irq_lookup_table[alias];
3300 	if (table) {
3301 		set_remap_table_entry(iommu, devid, table);
3302 		goto out_wait;
3303 	}
3304 
3305 	table = new_table;
3306 	new_table = NULL;
3307 
3308 	if (pdev)
3309 		pci_for_each_dma_alias(pdev, set_remap_table_entry_alias,
3310 				       table);
3311 	else
3312 		set_remap_table_entry(iommu, devid, table);
3313 
3314 	if (devid != alias)
3315 		set_remap_table_entry(iommu, alias, table);
3316 
3317 out_wait:
3318 	iommu_completion_wait(iommu);
3319 
3320 out_unlock:
3321 	spin_unlock_irqrestore(&iommu_table_lock, flags);
3322 
3323 	if (new_table) {
3324 		kmem_cache_free(amd_iommu_irq_cache, new_table->table);
3325 		kfree(new_table);
3326 	}
3327 	return table;
3328 }
3329 
alloc_irq_index(u16 devid,int count,bool align,struct pci_dev * pdev)3330 static int alloc_irq_index(u16 devid, int count, bool align,
3331 			   struct pci_dev *pdev)
3332 {
3333 	struct irq_remap_table *table;
3334 	int index, c, alignment = 1;
3335 	unsigned long flags;
3336 	struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
3337 
3338 	if (!iommu)
3339 		return -ENODEV;
3340 
3341 	table = alloc_irq_table(devid, pdev);
3342 	if (!table)
3343 		return -ENODEV;
3344 
3345 	if (align)
3346 		alignment = roundup_pow_of_two(count);
3347 
3348 	raw_spin_lock_irqsave(&table->lock, flags);
3349 
3350 	/* Scan table for free entries */
3351 	for (index = ALIGN(table->min_index, alignment), c = 0;
3352 	     index < MAX_IRQS_PER_TABLE;) {
3353 		if (!iommu->irte_ops->is_allocated(table, index)) {
3354 			c += 1;
3355 		} else {
3356 			c     = 0;
3357 			index = ALIGN(index + 1, alignment);
3358 			continue;
3359 		}
3360 
3361 		if (c == count)	{
3362 			for (; c != 0; --c)
3363 				iommu->irte_ops->set_allocated(table, index - c + 1);
3364 
3365 			index -= count - 1;
3366 			goto out;
3367 		}
3368 
3369 		index++;
3370 	}
3371 
3372 	index = -ENOSPC;
3373 
3374 out:
3375 	raw_spin_unlock_irqrestore(&table->lock, flags);
3376 
3377 	return index;
3378 }
3379 
modify_irte_ga(u16 devid,int index,struct irte_ga * irte,struct amd_ir_data * data)3380 static int modify_irte_ga(u16 devid, int index, struct irte_ga *irte,
3381 			  struct amd_ir_data *data)
3382 {
3383 	bool ret;
3384 	struct irq_remap_table *table;
3385 	struct amd_iommu *iommu;
3386 	unsigned long flags;
3387 	struct irte_ga *entry;
3388 
3389 	iommu = amd_iommu_rlookup_table[devid];
3390 	if (iommu == NULL)
3391 		return -EINVAL;
3392 
3393 	table = get_irq_table(devid);
3394 	if (!table)
3395 		return -ENOMEM;
3396 
3397 	raw_spin_lock_irqsave(&table->lock, flags);
3398 
3399 	entry = (struct irte_ga *)table->table;
3400 	entry = &entry[index];
3401 
3402 	ret = cmpxchg_double(&entry->lo.val, &entry->hi.val,
3403 			     entry->lo.val, entry->hi.val,
3404 			     irte->lo.val, irte->hi.val);
3405 	/*
3406 	 * We use cmpxchg16 to atomically update the 128-bit IRTE,
3407 	 * and it cannot be updated by the hardware or other processors
3408 	 * behind us, so the return value of cmpxchg16 should be the
3409 	 * same as the old value.
3410 	 */
3411 	WARN_ON(!ret);
3412 
3413 	if (data)
3414 		data->ref = entry;
3415 
3416 	raw_spin_unlock_irqrestore(&table->lock, flags);
3417 
3418 	iommu_flush_irt(iommu, devid);
3419 	iommu_completion_wait(iommu);
3420 
3421 	return 0;
3422 }
3423 
modify_irte(u16 devid,int index,union irte * irte)3424 static int modify_irte(u16 devid, int index, union irte *irte)
3425 {
3426 	struct irq_remap_table *table;
3427 	struct amd_iommu *iommu;
3428 	unsigned long flags;
3429 
3430 	iommu = amd_iommu_rlookup_table[devid];
3431 	if (iommu == NULL)
3432 		return -EINVAL;
3433 
3434 	table = get_irq_table(devid);
3435 	if (!table)
3436 		return -ENOMEM;
3437 
3438 	raw_spin_lock_irqsave(&table->lock, flags);
3439 	table->table[index] = irte->val;
3440 	raw_spin_unlock_irqrestore(&table->lock, flags);
3441 
3442 	iommu_flush_irt(iommu, devid);
3443 	iommu_completion_wait(iommu);
3444 
3445 	return 0;
3446 }
3447 
free_irte(u16 devid,int index)3448 static void free_irte(u16 devid, int index)
3449 {
3450 	struct irq_remap_table *table;
3451 	struct amd_iommu *iommu;
3452 	unsigned long flags;
3453 
3454 	iommu = amd_iommu_rlookup_table[devid];
3455 	if (iommu == NULL)
3456 		return;
3457 
3458 	table = get_irq_table(devid);
3459 	if (!table)
3460 		return;
3461 
3462 	raw_spin_lock_irqsave(&table->lock, flags);
3463 	iommu->irte_ops->clear_allocated(table, index);
3464 	raw_spin_unlock_irqrestore(&table->lock, flags);
3465 
3466 	iommu_flush_irt(iommu, devid);
3467 	iommu_completion_wait(iommu);
3468 }
3469 
irte_prepare(void * entry,u32 delivery_mode,u32 dest_mode,u8 vector,u32 dest_apicid,int devid)3470 static void irte_prepare(void *entry,
3471 			 u32 delivery_mode, u32 dest_mode,
3472 			 u8 vector, u32 dest_apicid, int devid)
3473 {
3474 	union irte *irte = (union irte *) entry;
3475 
3476 	irte->val                = 0;
3477 	irte->fields.vector      = vector;
3478 	irte->fields.int_type    = delivery_mode;
3479 	irte->fields.destination = dest_apicid;
3480 	irte->fields.dm          = dest_mode;
3481 	irte->fields.valid       = 1;
3482 }
3483 
irte_ga_prepare(void * entry,u32 delivery_mode,u32 dest_mode,u8 vector,u32 dest_apicid,int devid)3484 static void irte_ga_prepare(void *entry,
3485 			    u32 delivery_mode, u32 dest_mode,
3486 			    u8 vector, u32 dest_apicid, int devid)
3487 {
3488 	struct irte_ga *irte = (struct irte_ga *) entry;
3489 
3490 	irte->lo.val                      = 0;
3491 	irte->hi.val                      = 0;
3492 	irte->lo.fields_remap.int_type    = delivery_mode;
3493 	irte->lo.fields_remap.dm          = dest_mode;
3494 	irte->hi.fields.vector            = vector;
3495 	irte->lo.fields_remap.destination = APICID_TO_IRTE_DEST_LO(dest_apicid);
3496 	irte->hi.fields.destination       = APICID_TO_IRTE_DEST_HI(dest_apicid);
3497 	irte->lo.fields_remap.valid       = 1;
3498 }
3499 
irte_activate(void * entry,u16 devid,u16 index)3500 static void irte_activate(void *entry, u16 devid, u16 index)
3501 {
3502 	union irte *irte = (union irte *) entry;
3503 
3504 	irte->fields.valid = 1;
3505 	modify_irte(devid, index, irte);
3506 }
3507 
irte_ga_activate(void * entry,u16 devid,u16 index)3508 static void irte_ga_activate(void *entry, u16 devid, u16 index)
3509 {
3510 	struct irte_ga *irte = (struct irte_ga *) entry;
3511 
3512 	irte->lo.fields_remap.valid = 1;
3513 	modify_irte_ga(devid, index, irte, NULL);
3514 }
3515 
irte_deactivate(void * entry,u16 devid,u16 index)3516 static void irte_deactivate(void *entry, u16 devid, u16 index)
3517 {
3518 	union irte *irte = (union irte *) entry;
3519 
3520 	irte->fields.valid = 0;
3521 	modify_irte(devid, index, irte);
3522 }
3523 
irte_ga_deactivate(void * entry,u16 devid,u16 index)3524 static void irte_ga_deactivate(void *entry, u16 devid, u16 index)
3525 {
3526 	struct irte_ga *irte = (struct irte_ga *) entry;
3527 
3528 	irte->lo.fields_remap.valid = 0;
3529 	modify_irte_ga(devid, index, irte, NULL);
3530 }
3531 
irte_set_affinity(void * entry,u16 devid,u16 index,u8 vector,u32 dest_apicid)3532 static void irte_set_affinity(void *entry, u16 devid, u16 index,
3533 			      u8 vector, u32 dest_apicid)
3534 {
3535 	union irte *irte = (union irte *) entry;
3536 
3537 	irte->fields.vector = vector;
3538 	irte->fields.destination = dest_apicid;
3539 	modify_irte(devid, index, irte);
3540 }
3541 
irte_ga_set_affinity(void * entry,u16 devid,u16 index,u8 vector,u32 dest_apicid)3542 static void irte_ga_set_affinity(void *entry, u16 devid, u16 index,
3543 				 u8 vector, u32 dest_apicid)
3544 {
3545 	struct irte_ga *irte = (struct irte_ga *) entry;
3546 
3547 	if (!irte->lo.fields_remap.guest_mode) {
3548 		irte->hi.fields.vector = vector;
3549 		irte->lo.fields_remap.destination =
3550 					APICID_TO_IRTE_DEST_LO(dest_apicid);
3551 		irte->hi.fields.destination =
3552 					APICID_TO_IRTE_DEST_HI(dest_apicid);
3553 		modify_irte_ga(devid, index, irte, NULL);
3554 	}
3555 }
3556 
3557 #define IRTE_ALLOCATED (~1U)
irte_set_allocated(struct irq_remap_table * table,int index)3558 static void irte_set_allocated(struct irq_remap_table *table, int index)
3559 {
3560 	table->table[index] = IRTE_ALLOCATED;
3561 }
3562 
irte_ga_set_allocated(struct irq_remap_table * table,int index)3563 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
3564 {
3565 	struct irte_ga *ptr = (struct irte_ga *)table->table;
3566 	struct irte_ga *irte = &ptr[index];
3567 
3568 	memset(&irte->lo.val, 0, sizeof(u64));
3569 	memset(&irte->hi.val, 0, sizeof(u64));
3570 	irte->hi.fields.vector = 0xff;
3571 }
3572 
irte_is_allocated(struct irq_remap_table * table,int index)3573 static bool irte_is_allocated(struct irq_remap_table *table, int index)
3574 {
3575 	union irte *ptr = (union irte *)table->table;
3576 	union irte *irte = &ptr[index];
3577 
3578 	return irte->val != 0;
3579 }
3580 
irte_ga_is_allocated(struct irq_remap_table * table,int index)3581 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
3582 {
3583 	struct irte_ga *ptr = (struct irte_ga *)table->table;
3584 	struct irte_ga *irte = &ptr[index];
3585 
3586 	return irte->hi.fields.vector != 0;
3587 }
3588 
irte_clear_allocated(struct irq_remap_table * table,int index)3589 static void irte_clear_allocated(struct irq_remap_table *table, int index)
3590 {
3591 	table->table[index] = 0;
3592 }
3593 
irte_ga_clear_allocated(struct irq_remap_table * table,int index)3594 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
3595 {
3596 	struct irte_ga *ptr = (struct irte_ga *)table->table;
3597 	struct irte_ga *irte = &ptr[index];
3598 
3599 	memset(&irte->lo.val, 0, sizeof(u64));
3600 	memset(&irte->hi.val, 0, sizeof(u64));
3601 }
3602 
get_devid(struct irq_alloc_info * info)3603 static int get_devid(struct irq_alloc_info *info)
3604 {
3605 	switch (info->type) {
3606 	case X86_IRQ_ALLOC_TYPE_IOAPIC:
3607 	case X86_IRQ_ALLOC_TYPE_IOAPIC_GET_PARENT:
3608 		return get_ioapic_devid(info->devid);
3609 	case X86_IRQ_ALLOC_TYPE_HPET:
3610 	case X86_IRQ_ALLOC_TYPE_HPET_GET_PARENT:
3611 		return get_hpet_devid(info->devid);
3612 	case X86_IRQ_ALLOC_TYPE_PCI_MSI:
3613 	case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
3614 		return get_device_id(msi_desc_to_dev(info->desc));
3615 	default:
3616 		WARN_ON_ONCE(1);
3617 		return -1;
3618 	}
3619 }
3620 
get_irq_domain_for_devid(struct irq_alloc_info * info,int devid)3621 static struct irq_domain *get_irq_domain_for_devid(struct irq_alloc_info *info,
3622 						   int devid)
3623 {
3624 	struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
3625 
3626 	if (!iommu)
3627 		return NULL;
3628 
3629 	switch (info->type) {
3630 	case X86_IRQ_ALLOC_TYPE_IOAPIC_GET_PARENT:
3631 	case X86_IRQ_ALLOC_TYPE_HPET_GET_PARENT:
3632 		return iommu->ir_domain;
3633 	default:
3634 		WARN_ON_ONCE(1);
3635 		return NULL;
3636 	}
3637 }
3638 
get_irq_domain(struct irq_alloc_info * info)3639 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
3640 {
3641 	int devid;
3642 
3643 	if (!info)
3644 		return NULL;
3645 
3646 	devid = get_devid(info);
3647 	if (devid < 0)
3648 		return NULL;
3649 	return get_irq_domain_for_devid(info, devid);
3650 }
3651 
3652 struct irq_remap_ops amd_iommu_irq_ops = {
3653 	.prepare		= amd_iommu_prepare,
3654 	.enable			= amd_iommu_enable,
3655 	.disable		= amd_iommu_disable,
3656 	.reenable		= amd_iommu_reenable,
3657 	.enable_faulting	= amd_iommu_enable_faulting,
3658 	.get_irq_domain		= get_irq_domain,
3659 };
3660 
irq_remapping_prepare_irte(struct amd_ir_data * data,struct irq_cfg * irq_cfg,struct irq_alloc_info * info,int devid,int index,int sub_handle)3661 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3662 				       struct irq_cfg *irq_cfg,
3663 				       struct irq_alloc_info *info,
3664 				       int devid, int index, int sub_handle)
3665 {
3666 	struct irq_2_irte *irte_info = &data->irq_2_irte;
3667 	struct msi_msg *msg = &data->msi_entry;
3668 	struct IO_APIC_route_entry *entry;
3669 	struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
3670 
3671 	if (!iommu)
3672 		return;
3673 
3674 	data->irq_2_irte.devid = devid;
3675 	data->irq_2_irte.index = index + sub_handle;
3676 	iommu->irte_ops->prepare(data->entry, apic->irq_delivery_mode,
3677 				 apic->irq_dest_mode, irq_cfg->vector,
3678 				 irq_cfg->dest_apicid, devid);
3679 
3680 	switch (info->type) {
3681 	case X86_IRQ_ALLOC_TYPE_IOAPIC:
3682 		/* Setup IOAPIC entry */
3683 		entry = info->ioapic.entry;
3684 		info->ioapic.entry = NULL;
3685 		memset(entry, 0, sizeof(*entry));
3686 		entry->vector        = index;
3687 		entry->mask          = 0;
3688 		entry->trigger       = info->ioapic.trigger;
3689 		entry->polarity      = info->ioapic.polarity;
3690 		/* Mask level triggered irqs. */
3691 		if (info->ioapic.trigger)
3692 			entry->mask = 1;
3693 		break;
3694 
3695 	case X86_IRQ_ALLOC_TYPE_HPET:
3696 	case X86_IRQ_ALLOC_TYPE_PCI_MSI:
3697 	case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
3698 		msg->address_hi = MSI_ADDR_BASE_HI;
3699 		msg->address_lo = MSI_ADDR_BASE_LO;
3700 		msg->data = irte_info->index;
3701 		break;
3702 
3703 	default:
3704 		BUG_ON(1);
3705 		break;
3706 	}
3707 }
3708 
3709 struct amd_irte_ops irte_32_ops = {
3710 	.prepare = irte_prepare,
3711 	.activate = irte_activate,
3712 	.deactivate = irte_deactivate,
3713 	.set_affinity = irte_set_affinity,
3714 	.set_allocated = irte_set_allocated,
3715 	.is_allocated = irte_is_allocated,
3716 	.clear_allocated = irte_clear_allocated,
3717 };
3718 
3719 struct amd_irte_ops irte_128_ops = {
3720 	.prepare = irte_ga_prepare,
3721 	.activate = irte_ga_activate,
3722 	.deactivate = irte_ga_deactivate,
3723 	.set_affinity = irte_ga_set_affinity,
3724 	.set_allocated = irte_ga_set_allocated,
3725 	.is_allocated = irte_ga_is_allocated,
3726 	.clear_allocated = irte_ga_clear_allocated,
3727 };
3728 
irq_remapping_alloc(struct irq_domain * domain,unsigned int virq,unsigned int nr_irqs,void * arg)3729 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
3730 			       unsigned int nr_irqs, void *arg)
3731 {
3732 	struct irq_alloc_info *info = arg;
3733 	struct irq_data *irq_data;
3734 	struct amd_ir_data *data = NULL;
3735 	struct irq_cfg *cfg;
3736 	int i, ret, devid;
3737 	int index;
3738 
3739 	if (!info)
3740 		return -EINVAL;
3741 	if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_PCI_MSI &&
3742 	    info->type != X86_IRQ_ALLOC_TYPE_PCI_MSIX)
3743 		return -EINVAL;
3744 
3745 	/*
3746 	 * With IRQ remapping enabled, don't need contiguous CPU vectors
3747 	 * to support multiple MSI interrupts.
3748 	 */
3749 	if (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI)
3750 		info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
3751 
3752 	devid = get_devid(info);
3753 	if (devid < 0)
3754 		return -EINVAL;
3755 
3756 	ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
3757 	if (ret < 0)
3758 		return ret;
3759 
3760 	if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
3761 		struct irq_remap_table *table;
3762 		struct amd_iommu *iommu;
3763 
3764 		table = alloc_irq_table(devid, NULL);
3765 		if (table) {
3766 			if (!table->min_index) {
3767 				/*
3768 				 * Keep the first 32 indexes free for IOAPIC
3769 				 * interrupts.
3770 				 */
3771 				table->min_index = 32;
3772 				iommu = amd_iommu_rlookup_table[devid];
3773 				for (i = 0; i < 32; ++i)
3774 					iommu->irte_ops->set_allocated(table, i);
3775 			}
3776 			WARN_ON(table->min_index != 32);
3777 			index = info->ioapic.pin;
3778 		} else {
3779 			index = -ENOMEM;
3780 		}
3781 	} else if (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI ||
3782 		   info->type == X86_IRQ_ALLOC_TYPE_PCI_MSIX) {
3783 		bool align = (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI);
3784 
3785 		index = alloc_irq_index(devid, nr_irqs, align,
3786 					msi_desc_to_pci_dev(info->desc));
3787 	} else {
3788 		index = alloc_irq_index(devid, nr_irqs, false, NULL);
3789 	}
3790 
3791 	if (index < 0) {
3792 		pr_warn("Failed to allocate IRTE\n");
3793 		ret = index;
3794 		goto out_free_parent;
3795 	}
3796 
3797 	for (i = 0; i < nr_irqs; i++) {
3798 		irq_data = irq_domain_get_irq_data(domain, virq + i);
3799 		cfg = irq_data ? irqd_cfg(irq_data) : NULL;
3800 		if (!cfg) {
3801 			ret = -EINVAL;
3802 			goto out_free_data;
3803 		}
3804 
3805 		ret = -ENOMEM;
3806 		data = kzalloc(sizeof(*data), GFP_KERNEL);
3807 		if (!data)
3808 			goto out_free_data;
3809 
3810 		if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3811 			data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
3812 		else
3813 			data->entry = kzalloc(sizeof(struct irte_ga),
3814 						     GFP_KERNEL);
3815 		if (!data->entry) {
3816 			kfree(data);
3817 			goto out_free_data;
3818 		}
3819 
3820 		irq_data->hwirq = (devid << 16) + i;
3821 		irq_data->chip_data = data;
3822 		irq_data->chip = &amd_ir_chip;
3823 		irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
3824 		irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
3825 	}
3826 
3827 	return 0;
3828 
3829 out_free_data:
3830 	for (i--; i >= 0; i--) {
3831 		irq_data = irq_domain_get_irq_data(domain, virq + i);
3832 		if (irq_data)
3833 			kfree(irq_data->chip_data);
3834 	}
3835 	for (i = 0; i < nr_irqs; i++)
3836 		free_irte(devid, index + i);
3837 out_free_parent:
3838 	irq_domain_free_irqs_common(domain, virq, nr_irqs);
3839 	return ret;
3840 }
3841 
irq_remapping_free(struct irq_domain * domain,unsigned int virq,unsigned int nr_irqs)3842 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
3843 			       unsigned int nr_irqs)
3844 {
3845 	struct irq_2_irte *irte_info;
3846 	struct irq_data *irq_data;
3847 	struct amd_ir_data *data;
3848 	int i;
3849 
3850 	for (i = 0; i < nr_irqs; i++) {
3851 		irq_data = irq_domain_get_irq_data(domain, virq  + i);
3852 		if (irq_data && irq_data->chip_data) {
3853 			data = irq_data->chip_data;
3854 			irte_info = &data->irq_2_irte;
3855 			free_irte(irte_info->devid, irte_info->index);
3856 			kfree(data->entry);
3857 			kfree(data);
3858 		}
3859 	}
3860 	irq_domain_free_irqs_common(domain, virq, nr_irqs);
3861 }
3862 
3863 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
3864 			       struct amd_ir_data *ir_data,
3865 			       struct irq_2_irte *irte_info,
3866 			       struct irq_cfg *cfg);
3867 
irq_remapping_activate(struct irq_domain * domain,struct irq_data * irq_data,bool reserve)3868 static int irq_remapping_activate(struct irq_domain *domain,
3869 				  struct irq_data *irq_data, bool reserve)
3870 {
3871 	struct amd_ir_data *data = irq_data->chip_data;
3872 	struct irq_2_irte *irte_info = &data->irq_2_irte;
3873 	struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
3874 	struct irq_cfg *cfg = irqd_cfg(irq_data);
3875 
3876 	if (!iommu)
3877 		return 0;
3878 
3879 	iommu->irte_ops->activate(data->entry, irte_info->devid,
3880 				  irte_info->index);
3881 	amd_ir_update_irte(irq_data, iommu, data, irte_info, cfg);
3882 	return 0;
3883 }
3884 
irq_remapping_deactivate(struct irq_domain * domain,struct irq_data * irq_data)3885 static void irq_remapping_deactivate(struct irq_domain *domain,
3886 				     struct irq_data *irq_data)
3887 {
3888 	struct amd_ir_data *data = irq_data->chip_data;
3889 	struct irq_2_irte *irte_info = &data->irq_2_irte;
3890 	struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
3891 
3892 	if (iommu)
3893 		iommu->irte_ops->deactivate(data->entry, irte_info->devid,
3894 					    irte_info->index);
3895 }
3896 
3897 static const struct irq_domain_ops amd_ir_domain_ops = {
3898 	.alloc = irq_remapping_alloc,
3899 	.free = irq_remapping_free,
3900 	.activate = irq_remapping_activate,
3901 	.deactivate = irq_remapping_deactivate,
3902 };
3903 
amd_iommu_activate_guest_mode(void * data)3904 int amd_iommu_activate_guest_mode(void *data)
3905 {
3906 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3907 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3908 	u64 valid;
3909 
3910 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3911 	    !entry || entry->lo.fields_vapic.guest_mode)
3912 		return 0;
3913 
3914 	valid = entry->lo.fields_vapic.valid;
3915 
3916 	entry->lo.val = 0;
3917 	entry->hi.val = 0;
3918 
3919 	entry->lo.fields_vapic.valid       = valid;
3920 	entry->lo.fields_vapic.guest_mode  = 1;
3921 	entry->lo.fields_vapic.ga_log_intr = 1;
3922 	entry->hi.fields.ga_root_ptr       = ir_data->ga_root_ptr;
3923 	entry->hi.fields.vector            = ir_data->ga_vector;
3924 	entry->lo.fields_vapic.ga_tag      = ir_data->ga_tag;
3925 
3926 	return modify_irte_ga(ir_data->irq_2_irte.devid,
3927 			      ir_data->irq_2_irte.index, entry, ir_data);
3928 }
3929 EXPORT_SYMBOL(amd_iommu_activate_guest_mode);
3930 
amd_iommu_deactivate_guest_mode(void * data)3931 int amd_iommu_deactivate_guest_mode(void *data)
3932 {
3933 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3934 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3935 	struct irq_cfg *cfg = ir_data->cfg;
3936 	u64 valid;
3937 
3938 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3939 	    !entry || !entry->lo.fields_vapic.guest_mode)
3940 		return 0;
3941 
3942 	valid = entry->lo.fields_remap.valid;
3943 
3944 	entry->lo.val = 0;
3945 	entry->hi.val = 0;
3946 
3947 	entry->lo.fields_remap.valid       = valid;
3948 	entry->lo.fields_remap.dm          = apic->irq_dest_mode;
3949 	entry->lo.fields_remap.int_type    = apic->irq_delivery_mode;
3950 	entry->hi.fields.vector            = cfg->vector;
3951 	entry->lo.fields_remap.destination =
3952 				APICID_TO_IRTE_DEST_LO(cfg->dest_apicid);
3953 	entry->hi.fields.destination =
3954 				APICID_TO_IRTE_DEST_HI(cfg->dest_apicid);
3955 
3956 	return modify_irte_ga(ir_data->irq_2_irte.devid,
3957 			      ir_data->irq_2_irte.index, entry, ir_data);
3958 }
3959 EXPORT_SYMBOL(amd_iommu_deactivate_guest_mode);
3960 
amd_ir_set_vcpu_affinity(struct irq_data * data,void * vcpu_info)3961 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
3962 {
3963 	int ret;
3964 	struct amd_iommu *iommu;
3965 	struct amd_iommu_pi_data *pi_data = vcpu_info;
3966 	struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
3967 	struct amd_ir_data *ir_data = data->chip_data;
3968 	struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3969 	struct iommu_dev_data *dev_data = search_dev_data(irte_info->devid);
3970 
3971 	/* Note:
3972 	 * This device has never been set up for guest mode.
3973 	 * we should not modify the IRTE
3974 	 */
3975 	if (!dev_data || !dev_data->use_vapic)
3976 		return 0;
3977 
3978 	ir_data->cfg = irqd_cfg(data);
3979 	pi_data->ir_data = ir_data;
3980 
3981 	/* Note:
3982 	 * SVM tries to set up for VAPIC mode, but we are in
3983 	 * legacy mode. So, we force legacy mode instead.
3984 	 */
3985 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3986 		pr_debug("%s: Fall back to using intr legacy remap\n",
3987 			 __func__);
3988 		pi_data->is_guest_mode = false;
3989 	}
3990 
3991 	iommu = amd_iommu_rlookup_table[irte_info->devid];
3992 	if (iommu == NULL)
3993 		return -EINVAL;
3994 
3995 	pi_data->prev_ga_tag = ir_data->cached_ga_tag;
3996 	if (pi_data->is_guest_mode) {
3997 		ir_data->ga_root_ptr = (pi_data->base >> 12);
3998 		ir_data->ga_vector = vcpu_pi_info->vector;
3999 		ir_data->ga_tag = pi_data->ga_tag;
4000 		ret = amd_iommu_activate_guest_mode(ir_data);
4001 		if (!ret)
4002 			ir_data->cached_ga_tag = pi_data->ga_tag;
4003 	} else {
4004 		ret = amd_iommu_deactivate_guest_mode(ir_data);
4005 
4006 		/*
4007 		 * This communicates the ga_tag back to the caller
4008 		 * so that it can do all the necessary clean up.
4009 		 */
4010 		if (!ret)
4011 			ir_data->cached_ga_tag = 0;
4012 	}
4013 
4014 	return ret;
4015 }
4016 
4017 
amd_ir_update_irte(struct irq_data * irqd,struct amd_iommu * iommu,struct amd_ir_data * ir_data,struct irq_2_irte * irte_info,struct irq_cfg * cfg)4018 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
4019 			       struct amd_ir_data *ir_data,
4020 			       struct irq_2_irte *irte_info,
4021 			       struct irq_cfg *cfg)
4022 {
4023 
4024 	/*
4025 	 * Atomically updates the IRTE with the new destination, vector
4026 	 * and flushes the interrupt entry cache.
4027 	 */
4028 	iommu->irte_ops->set_affinity(ir_data->entry, irte_info->devid,
4029 				      irte_info->index, cfg->vector,
4030 				      cfg->dest_apicid);
4031 }
4032 
amd_ir_set_affinity(struct irq_data * data,const struct cpumask * mask,bool force)4033 static int amd_ir_set_affinity(struct irq_data *data,
4034 			       const struct cpumask *mask, bool force)
4035 {
4036 	struct amd_ir_data *ir_data = data->chip_data;
4037 	struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4038 	struct irq_cfg *cfg = irqd_cfg(data);
4039 	struct irq_data *parent = data->parent_data;
4040 	struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4041 	int ret;
4042 
4043 	if (!iommu)
4044 		return -ENODEV;
4045 
4046 	ret = parent->chip->irq_set_affinity(parent, mask, force);
4047 	if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4048 		return ret;
4049 
4050 	amd_ir_update_irte(data, iommu, ir_data, irte_info, cfg);
4051 	/*
4052 	 * After this point, all the interrupts will start arriving
4053 	 * at the new destination. So, time to cleanup the previous
4054 	 * vector allocation.
4055 	 */
4056 	send_cleanup_vector(cfg);
4057 
4058 	return IRQ_SET_MASK_OK_DONE;
4059 }
4060 
ir_compose_msi_msg(struct irq_data * irq_data,struct msi_msg * msg)4061 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4062 {
4063 	struct amd_ir_data *ir_data = irq_data->chip_data;
4064 
4065 	*msg = ir_data->msi_entry;
4066 }
4067 
4068 static struct irq_chip amd_ir_chip = {
4069 	.name			= "AMD-IR",
4070 	.irq_ack		= apic_ack_irq,
4071 	.irq_set_affinity	= amd_ir_set_affinity,
4072 	.irq_set_vcpu_affinity	= amd_ir_set_vcpu_affinity,
4073 	.irq_compose_msi_msg	= ir_compose_msi_msg,
4074 };
4075 
amd_iommu_create_irq_domain(struct amd_iommu * iommu)4076 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4077 {
4078 	struct fwnode_handle *fn;
4079 
4080 	fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index);
4081 	if (!fn)
4082 		return -ENOMEM;
4083 	iommu->ir_domain = irq_domain_create_tree(fn, &amd_ir_domain_ops, iommu);
4084 	if (!iommu->ir_domain) {
4085 		irq_domain_free_fwnode(fn);
4086 		return -ENOMEM;
4087 	}
4088 
4089 	iommu->ir_domain->parent = arch_get_ir_parent_domain();
4090 	iommu->msi_domain = arch_create_remap_msi_irq_domain(iommu->ir_domain,
4091 							     "AMD-IR-MSI",
4092 							     iommu->index);
4093 	return 0;
4094 }
4095 
amd_iommu_update_ga(int cpu,bool is_run,void * data)4096 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
4097 {
4098 	unsigned long flags;
4099 	struct amd_iommu *iommu;
4100 	struct irq_remap_table *table;
4101 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
4102 	int devid = ir_data->irq_2_irte.devid;
4103 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
4104 	struct irte_ga *ref = (struct irte_ga *) ir_data->ref;
4105 
4106 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
4107 	    !ref || !entry || !entry->lo.fields_vapic.guest_mode)
4108 		return 0;
4109 
4110 	iommu = amd_iommu_rlookup_table[devid];
4111 	if (!iommu)
4112 		return -ENODEV;
4113 
4114 	table = get_irq_table(devid);
4115 	if (!table)
4116 		return -ENODEV;
4117 
4118 	raw_spin_lock_irqsave(&table->lock, flags);
4119 
4120 	if (ref->lo.fields_vapic.guest_mode) {
4121 		if (cpu >= 0) {
4122 			ref->lo.fields_vapic.destination =
4123 						APICID_TO_IRTE_DEST_LO(cpu);
4124 			ref->hi.fields.destination =
4125 						APICID_TO_IRTE_DEST_HI(cpu);
4126 		}
4127 		ref->lo.fields_vapic.is_run = is_run;
4128 		barrier();
4129 	}
4130 
4131 	raw_spin_unlock_irqrestore(&table->lock, flags);
4132 
4133 	iommu_flush_irt(iommu, devid);
4134 	iommu_completion_wait(iommu);
4135 	return 0;
4136 }
4137 EXPORT_SYMBOL(amd_iommu_update_ga);
4138 #endif
4139